{"meta":{"query_hash":"e5a7119d20ad","filters":{"topic":"Clinical Reasoning and Diagnostic Skills"},"cohort_total":1287,"direct_labels_cover":7,"predictions_cover":1287,"exported":1287,"export_cap":100000,"truncated":false,"label_status":"direct model label, unvalidated","prediction_status":"machine_predicted_unvalidated (Codex and Gemma teacher distillation)","score_status":"score_only:v0-immature-baseline","snapshot":{"source":"OpenAlex, pinned release, all 482 partitions","release":"2026-06-24","frame_built":"2026-07-12"},"permalink":"https://metacan.xera.ac/q/e5a7119d20ad","api":"https://metacan.xera.ac/api/v1/cohort?topic=Clinical+Reasoning+and+Diagnostic+Skills"},"results":[{"id":"W100721478","doi":"10.15173/m.v1i16.750","title":"Paramedicine: Recognition beyond Algorithms","year":2012,"lang":"en","type":"article","venue":"The Meducator","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Computer science; Artificial intelligence; Algorithm; Medicine","score_opus":0.04848592275399413,"score_gpt":0.3588096241521197,"score_spread":0.3103237013981256,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W100721478","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9697086,0.0017382413,0.000055020344,0.014360672,0.0019263794,0.0002568344,0.0000061698106,0.0000694056,0.011878672],"genre_scores_gemma":[0.98756635,0.000140006,0.0013670523,0.0061412216,0.0029185612,0.00004799344,0.000050117767,0.000018821413,0.0017498843],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99910504,0.000048741545,0.00019196133,0.000116207404,0.00022984738,0.0003082239],"domain_scores_gemma":[0.99492604,0.0042857113,0.0000547432,0.00030675708,0.000053767475,0.0003729656],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0006749878,0.000097819706,0.00018955587,0.00003142078,0.00006622069,0.0000075332496,0.00006950867,0.00006618172,0.0009663902],"category_scores_gemma":[0.017838309,0.000054648313,0.00007863757,0.0001415754,0.00012886882,0.000037266385,0.000020594827,0.0002254603,0.0014704667],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00024087257,0.0023665698,0.11400991,0.000049178583,0.00021986515,0.000019728674,0.0054386454,2.0975799e-7,0.0005767141,0.00095425197,0.22600305,0.65012103],"study_design_scores_gemma":[0.0053251726,0.0010357171,0.8319882,0.0023116334,0.0018134625,0.0012000926,0.0054799593,0.0000795654,0.009002581,0.008522803,0.13253279,0.00070800306],"about_ca_topic_score_codex":0.000044988097,"about_ca_topic_score_gemma":5.911942e-7,"teacher_disagreement_score":0.7179783,"about_ca_system_score_codex":0.00003302386,"about_ca_system_score_gemma":0.00010037555,"threshold_uncertainty_score":0.99994683},"labels":[],"label_agreement":null},{"id":"W1011831166","doi":"","title":"Structured Clinical Insights Modules (SCIMs): High fidelity simulated scenarios for learning clinical skills and clinical reasoning","year":2010,"lang":"en","type":"article","venue":"Griffith Research Online (Griffith University, Queensland, Australia)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medical education; Likert scale; Psychology; Fidelity; Medicine; Computer science; Developmental psychology","score_opus":0.10772475226821003,"score_gpt":0.46141773735457237,"score_spread":0.3536929850863623,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1011831166","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99223274,0.000080077065,0.00023853376,0.0029075812,0.0016769061,0.0020312718,0.00038251613,0.00033273242,0.00011764336],"genre_scores_gemma":[0.968046,0.0017403192,0.01880431,0.0003035573,0.0031268273,0.000004236102,0.0011396805,0.00011221651,0.0067228237],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9892091,0.002451374,0.0025387479,0.002405158,0.0015987208,0.0017968557],"domain_scores_gemma":[0.946558,0.044892233,0.00082688103,0.001642497,0.0029243985,0.0031559814],"candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.008918467,0.00075117167,0.0022888824,0.0006405332,0.0010178315,0.00020113222,0.00086571334,0.0023682818,0.00024560103],"category_scores_gemma":[0.10436967,0.00065298413,0.0010127705,0.001072807,0.0029117486,0.00031994568,0.00083447574,0.009551493,0.00007996418],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0042385394,0.0039158603,0.9494915,0.00012615236,0.0007611293,0.0007538564,0.0001263005,0.000273785,0.00008990795,0.0011206395,0.009974098,0.029128226],"study_design_scores_gemma":[0.01765614,0.0042609335,0.9438482,0.0011169276,0.0006247092,0.00008786359,0.00040747668,0.0075705554,0.000024110315,0.00073010154,0.022920405,0.0007525599],"about_ca_topic_score_codex":0.001751138,"about_ca_topic_score_gemma":0.004174719,"teacher_disagreement_score":0.0954512,"about_ca_system_score_codex":0.00014644944,"about_ca_system_score_gemma":0.0011047621,"threshold_uncertainty_score":0.99980175},"labels":[],"label_agreement":null},{"id":"W105830209","doi":"","title":"\"Perpetual Problem-Solving\": An Ethnographic Study of Clinical Reasoning in a Therapeutic Recreation Setting.","year":2002,"lang":"en","type":"article","venue":"Therapeutic Recreation Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Recreation; Ethnography; Psychology; Observational study; Clinical Practice; Rehabilitation; Recreational therapy; Psychotherapist; Applied psychology; Medicine; Nursing; Sociology","score_opus":0.11374179819146207,"score_gpt":0.41596626362365585,"score_spread":0.3022244654321938,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W105830209","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9930374,0.0017118065,0.0010589283,0.0006894862,0.0002767278,0.0009219793,0.0000016166372,0.00007780293,0.0022242258],"genre_scores_gemma":[0.9939627,0.0013209489,0.0034557527,0.00041664363,0.00044485365,0.000031980126,0.000013603735,0.00005337429,0.00030011256],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99454343,0.0015019617,0.0021895946,0.00049603113,0.0008196556,0.00044930156],"domain_scores_gemma":[0.9932593,0.0043359874,0.0009763749,0.0005500056,0.0004506628,0.00042765256],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005123071,0.0003053163,0.0009435193,0.0005211332,0.00019725757,0.00008582417,0.00023541476,0.00030476268,0.0006773208],"category_scores_gemma":[0.005893661,0.00024201708,0.00028618355,0.00085440953,0.00016673129,0.00027447284,0.000027279151,0.00124,0.000020046175],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007948623,0.005033388,0.659943,0.000015472777,0.0002918983,0.000054079483,0.008866097,0.0000966199,0.00032635636,0.00022306248,0.00005169408,0.32430348],"study_design_scores_gemma":[0.011766907,0.009139618,0.9534998,0.0024286192,0.00066767854,0.0004981407,0.00948927,0.009483541,0.00006268848,0.0023577956,0.00024239627,0.0003635599],"about_ca_topic_score_codex":0.00011834991,"about_ca_topic_score_gemma":0.000043741256,"teacher_disagreement_score":0.32393992,"about_ca_system_score_codex":0.000103784085,"about_ca_system_score_gemma":0.00019400928,"threshold_uncertainty_score":0.986917},"labels":[],"label_agreement":null},{"id":"W1161382390","doi":"10.1051/pmed/2015012","title":"Le potentiel diagnostique des questions d’un test de concordance de scripts pour évaluer le raisonnement clinique infirmier","year":2015,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Humanities; Concordance; Psychology; Philosophy; Medicine; Internal medicine","score_opus":0.07661595466340172,"score_gpt":0.3710953256722419,"score_spread":0.2944793710088402,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1161382390","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6476399,0.09720513,0.09625342,0.119150154,0.00459018,0.001647634,0.0003286995,0.0006661817,0.032518703],"genre_scores_gemma":[0.895386,0.010592575,0.03578562,0.010437442,0.002621135,0.00032893475,0.0000854868,0.00015371978,0.044609096],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.99480677,0.0007958013,0.0012108214,0.0008437487,0.0006410303,0.0017018574],"domain_scores_gemma":[0.97094655,0.024170255,0.0003494345,0.00095182675,0.0006960662,0.0028858436],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.003635019,0.000546475,0.0009313156,0.00013202518,0.0004279991,0.00013729064,0.00047648247,0.00075797236,0.0005359042],"category_scores_gemma":[0.2646522,0.00057490164,0.00041266318,0.00044793336,0.0016273139,0.0002247632,0.000306108,0.0012488151,0.00087731035],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00023707986,0.010972877,0.5512224,0.0002466949,0.0002977144,0.001933306,0.00276916,0.0005330625,0.00083972974,0.01209368,0.20750226,0.21135205],"study_design_scores_gemma":[0.021043187,0.0033286605,0.33109668,0.021065773,0.001411966,0.0012512333,0.0065688924,0.009004695,0.013040245,0.0686201,0.52150166,0.0020668758],"about_ca_topic_score_codex":0.007598747,"about_ca_topic_score_gemma":0.00034155796,"teacher_disagreement_score":0.3139994,"about_ca_system_score_codex":0.00063238683,"about_ca_system_score_gemma":0.010584134,"threshold_uncertainty_score":0.99990064},"labels":[],"label_agreement":null},{"id":"W116233781","doi":"10.1007/978-90-481-9501-5_12","title":"Introduction into Prognostic Clinical Research","year":2010,"lang":"en","type":"book-chapter","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Psychological intervention; Intervention (counseling); Term (time); Clinical trial; Medicine; Intensive care medicine; Psychology; Psychiatry; Pathology","score_opus":0.12215759656873637,"score_gpt":0.4679210192135133,"score_spread":0.34576342264477694,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W116233781","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0016326266,0.00038930282,0.00006751803,0.023137942,0.0028783805,0.0010117589,0.000004478057,0.00022364172,0.97065437],"genre_scores_gemma":[0.0042759,0.0009858462,0.0045958655,0.000518964,0.018837016,0.00002670371,0.00023051313,0.00010196212,0.9704272],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9966732,0.000071103816,0.0009832537,0.00093947846,0.00094753527,0.00038543044],"domain_scores_gemma":[0.9743864,0.02276112,0.00015615107,0.0011115102,0.00096531207,0.0006194843],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity","insufficient_payload"],"category_scores_codex":[0.003501457,0.0002857544,0.00082127604,0.00025725216,0.0001246339,0.000040530285,0.00017173715,0.00157782,0.0113180475],"category_scores_gemma":[0.17215097,0.00021851316,0.0003840174,0.000064423926,0.0009076498,0.000026333453,0.00017903549,0.0048753736,0.007834455],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00047649935,0.0006220981,0.004065813,0.00012527224,0.0002929951,0.0003348958,0.000041940573,9.180057e-8,0.00005721495,0.21863309,0.5492304,0.22611971],"study_design_scores_gemma":[0.00088341534,0.0014145097,0.002854937,0.0015147328,0.0002873554,0.00010222744,0.0000101106125,0.0000075542407,0.000041559517,0.043108605,0.9495209,0.00025412822],"about_ca_topic_score_codex":0.000047810274,"about_ca_topic_score_gemma":0.000052841224,"teacher_disagreement_score":0.4002905,"about_ca_system_score_codex":0.00007403717,"about_ca_system_score_gemma":0.00060656585,"threshold_uncertainty_score":0.99971837},"labels":[],"label_agreement":null},{"id":"W1223080694","doi":"10.1167/15.12.1131","title":"The neural correlates of medical expertise.","year":2015,"lang":"en","type":"article","venue":"Journal of Vision","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia; University of Victoria; University of Alberta","funders":"","keywords":"Curran; Pulmonologists; Categorization; Medicine; Psychology; Audiology; Artificial intelligence; Computer science","score_opus":0.04134611033713565,"score_gpt":0.40333327534853064,"score_spread":0.361987165011395,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1223080694","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9809767,0.003926199,0.00009743858,0.013541355,0.00090958155,0.000034434128,2.3442597e-7,0.000003921088,0.0005101629],"genre_scores_gemma":[0.9984888,0.0006813688,0.0001229934,0.0002690979,0.0003167102,2.0794441e-7,3.726505e-7,0.0000050291114,0.00011546303],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9983787,0.000050807324,0.00048505922,0.000043870543,0.000953656,0.00008792693],"domain_scores_gemma":[0.99431425,0.0046747145,0.00025281668,0.000112592636,0.00028628847,0.00035930765],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0013938067,0.000046833444,0.00021474583,0.000029936838,0.000022910259,0.000009475383,0.00012164742,0.00008290386,0.000058959464],"category_scores_gemma":[0.075229995,0.000021802212,0.00012047081,0.00006622442,0.00009407474,0.00003273552,0.00003504829,0.00027007828,0.000009537757],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002999838,0.0012204964,0.15227659,0.000018096656,0.00014351019,0.00091021875,0.0013357259,0.00007239238,0.00065551075,0.00050034595,0.2529469,0.5869204],"study_design_scores_gemma":[0.02098721,0.018632876,0.8691577,0.014680294,0.00040857564,0.006145353,0.0024275517,0.03227428,0.0015921497,0.0031407021,0.0302858,0.00026751967],"about_ca_topic_score_codex":0.000007729437,"about_ca_topic_score_gemma":7.535484e-7,"teacher_disagreement_score":0.7168811,"about_ca_system_score_codex":0.000020626952,"about_ca_system_score_gemma":0.00019656555,"threshold_uncertainty_score":0.9325597},"labels":[],"label_agreement":null},{"id":"W122514425","doi":"10.1016/s0028-3770(04)98309-9","title":"Auto-évaluation des internes en neurochirurgie par tests de concordance de script (TCS)","year":2004,"lang":"fr","type":"article","venue":"Neurochirurgie","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Concordance; Elaboration; Competence (human resources); Neurosurgery; Medicine; Medical physics; Psychology; Radiology; Social psychology; Humanities; Internal medicine","score_opus":0.035057131117637454,"score_gpt":0.34102517100163143,"score_spread":0.30596803988399396,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W122514425","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9570302,0.01237512,0.006995137,0.009794989,0.0067882445,0.0006892873,0.000048842863,0.00043486958,0.005843306],"genre_scores_gemma":[0.96994257,0.006781298,0.006543568,0.008399604,0.0031364409,0.00008799608,0.000044084823,0.00017611014,0.004888301],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99464685,0.0007105327,0.0010919179,0.0010453339,0.00073109777,0.0017742722],"domain_scores_gemma":[0.9842829,0.012961189,0.00034845274,0.0008996772,0.00028129874,0.0012264883],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0015513786,0.0006865487,0.00089119177,0.00020564283,0.00029662318,0.00017990114,0.00055430504,0.00050179654,0.00043131877],"category_scores_gemma":[0.10396588,0.0007194877,0.0004459491,0.00060450967,0.0009029219,0.00029029546,0.00022649702,0.0013378396,0.001104437],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0015609304,0.0060553695,0.26506227,0.0012451754,0.0004769285,0.010054188,0.0087946495,0.017486712,0.046857126,0.0118294135,0.025029104,0.60554814],"study_design_scores_gemma":[0.01035288,0.0030791538,0.88313305,0.016484624,0.0012522056,0.002231757,0.00032591092,0.006587339,0.012144236,0.033659063,0.029622125,0.0011276603],"about_ca_topic_score_codex":0.0012927582,"about_ca_topic_score_gemma":0.00016956225,"teacher_disagreement_score":0.6180708,"about_ca_system_score_codex":0.00083805836,"about_ca_system_score_gemma":0.002328536,"threshold_uncertainty_score":0.9996733},"labels":[],"label_agreement":null},{"id":"W123476994","doi":"10.1017/s1481803500009891","title":"Incorporation bias in studies of diagnostic tests: how to avoid being biased about bias","year":2008,"lang":"en","type":"review","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":105,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Action (physics)","score_opus":0.38847064668518627,"score_gpt":0.46501192703154776,"score_spread":0.07654128034636148,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W123476994","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00406207,0.9856868,0.000019933486,0.0030162886,0.0058442703,0.00084402005,0.00004045296,0.000008275162,0.00047793335],"genre_scores_gemma":[0.004853368,0.99169135,0.00018623435,0.00022608158,0.0023560845,0.00002888514,0.00007578224,0.000074804426,0.00050743093],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9943147,0.0003940591,0.0035879938,0.000396727,0.00073676545,0.0005697558],"domain_scores_gemma":[0.96504796,0.02781245,0.002708977,0.0006010105,0.0015684834,0.0022611392],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.002059035,0.0005769265,0.004469772,0.0023127014,0.00009257107,0.0000056513213,0.00036555214,0.00036754247,0.00039009211],"category_scores_gemma":[0.58430207,0.00040456653,0.0006506002,0.0021441756,0.00030461597,0.00008588229,0.000028178016,0.0010475707,0.000019014342],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006422554,0.00028819704,0.04203174,0.013381369,0.0014203084,0.00676634,0.003500677,0.000031148287,0.0000037441532,0.0001534529,0.29804164,0.63431716],"study_design_scores_gemma":[0.0013559066,0.0021354503,0.0039762263,0.363072,0.0019138973,0.00037884476,0.0004840497,0.0000021782535,0.0000017764747,0.0001735333,0.62611634,0.00038978725],"about_ca_topic_score_codex":0.0037423086,"about_ca_topic_score_gemma":0.018040242,"teacher_disagreement_score":0.63392735,"about_ca_system_score_codex":0.000601894,"about_ca_system_score_gemma":0.005175557,"threshold_uncertainty_score":0.999878},"labels":[],"label_agreement":null},{"id":"W127775744","doi":"10.1017/s1481803500007910","title":"Disaster Medicine Online: evaluation of an online, modular, interactive, asynchronous curriculum","year":2002,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Facilitator; Distance education; Medical education; Curriculum; Usability; Medicine; Asynchronous learning; Health care; Disaster medicine; Multidisciplinary approach; Professional development; Nursing; Psychology; Medical emergency; Teaching method; Poison control; Suicide prevention; Pedagogy; Computer science; Synchronous learning","score_opus":0.15728296320276886,"score_gpt":0.41609097124757305,"score_spread":0.2588080080448042,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W127775744","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97449553,0.007848808,0.00013103428,0.007342583,0.006347013,0.0003090033,0.000042425185,0.000009286966,0.0034743017],"genre_scores_gemma":[0.99393225,0.0010241376,0.00040424254,0.00043949095,0.003660403,0.0000031173759,0.00013446102,0.0000372944,0.00036461407],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9961267,0.00023954251,0.0018060709,0.00027940253,0.0011425663,0.00040572428],"domain_scores_gemma":[0.9936918,0.00066387176,0.0008684567,0.0004746541,0.0024787032,0.0018225205],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0015180059,0.00027681064,0.00094993855,0.00060743705,0.00006842343,0.000002618637,0.0002787073,0.00015592868,0.026158366],"category_scores_gemma":[0.052779112,0.00019273665,0.00019830173,0.00058863946,0.00032746725,0.00019052546,0.0000118206535,0.00059180456,0.000013669512],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00043047484,0.0062038847,0.2249498,0.0003953346,0.0018086071,0.0014298167,0.021074258,0.0014126678,0.0031424607,0.00032411635,0.31259823,0.42623034],"study_design_scores_gemma":[0.033929095,0.02828223,0.73904693,0.03239801,0.010055997,0.0015976419,0.023901913,0.09865577,0.00014309159,0.002587603,0.028122414,0.0012793149],"about_ca_topic_score_codex":0.007878627,"about_ca_topic_score_gemma":0.014053433,"teacher_disagreement_score":0.5140971,"about_ca_system_score_codex":0.00026418973,"about_ca_system_score_gemma":0.00065930904,"threshold_uncertainty_score":0.998728},"labels":[],"label_agreement":null},{"id":"W134948774","doi":"10.1007/978-90-481-9501-5_19","title":"Some Elementary Concepts of Medicine, According to the Students","year":2010,"lang":"en","type":"book-chapter","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Set (abstract data type); Elementary mathematics; Mathematics education; Variance (accounting); Medical education; Computer science; Medicine; Psychology","score_opus":0.0375066956050779,"score_gpt":0.3974712309132003,"score_spread":0.3599645353081224,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W134948774","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.018784186,0.0018107452,0.00008648447,0.030581476,0.0036497184,0.002274178,0.000076407276,0.00012427644,0.9426125],"genre_scores_gemma":[0.118208475,0.0011931476,0.00094381213,0.032299645,0.005568866,0.000030622836,0.00018118633,0.000107631546,0.8414666],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9978623,0.000011745467,0.0006792325,0.0003462561,0.00088643096,0.00021404604],"domain_scores_gemma":[0.993146,0.005519919,0.00022281749,0.0006669247,0.00015352714,0.00029081153],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0005795373,0.00026965904,0.0007842944,0.000102199665,0.000048919464,0.000007986024,0.00037168417,0.0002827318,0.0038714972],"category_scores_gemma":[0.008023108,0.00014725394,0.00018492772,0.000030112782,0.00030337513,0.00001858118,0.00022879885,0.00067281094,0.00030817802],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00087919954,0.00060737436,0.046124216,0.00026474305,0.0024495537,0.00038752417,0.0013307695,0.000002493165,0.0012777129,0.2779421,0.49160478,0.17712954],"study_design_scores_gemma":[0.009098773,0.003967641,0.0144419605,0.022377115,0.0028739555,0.000087478686,0.0011975178,0.0000044053313,0.00090077927,0.022770656,0.92139256,0.00088715804],"about_ca_topic_score_codex":0.00018992125,"about_ca_topic_score_gemma":0.00005010545,"teacher_disagreement_score":0.42978778,"about_ca_system_score_codex":0.000033816686,"about_ca_system_score_gemma":0.00009823842,"threshold_uncertainty_score":0.9970391},"labels":[],"label_agreement":null},{"id":"W1448285687","doi":"10.1177/0004867415597073","title":"The Neuropsychiatric Effects of CNS Structural Abnormalities (NECSA) classification: An aid to differential diagnosis","year":2015,"lang":"en","type":"letter","venue":"Australian & New Zealand Journal of Psychiatry","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Differential diagnosis; Medicine; Neuroscience; Psychology; Pathology","score_opus":0.03203013058594898,"score_gpt":0.32875882672114126,"score_spread":0.29672869613519226,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1448285687","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.34809166,0.0007811678,0.000009009628,0.6388267,0.011693581,0.000436979,0.000052447234,0.000023555416,0.00008488709],"genre_scores_gemma":[0.41020104,0.0038714544,0.0075200847,0.3279317,0.17707615,0.000083555824,0.00070300844,0.0005921225,0.07202088],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.995644,0.00035388226,0.0016860876,0.0004081078,0.0012700758,0.0006378515],"domain_scores_gemma":[0.9928151,0.0033416264,0.0015252993,0.0008338972,0.0005076132,0.000976497],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0004369885,0.00052995945,0.0011985316,0.00033723135,0.00014381587,0.00014316698,0.00071718136,0.000689767,0.0001425058],"category_scores_gemma":[0.0032577303,0.00033271892,0.00067963026,0.00038216214,0.00017625737,0.00013104558,0.000052213163,0.0024707015,0.00002367904],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00046621304,0.0001602278,0.11246345,0.00016870526,0.00030503198,0.00013290757,0.00016871122,0.0000041634426,0.000006070993,0.000053020645,0.8829052,0.0031663335],"study_design_scores_gemma":[0.0055636833,0.0074153985,0.46383274,0.0037685,0.0023419685,0.0010597348,0.00017494414,0.0000034074556,0.000034416742,0.0027170617,0.51256984,0.00051830756],"about_ca_topic_score_codex":0.0002195263,"about_ca_topic_score_gemma":0.000057791494,"teacher_disagreement_score":0.3703353,"about_ca_system_score_codex":0.00007536899,"about_ca_system_score_gemma":0.0013491912,"threshold_uncertainty_score":0.9999125},"labels":[],"label_agreement":null},{"id":"W1453328089","doi":"10.71781/5319","title":"Le test de concordance de script (TCS) explore-t-il des processus de raisonnement clinique différents de ceux du QCM à contexte riche (basés sur des cas cliniques) ? : et si oui, quels sont-ils ?","year":2007,"lang":"fr","type":"dissertation","venue":"Open MIND","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Humanities; Philosophy","score_opus":0.08185480570175957,"score_gpt":0.4178973648667493,"score_spread":0.3360425591649897,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1453328089","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9679266,0.004752019,0.009106996,0.00084072415,0.0009901002,0.0028789854,0.00039751563,0.000045546858,0.0130614685],"genre_scores_gemma":[0.92843354,0.0066150953,0.048774965,0.0028936162,0.0010550162,0.00072998233,0.0011858493,0.0002592004,0.010052729],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9897249,0.0012490696,0.0029305485,0.0017187373,0.000833084,0.0035436673],"domain_scores_gemma":[0.9541354,0.038477648,0.0018737572,0.0010207007,0.0015603453,0.0029321485],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.014521392,0.0012507514,0.002433371,0.0002338708,0.00078533444,0.0005892751,0.0014739814,0.0015688698,0.0012670561],"category_scores_gemma":[0.23128976,0.0012677913,0.0006110389,0.0005501078,0.0010235971,0.0005458745,0.00032162585,0.0020146845,0.00032454132],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0023583893,0.00515259,0.6503411,0.0008606674,0.00045628907,0.0017535784,0.04100266,0.00010308762,0.0027377452,0.00017344393,0.0017060556,0.29335442],"study_design_scores_gemma":[0.022979673,0.005253243,0.6932206,0.07382053,0.0030462209,0.00064131373,0.061446667,0.0030453308,0.09771777,0.004676123,0.030918289,0.0032342367],"about_ca_topic_score_codex":0.027192773,"about_ca_topic_score_gemma":0.06877496,"teacher_disagreement_score":0.29012018,"about_ca_system_score_codex":0.0016597222,"about_ca_system_score_gemma":0.016835852,"threshold_uncertainty_score":0.9997273},"labels":[],"label_agreement":null},{"id":"W1485548027","doi":"","title":"Commentary Communication in the Toronto critical care community: important lessons learned during SARS","year":2003,"lang":"en","type":"article","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Staffing; Critically ill; Economic shortage; Public relations; Coronavirus disease 2019 (COVID-19); Nursing; Medicine; Political science; Intensive care medicine","score_opus":0.08355465391834051,"score_gpt":0.42506891656914886,"score_spread":0.34151426265080836,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1485548027","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.90707004,0.0020848003,0.000010882275,0.054286137,0.000061587554,0.0002698749,0.0000069918797,0.00005003304,0.03615965],"genre_scores_gemma":[0.9882922,0.00065158214,0.0007356431,0.010119215,0.000019948215,0.000028826062,0.00007889822,0.000012814216,0.00006086702],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99826086,0.00081162446,0.00036122033,0.00012434657,0.00020940282,0.00023257954],"domain_scores_gemma":[0.98816216,0.010694128,0.00003675885,0.00094215747,0.00006384621,0.00010092995],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010350436,0.000111343084,0.00022880337,0.000020223992,0.00027307044,0.000027876948,0.00020894084,0.000085588974,0.00024694676],"category_scores_gemma":[0.018607335,0.00007684784,0.00007822836,0.00008935016,0.00014835132,0.000059302496,0.00006447701,0.0006112444,0.000010470745],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004506399,0.005852742,0.81966084,0.00031757573,0.00012813034,0.00029759682,0.045399502,0.000009195856,0.00083955505,0.085562184,0.027338888,0.014143125],"study_design_scores_gemma":[0.006818062,0.00074922596,0.7418327,0.0023254761,0.00030416748,0.00023563346,0.22977893,0.0000670573,0.0018172407,0.0043563554,0.011240865,0.00047430964],"about_ca_topic_score_codex":0.008883945,"about_ca_topic_score_gemma":0.0094789835,"teacher_disagreement_score":0.18437943,"about_ca_system_score_codex":0.00016794373,"about_ca_system_score_gemma":0.000051269828,"threshold_uncertainty_score":0.997716},"labels":[],"label_agreement":null},{"id":"W1488584250","doi":"10.1111/j.1365-2923.2010.03896.x","title":"Role of clinical context in residents’ physical examination diagnostic accuracy","year":2011,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":37,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Western Hospital; University of Toronto; University Health Network","funders":"","keywords":"Context (archaeology); Physical examination; Medical diagnosis; Medicine; Medical history; Diagnostic accuracy; United States Medical Licensing Examination; Checklist; Physical therapy; Surgery; Internal medicine; Psychology; Radiology; Medical education","score_opus":0.04642157524422292,"score_gpt":0.42068884349860186,"score_spread":0.37426726825437895,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1488584250","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9883755,0.00040835707,0.000031299773,0.0010040333,0.00048790794,0.00029765954,0.0000012153739,0.000024395074,0.00936963],"genre_scores_gemma":[0.9973759,0.00023497095,0.00023139002,0.0014055131,0.00047956195,0.00005497035,0.000041764728,0.000012720323,0.00016318189],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9979163,0.0002377625,0.00082328287,0.00027351378,0.00057388935,0.00017527651],"domain_scores_gemma":[0.9718634,0.026885381,0.0002767422,0.00036381022,0.00022992538,0.00038075823],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001451552,0.00010360575,0.00045710852,0.00010226177,0.00001583958,0.000003984932,0.0001284659,0.00020318931,0.0007612048],"category_scores_gemma":[0.53618807,0.00008642368,0.0001243473,0.00023121972,0.00020062325,0.00006961863,0.00003744718,0.00038145535,0.00011056927],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007759483,0.00440271,0.42143375,0.00002075295,0.000011672362,0.000007807362,0.0012720654,5.1153314e-8,0.000012022162,0.0013302311,0.0014672487,0.5699641],"study_design_scores_gemma":[0.0013789037,0.00034724467,0.9928433,0.0017770333,0.00006112121,0.000010646152,0.0006947279,0.0003773648,0.0003226,0.001662922,0.00044685559,0.000077291064],"about_ca_topic_score_codex":0.0010330582,"about_ca_topic_score_gemma":0.00006652476,"teacher_disagreement_score":0.5714095,"about_ca_system_score_codex":0.00003735129,"about_ca_system_score_gemma":0.0013398366,"threshold_uncertainty_score":0.83346593},"labels":[],"label_agreement":null},{"id":"W1490393546","doi":"10.1007/978-1-59745-008-9_3","title":"The Philosophy of Evidence-Based Medicine","year":2007,"lang":"en","type":"book-chapter","venue":"Contemporary Endocrinology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":61,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Clinical Practice; Alternative medicine; Medical education; Evidence-based medicine; Best evidence; Engineering ethics; Medical literature; Medicine; Psychology; Family medicine; Engineering; Pathology","score_opus":0.336658441712103,"score_gpt":0.40064888766814083,"score_spread":0.06399044595603781,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1490393546","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00009206202,0.07455109,0.00007661841,0.034063403,0.0010879431,0.00066990085,0.000018882309,0.00010097119,0.88933915],"genre_scores_gemma":[0.41382623,0.014733719,0.00019948791,0.009002945,0.0029138078,0.000039373354,0.00021624964,0.00016423927,0.55890393],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.9972944,0.000060220147,0.0012227951,0.0005090533,0.00051635935,0.0003971268],"domain_scores_gemma":[0.9616727,0.035971098,0.00066579867,0.0010451454,0.000401933,0.00024327831],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0005890001,0.00040986217,0.0012786237,0.00025956362,0.000087254615,0.0000044028407,0.00033966338,0.00044712392,0.00053994043],"category_scores_gemma":[0.02429453,0.00022397727,0.0003824914,0.00005813276,0.0021738077,0.00003021139,0.000101950645,0.00089090894,0.00013883006],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0043002944,0.00019485246,0.005412273,0.00041633545,0.00097477797,0.0038939079,0.00006214492,0.0000010344437,0.00007284977,0.70924175,0.14927267,0.12615715],"study_design_scores_gemma":[0.003889695,0.0036822073,0.0015157777,0.023069661,0.00046599616,0.00022295387,0.000017683522,0.000010305001,0.0002497256,0.03633713,0.93017006,0.00036879187],"about_ca_topic_score_codex":0.00006965091,"about_ca_topic_score_gemma":0.0000065580525,"teacher_disagreement_score":0.78089744,"about_ca_system_score_codex":0.000059250087,"about_ca_system_score_gemma":0.0007177614,"threshold_uncertainty_score":0.98392427},"labels":[],"label_agreement":null},{"id":"W1495618792","doi":"10.1111/jep.12362","title":"Unpacking the process of interpretation in evidence‐based decision making","year":2015,"lang":"en","type":"article","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":39,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Unpacking; Interpretation (philosophy); Process (computing); Decision-making; Computer science; Psychology; Medicine; Process management; Management science; Business; Operations management; Engineering; Philosophy; Linguistics","score_opus":0.2824743720278754,"score_gpt":0.6043669668125549,"score_spread":0.32189259478467946,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1495618792","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9765748,0.00145133,0.0082811965,0.011659661,0.0008174731,0.00038813034,2.808868e-7,0.0000033390847,0.0008237971],"genre_scores_gemma":[0.98971295,0.00014704165,0.008079667,0.0018206553,0.00022383775,0.0000055552555,5.338447e-7,0.0000075891526,0.0000021670733],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9940415,0.0015988314,0.002501866,0.00015205372,0.0015950691,0.00011068754],"domain_scores_gemma":[0.6373736,0.35309795,0.0037342089,0.0003704188,0.005238194,0.00018559284],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.054698277,0.000078890385,0.00041299444,0.0001972215,0.000014690323,0.000023444014,0.00013151269,0.0001290701,0.000045631772],"category_scores_gemma":[0.90707564,0.000050305145,0.00014108581,0.0005300445,0.00008286225,0.0005747222,0.00002312966,0.0008113549,0.000005932869],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.016733238,0.0019005617,0.39737463,0.000048293125,0.000058353264,0.00010776265,0.0017391492,0.051123388,0.000011288909,0.00006878985,0.00092991977,0.5299046],"study_design_scores_gemma":[0.012239189,0.002830505,0.6369308,0.039754145,0.0007954082,0.00019591242,0.0037028182,0.2917614,0.000048772523,0.011428565,0.0001879393,0.0001245051],"about_ca_topic_score_codex":0.000014526762,"about_ca_topic_score_gemma":0.000016634922,"teacher_disagreement_score":0.85237736,"about_ca_system_score_codex":0.00017558274,"about_ca_system_score_gemma":0.0015280687,"threshold_uncertainty_score":0.97338706},"labels":[],"label_agreement":null},{"id":"W1503803862","doi":"10.1111/j.1365-2923.2010.03622.x","title":"Teaching and learning in morbidity and mortality rounds: an ethnographic study","year":2010,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":53,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"St. Michael's Hospital; The Wilson Centre; Canadian Patient Safety Institute; University Health Network; University of Toronto; Health Sciences Centre; Sunnybrook Health Science Centre","funders":"","keywords":"Context (archaeology); Medical education; Patient safety; Perception; Psychology; Quality (philosophy); Medicine; Content analysis; Ethnography; Process (computing); Nursing; Health care; Sociology; Computer science","score_opus":0.03147121830119933,"score_gpt":0.425181517134859,"score_spread":0.3937102988336597,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1503803862","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9962639,0.00008141502,0.000014053034,0.0026200598,0.0003268374,0.0001927381,1.8034076e-7,0.000039557803,0.0004612556],"genre_scores_gemma":[0.99818456,0.000029003813,0.00031914236,0.0010559333,0.00026909134,0.000023265773,0.000015944323,0.000007297816,0.00009573254],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99886024,0.0002404919,0.00022009222,0.00025226187,0.0003030195,0.00012389597],"domain_scores_gemma":[0.99748784,0.001935804,0.000044052835,0.00016828097,0.000033577406,0.00033042283],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0029329355,0.00007659302,0.00017781055,0.000081791244,0.00010380139,0.000023332403,0.00004004726,0.0001317249,0.00007232798],"category_scores_gemma":[0.09505304,0.000063462416,0.0000155796,0.00009848365,0.0001276137,0.0000699981,0.000028641387,0.0015939354,0.0000016111603],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000009658277,0.0014484982,0.84692174,0.000011060296,0.000005965377,0.000005981174,0.0023985389,7.268539e-8,0.00003270629,0.00007682744,0.00006220232,0.14902674],"study_design_scores_gemma":[0.0007478636,0.0002891761,0.99452394,0.00019892395,0.00004070242,0.00002473525,0.002779462,0.0006436108,0.0000018937776,0.00032109936,0.00036449154,0.00006411119],"about_ca_topic_score_codex":0.008344971,"about_ca_topic_score_gemma":0.002895742,"teacher_disagreement_score":0.14896262,"about_ca_system_score_codex":0.000009836585,"about_ca_system_score_gemma":0.00032128306,"threshold_uncertainty_score":0.99825853},"labels":[],"label_agreement":null},{"id":"W150570998","doi":"10.1007/978-94-007-1682-7_10","title":"Self-monitoring in Surgical Practice: Slowing Down When You Should","year":2011,"lang":"en","type":"book-chapter","venue":"Advances in medical education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":20,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; University of Toronto","funders":"","keywords":"Judgement; Autonomy; Cornerstone; Psychology; Cognition; Control (management); Engineering ethics; Political science; Engineering; Computer science; Law; Neuroscience; Art; Artificial intelligence","score_opus":0.032566353572094135,"score_gpt":0.4097072200423877,"score_spread":0.37714086647029355,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W150570998","genre_codex":"other","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0016047708,0.03882191,0.00003602318,0.006201807,0.0053619803,0.0006386028,0.0000025778713,0.00012170122,0.9472106],"genre_scores_gemma":[0.17116214,0.44832897,0.050216597,0.011592511,0.024596676,0.00083248253,0.0009158363,0.00059331785,0.29176146],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9968711,0.000071280825,0.0009541505,0.00062966207,0.0011042282,0.00036954417],"domain_scores_gemma":[0.9903489,0.008187778,0.00029001696,0.0004468579,0.00017085884,0.00055562856],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0013696773,0.0003253684,0.0007156298,0.00027378948,0.000035374676,0.000016409049,0.00021946794,0.0008951634,0.0015764601],"category_scores_gemma":[0.067286216,0.00028981888,0.0001405602,0.00009610203,0.00015281259,0.00028143634,0.000077864155,0.0019489916,0.00014727557],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00032593976,0.0016137818,0.017181238,0.00026342552,0.000054087635,0.0009930821,0.0012638449,0.0000017143095,2.399212e-7,0.020504165,0.001050856,0.95674765],"study_design_scores_gemma":[0.0016124477,0.00020811369,0.0014814792,0.018387793,0.00017770036,0.00025875797,0.00021645872,0.000020358058,0.00000541119,0.016291907,0.96097994,0.00035961322],"about_ca_topic_score_codex":0.000121736484,"about_ca_topic_score_gemma":0.000032651413,"teacher_disagreement_score":0.9599291,"about_ca_system_score_codex":0.00039702884,"about_ca_system_score_gemma":0.0020559314,"threshold_uncertainty_score":0.9999554},"labels":[],"label_agreement":null},{"id":"W1509997342","doi":"10.1111/j.1365-2753.2010.01389.x","title":"From Popperian science to normal science. Commentary on Sestini (2009) ‘Epistemology and ethics of evidence‐based medicine’","year":2010,"lang":"en","type":"letter","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Guelph","funders":"","keywords":"Maya; Citation; Psychology; Library science; Philosophy; Computer science; Epistemology","score_opus":0.36443691980397325,"score_gpt":0.5833568533284573,"score_spread":0.21891993352448408,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1509997342","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.099331826,0.00028910357,0.0001215291,0.8944907,0.0046136533,0.00052624237,0.000013246798,0.0000055704277,0.0006081205],"genre_scores_gemma":[0.12625138,0.0003381155,0.014429408,0.851288,0.0076217707,0.0000073500746,0.000016301605,0.000021297903,0.000026324218],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9867964,0.002599275,0.0037802404,0.0007364178,0.0056542386,0.00043344978],"domain_scores_gemma":[0.47812232,0.506146,0.00507007,0.0010422005,0.008785276,0.0008341137],"candidate_categories":["metaresearch","sts","research_integrity"],"consensus_categories":["metaresearch","research_integrity"],"category_scores_codex":[0.107263185,0.00027283456,0.0013193887,0.00098136,0.00016635485,0.000060395127,0.0006019814,0.0014135643,0.00039349636],"category_scores_gemma":[0.92910427,0.00018583164,0.00017351413,0.0010861098,0.0038456281,0.00059557025,0.00016144033,0.013135802,0.0000151258],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0056690266,0.0016214802,0.03970137,0.00015882646,0.00014390767,0.0012832509,0.0011056232,0.00022889827,0.00070523686,0.000044750905,0.911824,0.03751364],"study_design_scores_gemma":[0.022741621,0.04299087,0.4253926,0.08325542,0.006818719,0.0015609224,0.0012439279,0.0063546207,0.00046818997,0.003778259,0.4042906,0.0011042703],"about_ca_topic_score_codex":0.0006179873,"about_ca_topic_score_gemma":0.000024791394,"teacher_disagreement_score":0.82184106,"about_ca_system_score_codex":0.00025765077,"about_ca_system_score_gemma":0.0077951625,"threshold_uncertainty_score":0.9998828},"labels":[],"label_agreement":null},{"id":"W1515241746","doi":"10.1002/9781444303674.ch3","title":"Critical Appraisal: General Issues in Emergency Medicine","year":2008,"lang":"en","type":"other","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Critical appraisal; Psychology; Medicine; Engineering ethics; Alternative medicine; Epistemology; Engineering; Pathology; Philosophy","score_opus":0.04522030767744277,"score_gpt":0.4416595787963401,"score_spread":0.3964392711188973,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1515241746","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00022102882,0.009890329,0.000079346326,0.010321186,0.003031235,0.00020730402,0.000008372776,0.00031515356,0.97592604],"genre_scores_gemma":[0.00042681996,0.02082881,0.0019637463,0.0005701489,0.004860332,0.000023852312,0.00011377377,0.00035180923,0.9708607],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9980872,0.0000492193,0.0006194587,0.00046554717,0.00042159983,0.0003569496],"domain_scores_gemma":[0.99451673,0.004467377,0.00008191492,0.00049951265,0.000065302454,0.00036915808],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00016655204,0.00029587056,0.0009001426,0.00028218713,0.000013542662,0.0000020653617,0.00010667024,0.00050071557,0.08736264],"category_scores_gemma":[0.106451124,0.00020962127,0.0001315462,0.00022446657,0.00027137832,0.000011388321,0.000040374747,0.0004063904,0.00072389503],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00001971454,0.00026709525,0.008095612,0.00006407438,0.0000325398,0.00026807428,0.000049027763,9.0351676e-8,0.0000048747256,0.0018930968,0.9880407,0.0012650937],"study_design_scores_gemma":[0.0013343572,0.00032907206,0.0073012905,0.003958364,0.000102187085,0.00006830311,0.000032141907,0.00003183169,0.000005342232,0.00018428544,0.9863886,0.00026417617],"about_ca_topic_score_codex":0.003956668,"about_ca_topic_score_gemma":0.0002747262,"teacher_disagreement_score":0.10628457,"about_ca_system_score_codex":0.000021776077,"about_ca_system_score_gemma":0.000079520716,"threshold_uncertainty_score":0.9304448},"labels":[],"label_agreement":null},{"id":"W1528766431","doi":"10.1016/j.annemergmed.2015.07.023","title":"From Context Comes Expertise: How Do Expert Emergency Physicians Use Their Know-Who to Make Decisions?","year":2015,"lang":"en","type":"article","venue":"Annals of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Université de Montréal; Université de Sherbrooke","funders":"","keywords":"Seriousness; Emergency department; Medicine; Perspective (graphical); Context (archaeology); Qualitative research; Health care; Point (geometry); Medical emergency; Nursing","score_opus":0.2941979166589376,"score_gpt":0.4538951242605257,"score_spread":0.15969720760158812,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1528766431","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8775202,0.03550984,0.0010233193,0.076763086,0.0057173287,0.0007101175,0.00015009272,0.00011906309,0.00248692],"genre_scores_gemma":[0.95503545,0.028443044,0.00032959724,0.0051257606,0.0028968675,0.00007884893,0.00016034553,0.00008183486,0.007848252],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9960704,0.00016238926,0.0013498924,0.00072922674,0.0010542659,0.0006338232],"domain_scores_gemma":[0.9920708,0.003504135,0.00038322396,0.0012729255,0.00150359,0.001265302],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0007156838,0.0004838381,0.0014437069,0.0002397428,0.000079949124,0.000008156621,0.00039970936,0.00021348581,0.0033240346],"category_scores_gemma":[0.094639786,0.0003405306,0.0004358363,0.0007645889,0.0001762609,0.00014794306,0.00015280742,0.0002818651,0.00012572322],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00050172134,0.0005988709,0.019511644,0.000009837819,0.0002489797,0.00003209318,0.009149679,0.0000039675606,0.0015336081,0.00012570812,0.83413607,0.13414782],"study_design_scores_gemma":[0.008099117,0.0074419947,0.16178092,0.01685789,0.00045225557,0.0000151898985,0.027932217,0.00034335066,0.011728132,0.0049784863,0.7589457,0.0014247275],"about_ca_topic_score_codex":0.0027446318,"about_ca_topic_score_gemma":0.00021590738,"teacher_disagreement_score":0.14226928,"about_ca_system_score_codex":0.000030146244,"about_ca_system_score_gemma":0.000108070744,"threshold_uncertainty_score":0.9999047},"labels":[],"label_agreement":null},{"id":"W1529830748","doi":"10.1111/j.1365-2753.2012.01898.x","title":"Conciliating cognition and consciousness: the perceptual foundations of clinical reasoning","year":2012,"lang":"en","type":"article","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Psychology; Consciousness; Cognition; Empathy; Analytic reasoning; Perception; Cognitive science; Cognitive psychology; Epistemology; Deductive reasoning; Social psychology; Psychiatry","score_opus":0.19894625339464805,"score_gpt":0.553040400295548,"score_spread":0.35409414690089996,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1529830748","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9838259,0.0012474542,0.00071515073,0.0075824438,0.0014086172,0.00038984016,0.0000025898019,0.000005811161,0.004822173],"genre_scores_gemma":[0.9861685,0.000943532,0.009527197,0.0019613586,0.00135422,0.0000062172408,0.000006217826,0.00001028071,0.000022477587],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9923597,0.0026704716,0.0035900746,0.00015924765,0.0010354966,0.00018498691],"domain_scores_gemma":[0.5339198,0.4528526,0.006554115,0.00048223865,0.0055820583,0.0006091765],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.06373685,0.00010122674,0.0006297797,0.000080346275,0.00008113406,0.000025359812,0.0000717025,0.00021006944,0.0002307936],"category_scores_gemma":[0.8383938,0.000067894805,0.00024511485,0.00022754971,0.00043250035,0.00051919726,0.00004390172,0.0011452705,0.000013558994],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012762677,0.0020295647,0.86959815,0.000021417058,0.00024331424,0.000012338971,0.0010457253,0.000041957024,0.000024904928,0.0009821197,0.0018440753,0.12288019],"study_design_scores_gemma":[0.006927843,0.0010981007,0.97576374,0.0014281888,0.0021499887,0.0005215643,0.005105537,0.003694062,0.000008722222,0.0006870824,0.002520664,0.00009452111],"about_ca_topic_score_codex":0.000022796426,"about_ca_topic_score_gemma":0.000008042251,"teacher_disagreement_score":0.77465695,"about_ca_system_score_codex":0.000043829612,"about_ca_system_score_gemma":0.00072823634,"threshold_uncertainty_score":0.9640799},"labels":[],"label_agreement":null},{"id":"W1538633966","doi":"10.1111/j.1553-2712.2011.01123.x","title":"Getting the Evidence Straight in Emergency Diagnostics","year":2011,"lang":"en","type":"article","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary; Alberta Health Services","funders":"","keywords":"Medicine; Modalities; Work (physics); Task (project management); Test (biology); Quality (philosophy); Interpretation (philosophy); Evidence-based medicine; Emergency department; Emergency physician; Point (geometry); MEDLINE; Medical emergency; Engineering ethics; Alternative medicine; Pathology; Epistemology; Computer science; Nursing; Management","score_opus":0.1659709050698813,"score_gpt":0.41521142008851236,"score_spread":0.24924051501863106,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1538633966","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.88921785,0.055173084,0.00022525407,0.016622134,0.004560461,0.0008917802,0.000004423744,0.00014988566,0.03315511],"genre_scores_gemma":[0.8378024,0.15723126,0.0001457045,0.00090721634,0.0014990759,0.00007484516,0.000011796994,0.00003924656,0.0022884295],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9960043,0.00015399436,0.0019500669,0.0005473571,0.00068932417,0.0006549482],"domain_scores_gemma":[0.9929621,0.0053200927,0.00041341654,0.0007010663,0.00019834194,0.00040497902],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0023116285,0.00031732413,0.000559671,0.00016151152,0.00011060137,3.8756463e-7,0.0004927592,0.00033063325,0.014318295],"category_scores_gemma":[0.15372579,0.0001925229,0.00014410658,0.00091649615,0.00026571454,0.0001228467,0.00011897824,0.0015702421,0.00019464357],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008722125,0.00013149228,0.9175452,0.00006414909,0.00004540067,0.000072928815,0.003634908,0.0000014939019,0.0005791672,0.0014573634,0.07177241,0.004608276],"study_design_scores_gemma":[0.0009770038,0.0005714497,0.98206514,0.005997615,0.00035654887,0.000049739963,0.00077888643,0.00014177295,0.00024282232,0.005397524,0.0031238256,0.00029764877],"about_ca_topic_score_codex":0.00061752595,"about_ca_topic_score_gemma":0.0000632168,"teacher_disagreement_score":0.15141416,"about_ca_system_score_codex":0.00005745292,"about_ca_system_score_gemma":0.00011296715,"threshold_uncertainty_score":0.98658276},"labels":[],"label_agreement":null},{"id":"W1541980413","doi":"","title":"Applying structural equation modeling to Canadian Chiropractic Examining Board measures.","year":2006,"lang":"en","type":"article","venue":"PubMed","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Calgary","funders":"","keywords":"Chiropractic; Structural equation modeling; Computer science; Licensure; Latent variable; Path analysis (statistics); Matching (statistics); Variable (mathematics); Psychology; Applied psychology; Medical education; Artificial intelligence; Medicine; Machine learning; Alternative medicine; Mathematics; Pathology","score_opus":0.10597374117822615,"score_gpt":0.29989516413583966,"score_spread":0.1939214229576135,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1541980413","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98483145,0.00014352822,0.0026426082,0.0019049667,0.0004271687,0.001304344,0.0000048192355,0.00009366492,0.008647465],"genre_scores_gemma":[0.9947681,0.000003573963,0.0014676106,0.0015103774,0.0005492976,0.0012116348,0.00004073748,0.000025310275,0.0004233775],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9984945,0.000027311657,0.00030806576,0.0002882171,0.00034623555,0.0005356369],"domain_scores_gemma":[0.99816245,0.000827811,0.000053993135,0.00022264945,0.000114508315,0.00061856466],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0005145635,0.00012861914,0.0002255501,0.00016501405,0.0001124289,0.00005706127,0.0000669974,0.000092776645,0.00002604345],"category_scores_gemma":[0.022506837,0.000117874944,0.000051860232,0.0002163189,0.000014689632,0.00007155802,0.000018294762,0.00019769292,0.000042976462],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011703666,0.000043835,0.0949547,0.000018308006,0.000052013977,0.00009448461,0.00017330569,0.030560004,0.00007457254,0.0003863096,0.0017486423,0.87177676],"study_design_scores_gemma":[0.0014973454,0.0000561683,0.9247116,0.00024187117,0.00018181112,0.000047266796,0.00012219601,0.06915953,0.00009430694,0.0011303317,0.0024053142,0.00035229718],"about_ca_topic_score_codex":0.13409816,"about_ca_topic_score_gemma":0.018907076,"teacher_disagreement_score":0.8714245,"about_ca_system_score_codex":0.00030530457,"about_ca_system_score_gemma":0.00016589698,"threshold_uncertainty_score":0.9989953},"labels":[],"label_agreement":null},{"id":"W1544658434","doi":"10.1056/nejmcps0706713","title":"A Gut Feeling","year":2008,"lang":"en","type":"article","venue":"New England Journal of Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University Health Network; Centre for Global Health Research; William Osler Health System; University of Toronto","funders":"","keywords":"Feeling; Weight loss; Personality; Shame","score_opus":0.04717432239776368,"score_gpt":0.33900298879434837,"score_spread":0.29182866639658467,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1544658434","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9647632,0.0032469258,0.00097552314,0.023498511,0.00092052924,0.000065962166,4.1021687e-7,0.000017768834,0.0065111606],"genre_scores_gemma":[0.9834242,0.0017705024,0.0018606253,0.0020265447,0.009175293,2.5432286e-7,0.0000024925835,0.000012672536,0.0017274525],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99859196,0.000039404567,0.0005937538,0.00009386505,0.0005136511,0.00016734727],"domain_scores_gemma":[0.9950021,0.0037902554,0.00024084393,0.00014514256,0.00023902467,0.000582627],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00076834863,0.000105084946,0.0005321477,0.0001333561,0.000044817443,0.0000023700002,0.00007841951,0.00006267125,0.00046824527],"category_scores_gemma":[0.039594885,0.000061681974,0.00010668101,0.00015034458,0.00012622066,0.000035749057,0.00001153561,0.00040579942,0.000020305435],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0029168997,0.00041565366,0.16308266,0.000036571662,0.00030918006,0.010164629,0.003200331,0.00007574499,0.0037189557,0.0002831632,0.2764824,0.5393138],"study_design_scores_gemma":[0.25482953,0.010902643,0.34634367,0.011718715,0.000990585,0.08706003,0.0003570592,0.00015950194,0.00041461663,0.0020912716,0.28481832,0.00031408208],"about_ca_topic_score_codex":0.000054322856,"about_ca_topic_score_gemma":0.0000026541434,"teacher_disagreement_score":0.53899974,"about_ca_system_score_codex":0.000026683323,"about_ca_system_score_gemma":0.0003088115,"threshold_uncertainty_score":0.968495},"labels":[],"label_agreement":null},{"id":"W154754512","doi":"10.3233/978-1-61499-488-6-160","title":"Modelling Clinical Diagnostic Errors: A System Dynamics Approach","year":2015,"lang":"en","type":"article","venue":"Studies in health technology and informatics","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria","funders":"","keywords":"Computer science; Tracing; Process (computing); System dynamics; Key (lock); Data mining; Artificial intelligence","score_opus":0.1562475969601111,"score_gpt":0.438855780286302,"score_spread":0.2826081833261909,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W154754512","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9192941,0.02312271,0.03941946,0.00795843,0.0013751658,0.0015900823,0.000017306353,0.00067426346,0.006548483],"genre_scores_gemma":[0.95107013,0.007872752,0.040006906,0.0008340239,0.000063262756,0.00007284121,0.000018112463,0.000012290852,0.000049695496],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9976419,0.00005735626,0.0015666144,0.0001646453,0.00017872846,0.00039073516],"domain_scores_gemma":[0.99397355,0.0049701105,0.00031682657,0.00032465433,0.00017894496,0.00023590578],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0022513822,0.00016266486,0.00085387373,0.00032386987,0.00010226297,0.000006162509,0.00009680584,0.00034135178,2.6544205e-7],"category_scores_gemma":[0.044317458,0.00012737558,0.000046150864,0.0005205779,0.0006591658,0.00006405303,0.00018601236,0.00070836634,0.000012698611],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002895289,0.00095042813,0.7155057,0.0072353156,0.0003841897,0.00010129169,0.010841285,0.010367139,4.8862088e-9,0.1734642,0.0063208817,0.07454004],"study_design_scores_gemma":[0.003473488,0.0015527039,0.0010604291,0.0065177963,0.00006933922,0.00035979773,0.10152924,0.8809281,2.7029927e-7,0.0032122189,0.0010802826,0.00021632621],"about_ca_topic_score_codex":0.00003403754,"about_ca_topic_score_gemma":0.000007621919,"teacher_disagreement_score":0.87056094,"about_ca_system_score_codex":0.00027540672,"about_ca_system_score_gemma":0.00026048342,"threshold_uncertainty_score":0.96373266},"labels":[],"label_agreement":null},{"id":"W1549737310","doi":"10.82308/24767","title":"Making a case: validating criterion-referenced guidelines for virtual patient case authoring","year":2011,"lang":"en","type":"dissertation","venue":"eScholarship@McGill (McGill)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Computer science","score_opus":0.17694758294980972,"score_gpt":0.40963759624307194,"score_spread":0.23269001329326222,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1549737310","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97986186,0.00037629998,0.0000062869476,0.000037050693,0.003321271,0.0017041155,0.0016247102,0.00040870797,0.012659689],"genre_scores_gemma":[0.97907025,0.00009622316,0.013274103,0.0008698952,0.00033308115,0.0006023754,0.0017796032,0.00034129413,0.0036331904],"study_design_codex":"design_other","study_design_gemma":"case_report","domain_scores_codex":[0.9938659,0.00025857784,0.0024265687,0.0016076663,0.00079028733,0.0010509891],"domain_scores_gemma":[0.98848337,0.0061704,0.00116673,0.0010117124,0.0024567896,0.000710983],"candidate_categories":["metaresearch","metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.0015341165,0.0010442913,0.0015192535,0.00046580526,0.0014396289,0.00012646121,0.0002928054,0.0012164045,0.00038254104],"category_scores_gemma":[0.12688872,0.001001599,0.0008097062,0.00042777235,0.000064060696,0.00031652275,0.00017428564,0.0016289931,0.00010208859],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012802568,0.0005904611,0.000063552325,0.000801369,0.0005176945,0.02968566,0.00012635658,0.00001890916,0.0042913174,0.012829691,0.00009334649,0.94970137],"study_design_scores_gemma":[0.03728679,0.022712225,0.0015339595,0.20017931,0.021877736,0.23659627,0.039970506,0.0034055994,0.21611455,0.040175375,0.16122599,0.0189217],"about_ca_topic_score_codex":0.0009368721,"about_ca_topic_score_gemma":0.0011009645,"teacher_disagreement_score":0.9307797,"about_ca_system_score_codex":0.00044507626,"about_ca_system_score_gemma":0.00018782166,"threshold_uncertainty_score":0.99986035},"labels":[],"label_agreement":null},{"id":"W1552977961","doi":"10.1007/978-3-642-13437-1_32","title":"Using Expert Models to Provide Feedback on Clinical Reasoning Skills","year":2010,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Computer science; Process (computing); Outcome (game theory); Expert system; Model-based reasoning; Artificial intelligence; Management science; Data science; Knowledge representation and reasoning","score_opus":0.08075114636588483,"score_gpt":0.3851151452991617,"score_spread":0.30436399893327687,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1552977961","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015698718,0.00020255525,0.97074205,0.0018146324,0.0035738954,0.001040646,0.000008755276,0.00014321355,0.006775532],"genre_scores_gemma":[0.17419131,0.000116262556,0.79609185,0.024241075,0.004098551,0.000014330792,0.0000147266455,0.00012979617,0.001102111],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9954677,0.0000387747,0.00096262374,0.0017171229,0.0011009393,0.0007128324],"domain_scores_gemma":[0.9878661,0.009473313,0.00027596217,0.0013153453,0.00031688728,0.0007524044],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0015639601,0.0005498916,0.0010766368,0.00046887292,0.00017764149,0.00015978357,0.0007162156,0.0008523681,0.00006544774],"category_scores_gemma":[0.01654842,0.0004434049,0.00030766486,0.00030510785,0.0008016669,0.000112939604,0.0004947492,0.00219423,0.00010678313],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018606482,0.00035498448,0.00061625044,0.00002403685,0.00003531523,0.000370395,0.00079432054,0.029290449,0.00025588996,0.0016095256,0.00013593175,0.96632683],"study_design_scores_gemma":[0.0066461833,0.008181243,0.0033075893,0.08119881,0.00033622608,0.0009925799,0.0000027412148,0.7278125,0.005642128,0.14655234,0.014756349,0.004571258],"about_ca_topic_score_codex":0.000072774405,"about_ca_topic_score_gemma":0.000049568567,"teacher_disagreement_score":0.9617556,"about_ca_system_score_codex":0.0002507064,"about_ca_system_score_gemma":0.0010358104,"threshold_uncertainty_score":0.99980175},"labels":[],"label_agreement":null},{"id":"W1553259577","doi":"","title":"Revisiting the modern scientific physician","year":2002,"lang":"en","type":"article","venue":"PubMed Central","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Computer science; Context (archaeology); Presentation (obstetrics); Action (physics); Epistemology; Subject (documents); Data science; Engineering ethics; Management science; Medicine; World Wide Web; Philosophy; History","score_opus":0.039073349309110925,"score_gpt":0.2761523519803906,"score_spread":0.23707900267127968,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1553259577","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95187426,0.0014389453,0.00014286308,0.01526977,0.0011534055,0.00042125586,0.0000067258475,0.00011242241,0.029580384],"genre_scores_gemma":[0.9914973,0.000041790103,0.00008548541,0.001771223,0.0018966382,0.00003228905,0.000010164647,0.000012363727,0.0046527428],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9983833,0.000031770684,0.00017681948,0.00023367933,0.00028435176,0.0008900854],"domain_scores_gemma":[0.9980355,0.0011897696,0.000050092945,0.0003259587,0.00003419865,0.0003644821],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00045943834,0.00008659991,0.00015242613,0.000021885768,0.00016631874,0.000091338574,0.00010396985,0.00003769002,0.00013066108],"category_scores_gemma":[0.010577948,0.0000548128,0.000119833996,0.00019172342,0.0001524008,0.0000400423,0.000028956842,0.00019074182,0.000138538],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000013831804,0.00012443698,0.039858084,0.000011907721,0.0000316589,0.000030883355,0.0005224555,0.000025888456,0.000056611236,0.00052632816,0.01305561,0.9457423],"study_design_scores_gemma":[0.0015179268,0.000031066167,0.94072026,0.0006088159,0.00020142646,0.0000422239,0.00009371465,0.018079497,0.0004350637,0.0013820818,0.036666926,0.00022097709],"about_ca_topic_score_codex":0.000006833452,"about_ca_topic_score_gemma":9.194774e-7,"teacher_disagreement_score":0.94552135,"about_ca_system_score_codex":0.00006362347,"about_ca_system_score_gemma":0.00002521219,"threshold_uncertainty_score":0.99775636},"labels":[],"label_agreement":null},{"id":"W1561989677","doi":"10.1007/978-1-4419-5562-3_54","title":"Communication in Crisis: The Importance of ‘Verbal Dexterity’","year":2010,"lang":"en","type":"book-chapter","venue":"Intensive Care Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta Hospital","funders":"","keywords":"Aviation; Crash; Psychology; Medicine; Public relations; Engineering; Political science; Computer science","score_opus":0.03116236581123113,"score_gpt":0.34267192064007923,"score_spread":0.3115095548288481,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1561989677","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0773556,0.10106443,0.000060131213,0.16848579,0.003100772,0.0029750199,0.00018395734,0.00015501474,0.64661926],"genre_scores_gemma":[0.9294368,0.0056972248,0.00022682462,0.036930624,0.0007772217,0.000037794078,0.000627137,0.00009797309,0.02616837],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99810326,0.000032286716,0.00087682216,0.00033620623,0.00046336165,0.00018805276],"domain_scores_gemma":[0.9893486,0.0047867363,0.0004783095,0.0013091558,0.0039370353,0.00014019253],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00038315298,0.0002941437,0.0010455974,0.0001392731,0.00003783911,0.0000037214775,0.00030053602,0.0004973644,0.00063070556],"category_scores_gemma":[0.048322517,0.00018015016,0.0001887705,0.000062457,0.0009568265,0.000021063343,0.00013342207,0.0015892494,0.000027625958],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004605467,0.00045361303,0.13774271,0.002466456,0.0019449199,0.004110358,0.07481135,0.000004372923,0.0031303032,0.13031465,0.5313513,0.10906448],"study_design_scores_gemma":[0.018638024,0.0061532003,0.12471111,0.19590043,0.005481679,0.0014403267,0.092057824,0.000036697245,0.0012336406,0.068876915,0.48352244,0.0019476976],"about_ca_topic_score_codex":0.00045058233,"about_ca_topic_score_gemma":0.00076769525,"teacher_disagreement_score":0.85208124,"about_ca_system_score_codex":0.000092060596,"about_ca_system_score_gemma":0.00018170138,"threshold_uncertainty_score":0.95969385},"labels":[],"label_agreement":null},{"id":"W1563868984","doi":"10.1111/j.1365-2923.2010.03799.x","title":"The influence of familiar non‐diagnostic information on the diagnostic decisions of novices","year":2011,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; McGill University","funders":"","keywords":"Medical diagnosis; Psychology; Competence (human resources); Diagnostic test; Cognitive psychology; Clinical psychology; Medicine; Social psychology; Pediatrics","score_opus":0.019073545800488263,"score_gpt":0.3307955772174822,"score_spread":0.31172203141699395,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1563868984","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9901154,0.00012724767,0.00014731732,0.005240527,0.00042496252,0.00041533168,0.000004471981,0.000013171833,0.0035115732],"genre_scores_gemma":[0.9943597,0.0012386189,0.000157673,0.0040089902,0.000078010424,0.00009107323,0.000020965288,0.000005938835,0.00003907832],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99833924,0.00007100326,0.0006360495,0.00010850065,0.0006912838,0.00015390286],"domain_scores_gemma":[0.8290276,0.16859253,0.00050666404,0.0008563779,0.0007186685,0.00029816746],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010045054,0.000098080054,0.00019729181,0.00006406909,0.00009740078,0.000008121992,0.0002675788,0.00011547539,0.00017191519],"category_scores_gemma":[0.66277826,0.000049194205,0.000079307916,0.000287158,0.0003832308,0.00008115172,0.000042297434,0.00025117464,0.0001225256],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00039737186,0.0040289965,0.19250686,0.00013674093,0.00012895052,0.0000060317775,0.008242409,0.000077430086,0.000048582562,0.018443068,0.05668586,0.7192977],"study_design_scores_gemma":[0.00044514387,0.00042982306,0.9894869,0.0037787526,0.00009507997,0.000008249592,0.0008629529,0.00020506322,0.00035206552,0.0012653386,0.0030061419,0.00006451859],"about_ca_topic_score_codex":0.00054641423,"about_ca_topic_score_gemma":0.000028020662,"teacher_disagreement_score":0.79698,"about_ca_system_score_codex":0.000024286726,"about_ca_system_score_gemma":0.0010039666,"threshold_uncertainty_score":0.3400623},"labels":[],"label_agreement":null},{"id":"W1570982714","doi":"","title":"UNDA academic collects top accolade for new assessment model","year":2011,"lang":"en","type":"article","venue":"ResearchOnline - ND (The University of Notre Dame Australia)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Computer science","score_opus":0.269251184019913,"score_gpt":0.4408702262389802,"score_spread":0.1716190422190672,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1570982714","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96832,0.00007073718,0.0071976837,0.018064922,0.00013569182,0.0016070378,0.00025520258,0.000082692764,0.0042660087],"genre_scores_gemma":[0.8540265,0.00035235347,0.03949421,0.00028905447,0.00021850849,0.0000022095296,0.00014061382,0.000027322694,0.10544924],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9983744,0.00007059224,0.00024572364,0.0003385099,0.00050809124,0.00046270262],"domain_scores_gemma":[0.99595964,0.0027070954,0.00012519464,0.00046984165,0.0002502462,0.0004879533],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00052765134,0.0001535248,0.0003760251,0.00012574028,0.00018628401,0.000009018425,0.00045967422,0.00023051738,0.0007129226],"category_scores_gemma":[0.0042352006,0.00012971806,0.00023479311,0.00034055675,0.0003290215,0.000084539235,0.00016436403,0.0007011758,0.0000527457],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0073716175,0.0027675976,0.06936845,0.00041468922,0.0010324211,0.0003155487,0.004926363,0.001348143,0.00582578,0.007480506,0.8667405,0.03240833],"study_design_scores_gemma":[0.044563882,0.009098534,0.5872155,0.006075346,0.0029950866,0.00011525336,0.007308462,0.14162393,0.007247795,0.01909456,0.17279358,0.0018680994],"about_ca_topic_score_codex":0.0011616095,"about_ca_topic_score_gemma":0.00007870365,"teacher_disagreement_score":0.69394696,"about_ca_system_score_codex":0.00015103948,"about_ca_system_score_gemma":0.0009762248,"threshold_uncertainty_score":0.78060025},"labels":[],"label_agreement":null},{"id":"W1571663093","doi":"10.3917/rfla.181.0093","title":"Des analogies entre le raisonnement médical et l'évaluation formative","year":2013,"lang":"fr","type":"article","venue":"Revue française de linguistique appliquée","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Formative assessment; Humanities; Valuation (finance); Philosophy; Psychology; Economics; Pedagogy","score_opus":0.024865769517826108,"score_gpt":0.3345235094149087,"score_spread":0.3096577398970826,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1571663093","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.52153885,0.10432298,0.16078971,0.06966664,0.004838836,0.0056286585,0.0002015108,0.00077338953,0.1322394],"genre_scores_gemma":[0.963774,0.0044350307,0.019945113,0.002499841,0.0028881468,0.0005286254,0.00020062667,0.00008399409,0.0056445966],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9964988,0.0004517987,0.0010362217,0.00059289817,0.00041677937,0.001003479],"domain_scores_gemma":[0.9868545,0.01074207,0.0003493689,0.0005908243,0.0006960397,0.000767175],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0018384892,0.00046888657,0.000823023,0.00013825447,0.00025381058,0.00012999559,0.00026640648,0.0005937468,0.0019784523],"category_scores_gemma":[0.07461832,0.00047468152,0.000363919,0.00035342047,0.00064014073,0.00013932401,0.00015048293,0.000917274,0.0017341148],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003859908,0.004640411,0.04888592,0.0030250133,0.00084592553,0.00074951426,0.056194868,0.0025720168,0.0006941843,0.341512,0.14378837,0.3967058],"study_design_scores_gemma":[0.014792551,0.0029773898,0.24446894,0.054607384,0.0022446285,0.0005953476,0.010724711,0.07290617,0.004495472,0.1825958,0.4067343,0.0028573186],"about_ca_topic_score_codex":0.008718789,"about_ca_topic_score_gemma":0.00046786928,"teacher_disagreement_score":0.44223514,"about_ca_system_score_codex":0.00031793377,"about_ca_system_score_gemma":0.001337418,"threshold_uncertainty_score":0.99977046},"labels":[],"label_agreement":null},{"id":"W1575075084","doi":"","title":"Measuring philosophy: a philosophy index.","year":2002,"lang":"en","type":"article","venue":"PubMed Central","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Chiropractic; Meaning (existential); Empiricism; Medicine; Index (typography); Epistemology; Alternative medicine; Philosophy; Pathology; Computer science","score_opus":0.08217778038973296,"score_gpt":0.2678684680850567,"score_spread":0.18569068769532376,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1575075084","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8599233,0.0015479038,0.00010918204,0.01696724,0.0014562514,0.00073440856,0.000014479137,0.0003383277,0.118908904],"genre_scores_gemma":[0.9942783,0.00017368904,0.0001659732,0.0021519212,0.002396503,0.00008178157,0.000012428799,0.000031489133,0.0007079218],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9976056,0.000032378135,0.00031720242,0.00034862012,0.0004489706,0.0012472128],"domain_scores_gemma":[0.997604,0.00082397327,0.00006825359,0.00034625683,0.000045216155,0.0011123096],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00025813453,0.00018347501,0.0003322988,0.00008081811,0.00006700852,0.000026161775,0.00011413078,0.00013083743,0.0005978756],"category_scores_gemma":[0.012943252,0.00015990346,0.00020155418,0.0002070688,0.00009875471,0.00006825547,0.000040609233,0.00034841878,0.00026640625],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00019559814,0.0011329503,0.9276418,0.00006619622,0.0001734828,0.0003891571,0.0002837095,0.000041546573,0.000015956657,0.003414959,0.0074760574,0.05916857],"study_design_scores_gemma":[0.003268356,0.00011066034,0.98688847,0.0005826628,0.00012460697,0.000105471765,0.000009149823,0.00070001645,0.00013877492,0.0042475276,0.0035642877,0.00026000867],"about_ca_topic_score_codex":0.000021042364,"about_ca_topic_score_gemma":0.0000021907038,"teacher_disagreement_score":0.134355,"about_ca_system_score_codex":0.00016423404,"about_ca_system_score_gemma":0.00003919463,"threshold_uncertainty_score":0.99537116},"labels":[],"label_agreement":null},{"id":"W1577673877","doi":"10.1023/a:1026006801902","title":"Argumentation and Evidence","year":2003,"lang":"en","type":"article","venue":"Metamedicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":72,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; University of Toronto; Sunnybrook Health Science Centre","funders":"Canadian Institutes of Health Research","keywords":"Philosophy of medicine; Dialogical self; Informal logic; Epistemology; Relevance (law); Context (archaeology); Empirical evidence; Evidence-based medicine; Rationality; Psychology; Engineering ethics; MEDLINE; Alternative medicine; Medicine; Philosophy; Political science; Law; Argumentation theory","score_opus":0.06111398801593036,"score_gpt":0.3871033470182868,"score_spread":0.3259893590023564,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1577673877","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9786021,0.009047415,0.0011308091,0.0050517744,0.00022740844,0.00020409435,3.3752175e-7,0.000044106706,0.0056919367],"genre_scores_gemma":[0.99232,0.00076156494,0.0026383123,0.0026433629,0.00008492011,0.000008531765,0.000003047842,0.0000067716915,0.0015334697],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99932235,0.000040003943,0.00017314096,0.00016230492,0.00018930649,0.00011290664],"domain_scores_gemma":[0.9962504,0.0033416795,0.000038734142,0.00014748225,0.000045473025,0.00017622863],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0005296031,0.000068749956,0.00020670534,0.00004109123,0.000025871455,0.0000056867216,0.00001629479,0.000034676603,0.00034540662],"category_scores_gemma":[0.0591967,0.000047778725,0.000027648495,0.00011562337,0.0000850554,0.000034399218,0.0000062693334,0.000093132316,0.000046496596],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00031155496,0.00048840244,0.78751916,0.00020262325,0.00021378009,0.0003117564,0.0007752859,0.0000025093898,0.013268598,0.012233021,0.016462062,0.16821124],"study_design_scores_gemma":[0.008181173,0.0022745917,0.9382895,0.0081558805,0.0013027836,0.0007561911,0.0005393272,0.00018734798,0.012471121,0.002801632,0.02473713,0.00030328459],"about_ca_topic_score_codex":0.000015798989,"about_ca_topic_score_gemma":0.0000010964868,"teacher_disagreement_score":0.16790795,"about_ca_system_score_codex":0.000013457483,"about_ca_system_score_gemma":0.00003899828,"threshold_uncertainty_score":0.9487281},"labels":[],"label_agreement":null},{"id":"W1582586801","doi":"10.18438/b8tw52","title":"What is Evidence?","year":2011,"lang":"en","type":"article","venue":"Evidence Based Library and Information Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"History","score_opus":0.051228912998282056,"score_gpt":0.3303559925325011,"score_spread":0.279127079534219,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1582586801","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0696811,0.028377663,0.0058024875,0.8377977,0.0020901437,0.0017266228,0.000014896094,0.0008047961,0.053704593],"genre_scores_gemma":[0.35847318,0.048205674,0.023089202,0.56893015,0.00017350253,0.000040078452,0.000032665863,0.000014560774,0.0010409831],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99888676,0.00010008939,0.00042346324,0.00015044157,0.00028182904,0.00015743966],"domain_scores_gemma":[0.9819398,0.017128767,0.0002522392,0.00032178083,0.000090913454,0.0002665172],"candidate_categories":["metaresearch","scholarly_communication","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00056519016,0.00011745378,0.00016457327,0.00010476959,0.000083104664,0.00029275863,0.00009459105,0.00010174443,0.002402958],"category_scores_gemma":[0.04899195,0.000095111754,0.00005829799,0.00026958925,0.0000800275,0.27016398,0.000059709797,0.00023296714,0.00039139687],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0138686495,0.00097393134,0.062083688,0.0010197238,0.00021846642,0.00022080478,0.00868845,0.000009507475,0.00007455864,0.21114677,0.11813471,0.5835607],"study_design_scores_gemma":[0.0022223648,0.002598298,0.18274686,0.04188326,0.00058619044,0.00041852897,0.00463093,0.003198531,0.011975331,0.0010653355,0.74806225,0.0006120917],"about_ca_topic_score_codex":0.000009801431,"about_ca_topic_score_gemma":5.809653e-9,"teacher_disagreement_score":0.6299276,"about_ca_system_score_codex":0.0000063950233,"about_ca_system_score_gemma":0.00025226615,"threshold_uncertainty_score":0.998509},"labels":[],"label_agreement":null},{"id":"W1603256161","doi":"10.1111/j.1365-2753.2002.00433.x","title":"Knowledge base of scientific gnosis: IV. Knowledge base of scientific gnosis <i>vis‐à‐vis</i> evidence base of this","year":2004,"lang":"en","type":"article","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Biostatistics; Epidemiology; Library science; Face (sociological concept); Medicine; Sociology; Computer science; Social science; Pathology","score_opus":0.20875788186395522,"score_gpt":0.5143519475341006,"score_spread":0.3055940656701454,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1603256161","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9743716,0.009650702,0.00041134076,0.007451629,0.0047066836,0.0009495619,0.000037903224,0.00001574869,0.002404837],"genre_scores_gemma":[0.98843664,0.0012253795,0.009062046,0.00043004175,0.0005016112,0.000021100572,0.000015330152,0.000035440138,0.0002724365],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9854111,0.003209961,0.006827563,0.0008539775,0.0031722942,0.0005250941],"domain_scores_gemma":[0.67743963,0.28116566,0.0121382065,0.0024906178,0.025192367,0.0015735099],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.08873701,0.00032794382,0.0016174082,0.00096797507,0.00016827129,0.00009873746,0.00055096,0.00041024855,0.0010291231],"category_scores_gemma":[0.7845388,0.00027480454,0.0009336123,0.0026485247,0.0014697631,0.0011968138,0.00020145839,0.0014655768,0.00013169214],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.024613615,0.15538521,0.33172986,0.0043420773,0.0022429884,0.0007188442,0.010108148,0.005531786,0.0419184,0.002665098,0.16704176,0.25370222],"study_design_scores_gemma":[0.10212441,0.03150063,0.4103216,0.21955624,0.030653542,0.0011967674,0.0064372327,0.039419223,0.12062124,0.010203767,0.025378106,0.0025872455],"about_ca_topic_score_codex":0.00010427268,"about_ca_topic_score_gemma":0.00007482017,"teacher_disagreement_score":0.6958018,"about_ca_system_score_codex":0.00040443562,"about_ca_system_score_gemma":0.0074924356,"threshold_uncertainty_score":0.99997044},"labels":[],"label_agreement":null},{"id":"W1603423056","doi":"10.1186/1472-6920-6-45","title":"On line clinical reasoning assessment with Script Concordance test in urology: results of a French pilot study","year":2006,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":69,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Bonferroni correction; Cronbach's alpha; Test (biology); Concordance; Medicine; Context (archaeology); Medical education; Reliability (semiconductor); Family medicine; Statistics; Clinical psychology; Psychometrics; Mathematics; Internal medicine","score_opus":0.04963797892698821,"score_gpt":0.42678152117102125,"score_spread":0.377143542244033,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1603423056","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9903619,0.00012147838,0.00064258854,0.0024383038,0.00053838146,0.0006614214,0.000005619568,0.000036178342,0.0051941015],"genre_scores_gemma":[0.9932879,0.000028688546,0.004048537,0.0013278453,0.00057702017,0.00007218623,0.000106346626,0.000016557324,0.0005348965],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9969665,0.00030630227,0.0012036738,0.00049222284,0.00079239084,0.00023895029],"domain_scores_gemma":[0.9681628,0.030226808,0.0004086494,0.00054926437,0.00026135257,0.0003911496],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0027089233,0.00015738868,0.0005516855,0.000116086,0.000027731228,0.000008383051,0.00012871125,0.00012722338,0.00010650039],"category_scores_gemma":[0.22269768,0.0001146581,0.00004811743,0.00035853908,0.00023571668,0.000026791195,0.000029495724,0.00061298406,0.000017091576],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011464554,0.031101875,0.95223635,0.000025231693,0.000013516157,0.00003082957,0.00005430344,0.00004157006,0.0000044201574,0.00046265183,0.009019576,0.005863235],"study_design_scores_gemma":[0.009961246,0.022766879,0.9624063,0.0023477403,0.00009142068,0.000020697857,0.00017427187,0.0017021507,0.0000068826034,0.00025150887,0.00016360678,0.0001073107],"about_ca_topic_score_codex":0.0027315915,"about_ca_topic_score_gemma":0.0030876785,"teacher_disagreement_score":0.21998876,"about_ca_system_score_codex":0.00008317541,"about_ca_system_score_gemma":0.0047124317,"threshold_uncertainty_score":0.8359651},"labels":[],"label_agreement":null},{"id":"W1604745203","doi":"10.1111/j.1365-2923.2011.04007.x","title":"Highway spotters and traffic controllers: further reflections on complexity","year":2011,"lang":"en","type":"letter","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Scholarship; Columbia university; Library science; Citation; Curriculum; Sociology; Media studies; Political science; Law; Computer science","score_opus":0.06698879908442568,"score_gpt":0.3832318692433894,"score_spread":0.31624307015896375,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1604745203","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0164443,0.00029237766,0.00002744241,0.96771604,0.0026153913,0.00046093584,0.000012501744,0.00012984705,0.012301169],"genre_scores_gemma":[0.031904608,0.00023727286,0.0003189152,0.9554326,0.0063046403,0.000105010906,0.00069688517,0.000049632465,0.0049504302],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99799174,0.00014248048,0.00041946708,0.0004921776,0.0006512568,0.00030289],"domain_scores_gemma":[0.9956405,0.0033342615,0.00015109431,0.00040044432,0.00014339542,0.0003303155],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00036056785,0.00025977538,0.0005575887,0.00017005138,0.000081019396,0.000023794424,0.000111810055,0.000956785,0.0030026445],"category_scores_gemma":[0.030938558,0.00020130548,0.000131526,0.00015793573,0.00042603866,0.000021062146,0.000014726166,0.001915647,0.00028027705],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009005007,0.0006048451,0.00010527583,0.00005193491,0.00009066615,0.000051013812,0.00034209335,1.0664004e-7,8.06992e-7,0.000074944226,0.9677254,0.030862885],"study_design_scores_gemma":[0.0019441779,0.0006337795,0.006186301,0.0034690364,0.00042763646,0.00015353395,0.00010709684,0.000044771095,0.0000029726293,0.00043030625,0.98632586,0.0002745549],"about_ca_topic_score_codex":0.00018975772,"about_ca_topic_score_gemma":0.00002686632,"teacher_disagreement_score":0.03058833,"about_ca_system_score_codex":0.00011698516,"about_ca_system_score_gemma":0.0014634684,"threshold_uncertainty_score":0.9979088},"labels":[],"label_agreement":null},{"id":"W1605872901","doi":"10.1111/j.1365-2923.2007.02913.x","title":"Found in translation: the impact of familiar symptom descriptions on diagnosis in novices","year":2007,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medical diagnosis; Context (archaeology); Medicine; Psychology; Clinical psychology; Radiology","score_opus":0.036140614749233406,"score_gpt":0.4224017399521412,"score_spread":0.3862611252029078,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1605872901","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98947805,0.0004419213,0.00012664316,0.0065847435,0.00027975856,0.0002689794,0.0000014877182,0.0000105774825,0.0028078293],"genre_scores_gemma":[0.99798775,0.0002478052,0.00014478984,0.0013142405,0.00018574385,0.000047521742,0.000026687745,0.0000069813946,0.000038494734],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9988118,0.00005508559,0.00043253135,0.00014965136,0.0003829736,0.00016795855],"domain_scores_gemma":[0.9909109,0.00857852,0.00007117689,0.00021017127,0.000076747834,0.00015248843],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0011635574,0.000080752536,0.00018748739,0.00016045688,0.00002182916,0.000005643505,0.000081129525,0.00012930755,0.00031184734],"category_scores_gemma":[0.03178308,0.000050732237,0.0000795706,0.00056723016,0.000107069085,0.000034240704,0.0000057578104,0.00027559508,0.00001799431],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010888392,0.0023824305,0.76176363,0.000016253953,0.000011827302,0.0000057322873,0.0012461382,0.000039737944,0.00000994584,0.00022534477,0.0021414314,0.23204866],"study_design_scores_gemma":[0.0009029232,0.00024577195,0.9952767,0.0018237698,0.000018999834,0.0000063423645,0.00038814754,0.0005192446,0.000023800243,0.00029307025,0.0004533641,0.00004789988],"about_ca_topic_score_codex":0.003107765,"about_ca_topic_score_gemma":0.0012086607,"teacher_disagreement_score":0.23351306,"about_ca_system_score_codex":0.00014397752,"about_ca_system_score_gemma":0.00094730256,"threshold_uncertainty_score":0.9763726},"labels":[],"label_agreement":null},{"id":"W1608845916","doi":"10.1111/ans.12180","title":"Clinical decision making: how surgeons do it","year":2013,"lang":"en","type":"review","venue":"ANZ Journal of Surgery","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":94,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Royal Australasian College of Physicians; Royal College of Physicians and Surgeons of Canada; Royal Australasian College of Surgeons; U.S. Consumer Product Safety Commission","keywords":"Subconscious; Medicine; Clinical decision making; Cognition; Process (computing); Coping (psychology); Management science; Computer science; Pathology; Psychiatry","score_opus":0.2294175258277729,"score_gpt":0.4771873679789461,"score_spread":0.24776984215117323,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1608845916","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0006948222,0.9931838,0.0001268892,0.0009500326,0.0042633354,0.00027166528,0.000011834183,0.000019712717,0.00047791787],"genre_scores_gemma":[0.0005751543,0.99467784,0.0010657763,0.0006128489,0.0022988273,0.000007296757,0.000019440191,0.0000837424,0.00065905787],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9939552,0.0004971594,0.0037184234,0.00040087206,0.00094549207,0.00048286337],"domain_scores_gemma":[0.7711087,0.22284155,0.0036070189,0.0007395029,0.00081796065,0.00088526553],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0064912573,0.00048358383,0.0062186727,0.00061747694,0.000059838923,0.00013039018,0.00028204615,0.0008910306,0.0009412258],"category_scores_gemma":[0.21148123,0.00030328505,0.0053085038,0.00046894889,0.00017324039,0.00012931133,0.000106109364,0.001931981,0.0003775509],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000039605777,0.00024147151,0.0059109717,0.0007052833,0.0003315034,0.0012215736,0.0000027589454,4.1314685e-8,5.781008e-9,0.000004905719,0.31648642,0.67505544],"study_design_scores_gemma":[0.0002705799,0.00013343462,0.0015207258,0.22756068,0.0018160492,0.002811455,0.000011108516,5.1835997e-7,3.2487595e-8,0.00008150886,0.76557577,0.00021812793],"about_ca_topic_score_codex":0.0000036367644,"about_ca_topic_score_gemma":5.6174287e-7,"teacher_disagreement_score":0.67483735,"about_ca_system_score_codex":0.00009327148,"about_ca_system_score_gemma":0.0020225937,"threshold_uncertainty_score":0.99997205},"labels":[],"label_agreement":null},{"id":"W1634504077","doi":"10.1136/emermed-2012-201737","title":"Assessing clinical reasoning using a script concordance test with electrocardiogram in an emergency medicine clerkship rotation","year":2013,"lang":"en","type":"article","venue":"Emergency Medicine Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Concordance; Cronbach's alpha; Medicine; Test (biology); Context (archaeology); Reliability (semiconductor); Emergency department; Inter-rater reliability; Post-hoc analysis; Psychology; Internal medicine; Clinical psychology; Psychometrics; Psychiatry","score_opus":0.11019084544569406,"score_gpt":0.45341518134835057,"score_spread":0.3432243359026565,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1634504077","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96964186,0.0032028921,0.0102802,0.003533455,0.0039215353,0.000708586,8.671294e-7,0.00009878886,0.008611791],"genre_scores_gemma":[0.982618,0.0034926676,0.003992668,0.00022557363,0.0076587,0.000040160587,0.000039242703,0.00010450049,0.0018284892],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99307495,0.00056904077,0.003344682,0.00073676923,0.0013056014,0.000968924],"domain_scores_gemma":[0.9929532,0.0023234147,0.0013231515,0.00062033365,0.0013726717,0.0014072658],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004337625,0.00051268213,0.0014478903,0.00051214005,0.00031575415,0.00003099446,0.0003017033,0.00027693986,0.017709324],"category_scores_gemma":[0.06401458,0.0003449679,0.00024593386,0.0015738806,0.00038988015,0.00094758876,0.000038485934,0.0020128845,0.00003213769],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001656139,0.0005334738,0.9490175,0.00004080007,0.00011198559,0.0000518592,0.0005269477,0.000060321152,0.002312534,0.000026070978,0.03750723,0.009645639],"study_design_scores_gemma":[0.0063625053,0.006559422,0.9628661,0.008611794,0.00064174354,0.0010198756,0.0030301395,0.008517804,0.000022646473,0.0006977862,0.0012126402,0.00045750034],"about_ca_topic_score_codex":0.0010852297,"about_ca_topic_score_gemma":0.00015554897,"teacher_disagreement_score":0.05967695,"about_ca_system_score_codex":0.00020849468,"about_ca_system_score_gemma":0.0004868117,"threshold_uncertainty_score":0.9999002},"labels":[],"label_agreement":null},{"id":"W1648486913","doi":"10.1093/oxfordhb/9780199734689.013.0037","title":"Medical Reasoning and Thinking","year":2012,"lang":"en","type":"book-chapter","venue":"Oxford University Press eBooks","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":43,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Waterloo","funders":"","keywords":"Cognition; Focus (optics); Engineering ethics; Cognitive science; Management science; Psychology; Engineering; Neuroscience","score_opus":0.026932365206461088,"score_gpt":0.2613796849403138,"score_spread":0.2344473197338527,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1648486913","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0007540162,0.00089849514,0.00010722029,0.00013833075,0.00017480219,0.00022538716,0.000019232957,0.00013827781,0.9975442],"genre_scores_gemma":[0.0034461683,0.0026369577,0.0008356942,0.00041793933,0.0004324576,3.9554044e-7,0.000048900598,0.0000516197,0.99212986],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.99851143,0.000021872558,0.0002012998,0.0003791964,0.000585596,0.0003006059],"domain_scores_gemma":[0.9968831,0.0018378943,0.00015445372,0.00035735796,0.000080397054,0.0006868072],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0003102832,0.00027631057,0.0005392232,0.00009512276,0.00014717659,0.000021231306,0.0001805276,0.00075868407,0.00025676147],"category_scores_gemma":[0.0017596136,0.00027117116,0.00017503915,0.000004906824,0.0002864699,0.00003777787,0.0003875674,0.00085189095,0.0000029665894],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00023227905,0.000033414755,0.00054606283,0.00009260072,0.00027099022,0.0010196497,0.000253428,1.1896404e-7,0.0000010329181,0.90396553,0.002849744,0.09073515],"study_design_scores_gemma":[0.001208373,0.0001045488,0.0002914058,0.0047969343,0.00068423816,0.00018571832,0.000021588507,0.000028904748,0.0000071656477,0.0006305023,0.99177206,0.00026854387],"about_ca_topic_score_codex":0.000050550792,"about_ca_topic_score_gemma":0.0000066117987,"teacher_disagreement_score":0.98892236,"about_ca_system_score_codex":0.00007775091,"about_ca_system_score_gemma":0.00017756397,"threshold_uncertainty_score":0.9999741},"labels":[],"label_agreement":null},{"id":"W1693843604","doi":"10.1111/jep.12060","title":"Explanation, understanding, objectivity and experience","year":2013,"lang":"en","type":"editorial","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Guelph; Dalhousie University; De Veber; University of Toronto","funders":"","keywords":"Epistemology; Reductionism; Judgement; Objectivity (philosophy); Neglect; Set (abstract data type); Psychology; Sociology; Philosophy; Computer science; Psychiatry","score_opus":0.14716774827234003,"score_gpt":0.5305246916959345,"score_spread":0.3833569434235945,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1693843604","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.023147801,0.0025122077,0.00180527,0.0078375675,0.9570304,0.0009510637,0.0000097955835,0.000016882836,0.006688976],"genre_scores_gemma":[0.17666231,0.011979273,0.0058124987,0.0012334962,0.80375475,0.000044034718,0.000056293968,0.00005846141,0.00039889847],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99257416,0.0014650546,0.0026621367,0.0003711798,0.0027362695,0.00019117541],"domain_scores_gemma":[0.42010012,0.5639403,0.0069791623,0.00065534224,0.007587389,0.00073769887],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.024483971,0.00020286531,0.0009386301,0.00024296694,0.000053655047,0.00009725481,0.00011297035,0.00095994206,0.0003135585],"category_scores_gemma":[0.90471023,0.00016408491,0.00020739884,0.0002336383,0.00019829352,0.00062606944,0.000054306936,0.0027293738,0.000021222686],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00081413117,0.0008333093,0.0046362244,0.0000322256,0.00013374467,0.00010207187,0.00030952142,0.000014239285,8.5701805e-7,0.0000772691,0.9842317,0.00881467],"study_design_scores_gemma":[0.022458786,0.0057819905,0.052828927,0.01088398,0.0042161285,0.0006145308,0.00477431,0.0031475388,0.0000045608863,0.013166848,0.8814159,0.00070647674],"about_ca_topic_score_codex":0.00008280306,"about_ca_topic_score_gemma":0.00001880652,"teacher_disagreement_score":0.88022625,"about_ca_system_score_codex":0.0005643691,"about_ca_system_score_gemma":0.0020193378,"threshold_uncertainty_score":0.9995714},"labels":[],"label_agreement":null},{"id":"W1708384859","doi":"","title":"Learning to act like a doctor","year":2000,"lang":"en","type":"article","venue":"Europe PMC (PubMed Central)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Misfortune; Worry; Computer science; Medicine; World Wide Web; Artificial intelligence; Psychiatry","score_opus":0.01960157858513427,"score_gpt":0.2778056714693966,"score_spread":0.2582040928842623,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1708384859","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95674986,0.00015938861,0.0000142262,0.0028791274,0.00046382516,0.0004535105,0.000004370514,0.0002561944,0.03901949],"genre_scores_gemma":[0.92303824,0.00047697235,0.000476553,0.0037058846,0.00073883863,0.000040913477,0.00002219421,0.000055410455,0.07144501],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99777544,0.00010638342,0.00033205946,0.0004410309,0.00034145507,0.0010036523],"domain_scores_gemma":[0.99697095,0.0013040723,0.0000428498,0.00036793394,0.00006446586,0.0012497209],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00033316738,0.00019611485,0.00032478516,0.00005886436,0.00007456206,0.00005197407,0.00014542829,0.00007380045,0.0028581393],"category_scores_gemma":[0.019626526,0.00016274117,0.00012777875,0.00042345328,0.00004467772,0.00004738474,0.00004114219,0.00041977994,0.0034722625],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007291813,0.00052600144,0.21508375,0.00002163077,0.00009073294,0.0006835448,0.0038796023,0.0001861808,0.00018333264,0.00008208512,0.02885077,0.7496832],"study_design_scores_gemma":[0.0009904766,0.00026254077,0.43998054,0.00028348062,0.000059051577,0.000042784715,0.000019734249,0.000030038622,0.00021982734,0.0000069406083,0.55795276,0.00015184142],"about_ca_topic_score_codex":0.000029611967,"about_ca_topic_score_gemma":0.0000028270167,"teacher_disagreement_score":0.7495313,"about_ca_system_score_codex":0.00007387957,"about_ca_system_score_gemma":0.00011053227,"threshold_uncertainty_score":0.9980534},"labels":[],"label_agreement":null},{"id":"W1751208519","doi":"10.1186/1472-6947-5-11","title":"Optimal search strategies for identifying sound clinical prediction studies in EMBASE","year":2005,"lang":"en","type":"article","venue":"BMC Medical Informatics and Decision Making","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":27,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"U.S. National Library of Medicine","keywords":"MEDLINE; Information retrieval; Computer science; Health informatics; Presentation (obstetrics); Publishing; Gold standard (test); Outcome (game theory); Medical physics; Index (typography); Medicine; World Wide Web; Pathology; Public health; Surgery","score_opus":0.27825110761980293,"score_gpt":0.5364455274296518,"score_spread":0.25819441980984886,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1751208519","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6297126,0.0005453526,0.36892483,0.000053811986,0.00025244954,0.00024981232,0.000005057634,0.000030022911,0.00022607986],"genre_scores_gemma":[0.6781966,0.00088329025,0.31964985,0.0007755949,0.00043329594,0.00002352113,0.00001130739,0.00001023342,0.00001632332],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99695253,0.0000465725,0.0016519082,0.00019459239,0.0008386899,0.00031568512],"domain_scores_gemma":[0.96604913,0.03306818,0.00015617134,0.0002098941,0.00018645327,0.00033014227],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0043703713,0.00014810635,0.0005267921,0.0001796485,0.000115350056,0.0001088125,0.00010057798,0.00024684277,0.000056594665],"category_scores_gemma":[0.04448477,0.00010890751,0.00013960729,0.00016637372,0.00025686115,0.00026229888,0.00015895414,0.00041111023,0.000019595816],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013211297,0.00045598118,0.05590519,0.00070556754,0.00009759375,0.000046865684,0.002770846,0.005139431,7.701458e-7,0.0024882427,0.005052587,0.9260158],"study_design_scores_gemma":[0.005997232,0.0004957712,0.01367747,0.008888498,0.00008320811,0.0001347579,0.01189644,0.9530643,0.0000024455794,0.0051536513,0.00043998455,0.00016620175],"about_ca_topic_score_codex":0.0000024484575,"about_ca_topic_score_gemma":0.00003801395,"teacher_disagreement_score":0.9479249,"about_ca_system_score_codex":0.00005113285,"about_ca_system_score_gemma":0.00041320344,"threshold_uncertainty_score":0.9635639},"labels":[],"label_agreement":null},{"id":"W1772951434","doi":"10.1111/j.1365-2923.2011.04151.x","title":"What is so difficult about managing clinical reasoning difficulties?","year":2012,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":58,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Competence (human resources); Apprenticeship; Medical education; Psychology; Focus group; Cornerstone; Process (computing); Medicine; Social psychology; Sociology; Computer science","score_opus":0.022698909181947665,"score_gpt":0.4004133044897559,"score_spread":0.3777143953078082,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1772951434","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93221116,0.008779282,0.0003310311,0.04598829,0.00942062,0.0003426096,0.0000015273268,0.00017351082,0.0027519646],"genre_scores_gemma":[0.9335616,0.0066498113,0.001568754,0.04776698,0.0051174187,0.000066903456,0.0001772745,0.000040823008,0.0050504133],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9972818,0.0001361942,0.0007587738,0.0003864648,0.0008932645,0.00054350577],"domain_scores_gemma":[0.99421084,0.003678917,0.00019000284,0.00048150422,0.00022145426,0.001217296],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0015898199,0.00020245425,0.00045514113,0.00007455568,0.00015063737,0.00010764568,0.00015382786,0.00036993346,0.0019053258],"category_scores_gemma":[0.084872164,0.00015446734,0.00021033388,0.00028033438,0.00029023862,0.000288752,0.00007691973,0.00066452386,0.0006132285],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000056604236,0.0027924946,0.22500812,0.000044443936,0.000058259833,0.0000068832264,0.0016228189,2.1391969e-7,0.00000393263,0.00057423674,0.19447048,0.57536155],"study_design_scores_gemma":[0.0017328724,0.00017277301,0.84532243,0.0073407553,0.00027007164,0.00014145894,0.0025965476,0.00053206977,0.000048475977,0.00009676742,0.14143977,0.00030600297],"about_ca_topic_score_codex":0.00008557312,"about_ca_topic_score_gemma":0.0000038304165,"teacher_disagreement_score":0.6203143,"about_ca_system_score_codex":0.00009868922,"about_ca_system_score_gemma":0.0010397598,"threshold_uncertainty_score":0.99900705},"labels":[],"label_agreement":null},{"id":"W1783297522","doi":"10.3917/spub.138.0227","title":"Taxonomies et typologies: quelques clarifications complémentaires","year":2014,"lang":"fr","type":"article","venue":"Santé Publique","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Philosophy","score_opus":0.037945205226635276,"score_gpt":0.36970379314463125,"score_spread":0.33175858791799595,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1783297522","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.10236994,0.048749454,0.02601762,0.43918142,0.011159853,0.0016162848,0.00037229553,0.001018453,0.36951467],"genre_scores_gemma":[0.95034176,0.0060981717,0.0057928227,0.028745975,0.0016478605,0.00012065478,0.00029632068,0.000045708915,0.0069107176],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.9973617,0.000618817,0.0006264735,0.0005008274,0.00023168992,0.00066049135],"domain_scores_gemma":[0.97753507,0.020828528,0.00021788504,0.0007581953,0.0002248028,0.00043551865],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0016158761,0.000292695,0.00055005436,0.00013813976,0.00012585215,0.00017127127,0.00028508483,0.00046122342,0.002280871],"category_scores_gemma":[0.08717016,0.0002661236,0.00020511622,0.00021995431,0.00064932776,0.00022764939,0.00013323405,0.00069767464,0.0013085547],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000573757,0.00074022193,0.04507093,0.000091881906,0.0001918645,0.000016445256,0.00042679807,0.000027481847,0.0000697258,0.6220049,0.25761652,0.07368583],"study_design_scores_gemma":[0.0012585882,0.00067261193,0.17429475,0.0024755765,0.00026027398,0.00005008843,0.0009655924,0.0008114496,0.0004317626,0.024900071,0.7934899,0.0003893555],"about_ca_topic_score_codex":0.0018323336,"about_ca_topic_score_gemma":0.00031000376,"teacher_disagreement_score":0.8479718,"about_ca_system_score_codex":0.00018453562,"about_ca_system_score_gemma":0.0005694178,"threshold_uncertainty_score":0.9999791},"labels":[],"label_agreement":null},{"id":"W1789151472","doi":"10.1111/j.1365-2923.2011.03997.x","title":"The biasing effect of clinical history on physical examination diagnostic accuracy","year":2011,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":43,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Physical examination; Interquartile range; Medical diagnosis; Medicine; Medical history; Diagnostic accuracy; Physical therapy; Test (biology); Surgery; Internal medicine; Pathology","score_opus":0.05637453176130086,"score_gpt":0.42281508343669205,"score_spread":0.3664405516753912,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1789151472","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9847772,0.00040553047,0.00004548695,0.0015143333,0.0025319464,0.00033641342,4.6859873e-7,0.000037892096,0.0103507275],"genre_scores_gemma":[0.99686563,0.00026825932,0.00006803759,0.0013836924,0.00095163065,0.00006006895,0.000020477693,0.000014943747,0.0003672319],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99783194,0.0005042959,0.0005792881,0.00024519407,0.00066653494,0.0001727678],"domain_scores_gemma":[0.85505944,0.1435051,0.00034787052,0.0005206595,0.00018770654,0.00037922466],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0028185977,0.00011946042,0.0003677358,0.00004902307,0.000054627802,0.000004270562,0.00014247038,0.00016057656,0.00028023054],"category_scores_gemma":[0.68041813,0.00006988713,0.00018298723,0.00010890927,0.00038053765,0.0000355377,0.000026454662,0.00043844548,0.000107209846],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000242689,0.0020757932,0.04261588,0.00003494991,0.000027568925,0.000006716694,0.00052215037,9.195498e-8,0.000006204902,0.0009710897,0.020917062,0.9325798],"study_design_scores_gemma":[0.0016159714,0.0027362683,0.97693497,0.0023147769,0.00023629774,0.000016121881,0.00006107558,0.0004994864,0.0006077487,0.00025360545,0.014612748,0.00011092482],"about_ca_topic_score_codex":0.00011253682,"about_ca_topic_score_gemma":0.0000034066836,"teacher_disagreement_score":0.9343191,"about_ca_system_score_codex":0.00010597395,"about_ca_system_score_gemma":0.001185879,"threshold_uncertainty_score":0.32227382},"labels":[],"label_agreement":null},{"id":"W1795550258","doi":"10.1007/978-3-642-15231-3_3","title":"Contributions of Psychology to the Design of Diagnostic Decision Support Systems","year":2010,"lang":"en","type":"book-chapter","venue":"IFIP International Federation for Information Processing/IFIP","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Izaak Walton Killam Health Centre; Carleton University","funders":"","keywords":"Bayesian probability; Formative assessment; Decision support system; Set (abstract data type); Computer science; Psychology; Cognition; Artificial intelligence; Management science; Applied psychology; Cognitive psychology; Machine learning; Engineering; Mathematics education","score_opus":0.03638559969736546,"score_gpt":0.37784121943361837,"score_spread":0.34145561973625294,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1795550258","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00020081372,0.00020161242,0.9070671,0.0049853367,0.0033347227,0.0033016482,0.0014812009,0.00007776701,0.07934977],"genre_scores_gemma":[0.8860118,0.00052110973,0.02776296,0.006982688,0.0020636125,0.0013190837,0.012190718,0.00014712705,0.06300091],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.9967865,0.000022262771,0.0018950133,0.0002545417,0.00084712036,0.00019458524],"domain_scores_gemma":[0.9848306,0.009073226,0.0013958202,0.00036461424,0.004178336,0.00015736488],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0012195738,0.00028425496,0.00058119296,0.00040825357,0.00022514888,0.00017042783,0.00031795926,0.00054955174,0.00027900387],"category_scores_gemma":[0.04967497,0.00021656459,0.00022703108,0.00008427269,0.00015290739,0.00037208435,0.000053086453,0.00039890295,0.00017620959],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0052981446,0.000850796,0.00035615088,0.0012025451,0.0010407172,0.000010611116,0.0023383666,0.01189641,0.0013422655,0.4136499,0.22649917,0.33551493],"study_design_scores_gemma":[0.0052629826,0.0017631066,0.00091706397,0.0072374307,0.00060302817,0.00022685798,0.00012200085,0.011758323,0.0029166394,0.014613469,0.9539507,0.00062837516],"about_ca_topic_score_codex":0.000012644414,"about_ca_topic_score_gemma":0.000006300695,"teacher_disagreement_score":0.885811,"about_ca_system_score_codex":0.00010946536,"about_ca_system_score_gemma":0.0006577176,"threshold_uncertainty_score":0.95833004},"labels":[],"label_agreement":null},{"id":"W1797266117","doi":"10.1111/j.1365-2753.2007.00893.x","title":"What’s in a name? A commentary on Tonelli (2007) ‘Advancing a casuistic model of clinical decision making: a response to commentators’","year":2007,"lang":"en","type":"letter","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Women's College Hospital; University of Toronto","funders":"","keywords":"Citation; Clinical Practice; Library science; Mental health; Psychology; Medicine; Psychiatry; Family medicine; Computer science","score_opus":0.18538692598523762,"score_gpt":0.5743430294043432,"score_spread":0.3889561034191056,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1797266117","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.19835046,0.00072415714,0.0047492804,0.78649235,0.0075965296,0.0016934882,0.000021151012,0.000011542574,0.00036101943],"genre_scores_gemma":[0.0784226,0.0016156832,0.04392192,0.8716377,0.0041978224,0.000023758897,0.000027943992,0.000080313504,0.00007224135],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9730175,0.008139941,0.012822077,0.0008869744,0.00444453,0.0006889738],"domain_scores_gemma":[0.4534134,0.5359556,0.006445372,0.000990857,0.0026446742,0.00055009295],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["metaresearch","research_integrity"],"category_scores_codex":[0.12969902,0.00046888043,0.002812309,0.0012282316,0.00004657402,0.00007720257,0.00040662973,0.001595473,0.0002587285],"category_scores_gemma":[0.83645636,0.00040804883,0.0009274034,0.0006941015,0.00021308417,0.00048921455,0.00017478812,0.009356645,0.000052520652],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.071563646,0.0045357305,0.012289998,0.000070202484,0.00023430407,0.0040248223,0.0003296094,0.008379484,0.0000016546476,0.0000029659739,0.84183437,0.056733187],"study_design_scores_gemma":[0.070941016,0.03821372,0.08308777,0.16404429,0.0070140082,0.00172595,0.003142211,0.111141235,0.000008303484,0.005773784,0.5131564,0.0017513195],"about_ca_topic_score_codex":0.000056320885,"about_ca_topic_score_gemma":0.000034618235,"teacher_disagreement_score":0.70675737,"about_ca_system_score_codex":0.0011120064,"about_ca_system_score_gemma":0.0023968485,"threshold_uncertainty_score":0.99983716},"labels":[],"label_agreement":null},{"id":"W1811267435","doi":"10.7202/1072334ar","title":"Socratic Open-mindedness","year":2020,"lang":"en","type":"article","venue":"Paideusis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"Mount Saint Vincent University","funders":"Mount Saint Vincent University","keywords":"Socratic method; SOCRATES; Skepticism; Epistemology; Argument (complex analysis); Philosophy; Context (archaeology); Ideal (ethics); Interpretation (philosophy); Socratic questioning; Value (mathematics); Computer science; Medicine; History","score_opus":0.09372993926913116,"score_gpt":0.37818597362078166,"score_spread":0.2844560343516505,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1811267435","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.80988735,0.0004174659,0.002006426,0.11246995,0.00037090076,0.0008003723,0.0000153092,0.0002474459,0.07378476],"genre_scores_gemma":[0.9758331,0.000030649713,0.001730956,0.021037241,0.0002539235,0.00002095333,0.000016492142,0.000016225546,0.0010604272],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9992054,0.00002886505,0.00020834083,0.0002326758,0.00016375976,0.00016096112],"domain_scores_gemma":[0.9968779,0.0024565991,0.000043036445,0.00020147128,0.000051797328,0.0003691809],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00016510415,0.00008719348,0.00028225878,0.0000137939005,0.00004420247,0.00004092162,0.00015316364,0.00005796052,0.0013653294],"category_scores_gemma":[0.0381406,0.000069918155,0.0000792948,0.00017631354,0.00003969198,0.000041273724,0.0001319484,0.00013185569,0.00089196675],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00076070376,0.0012109041,0.31212288,0.00021730473,0.00033948862,0.00081414875,0.0027234969,0.000025772346,0.0019851439,0.0024487083,0.4203167,0.25703475],"study_design_scores_gemma":[0.03411309,0.006601489,0.6174706,0.007413642,0.0029473167,0.00027299387,0.003732629,0.01528802,0.015814189,0.011095499,0.28271988,0.0025306505],"about_ca_topic_score_codex":0.00005544481,"about_ca_topic_score_gemma":0.000003091664,"teacher_disagreement_score":0.3053477,"about_ca_system_score_codex":0.000014268396,"about_ca_system_score_gemma":0.000108317814,"threshold_uncertainty_score":0.999886},"labels":[],"label_agreement":null},{"id":"W1811945388","doi":"10.1111/j.1365-2753.2007.00877.x","title":"A commentary on Murray <i>et al</i>. (2007) ‘No exit? Intellectual integrity under the regime of “evidence” and “best practices”’","year":2007,"lang":"en","type":"letter","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Associate editor; Citation; Library science; Psychology; Computer science","score_opus":0.4025175245176905,"score_gpt":0.5706118379101258,"score_spread":0.1680943133924353,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1811945388","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.007000943,0.0024871,0.0003026891,0.98118955,0.002958912,0.0008431886,0.00000758613,0.000007555668,0.005202453],"genre_scores_gemma":[0.0045181406,0.022240825,0.0033944445,0.9648362,0.004570827,0.000011074959,0.000029817369,0.000044328255,0.00035433483],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9840186,0.0073646777,0.0043560197,0.00055506965,0.0033452914,0.00036030856],"domain_scores_gemma":[0.24071366,0.74208057,0.011189776,0.0009270541,0.004794338,0.00029462433],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.07857837,0.00037106863,0.001336922,0.00048620295,0.000072861025,0.00008498383,0.00034646754,0.0009685982,0.00071159797],"category_scores_gemma":[0.88475865,0.00021804565,0.00047685814,0.00054573925,0.00061588304,0.00059402315,0.0001334802,0.0146587035,0.000090639085],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0044637946,0.0022450944,0.0011501372,0.00005947765,0.0004326003,0.0004328983,0.00016354854,0.000024607789,0.0000021246638,0.000012768095,0.9835926,0.007420291],"study_design_scores_gemma":[0.0040714256,0.0074297753,0.0028407613,0.012884046,0.0030320666,0.0011165625,0.0008436889,0.00045629495,0.000005369464,0.00043715848,0.96664876,0.00023411363],"about_ca_topic_score_codex":0.00039298087,"about_ca_topic_score_gemma":0.00002434028,"teacher_disagreement_score":0.8061803,"about_ca_system_score_codex":0.00020922333,"about_ca_system_score_gemma":0.0014480099,"threshold_uncertainty_score":0.9876145},"labels":[],"label_agreement":null},{"id":"W1836271141","doi":"10.1111/medu.12080","title":"Checklists improve experts’ diagnostic decisions","year":2013,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":93,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; University Health Network","funders":"","keywords":"Checklist; Context (archaeology); Interpretation (philosophy); Computer science; Cognition; Diagnostic accuracy; Medical physics; Medicine; Psychology; Cognitive psychology; Radiology; Psychiatry","score_opus":0.015139960759559618,"score_gpt":0.3605869921277853,"score_spread":0.34544703136822563,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1836271141","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9406537,0.0006638164,0.0003680169,0.049886357,0.0021587487,0.00056004873,0.0000013374507,0.0001053281,0.0056026853],"genre_scores_gemma":[0.97192496,0.0002734278,0.001430096,0.021684581,0.001035632,0.00040830337,0.00008835031,0.000020818074,0.0031338288],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9985008,0.000029872694,0.00035671834,0.00027566674,0.0005925174,0.00024437258],"domain_scores_gemma":[0.9733502,0.024393396,0.0000997233,0.0005994131,0.00036022003,0.0011970401],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00025055613,0.0001164929,0.00022302946,0.000062756044,0.00004903516,0.00002940792,0.00011269171,0.00019738774,0.0076864534],"category_scores_gemma":[0.5659319,0.00008633298,0.00009242405,0.00016713243,0.000092651084,0.000060886876,0.000040267943,0.00022375796,0.0015505061],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000094365205,0.0014998225,0.022108693,0.000009660219,0.000014760801,0.000009702184,0.00017420012,1.3351764e-7,0.00010981013,0.00034816714,0.51552504,0.4601906],"study_design_scores_gemma":[0.0025097867,0.00052034925,0.88249075,0.004967599,0.00014817463,0.00017705608,0.0005204818,0.0014125473,0.00047395148,0.0025200103,0.103868805,0.00039049942],"about_ca_topic_score_codex":0.00045671352,"about_ca_topic_score_gemma":0.00000637996,"teacher_disagreement_score":0.860382,"about_ca_system_score_codex":0.0000682389,"about_ca_system_score_gemma":0.0014171447,"threshold_uncertainty_score":0.9992269},"labels":[],"label_agreement":null},{"id":"W1848713823","doi":"","title":"How doctors think","year":2009,"lang":"en","type":"article","venue":"Canadian Family Physician","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Strengths and weaknesses; Computer science; Data science; Library science; World Wide Web; Psychology","score_opus":0.023518676903017843,"score_gpt":0.2793744724149954,"score_spread":0.25585579551197757,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1848713823","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.47841528,0.0007918145,0.000025536407,0.032440264,0.0005190561,0.00033678763,0.000040325154,0.00015155395,0.4872794],"genre_scores_gemma":[0.76007426,0.000019489695,0.00015021075,0.23693758,0.00052475877,0.0000022841352,0.000048082402,0.0000137321895,0.002229627],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9989623,0.000015998843,0.00013215483,0.00025863323,0.00018000341,0.00045094977],"domain_scores_gemma":[0.9980949,0.00039229542,0.000036646445,0.00037118752,0.000082155704,0.0010227924],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.000072421055,0.00015071066,0.00027251162,0.00010977665,0.00008889145,0.000073569565,0.00010425376,0.00010311702,0.000004215085],"category_scores_gemma":[0.0014393462,0.00013577328,0.00013868221,0.00024917396,0.000056643352,0.00007038759,0.0000049061555,0.0002201306,0.00017723633],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000029939823,0.0001732321,0.00096797623,0.000007803279,0.000061901475,0.0005081798,0.00011865154,0.0000039047463,0.0013825442,0.01075978,0.26499122,0.72099483],"study_design_scores_gemma":[0.0018587294,0.00086076075,0.4929213,0.0009828699,0.00018521388,0.000007050783,0.00063302525,0.00010695022,0.00054637785,0.0066436958,0.49469197,0.00056205393],"about_ca_topic_score_codex":0.003379922,"about_ca_topic_score_gemma":0.0012720871,"teacher_disagreement_score":0.7204328,"about_ca_system_score_codex":0.0001637474,"about_ca_system_score_gemma":0.0005766576,"threshold_uncertainty_score":0.5536674},"labels":[],"label_agreement":null},{"id":"W1867292416","doi":"10.36834/cmej.36589","title":"Experienced physicians benefit from analyzing initial diagnostic hypotheses","year":2013,"lang":"en","type":"article","venue":"Canadian Medical Education Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"University of Calgary","funders":"","keywords":"Odds ratio; Nephrology; Medicine; Odds; Confidence interval; Medical diagnosis; Internal medicine; Medical decision making; Cognition; Closure (psychology); Metacognition; Family medicine; Logistic regression; Pathology; Psychiatry","score_opus":0.01870167998650206,"score_gpt":0.33143537246929283,"score_spread":0.31273369248279076,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1867292416","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9472001,0.0008545416,0.00009457703,0.045362625,0.0019805932,0.0002001358,0.0000095467,0.000025044044,0.0042727967],"genre_scores_gemma":[0.93828523,0.00023128292,0.00062107417,0.056944165,0.0033892507,0.000054044074,0.000046025183,0.00002230753,0.0004065979],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.998217,0.00005335129,0.00046063837,0.00023156244,0.0005935196,0.00044396735],"domain_scores_gemma":[0.97976404,0.006872748,0.00017165348,0.00032314274,0.0005356474,0.012332779],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00022433646,0.00014790619,0.0002962307,0.00019604781,0.00021404232,0.00013279433,0.00020255137,0.00019421407,0.07475568],"category_scores_gemma":[0.33370057,0.00012724208,0.00013103474,0.00023520025,0.00017657113,0.00014506167,0.000013050263,0.0006665812,0.0005922691],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000010886756,0.0003489621,0.03195932,0.00000481471,0.000075708034,0.00018743271,0.0015033805,0.000003363169,0.000026538928,0.0002240132,0.17383562,0.79181993],"study_design_scores_gemma":[0.0029262954,0.00047346792,0.88034755,0.0085235955,0.00033587712,0.0012070177,0.009523459,0.0006939392,0.00023368491,0.008923915,0.086052366,0.00075881527],"about_ca_topic_score_codex":0.053073477,"about_ca_topic_score_gemma":0.0059885564,"teacher_disagreement_score":0.84838825,"about_ca_system_score_codex":0.00025018377,"about_ca_system_score_gemma":0.01655443,"threshold_uncertainty_score":0.98902076},"labels":[],"label_agreement":null},{"id":"W1875877400","doi":"10.1111/j.1553-2712.2011.01051.x","title":"In Support of a Journey of Metacognition","year":2011,"lang":"en","type":"article","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Public Health Ontario; University of Toronto","funders":"","keywords":"Medicine; Chart; Recall; Medical diagnosis; Medical education; Psychology; Cognitive psychology","score_opus":0.15759858813314523,"score_gpt":0.4239776352404441,"score_spread":0.26637904710729887,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1875877400","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9600786,0.0021570085,0.0000860646,0.0016160848,0.0007748653,0.00023232841,0.000006759782,0.000015966232,0.035032343],"genre_scores_gemma":[0.98888093,0.009331413,0.00021878908,0.0004735756,0.0002264832,0.0000103006405,0.00001853915,0.000013698,0.00082624814],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.997743,0.000046197787,0.0014778367,0.00018311896,0.00035416437,0.00019566227],"domain_scores_gemma":[0.99831825,0.0007198517,0.0004048075,0.00020644497,0.00017898204,0.00017166643],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0010825848,0.00011697018,0.0005883269,0.00023100675,0.0000073383612,2.0850887e-8,0.00010966657,0.00019439029,0.011048571],"category_scores_gemma":[0.026764013,0.00008655704,0.0000979229,0.000448239,0.00017148671,0.000054285883,0.000025846928,0.00043820296,0.000023099738],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005031724,0.0005498626,0.8942009,0.00026478033,0.00013692008,0.00004641093,0.005023618,3.2414286e-7,0.018483816,0.0018427923,0.07735598,0.0015913983],"study_design_scores_gemma":[0.008928665,0.004759407,0.9030351,0.0068104165,0.0014903621,0.00012714782,0.001268061,0.00004932768,0.048569746,0.019793244,0.004815481,0.00035302786],"about_ca_topic_score_codex":0.00037276384,"about_ca_topic_score_gemma":0.0000061351225,"teacher_disagreement_score":0.0725405,"about_ca_system_score_codex":0.000012832135,"about_ca_system_score_gemma":0.00007621713,"threshold_uncertainty_score":0.98985547},"labels":[],"label_agreement":null},{"id":"W1900949867","doi":"10.1111/j.1365-2923.2012.04311.x","title":"Should we teach using schemas? Evidence from a randomised trial","year":2012,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":39,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Western Hospital; University of Toronto; University Health Network","funders":"","keywords":"Schema (genetic algorithms); Medical diagnosis; Medical knowledge; Psychology; Diagnostic accuracy; Medicine; Medical education; Computer science; Internal medicine; Radiology; Machine learning","score_opus":0.12273860134725248,"score_gpt":0.46001974535068607,"score_spread":0.33728114400343356,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1900949867","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97084856,0.004705367,0.0017636843,0.018335056,0.0030960855,0.0006053109,0.0000010648456,0.00005773494,0.00058711553],"genre_scores_gemma":[0.98318875,0.00038082423,0.0051251203,0.004752205,0.00592945,0.00008760835,0.000048254977,0.000018543164,0.00046922878],"study_design_codex":"design_other","study_design_gemma":"randomized_trial","domain_scores_codex":[0.99815595,0.00014239612,0.00043353665,0.00022385114,0.0007681199,0.0002761518],"domain_scores_gemma":[0.98632234,0.012291222,0.00013056633,0.00035681354,0.0001006188,0.00079841673],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0014449792,0.000120945675,0.00034753364,0.000052525553,0.00005514011,0.000013560594,0.000086398606,0.00025343904,0.0029582176],"category_scores_gemma":[0.33342808,0.00009044212,0.00013120295,0.00015472122,0.000095998665,0.00011056215,0.000030208901,0.000387859,0.000148469],"study_design_candidate":"randomized_trial","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.10352687,0.012954945,0.12901197,0.00013098202,0.00025406422,0.00001772791,0.005007956,0.0000017244371,0.0017747445,0.00034315043,0.19324666,0.55372924],"study_design_scores_gemma":[0.7464346,0.0014204172,0.07837397,0.053146344,0.0032130457,0.00021719573,0.002026934,0.007427635,0.0041901744,0.00095089315,0.10127697,0.0013218125],"about_ca_topic_score_codex":0.00046047985,"about_ca_topic_score_gemma":0.0000029594246,"teacher_disagreement_score":0.64290774,"about_ca_system_score_codex":0.000106377105,"about_ca_system_score_gemma":0.001890821,"threshold_uncertainty_score":0.99795324},"labels":[],"label_agreement":null},{"id":"W1911295571","doi":"10.1111/j.1365-2753.2011.01727.x","title":"Mere anecdote: evidence and stories in medicine","year":2011,"lang":"en","type":"article","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":30,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria; University of Toronto","funders":"","keywords":"Anecdote; Alternative medicine; Medicine; Psychology; Literature; Art; Pathology","score_opus":0.3799160249335071,"score_gpt":0.5787985601705926,"score_spread":0.1988825352370855,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1911295571","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9580381,0.008282732,0.00018026498,0.025560392,0.0011871678,0.00034412846,3.531336e-7,0.0000061802452,0.0064007053],"genre_scores_gemma":[0.98363835,0.0059006126,0.0077796327,0.0021085977,0.00051025295,0.0000040283803,6.3832886e-7,0.000008290793,0.00004958993],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9954672,0.0011613829,0.0021026074,0.00019815679,0.0009336629,0.00013702159],"domain_scores_gemma":[0.83251935,0.16301343,0.0018365124,0.00032633042,0.0019333478,0.0003710422],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.033305325,0.00009781936,0.0005808698,0.00017376877,0.000019171855,0.000008706656,0.00007561513,0.0001667216,0.0004980977],"category_scores_gemma":[0.83205354,0.00007030199,0.00007366846,0.00030251223,0.000214535,0.0005200933,0.000030638275,0.0009619787,0.0000100133275],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004782797,0.0018677152,0.86163056,0.000040954335,0.000090720445,0.00072818383,0.0027196656,0.000020522153,0.000021597361,0.00052746275,0.0045397384,0.12303007],"study_design_scores_gemma":[0.005412503,0.0029505522,0.9778184,0.0053747026,0.0005983329,0.0005879969,0.001519593,0.00061497826,0.000008003855,0.0034171746,0.0016161695,0.00008154059],"about_ca_topic_score_codex":0.00020557058,"about_ca_topic_score_gemma":0.000054504948,"teacher_disagreement_score":0.79874825,"about_ca_system_score_codex":0.00008937587,"about_ca_system_score_gemma":0.00045853894,"threshold_uncertainty_score":0.99541557},"labels":[],"label_agreement":null},{"id":"W1917239007","doi":"10.1002/j.0022-0337.2009.73.10.tb04810.x","title":"The Role of Basic Sciences in Diagnostic Oral Radiology","year":2009,"lang":"en","type":"article","venue":"Journal of Dental Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":56,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Feature (linguistics); Mathematics education; Test (biology); Computer science; Coherence (philosophical gambling strategy); Medicine; Psychology; Mathematics; Statistics; Linguistics","score_opus":0.014396167648183283,"score_gpt":0.36496464919589944,"score_spread":0.35056848154771614,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1917239007","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9920281,0.0036862313,0.000003053829,0.0022043502,0.00051879534,0.00005743474,2.540817e-7,0.0000012168945,0.0015005532],"genre_scores_gemma":[0.9986431,0.0005049825,0.00025711244,0.00034494052,0.00017766848,7.7980155e-7,7.431657e-7,0.0000015597677,0.000069113376],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9992982,0.00004599109,0.0003641598,0.000049338178,0.00015857868,0.00008372179],"domain_scores_gemma":[0.9942923,0.005294581,0.0002301756,0.000056370154,0.00007385877,0.00005267774],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00062704086,0.000037111044,0.00014429924,0.000081656915,0.000030513062,0.0000085208085,0.00008006099,0.00003050072,0.000024020033],"category_scores_gemma":[0.03765354,0.000022215272,0.000057345722,0.00016149376,0.00010306297,0.000044925797,0.0000046513287,0.00012071304,0.0000057573507],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011653544,0.00090003776,0.71501076,0.0000023420496,0.000009562376,0.000007658009,0.0002382455,0.000018947974,0.00095819484,0.0011124961,0.0025039078,0.2791213],"study_design_scores_gemma":[0.00055037974,0.0011128274,0.9893271,0.0005607697,0.00003672069,0.0005166953,0.0016348462,0.000030524814,0.0010728638,0.0042013605,0.0009277444,0.000028182052],"about_ca_topic_score_codex":0.000031816213,"about_ca_topic_score_gemma":0.000009924785,"teacher_disagreement_score":0.27909312,"about_ca_system_score_codex":0.000044542674,"about_ca_system_score_gemma":0.00044479407,"threshold_uncertainty_score":0.9704527},"labels":[],"label_agreement":null},{"id":"W1917449996","doi":"10.1111/j.1365-2753.2011.01749.x","title":"Four alternatives to a reductive view of knowledge: a commentary","year":2011,"lang":"en","type":"letter","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Citation; Library science; Psychology; Medicine; Computer science","score_opus":0.23470678065379585,"score_gpt":0.5469815447127633,"score_spread":0.3122747640589674,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1917449996","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.011536024,0.0026162793,0.0003280453,0.9716915,0.0037530593,0.0011368105,0.000015911392,0.0000069480557,0.008915406],"genre_scores_gemma":[0.02566011,0.0035070344,0.023686262,0.93307126,0.013413566,0.000044247612,0.000039797484,0.00007393834,0.0005037758],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9897471,0.0037667088,0.0042310045,0.00039173427,0.001611836,0.00025159377],"domain_scores_gemma":[0.8347948,0.15329175,0.005312622,0.0005960789,0.005676292,0.00032846176],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.023281807,0.00026738818,0.0016289877,0.00043606266,0.000024368926,0.00001679175,0.0002888741,0.00067636144,0.0006106632],"category_scores_gemma":[0.60741174,0.00021247992,0.00063150894,0.00039990523,0.00018553293,0.0002721993,0.00011340076,0.004607209,0.00007558436],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018863459,0.0021399586,0.0019910655,0.00010864365,0.0007041208,0.00066821155,0.00073090737,0.00000944792,0.0000029960322,0.000009843557,0.92216086,0.069587566],"study_design_scores_gemma":[0.008230906,0.0074503636,0.03962638,0.018917933,0.0052729123,0.0010795394,0.00042817535,0.00015067503,0.000029015579,0.0028914204,0.91558146,0.00034124384],"about_ca_topic_score_codex":0.000082288956,"about_ca_topic_score_gemma":0.0000051548864,"teacher_disagreement_score":0.5841299,"about_ca_system_score_codex":0.0002867052,"about_ca_system_score_gemma":0.0012981104,"threshold_uncertainty_score":0.9976892},"labels":[],"label_agreement":null},{"id":"W1936375832","doi":"10.1111/jep.12040","title":"A short note on probability in clinical medicine","year":2013,"lang":"en","type":"article","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Probabilistic logic; Interpretation (philosophy); Epistemology; Psychology; Medicine; Philosophy; Linguistics","score_opus":0.214367857579418,"score_gpt":0.5743568763757636,"score_spread":0.3599890187963456,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1936375832","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9109524,0.000277176,0.00018752021,0.07295019,0.0019580047,0.0011863653,6.563712e-7,0.000011192746,0.012476507],"genre_scores_gemma":[0.9826376,0.00073899055,0.00604422,0.008782675,0.0016784946,0.000028689608,0.0000044778044,0.000017843658,0.000067050125],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.98737764,0.003565001,0.0064200847,0.0004408585,0.0019115945,0.0002847905],"domain_scores_gemma":[0.5345737,0.45411083,0.0034940126,0.0011944057,0.0054985685,0.0011285113],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.0748847,0.00018713456,0.0011944757,0.00024915475,0.00002608951,0.000023787383,0.00017000188,0.0004473902,0.00095860363],"category_scores_gemma":[0.8870502,0.00012964182,0.00036874574,0.00049653975,0.0003192905,0.00043140948,0.00004611285,0.0028117497,0.00018767793],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0031902727,0.0071021398,0.54582214,0.000022818193,0.0000883035,0.00021417187,0.00013770687,0.0003455508,0.000006920519,0.0001941733,0.014652714,0.4282231],"study_design_scores_gemma":[0.009146026,0.005136672,0.96581095,0.0025908896,0.00043679622,0.00017615876,0.00016275024,0.005577456,0.00000262656,0.00798041,0.0028561205,0.00012314373],"about_ca_topic_score_codex":0.00011164895,"about_ca_topic_score_gemma":0.000032330878,"teacher_disagreement_score":0.8121655,"about_ca_system_score_codex":0.0002797944,"about_ca_system_score_gemma":0.00085781515,"threshold_uncertainty_score":0.99995464},"labels":[],"label_agreement":null},{"id":"W1936663290","doi":"10.1007/s11606-015-3369-4","title":"Reflecting on Diagnostic Errors: Taking a Second Look is Not Enough","year":2015,"lang":"en","type":"article","venue":"Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":67,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"University of Toronto","keywords":"Medicine; MEDLINE","score_opus":0.15217967745931776,"score_gpt":0.44865656682016825,"score_spread":0.2964768893608505,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1936663290","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95110273,0.0016356456,0.00037532594,0.02034708,0.0029057078,0.0001341893,0.0000040730692,0.000023695748,0.023471551],"genre_scores_gemma":[0.96078426,0.0001903769,0.0016595855,0.018561695,0.0072959713,0.0000028250745,0.0000033510148,0.000039873135,0.011462059],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9969513,0.000105426654,0.0012618977,0.00026038266,0.0010574209,0.0003635775],"domain_scores_gemma":[0.9897293,0.0072318306,0.0011312234,0.00027921662,0.0007172691,0.00091115694],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0015318872,0.00026407218,0.0008945076,0.0003059698,0.000046171324,0.000023840797,0.0002364928,0.00014252405,0.0011932291],"category_scores_gemma":[0.123704076,0.00017121229,0.00024559168,0.0002114978,0.00015354015,0.000098826444,0.00005723505,0.0010114904,0.000079526326],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.007543362,0.0014547842,0.07118932,0.00014456002,0.0012125799,0.015515716,0.007969834,0.0002102704,0.014275718,0.0005315451,0.78960896,0.09034335],"study_design_scores_gemma":[0.08538891,0.10528559,0.18303579,0.1432647,0.0040856134,0.06266767,0.007109354,0.0036365371,0.11068742,0.0051340605,0.2875352,0.002169157],"about_ca_topic_score_codex":0.0002513855,"about_ca_topic_score_gemma":0.000014204894,"teacher_disagreement_score":0.50207376,"about_ca_system_score_codex":0.00024958432,"about_ca_system_score_gemma":0.00026849413,"threshold_uncertainty_score":0.9997198},"labels":[],"label_agreement":null},{"id":"W1947015618","doi":"10.1002/jhm.2505","title":"Incorporating metacognition into morbidity and mortality rounds: The next frontier in quality improvement","year":2015,"lang":"en","type":"review","venue":"Journal of Hospital Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University Health Network; Mount Sinai Hospital; University of Toronto","funders":"","keywords":"Heuristics; Metacognition; Context (archaeology); Medicine; Frontier; Quality management; Cognition; Perspective (graphical); Quality (philosophy); Medical education; Management science; Cognitive psychology; Psychology; Computer science; Artificial intelligence; Psychiatry; Operations management; Epistemology","score_opus":0.14984778095860188,"score_gpt":0.4540727786776215,"score_spread":0.3042249977190197,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1947015618","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.03181347,0.96289015,0.000056107703,0.0026245767,0.0016087381,0.0007766231,0.000008034318,0.000009317383,0.00021297311],"genre_scores_gemma":[0.014053704,0.9828277,0.00038753246,0.0003147568,0.0022530488,0.000024998479,0.00004774302,0.000032175532,0.00005830034],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.995311,0.000403073,0.0026441594,0.00031334592,0.0010876356,0.0002407606],"domain_scores_gemma":[0.9921214,0.0038420355,0.002559235,0.00041314057,0.00063318893,0.0004310167],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0074963183,0.00039436168,0.0033193682,0.00020480204,0.00005349799,0.000033960772,0.00021115711,0.0002879172,0.000030534295],"category_scores_gemma":[0.050961398,0.00019465254,0.00038567817,0.00035393826,0.00039239763,0.00017958257,0.00011613352,0.0014126116,0.0000045608085],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000049102167,0.00042255843,0.0032645557,0.002502509,0.00045505143,0.00025244593,0.00035067066,1.8611307e-7,0.0000011728137,0.0000586834,0.002973833,0.9896692],"study_design_scores_gemma":[0.016969347,0.028263628,0.04768531,0.401536,0.02580899,0.0009822211,0.006337662,0.000050613016,0.000005075086,0.026397083,0.44438142,0.0015826593],"about_ca_topic_score_codex":0.0008338364,"about_ca_topic_score_gemma":0.000051150168,"teacher_disagreement_score":0.9880866,"about_ca_system_score_codex":0.00033664715,"about_ca_system_score_gemma":0.0006505875,"threshold_uncertainty_score":0.95703274},"labels":[],"label_agreement":null},{"id":"W1960219531","doi":"10.5430/jnep.v5n12p96","title":"Training paradigms to enhance clinical observational skills in clinical practice: A scoping review","year":2015,"lang":"en","type":"review","venue":"Journal of Nursing Education and Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Observational study; Perception; Psychology; Clinical Practice; Medical education; Conflation; Applied psychology; Medicine; Nursing; Pathology","score_opus":0.5416661527596015,"score_gpt":0.6847609140823491,"score_spread":0.14309476132274757,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1960219531","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000019895451,0.9609478,0.0001318297,0.026314896,0.004050657,0.0012003095,0.0000029080527,0.000010570222,0.0073211268],"genre_scores_gemma":[0.000007493239,0.9381954,0.043557227,0.0150221065,0.0028324202,0.000028526438,0.000031718468,0.000040726005,0.00028439536],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9905794,0.0028862364,0.0049606916,0.00048601232,0.00079908624,0.00028857222],"domain_scores_gemma":[0.83377236,0.15540184,0.0065726386,0.00055056426,0.0023270005,0.001375589],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.025523726,0.00032964625,0.0032196264,0.00028106032,0.00008164142,0.0000757334,0.00019586027,0.0004576584,0.000036776062],"category_scores_gemma":[0.6824924,0.00026499806,0.00055994454,0.00070365577,0.0001310445,0.0005852219,0.000050782084,0.0022107025,0.000035042434],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001812041,0.0039128824,0.000031278047,0.0047958824,0.0001198547,0.000059138343,0.0005165103,2.5625343e-7,7.87362e-9,0.00006072652,0.030449169,0.9598731],"study_design_scores_gemma":[0.0002668465,0.00048103352,0.00014559859,0.4776247,0.002288844,0.0059337574,0.0002785722,5.551319e-7,7.31769e-9,0.000052372485,0.51279485,0.00013280493],"about_ca_topic_score_codex":0.0000047429608,"about_ca_topic_score_gemma":7.771596e-7,"teacher_disagreement_score":0.9597403,"about_ca_system_score_codex":0.00034566442,"about_ca_system_score_gemma":0.016104175,"threshold_uncertainty_score":0.9999802},"labels":[],"label_agreement":null},{"id":"W1964143014","doi":"10.1111/j.1553-2712.2002.tb01574.x","title":"Achieving Quality in Clinical Decision Making: Cognitive Strategies and Detection of Bias","year":2002,"lang":"en","type":"article","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":692,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Heuristics; Cornerstone; Cognition; Clinical decision making; Medicine; Quality (philosophy); Cognitive bias; Emergency department; Resource (disambiguation); Cognitive load; Principal (computer security); Quality management; Cognitive psychology; Psychology; Operations management; Intensive care medicine; Computer science; Psychiatry; Computer security","score_opus":0.20829777717127104,"score_gpt":0.4949761063994055,"score_spread":0.2866783292281344,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1964143014","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98465335,0.010234818,0.0009954291,0.00065607147,0.00067988224,0.00027133516,0.0000025024244,0.000036462152,0.0024701343],"genre_scores_gemma":[0.97413576,0.024994668,0.000098784796,0.00021063947,0.00044582324,0.0000090998665,0.0000046393,0.000014939829,0.000085638225],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99645495,0.00025027725,0.0022208898,0.00039697907,0.00043358895,0.0002433046],"domain_scores_gemma":[0.982451,0.016498141,0.00049906137,0.00020224942,0.00015662039,0.00019291209],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00319583,0.00017195927,0.00071367854,0.0001803801,0.000036279904,6.4183774e-7,0.00008064549,0.00036778676,0.0006587207],"category_scores_gemma":[0.17284155,0.00012954454,0.00009793556,0.00045239166,0.00031929262,0.00010996584,0.00005304655,0.00096825627,0.000013617313],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00032049904,0.00014905457,0.86585575,0.000072201336,0.000034401703,0.000010431985,0.00065363606,0.0000012065472,0.0009356578,0.00018056095,0.000384076,0.13140252],"study_design_scores_gemma":[0.0026485967,0.0008803215,0.98566234,0.006722609,0.00015977105,0.000020570575,0.0019464444,0.00038535715,0.00006957038,0.001258618,0.000119545344,0.00012624555],"about_ca_topic_score_codex":0.00013605042,"about_ca_topic_score_gemma":0.00004218319,"teacher_disagreement_score":0.16964573,"about_ca_system_score_codex":0.000021218311,"about_ca_system_score_gemma":0.000032246204,"threshold_uncertainty_score":0.83412594},"labels":[],"label_agreement":null},{"id":"W1964392332","doi":"10.1007/s10459-014-9504-x","title":"Information processing, specificity of practice, and the transfer of learning: considerations for reconsidering fidelity","year":2014,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":79,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Fidelity; Context (archaeology); Computer science; Construct (python library); Psychology","score_opus":0.04223478773428864,"score_gpt":0.41691386400923747,"score_spread":0.37467907627494884,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1964392332","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.88780254,0.00980369,0.041945223,0.049588397,0.00075596233,0.0024317764,0.000005790215,0.000029631869,0.0076369788],"genre_scores_gemma":[0.98217475,0.0015968253,0.015301734,0.00084282283,0.000032503795,0.000039437255,0.0000034257512,0.0000014165233,0.0000070595843],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99901986,0.000107789376,0.00055352144,0.00008853734,0.00013779117,0.00009252154],"domain_scores_gemma":[0.98338103,0.015860194,0.00032307717,0.00008176494,0.00030359742,0.000050304177],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.003093279,0.000042563388,0.00018350211,0.00006873098,0.00013242471,0.000013234881,0.000028234996,0.000022707427,0.0000053612835],"category_scores_gemma":[0.08100653,0.000029237997,0.00001979636,0.0001787062,0.00042583016,0.00046993917,0.0000046615523,0.00008350059,2.1366816e-7],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00042787066,0.00041819765,0.08290968,0.0015761871,0.0000057131033,2.2635348e-8,0.017516764,0.004727457,0.000020665288,0.10619195,0.00022768507,0.7859778],"study_design_scores_gemma":[0.015964042,0.006533596,0.3650884,0.017360654,0.00020492806,0.00015984988,0.08075518,0.09969362,0.0025991264,0.28447178,0.12650098,0.00066784053],"about_ca_topic_score_codex":0.00015764736,"about_ca_topic_score_gemma":0.0000625182,"teacher_disagreement_score":0.78531,"about_ca_system_score_codex":0.00001819644,"about_ca_system_score_gemma":0.00072511745,"threshold_uncertainty_score":0.92673457},"labels":[],"label_agreement":null},{"id":"W1964658130","doi":"10.1111/j.1440-1800.2007.00376.x","title":"SCYLLA OR CHARYBDIS: NAVIGATING BETWEEN EXCESSIVE EXAMINATION AND NAÏVE RELIANCE ON SELF‐ASSESSMENT","year":2007,"lang":"en","type":"editorial","venue":"Nursing Inquiry","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Licensure; Certification; Competence (human resources); Psychology; Medical education; Social psychology; Medicine; Law; Political science","score_opus":0.050685186122321216,"score_gpt":0.44416938019202623,"score_spread":0.39348419406970503,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1964658130","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.1760417,0.0008585337,0.000432131,0.00062688376,0.8171572,0.0008791921,0.00009332097,0.00037366635,0.0035373843],"genre_scores_gemma":[0.10575998,0.0007280434,0.004007945,0.00024351278,0.8870606,0.000042268253,0.0013520973,0.00013289248,0.0006726478],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.99605954,0.0001556838,0.00076646183,0.000907796,0.001577748,0.00053278345],"domain_scores_gemma":[0.96620595,0.031841088,0.0005800848,0.00049662165,0.0004831906,0.00039308492],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0016013602,0.00047218776,0.0009802976,0.00016320813,0.00021821956,0.00008474082,0.00015316864,0.0013744379,0.000025911091],"category_scores_gemma":[0.03511612,0.00039144748,0.000120401084,0.00031057018,0.0003815092,0.00008177073,0.00004322409,0.0025084908,0.000026627515],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00023013978,0.00042326935,0.0015190117,0.00026464125,0.00011009713,0.00020752079,0.002683564,0.0000013135794,0.000008420744,0.000016599955,0.6816029,0.31293255],"study_design_scores_gemma":[0.01319931,0.009930768,0.38116586,0.4486656,0.0034166507,0.000087400316,0.004257211,0.0003335573,0.00041126274,0.0004241973,0.13564602,0.0024621787],"about_ca_topic_score_codex":0.000037401704,"about_ca_topic_score_gemma":0.0000013775898,"teacher_disagreement_score":0.54595685,"about_ca_system_score_codex":0.0008294199,"about_ca_system_score_gemma":0.00077163195,"threshold_uncertainty_score":0.999922},"labels":[],"label_agreement":null},{"id":"W1964933620","doi":"10.3138/jvme.0113-008r","title":"Clinical Reasoning and Case-Based Decision Making: The Fundamental Challenge to Veterinary Educators","year":2013,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":59,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Medical education; Veterinary education; Clinical decision making; Veterinary medicine; Medicine; Psychology; Pedagogy; Curriculum; Family medicine","score_opus":0.09723622214676948,"score_gpt":0.4700354031166233,"score_spread":0.3727991809698538,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1964933620","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97467345,0.0012300465,0.00032932326,0.020674342,0.0024856604,0.0003396201,9.320934e-7,0.000014156299,0.0002524457],"genre_scores_gemma":[0.9814095,0.000285601,0.010137829,0.0065788566,0.0014924791,0.000029539075,0.0000028018055,0.000023125283,0.000040294985],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99748564,0.00025829702,0.0010472441,0.00025318228,0.0007062482,0.00024938263],"domain_scores_gemma":[0.9896316,0.008223309,0.0003800322,0.0003255869,0.00030347434,0.0011360152],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0020412796,0.00017435328,0.00043846515,0.00015843452,0.00013384056,0.000059219616,0.00018555507,0.00019732339,0.0010214932],"category_scores_gemma":[0.04910664,0.00011014907,0.0001981057,0.00017987638,0.00020590186,0.00012189288,0.000105066974,0.00068651815,0.000049604943],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009854583,0.0022541576,0.02653061,0.000052895542,0.000066780325,0.0017642971,0.0005430745,0.0000017629739,0.000037479545,0.000041351424,0.035771366,0.93195075],"study_design_scores_gemma":[0.004622524,0.030863306,0.72676015,0.020100998,0.0003692534,0.14308816,0.0051786923,0.0012260692,0.0000071088907,0.00057569775,0.06675004,0.00045797497],"about_ca_topic_score_codex":0.000040563395,"about_ca_topic_score_gemma":0.0000013546959,"teacher_disagreement_score":0.9314928,"about_ca_system_score_codex":0.000111771864,"about_ca_system_score_gemma":0.0015503499,"threshold_uncertainty_score":0.9998917},"labels":[],"label_agreement":null},{"id":"W1965296201","doi":"10.1371/journal.pone.0005881","title":"Categorizing Patients in a Forced-Choice Triad Task: The Integration of Context in Patient Management","year":2009,"lang":"en","type":"article","venue":"PLoS ONE","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Context (archaeology); Triad (sociology); Task (project management); Medicine; Two-alternative forced choice; MEDLINE; Psychology; Cognitive psychology","score_opus":0.04961863639232448,"score_gpt":0.301212090390176,"score_spread":0.2515934539978515,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1965296201","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9956782,0.00015042789,0.000022861426,0.0010609056,0.000032071846,0.00076238724,0.0000033412493,0.000010571256,0.002279239],"genre_scores_gemma":[0.99851483,0.000102985934,0.000323189,0.00086219487,0.000023282886,0.000035076024,0.00003911715,0.000005734435,0.000093559654],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99891055,0.000049525734,0.00044679083,0.00015174315,0.00029540254,0.00014599781],"domain_scores_gemma":[0.99783605,0.001731248,0.000115144794,0.00019853274,0.00007292278,0.000046100762],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00022048944,0.00008188809,0.00026983873,0.00009608202,0.000014224475,0.000006166955,0.0000587186,0.000048059876,0.000016775526],"category_scores_gemma":[0.009771982,0.000055707194,0.00004162154,0.00022627405,0.00002154425,0.000032987733,0.000018656228,0.00016446902,0.000009479657],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009904644,0.016743079,0.42072353,0.00009119458,0.0001281505,0.000041723597,0.0035923095,0.000028561637,0.0010445502,0.0019456311,0.00050039275,0.5541704],"study_design_scores_gemma":[0.0063965004,0.0015564511,0.9808336,0.0057104807,0.00015298121,3.5865204e-7,0.00070238655,0.0012764999,0.002605265,0.0006050914,0.000053634423,0.00010674478],"about_ca_topic_score_codex":0.00048118987,"about_ca_topic_score_gemma":0.00018838597,"teacher_disagreement_score":0.5601101,"about_ca_system_score_codex":0.000075954296,"about_ca_system_score_gemma":0.000016547381,"threshold_uncertainty_score":0.99856913},"labels":[],"label_agreement":null},{"id":"W1965806893","doi":"10.1016/j.seizure.2014.09.017","title":"The standardization debate: A conflation trap in critical care electroencephalography","year":2014,"lang":"en","type":"review","venue":"Seizure","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Manitoba; Health Sciences Centre","funders":"National Institute of Neurological Disorders and Stroke","keywords":"Conflation; Terminology; Standardization; Confusion; Critically ill; Electroencephalography; Set (abstract data type); Intensive care unit; Protocol (science); Computer science; Epistemology; Psychology; Medicine; Intensive care medicine; Psychiatry; Philosophy; Alternative medicine; Pathology; Linguistics","score_opus":0.04102723312221472,"score_gpt":0.4338374557048836,"score_spread":0.39281022258266884,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1965806893","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000012494503,0.9965899,0.00065446115,0.00035023922,0.00023831699,0.00066109793,0.000027856966,0.000049172402,0.001416472],"genre_scores_gemma":[0.0028411679,0.99593616,0.00009309193,0.00014822351,0.00032648537,0.00009217342,0.00044944094,0.00003576776,0.000077467936],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9981557,0.00025978938,0.0005846227,0.00035154342,0.00032364784,0.00032467942],"domain_scores_gemma":[0.98879516,0.010382271,0.00012388658,0.00035818043,0.000204033,0.0001364672],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00047586794,0.00024456586,0.0010442447,0.00012784553,0.00010587501,0.00004952367,0.0001125582,0.00041335344,0.000038685357],"category_scores_gemma":[0.021364067,0.0001465134,0.00040118507,0.0003452727,0.00015166253,0.000020728685,0.000016719883,0.0006260477,0.000040388404],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000025986034,0.000031157542,0.00026849305,0.0028306853,0.000039469825,0.000042259708,0.0000715405,8.174889e-7,3.7750418e-8,0.0020029237,0.0012672337,0.9934194],"study_design_scores_gemma":[0.00042146633,0.00025039786,0.00029394153,0.024612,0.0007615826,0.000071017224,0.000045721383,0.000014737633,3.1338786e-7,0.00032934768,0.97304165,0.00015781132],"about_ca_topic_score_codex":0.000010251262,"about_ca_topic_score_gemma":0.000041720185,"teacher_disagreement_score":0.9932616,"about_ca_system_score_codex":0.000101542115,"about_ca_system_score_gemma":0.00043662023,"threshold_uncertainty_score":0.9868794},"labels":[],"label_agreement":null},{"id":"W1965964514","doi":"10.2147/amep.s38372","title":"Problem-solving strategies in psychiatry: differences between experts and novices in diagnostic accuracy and reasoning","year":2013,"lang":"en","type":"article","venue":"Advances in Medical Education and Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"Royal College of Physicians and Surgeons of Canada; University of Calgary","keywords":"Medical diagnosis; Think aloud protocol; Psychiatric diagnosis; Medicine; Clinical judgment; Deductive reasoning; Psychology; Psychiatry; Medical education; Computer science; Medical physics; Artificial intelligence; Cognition; Pathology","score_opus":0.017988893659371646,"score_gpt":0.40783883521550346,"score_spread":0.38984994155613184,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1965964514","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93433744,0.03542468,0.000055311546,0.026117034,0.0002451077,0.00038333595,5.2606583e-7,0.000015158302,0.0034213883],"genre_scores_gemma":[0.944411,0.04596462,0.0072035477,0.0021219153,0.00015375105,0.00011164692,0.000008609972,0.00000788129,0.000017026105],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99850744,0.00021122501,0.00046496102,0.0003389829,0.0002597607,0.00021761563],"domain_scores_gemma":[0.91373885,0.08535645,0.00024204857,0.00016196592,0.00007924352,0.0004214631],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00091989007,0.00013510676,0.00030545384,0.00012961536,0.00004446988,0.00009612784,0.00006109732,0.00013069069,0.000099836725],"category_scores_gemma":[0.24974194,0.00010652386,0.000010208326,0.00024047839,0.00019240953,0.0011363052,0.000044792418,0.00040155914,0.0000020971684],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000027859836,0.00030176295,0.6681472,0.00008432873,0.0000036905078,0.000006432685,0.001901631,5.737203e-7,0.0000011575999,0.001095219,0.000084451465,0.3283457],"study_design_scores_gemma":[0.0012776526,0.00017035715,0.95726824,0.00686715,0.000028883329,0.00008334665,0.022097005,0.0002170807,8.717559e-7,0.009136894,0.0026979141,0.00015463113],"about_ca_topic_score_codex":0.0017135781,"about_ca_topic_score_gemma":0.0008397031,"teacher_disagreement_score":0.32819107,"about_ca_system_score_codex":0.000021572749,"about_ca_system_score_gemma":0.0005327574,"threshold_uncertainty_score":0.7565778},"labels":[],"label_agreement":null},{"id":"W1966803065","doi":"10.1503/cmaj.060752","title":"Physician experience, public health and the management of tuberculosis","year":2006,"lang":"en","type":"letter","venue":"Canadian Medical Association Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"University of Alberta","funders":"","keywords":"Tuberculosis; Proxy (statistics); Public health; Medicine; Family medicine; Active tuberculosis; Pathology; Mycobacterium tuberculosis; Computer science","score_opus":0.01613815744019857,"score_gpt":0.29012510967185906,"score_spread":0.2739869522316605,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1966803065","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0036812765,0.0012635855,0.000019076546,0.986218,0.0005819728,0.0003012574,0.000029988358,0.000009342187,0.007895532],"genre_scores_gemma":[0.005718868,0.0015857343,0.000102239064,0.98621786,0.004203752,0.000028394987,0.00011862215,0.000023725657,0.002000818],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99603695,0.0003045101,0.00076739775,0.00020535321,0.0020816193,0.0006041977],"domain_scores_gemma":[0.99495226,0.0025909678,0.0005785448,0.00018706414,0.00028587546,0.0014052903],"candidate_categories":["research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0023399116,0.00017258144,0.0006557231,0.0001647359,0.00021202114,0.00008865792,0.00021706155,0.0006691735,0.000530943],"category_scores_gemma":[0.006106346,0.00011451978,0.000231181,0.00025141216,0.00023318373,0.00003432099,0.000027673042,0.0024734822,0.000021159307],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000004009384,0.000022879978,0.003822597,0.000051648716,0.00032868003,0.0004790948,0.00021879935,6.105931e-8,5.73851e-9,0.00017175,0.94128954,0.053610947],"study_design_scores_gemma":[0.0022182001,0.00006269912,0.009898323,0.0017571094,0.00011556288,0.00014167449,0.0003718499,0.000023572717,1.008724e-7,0.0002145374,0.9850909,0.00010548008],"about_ca_topic_score_codex":0.0035469441,"about_ca_topic_score_gemma":0.00084714964,"teacher_disagreement_score":0.05350547,"about_ca_system_score_codex":0.0007502462,"about_ca_system_score_gemma":0.0030043782,"threshold_uncertainty_score":0.99982786},"labels":[],"label_agreement":null},{"id":"W1966986403","doi":"10.1051/pmed/2013040","title":"« Ils sont carrément incurables » : comment les métaphores des cliniciens enseignants révèlent leur malaise dans la gestion des difficultés de raisonnement clinique de leurs internes","year":2013,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Humanities; Philosophy","score_opus":0.14476721937165815,"score_gpt":0.3747158561215909,"score_spread":0.22994863674993277,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1966986403","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9474323,0.02206426,0.008174556,0.018466933,0.0011688485,0.0013307842,0.0001897476,0.00026916107,0.00090345583],"genre_scores_gemma":[0.9506547,0.03126372,0.0057798084,0.002937531,0.00094801676,0.0006032271,0.000095175616,0.00013122836,0.0075866194],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99200195,0.0016268953,0.0018785865,0.0010761684,0.000853569,0.0025628386],"domain_scores_gemma":[0.97635114,0.019690007,0.00047796,0.0008595378,0.00041791596,0.0022034433],"candidate_categories":["metaresearch","metaepi_narrow","sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0037975956,0.0008346742,0.0012062376,0.00017841757,0.0006025963,0.00022941908,0.0006021212,0.0008695753,0.0014575248],"category_scores_gemma":[0.03934441,0.00071248173,0.0006430834,0.00028186568,0.0030790835,0.00019571632,0.00049041794,0.0012421905,0.00037053652],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00024589003,0.0062214234,0.45280656,0.0005508599,0.00038896388,0.0005073282,0.024085285,0.0005971973,0.0017081837,0.00035681442,0.043336116,0.4691954],"study_design_scores_gemma":[0.007915734,0.0037762984,0.8302719,0.03533333,0.0015957237,0.00043552477,0.041842792,0.009329656,0.011384679,0.0048335367,0.051695943,0.0015849099],"about_ca_topic_score_codex":0.017339926,"about_ca_topic_score_gemma":0.0019236924,"teacher_disagreement_score":0.46761048,"about_ca_system_score_codex":0.0019701377,"about_ca_system_score_gemma":0.0009768355,"threshold_uncertainty_score":0.99963397},"labels":[],"label_agreement":null},{"id":"W1967006624","doi":"10.1002/chp.21166","title":"The Practicum Script Concordance Test: An Online Continuing Professional Development Format to Foster Reflection on Clinical Practice","year":2013,"lang":"en","type":"article","venue":"Journal of Continuing Education in the Health Professions","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":39,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; College of Family Physicians of Canada","funders":"","keywords":"Practicum; Concordance; Test (biology); Medical education; Interim; Competence (human resources); Professional development; TUTOR; Continuing professional development; Psychology; Medicine; Pedagogy; Social psychology","score_opus":0.10821826301238563,"score_gpt":0.5215111941216675,"score_spread":0.41329293110928184,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1967006624","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7511037,0.00041914106,0.00021410198,0.24006186,0.006289332,0.0015508208,0.0000013101455,0.000017351005,0.00034237432],"genre_scores_gemma":[0.9186683,0.00037012738,0.01887412,0.05508449,0.0017409959,0.00016799849,0.000015477142,0.000030367231,0.0050481055],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99484146,0.0014992333,0.0022750043,0.0002244756,0.00073415285,0.00042564954],"domain_scores_gemma":[0.94915324,0.04469695,0.0028296807,0.0005657449,0.0021092773,0.00064509147],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.008885505,0.00017632527,0.00045275502,0.00016328134,0.00045627324,0.00012923608,0.0002967016,0.00013973596,0.000017142273],"category_scores_gemma":[0.1451141,0.000090622605,0.0001007569,0.00034738667,0.0000693573,0.0005495188,0.000053663043,0.0017715032,0.00003880557],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00084163976,0.005664843,0.034478445,0.00004232766,0.000027152528,0.0000062529716,0.023637002,0.0000024219642,0.000058899077,0.00020097883,0.16653578,0.76850426],"study_design_scores_gemma":[0.002290836,0.004004781,0.53543264,0.020615024,0.00006840184,0.00045306902,0.11634883,0.000060935894,0.0000431462,0.0005019455,0.31998497,0.00019541403],"about_ca_topic_score_codex":0.00003292536,"about_ca_topic_score_gemma":0.000067453395,"teacher_disagreement_score":0.7683088,"about_ca_system_score_codex":0.00027244026,"about_ca_system_score_gemma":0.0030438502,"threshold_uncertainty_score":0.86208695},"labels":[],"label_agreement":null},{"id":"W1967291734","doi":"10.1097/00001888-200308000-00003","title":"The Importance of Cognitive Errors in Diagnosis and Strategies to Minimize Them","year":2003,"lang":"en","type":"review","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1598,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dartmouth General Hospital; Dalhousie University","funders":"Agency for Healthcare Research and Quality","keywords":"Debiasing; Cognitive bias; Heuristics; Cognition; Context (archaeology); Cognitive psychology; Process (computing); Pessimism; Confirmation bias; Metacognition; Perception; Computer science; Overconfidence effect; Psychology; Social psychology; Psychiatry","score_opus":0.13527516164826572,"score_gpt":0.4492700642212217,"score_spread":0.3139949025729559,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1967291734","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0023255658,0.9913857,0.000012483477,0.0016507707,0.00024265662,0.001320703,0.000020617459,0.000016826763,0.0030246591],"genre_scores_gemma":[0.0011319394,0.99628,0.000048543712,0.0012062365,0.00024970717,0.00038920811,0.00004171963,0.000041610554,0.0006109959],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.996927,0.00021458096,0.001529153,0.00050178124,0.00043956668,0.00038788377],"domain_scores_gemma":[0.9426552,0.05603864,0.00054028566,0.000331303,0.00009985912,0.0003347294],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0016745517,0.00040022045,0.0025122552,0.00014662251,0.000043137785,0.00000479018,0.00021266766,0.00057999924,0.00013398369],"category_scores_gemma":[0.10771862,0.00020996119,0.00014442767,0.0006106024,0.00060453295,0.000028151122,0.000074248666,0.0013685505,0.000021207323],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010505422,0.00005062806,0.026046842,0.0037122683,0.00027903207,0.0001247098,0.0016382708,9.743172e-8,1.5439508e-7,0.0016875708,0.012426271,0.9539291],"study_design_scores_gemma":[0.0022754655,0.0005252099,0.0028479085,0.4197487,0.002811632,0.0002144334,0.0049151247,0.0000010333052,0.0000015510539,0.0012052517,0.56509966,0.00035405986],"about_ca_topic_score_codex":0.000048513215,"about_ca_topic_score_gemma":0.000012646526,"teacher_disagreement_score":0.953575,"about_ca_system_score_codex":0.000054814365,"about_ca_system_score_gemma":0.00046159595,"threshold_uncertainty_score":0.89979744},"labels":[],"label_agreement":null},{"id":"W1967584589","doi":"10.1007/s10459-012-9388-6","title":"Cardiac examination and the effect of dual-processing instruction in a cardiopulmonary simulator","year":2012,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; University of British Columbia; University Health Network","funders":"Royal College of Physicians and Surgeons of Canada","keywords":"Dual (grammatical number); Computer science; Cognition; Medical physics; Medicine; Psychiatry","score_opus":0.012818976619416259,"score_gpt":0.394585856172095,"score_spread":0.38176687955267874,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1967584589","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9569565,0.040778857,0.00001879394,0.0004890514,0.00075112615,0.0004214986,4.7140466e-7,0.0000062547374,0.00057745015],"genre_scores_gemma":[0.99669206,0.0026532602,0.0003290951,0.00010312709,0.0001558927,0.00005106957,0.0000036649137,0.000002982373,0.000008842043],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99884707,0.00028797353,0.00033144848,0.00014885675,0.00020901224,0.0001756398],"domain_scores_gemma":[0.9959056,0.0037361386,0.00016983532,0.00008537744,0.000029669689,0.00007337515],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0039876276,0.00006240915,0.00025520512,0.0001328515,0.00007701243,0.000006810797,0.000031572563,0.000033176104,0.00000142077],"category_scores_gemma":[0.0072559994,0.00003982298,0.0000238823,0.00054009556,0.00033780656,0.0003440326,0.000012425562,0.000100055884,4.5880566e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000039778504,0.000028089858,0.61187094,0.00013153013,4.896992e-7,4.0003258e-8,0.0005982268,0.00017939837,0.0000020875075,0.00021499944,0.0000017943516,0.3869326],"study_design_scores_gemma":[0.0009319703,0.00028307067,0.99406594,0.0015150883,0.000010070151,0.000010111954,0.00096946856,0.0015350744,0.00007537424,0.0002931206,0.0002655056,0.000045181885],"about_ca_topic_score_codex":0.00038026206,"about_ca_topic_score_gemma":0.000012297573,"teacher_disagreement_score":0.38688743,"about_ca_system_score_codex":0.00008469845,"about_ca_system_score_gemma":0.00023218102,"threshold_uncertainty_score":0.8686633},"labels":[],"label_agreement":null},{"id":"W1967910943","doi":"10.1046/j.1365-2923.2000.00572.x","title":"Use of scheme‐based problem solving: an evaluation of the implementation and utilization of schemes in a clinical presentation curriculum","year":2000,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":51,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Calgary","funders":"","keywords":"Curriculum; Scheme (mathematics); Multivariate analysis of variance; Presentation (obstetrics); Mathematics education; Problem-based learning; Computer science; Medical education; Scale (ratio); Psychology; Medicine; Mathematics; Pedagogy; Machine learning","score_opus":0.10997420578870158,"score_gpt":0.5004821997414044,"score_spread":0.3905079939527028,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1967910943","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.997594,0.00013038734,0.00011398981,0.0012152363,0.00010187386,0.00073301716,0.0000031762422,0.000006384508,0.00010191035],"genre_scores_gemma":[0.9963481,0.000120549455,0.003001233,0.00025750042,0.000049976876,0.000039285493,0.00015602245,0.000006968769,0.000020356756],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99753577,0.00039145342,0.00091375405,0.00018912923,0.000879807,0.000090095484],"domain_scores_gemma":[0.99776804,0.0010581891,0.00033950974,0.00022503128,0.00051717315,0.000092024595],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002114993,0.000072362076,0.00022957349,0.00008247365,0.00001871113,0.000005358227,0.000060408216,0.00012464618,0.0006871209],"category_scores_gemma":[0.026704708,0.000053630898,0.00005767894,0.00034441578,0.0001820431,0.000126077,0.000012686388,0.00012530992,5.5931747e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005755059,0.0016463861,0.60147834,0.000070415146,0.000009776104,5.6712274e-8,0.00034246693,0.000023725708,0.00013321658,0.00014921461,0.00034382477,0.39574504],"study_design_scores_gemma":[0.0024892304,0.00030890486,0.93778574,0.0016077874,0.00017934361,0.0000020744062,0.00059391913,0.053907186,0.0025354938,0.0003358919,0.00019703532,0.00005741427],"about_ca_topic_score_codex":0.0008169234,"about_ca_topic_score_gemma":0.00015815411,"teacher_disagreement_score":0.39568764,"about_ca_system_score_codex":0.00003534367,"about_ca_system_score_gemma":0.001746029,"threshold_uncertainty_score":0.9814938},"labels":[],"label_agreement":null},{"id":"W1968036635","doi":"10.1353/pbm.2013.0009","title":"Embracing the Certainty of Uncertainty: Implications for Health Care and Research","year":2013,"lang":"en","type":"article","venue":"Perspectives in biology and medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":36,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Certainty; Surprise; Deliberation; Health care; Principal (computer security); Psychology; Epistemology; Management science; Engineering ethics; Computer science; Social psychology; Political science; Economics","score_opus":0.0816433763945975,"score_gpt":0.5148732872243384,"score_spread":0.43322991082974094,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1968036635","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7298412,0.026457844,0.00009987968,0.23894271,0.00006801434,0.0011055511,0.000013705182,0.000011578613,0.003459565],"genre_scores_gemma":[0.99566406,0.0024968935,0.00025413834,0.0012255537,0.0001657841,0.00012389844,0.00001789875,0.000004635904,0.00004712322],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"observational","domain_scores_codex":[0.99918413,0.00012524068,0.00021348568,0.00022732091,0.00004785168,0.00020199604],"domain_scores_gemma":[0.98789614,0.011493693,0.000051684423,0.00016918007,0.00028745443,0.000101836304],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00090968754,0.00006446879,0.0002906766,0.00009829566,0.00013160003,0.000002174541,0.00004574745,0.00009541474,0.000018397086],"category_scores_gemma":[0.023654826,0.000034925713,0.000020650974,0.00014771457,0.0012017642,0.000009670188,0.00003551881,0.0002829827,4.6225816e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006612084,0.00023655912,0.20117167,0.00028973178,0.0001291635,9.671778e-7,0.08020619,0.0000039467664,0.002244809,0.50731397,0.009732468,0.19800934],"study_design_scores_gemma":[0.0025876663,0.002729118,0.82652575,0.0008840692,0.000029657873,0.000025372588,0.07615268,0.00007416342,0.000020152313,0.08973141,0.0011780916,0.000061838124],"about_ca_topic_score_codex":0.0018538694,"about_ca_topic_score_gemma":0.00011837936,"teacher_disagreement_score":0.6253541,"about_ca_system_score_codex":0.00005229943,"about_ca_system_score_gemma":0.00010042295,"threshold_uncertainty_score":0.9845694},"labels":[],"label_agreement":null},{"id":"W1968386328","doi":"10.1007/s10459-013-9453-9","title":"Exploring the role of first impressions in rater-based assessments","year":2013,"lang":"en","type":"review","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":110,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Psychology; Cognition; Competence (human resources); Applied psychology; Inter-rater reliability; Social cognition; Cognitive skill; Cognitive psychology; Social psychology; Rating scale; Developmental psychology","score_opus":0.1735800993244859,"score_gpt":0.5319678283559396,"score_spread":0.35838772903145366,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1968386328","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0017267488,0.99485314,0.000009363985,0.0005150292,0.0010929095,0.0011439864,0.0000033455567,0.000008200048,0.0006472693],"genre_scores_gemma":[0.0047252597,0.9923716,0.0015109773,0.00013956495,0.000102492995,0.0010866069,0.000025247855,0.0000088876905,0.000029352526],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99791455,0.00019622482,0.00094606815,0.00034695878,0.00030821472,0.00028795894],"domain_scores_gemma":[0.99021375,0.008685256,0.00060825713,0.0003116236,0.000055734443,0.00012539336],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001321419,0.00015270928,0.0007416943,0.00028348886,0.000111197594,0.000017117503,0.0002639108,0.000055969795,0.000029997018],"category_scores_gemma":[0.0062237843,0.00008847002,0.00008506685,0.0010251488,0.00020229306,0.00025323633,0.00003155007,0.0003063615,0.000008963136],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000017482957,0.00024140945,0.010546921,0.002605385,0.000001138088,1.3699439e-7,0.00007309398,0.00004732333,2.4129022e-8,0.000116771786,0.000023356324,0.98634267],"study_design_scores_gemma":[0.00019072475,0.00030396646,0.0039763935,0.18859924,0.000029946243,0.000003477294,0.00075619604,0.00032227786,0.0000020057266,0.00054834335,0.8051374,0.00013005343],"about_ca_topic_score_codex":0.0012442687,"about_ca_topic_score_gemma":0.00021579,"teacher_disagreement_score":0.9862126,"about_ca_system_score_codex":0.00021203673,"about_ca_system_score_gemma":0.0034922296,"threshold_uncertainty_score":0.74509007},"labels":[],"label_agreement":null},{"id":"W1968642299","doi":"10.1111/j.1365-2923.2007.02914.x","title":"Non‐analytical models of clinical reasoning: the role of experience","year":2007,"lang":"en","type":"review","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":402,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Introspection; Process (computing); Centrality; Similarity (geometry); Component (thermodynamics); Psychology; Interpretation (philosophy); Medical diagnosis; Computer science; Artificial intelligence; Cognitive psychology; Cognitive science; Medicine","score_opus":0.1103726393677781,"score_gpt":0.5216282285392397,"score_spread":0.41125558917146166,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1968642299","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0007926838,0.98540694,0.00023299026,0.0003898491,0.0008516373,0.00059816655,0.0000044916046,0.000017054725,0.011706221],"genre_scores_gemma":[0.0042047394,0.99300057,0.00052256306,0.0005998018,0.000993447,0.00007025084,0.000092285205,0.000032416952,0.0004839268],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9952396,0.0001941342,0.0025713914,0.00043958268,0.001281112,0.0002741728],"domain_scores_gemma":[0.9790822,0.017831495,0.0010594452,0.0010049499,0.00040387153,0.00061798765],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0032551903,0.00025703644,0.0023809946,0.00012381874,0.00003167308,0.000006226861,0.0004420323,0.0008661073,0.00048735854],"category_scores_gemma":[0.1656782,0.00014593931,0.0009279367,0.00049951597,0.0008136934,0.000029088182,0.00010049755,0.0010692114,0.00003116685],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00002669373,0.0012674357,0.0006444173,0.0010455486,0.00010135912,0.0000036921178,0.00017794051,1.898651e-7,1.120516e-8,0.0021123008,0.0028542883,0.9917661],"study_design_scores_gemma":[0.0004004639,0.00025836576,0.0004536358,0.08371438,0.0018241142,0.000096280004,0.0003352013,0.0018350438,0.0000019417553,0.00041219083,0.91049004,0.0001783246],"about_ca_topic_score_codex":0.00015319625,"about_ca_topic_score_gemma":0.0000024697338,"teacher_disagreement_score":0.9915878,"about_ca_system_score_codex":0.00004829453,"about_ca_system_score_gemma":0.007876849,"threshold_uncertainty_score":0.99774754},"labels":[],"label_agreement":null},{"id":"W1968991973","doi":"10.1007/s10459-012-9396-6","title":"Mental workload as a tool for understanding dual processes in rater-based assessments","year":2012,"lang":"en","type":"letter","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Workload; Inter-rater reliability; Psychology; Mental health; Medical education; Dual (grammatical number); Applied psychology; Computer science; Medicine; Rating scale; Psychiatry; Developmental psychology","score_opus":0.09136439521427846,"score_gpt":0.492049572093889,"score_spread":0.4006851768796105,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1968991973","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0420494,0.027838174,0.0019423652,0.91225094,0.008798407,0.0050372756,0.000045404522,0.00007432453,0.0019637097],"genre_scores_gemma":[0.35964328,0.0052191718,0.013970248,0.61495674,0.003213246,0.0015025551,0.0007388798,0.000047164907,0.0007087236],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.997339,0.000108724,0.0007403156,0.0005764375,0.0005472388,0.0006882716],"domain_scores_gemma":[0.9917129,0.007441237,0.00045467698,0.00018368257,0.000079944795,0.00012754885],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0015327226,0.00022219466,0.000456516,0.00040512707,0.0001717308,0.00005847693,0.00015236021,0.00022587329,0.00004939508],"category_scores_gemma":[0.008525139,0.00019183141,0.000051207207,0.00086058065,0.000237085,0.00039551,0.000017947812,0.00055846083,0.0000094762745],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00048220923,0.0029873056,0.6189778,0.0108251525,0.00002046522,0.00002461458,0.0020624236,0.00030406454,0.0000033306806,0.00079772325,0.23657608,0.12693885],"study_design_scores_gemma":[0.007531843,0.0063229203,0.022721913,0.12998295,0.000119552984,0.000097223616,0.0062590223,0.0011585477,0.000079038786,0.034032553,0.7899266,0.0017678573],"about_ca_topic_score_codex":0.00025896842,"about_ca_topic_score_gemma":0.00015526125,"teacher_disagreement_score":0.59625584,"about_ca_system_score_codex":0.0012349931,"about_ca_system_score_gemma":0.007767952,"threshold_uncertainty_score":0.9998265},"labels":[],"label_agreement":null},{"id":"W1969193293","doi":"10.1051/pmed/2011111","title":"Aborder le raisonnement clinique d’un point de vue pédagogique","year":2011,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Philosophy; Humanities","score_opus":0.05909274047187839,"score_gpt":0.3502573294590347,"score_spread":0.2911645889871563,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1969193293","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.38613793,0.11868466,0.055279948,0.09516865,0.011951355,0.0030657998,0.0002977454,0.0012889007,0.32812503],"genre_scores_gemma":[0.77986944,0.0317353,0.044885002,0.024616504,0.0053736074,0.00041729005,0.00013821891,0.00033337838,0.11263128],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99411684,0.0007038507,0.0015345331,0.0010907416,0.0006492214,0.0019048305],"domain_scores_gemma":[0.9865167,0.009628432,0.0003672272,0.0012804815,0.00030212972,0.0019050117],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0032295443,0.0006604335,0.0011480117,0.00015995765,0.00027990987,0.00006338092,0.00057545624,0.0011184708,0.025952108],"category_scores_gemma":[0.046803817,0.0006426174,0.0006445099,0.00043744053,0.001003984,0.00015853145,0.00034440638,0.0015894513,0.002502032],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010617229,0.016876364,0.059875686,0.00053668674,0.0010222669,0.0039022423,0.027803784,0.000022885952,0.000526052,0.022779226,0.18259218,0.6830009],"study_design_scores_gemma":[0.019201176,0.0058067157,0.14266446,0.015552212,0.0021268404,0.0009183671,0.016008385,0.0021148503,0.025093911,0.05082505,0.7167397,0.0029483298],"about_ca_topic_score_codex":0.0020274841,"about_ca_topic_score_gemma":0.00011769892,"teacher_disagreement_score":0.6800526,"about_ca_system_score_codex":0.0002964605,"about_ca_system_score_gemma":0.0021447833,"threshold_uncertainty_score":0.9996025},"labels":[],"label_agreement":null},{"id":"W1969750362","doi":"10.1017/s031716710000706x","title":"The Script Concordance Test: A New Tool Assessing Clinical Judgement in Neurology","year":2009,"lang":"en","type":"article","venue":"Canadian Journal of Neurological Sciences / Journal Canadien des Sciences Neurologiques","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":72,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"McGill University; Université de Montréal; McGill University Health Centre","funders":"Dalhousie University; McGill University; American Academy of Neurology; Mayo Clinic","keywords":"Neurology; Concordance; Test (biology); Cronbach's alpha; Judgement; Psychology; Construct validity; Construct (python library); Clinical neurology; Medicine; Medical education; Clinical psychology; Psychiatry; Psychometrics; Internal medicine","score_opus":0.06986137776209353,"score_gpt":0.3734645410053637,"score_spread":0.3036031632432702,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1969750362","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93769145,0.0017638991,0.000034931487,0.0564681,0.0015162413,0.00020848982,0.0000019519914,0.00001662281,0.0022982953],"genre_scores_gemma":[0.96783864,0.001069177,0.0014583749,0.028866874,0.0006929621,0.0000013445458,1.7728506e-7,0.000008587402,0.00006384083],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99389136,0.0011208133,0.0018784322,0.0006314159,0.00086480717,0.0016131486],"domain_scores_gemma":[0.98141444,0.014118604,0.0010413198,0.00026232802,0.00043690516,0.002726391],"candidate_categories":["metaresearch","sts","research_integrity"],"consensus_categories":["sts"],"category_scores_codex":[0.01035514,0.00032630516,0.0007787459,0.0005380973,0.0019633018,0.00090909464,0.0016638329,0.00027144596,0.00007432267],"category_scores_gemma":[0.097536094,0.0001868545,0.00036596926,0.0014294728,0.0062158597,0.0006087356,0.000056106706,0.0025369895,0.0000059332333],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015536556,0.00008649793,0.93705904,0.0000012291227,0.0000068074482,0.008419204,0.00007378341,0.00036378615,0.000008703463,0.00070475,0.003852912,0.04926794],"study_design_scores_gemma":[0.00086928817,0.085513346,0.86832744,0.00014927473,0.000039859704,0.017618872,0.00007508865,0.0009929336,0.000008775886,0.018571554,0.007634555,0.0001990445],"about_ca_topic_score_codex":0.0009402106,"about_ca_topic_score_gemma":0.027230317,"teacher_disagreement_score":0.08718096,"about_ca_system_score_codex":0.00016304827,"about_ca_system_score_gemma":0.008406774,"threshold_uncertainty_score":0.9997642},"labels":[],"label_agreement":null},{"id":"W1969927924","doi":"10.1017/s0317167100006946","title":"The Script Concordance Test","year":2009,"lang":"fr","type":"letter","venue":"Canadian Journal of Neurological Sciences / Journal Canadien des Sciences Neurologiques","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Calgary Laboratory Services","funders":"","keywords":"Concordance; Test (biology); Psychology; Computer science; Natural language processing; Medicine; Biology; Internal medicine","score_opus":0.04102365242630865,"score_gpt":0.3068804626502669,"score_spread":0.26585681022395824,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1969927924","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.578118,0.027074575,0.000015726233,0.3754443,0.0101055475,0.00037637696,0.000054394655,0.000025167667,0.008785856],"genre_scores_gemma":[0.67403674,0.012272443,0.0012785326,0.3004776,0.010500267,0.000006243122,0.0000021984229,0.00004772343,0.001378232],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.9857509,0.0027917996,0.0030161454,0.0012627195,0.0024322243,0.004746215],"domain_scores_gemma":[0.93692887,0.0514697,0.0031565747,0.0005473651,0.0017087816,0.0061887177],"candidate_categories":["metaresearch","metaepi_narrow","sts","scholarly_communication","open_science","research_integrity"],"consensus_categories":["sts"],"category_scores_codex":[0.014557309,0.0010544924,0.0017660211,0.0010359656,0.009896739,0.0029545564,0.006112003,0.0012082492,0.0005136796],"category_scores_gemma":[0.20646593,0.0005989143,0.0010888065,0.0024386914,0.038718034,0.0008718456,0.00014686424,0.0098906215,0.00006683957],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015006529,0.00010975736,0.6064015,0.000021102745,0.00006546428,0.079584055,0.00022846021,0.0009096552,0.000008296157,0.0011408401,0.27006003,0.041320767],"study_design_scores_gemma":[0.00089446775,0.21166226,0.18159696,0.0020418148,0.00042934855,0.12911047,0.00027085922,0.001105549,0.000023184066,0.02618259,0.44576356,0.0009189179],"about_ca_topic_score_codex":0.0029337476,"about_ca_topic_score_gemma":0.06690492,"teacher_disagreement_score":0.42480454,"about_ca_system_score_codex":0.0009636887,"about_ca_system_score_gemma":0.020454735,"threshold_uncertainty_score":0.99964625},"labels":[],"label_agreement":null},{"id":"W1970175752","doi":"10.1007/s10459-010-9225-8","title":"Using signal detection theory to model changes in serial learning of radiological image interpretation","year":2010,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":40,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University; SickKids Foundation; Hospital for Sick Children; University of Toronto","funders":"","keywords":"Radiological weapon; Abnormality; Radiography; Medicine; Ankle; Medical physics; Radiology; Psychology; Physical therapy; Surgery; Psychiatry","score_opus":0.03393052233658283,"score_gpt":0.4453861459298375,"score_spread":0.41145562359325466,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1970175752","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9658708,0.00029151654,0.031958506,0.0006319075,0.00069467427,0.00028514353,5.20272e-7,0.000011778507,0.00025517188],"genre_scores_gemma":[0.96401244,0.00018757097,0.035256922,0.0003741129,0.00012564856,0.000026455242,0.0000028334737,0.000003294445,0.000010729482],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99899995,0.00012201396,0.0003211448,0.00022872703,0.00014947173,0.00017870172],"domain_scores_gemma":[0.99791545,0.0017033301,0.00015902534,0.00006795869,0.000060998842,0.00009326333],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001979045,0.000061141196,0.00017672458,0.00022275996,0.00006243771,0.000008603321,0.00006525413,0.000052155105,0.000020272759],"category_scores_gemma":[0.012829976,0.00005119294,0.000017186723,0.00045714065,0.00016009649,0.00016911163,0.000014419944,0.00023458012,0.0000013021908],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00046047606,0.00033782973,0.07642748,0.00011628479,0.0000011284917,5.182889e-7,0.003705076,0.09305703,0.035824817,0.0020541716,0.0000030637154,0.78801215],"study_design_scores_gemma":[0.001026006,0.0028047522,0.09523548,0.0035569954,0.000012316856,0.000036277008,0.004714663,0.857519,0.008600285,0.026092531,0.00013947835,0.00026224568],"about_ca_topic_score_codex":0.0002031768,"about_ca_topic_score_gemma":0.0005640779,"teacher_disagreement_score":0.7877499,"about_ca_system_score_codex":0.000080718295,"about_ca_system_score_gemma":0.0005108204,"threshold_uncertainty_score":0.99548537},"labels":[],"label_agreement":null},{"id":"W1970818269","doi":"10.1186/1471-2296-15-92","title":"Identification of factors associated with diagnostic error in primary care","year":2014,"lang":"en","type":"article","venue":"BMC Family Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institut National de Santé Publique du Québec; Université de Montréal","funders":"European Regional Development Fund","keywords":"Medicine; Medical diagnosis; Representativeness heuristic; Medical record; Audit; MEDLINE; Family medicine; Surgery; Pathology","score_opus":0.047999351886975496,"score_gpt":0.3352600772094862,"score_spread":0.2872607253225107,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1970818269","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99248797,0.00032977294,0.00071882695,0.0001590625,0.0001407214,0.0003512737,0.000012868255,0.000046023433,0.005753471],"genre_scores_gemma":[0.9985065,0.000041927626,0.0005080475,0.00052772433,0.000036975565,0.000022178834,0.00017615272,0.000021991611,0.00015851315],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9984321,0.00025534222,0.00046856553,0.00026563098,0.00039783836,0.00018054839],"domain_scores_gemma":[0.92230874,0.076452516,0.00043251624,0.0003512465,0.00034737712,0.0001076136],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00078199385,0.00012648261,0.000345026,0.00008743734,0.000028689847,0.00001505548,0.00008005617,0.0001221118,0.0000066872685],"category_scores_gemma":[0.36820135,0.0001000965,0.000058466794,0.0003083177,0.000073352196,0.00015425708,0.000024411958,0.0002373061,0.000016979164],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003671202,0.0008906063,0.99437267,0.0001378064,0.000055490647,0.00001549808,0.0007274971,0.00014810915,0.001027739,0.00013168351,0.00027841685,0.0018473468],"study_design_scores_gemma":[0.0021994119,0.00038386695,0.9930345,0.0014921355,0.00032719498,0.0000030823992,0.0016790866,0.00014837913,0.00038023875,0.000016625438,0.00022207398,0.00011342832],"about_ca_topic_score_codex":0.00017103608,"about_ca_topic_score_gemma":0.00006518803,"teacher_disagreement_score":0.36741936,"about_ca_system_score_codex":0.00011995458,"about_ca_system_score_gemma":0.00024462526,"threshold_uncertainty_score":0.63712054},"labels":[],"label_agreement":null},{"id":"W1971457640","doi":"10.1111/medu.12509","title":"Key‐feature questions for assessment of clinical reasoning: a literature review","year":2014,"lang":"en","type":"review","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":125,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; University of Waterloo","funders":"","keywords":"Key (lock); Feature (linguistics); MEDLINE; Psychology; Computer science; Management science; Medical education; Data science; Medicine; Political science; Linguistics; Engineering","score_opus":0.06395609745718979,"score_gpt":0.5513012656022689,"score_spread":0.4873451681450791,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1971457640","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[9.005665e-7,0.98370093,0.0006286526,0.00914032,0.0023136023,0.0022319534,0.00004933313,0.00005348804,0.0018808454],"genre_scores_gemma":[0.0000015263261,0.97112775,0.011932293,0.0067591765,0.0029338503,0.00087608007,0.0037747393,0.000059358208,0.0025352407],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9953542,0.00059512624,0.0022324817,0.0006731156,0.0008613996,0.00028370396],"domain_scores_gemma":[0.97070724,0.023980606,0.0018263242,0.0013416049,0.0011094989,0.0010347345],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.004583593,0.00039735466,0.0041383025,0.00014438668,0.00005157013,0.000020746782,0.00029715462,0.0014385049,0.00028208655],"category_scores_gemma":[0.42858228,0.00026595435,0.0016774312,0.0005494004,0.0001768226,0.000029664112,0.000054835957,0.0016966505,0.000022390494],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000048666047,0.00076174317,0.00006344862,0.0641136,0.0000975025,0.000003231574,0.000006206666,4.6248507e-9,1.3254424e-9,0.001217462,0.28397125,0.64976066],"study_design_scores_gemma":[0.00027883184,0.00021108481,0.00014425174,0.45898277,0.0022707218,0.00010650315,0.0000013457781,0.000019169527,5.854055e-9,0.00003399381,0.53783655,0.00011480816],"about_ca_topic_score_codex":0.0000068042596,"about_ca_topic_score_gemma":0.0000019282759,"teacher_disagreement_score":0.64964586,"about_ca_system_score_codex":0.000100786674,"about_ca_system_score_gemma":0.013506775,"threshold_uncertainty_score":0.99997926},"labels":[],"label_agreement":null},{"id":"W1971803399","doi":"10.1176/ajp.2006.163.5.913","title":"The Persistence of Mind-Brain Dualism in Psychiatric Reasoning About Clinical Scenarios","year":2006,"lang":"en","type":"article","venue":"American Journal of Psychiatry","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":233,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McGill University","funders":"Canadian Institutes of Health Research","keywords":"Psychology; Biopsychosocial model; Blame; Attribution; Cognition; Social cognition; Clinical psychology; Theory of mind; Psychiatry; Social psychology","score_opus":0.015459644911953984,"score_gpt":0.3398524634842877,"score_spread":0.3243928185723337,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1971803399","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9665609,0.004236838,0.00011645047,0.025767934,0.0016417103,0.00012337783,0.0000035548926,0.0000068301383,0.0015423913],"genre_scores_gemma":[0.97534454,0.0005503936,0.021310568,0.0012191095,0.0013539137,0.0000018741163,0.0000020578511,0.000023394172,0.00019415574],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9965702,0.00030251962,0.0019841543,0.00023760289,0.0005466137,0.00035887185],"domain_scores_gemma":[0.9901759,0.007216291,0.0017227054,0.00038987404,0.0002507408,0.00024446903],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0025809037,0.0001785285,0.0008049903,0.00019774235,0.00008558738,0.000021487253,0.0003000244,0.000097447955,0.00002734608],"category_scores_gemma":[0.011482708,0.00011792141,0.00060687604,0.0007703208,0.00074110326,0.000052027273,0.000030269006,0.0008177768,0.000008935295],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006793495,0.00075257965,0.94043773,0.000011111805,0.00007922206,0.00003808275,0.00011229944,0.000043881744,0.00000949005,0.0009339863,0.013113835,0.043788455],"study_design_scores_gemma":[0.0031528084,0.0028257973,0.98462486,0.0019639784,0.00022979625,0.00050184614,0.0021764436,0.00011859932,0.0000058820933,0.0013419259,0.002895361,0.00016267726],"about_ca_topic_score_codex":0.00027193705,"about_ca_topic_score_gemma":0.00018772348,"teacher_disagreement_score":0.044187177,"about_ca_system_score_codex":0.000043182714,"about_ca_system_score_gemma":0.00085035193,"threshold_uncertainty_score":0.996844},"labels":[],"label_agreement":null},{"id":"W1972679656","doi":"10.1007/s10459-009-9175-1","title":"Two models of raters in a structured oral examination: does it make a difference?","year":2009,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; University of Calgary; Medical Council of Canada; Ottawa Hospital; University of Ottawa","funders":"","keywords":"Reliability (semiconductor); Oral examination; Inter-rater reliability; Internal consistency; Objective structured clinical examination; Physical examination; Psychology; Medicine; Clinical psychology; Psychometrics; Developmental psychology; Rating scale; Family medicine; Oral health; Psychiatry","score_opus":0.03388845178213149,"score_gpt":0.43300130331885756,"score_spread":0.39911285153672604,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1972679656","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9868634,0.0020674558,0.00023481819,0.008058953,0.0005733922,0.0003841482,0.0000019571116,0.000011352397,0.00180456],"genre_scores_gemma":[0.9887286,0.0014997453,0.0073790974,0.0022148278,0.000053177875,0.000021243108,0.000009919067,0.0000023375806,0.000091055845],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99857795,0.000096729804,0.0005251193,0.00028236432,0.00028485683,0.0002330025],"domain_scores_gemma":[0.9983866,0.0010917203,0.00020398041,0.00013964469,0.000069378904,0.00010862534],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009457832,0.00008509978,0.00026309842,0.00022963743,0.00004540508,0.000011338739,0.000111742025,0.000034534798,0.000017104137],"category_scores_gemma":[0.0030348797,0.00005913627,0.000021976006,0.0007238015,0.0001727058,0.00019370829,0.0000082382185,0.00011293537,6.640741e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005125368,0.0004977338,0.1476388,0.000090820366,8.554902e-7,0.0000011531789,0.002791071,0.0030978143,0.00003471699,0.00946745,0.000024582156,0.8363038],"study_design_scores_gemma":[0.0011500185,0.0008192843,0.89069,0.003163278,0.0000047849867,0.000008240372,0.003088069,0.00889078,0.00014125118,0.09178349,0.00013704266,0.00012375263],"about_ca_topic_score_codex":0.0005378243,"about_ca_topic_score_gemma":0.0013486306,"teacher_disagreement_score":0.83618,"about_ca_system_score_codex":0.00011514591,"about_ca_system_score_gemma":0.0009128631,"threshold_uncertainty_score":0.3633254},"labels":[],"label_agreement":null},{"id":"W1973603798","doi":"10.3138/jvme.35.4.622","title":"Attitudes of Veterinary Faculty to the Assessment of Clinical Reasoning Using Extended Matching Questions","year":2008,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"KU Leuven","keywords":"Medical education; Matching (statistics); Promotion (chess); Psychology; Identification (biology); Test (biology); Scientific reasoning; Variety (cybernetics); Mathematics education; Computer science; Medicine; Artificial intelligence; Pathology","score_opus":0.24347982579857388,"score_gpt":0.559914772855029,"score_spread":0.31643494705645514,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1973603798","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9916409,0.00027982736,0.0031842147,0.0031679445,0.001357609,0.00017049549,0.0000045576166,0.000008219258,0.00018621016],"genre_scores_gemma":[0.9536083,0.00038362812,0.0445658,0.0005634082,0.0008073972,0.0000049301625,0.000008730468,0.000014314351,0.000043530257],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9966505,0.0003377714,0.0016753991,0.00016956865,0.0009960548,0.00017068676],"domain_scores_gemma":[0.994368,0.0032333175,0.0009627125,0.00030468902,0.00058420876,0.00054704084],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0027430023,0.0001311794,0.0006714086,0.0001345475,0.00009790455,0.0000067702094,0.00023642654,0.00015163393,0.00015482282],"category_scores_gemma":[0.041820552,0.00008662826,0.00031086864,0.00023350718,0.00022760687,0.00009793095,0.00009552664,0.0006224601,0.0000021777234],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.003129039,0.022569835,0.6973084,0.0006582271,0.00093975104,0.001352936,0.0037368208,0.00082443195,0.0096291425,0.0008047712,0.011231358,0.24781528],"study_design_scores_gemma":[0.0010293055,0.004620034,0.9775214,0.007302309,0.00016490412,0.0067278533,0.00033972267,0.00059247913,0.000031014315,0.000047745234,0.0015369102,0.00008632579],"about_ca_topic_score_codex":0.000104479266,"about_ca_topic_score_gemma":0.000001351135,"teacher_disagreement_score":0.280213,"about_ca_system_score_codex":0.00008881232,"about_ca_system_score_gemma":0.003533441,"threshold_uncertainty_score":0.9662506},"labels":[],"label_agreement":null},{"id":"W1973816053","doi":"10.1007/s11251-013-9290-5","title":"Do expert clinical teachers have a shared understanding of what constitutes a competent reasoning performance in case-based teaching?","year":2013,"lang":"en","type":"article","venue":"Instructional Science","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":17,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Alberta Advanced Education; University of Alberta","funders":"","keywords":"Educational psychology; Set (abstract data type); Process (computing); Key (lock); Psychology; Mathematics education; Logical reasoning; Computer science","score_opus":0.07755019330662685,"score_gpt":0.3782165528307532,"score_spread":0.30066635952412635,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1973816053","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9958323,0.00013203947,0.00043254028,0.00088804495,0.00054936024,0.00026447728,0.0000024776107,0.00003103499,0.001867686],"genre_scores_gemma":[0.980207,0.00005541075,0.019234741,0.00034800178,0.00010384984,0.000022229311,0.0000047468798,0.000006769343,0.000017197666],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9979985,0.000064328895,0.0005993117,0.00042884282,0.000595016,0.00031403603],"domain_scores_gemma":[0.99641645,0.0027669815,0.00017616266,0.000241368,0.00011747817,0.0002815498],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0015190837,0.00013026522,0.00032081897,0.00020802804,0.00022783558,0.00012676885,0.00015922026,0.00008908711,0.00013752433],"category_scores_gemma":[0.014308409,0.00010656239,0.000099519224,0.00032445553,0.0018383933,0.0006718028,0.000059910213,0.000427863,0.000011760831],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000649238,0.0002750961,0.9511295,0.00001351896,0.00001407991,0.00011081052,0.0005033736,0.00018346436,0.0003682502,0.0032411288,0.00015928932,0.04393655],"study_design_scores_gemma":[0.004988174,0.00071132043,0.9109801,0.0099304505,0.000036043697,0.0029409002,0.011644892,0.057057105,0.00040479485,0.0005244044,0.00039290832,0.00038889618],"about_ca_topic_score_codex":0.00055329554,"about_ca_topic_score_gemma":0.000029857214,"teacher_disagreement_score":0.056873642,"about_ca_system_score_codex":0.00033909746,"about_ca_system_score_gemma":0.0007113036,"threshold_uncertainty_score":0.9939945},"labels":[],"label_agreement":null},{"id":"W1974039834","doi":"10.1097/acm.0000000000000105","title":"The Etiology of Diagnostic Errors","year":2013,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":173,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University; ECW Press (Canada); University of Ottawa; McGill University; Ethica (Canada); Medical Council of Canada","funders":"McMaster University; McGill University","keywords":"Cohort; Diagnostic accuracy; Medicine; Licensure; Cohort study; Etiology; Psychology; Psychiatry; Medical education; Pathology; Internal medicine","score_opus":0.03236790360632494,"score_gpt":0.36296652753934233,"score_spread":0.3305986239330174,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1974039834","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9180531,0.004953838,0.00014265727,0.06863244,0.00065731525,0.00052085606,0.0000011838366,0.00005067636,0.0069879517],"genre_scores_gemma":[0.9853139,0.0066891923,0.000052106447,0.0036156098,0.00056320336,0.00006284358,0.000013466159,0.0000141241035,0.00367555],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99849063,0.00007860067,0.0006373734,0.00019494895,0.0003014615,0.00029698003],"domain_scores_gemma":[0.92866045,0.070398465,0.00018236016,0.0003498205,0.0001580455,0.00025084312],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00077886425,0.0001230503,0.00041855464,0.000050339975,0.00006065199,0.0000013407307,0.00018587337,0.00022903287,0.0008007338],"category_scores_gemma":[0.29701284,0.0000617571,0.000056838682,0.00018732522,0.0008373265,0.000023738425,0.00004869935,0.0006350193,0.00028022964],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012473117,0.00005220136,0.28152922,0.00006512164,0.00016759906,0.000030206096,0.0014123785,0.0000048682837,0.0041278093,0.012906329,0.61861116,0.08096835],"study_design_scores_gemma":[0.00729372,0.00287839,0.8778686,0.007193528,0.0007037826,0.00037074237,0.0033548912,0.0004277352,0.0014731652,0.028333802,0.06976167,0.0003399481],"about_ca_topic_score_codex":0.00027419886,"about_ca_topic_score_gemma":0.0000022445759,"teacher_disagreement_score":0.5963394,"about_ca_system_score_codex":0.000018572706,"about_ca_system_score_gemma":0.00007176328,"threshold_uncertainty_score":0.87674737},"labels":[],"label_agreement":null},{"id":"W1974061733","doi":"10.5430/jha.v3n2p10","title":"Implementing combined decision models in healthcare settings: the Simon and Pauker- Kassirer models","year":2013,"lang":"en","type":"article","venue":"Journal of Hospital Administration","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Computer science; Process (computing); Management science; Decision model; Medical decision making; Influence diagram; Decision-making; Bridge (graph theory); Decision engineering; Decision analysis; Decision tree; Artificial intelligence; Decision support system; Data science; Business decision mapping; Machine learning; Engineering; Medicine; Mathematics; Operations management","score_opus":0.02018405473661035,"score_gpt":0.32566818766492506,"score_spread":0.3054841329283147,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1974061733","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9699559,0.0003272665,0.0008860755,0.028089844,0.00015817131,0.0003431365,0.0000024830008,0.000007520432,0.0002296037],"genre_scores_gemma":[0.9975837,0.00017538293,0.0014199611,0.0006288965,0.00013629763,0.000008696633,0.0000077758,0.000011019251,0.000028282018],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99842584,0.000049075756,0.00078999443,0.00013688557,0.00037056083,0.00022765245],"domain_scores_gemma":[0.99708754,0.0018885966,0.00041398098,0.00013096242,0.00029886665,0.00018004839],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.000843253,0.000111414374,0.0002767242,0.000064078595,0.00008295916,0.00008252719,0.00007176886,0.000084371255,0.000017281282],"category_scores_gemma":[0.0031800515,0.0000695904,0.000084689986,0.00010920729,0.000041581472,0.0003781364,0.000036606223,0.00032018023,0.0000027714098],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016801659,0.0049670828,0.63144565,0.00025583064,0.000274232,0.00059939903,0.0048809266,0.0017462289,0.0003707516,0.013496805,0.019719133,0.3205638],"study_design_scores_gemma":[0.010546079,0.0367115,0.52627456,0.006143191,0.00026259563,0.00032703308,0.0055737444,0.13567272,0.0006590926,0.27714235,0.00015064067,0.00053649145],"about_ca_topic_score_codex":0.000077940786,"about_ca_topic_score_gemma":0.00003809236,"teacher_disagreement_score":0.32002732,"about_ca_system_score_codex":0.000049851,"about_ca_system_score_gemma":0.00017542033,"threshold_uncertainty_score":0.38070482},"labels":[],"label_agreement":null},{"id":"W1974149035","doi":"10.1097/acm.0b013e3181e16103","title":"Swapping Horses Midstream: Factors Related to Physiciansʼ Changing Their Minds About a Diagnosis","year":2010,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":40,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"National Heart, Lung, and Blood Institute","keywords":"Medical diagnosis; Vignette; Medicine; Generalizability theory; Socioeconomic status; Disease; Medical history; Psychology; Clinical psychology; Family medicine; Social psychology; Developmental psychology; Surgery","score_opus":0.031836931970823774,"score_gpt":0.35212387715130156,"score_spread":0.3202869451804778,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1974149035","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9763803,0.0008682153,0.00007704243,0.015375498,0.0015871316,0.000614863,0.000014411012,0.00026076706,0.004821795],"genre_scores_gemma":[0.98823595,0.00074084295,0.000102325,0.0059174146,0.0015665243,0.000108832486,0.00010565896,0.00006599954,0.0031564715],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9971368,0.000056289093,0.00080895494,0.0006399194,0.00055235263,0.0008056907],"domain_scores_gemma":[0.98948944,0.008699015,0.00021742129,0.0005783896,0.00014223099,0.0008734851],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0009592188,0.00041496457,0.000915438,0.00044772952,0.0001542687,0.000008973854,0.0003051805,0.0005871007,0.0010192862],"category_scores_gemma":[0.052026615,0.00028076908,0.00016319918,0.0010706987,0.00030592078,0.00007739094,0.00011572915,0.002014307,0.00022140714],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021580026,0.00024209393,0.68122846,0.00015206548,0.0005050454,0.0001828647,0.051717855,0.0000075431985,0.05608221,0.0020395075,0.08373209,0.12389447],"study_design_scores_gemma":[0.014249879,0.0036685213,0.653735,0.051065207,0.0022476453,0.00041386907,0.028563207,0.0006106492,0.08683801,0.0046521565,0.15150385,0.002452057],"about_ca_topic_score_codex":0.00017244059,"about_ca_topic_score_gemma":0.0000100046655,"teacher_disagreement_score":0.121442415,"about_ca_system_score_codex":0.000052850544,"about_ca_system_score_gemma":0.000097347605,"threshold_uncertainty_score":0.9999645},"labels":[],"label_agreement":null},{"id":"W1974436001","doi":"10.1207/s15328015tlm1403_3","title":"Comparison of an Aggregate Scoring Method With a Consensus Scoring Method in a Measure of Clinical Reasoning Capacity","year":2002,"lang":"en","type":"article","venue":"Teaching and Learning in Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":74,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Université du Québec à Montréal","funders":"","keywords":"Concordance; Context (archaeology); Test (biology); Psychology; Medical education; Computer science; Medicine","score_opus":0.15035216728360518,"score_gpt":0.47140274565511475,"score_spread":0.3210505783715096,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1974436001","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9893436,0.0013276946,0.005965851,0.00067106704,0.00010469699,0.00025264802,7.397197e-7,0.00006837653,0.002265334],"genre_scores_gemma":[0.8184233,0.00007659307,0.18113387,0.00008499984,0.00015204037,0.000005799668,0.0000040049026,0.000028591561,0.00009083105],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9935829,0.0032116796,0.0016318659,0.00059001934,0.0005881652,0.00039534282],"domain_scores_gemma":[0.9707508,0.027685702,0.000749406,0.0003685741,0.00012453372,0.00032100172],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.019250825,0.00026451534,0.0020024804,0.00036665375,0.0001118922,0.000007647782,0.00011459132,0.00022363356,0.000029036682],"category_scores_gemma":[0.16824566,0.00019817278,0.000084411586,0.000396397,0.00036249144,0.000051146522,0.00004583455,0.0036838178,7.9228914e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004885209,0.00046511122,0.6991663,0.0001420209,0.00005734592,0.000089711066,0.010399317,0.0024370565,0.0006143644,0.00011742582,0.000015347186,0.2860075],"study_design_scores_gemma":[0.016403623,0.006585323,0.58692193,0.057759177,0.00050943403,0.00041296243,0.01075419,0.31932887,0.00047819462,0.000080613405,0.0003208893,0.00044481363],"about_ca_topic_score_codex":0.0047123968,"about_ca_topic_score_gemma":0.00018645263,"teacher_disagreement_score":0.31689182,"about_ca_system_score_codex":0.00004143246,"about_ca_system_score_gemma":0.00006508555,"threshold_uncertainty_score":0.9986147},"labels":[],"label_agreement":null},{"id":"W1974476540","doi":"10.1207/s15328015tlm1302_6","title":"Coordination of Analytic and Similarity-Based Processing Strategies and Expertise in Dermatological Diagnosis","year":2001,"lang":"en","type":"article","venue":"Teaching and Learning in Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":96,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Categorization; Similarity (geometry); Task (project management); Medical diagnosis; Psychology; Cognitive psychology; Contrast (vision); Computer science; Medicine; Artificial intelligence; Pathology; Image (mathematics)","score_opus":0.03750432564261896,"score_gpt":0.3664209226835502,"score_spread":0.32891659704093124,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1974476540","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9885758,0.0039761374,0.0006284513,0.0062438957,0.000012723316,0.00010956147,1.2027873e-7,0.000032827076,0.00042044345],"genre_scores_gemma":[0.9984577,0.00040331017,0.000653958,0.00039741842,0.000032602675,0.000011176713,0.000007482781,0.000007690695,0.00002871429],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9988551,0.0002629551,0.0003432094,0.00022916304,0.0001520282,0.00015751334],"domain_scores_gemma":[0.99553746,0.004180797,0.00008942811,0.0000630253,0.000027003975,0.000102283906],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0014777759,0.0001129467,0.00044975933,0.00021193954,0.00008499429,0.000017507002,0.000023334462,0.00010132907,0.0000122027295],"category_scores_gemma":[0.029100725,0.0000839171,0.0000135337095,0.00015508075,0.0002523884,0.00007093628,0.00001700178,0.00077648397,1.2113223e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000077816796,0.00010924015,0.85335255,0.00011539088,0.00000412894,0.000090704365,0.0023346292,0.00014680966,0.00008156801,0.00006473705,0.000022498241,0.14359993],"study_design_scores_gemma":[0.003626501,0.00053777057,0.8886616,0.0063095894,0.000061043866,0.000092115515,0.004469752,0.095841184,0.000010864824,0.00016219655,0.00012428797,0.000103069244],"about_ca_topic_score_codex":0.0007654743,"about_ca_topic_score_gemma":0.000057689173,"teacher_disagreement_score":0.14349686,"about_ca_system_score_codex":0.000019231778,"about_ca_system_score_gemma":0.00004287406,"threshold_uncertainty_score":0.9790776},"labels":[],"label_agreement":null},{"id":"W1974851520","doi":"10.3138/jvme.1113-151r1","title":"Development and Pilot of Case Manager: A Virtual-Patient Experience for Veterinary Students","year":2014,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Medical education; Documentation; Presentation (obstetrics); Quality (philosophy); Referral; Variety (cybernetics); Test (biology); Medicine; Psychology; Computer science; Nursing; Surgery","score_opus":0.09281422668600062,"score_gpt":0.44458057631748893,"score_spread":0.35176634963148834,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1974851520","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99688816,0.00032347636,0.0011141959,0.00046906667,0.0009343998,0.00019222542,6.9996787e-7,0.0000050894705,0.00007266307],"genre_scores_gemma":[0.98913485,0.00007014411,0.009868753,0.0005868568,0.00021707102,0.0000325395,0.000003364634,0.000010295014,0.00007611921],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99841726,0.00006223974,0.0007026467,0.00013596183,0.0005518737,0.00013000068],"domain_scores_gemma":[0.9972876,0.0015820224,0.00034505915,0.00013346414,0.00019987508,0.0004520117],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.000743293,0.00009947772,0.00031063575,0.00010406759,0.000047254525,0.000010315299,0.000097020486,0.00006275267,0.00007467727],"category_scores_gemma":[0.014778927,0.000077630364,0.00006113123,0.00006825234,0.00009874022,0.00006254825,0.000067780406,0.00013876514,0.0000015210468],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002194034,0.004240084,0.004811711,0.00019634042,0.00011130738,0.0011154874,0.0062583894,9.0615777e-7,0.0009338716,0.00013719637,0.0012520023,0.9787487],"study_design_scores_gemma":[0.020056257,0.20189027,0.2932305,0.026416441,0.0007748869,0.19866896,0.022522062,0.00073587504,0.0019347528,0.0002906626,0.2325706,0.0009087292],"about_ca_topic_score_codex":0.000009760435,"about_ca_topic_score_gemma":4.3508183e-7,"teacher_disagreement_score":0.97783995,"about_ca_system_score_codex":0.000050971703,"about_ca_system_score_gemma":0.00044522644,"threshold_uncertainty_score":0.99352},"labels":[],"label_agreement":null},{"id":"W1974905690","doi":"10.1016/j.nedt.2014.12.006","title":"Absent bystanders and cognitive dissonance: A comment on Timmins &amp; de Vries","year":2014,"lang":"en","type":"letter","venue":"Nurse Education Today","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Cognitive dissonance; Phenomenon; Bystander effect; Psychology; Compassion; Cognition; Social psychology; Psychoanalysis; Psychotherapist; Epistemology; Philosophy; Psychiatry; Theology","score_opus":0.029556240493466755,"score_gpt":0.3655150700068181,"score_spread":0.3359588295133513,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1974905690","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008264059,0.000438288,0.00017903416,0.97916836,0.0012887911,0.00053146307,0.00003692513,0.000052636027,0.010040436],"genre_scores_gemma":[0.07336364,0.00020119992,0.00019154222,0.8986145,0.00310956,0.00015908542,0.0020926301,0.000049015063,0.022218812],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99853367,0.00010953032,0.00029442777,0.0004206934,0.00033933247,0.0003023248],"domain_scores_gemma":[0.99222237,0.0069064507,0.00018293045,0.0003226484,0.00016054764,0.00020505366],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00022285401,0.00026531334,0.00042413917,0.0001002326,0.00008646594,0.00004476785,0.00006357924,0.00043902002,0.00029682313],"category_scores_gemma":[0.010781411,0.00022498368,0.00010410823,0.00007350995,0.00019826749,0.000017663417,0.000010871056,0.0009217511,0.0003203061],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007300544,0.00039060344,0.0018904901,0.00008245121,0.000060844774,0.000012273886,0.0009240356,1.7726963e-7,0.0000013568441,0.00016352473,0.97905254,0.017348722],"study_design_scores_gemma":[0.00087048305,0.0002801931,0.018168522,0.007037433,0.00038567546,0.000040012583,0.00055736606,0.0000046823725,0.00001428766,0.0004949423,0.9719088,0.0002375903],"about_ca_topic_score_codex":0.00008641169,"about_ca_topic_score_gemma":0.0000057938437,"teacher_disagreement_score":0.08055385,"about_ca_system_score_codex":0.00018051053,"about_ca_system_score_gemma":0.0005791527,"threshold_uncertainty_score":0.9975512},"labels":[],"label_agreement":null},{"id":"W1975074072","doi":"10.4997/jrcpe.2011.208","title":"Better clinical decision making and reducing diagnostic error","year":2011,"lang":"en","type":"review","venue":"The Journal of the Royal College of Physicians of Edinburgh","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":150,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; Dalhousie University","funders":"","keywords":"Assertion; Process (computing); Nihilism; Clinical decision making; Computer science; Decision-making; Dual process theory (moral psychology); Management science; Psychology; Medicine; Epistemology; Cognition; Intensive care medicine; Operations management; Psychiatry","score_opus":0.058599249213195995,"score_gpt":0.3910490577982215,"score_spread":0.33244980858502554,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1975074072","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.022385515,0.9732173,0.000064946486,0.00048796815,0.002213866,0.00074546353,0.00009184424,0.0000065710437,0.0007865263],"genre_scores_gemma":[0.04568219,0.95035124,0.00089792715,0.0002833461,0.0022587155,0.0000044660496,0.0000014503139,0.000060529524,0.00046014372],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9956355,0.00041952677,0.0027387522,0.00023393778,0.0006821135,0.0002901643],"domain_scores_gemma":[0.9322985,0.062466495,0.0038811527,0.0008067607,0.0004228031,0.00012427656],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0033758052,0.00038703645,0.0032797565,0.0001522772,0.00010478746,0.0000072683883,0.00083719497,0.00032761254,0.00011926906],"category_scores_gemma":[0.053739313,0.00018620846,0.0017786883,0.0003628491,0.00061679876,0.000044254615,0.0003579621,0.001379194,0.0000017591168],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00047439325,0.0006754332,0.0004262588,0.000984338,0.00096163835,0.000050201797,0.00018188036,0.000032136115,9.317819e-7,0.000037151436,0.07958723,0.9165884],"study_design_scores_gemma":[0.0028636376,0.003180668,0.0037430702,0.7750512,0.027562117,0.0005821646,0.00023341502,0.00011804242,0.000022069074,0.002966868,0.18312494,0.00055184204],"about_ca_topic_score_codex":0.000017641474,"about_ca_topic_score_gemma":0.0000019963866,"teacher_disagreement_score":0.91603655,"about_ca_system_score_codex":0.000053892178,"about_ca_system_score_gemma":0.0004723621,"threshold_uncertainty_score":0.95423144},"labels":[],"label_agreement":null},{"id":"W1977281256","doi":"10.1207/s15328015tlm1701_3","title":"Use of \"Standardized Examinees\" to Screen for Standardized-Patient Scoring Bias in a Clinical Skills Examination","year":2005,"lang":"en","type":"article","venue":"Teaching and Learning in Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"College of Physicians and Surgeons of Ontario","funders":"","keywords":"Standardized test; Medicine; Psychology; Physical examination; Medical education; Medical physics; Family medicine; Radiology; Mathematics education","score_opus":0.08751713581490382,"score_gpt":0.4095279154759775,"score_spread":0.32201077966107367,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1977281256","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9916499,0.00030503585,0.0047279536,0.0020675634,0.00013644285,0.0006933308,0.000010182031,0.00006166493,0.00034792296],"genre_scores_gemma":[0.971014,0.00015535504,0.027173465,0.0008130218,0.0003422496,0.00002483322,0.00005340285,0.00003109112,0.00039256207],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99661124,0.0008564653,0.0012723912,0.00045467244,0.00046771453,0.00033753656],"domain_scores_gemma":[0.9701877,0.028932368,0.0002499026,0.00023347701,0.00012937054,0.00026712954],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.010134002,0.00020090274,0.0010003822,0.0004577975,0.0000878786,0.000014521005,0.00006571158,0.00015681854,0.000023288421],"category_scores_gemma":[0.2987177,0.00016053182,0.000091000235,0.0002539108,0.00013663925,0.00008532769,0.00005543635,0.0012455035,0.0000014808027],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006440417,0.00039275756,0.2854784,0.000049428152,0.000025508933,0.000016680382,0.0047997087,0.0011523673,0.00018054657,0.00006288536,0.0003836677,0.706814],"study_design_scores_gemma":[0.011704341,0.004476918,0.9438555,0.013098621,0.0001236192,0.000020666623,0.0017077596,0.0053125313,0.00007763032,0.000020151205,0.019366397,0.00023586572],"about_ca_topic_score_codex":0.0007095012,"about_ca_topic_score_gemma":0.00012868203,"teacher_disagreement_score":0.70657814,"about_ca_system_score_codex":0.00012241316,"about_ca_system_score_gemma":0.00008523024,"threshold_uncertainty_score":0.7071895},"labels":[],"label_agreement":null},{"id":"W1977627312","doi":"10.1097/pts.0b013e3181f6c01a","title":"Emotional Influences in Patient Safety","year":2010,"lang":"en","type":"review","venue":"Journal of Patient Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":143,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"Agency for Healthcare Research and Quality","keywords":"Feeling; Compromise; Health care; Psychology; Openness to experience; Patient safety; Social psychology","score_opus":0.03612252521162444,"score_gpt":0.3732221828609067,"score_spread":0.33709965764928224,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1977627312","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0039607477,0.9914043,0.000019462224,0.00030052522,0.0021072752,0.00059676653,0.000078387064,0.000011207249,0.0015213283],"genre_scores_gemma":[0.0005850879,0.99735785,0.0010371989,0.00030432607,0.000530734,0.0000071636828,0.00008004448,0.00003600897,0.00006156424],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99478066,0.0002460021,0.0033609383,0.00029013213,0.0009885152,0.0003337351],"domain_scores_gemma":[0.9891673,0.0072501507,0.0023943526,0.00034362942,0.0003994339,0.0004451381],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00085357117,0.0003948661,0.0025254246,0.0004262591,0.000058927348,0.000017254779,0.00021362245,0.0005654189,0.00043782522],"category_scores_gemma":[0.024487546,0.0002560093,0.0011372152,0.000396565,0.0001206882,0.00008655925,0.00008960152,0.00217464,0.000075911514],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020178745,0.00063451566,0.0013880075,0.0008236236,0.000118450735,0.00043506952,0.0001124643,0.000020212095,1.2531018e-7,0.00006437739,0.0007644889,0.9954369],"study_design_scores_gemma":[0.0010153649,0.001266047,0.0037654615,0.0998503,0.00057039334,0.0010329094,0.000029735023,0.0000025234729,4.6171172e-7,0.00007468232,0.89217085,0.00022128991],"about_ca_topic_score_codex":0.000015325637,"about_ca_topic_score_gemma":0.000009637795,"teacher_disagreement_score":0.9952156,"about_ca_system_score_codex":0.00034902853,"about_ca_system_score_gemma":0.001307677,"threshold_uncertainty_score":0.9999892},"labels":[],"label_agreement":null},{"id":"W19780888","doi":"10.3233/978-1-60750-028-5-781","title":"Learning from Feedback In BioWorld","year":2009,"lang":"en","type":"book-chapter","venue":"Frontiers in artificial intelligence and applications","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Computer science; Psychology","score_opus":0.05002268331553205,"score_gpt":0.3191164113967818,"score_spread":0.26909372808124976,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W19780888","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0049877865,0.01832671,0.124137886,0.008176147,0.0010426213,0.004512789,0.00013370933,0.00032626445,0.8383561],"genre_scores_gemma":[0.41677666,0.044033367,0.039496787,0.0036485654,0.004033946,0.000697306,0.002279095,0.00029194474,0.48874232],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9984851,0.000013178626,0.00062151876,0.0005081918,0.0001579418,0.00021406803],"domain_scores_gemma":[0.9985174,0.0008915606,0.00013083166,0.0002732797,0.000044116525,0.00014280598],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0001687958,0.00021801844,0.0005083086,0.00026404258,0.00005651709,0.000031491047,0.00010643016,0.00033601414,0.0001635283],"category_scores_gemma":[0.0007990283,0.0002238265,0.000089885645,0.00013120711,0.00019560497,0.000026743926,0.000032974935,0.0007874729,0.00016428115],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000060570197,0.00012093878,0.0031012907,0.000008309194,0.000018713285,0.000020315743,0.000117732394,0.000101470476,0.00000801053,0.02939642,0.0011619843,0.96588427],"study_design_scores_gemma":[0.00017354959,0.00022382067,0.0033934803,0.0031784512,0.00015014049,0.0000049364835,0.00051530614,0.0024475928,0.00022864837,0.8039706,0.18517005,0.00054343225],"about_ca_topic_score_codex":0.00013024044,"about_ca_topic_score_gemma":0.00014304268,"teacher_disagreement_score":0.9653408,"about_ca_system_score_codex":0.00007948696,"about_ca_system_score_gemma":0.0000766801,"threshold_uncertainty_score":0.91273797},"labels":[],"label_agreement":null},{"id":"W1978234974","doi":"10.3138/jvme.33.3.426","title":"Increased Student Self-Confidence in Clinical Reasoning Skills Associated with Case-Based Learning (CBL)","year":2006,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":45,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"University Research Committee, Emory University","keywords":"Competence (human resources); Psychology; Self-confidence; Medical education; Confidence interval; Low Confidence; Mathematics education; Medicine; Internal medicine; Social psychology","score_opus":0.03197173019140101,"score_gpt":0.4112468638534087,"score_spread":0.3792751336620077,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1978234974","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9968617,0.00032224678,0.0002523341,0.0014588538,0.0005410009,0.00019901205,6.943152e-7,0.00003792436,0.00032622542],"genre_scores_gemma":[0.99488217,0.000084516185,0.0031603028,0.000996456,0.0007374023,0.000011247655,0.00002524267,0.00002508915,0.000077558376],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.996255,0.00061696203,0.0014897696,0.0002592689,0.0010849177,0.00029412037],"domain_scores_gemma":[0.9885188,0.009191607,0.0009279693,0.00020208664,0.00051301817,0.00064651616],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0043135136,0.00018566966,0.0006569443,0.00022983245,0.000072168536,0.000033413435,0.00014866474,0.0002791882,0.00020905753],"category_scores_gemma":[0.076451205,0.00014057972,0.00018721994,0.00038423744,0.00013770488,0.00010808229,0.00002877568,0.0012715989,0.000007279706],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007247487,0.01055178,0.95879275,0.000033910957,0.00009471616,0.014784069,0.00019585183,0.00010411892,0.000025993024,0.00003958466,0.0012209436,0.01343153],"study_design_scores_gemma":[0.006167735,0.0058077592,0.9617562,0.010548305,0.00023965008,0.013282857,0.00032073472,0.000903311,0.000013640355,0.00003623706,0.00075191655,0.0001716473],"about_ca_topic_score_codex":0.00030360333,"about_ca_topic_score_gemma":0.000049157883,"teacher_disagreement_score":0.07213769,"about_ca_system_score_codex":0.00030977107,"about_ca_system_score_gemma":0.003933572,"threshold_uncertainty_score":0.93132824},"labels":[],"label_agreement":null},{"id":"W1978440292","doi":"10.1034/j.1600-0579.2000.040401.x","title":"Clinical decision‐making in restorative dentistry. Content‐analysis of diagnostic thinking processes and concurrent concepts used in an educational environment","year":2000,"lang":"en","type":"article","venue":"European Journal Of Dental Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":37,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Context (archaeology); Psychology; Recall; Heuristics; Cognition; Session (web analytics); Introspection; Medical education; Medicine; Dentistry; Cognitive psychology; Computer science; Psychiatry","score_opus":0.06419530810166842,"score_gpt":0.428129106286398,"score_spread":0.3639337981847296,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1978440292","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99702996,0.0019345592,0.00016792827,0.00021252003,0.00035095852,0.00015275551,0.000005179836,0.0000022529614,0.00014386409],"genre_scores_gemma":[0.99745786,0.00084500114,0.0012912933,0.00016030164,0.00016569997,0.000002757124,0.000032254793,0.000011317584,0.00003352275],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99756736,0.00045233648,0.0012809952,0.00020911408,0.0003767791,0.000113389935],"domain_scores_gemma":[0.9896587,0.009332503,0.00056958554,0.00012821727,0.00013707475,0.00017391413],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001350562,0.00011073659,0.0004538772,0.00029592734,0.00002941143,0.000032124924,0.0001218663,0.00003379291,0.00036724264],"category_scores_gemma":[0.022385549,0.00009643982,0.00010510174,0.00036015964,0.00013473195,0.00017350349,0.000021906453,0.0002733837,0.00000981464],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00016414253,0.0018572968,0.87024033,0.000014071248,0.00009142676,0.00005345185,0.0018819548,0.0001640617,0.000017725708,0.000027109703,0.000053122698,0.12543531],"study_design_scores_gemma":[0.0015922511,0.00048958446,0.9914079,0.003296178,0.00041988265,0.00008498504,0.0023953079,0.00007213714,0.00001302081,0.00008445192,0.000063313084,0.00008099733],"about_ca_topic_score_codex":0.000017693166,"about_ca_topic_score_gemma":0.000081087725,"teacher_disagreement_score":0.1253543,"about_ca_system_score_codex":0.00010967543,"about_ca_system_score_gemma":0.00043422505,"threshold_uncertainty_score":0.9858493},"labels":[],"label_agreement":null},{"id":"W1978509004","doi":"10.1016/s0006-3223(00)00566-7","title":"302. Treating dysthymia in primary care: sertraline vs IPT (a two-year follow up)","year":2000,"lang":"en","type":"article","venue":"Biological Psychiatry","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Sertraline; Primary care; Medicine; Psychiatry; Psychology; Pediatrics; Internal medicine; Family medicine; Anxiety; Antidepressant","score_opus":0.023933493241069966,"score_gpt":0.31478579433899384,"score_spread":0.29085230109792387,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1978509004","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9520604,0.0015192284,0.0000045043657,0.0021569782,0.0007138671,0.0003225516,0.000020441084,0.00015846078,0.043043587],"genre_scores_gemma":[0.988214,0.00033483913,0.0047863885,0.0040998426,0.0008483919,0.00002851235,0.00030004408,0.000022195858,0.0013657743],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9980979,0.000109346955,0.0005728632,0.00055294647,0.00020174617,0.00046522118],"domain_scores_gemma":[0.99797827,0.0012670562,0.00007280547,0.00036275806,0.000040484352,0.00027862357],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00026478493,0.00024981663,0.000589323,0.00005803724,0.000065750806,0.000017307744,0.00014356572,0.0003331397,0.0013479193],"category_scores_gemma":[0.0021076712,0.00016928652,0.00025527668,0.0002956525,0.00011142251,0.000031222393,0.000031649564,0.00043672192,0.00036657482],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016450449,0.00052528695,0.9604141,0.00003662794,0.000028544155,0.000050286515,0.00009252661,0.000009626623,0.00008591437,0.0004474342,0.0027613947,0.033903204],"study_design_scores_gemma":[0.009183934,0.0016784949,0.98277074,0.0012073152,0.000061490566,0.00004403248,0.00022136547,0.000024761543,0.00002242085,0.00080806087,0.0036702452,0.0003071518],"about_ca_topic_score_codex":0.0001451202,"about_ca_topic_score_gemma":0.000053255197,"teacher_disagreement_score":0.04167781,"about_ca_system_score_codex":0.000098037955,"about_ca_system_score_gemma":0.00015607882,"threshold_uncertainty_score":0.999565},"labels":[],"label_agreement":null},{"id":"W1978687968","doi":"10.1007/s10459-015-9585-1","title":"Why verifying diagnostic decisions with a checklist can help: insights from eye tracking","year":2015,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":35,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Hamilton General Hospital","funders":"","keywords":"Checklist; Scrutiny; Key (lock); Process (computing); Computer science; Variable (mathematics); Eye tracking; Psychology; Medical physics; Artificial intelligence; Medicine; Cognitive psychology; Mathematics; Computer security; Programming language; Law","score_opus":0.050827324082609183,"score_gpt":0.43009725880824107,"score_spread":0.37926993472563186,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1978687968","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9604104,0.022510577,0.00086570275,0.012571089,0.0014904893,0.0004952974,0.000004364289,0.000047897345,0.001604139],"genre_scores_gemma":[0.97838867,0.0029112997,0.011595425,0.0066328263,0.00030172692,0.00008079706,0.00004133676,0.000010208402,0.000037708523],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9981238,0.000088646484,0.00046894065,0.00046657064,0.0005260473,0.00032596823],"domain_scores_gemma":[0.98789173,0.010837474,0.0002449266,0.00027210196,0.0001886055,0.00056514895],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00057603803,0.00013248417,0.00029354173,0.00017482572,0.00020288769,0.000051216844,0.00015953308,0.00005126659,0.000014119704],"category_scores_gemma":[0.05209529,0.00009756307,0.000025678033,0.00092863454,0.00028527156,0.0003607271,0.000022877808,0.00020885046,0.000010830589],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000088594665,0.0007145287,0.79787743,0.00004364286,0.000004336896,0.000011903349,0.0054774135,0.001690646,0.000006105214,0.00092772796,0.0016846876,0.191473],"study_design_scores_gemma":[0.0030010191,0.0027402786,0.8538432,0.02431793,0.0000494254,0.000040726773,0.019346464,0.0017931456,0.00015803709,0.02260292,0.07154663,0.0005602393],"about_ca_topic_score_codex":0.0062362715,"about_ca_topic_score_gemma":0.0055733575,"teacher_disagreement_score":0.19091277,"about_ca_system_score_codex":0.00028042047,"about_ca_system_score_gemma":0.0029541522,"threshold_uncertainty_score":0.95588934},"labels":[],"label_agreement":null},{"id":"W1978974820","doi":"10.1097/00001888-200110001-00029","title":"Difficulties in Recognizing Oneʼs Own Incompetence","year":2001,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":279,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto General Hospital","funders":"","keywords":"Quartile; Competence (human resources); Psychology; Set (abstract data type); Clinical psychology; Social psychology; Cognitive psychology; Applied psychology; Developmental psychology; Medicine; Computer science","score_opus":0.06041760094654559,"score_gpt":0.3718384539956498,"score_spread":0.3114208530491042,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1978974820","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.971138,0.0025258465,0.000091928574,0.016447527,0.0004213282,0.00025233362,7.8532247e-7,0.000109648594,0.0090126125],"genre_scores_gemma":[0.9828412,0.0067467224,0.00025278958,0.0053611486,0.0009797773,0.000021815353,0.00002941458,0.000020686723,0.003746466],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9980456,0.000057836736,0.00069298665,0.0003718547,0.00042151954,0.00041015254],"domain_scores_gemma":[0.9930115,0.0062207785,0.00012459263,0.00027660342,0.000071307855,0.00029521406],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00086873834,0.00018701275,0.0005768738,0.00019170849,0.0000400841,0.0000027306526,0.0001555027,0.00027784213,0.00045528027],"category_scores_gemma":[0.054686364,0.00012983599,0.00004834275,0.0005105136,0.00026382037,0.000053798874,0.00005711429,0.0010873072,0.00015796581],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00034370492,0.000087099485,0.89192957,0.000050191058,0.000036752186,0.00069290044,0.004523956,0.000007992078,0.0035112672,0.0004794904,0.021467255,0.07686983],"study_design_scores_gemma":[0.00973029,0.0009273788,0.93965447,0.020150214,0.0002020858,0.0009528247,0.0053958977,0.00052917015,0.00042484613,0.0026471152,0.018958036,0.0004276966],"about_ca_topic_score_codex":0.0003599576,"about_ca_topic_score_gemma":0.000022342609,"teacher_disagreement_score":0.07644213,"about_ca_system_score_codex":0.00007945841,"about_ca_system_score_gemma":0.00007813927,"threshold_uncertainty_score":0.9532764},"labels":[],"label_agreement":null},{"id":"W1979178476","doi":"10.1080/10401330709336621","title":"Expert-Type Knowledge Structure in Medical Students is Associated With Increased Odds of Diagnostic Success","year":2007,"lang":"en","type":"article","venue":"Teaching and Learning in Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":30,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"South Health Campus; Alberta Health Services; University of Calgary","funders":"","keywords":"Odds; Medicine; Odds ratio; Medical diagnosis; Logistic regression; Internal medicine; Pathology","score_opus":0.017147574943745116,"score_gpt":0.38419507119419877,"score_spread":0.36704749625045363,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1979178476","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9951196,0.0022184802,0.00007929682,0.0010715995,0.0001328091,0.00020076316,9.557741e-7,0.000056598255,0.0011199038],"genre_scores_gemma":[0.99800295,0.00026225302,0.0001295706,0.0011075736,0.0002026359,0.0000030439742,0.000050674436,0.000028046441,0.00021323198],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9970243,0.0005686021,0.00067491626,0.00036182124,0.0009838429,0.00038654418],"domain_scores_gemma":[0.9533167,0.045810957,0.00019192602,0.00019334337,0.00010568305,0.00038140902],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.007723229,0.00021911357,0.0008237575,0.00033413092,0.0000890607,0.000009100795,0.00018510976,0.00028293172,0.00024645746],"category_scores_gemma":[0.3246136,0.0001502036,0.00002948022,0.0004772835,0.00030085942,0.000034632318,0.000074672826,0.002162212,0.0000014377517],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00033306677,0.00052119105,0.95795375,0.000046297788,0.000054143664,0.0004952675,0.009558878,0.00001088779,0.0001019508,0.00002603844,0.00036600197,0.030532526],"study_design_scores_gemma":[0.008420246,0.001500063,0.9696973,0.017423479,0.00007999826,0.00006504258,0.0015765814,0.0005725864,0.00003605814,0.000022343896,0.00044738388,0.00015890352],"about_ca_topic_score_codex":0.003141624,"about_ca_topic_score_gemma":0.00091266795,"teacher_disagreement_score":0.31689036,"about_ca_system_score_codex":0.000088914014,"about_ca_system_score_gemma":0.00015051245,"threshold_uncertainty_score":0.93938553},"labels":[],"label_agreement":null},{"id":"W1979516180","doi":"10.1136/emermed-2012-202303","title":"How emergency physicians use biomarkers: insights from a qualitative assessment of script concordance tests","year":2013,"lang":"en","type":"article","venue":"Emergency Medicine Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Concordance; Medicine; Intuition; Cognition; Emergency department; Bayesian probability; Clinical decision making; Medical emergency; Family medicine; Artificial intelligence; Psychiatry; Psychology; Computer science; Internal medicine","score_opus":0.07805561048402467,"score_gpt":0.42253630409568993,"score_spread":0.3444806936116653,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1979516180","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9690762,0.0042923493,0.0040536285,0.0104772085,0.005418957,0.0005148249,0.000026454238,0.000045929602,0.0060944483],"genre_scores_gemma":[0.9761817,0.0042882175,0.0027457755,0.00016323481,0.0021854858,0.00003744123,0.00006385313,0.00004931763,0.014284993],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9959853,0.00040429487,0.0015870468,0.00042355704,0.0011212623,0.00047852963],"domain_scores_gemma":[0.99320513,0.0030022128,0.0011755886,0.00053058186,0.0014024444,0.00068406976],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00059850374,0.00036558253,0.0010285343,0.00023077117,0.00014009455,0.000017467994,0.00024659606,0.00014087268,0.018900039],"category_scores_gemma":[0.035289798,0.00025330504,0.00034149826,0.0006337867,0.00028390548,0.00048442298,0.000052663316,0.0006836099,0.000032780332],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000098136385,0.0009956525,0.15414198,0.000062869716,0.001032077,0.000025091334,0.008958168,0.000002952038,0.034883514,0.0002993427,0.7951411,0.0043591377],"study_design_scores_gemma":[0.0039809668,0.0023294552,0.9551721,0.0048610372,0.0006821192,0.000045605742,0.015651034,0.0005655476,0.0003342754,0.006976884,0.009002811,0.00039812346],"about_ca_topic_score_codex":0.0017298709,"about_ca_topic_score_gemma":0.00008683576,"teacher_disagreement_score":0.80103016,"about_ca_system_score_codex":0.00010810866,"about_ca_system_score_gemma":0.00022286458,"threshold_uncertainty_score":0.9999919},"labels":[],"label_agreement":null},{"id":"W1980216664","doi":"10.3402/meo.v16i0.5890","title":"An analysis of clinical reasoning through a recent and comprehensive approach: the dual-process theory","year":2011,"lang":"en","type":"article","venue":"Medical Education Online","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":378,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Université de Sherbrooke","funders":"","keywords":"Dual (grammatical number); Process (computing); Psychology; Medical education; Engineering ethics; Management science; Computer science; Medicine; Engineering; Philosophy","score_opus":0.11425318485257535,"score_gpt":0.47204487961927066,"score_spread":0.3577916947666953,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1980216664","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9913381,0.0008290286,0.0015128785,0.0030364322,0.0003250441,0.00022946247,0.000014343924,0.000038609043,0.0026761354],"genre_scores_gemma":[0.9805492,0.001736346,0.008519814,0.0081652645,0.0004583559,0.000023518623,0.00044560962,0.000015836404,0.00008600839],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9976113,0.0003939019,0.0008689284,0.00038863518,0.000559198,0.00017805511],"domain_scores_gemma":[0.99278206,0.00532991,0.00031099614,0.0005576752,0.0004988896,0.00052046805],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0016187368,0.0001404504,0.00067979645,0.00008694057,0.000064814914,0.000007485385,0.00015631148,0.00020970854,0.0014901202],"category_scores_gemma":[0.07240849,0.000084077525,0.00016359364,0.0006752726,0.0006560719,0.000050835708,0.00004113872,0.0004537343,0.0000033122928],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001168759,0.02904298,0.42972592,0.00010441059,0.0019546377,0.000017034094,0.012575912,0.000015121905,0.0000053664303,0.0148834195,0.0012702508,0.50923616],"study_design_scores_gemma":[0.001317743,0.0004985949,0.9705809,0.0004889083,0.0027475653,0.000040107414,0.010912547,0.009658021,0.000016744805,0.0018446934,0.0017474683,0.00014671352],"about_ca_topic_score_codex":0.00019636087,"about_ca_topic_score_gemma":0.000022328497,"teacher_disagreement_score":0.540855,"about_ca_system_score_codex":0.000013095635,"about_ca_system_score_gemma":0.0011111973,"threshold_uncertainty_score":0.99942267},"labels":[],"label_agreement":null},{"id":"W1980329199","doi":"10.1097/00001888-200303000-00005","title":"Development of an Undergraduate Medical Curriculum","year":2003,"lang":"en","type":"review","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McGill University","funders":"Molson Foundation; McGill University","keywords":"Curriculum; Mainstream; Interactivity; Variety (cybernetics); Medical education; Learning styles; Computer science; Multimedia; Medicine; Mathematics education; Pedagogy; Psychology; Political science","score_opus":0.09565868148907192,"score_gpt":0.4588380886805036,"score_spread":0.36317940719143166,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1980329199","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00016201584,0.9935786,0.00022249745,0.00066151895,0.0008254458,0.00073689426,0.00000481785,0.00008195059,0.0037262419],"genre_scores_gemma":[0.000019792426,0.9939568,0.0017893268,0.00090728374,0.0009639642,0.00007410001,0.0005423222,0.000085690896,0.0016607099],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99372184,0.00028266598,0.0028847395,0.0007394234,0.0018295249,0.000541811],"domain_scores_gemma":[0.9913782,0.005410322,0.0009168576,0.0007017284,0.00015760065,0.0014353027],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0028176422,0.0006146516,0.0042600157,0.0003289175,0.000058163354,0.0000021634858,0.0005004716,0.0017107093,0.001153368],"category_scores_gemma":[0.05928909,0.0003782057,0.00030137348,0.0006692658,0.0004799742,0.00003552853,0.00010630956,0.0026077307,0.0001810785],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000063614148,0.000096686206,0.00003470622,0.0055654636,0.00023798228,0.00013244613,0.00018521278,1.6224282e-8,2.340639e-7,0.0007694442,0.010097703,0.98287374],"study_design_scores_gemma":[0.001187269,0.00023973224,0.0000058328333,0.1983338,0.0017901341,0.00066266826,0.00009997743,0.0000056855292,0.0000013102482,0.00015694373,0.79725325,0.00026337174],"about_ca_topic_score_codex":0.000022590608,"about_ca_topic_score_gemma":0.000003120638,"teacher_disagreement_score":0.98261034,"about_ca_system_score_codex":0.00015683025,"about_ca_system_score_gemma":0.0030377703,"threshold_uncertainty_score":0.99986696},"labels":[],"label_agreement":null},{"id":"W1980392619","doi":"10.1007/s40037-012-0017-0","title":"Effects of two different instructional formats on scores and reliability of a script concordance test","year":2012,"lang":"en","type":"article","venue":"Perspectives on Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"University of Adelaide; Universitair Medisch Centrum Utrecht; Universiteit Utrecht; McGill University","keywords":"Concordance; Test (biology); Reliability (semiconductor); Computer science; Medical education; Psychology; Medical physics; Natural language processing; Medicine; Internal medicine","score_opus":0.010085203556084886,"score_gpt":0.3485447254425015,"score_spread":0.3384595218864166,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1980392619","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9959408,0.0007631792,0.00006995748,0.0015763829,0.0005147089,0.00026217516,0.0000038348667,0.0000141838755,0.00085480016],"genre_scores_gemma":[0.9983777,0.00032056376,0.0004935506,0.00038797664,0.00033701488,0.000031712094,0.0000077002005,0.000006924357,0.000036874844],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99887496,0.000056426354,0.00026847023,0.00017988871,0.00048423852,0.00013601447],"domain_scores_gemma":[0.99136066,0.007814676,0.0001323404,0.00016941613,0.00017019553,0.0003527358],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00027683924,0.00010091221,0.00028117382,0.00007119364,0.00002722591,0.0000027147512,0.00004476422,0.000071881324,0.00009279529],"category_scores_gemma":[0.12250681,0.00007150479,0.000059571666,0.00009595865,0.0003814421,0.000048202026,0.000017952805,0.00021905171,0.000004153092],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00060966675,0.014264834,0.89048696,0.0004687505,0.000039822968,8.8981625e-7,0.005342027,9.937858e-7,0.000729604,0.02492734,0.0010830472,0.062046077],"study_design_scores_gemma":[0.0015575377,0.0009497331,0.9886953,0.0024033447,0.00005796305,0.000013663681,0.0010702239,0.00006101545,0.0039752955,0.0011137754,0.00003904256,0.00006309802],"about_ca_topic_score_codex":0.00008537713,"about_ca_topic_score_gemma":0.0000016379095,"teacher_disagreement_score":0.12222997,"about_ca_system_score_codex":0.00009932859,"about_ca_system_score_gemma":0.0003177257,"threshold_uncertainty_score":0.88488466},"labels":[],"label_agreement":null},{"id":"W1980536747","doi":"10.12968/bjtr.2003.10.11.13456","title":"The role of experience in the development of clinical reasoning","year":2003,"lang":"en","type":"article","venue":"British Journal of Therapy and Rehabilitation","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Psychology; Signs and symptoms; Medicine; Medical education; Surgery","score_opus":0.03546725680112975,"score_gpt":0.3779739300421593,"score_spread":0.34250667324102957,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1980536747","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9916453,0.007580474,0.000042060456,0.00026623846,0.00008317725,0.0001101322,2.7406e-7,8.557021e-7,0.0002714955],"genre_scores_gemma":[0.9893736,0.0023620431,0.008175926,0.00004649355,0.000027262937,0.000003400273,2.4206636e-7,0.0000027954188,0.000008183477],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99845517,0.00030930093,0.0008915636,0.000060992203,0.00021632473,0.00006663274],"domain_scores_gemma":[0.9854084,0.01394883,0.0003527262,0.000075253985,0.00017652707,0.000038220467],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0042494615,0.000037361762,0.00021353342,0.000022709635,0.000055382454,0.000010055996,0.00006241786,0.000044137352,0.0000059093163],"category_scores_gemma":[0.042368,0.00002181794,0.00009293198,0.00009157198,0.00020867132,0.000035519915,0.0000039713295,0.00016275638,9.311073e-8],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00029841255,0.00035008386,0.2973087,0.000007839201,0.000023036786,0.0000043851705,0.00868022,0.0000035321145,0.00018129138,0.0004746882,0.000039772967,0.692628],"study_design_scores_gemma":[0.0015609608,0.00077511865,0.9793077,0.001672924,0.000008929361,0.00016779064,0.011500237,0.000012521066,0.0004532146,0.001647164,0.0028635103,0.00002992195],"about_ca_topic_score_codex":0.000014763034,"about_ca_topic_score_gemma":0.000014349305,"teacher_disagreement_score":0.6925981,"about_ca_system_score_codex":0.0000069188404,"about_ca_system_score_gemma":0.00013039066,"threshold_uncertainty_score":0.96569854},"labels":[],"label_agreement":null},{"id":"W1982729363","doi":"10.1097/opx.0b013e3182776002","title":"Optometrists’ Clinical Reasoning Made Explicit","year":2012,"lang":"en","type":"article","venue":"Optometry and Vision Science","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Sherbrooke; Université de Montréal","funders":"Social Sciences and Humanities Research Council of Canada; Canadian Optometric Education Trust Fund","keywords":"Optometry; Computer science; Psychology; Medicine","score_opus":0.049997534920460024,"score_gpt":0.5305545808577723,"score_spread":0.4805570459373123,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1982729363","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98494893,0.0009687508,0.0008510002,0.00041082656,0.00070745585,0.00013156951,0.0000020442308,0.00006487134,0.01191456],"genre_scores_gemma":[0.9893764,0.00039384616,0.008108564,0.0008540755,0.00043187934,0.0000046330124,0.0000026774487,0.000010451923,0.0008174278],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9976801,0.000051559942,0.00046006893,0.0004777185,0.00076082,0.00056971866],"domain_scores_gemma":[0.9921323,0.0059106015,0.00012204934,0.0004534288,0.00015505504,0.001226552],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005390493,0.00014014619,0.00036345137,0.00044695783,0.00026684155,0.00009490602,0.0001885459,0.00012937731,0.0001680938],"category_scores_gemma":[0.071225114,0.00010389725,0.00010279611,0.0025274954,0.0006313238,0.00041725585,0.00023371581,0.00033787527,0.00018524863],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000060949238,0.00031784736,0.9168903,0.0000106179705,0.00000635657,0.000005168226,0.000061080435,1.3812135e-7,0.00087735435,0.0007023207,0.0006169961,0.080450855],"study_design_scores_gemma":[0.00095188234,0.00037017794,0.99278766,0.0005879467,0.00003483801,0.00010530979,0.00015966182,0.00035020502,0.0012393731,0.000017540617,0.0032497135,0.0001457102],"about_ca_topic_score_codex":0.0000060346083,"about_ca_topic_score_gemma":1.6790707e-8,"teacher_disagreement_score":0.080305144,"about_ca_system_score_codex":0.00003447692,"about_ca_system_score_gemma":0.00009066622,"threshold_uncertainty_score":0.93659836},"labels":[],"label_agreement":null},{"id":"W1985476543","doi":"10.1197/j.aem.2007.04.011","title":"Profiles in Patient Safety: A “Perfect Storm” in the Emergency Department","year":2007,"lang":"en","type":"review","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":81,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Medicine; Emergency department; Medical emergency; Process (computing); Commission; Patient safety; Intensive care medicine; Nursing; Law","score_opus":0.1125582382872096,"score_gpt":0.4546555336715895,"score_spread":0.34209729538437994,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1985476543","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004093622,0.9825062,0.000008903302,0.0011682101,0.0030538887,0.0038584312,0.000034535566,0.0000598651,0.0052163224],"genre_scores_gemma":[0.001116983,0.99458337,0.000024247858,0.0004331365,0.0017947939,0.00072699273,0.0006513585,0.000098040844,0.00057109597],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99063236,0.0006562645,0.005286921,0.001013168,0.0013595137,0.0010517971],"domain_scores_gemma":[0.99217093,0.005333798,0.0010104086,0.0009765515,0.00010848393,0.00039983622],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0047217156,0.00088988955,0.0032264607,0.0007736869,0.00007799434,6.21583e-7,0.0006553248,0.0011931619,0.0028567417],"category_scores_gemma":[0.035669215,0.0004972381,0.0006135627,0.0023088728,0.00017312651,0.00005284123,0.00013767219,0.0039533013,0.00020462931],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013068301,0.00060604594,0.024937637,0.0053564645,0.00017878169,0.0005120395,0.0017931658,0.0000010389119,6.658656e-7,0.00025359634,0.06503859,0.9011913],"study_design_scores_gemma":[0.0011550105,0.0008966755,0.0052020894,0.07221723,0.0011655752,0.00016569896,0.00033061835,0.0000045080947,3.426036e-7,0.00014132405,0.9182663,0.0004546083],"about_ca_topic_score_codex":0.00047921168,"about_ca_topic_score_gemma":0.00022230102,"teacher_disagreement_score":0.9007367,"about_ca_system_score_codex":0.0004523973,"about_ca_system_score_gemma":0.0004351061,"threshold_uncertainty_score":0.99974793},"labels":[],"label_agreement":null},{"id":"W1985501359","doi":"10.1016/j.pain.2006.04.003","title":"Testing two accounts of pain underestimation","year":2006,"lang":"en","type":"article","venue":"Pain","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":132,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Northern British Columbia","funders":"","keywords":"Cheating; Psychology; Context (archaeology); Health care; Priming (agriculture); Health professionals; Medicine; Social psychology; Clinical psychology","score_opus":0.039105549891724774,"score_gpt":0.3417666516562434,"score_spread":0.30266110176451866,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1985501359","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.89691776,0.00021238676,0.08060796,0.0014512022,0.00009126483,0.00025641674,0.0000062446593,0.00013097661,0.02032579],"genre_scores_gemma":[0.988983,9.532953e-7,0.010050438,0.0005552407,0.00015512381,0.000005119383,0.00003010908,0.000008738095,0.00021128077],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9991753,0.00020527691,0.00023807512,0.00010768037,0.00015592447,0.0001177492],"domain_scores_gemma":[0.96246743,0.037148286,0.00009215373,0.0001495509,0.00009133522,0.00005124889],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0038343018,0.000058543817,0.00014945169,0.000034223423,0.000021343141,0.0000063673283,0.00003062054,0.00003846876,0.00005314588],"category_scores_gemma":[0.20111345,0.000049827675,0.00003788924,0.00014374283,0.000035210367,0.000016819728,0.0000097589555,0.00006601023,0.000024957833],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000024213105,0.00027486554,0.7399087,0.000057139416,0.000011423587,0.00001760142,0.000030730447,0.00032974276,0.0021120163,0.00083763537,0.008818131,0.24757783],"study_design_scores_gemma":[0.0022340952,0.00037759566,0.8559858,0.0033352033,0.00009371263,0.00001265426,0.00009009588,0.12263053,0.0011173425,0.013271718,0.0006685249,0.00018270916],"about_ca_topic_score_codex":0.0007064554,"about_ca_topic_score_gemma":0.000024739953,"teacher_disagreement_score":0.24739513,"about_ca_system_score_codex":0.000024783885,"about_ca_system_score_gemma":0.000055149798,"threshold_uncertainty_score":0.8056159},"labels":[],"label_agreement":null},{"id":"W1987173073","doi":"10.1037/h0087458","title":"From theory to application and back again: Implications of research on medical expertise for psychological theory.","year":2005,"lang":"en","type":"article","venue":"Canadian Journal of Experimental Psychology/Revue canadienne de psychologie expérimentale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Psychology; Perspective (graphical); Psychological research; Psychological Theory; Point (geometry); Cognition; Domain (mathematical analysis); Epistemology; Cognitive psychology; Social psychology; Psychiatry; Computer science","score_opus":0.11304143951316681,"score_gpt":0.46840117181639024,"score_spread":0.35535973230322343,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1987173073","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96873665,0.0025208972,0.0043675257,0.018101575,0.0006009353,0.00096891174,0.00015505769,0.0000166071,0.0045318217],"genre_scores_gemma":[0.97845125,0.00007484164,0.0083238995,0.011710771,0.0008853966,0.00019310936,0.000048100086,0.000048676055,0.00026398044],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9967441,0.00035107858,0.0009496819,0.00077273377,0.00020193156,0.0009804831],"domain_scores_gemma":[0.99187684,0.003586233,0.00026824072,0.00083294837,0.00027982093,0.0031558948],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0030495427,0.00027991988,0.00063181523,0.00059952505,0.00017940257,0.000026076536,0.0006575996,0.00045403742,0.0013063665],"category_scores_gemma":[0.006579455,0.00025565742,0.00023021098,0.00039493752,0.0008947366,0.00007069901,0.000040087354,0.0006420606,0.000088169145],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.011993635,0.0070827226,0.031677805,0.00005217686,0.0008023687,0.00062425656,0.023065194,0.000038935577,0.18719223,0.04170685,0.24461195,0.45115188],"study_design_scores_gemma":[0.04735482,0.049113154,0.4997954,0.007227193,0.00051477714,0.009058822,0.04127881,0.00020018454,0.0970702,0.10913882,0.13641807,0.0028297398],"about_ca_topic_score_codex":0.0008211436,"about_ca_topic_score_gemma":0.002179715,"teacher_disagreement_score":0.4681176,"about_ca_system_score_codex":0.0007248713,"about_ca_system_score_gemma":0.0003665888,"threshold_uncertainty_score":0.99998957},"labels":[],"label_agreement":null},{"id":"W1987222414","doi":"10.1051/pmed/2012015","title":"Aborder le raisonnement clinique du point de vue pédagogique","year":2012,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Humanities; Philosophy","score_opus":0.039356311083504326,"score_gpt":0.36381620913121865,"score_spread":0.3244598980477143,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1987222414","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3982676,0.20325738,0.039880488,0.23486461,0.017563542,0.0022304354,0.00022610986,0.0008738465,0.10283601],"genre_scores_gemma":[0.8442542,0.027145552,0.010718736,0.022635687,0.01601852,0.00027630504,0.0001402672,0.00022126322,0.07858948],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9939491,0.00076280936,0.0013598906,0.0007429188,0.00070184085,0.0024834105],"domain_scores_gemma":[0.97879654,0.016951386,0.00032422383,0.0010361396,0.00024884575,0.00264288],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.004763875,0.0006192605,0.001081337,0.00014070707,0.0002868684,0.00008299341,0.0004205311,0.0010297544,0.009857202],"category_scores_gemma":[0.07338615,0.00059604633,0.0005973869,0.00040923874,0.0008038801,0.00026800347,0.00033123075,0.0014482298,0.0030281793],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004271191,0.0137556745,0.22814932,0.00046647113,0.0005690587,0.00058021746,0.012068621,0.000069218026,0.00047969216,0.019369483,0.43802917,0.28603595],"study_design_scores_gemma":[0.005734955,0.00094307546,0.07072619,0.00467687,0.00066591153,0.0003728625,0.004562546,0.0008373698,0.0024607095,0.0022457237,0.9058205,0.000953298],"about_ca_topic_score_codex":0.0009767693,"about_ca_topic_score_gemma":0.00006615445,"teacher_disagreement_score":0.46779132,"about_ca_system_score_codex":0.00046045092,"about_ca_system_score_gemma":0.0017280233,"threshold_uncertainty_score":0.9996491},"labels":[],"label_agreement":null},{"id":"W1987462158","doi":"10.1080/01421590801953018","title":"Teaching and assessment of explanation and planning in medical schools in the United Kingdom: Cross sectional questionnaire survey","year":2009,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":28,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Summative assessment; Formative assessment; Medical education; Medicine; Cross-sectional study; Questionnaire; Medical school; Family medicine; Psychology; Pedagogy; Sociology","score_opus":0.08789489904866946,"score_gpt":0.46024739474790677,"score_spread":0.37235249569923734,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1987462158","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.992607,0.00014246862,0.00035934523,0.0060899267,0.000065709486,0.00013241138,0.0000010633973,0.000016631431,0.0005854481],"genre_scores_gemma":[0.998218,0.000029312954,0.00015497369,0.0013434541,0.0001502146,0.000008049678,0.000058610323,0.0000053123213,0.000032054508],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99742246,0.0007711025,0.00046761913,0.00021741136,0.00095388567,0.00016753047],"domain_scores_gemma":[0.99540794,0.004127786,0.000074034746,0.0001151415,0.000046832534,0.00022829243],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.009356361,0.00009702098,0.00024644364,0.00015526469,0.000052550993,0.000021611882,0.00008538987,0.00025679302,0.000098721895],"category_scores_gemma":[0.12324449,0.000064632586,0.00002291724,0.00021011327,0.0001342984,0.000045382934,0.000027387692,0.0015819928,5.897159e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000048159814,0.00032474787,0.98777086,0.00001150355,0.0000074930476,0.000107283944,0.0003607557,0.000006341979,0.0000013545933,0.0005712566,0.00018646802,0.010603768],"study_design_scores_gemma":[0.0020812012,0.00015268174,0.98050255,0.0019712932,0.000007650783,0.00008880515,0.00013182225,0.0146035,4.3240837e-7,0.0003101682,0.00008819243,0.00006170155],"about_ca_topic_score_codex":0.0027840934,"about_ca_topic_score_gemma":0.00016446445,"teacher_disagreement_score":0.11388812,"about_ca_system_score_codex":0.000048228754,"about_ca_system_score_gemma":0.00020524002,"threshold_uncertainty_score":0.8841408},"labels":[],"label_agreement":null},{"id":"W1987532297","doi":"10.1037/a0024461","title":"Teaching posttraining: Influencing diagnostic strategy with instructions at test.","year":2011,"lang":"en","type":"article","venue":"Journal of Experimental Psychology Applied","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Hamilton Health Sciences; McMaster University Medical Centre","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Similarity (geometry); Dual (grammatical number); Computer science; Test (biology); Natural language processing; Diagnostic accuracy; Medical diagnosis; Artificial intelligence; Cognitive psychology; Psychology; Machine learning; Data mining; Information retrieval; Medicine; Linguistics; Radiology","score_opus":0.058082809693292756,"score_gpt":0.37343723061223366,"score_spread":0.3153544209189409,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1987532297","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93433213,0.00030851754,0.000074421696,0.00015626708,0.00025577212,0.00016827264,0.0000019398065,0.000029400206,0.06467325],"genre_scores_gemma":[0.99167377,0.000020354157,0.0064247525,0.0014775645,0.00025067272,0.000013239153,0.0000033818612,0.000029718827,0.00010652492],"study_design_codex":"bench_or_experimental","study_design_gemma":"observational","domain_scores_codex":[0.9985827,0.000036963185,0.00062316895,0.00023886094,0.00024188399,0.00027640857],"domain_scores_gemma":[0.99652094,0.0024635585,0.00038538512,0.00023919916,0.00006954935,0.0003213641],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00033365918,0.0001930189,0.00045503976,0.00014883604,0.00013343054,0.000009571737,0.00013442143,0.00015310949,0.00046205372],"category_scores_gemma":[0.0023150297,0.0001423618,0.000116586074,0.00010908623,0.0002576915,0.00007328368,0.000036365844,0.0006452992,0.000041352316],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008189711,0.0054959515,0.34720862,0.000027953021,0.0009478742,0.0026791824,0.01631385,0.00001838683,0.5881725,0.004974998,0.0057233553,0.020247586],"study_design_scores_gemma":[0.024084948,0.020325525,0.8127215,0.0017645215,0.0006450329,0.02609385,0.011727517,0.000009914929,0.09950074,0.0012895578,0.0010719249,0.0007649501],"about_ca_topic_score_codex":0.00001850861,"about_ca_topic_score_gemma":0.0000016538083,"teacher_disagreement_score":0.48867178,"about_ca_system_score_codex":0.00009714447,"about_ca_system_score_gemma":0.00009012557,"threshold_uncertainty_score":0.5805346},"labels":[],"label_agreement":null},{"id":"W1987548299","doi":"10.1007/s10459-012-9370-3","title":"Exploring the impact of mental workload on rater-based assessments","year":2012,"lang":"en","type":"review","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":112,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia; Centennial College; University of Toronto","funders":"","keywords":"Workload; Mental health; Psychology; Inter-rater reliability; Medical education; Applied psychology; Medicine; Computer science; Psychiatry; Rating scale","score_opus":0.28063833374535374,"score_gpt":0.5869026846186868,"score_spread":0.30626435087333304,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1987548299","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0035587149,0.9923993,0.000008076332,0.00024718032,0.0020909605,0.0009789107,0.0000072365574,0.000010199958,0.00069942203],"genre_scores_gemma":[0.015798293,0.98283315,0.0005477638,0.00017641574,0.0002530706,0.0003087633,0.000040862447,0.000010864532,0.00003082538],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99790365,0.00024483708,0.00078108296,0.0003160536,0.00039623852,0.00035817194],"domain_scores_gemma":[0.9927437,0.0060477206,0.0006835137,0.00030224276,0.000040330335,0.00018250551],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001791684,0.00019930438,0.0007828951,0.00023678472,0.0001232013,0.000016691547,0.00022004038,0.000049480313,0.000036431942],"category_scores_gemma":[0.0024291493,0.00010546265,0.000205547,0.00090123527,0.00021932287,0.00022976633,0.000019220204,0.00029934122,0.000016958125],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000009975186,0.00045541077,0.0044772183,0.0015003555,0.000005155985,1.1677358e-7,0.00006645046,0.00004872527,1.2112271e-8,0.000047543508,0.000043853677,0.9933452],"study_design_scores_gemma":[0.0007045152,0.004113953,0.026186634,0.50871384,0.00020453641,0.000022456396,0.00055120914,0.0001270851,0.000002784891,0.0002263774,0.4586933,0.00045332633],"about_ca_topic_score_codex":0.00040201886,"about_ca_topic_score_gemma":0.000011634329,"teacher_disagreement_score":0.99289185,"about_ca_system_score_codex":0.00055670284,"about_ca_system_score_gemma":0.0045702714,"threshold_uncertainty_score":0.8107465},"labels":[],"label_agreement":null},{"id":"W1988427661","doi":"10.1067/mem.2002.121919","title":"Brain cramp: The emergency physician's worst nightmare","year":2002,"lang":"en","type":"article","venue":"Annals of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Thunder Bay Regional Research Institute; Thunder Bay Regional Health Sciences Centre","funders":"","keywords":"Medicine; Nightmare; Emergency physician; Emergency department; Medical emergency; Emergency medicine; Physical medicine and rehabilitation; Psychiatry","score_opus":0.17139497098176965,"score_gpt":0.4236192777123412,"score_spread":0.25222430673057156,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1988427661","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.39360785,0.025764732,0.0001473122,0.38770092,0.0057919817,0.000876352,0.000032897377,0.00016717493,0.1859108],"genre_scores_gemma":[0.93709964,0.01595646,0.000032914042,0.007260163,0.0027657899,0.000033642922,0.00005774514,0.000053120286,0.036740504],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9970899,0.000115024144,0.0011488287,0.0003945141,0.0007442918,0.00050741836],"domain_scores_gemma":[0.9959262,0.0019698548,0.00036215308,0.0008761872,0.00051939033,0.0003461918],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0008046279,0.0002918482,0.0006864311,0.000113515016,0.00013606294,0.0000015825888,0.0003163071,0.00013087569,0.034781363],"category_scores_gemma":[0.02854514,0.00016986526,0.00040599648,0.00082313886,0.00024029425,0.00006114874,0.000060450257,0.00035604753,0.00029439924],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000060287635,0.0005133465,0.010356421,0.000051566956,0.00016613714,0.000016040714,0.00088234636,0.0000036210936,0.00049095176,0.0011492766,0.97186196,0.014448035],"study_design_scores_gemma":[0.00428024,0.0051374785,0.20603023,0.004646281,0.00090512406,0.000030555082,0.0018059281,0.0011672508,0.0018318214,0.013146968,0.7600297,0.0009884218],"about_ca_topic_score_codex":0.00017362804,"about_ca_topic_score_gemma":0.000022577702,"teacher_disagreement_score":0.5434918,"about_ca_system_score_codex":0.000008258313,"about_ca_system_score_gemma":0.000023688885,"threshold_uncertainty_score":0.97963786},"labels":[],"label_agreement":null},{"id":"W1989196101","doi":"10.1017/s0317167100008921","title":"Diagnostic Accuracy of Neurological Problems in the Emergency Department","year":2008,"lang":"en","type":"article","venue":"Canadian Journal of Neurological Sciences / Journal Canadien des Sciences Neurologiques","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":104,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Queen Elizabeth II Health Sciences Centre; Dalhousie University","funders":"","keywords":"Medicine; Emergency department; Psychogenic disease; Neurology; Vertigo; Stroke (engine); Migraine; Pediatrics; Neurological examination; Medical diagnosis; Neurological disorder; Emergency medicine; Central nervous system disease; Surgery; Psychiatry","score_opus":0.06489627169033615,"score_gpt":0.32505335037404437,"score_spread":0.2601570786837082,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1989196101","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9863357,0.0019690336,0.000012841281,0.0090623945,0.0006931093,0.00034776964,0.000008780145,0.0000104320125,0.0015599658],"genre_scores_gemma":[0.9924919,0.0027746747,0.0005248168,0.003928537,0.00025094423,0.0000079295305,5.9739585e-7,0.000008843673,0.000011745002],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9943263,0.0012595532,0.0015494784,0.0005019847,0.0011322527,0.0012304552],"domain_scores_gemma":[0.98321426,0.013515223,0.001026378,0.00024117235,0.00045942987,0.0015435447],"candidate_categories":["metaresearch","sts"],"consensus_categories":["sts"],"category_scores_codex":[0.0053318967,0.00031693897,0.00072647654,0.00066669355,0.0013032542,0.00011177201,0.0017985889,0.00021162585,0.0002548544],"category_scores_gemma":[0.098184876,0.000169214,0.00038010153,0.0017491616,0.0063738553,0.00040297303,0.00006464816,0.0015799764,0.000004547329],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000072393144,0.00016269491,0.983444,0.000006852356,0.000009308527,0.011088194,0.0004197976,0.0017999756,0.000007155131,0.00027138088,0.0012296896,0.0014885612],"study_design_scores_gemma":[0.0004772562,0.06436912,0.8819467,0.00011797447,0.000049328155,0.045567784,0.00009908717,0.00030324265,0.000014704192,0.005896367,0.000993563,0.00016487205],"about_ca_topic_score_codex":0.00101364,"about_ca_topic_score_gemma":0.016652146,"teacher_disagreement_score":0.1014973,"about_ca_system_score_codex":0.00008362437,"about_ca_system_score_gemma":0.0029745295,"threshold_uncertainty_score":0.9999969},"labels":[],"label_agreement":null},{"id":"W1989524138","doi":"10.1016/j.annemergmed.2012.04.013","title":"Mapping Out the Emergency Department Disposition Decision for High-Acuity Patients","year":2012,"lang":"en","type":"article","venue":"Annals of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":84,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; Ottawa Hospital","funders":"","keywords":"Medicine; Emergency department; Disposition; Medical emergency; Visual acuity; Emergency medicine; Optometry; Ophthalmology; Nursing","score_opus":0.14911414376416077,"score_gpt":0.44103702463401695,"score_spread":0.2919228808698562,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1989524138","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9762999,0.0031480817,0.003856641,0.0072070044,0.0071377982,0.00095331966,0.00004461089,0.00003116854,0.0013214904],"genre_scores_gemma":[0.9926151,0.003998978,0.0002406478,0.00051043346,0.0016185518,0.00009482521,0.00032913158,0.000023517805,0.00056878675],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9976064,0.00005720077,0.0010651853,0.00024104354,0.00057501247,0.00045513545],"domain_scores_gemma":[0.99641305,0.001833319,0.00035456114,0.00045012266,0.0006101065,0.00033881338],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012189017,0.00020225617,0.00047009997,0.00008574309,0.00013113454,9.4284826e-7,0.00014063602,0.000100840625,0.0018540026],"category_scores_gemma":[0.027715625,0.00011805997,0.00024559538,0.00027351858,0.00007272278,0.00008432069,0.000050225397,0.00012747195,0.000035300425],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005582556,0.0016363519,0.5157143,0.00010981481,0.00022878601,8.3330013e-7,0.001387796,0.0000025531506,0.00039798912,0.0010148763,0.42269558,0.056252867],"study_design_scores_gemma":[0.001692657,0.001403751,0.97302693,0.0010392076,0.000191436,8.886907e-7,0.00020051734,0.000028509563,0.0007657584,0.0030253204,0.018469542,0.00015549996],"about_ca_topic_score_codex":0.00007092656,"about_ca_topic_score_gemma":0.00001231556,"teacher_disagreement_score":0.4573126,"about_ca_system_score_codex":0.000016382415,"about_ca_system_score_gemma":0.000026163952,"threshold_uncertainty_score":0.9990584},"labels":[],"label_agreement":null},{"id":"W1989691238","doi":"10.1097/00001888-200210001-00002","title":"The Aging Physician","year":2002,"lang":"en","type":"review","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":190,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"","keywords":"Population ageing; Context (archaeology); Gerontology; Health care; Developed country; Population; Cohort; Medicine; Geriatrics; Psychology; Cohort effect; Family medicine; Demography; Political science; Geography; Environmental health; Psychiatry; Sociology","score_opus":0.12676009663902119,"score_gpt":0.4540729888525732,"score_spread":0.327312892213552,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1989691238","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000015772395,0.97546315,0.000015301957,0.0049249767,0.001065028,0.00072545744,0.0000042495944,0.00012055514,0.017679712],"genre_scores_gemma":[0.000008604696,0.9708418,0.000010515346,0.0032314905,0.006387855,0.00008697759,0.00009130294,0.0000830254,0.019258408],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99675786,0.00016716772,0.0012802428,0.0005304693,0.00068218267,0.0005820612],"domain_scores_gemma":[0.96795005,0.030203218,0.00054392405,0.00080236106,0.000073103,0.0004273347],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001182662,0.0004958271,0.0025406147,0.00012224884,0.00019290131,0.000008508651,0.00046253589,0.0006979348,0.0002073008],"category_scores_gemma":[0.04707597,0.00023380268,0.00042389956,0.0005310581,0.000536653,0.000022361985,0.00010391566,0.0031162656,0.0008704982],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000024955732,0.0000063686875,0.000005427025,0.0012570594,0.00012934925,0.000046402798,0.0000855175,1.24986625e-8,5.163791e-8,0.00043252853,0.13247997,0.8655548],"study_design_scores_gemma":[0.00046789495,0.00011859477,0.000004574501,0.17925316,0.0025122056,0.00018431169,0.000060601174,0.0000062979275,1.0281067e-7,0.00020316459,0.8170174,0.00017174013],"about_ca_topic_score_codex":0.000015957712,"about_ca_topic_score_gemma":3.715702e-7,"teacher_disagreement_score":0.8653831,"about_ca_system_score_codex":0.00011989626,"about_ca_system_score_gemma":0.00019315988,"threshold_uncertainty_score":0.99990743},"labels":[],"label_agreement":null},{"id":"W1989953251","doi":"10.1097/acm.0000000000000074","title":"Am I Right When I Am Sure? Data Consistency Influences the Relationship Between Diagnostic Accuracy and Certainty","year":2013,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":32,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Western Hospital","funders":"","keywords":"Certainty; Medical diagnosis; Diagnostic accuracy; Consistency (knowledge bases); Context (archaeology); Medicine; Statistics; Psychology; Computer science; Mathematics; Artificial intelligence; Internal medicine; Pathology","score_opus":0.13436026941324453,"score_gpt":0.3999045361787397,"score_spread":0.2655442667654952,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1989953251","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8841451,0.0073168855,0.00009892715,0.10534923,0.0002811499,0.0009917313,0.000028073926,0.000099765304,0.0016891869],"genre_scores_gemma":[0.98900497,0.0018280792,0.000106250074,0.0060929563,0.0014577046,0.00006431521,0.00034498642,0.000024113544,0.0010766436],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9971551,0.00024269646,0.00095823687,0.0005932111,0.00061937724,0.00043141525],"domain_scores_gemma":[0.6683883,0.32852527,0.00044373857,0.0016448552,0.0002369587,0.0007608798],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0019460999,0.0002710202,0.00065280887,0.000097865035,0.0002547889,0.000022619954,0.00061643193,0.0003676876,0.000778344],"category_scores_gemma":[0.55790836,0.00014545141,0.00004225283,0.00027900047,0.0012163521,0.00026875382,0.00029643674,0.0014869272,0.00024993613],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000011296829,0.000010894532,0.88984907,0.000038849215,0.000057300775,0.000013695038,0.00081829657,3.259509e-7,0.000014382974,0.0012574029,0.10250824,0.0054202564],"study_design_scores_gemma":[0.0016765828,0.00020611902,0.9610657,0.0031476181,0.0006540449,0.00008563787,0.0006607139,0.00007032809,0.000009139487,0.023927791,0.008325759,0.00017056701],"about_ca_topic_score_codex":0.0014362738,"about_ca_topic_score_gemma":0.000016156813,"teacher_disagreement_score":0.55596226,"about_ca_system_score_codex":0.00003174388,"about_ca_system_score_gemma":0.00017214505,"threshold_uncertainty_score":0.8522321},"labels":[],"label_agreement":null},{"id":"W1989990663","doi":"10.1007/s11606-012-2097-2","title":"Judging Whether a Patient is Actually Improving: More Pitfalls from the Science of Human Perception","year":2012,"lang":"en","type":"review","venue":"Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute for Clinical Evaluative Sciences; University of Toronto","funders":"Canadian Institutes of Health Research; Sunnybrook Research Institute","keywords":"Medicine; Overconfidence effect; Recall; Conformist; Comprehension; Perception; Blinding; Anxiety; Cognitive psychology; Social psychology; Psychology; Psychiatry; Clinical trial; Pathology","score_opus":0.08853590523835074,"score_gpt":0.4210995735265405,"score_spread":0.33256366828818973,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1989990663","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.066358544,0.92919326,0.0001275001,0.0014436418,0.0018832184,0.00035601744,0.000020701273,0.000008587113,0.0006085451],"genre_scores_gemma":[0.04157169,0.94711864,0.0007995273,0.0018280117,0.0077293054,0.000006590215,0.000020454676,0.00006648412,0.00085929164],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9954516,0.00015228087,0.0021835254,0.00030514417,0.0015318014,0.0003756157],"domain_scores_gemma":[0.9926751,0.0027341272,0.0028048952,0.00052011875,0.00076097396,0.0005047846],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0018745457,0.00040042726,0.002252768,0.0003388561,0.000094864896,0.000027739716,0.00068033446,0.00022393056,0.0009609828],"category_scores_gemma":[0.010267744,0.00018709611,0.00075345015,0.00031077297,0.0008588998,0.000120890654,0.00018024845,0.0012863987,0.000015106124],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007850584,0.00022632917,0.00028894577,0.0005399457,0.00034963468,0.000069265356,0.0028994742,3.7636678e-7,0.0020756607,0.000014791945,0.0071011535,0.9863559],"study_design_scores_gemma":[0.00514462,0.011286883,0.007480989,0.53683937,0.019487834,0.0037295583,0.0022318242,0.00011293173,0.0004544208,0.00024883624,0.4120639,0.000918813],"about_ca_topic_score_codex":0.0008757367,"about_ca_topic_score_gemma":0.000004107169,"teacher_disagreement_score":0.9854371,"about_ca_system_score_codex":0.0003060579,"about_ca_system_score_gemma":0.0005915521,"threshold_uncertainty_score":0.99995226},"labels":[],"label_agreement":null},{"id":"W1990245305","doi":"10.1007/s10459-011-9320-5","title":"Feasibility of self-reflection as a tool to balance clinical reasoning strategies","year":2011,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":37,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; University Health Network","funders":"","keywords":"Balance (ability); Reflection (computer programming); Self-reflection; Psychology; Computer science; Medicine; Physical therapy","score_opus":0.08900560466680295,"score_gpt":0.5142230913752249,"score_spread":0.42521748670842197,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1990245305","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98246926,0.0016843077,0.00039704642,0.0010469321,0.0012731606,0.0006070545,9.414139e-7,0.000046352034,0.012474959],"genre_scores_gemma":[0.94203275,0.0018198829,0.054733872,0.0011892371,0.00012646089,0.000043383683,0.0000026878488,0.0000043915047,0.00004733111],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99806094,0.00013848077,0.00082208606,0.00043016436,0.000281778,0.00026654618],"domain_scores_gemma":[0.99736845,0.0016818979,0.00031680032,0.00025599933,0.00016121649,0.00021560967],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0028335792,0.000090506546,0.00030805822,0.00012440853,0.00009005892,0.000012171098,0.00012842282,0.000058579048,0.000048384547],"category_scores_gemma":[0.01908971,0.000076546465,0.00004912434,0.0008089282,0.00020183907,0.00035119068,0.000021199496,0.00016443654,0.000018405168],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011643604,0.0008664113,0.92157996,0.00009802393,0.0000018771995,4.2033403e-7,0.0012035544,0.000036715566,0.0000044150183,0.0074741864,0.00010785565,0.06851012],"study_design_scores_gemma":[0.0004208586,0.002237065,0.97994244,0.0021183605,0.0000085934125,0.000014464449,0.0022199473,0.00021019457,0.000068198286,0.011322109,0.0013313708,0.000106378335],"about_ca_topic_score_codex":0.0011504083,"about_ca_topic_score_gemma":0.00019252226,"teacher_disagreement_score":0.06840374,"about_ca_system_score_codex":0.00013337084,"about_ca_system_score_gemma":0.0030981528,"threshold_uncertainty_score":0.98917294},"labels":[],"label_agreement":null},{"id":"W1991122888","doi":"10.1051/pmed:2004019","title":"Un programme pour aider les étudiants en médecine àdévelopper leurs stratégies d'apprentissage","year":2004,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université de Sherbrooke","funders":"","keywords":"Humanities; Political science; Philosophy","score_opus":0.09277218361032437,"score_gpt":0.3633622371688279,"score_spread":0.2705900535585035,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1991122888","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.74452716,0.08798574,0.0039022707,0.15161262,0.004191066,0.0012120849,0.00018024333,0.000555953,0.0058328467],"genre_scores_gemma":[0.9355837,0.015789622,0.013022562,0.0013795241,0.0036032356,0.00013021984,0.00018564792,0.00018647923,0.030118976],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99423766,0.0002711994,0.0012469393,0.0011470275,0.0010760508,0.0020211271],"domain_scores_gemma":[0.98862654,0.008408684,0.00026063647,0.00093202293,0.00024001951,0.001532114],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0011898344,0.00075491954,0.0011884506,0.00017298825,0.0002923,0.00017466726,0.0005355711,0.00089669763,0.0044867396],"category_scores_gemma":[0.040404953,0.0006782762,0.0005393793,0.00064242084,0.0010382673,0.0002401619,0.00032189695,0.0013710844,0.0026317686],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00019619113,0.0068803853,0.17644446,0.00029009988,0.0006054323,0.0020937622,0.00084127503,0.00016677876,0.00042847497,0.004633276,0.01825294,0.7891669],"study_design_scores_gemma":[0.035967212,0.0027801497,0.8103938,0.021511126,0.0024975387,0.0008577871,0.010491859,0.00019163146,0.0063407775,0.022337746,0.08428365,0.002346725],"about_ca_topic_score_codex":0.00041325105,"about_ca_topic_score_gemma":0.00007578841,"teacher_disagreement_score":0.78682023,"about_ca_system_score_codex":0.00039590927,"about_ca_system_score_gemma":0.0013247658,"threshold_uncertainty_score":0.99956685},"labels":[],"label_agreement":null},{"id":"W1992106100","doi":"10.1186/s12909-015-0308-3","title":"Example-based learning: comparing the effects of additionally providing three different integrative learning activities on physiotherapy intervention knowledge","year":2015,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":33,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"Université de Montréal","keywords":"Concept map; Intervention (counseling); Test (biology); Process (computing); Mathematics education; Psychology; Computer science; Procedural knowledge; Artificial intelligence; Domain knowledge","score_opus":0.05075087183120075,"score_gpt":0.3924551387216569,"score_spread":0.34170426689045613,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1992106100","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9860037,0.00041505115,0.01053041,0.00046337058,0.00074178807,0.00047634967,4.8951335e-7,0.00006533756,0.0013035202],"genre_scores_gemma":[0.99815035,0.000025307863,0.00022862968,0.00020920375,0.0004691957,0.0001902113,0.00014969152,0.000019429379,0.00055798],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9981502,0.0005287032,0.00033260576,0.00023618987,0.0005937903,0.00015850269],"domain_scores_gemma":[0.9747312,0.024362996,0.00024119447,0.00017927666,0.00020578605,0.00027953606],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0008234748,0.00016073948,0.00034589675,0.00009176659,0.00008893631,0.000021391783,0.0001131398,0.000104915984,0.00014181504],"category_scores_gemma":[0.11323373,0.00009672888,0.0001622311,0.00015491404,0.00016473206,0.000042384872,0.000027741371,0.0005660024,0.000015575744],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0031908148,0.013890402,0.57853884,0.0016750798,0.00027768675,0.0000038297785,0.007835874,0.00046717902,0.0009637972,0.00973197,0.012609713,0.37081483],"study_design_scores_gemma":[0.013881896,0.009226228,0.8320217,0.067358464,0.0003703419,0.000016488479,0.0066043288,0.051632136,0.012079024,0.0025615483,0.0037052587,0.0005425885],"about_ca_topic_score_codex":0.00016049609,"about_ca_topic_score_gemma":0.00017902377,"teacher_disagreement_score":0.37027222,"about_ca_system_score_codex":0.00016726824,"about_ca_system_score_gemma":0.0015083912,"threshold_uncertainty_score":0.8942359},"labels":[],"label_agreement":null},{"id":"W1992937121","doi":"10.1097/00001888-200010001-00026","title":"The Under-weighting of Implicitly Generated Diagnoses","year":2000,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medical diagnosis; Weighting; Psychology; Isolation (microbiology); Medicine; Cognitive psychology; Computer science; Pathology; Radiology","score_opus":0.041614375988241345,"score_gpt":0.369922601733617,"score_spread":0.32830822574537566,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1992937121","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95901567,0.007517068,0.00006546974,0.020520825,0.00022976764,0.00024824584,0.0000028820753,0.00007090762,0.012329192],"genre_scores_gemma":[0.97441566,0.014752299,0.000050567767,0.0043842345,0.0010352081,0.000018381408,0.000024752066,0.000022520528,0.005296375],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99814504,0.000071840725,0.0007516872,0.00025173475,0.00043223653,0.00034743678],"domain_scores_gemma":[0.983429,0.015678305,0.0001560571,0.00037082288,0.00010950446,0.00025631738],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0008761157,0.00016085415,0.00045875914,0.000047131398,0.00012614763,0.0000031036473,0.00019503919,0.0002137201,0.001218135],"category_scores_gemma":[0.023997124,0.00008594149,0.000068964684,0.0003427889,0.00045694734,0.000026061629,0.000026077278,0.0006206122,0.00008925426],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007730108,0.00013442949,0.08586167,0.00008771618,0.00041030653,0.0001036323,0.002299188,0.00010300232,0.020806346,0.013571735,0.26797053,0.60787845],"study_design_scores_gemma":[0.026960205,0.005833964,0.5341733,0.025943572,0.002577636,0.001246502,0.005455961,0.003079247,0.03995696,0.033070903,0.3203701,0.0013316197],"about_ca_topic_score_codex":0.00018733717,"about_ca_topic_score_gemma":0.000004165439,"teacher_disagreement_score":0.6065468,"about_ca_system_score_codex":0.0000284716,"about_ca_system_score_gemma":0.00010310339,"threshold_uncertainty_score":0.9996949},"labels":[],"label_agreement":null},{"id":"W1993283243","doi":"10.1016/s0140-6736(08)61500-7","title":"How doctors feel: affective issues in patients' safety","year":2008,"lang":"en","type":"article","venue":"The Lancet","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":115,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Scopus; Patient safety; Psychology; Neglect; Medicine; Cognition; Alternative medicine; MEDLINE; Medical education; Health care; Psychiatry; Political science; Pathology; Law","score_opus":0.030335783140816044,"score_gpt":0.32059711972808724,"score_spread":0.2902613365872712,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1993283243","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9788799,0.00024278753,0.000009078946,0.015939787,0.00015650506,0.00021394635,0.00001057951,0.000048985174,0.0044984203],"genre_scores_gemma":[0.99601966,0.00041523192,0.00012384853,0.0012661907,0.0005930197,0.000008155714,0.000018627023,0.000009280956,0.0015459976],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9993267,0.000064202955,0.000111629655,0.0001338321,0.00016705459,0.00019658514],"domain_scores_gemma":[0.9967832,0.0027671445,0.00003744631,0.00031502155,0.000038188264,0.00005903141],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00021013056,0.000079297875,0.00034377183,0.000021528298,0.000047187386,0.000008664434,0.000089538946,0.00004888917,0.00003892033],"category_scores_gemma":[0.015504145,0.000045589986,0.000050484334,0.00012728514,0.00009800219,0.000025960671,0.000029038472,0.0001974943,0.000051557967],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003879482,0.00017384323,0.9613621,0.0000056633858,0.000018045674,0.000023982964,0.00060819456,0.000002923644,0.0000041463154,0.000246698,0.03258973,0.0045767007],"study_design_scores_gemma":[0.0030275348,0.00022086933,0.9807743,0.0002546838,0.000015326197,0.000005234584,0.000045710378,0.000010278929,0.00009668107,0.00019556114,0.015295154,0.000058703226],"about_ca_topic_score_codex":0.00012249863,"about_ca_topic_score_gemma":0.000019811314,"teacher_disagreement_score":0.01941214,"about_ca_system_score_codex":0.000036289635,"about_ca_system_score_gemma":0.000023558357,"threshold_uncertainty_score":0.9927887},"labels":[],"label_agreement":null},{"id":"W1993548098","doi":"10.1515/dx-2013-0028","title":"Bias: a normal operating characteristic of the diagnosing brain","year":2014,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":47,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Debiasing; Confirmation bias; Psychology; Variety (cybernetics); Process (computing); Critical thinking; Cognitive psychology; Social psychology; Computer science; Artificial intelligence; Mathematics education","score_opus":0.03206950935699471,"score_gpt":0.31162862316744133,"score_spread":0.2795591138104466,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1993548098","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98828995,0.00011644571,0.00010562411,0.008925343,0.0003788542,0.00026802215,0.000014069797,0.000043766533,0.0018579451],"genre_scores_gemma":[0.99350166,0.00007494392,0.00037652726,0.0053872042,0.00034442559,0.00010464225,0.000014134742,0.000023606754,0.00017285171],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9986087,0.00014626855,0.00044985325,0.00024054086,0.00029503595,0.0002595742],"domain_scores_gemma":[0.971768,0.027260013,0.00019244895,0.0004978138,0.00011708671,0.00016460865],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0006416512,0.00014567084,0.0003874029,0.00004743757,0.00011092338,0.00002824345,0.00015662456,0.00009546862,0.00021951858],"category_scores_gemma":[0.16329853,0.00009574104,0.00021355432,0.0002492883,0.00012849671,0.000045534496,0.00009878011,0.00021717882,0.00004635378],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000013080226,0.00031550147,0.95578754,0.00004248349,0.000029966544,0.0000060581106,0.00026817125,0.000013717365,0.00036937682,0.00035426093,0.0055777365,0.037222087],"study_design_scores_gemma":[0.0010829397,0.0002814787,0.97992766,0.0032692382,0.0001372612,0.000019116293,0.000055292076,0.0007424928,0.009469577,0.000069492344,0.0048027406,0.00014270886],"about_ca_topic_score_codex":0.00015496253,"about_ca_topic_score_gemma":0.000014201769,"teacher_disagreement_score":0.16265689,"about_ca_system_score_codex":0.000026001391,"about_ca_system_score_gemma":0.00006461285,"threshold_uncertainty_score":0.84374934},"labels":[],"label_agreement":null},{"id":"W1994121405","doi":"10.1007/s10067-004-1051-z","title":"A brief history of medical taxonomy and diagnosis","year":2006,"lang":"en","type":"article","venue":"Clinical Rheumatology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Rheumatology; Taxonomy (biology); Internal medicine; Medical physics; Zoology","score_opus":0.05980382989357903,"score_gpt":0.3582221679300058,"score_spread":0.29841833803642676,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1994121405","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9756679,0.006338679,0.00028689046,0.0073855817,0.0010568739,0.00025810688,0.0000044600697,0.00006979058,0.00893175],"genre_scores_gemma":[0.9925424,0.0039014677,0.0010605762,0.0015433588,0.00011411531,0.00007729592,0.000019141193,0.000017760325,0.00072389515],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9975452,0.00015381268,0.0013167555,0.00039473947,0.00033447452,0.00025505596],"domain_scores_gemma":[0.9657061,0.033177264,0.00026593878,0.00034193715,0.00009473061,0.00041402993],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012024974,0.00013652514,0.0010153692,0.00006310645,0.000015776945,0.0000028872835,0.00013225063,0.0005640546,0.0011206298],"category_scores_gemma":[0.09308584,0.000114167706,0.00021964076,0.000068713845,0.0012831311,0.000019902467,0.0001159107,0.0004332793,0.00005040233],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010523986,0.00075425813,0.821957,0.00004321904,0.000040853967,0.00008606222,0.00000934599,7.933262e-8,5.2984853e-7,0.010834043,0.13635287,0.029816488],"study_design_scores_gemma":[0.0061130757,0.00077332597,0.5736393,0.001545107,0.00016195649,0.0005591304,0.000019938636,0.000249255,0.000010248656,0.0014303176,0.41531175,0.00018662128],"about_ca_topic_score_codex":0.0010887611,"about_ca_topic_score_gemma":0.00006571734,"teacher_disagreement_score":0.2789589,"about_ca_system_score_codex":0.000046001835,"about_ca_system_score_gemma":0.0006153783,"threshold_uncertainty_score":0.99979246},"labels":[],"label_agreement":null},{"id":"W1995635137","doi":"10.1007/s10459-011-9331-2","title":"Expanding the basic science debate: the role of physics knowledge in interpreting clinical findings","year":2011,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":51,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; University of Toronto; MacEwan University; Western University","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"Science education; MEDLINE; Mathematics education; Psychology; Medicine; Medical education; Engineering ethics; Political science; Engineering","score_opus":0.05200758058814899,"score_gpt":0.46930268499525407,"score_spread":0.4172951044071051,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1995635137","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9811145,0.006480559,0.000099804936,0.0010080233,0.0011116213,0.00034944425,3.8159948e-7,0.000009142761,0.009826531],"genre_scores_gemma":[0.996171,0.0016208023,0.0013984275,0.0006411404,0.00011008908,0.000034484437,5.7616563e-7,0.0000036018482,0.000019864758],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.998299,0.0001476788,0.0006673345,0.00031215267,0.00026292837,0.00031087946],"domain_scores_gemma":[0.9919629,0.007314248,0.0002736285,0.0002480336,0.00009930921,0.000101852864],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.007012032,0.000074803844,0.00020196359,0.00011193891,0.00023640016,0.000015649428,0.00039817358,0.0000287698,0.000015050527],"category_scores_gemma":[0.01864856,0.00004169046,0.000038810704,0.0015357095,0.001656327,0.00028519888,0.00006914673,0.0002636731,0.000008154226],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000015639798,0.00026770702,0.6845488,0.00001968054,4.70927e-7,9.187439e-8,0.00583404,0.000016677004,0.000011215114,0.0048640366,0.000015242339,0.30440643],"study_design_scores_gemma":[0.0004901475,0.0008254249,0.92970645,0.0070855464,0.000010379404,0.0000133002795,0.025496654,0.0043043816,0.0020905049,0.02848644,0.0013394455,0.00015129693],"about_ca_topic_score_codex":0.00046954592,"about_ca_topic_score_gemma":0.00010689868,"teacher_disagreement_score":0.30425513,"about_ca_system_score_codex":0.00010819152,"about_ca_system_score_gemma":0.0020197649,"threshold_uncertainty_score":0.98961776},"labels":[],"label_agreement":null},{"id":"W1995786180","doi":"10.1111/jan.12379","title":"A concept analysis of abductive reasoning","year":2014,"lang":"en","type":"article","venue":"Journal of Advanced Nursing","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":57,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Regina; McMaster University","funders":"","keywords":"Abductive reasoning; Deductive reasoning; Inference; Formal concept analysis; Psychology; Epistemology; Computer science; Artificial intelligence","score_opus":0.012837326623144825,"score_gpt":0.36271558557828715,"score_spread":0.3498782589551423,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1995786180","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96881485,0.0008998816,0.027124252,0.0007406944,0.0004232595,0.000055566314,0.0000021871874,0.00000924209,0.0019300515],"genre_scores_gemma":[0.9674297,0.000050808663,0.03207139,0.00012476972,0.00025658746,2.8539196e-7,0.0000028208253,0.000009467734,0.000054187203],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9988032,0.000055988658,0.00056787464,0.00010971468,0.0003160581,0.000147186],"domain_scores_gemma":[0.99524623,0.0033123451,0.00067265786,0.00015948288,0.0004321715,0.00017710199],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00043769565,0.00008655967,0.0007063907,0.0002950989,0.000029324608,0.0000059835916,0.0000630961,0.000054334076,0.000039083807],"category_scores_gemma":[0.02750661,0.00006820776,0.0003607638,0.0005918278,0.00013379072,0.000084436426,0.00000590828,0.0002306287,8.887997e-7],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002240061,0.0011781408,0.031184787,0.00002466328,0.0026051768,0.00007413151,0.0050634956,0.030401213,0.015439896,0.0033366333,0.0007837346,0.90766805],"study_design_scores_gemma":[0.017505491,0.010014353,0.814043,0.043517098,0.028673451,0.0009051418,0.011691337,0.018067354,0.041783888,0.01010533,0.0029613227,0.000732241],"about_ca_topic_score_codex":0.0000033491726,"about_ca_topic_score_gemma":4.0716495e-7,"teacher_disagreement_score":0.9069358,"about_ca_system_score_codex":0.00008144494,"about_ca_system_score_gemma":0.000059427424,"threshold_uncertainty_score":0.9806851},"labels":[],"label_agreement":null},{"id":"W1996181808","doi":"10.1197/j.aem.2005.04.024","title":"Clinical Decision Rules \"in the Real World\": How a Widely Disseminated Rule Is Used in Everyday Practice","year":2005,"lang":"en","type":"article","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":115,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Ottawa","funders":"","keywords":"Medicine; Decision rule; Family medicine; Actuarial science; Statistics","score_opus":0.07453567313301618,"score_gpt":0.4708320477331544,"score_spread":0.3962963746001382,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1996181808","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8612755,0.0046241516,0.000082218874,0.123341344,0.0009749928,0.00064478815,0.0000069534344,0.000067340916,0.008982724],"genre_scores_gemma":[0.945463,0.040477213,0.00083890185,0.006559307,0.0019998106,0.000061366074,0.00006463223,0.000042581298,0.0044931686],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99456257,0.00053385145,0.0023608669,0.0007400642,0.001158981,0.0006436552],"domain_scores_gemma":[0.96390975,0.034270626,0.0005201978,0.0007303738,0.00018876207,0.00038032245],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.008015345,0.00034737223,0.0009306111,0.0003559825,0.000068428235,0.0000024067028,0.00050939753,0.0005060181,0.001178028],"category_scores_gemma":[0.18389288,0.00021458394,0.0001960971,0.001431005,0.00027234718,0.00024313504,0.00008683441,0.0025696054,0.00019505952],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00056157174,0.00051845826,0.74598944,0.000017726185,0.000039528844,0.00012596678,0.002259492,0.0000035812536,0.00008378603,0.00025624782,0.21770975,0.03243446],"study_design_scores_gemma":[0.0052474537,0.0005111618,0.89529556,0.0047978493,0.00034388425,0.00005201003,0.0017607749,0.00059973693,0.00003422432,0.0013918021,0.0896756,0.00028991816],"about_ca_topic_score_codex":0.0006759393,"about_ca_topic_score_gemma":0.00055768935,"teacher_disagreement_score":0.17587753,"about_ca_system_score_codex":0.0001266639,"about_ca_system_score_gemma":0.00016939014,"threshold_uncertainty_score":0.99973506},"labels":[],"label_agreement":null},{"id":"W1996488277","doi":"10.1051/pmed/2012010","title":"Aborder le raisonnement clinique du point de vue pédagogique","year":2012,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Philosophy; Humanities","score_opus":0.039356311083504326,"score_gpt":0.36381620913121865,"score_spread":0.3244598980477143,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1996488277","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3982676,0.20325738,0.039880488,0.23486461,0.017563542,0.0022304354,0.00022610986,0.0008738465,0.10283601],"genre_scores_gemma":[0.8442542,0.027145552,0.010718736,0.022635687,0.01601852,0.00027630504,0.0001402672,0.00022126322,0.07858948],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9939491,0.00076280936,0.0013598906,0.0007429188,0.00070184085,0.0024834105],"domain_scores_gemma":[0.97879654,0.016951386,0.00032422383,0.0010361396,0.00024884575,0.00264288],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.004763875,0.0006192605,0.001081337,0.00014070707,0.0002868684,0.00008299341,0.0004205311,0.0010297544,0.009857202],"category_scores_gemma":[0.07338615,0.00059604633,0.0005973869,0.00040923874,0.0008038801,0.00026800347,0.00033123075,0.0014482298,0.0030281793],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004271191,0.0137556745,0.22814932,0.00046647113,0.0005690587,0.00058021746,0.012068621,0.000069218026,0.00047969216,0.019369483,0.43802917,0.28603595],"study_design_scores_gemma":[0.005734955,0.00094307546,0.07072619,0.00467687,0.00066591153,0.0003728625,0.004562546,0.0008373698,0.0024607095,0.0022457237,0.9058205,0.000953298],"about_ca_topic_score_codex":0.0009767693,"about_ca_topic_score_gemma":0.00006615445,"teacher_disagreement_score":0.46779132,"about_ca_system_score_codex":0.00046045092,"about_ca_system_score_gemma":0.0017280233,"threshold_uncertainty_score":0.9996491},"labels":[],"label_agreement":null},{"id":"W1996894586","doi":"10.4321/s1575-18132005000200003","title":"Lo que todo profesor necesita saber sobre el razonamiento clínico","year":2005,"lang":"es","type":"article","venue":"Educación Médica","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Computer science; Philosophy","score_opus":0.018852578767989212,"score_gpt":0.3495219426590169,"score_spread":0.3306693638910277,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1996894586","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7228032,0.014937545,0.000087173314,0.13382077,0.0062700235,0.0016700694,0.00014273821,0.0004055199,0.119862966],"genre_scores_gemma":[0.92795473,0.0022840535,0.0012774417,0.010900302,0.010635193,0.00011229921,0.0001823695,0.00015223566,0.046501372],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99462503,0.0002988218,0.0013025879,0.0012553982,0.0010719033,0.0014462458],"domain_scores_gemma":[0.98726666,0.008955533,0.0004058175,0.0015068315,0.0003277536,0.0015373772],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0011339836,0.0007313475,0.0011947288,0.0002495015,0.00032721963,0.00024359963,0.0006009383,0.00076686154,0.0077883117],"category_scores_gemma":[0.03866404,0.00064558815,0.00057121215,0.0006184469,0.00049205247,0.00023463281,0.00028421058,0.00134917,0.013328347],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005292004,0.0056462185,0.0740345,0.00024804144,0.00042260814,0.0001684425,0.0014179408,0.000018524785,0.00026200176,0.008125454,0.7194041,0.18972297],"study_design_scores_gemma":[0.0058135106,0.0010652897,0.15154755,0.007921428,0.0010845725,0.00013642709,0.0011254108,0.0004961045,0.0007325391,0.0007191126,0.82822865,0.0011294215],"about_ca_topic_score_codex":0.00033560023,"about_ca_topic_score_gemma":0.00004515887,"teacher_disagreement_score":0.20515154,"about_ca_system_score_codex":0.00035745715,"about_ca_system_score_gemma":0.0014905,"threshold_uncertainty_score":0.9995995},"labels":[],"label_agreement":null},{"id":"W1998095277","doi":"10.1016/s1607-551x(08)70131-1","title":"Current Trends in Developing Medical Students' Critical Thinking Abilities","year":2008,"lang":"en","type":"review","venue":"The Kaohsiung Journal of Medical Sciences","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":110,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Medicine; Presentation (obstetrics); Task (project management); Apprenticeship; Problem-based learning; Cognition; Cognitive psychology; Artificial intelligence; Medical education; Computer science; Psychology","score_opus":0.15949841274118146,"score_gpt":0.5292976113927271,"score_spread":0.36979919865154565,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1998095277","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0008289035,0.96548223,0.00011723581,0.030893944,0.00220203,0.00015121204,0.0000024091694,0.000018754277,0.00030330368],"genre_scores_gemma":[0.000425001,0.99641865,0.00035238772,0.00091996253,0.001817438,0.000009997179,0.000004102463,0.000020116542,0.000032336116],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9828203,0.00090943184,0.0026103058,0.00042982143,0.012523453,0.0007067197],"domain_scores_gemma":[0.9375499,0.059654865,0.0008447077,0.00037654166,0.00020089636,0.0013730767],"candidate_categories":["metaresearch","sts","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.018582124,0.00040584916,0.0025933818,0.0006150814,0.00034239903,0.0001416441,0.0034887197,0.0005959663,0.0010346101],"category_scores_gemma":[0.17881796,0.00018833212,0.00088438124,0.0014645513,0.003121888,0.000114439266,0.00059745193,0.003411364,0.000032651114],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000010776375,0.0003542683,0.0008896013,0.00068593095,0.00006410824,0.0011375953,0.0003537122,2.601535e-7,1.3195168e-9,0.002206352,0.004079601,0.9902178],"study_design_scores_gemma":[0.00096480694,0.0005056398,0.0010864711,0.28898203,0.00047630534,0.010441657,0.00019627418,0.0000310849,7.698955e-8,0.0013543667,0.6956611,0.00030014227],"about_ca_topic_score_codex":0.000040917137,"about_ca_topic_score_gemma":0.000022661367,"teacher_disagreement_score":0.98991764,"about_ca_system_score_codex":0.000268012,"about_ca_system_score_gemma":0.006777121,"threshold_uncertainty_score":0.9998786},"labels":[],"label_agreement":null},{"id":"W2001057515","doi":"10.1016/j.lpm.2009.05.016","title":"Campus numérique de neurochirurgie : mobiliser une discipline des sciences médicales autour d’un projet pédagogique fédérateur","year":2009,"lang":"fr","type":"article","venue":"La Presse Médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Specialty; Library science; Medical education; Medicine; Psychology; Computer science; Family medicine","score_opus":0.030298598718353095,"score_gpt":0.349618860018111,"score_spread":0.3193202612997579,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2001057515","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8552662,0.04326892,0.0039915442,0.049473796,0.0012074346,0.0012801514,0.00012381231,0.000563273,0.04482488],"genre_scores_gemma":[0.9483399,0.012125394,0.0063977726,0.0039970903,0.0021398864,0.00011389022,0.000048935155,0.000085083506,0.02675204],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99375296,0.0011937909,0.0011386375,0.0012565665,0.0009077397,0.0017503006],"domain_scores_gemma":[0.9835865,0.013469619,0.0003232509,0.00083124475,0.00020161795,0.0015877956],"candidate_categories":["metaresearch","metaepi_narrow","sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002923844,0.0006722675,0.0010386914,0.00016270847,0.00075264793,0.00029246853,0.00078334037,0.0008719918,0.0011499879],"category_scores_gemma":[0.040753525,0.00055498275,0.00040894328,0.0010837592,0.0048671826,0.00029899264,0.00028429014,0.0015191179,0.00022412495],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013557129,0.023677906,0.13714409,0.0012106675,0.00046247826,0.009954349,0.010672379,0.0049731233,0.006158584,0.054336872,0.075641036,0.6744128],"study_design_scores_gemma":[0.011704989,0.0091828955,0.5596571,0.030865116,0.0012503674,0.0029909064,0.003799463,0.03340263,0.013114201,0.20831625,0.12265202,0.0030641027],"about_ca_topic_score_codex":0.0003766687,"about_ca_topic_score_gemma":0.00008270249,"teacher_disagreement_score":0.6713487,"about_ca_system_score_codex":0.00014415674,"about_ca_system_score_gemma":0.001468157,"threshold_uncertainty_score":0.9997631},"labels":[],"label_agreement":null},{"id":"W2002197924","doi":"10.1007/s11606-009-0918-8","title":"Tips for Teachers of Evidence-based Medicine: Making Sense of Decision Analysis Using a Decision Tree","year":2009,"lang":"en","type":"article","venue":"Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Columbia College","funders":"","keywords":"Decision tree; Calculator; Medicine; Decision analysis; Relevance (law); Field (mathematics); Clinical decision making; Decision engineering; Process (computing); Medical education; Management science; Business decision mapping; Computer science; Artificial intelligence; Family medicine; Statistics; Mathematics","score_opus":0.17530255347463786,"score_gpt":0.47421303040823837,"score_spread":0.2989104769336005,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2002197924","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7799071,0.0029999947,0.21280125,0.0034373,0.00053148664,0.00016796526,0.0000032254418,0.0000059012,0.00014578571],"genre_scores_gemma":[0.94038963,0.0003990436,0.056303833,0.0012976999,0.0014453006,7.27559e-7,0.0000034886725,0.00001716558,0.00014310115],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9955626,0.000117816584,0.0024535996,0.00025217788,0.001355245,0.00025855724],"domain_scores_gemma":[0.9803154,0.015653536,0.0019450007,0.00035952663,0.0013789806,0.00034758073],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0032610225,0.00024878254,0.0018133417,0.0011825612,0.000034864057,0.0000061259793,0.00023600805,0.00016215409,0.00018838026],"category_scores_gemma":[0.09298326,0.0001371963,0.00073951413,0.0008367282,0.00026408318,0.000104437415,0.000024554622,0.0003812109,4.876299e-7],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.029809909,0.0010016315,0.07039654,0.00014466884,0.0023135322,0.0010126625,0.0014598308,0.009170463,0.12520689,0.000104562154,0.024126176,0.73525316],"study_design_scores_gemma":[0.04401379,0.058212504,0.3676746,0.2803905,0.030352691,0.0035192962,0.0013115961,0.18266027,0.024667382,0.0050442987,0.0014353513,0.000717722],"about_ca_topic_score_codex":0.00018443864,"about_ca_topic_score_gemma":0.000016834165,"teacher_disagreement_score":0.7345354,"about_ca_system_score_codex":0.00015869128,"about_ca_system_score_gemma":0.00022321403,"threshold_uncertainty_score":0.91465694},"labels":[],"label_agreement":null},{"id":"W2003211637","doi":"10.1097/acm.0b013e3181e1b229","title":"To Think Is Good: Querying an Initial Hypothesis Reduces Diagnostic Error in Medical Students","year":2010,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":50,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Calgary; Health Sciences Centre","funders":"","keywords":"Medical diagnosis; Medicine; Diagnostic accuracy; Concordance; Pediatrics; Radiology; Internal medicine","score_opus":0.06561682143505307,"score_gpt":0.446273172229573,"score_spread":0.38065635079452,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2003211637","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9419064,0.000285478,0.00002552303,0.053655908,0.001405447,0.0006402057,0.0000064573173,0.00013021161,0.0019443829],"genre_scores_gemma":[0.9545606,0.0005100582,0.00033643836,0.040796075,0.003131992,0.00007678455,0.000029677447,0.00005492367,0.00050346594],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99483335,0.0001582264,0.0011728008,0.0007702333,0.0023749059,0.0006904827],"domain_scores_gemma":[0.9680549,0.029041363,0.00017884736,0.00064594846,0.00012652503,0.0019524484],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0036803505,0.00034919733,0.0009353183,0.00037464863,0.00008207165,0.000018915229,0.00088164624,0.00088729314,0.002991091],"category_scores_gemma":[0.3497971,0.00026045874,0.00007993037,0.00058890803,0.00038154857,0.0001466057,0.00022441396,0.0029583182,0.00031853613],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00060227676,0.0007247896,0.76511765,0.00008755539,0.00015227695,0.002022934,0.02138369,0.000002651629,0.006308609,0.0006086423,0.049264543,0.1537244],"study_design_scores_gemma":[0.012583566,0.0027324674,0.94679344,0.016527975,0.0005550793,0.00073032576,0.0048682927,0.0002790814,0.0037881276,0.0030915316,0.0071852864,0.00086483447],"about_ca_topic_score_codex":0.0005938593,"about_ca_topic_score_gemma":0.00013244229,"teacher_disagreement_score":0.34611675,"about_ca_system_score_codex":0.00006599987,"about_ca_system_score_gemma":0.000357203,"threshold_uncertainty_score":0.99998474},"labels":[],"label_agreement":null},{"id":"W2003404308","doi":"10.1097/00005082-200401000-00005","title":"Patientsʼ Understanding of Cardiac Risk Factors","year":2004,"lang":"en","type":"article","venue":"The Journal of Cardiovascular Nursing","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":30,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University; Canadian Heart Research Centre; University of Ottawa","funders":"","keywords":"Medicine; Audit; Documentation; Risk factor; Recall; Internal medicine; Psychology","score_opus":0.03913278544899355,"score_gpt":0.30125672756925903,"score_spread":0.26212394212026546,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2003404308","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96739286,0.012311235,0.018141434,0.00019518542,0.0008595404,0.00012846517,0.0000040288974,0.000008482447,0.0009587786],"genre_scores_gemma":[0.99691516,0.0023717252,0.00043612524,0.000014987282,0.00023855871,1.5488519e-7,0.0000013007733,0.000017446706,0.0000045644406],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9982688,0.00020878913,0.00047771685,0.000078396544,0.000789161,0.00017717885],"domain_scores_gemma":[0.9971933,0.0017859044,0.00034140408,0.00034781298,0.00018612473,0.00014545121],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001792427,0.00011271908,0.0006826199,0.00009551146,0.00009221225,0.000008869227,0.00010411493,0.00007003435,0.000005326216],"category_scores_gemma":[0.006846006,0.00006676631,0.0014704289,0.00019656893,0.00019009071,0.00005479983,0.000012105339,0.0003628226,0.000001644664],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004394811,0.003425052,0.18637276,0.00029470175,0.06375531,0.00034809142,0.048320517,0.28840736,0.0023922161,0.0030047302,0.0028842278,0.3964002],"study_design_scores_gemma":[0.030541116,0.009311622,0.6111739,0.06455478,0.0992658,0.00256423,0.058766585,0.00019037127,0.06417711,0.053532448,0.004496734,0.0014252674],"about_ca_topic_score_codex":0.0000857027,"about_ca_topic_score_gemma":1.7110442e-7,"teacher_disagreement_score":0.42480117,"about_ca_system_score_codex":0.00031851005,"about_ca_system_score_gemma":0.00014588979,"threshold_uncertainty_score":0.8195803},"labels":[],"label_agreement":null},{"id":"W2003457141","doi":"10.1097/01.acm.0000236543.88782.b6","title":"Medical Students?? Clinical Reasoning Skills as a Function of Basic Science Achievement and Clinical Competency Measures: A Structural Equation Model","year":2006,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":24,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"Division of Undergraduate Education","keywords":"Aptitude; Psychology; Structural equation modeling; Confirmatory factor analysis; Achievement test; Medical education; Mathematics education; Medicine; Mathematics; Standardized test; Statistics; Developmental psychology","score_opus":0.05112432079455465,"score_gpt":0.4284400285451951,"score_spread":0.37731570775064044,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2003457141","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9845201,0.001092457,0.0067370944,0.004738137,0.00084313896,0.00059307617,0.0000035995095,0.00008227685,0.0013901622],"genre_scores_gemma":[0.99137944,0.0019819,0.0009070084,0.0036255578,0.0015082867,0.000014952848,0.000048694503,0.00002708316,0.00050708756],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9914139,0.00028671074,0.002796962,0.00090896763,0.0040564844,0.00053700607],"domain_scores_gemma":[0.98833394,0.008978859,0.0007071899,0.0005197496,0.0004980971,0.0009621754],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.014102539,0.00030809193,0.0011676907,0.00025068695,0.00017183463,0.000014192877,0.00052686577,0.00061722443,0.00018647898],"category_scores_gemma":[0.13037296,0.00021914301,0.00016591666,0.00057058426,0.0027324674,0.00014959149,0.00029091825,0.0016814161,0.000027437985],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005250912,0.0002797758,0.957518,0.000038554263,0.0000999294,0.000026251755,0.0002216896,0.000038122787,0.00023004157,0.009487767,0.0017407426,0.029794024],"study_design_scores_gemma":[0.010895262,0.0029135717,0.92304,0.0054020286,0.0008690452,0.00011089617,0.00026319607,0.047490396,0.00004188618,0.008493434,0.00021861942,0.00026168197],"about_ca_topic_score_codex":0.00023268415,"about_ca_topic_score_gemma":0.000010444459,"teacher_disagreement_score":0.11627042,"about_ca_system_score_codex":0.00010075479,"about_ca_system_score_gemma":0.0011988085,"threshold_uncertainty_score":0.9999815},"labels":[],"label_agreement":null},{"id":"W2004263323","doi":"10.1001/jama.2013.7572","title":"The Evolution of the Master Diagnostician","year":2013,"lang":"en","type":"article","venue":"JAMA","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":22,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; University Health Network; Mount Sinai Hospital","funders":"","keywords":"Medicine; Intensive care medicine; Physiology","score_opus":0.016071495654521862,"score_gpt":0.2778771693169596,"score_spread":0.26180567366243773,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2004263323","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94579273,0.00038547092,0.00005695885,0.038295165,0.0003894886,0.00029426836,0.0000021482533,0.00001489436,0.014768901],"genre_scores_gemma":[0.9949583,0.000031522824,0.000055930843,0.0009830703,0.00022958327,0.000022648892,9.2486613e-7,0.0000056714593,0.003712341],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99942654,0.000036559617,0.00016873796,0.00007624345,0.00017078767,0.00012114362],"domain_scores_gemma":[0.9905309,0.008889564,0.00006622372,0.00033250888,0.00012073797,0.000060039925],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0001673978,0.000046707602,0.00009205874,0.0000089797395,0.000062254156,0.000017975033,0.000094532064,0.000054591113,0.00014805462],"category_scores_gemma":[0.049270302,0.000020064557,0.000079352816,0.00009105293,0.00009596908,0.000023741226,0.00004426818,0.00012372545,0.00022736989],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004263998,0.00016405409,0.7003093,0.000012089204,0.000038923557,0.0000026650446,0.00015383492,0.000004527727,0.0003453355,0.0050136787,0.113297455,0.1806155],"study_design_scores_gemma":[0.00058638485,0.00007031925,0.9919536,0.0004070812,0.000036174057,0.0000052610935,0.0001625759,0.0001627156,0.00031184655,0.002156305,0.004120027,0.00002771916],"about_ca_topic_score_codex":0.00057745754,"about_ca_topic_score_gemma":0.000031159823,"teacher_disagreement_score":0.29164428,"about_ca_system_score_codex":0.00004409347,"about_ca_system_score_gemma":0.00006141198,"threshold_uncertainty_score":0.9587381},"labels":[],"label_agreement":null},{"id":"W2004913883","doi":"10.1111/j.1365-2929.2004.01979.x","title":"Self‐reflection on the quality of decisions in health care","year":2004,"lang":"en","type":"letter","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre","funders":"","keywords":"Citation; Quality (philosophy); Reflection (computer programming); Psychology; Health care; Library science; Medical education; Media studies; Sociology; Medicine; Computer science; Political science; Epistemology; Law; Philosophy","score_opus":0.06643083663075568,"score_gpt":0.4836338128284156,"score_spread":0.41720297619765995,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2004913883","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.016468322,0.0007856901,0.00004369126,0.9785282,0.0012980271,0.0005368225,0.0000076004103,0.00003273148,0.0022989411],"genre_scores_gemma":[0.05121652,0.00072916725,0.0003801775,0.94436634,0.0022407342,0.00008943128,0.0005811197,0.000022658756,0.00037383373],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9970272,0.00038270693,0.0008673542,0.00030932666,0.0011936412,0.00021975518],"domain_scores_gemma":[0.9831747,0.015609011,0.00034750282,0.00052054063,0.00019491419,0.00015334539],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001369884,0.0001498731,0.00049038517,0.0001543106,0.000051578034,0.000007178166,0.00014653409,0.00066162116,0.00034321236],"category_scores_gemma":[0.18606919,0.000094941766,0.00014524047,0.00032153702,0.00009669707,0.000010271076,0.00002019687,0.0019720565,0.000050346636],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000034887707,0.0008033255,0.000964567,0.00024096096,0.000017747692,0.0000073870583,0.0010884177,0.0000010612127,2.1266173e-7,0.00036599496,0.9507628,0.045712624],"study_design_scores_gemma":[0.0015097748,0.00076494477,0.044435307,0.024420453,0.000062888845,0.000025391457,0.0007001463,0.000012308413,0.00001607009,0.0016229905,0.92622715,0.00020254528],"about_ca_topic_score_codex":0.0017916972,"about_ca_topic_score_gemma":0.00014647155,"teacher_disagreement_score":0.1846993,"about_ca_system_score_codex":0.0006610165,"about_ca_system_score_gemma":0.009934893,"threshold_uncertainty_score":0.9956778},"labels":[],"label_agreement":null},{"id":"W2005030293","doi":"10.5688/aj7410194","title":"Evaluation of Clinical Reasoning in Basic Emergencies Using a Script Concordance Test","year":2010,"lang":"en","type":"article","venue":"American Journal of Pharmaceutical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Concordance; Test (biology); Pharmacy; Medicine; Medical education; Multiple choice; Psychology; Family medicine; Significant difference; Internal medicine","score_opus":0.13366497640161806,"score_gpt":0.5336216288282879,"score_spread":0.39995665242666983,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2005030293","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99561334,0.00044639822,0.00014672951,0.0011182845,0.0018889064,0.00018089719,0.0000018301702,0.0000051680254,0.00059845456],"genre_scores_gemma":[0.98910826,0.00018803623,0.009357207,0.0005785089,0.0007347361,0.000004235124,0.0000022929585,0.000013639065,0.000013050972],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9970955,0.00034286254,0.0013624557,0.00017389796,0.0008333443,0.00019190005],"domain_scores_gemma":[0.9902648,0.0057475986,0.001208052,0.00022662093,0.0020941354,0.000458839],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0070312694,0.00011180611,0.0005781655,0.00017814094,0.000025402356,0.000009032478,0.00011240699,0.00006328684,0.00029783905],"category_scores_gemma":[0.16411549,0.000095239535,0.00018099409,0.00052441633,0.00046549292,0.00011064908,0.000018038754,0.00084153167,0.000004644449],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018356566,0.0015606137,0.50537956,0.000010936881,0.000040867733,0.0000044454214,0.00019450487,0.000038691036,0.010213617,0.0000742452,0.00022841837,0.48207054],"study_design_scores_gemma":[0.0039068623,0.0015440106,0.94917655,0.0021385234,0.0014728669,0.0004579454,0.0022636447,0.033076704,0.0042544217,0.0003994065,0.0011114676,0.00019759816],"about_ca_topic_score_codex":0.0001501623,"about_ca_topic_score_gemma":0.000014456597,"teacher_disagreement_score":0.48187295,"about_ca_system_score_codex":0.00009081729,"about_ca_system_score_gemma":0.00362166,"threshold_uncertainty_score":0.8429255},"labels":[],"label_agreement":null},{"id":"W2006109236","doi":"10.1016/j.ajog.2010.04.010","title":"Assessment of intraoperative judgment during gynecologic surgery using the Script Concordance Test","year":2010,"lang":"en","type":"article","venue":"American Journal of Obstetrics and Gynecology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":39,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Concordance; Medicine; Cronbach's alpha; Test (biology); Construct validity; Reliability (semiconductor); Gynecologic surgical procedures; Obstetrics and gynaecology; Rating scale; Medical physics; General surgery; Physical therapy; Surgery; Psychometrics; Clinical psychology; Patient satisfaction; Internal medicine; Laparoscopy; Psychology","score_opus":0.023346619764998487,"score_gpt":0.3291954071689163,"score_spread":0.30584878740391785,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2006109236","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9966188,0.00041572683,0.00012796,0.0009855325,0.0015661252,0.000116825075,0.0000056910694,0.000003943283,0.00015936594],"genre_scores_gemma":[0.9949652,0.00078756805,0.003536792,0.00057920034,0.000096291275,0.0000025118127,9.192138e-7,0.000009875444,0.000021659038],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99857724,0.000106561434,0.0007130396,0.00014709479,0.0002245609,0.00023151099],"domain_scores_gemma":[0.8683901,0.12957756,0.001079931,0.00021002418,0.0005208619,0.00022154886],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0008509959,0.00012450687,0.00074308325,0.00017240967,0.00009557979,0.000015659803,0.000105114166,0.00006574902,0.00007149159],"category_scores_gemma":[0.16934009,0.0000763324,0.00013799759,0.00043437027,0.00074096967,0.000040724484,0.000055362314,0.0006256291,4.9082587e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000021848606,0.0004115394,0.9554345,0.000013281464,0.000106228894,0.00023495144,0.000112423506,0.00003845581,0.008129103,0.0001261775,0.000107800515,0.035263695],"study_design_scores_gemma":[0.0009951752,0.002061542,0.9935007,0.00013268481,0.0001726939,0.000488862,0.0008506962,0.0003287937,0.00087246724,0.00006219657,0.00043919202,0.00009496918],"about_ca_topic_score_codex":0.000027534052,"about_ca_topic_score_gemma":0.000007092507,"teacher_disagreement_score":0.1684891,"about_ca_system_score_codex":0.00008257507,"about_ca_system_score_gemma":0.00050871755,"threshold_uncertainty_score":0.8376569},"labels":[],"label_agreement":null},{"id":"W2006319269","doi":"10.1046/j.1365-2923.2003.01528.x","title":"Doggie diagnosis, diagnostic success and diagnostic reasoning strategies: an alternative view","year":2003,"lang":"en","type":"letter","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":32,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; McMaster University Medical Centre","funders":"","keywords":"Citation; Library science; Medicine; Operations research; Computer science; Mathematics","score_opus":0.021862347480758743,"score_gpt":0.37364387028317064,"score_spread":0.35178152280241193,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2006319269","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.07578735,0.025020214,0.0004168801,0.88373834,0.0065310653,0.0022928896,0.00006602071,0.00030293947,0.0058443155],"genre_scores_gemma":[0.046290115,0.029865904,0.0007400261,0.90616614,0.010821596,0.0017520204,0.0033121598,0.00019810151,0.0008539312],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9942759,0.0006504947,0.0010958877,0.0013607034,0.0017614227,0.00085555913],"domain_scores_gemma":[0.918083,0.078269094,0.00060161,0.0011807589,0.0005510253,0.0013145214],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0011285098,0.0007495595,0.0013050251,0.00028121777,0.00019477734,0.00027654477,0.0004338018,0.0017022863,0.0023477275],"category_scores_gemma":[0.4249523,0.0006432293,0.00021537507,0.00039954096,0.00059383136,0.00030278406,0.00010321171,0.0034031884,0.00016695399],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000016191332,0.0011354268,0.027423928,0.0005322916,0.00013162811,0.0009689087,0.00030537834,0.0000015758492,1.4402464e-7,0.00030849228,0.884536,0.08464002],"study_design_scores_gemma":[0.0023110856,0.0012065334,0.03802428,0.03536639,0.0015980166,0.0009011236,0.0005455141,0.00015932471,0.00001769501,0.0033541436,0.9152433,0.0012725542],"about_ca_topic_score_codex":0.001471384,"about_ca_topic_score_gemma":0.000116623356,"teacher_disagreement_score":0.4238238,"about_ca_system_score_codex":0.00027997768,"about_ca_system_score_gemma":0.005876778,"threshold_uncertainty_score":0.99975896},"labels":[],"label_agreement":null},{"id":"W2006961035","doi":"10.3138/jvme.0312.020r1","title":"Outcomes Assessment of Case-Based Writing Exercises in a Veterinary Clinical Pathology Course","year":2012,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Preparedness; Medical education; Curriculum; Class (philosophy); Cohort; Psychology; Medicine; Computer science; Pathology; Artificial intelligence; Pedagogy","score_opus":0.14714434747959615,"score_gpt":0.5373609658844491,"score_spread":0.390216618404853,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2006961035","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9927024,0.0008865221,0.0003253022,0.0036595303,0.0020683678,0.00015477215,0.0000023774471,0.000008327793,0.00019235107],"genre_scores_gemma":[0.986514,0.00025211764,0.011083692,0.0013995366,0.00069782225,0.000012307141,0.000008748369,0.000015913542,0.000015890058],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9967923,0.00047809005,0.0017125924,0.00014329016,0.0006010185,0.0002726792],"domain_scores_gemma":[0.98944825,0.008609701,0.00073216646,0.00024014052,0.0002613633,0.00070839206],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004646355,0.00014139099,0.0007966776,0.00019014034,0.000027286562,0.000005859496,0.000118794786,0.00025231484,0.00034697657],"category_scores_gemma":[0.0381428,0.00010852727,0.00027552957,0.00015884041,0.00020830325,0.00011722209,0.000046570116,0.00069792586,0.000003465616],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00039669775,0.008075803,0.8962854,0.000106299914,0.000042692147,0.004780696,0.0001344588,0.0000027635547,0.00009452219,0.000032384458,0.0008364988,0.08921178],"study_design_scores_gemma":[0.0028737455,0.004579276,0.95895404,0.0029079749,0.00020036806,0.02890642,0.00059912144,0.0002553753,0.000007179969,0.00002093984,0.00060165214,0.00009391788],"about_ca_topic_score_codex":0.000026198322,"about_ca_topic_score_gemma":0.0000011226338,"teacher_disagreement_score":0.08911786,"about_ca_system_score_codex":0.00009819316,"about_ca_system_score_gemma":0.0025467265,"threshold_uncertainty_score":0.9699593},"labels":[],"label_agreement":null},{"id":"W2007157716","doi":"10.1016/j.jclinepi.2013.04.018","title":"Diagnostic probability function for acute coronary heart disease garnered from experts' tacit knowledge","year":2013,"lang":"en","type":"article","venue":"Journal of Clinical Epidemiology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Tacit knowledge; Diagnostic test; Medicine; Set (abstract data type); Intensive care medicine; Computer science; Emergency medicine; Knowledge management","score_opus":0.1941329512350133,"score_gpt":0.49129977954003595,"score_spread":0.2971668283050226,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2007157716","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95011646,0.0027165527,0.009060038,0.03364207,0.003274678,0.0009744431,0.000040922925,0.000038860686,0.00013595818],"genre_scores_gemma":[0.9678692,0.00078569155,0.012103863,0.014854445,0.0036483207,0.00013533888,0.00008062629,0.000033552315,0.00048891205],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.991865,0.0020996237,0.004812292,0.0005293893,0.00020004633,0.0004936613],"domain_scores_gemma":[0.28976566,0.70340794,0.0024045403,0.0007920811,0.00124827,0.0023815092],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.008122078,0.00026230118,0.0030963377,0.000069958965,0.00007132925,0.000011198983,0.00019934234,0.0005033882,0.0008015032],"category_scores_gemma":[0.7989446,0.00017617081,0.00151045,0.000094371724,0.0004541878,0.00013649078,0.000075985154,0.0008071524,0.000166975],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0037614503,0.0017373329,0.816569,0.000027140484,0.00044754604,0.000040484007,0.00002110915,0.000020206113,0.000020156851,0.00029073746,0.14771184,0.029353034],"study_design_scores_gemma":[0.0045290412,0.0037859096,0.8971776,0.00078668544,0.0011298602,0.000061144936,0.000021024041,0.0007984949,0.0000025282552,0.08104906,0.010507412,0.00015123094],"about_ca_topic_score_codex":0.000053961743,"about_ca_topic_score_gemma":0.0000051562774,"teacher_disagreement_score":0.7908225,"about_ca_system_score_codex":0.000065306616,"about_ca_system_score_gemma":0.0006094429,"threshold_uncertainty_score":0.87758976},"labels":[],"label_agreement":null},{"id":"W2007560473","doi":"10.1007/s10459-013-9459-3","title":"Finding and fixing mistakes: do checklists work for clinicians with different levels of experience?","year":2013,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":36,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Western Hospital; University of Toronto; University Health Network","funders":"","keywords":"Checklist; Commit; MEDLINE; Psychology; Interpretation (philosophy); Medical education; Medicine; Applied psychology; Computer science; Cognitive psychology","score_opus":0.056096643530171325,"score_gpt":0.4434307088986136,"score_spread":0.3873340653684423,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2007560473","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9918067,0.004140446,0.00056972605,0.0019823208,0.0003661975,0.0006682055,0.000002094672,0.000009314483,0.00045500675],"genre_scores_gemma":[0.9885406,0.0006185834,0.009912255,0.0005632791,0.000077663346,0.00018544837,0.0000030653039,0.0000048043084,0.00009432983],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99880826,0.000026960253,0.0004463758,0.00028920922,0.00018604794,0.00024316677],"domain_scores_gemma":[0.995515,0.0038731429,0.000259063,0.00012131805,0.000084740765,0.00014676977],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004740719,0.0000843872,0.00025197893,0.000088867324,0.00011224699,0.000023294608,0.00007536968,0.000032931122,0.000028918512],"category_scores_gemma":[0.0054400205,0.00005810385,0.000021216678,0.00033577924,0.00031571437,0.00019561569,0.000007296203,0.00007420691,6.9432616e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000028880804,0.00017278244,0.8234327,0.00018358034,0.0000013159904,8.419299e-8,0.0034419792,0.000032163927,0.000008250604,0.0003939772,0.000058982976,0.17224532],"study_design_scores_gemma":[0.00060399086,0.0008776001,0.9787522,0.0066095795,0.00000595394,0.0000049178407,0.010360353,0.00018150048,0.00013628659,0.0018274076,0.00053418375,0.00010603478],"about_ca_topic_score_codex":0.00016278414,"about_ca_topic_score_gemma":0.000051390765,"teacher_disagreement_score":0.17213929,"about_ca_system_score_codex":0.000056192835,"about_ca_system_score_gemma":0.00042401164,"threshold_uncertainty_score":0.65126055},"labels":[],"label_agreement":null},{"id":"W2007568911","doi":"10.1136/bmj.326.7393.829","title":"How 9/11 thinking can impair doctors' clinical judgment","year":2003,"lang":"en","type":"article","venue":"BMJ","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Psychology; Phenomenon; Social psychology; Morality; Magical thinking; Medicine; Law; Alternative medicine; Epistemology; Political science; Pathology","score_opus":0.0605067135789006,"score_gpt":0.39609665128169785,"score_spread":0.33558993770279727,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2007568911","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94246835,0.00023479371,0.00024114484,0.04024003,0.0016888345,0.0005729144,0.0000062084664,0.00016443898,0.014383286],"genre_scores_gemma":[0.9844637,0.00006476414,0.0035874567,0.0052341735,0.0009108638,0.000023858676,0.000015838203,0.0000225163,0.005676822],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9983533,0.00014079579,0.0004977243,0.00031663637,0.0003745433,0.0003169883],"domain_scores_gemma":[0.993178,0.0057161343,0.00013833027,0.0004723708,0.00008057077,0.0004145585],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0016178875,0.00014114202,0.0004007863,0.0000361898,0.000065846645,0.00004758014,0.00008315698,0.0001676914,0.00017506267],"category_scores_gemma":[0.08434204,0.00010615331,0.00024625738,0.000118643155,0.00008374047,0.00002428519,0.000032875672,0.00029192024,0.00006813088],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000042515272,0.000512602,0.8424053,0.000016007933,0.00007619241,0.00024114769,0.00014204862,0.000002165996,0.00001382158,0.0049405717,0.1427288,0.0088787675],"study_design_scores_gemma":[0.006583295,0.0013014877,0.6487773,0.001958286,0.00038095415,0.00027510323,0.00050366187,0.00007479099,0.00066303834,0.0031309796,0.33585557,0.00049553666],"about_ca_topic_score_codex":0.00006306761,"about_ca_topic_score_gemma":0.000047046433,"teacher_disagreement_score":0.19362804,"about_ca_system_score_codex":0.000050530685,"about_ca_system_score_gemma":0.00024630644,"threshold_uncertainty_score":0.92337096},"labels":[],"label_agreement":null},{"id":"W2008259","doi":"","title":"Kvalitetsgranskning av svenska ambulanssjukvårdens behandlingsriktlinjer rörande patienter med svår sepsis och septisk chock","year":2012,"lang":"sv","type":"article","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Canadian Institutes of Health Research","keywords":"Medicine; Religious studies; Gynecology; Philosophy; Internal medicine","score_opus":0.03087689589121433,"score_gpt":0.33258600972362096,"score_spread":0.3017091138324066,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2008259","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.946489,0.0074244686,0.0012617859,0.0035415208,0.0055967583,0.0012313913,0.00018035246,0.00046814035,0.0338066],"genre_scores_gemma":[0.9515071,0.001176839,0.0017899253,0.005478293,0.0034656893,0.00007258272,0.00038063191,0.00021409878,0.03591484],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99236816,0.00043790162,0.0019986837,0.0011576198,0.0015143863,0.0025232623],"domain_scores_gemma":[0.9828407,0.012553682,0.0005047508,0.0013476693,0.0004971153,0.0022560935],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0020516664,0.0010343783,0.0018568847,0.00033287948,0.0004142442,0.00017355832,0.00041343537,0.0010103141,0.013394017],"category_scores_gemma":[0.02508333,0.00085582206,0.0010918716,0.0006470599,0.00041339075,0.0003092658,0.00026087143,0.0013389256,0.0045201606],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008110909,0.0034143245,0.8912209,0.0002666441,0.0013348794,0.00015767933,0.007846804,0.0000324936,0.0005520771,0.00039410542,0.06922742,0.024741564],"study_design_scores_gemma":[0.03902705,0.0045992075,0.70697105,0.025573358,0.012321229,0.0004009987,0.0054882825,0.0028319387,0.0072657513,0.00031825426,0.19011937,0.005083507],"about_ca_topic_score_codex":0.00093939534,"about_ca_topic_score_gemma":0.00009872136,"teacher_disagreement_score":0.18424986,"about_ca_system_score_codex":0.0002617604,"about_ca_system_score_gemma":0.00025343598,"threshold_uncertainty_score":0.99938923},"labels":[],"label_agreement":null},{"id":"W2009090944","doi":"10.1007/s10459-014-9498-4","title":"Is it time to move beyond errors in clinical reasoning and discuss accuracy?","year":2014,"lang":"en","type":"letter","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Computer science; MEDLINE; Data science; Medical education; Medicine","score_opus":0.04158929875756412,"score_gpt":0.4697403645437179,"score_spread":0.4281510657861538,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2009090944","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.028945243,0.0044595925,0.000027729837,0.9591932,0.001619788,0.0009399967,0.00000897441,0.00002208059,0.004783378],"genre_scores_gemma":[0.00912503,0.005037193,0.0071930266,0.9731548,0.0022171962,0.00012738079,0.000098293356,0.000024929817,0.0030221809],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9960182,0.0003118166,0.0014102568,0.001052533,0.00056740746,0.0006397956],"domain_scores_gemma":[0.98335046,0.01510573,0.00061981427,0.0004451305,0.00009225333,0.00038663304],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0038265935,0.0002596018,0.0008193652,0.00041879682,0.00012539762,0.000051987838,0.0002822456,0.000366406,0.00017128208],"category_scores_gemma":[0.049525153,0.00020695217,0.00007842309,0.0008313396,0.00043299329,0.00026366508,0.00007416912,0.0012865005,0.00017566817],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000033743898,0.00019682987,0.13370238,0.00031897868,0.0000034834238,0.0000092948385,0.0009011189,0.00004158861,1.4767402e-7,0.00004703331,0.6290531,0.23569228],"study_design_scores_gemma":[0.00055763754,0.0007758791,0.08737552,0.013082779,0.000017946813,0.000034654673,0.00040881583,0.00067987025,0.0000013545148,0.0017481886,0.8949769,0.00034044526],"about_ca_topic_score_codex":0.00068497914,"about_ca_topic_score_gemma":0.00012269247,"teacher_disagreement_score":0.2659238,"about_ca_system_score_codex":0.00021954897,"about_ca_system_score_gemma":0.0020475567,"threshold_uncertainty_score":0.9584811},"labels":[],"label_agreement":null},{"id":"W2010036053","doi":"10.1046/j.1365-2923.2003.01577.x","title":"Diagnostic reasoning strategies and diagnostic success","year":2003,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":364,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary; Foothills Medical Centre","funders":"","keywords":"MEDLINE; Psychology; Medical education; Medicine; Political science","score_opus":0.012492666209228618,"score_gpt":0.3514191654319889,"score_spread":0.3389264992227603,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2010036053","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9756231,0.0041666604,0.0006313634,0.0061049815,0.00091211067,0.00030596709,0.0000010832637,0.00008622906,0.0121685555],"genre_scores_gemma":[0.99339986,0.0011727322,0.0008824582,0.0037038778,0.00031559682,0.0000928096,0.000042829786,0.0000196207,0.0003702425],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99847114,0.00011254386,0.00032794685,0.0003227516,0.0004930771,0.00027252134],"domain_scores_gemma":[0.95819414,0.040203836,0.00011806443,0.00039541943,0.00019729417,0.0008912403],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00063528743,0.00014917104,0.0002739088,0.00007010095,0.000091595015,0.000077247794,0.00006850659,0.00018208605,0.0010486243],"category_scores_gemma":[0.5865093,0.00012330175,0.000050375093,0.00020635132,0.00020288817,0.00012484609,0.000020781132,0.0003137463,0.00007312693],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000033374192,0.0019518476,0.76184773,0.0001674108,0.000055415196,0.00013652821,0.0006030925,0.000004157118,0.000011752727,0.054385412,0.02723262,0.15357064],"study_design_scores_gemma":[0.0021474555,0.00042950732,0.95837456,0.006690484,0.0002975486,0.0006998502,0.001882626,0.00019321145,0.00014163206,0.005545726,0.023222692,0.00037473682],"about_ca_topic_score_codex":0.00019152903,"about_ca_topic_score_gemma":0.00002450307,"teacher_disagreement_score":0.58587396,"about_ca_system_score_codex":0.000044150238,"about_ca_system_score_gemma":0.0024273228,"threshold_uncertainty_score":0.9998646},"labels":[],"label_agreement":null},{"id":"W201261089","doi":"10.17483/2368-6669.1009","title":"Clinical Nursing Reasoning in Nursing Practice: A Cognitive Learning Model Based on a Think Aloud Methodology","year":2014,"lang":"en","type":"article","venue":"Quality Advancement in Nursing Education - Avancées en formation infirmière","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":29,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"Canadian Foundation for Healthcare Improvement","keywords":"Think aloud protocol; Cognition; Context (archaeology); Nursing; Psychology; Nurse education; Nursing practice; Medicine; Computer science","score_opus":0.0876819022634834,"score_gpt":0.5242238408014817,"score_spread":0.4365419385379983,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W201261089","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.32744595,0.00031987272,0.6232345,0.016616462,0.0018460967,0.0014325395,0.000006186111,0.00021383612,0.028884515],"genre_scores_gemma":[0.890211,0.00014453525,0.10543913,0.0032436927,0.00032785637,0.000076561286,0.00038035848,0.00004478815,0.00013209715],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99282205,0.0032175751,0.001999617,0.00069080346,0.0006861391,0.00058379985],"domain_scores_gemma":[0.9681258,0.029595979,0.0010610161,0.0004668036,0.00046697748,0.00028339832],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.008033443,0.00035714495,0.00082580245,0.0005925555,0.00022237691,0.00006926019,0.00013960306,0.00036332797,0.000039730985],"category_scores_gemma":[0.11994183,0.00037949983,0.00017527374,0.0006099561,0.00028817097,0.0007453837,0.000014016287,0.001347599,0.000042606796],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0022725987,0.009142269,0.008704486,0.00012896334,0.000018700692,0.0000023854209,0.021373644,0.05748215,0.000028712057,0.042315613,0.00028414934,0.8582463],"study_design_scores_gemma":[0.010784944,0.002298001,0.051224902,0.10584795,0.00034489532,0.00006519229,0.049442567,0.73805666,0.00025477647,0.039335363,0.0013732752,0.0009714548],"about_ca_topic_score_codex":0.000042424817,"about_ca_topic_score_gemma":0.0000087356075,"teacher_disagreement_score":0.8572749,"about_ca_system_score_codex":0.0013591402,"about_ca_system_score_gemma":0.00065195095,"threshold_uncertainty_score":0.9998657},"labels":[],"label_agreement":null},{"id":"W2013908670","doi":"10.1097/acm.0000000000000614","title":"Disrupting Diagnostic Reasoning","year":2015,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":67,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Test (biology); Affect (linguistics); Audiology; Emergency medicine; Medical emergency; Psychology","score_opus":0.08311560416684625,"score_gpt":0.40681146697699333,"score_spread":0.32369586281014706,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2013908670","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9198639,0.007941154,0.0018349029,0.022601351,0.0013555106,0.0004803569,0.0000026849118,0.0003843037,0.045535836],"genre_scores_gemma":[0.9866117,0.0008068228,0.0006353797,0.005541718,0.002728113,0.000028253737,0.00005641711,0.000039151575,0.0035524447],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99772686,0.000071154274,0.0006404123,0.00039989923,0.0006961939,0.00046549278],"domain_scores_gemma":[0.98202753,0.01607601,0.00016912301,0.00038897843,0.0001855815,0.0011527494],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0016476922,0.00021664698,0.0005853667,0.00011480747,0.000056545272,0.000004768089,0.00015959473,0.00027853658,0.00022759172],"category_scores_gemma":[0.4535148,0.00015426171,0.00006242186,0.00036955826,0.00027575114,0.00006704243,0.000082154445,0.0009860345,0.00035904086],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00028674566,0.00006662493,0.61355287,0.00007324441,0.00011260791,0.0007959981,0.0061323154,0.000026354252,0.00039251425,0.0065638986,0.3346731,0.037323713],"study_design_scores_gemma":[0.06368639,0.010042806,0.41690847,0.09636276,0.0044100797,0.0079889195,0.027419955,0.007428084,0.00180895,0.0331531,0.32793814,0.002852336],"about_ca_topic_score_codex":0.00013411735,"about_ca_topic_score_gemma":0.0000012143163,"teacher_disagreement_score":0.4518671,"about_ca_system_score_codex":0.000100935475,"about_ca_system_score_gemma":0.0002012114,"threshold_uncertainty_score":0.629061},"labels":[],"label_agreement":null},{"id":"W2014767534","doi":"10.3138/jvme.35.4.612","title":"Veterinary Students' Attitudes toward the Assessment of Clinical Reasoning Using Extended Matching Questions","year":2008,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Variety (cybernetics); Curriculum; Medical education; Test (biology); Perception; Matching (statistics); Psychology; Process (computing); Medicine; Mathematics education; Computer science; Pedagogy; Pathology; Artificial intelligence","score_opus":0.21219913666038392,"score_gpt":0.5581965472949041,"score_spread":0.34599741063452016,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2014767534","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9916612,0.0006576794,0.002281696,0.002957844,0.0021039299,0.00017035934,0.000001317523,0.000013686027,0.00015230798],"genre_scores_gemma":[0.97570145,0.0014058432,0.020782065,0.00076488365,0.0012686838,0.000006190412,0.0000053772633,0.000018734667,0.00004678959],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99600947,0.0004871886,0.0016762095,0.00019273188,0.0014333811,0.00020102956],"domain_scores_gemma":[0.99316174,0.0044832104,0.0009978388,0.0003273557,0.00047689685,0.00055298],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0035503195,0.00015347714,0.000664626,0.00012216742,0.00013524767,0.000017498454,0.0003448496,0.00018001284,0.00022384737],"category_scores_gemma":[0.03418064,0.00010090176,0.00035703593,0.00019143034,0.00034698757,0.00013860747,0.00013117131,0.0009214572,0.0000027556623],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011532496,0.013205706,0.900426,0.00025273193,0.00060750113,0.0016604678,0.0021318668,0.000083244224,0.0013914241,0.00033912313,0.0024018213,0.07634691],"study_design_scores_gemma":[0.0014550028,0.0037649965,0.97249717,0.006134534,0.00022657205,0.014009796,0.0005041145,0.00038755947,0.0000075462253,0.00010480045,0.0008082951,0.00009959697],"about_ca_topic_score_codex":0.00006789857,"about_ca_topic_score_gemma":5.125435e-7,"teacher_disagreement_score":0.07624731,"about_ca_system_score_codex":0.00012102446,"about_ca_system_score_gemma":0.004008694,"threshold_uncertainty_score":0.97395486},"labels":[],"label_agreement":null},{"id":"W2014910999","doi":"10.1111/j.1741-6787.2008.00138.x","title":"Nurses' Uncertainty in Decision‐Making: A Literature Review","year":2008,"lang":"en","type":"review","venue":"Worldviews on Evidence-Based Nursing","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":78,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria; University of Toronto","funders":"Canadian Institutes of Health Research; Ontario Ministry of Health and Long-Term Care","keywords":"CINAHL; Viewpoints; Heuristics; MEDLINE; Context (archaeology); Nursing literature; Psychology; Nursing; Medicine; Computer science; Alternative medicine; Psychological intervention","score_opus":0.09834335516667983,"score_gpt":0.46131675531539923,"score_spread":0.36297340014871937,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2014910999","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000044079957,0.99042636,0.000046210418,0.0021318218,0.0015853859,0.0040220474,0.0000336097,0.00015111757,0.00159901],"genre_scores_gemma":[0.000027362086,0.98838264,0.0017239682,0.007883857,0.0007015692,0.00043988245,0.0002490988,0.00016582862,0.00042577568],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9918126,0.00110928,0.0029490632,0.0017512237,0.0013881346,0.0009897079],"domain_scores_gemma":[0.9463699,0.049498204,0.0011622427,0.0020738386,0.0003038402,0.00059197686],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.001666029,0.0013079491,0.0064418055,0.0012218495,0.00016594362,0.000104400344,0.0006499701,0.00082153344,0.0003129187],"category_scores_gemma":[0.117017984,0.00092815945,0.0024818152,0.0043251445,0.00034750748,0.00015924685,0.000041360487,0.0028654872,0.00055576913],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020998962,0.0007519539,0.000024043748,0.023467792,0.000036739704,0.0015262313,0.000053817406,0.000017149127,1.1364326e-8,0.000042454918,0.05928642,0.9145834],"study_design_scores_gemma":[0.00042913735,0.0002885331,0.000027556775,0.5280286,0.0007751036,0.000437571,0.0000023976677,0.000022278266,4.7451103e-8,0.000045857276,0.46952194,0.0004209599],"about_ca_topic_score_codex":0.000008608401,"about_ca_topic_score_gemma":0.000009114275,"teacher_disagreement_score":0.91416246,"about_ca_system_score_codex":0.0016483023,"about_ca_system_score_gemma":0.0023866664,"threshold_uncertainty_score":0.9999672},"labels":[],"label_agreement":null},{"id":"W2015581685","doi":"10.1007/s11606-013-2742-4","title":"An HIV-Positive Patient with Back Pain: Too Many Choices","year":2014,"lang":"en","type":"article","venue":"Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; University of Ottawa","funders":"","keywords":"Medicine; Human immunodeficiency virus (HIV); MEDLINE; Family medicine; Physical therapy","score_opus":0.015337976667205135,"score_gpt":0.31555127568231833,"score_spread":0.3002132990151132,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2015581685","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9611751,0.00044944527,0.012545792,0.011267774,0.00078067236,0.00014579449,0.000003473757,0.000015102881,0.013616861],"genre_scores_gemma":[0.974649,0.00011228881,0.003789269,0.013684668,0.0042987987,0.0000013473428,0.0000130067165,0.000029121571,0.0034224738],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9977897,0.00022060565,0.0008203991,0.00020455728,0.0007074294,0.0002573072],"domain_scores_gemma":[0.9954428,0.0024544233,0.00060058443,0.00024392121,0.00057996955,0.00067830813],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0011573809,0.00021139787,0.00069504057,0.00015644234,0.000037022033,0.000021038813,0.00018486449,0.00008479823,0.00052236865],"category_scores_gemma":[0.00806805,0.00011466118,0.00013054349,0.000117444295,0.00019215832,0.00011592442,0.000022616405,0.00048985623,0.000033227323],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.012123118,0.002834279,0.33964977,0.00017100503,0.0014352488,0.004455306,0.0025887245,0.00063677947,0.019095384,0.0011883004,0.31962752,0.29619455],"study_design_scores_gemma":[0.040981326,0.24187,0.44637004,0.060986307,0.0021846087,0.02186102,0.0016579765,0.026155932,0.015711369,0.0009979585,0.13995022,0.0012732438],"about_ca_topic_score_codex":0.00017894137,"about_ca_topic_score_gemma":0.0000085830725,"teacher_disagreement_score":0.2949213,"about_ca_system_score_codex":0.00007382619,"about_ca_system_score_gemma":0.00007415531,"threshold_uncertainty_score":0.9658792},"labels":[],"label_agreement":null},{"id":"W2015662441","doi":"10.1016/j.eurpsy.2008.02.003","title":"Cognitive insight and verbal memory in first episode of psychosis","year":2008,"lang":"en","type":"article","venue":"European Psychiatry","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":78,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Douglas Mental Health University Institute","funders":"Canadian Institutes of Health Research; Canada Research Chairs","keywords":"Neurocognitive; Psychology; Cognition; Verbal memory; Cognitive psychology; Verbal learning; Schizophrenia (object-oriented programming); Working memory; Association (psychology); Psychosis; Developmental psychology; Clinical psychology; Psychiatry; Psychotherapist","score_opus":0.022717882862759126,"score_gpt":0.29279134211850405,"score_spread":0.27007345925574494,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2015662441","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9238041,0.0027968388,0.000022641758,0.0011064908,0.0004328723,0.00014320518,0.000009370068,0.000030220213,0.07165424],"genre_scores_gemma":[0.99699944,0.0005409007,0.0006038144,0.0012352116,0.00022935269,0.0000032050077,0.000011578051,0.000027150698,0.0003493276],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9989672,0.00008801171,0.00035910544,0.000269308,0.00015482448,0.00016155232],"domain_scores_gemma":[0.9983554,0.0011665153,0.000086709704,0.0001936435,0.000042782547,0.00015493548],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0002108918,0.00012245211,0.00027776553,0.00009154176,0.000045910478,0.0000033452495,0.000057480953,0.00004593684,0.00008696221],"category_scores_gemma":[0.0028866858,0.00010764662,0.00007567702,0.00018860659,0.00018596165,0.000029292698,0.00003739718,0.00021541529,0.00009028707],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000377003,0.0006756585,0.9804922,0.000047269274,0.000038850896,0.000134685,0.0018543317,9.696033e-7,0.000006836522,0.00008141204,0.011110331,0.00518046],"study_design_scores_gemma":[0.0039496436,0.00047932353,0.99044883,0.0034684178,0.00005722232,0.00013019978,0.00028023147,0.000015702864,0.000035651326,0.00005863368,0.0009513008,0.00012486694],"about_ca_topic_score_codex":0.000070090195,"about_ca_topic_score_gemma":0.000063366446,"teacher_disagreement_score":0.07319534,"about_ca_system_score_codex":0.000008424589,"about_ca_system_score_gemma":0.000034911733,"threshold_uncertainty_score":0.43897018},"labels":[],"label_agreement":null},{"id":"W2015679564","doi":"10.1207/s15328015tlm1203_4","title":"Two Techniques for Teaching the Estimation of Prior Probabilities","year":2000,"lang":"en","type":"article","venue":"Teaching and Learning in Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Health Sciences Centre; Sunnybrook Health Science Centre","funders":"","keywords":"Heuristic; Computer science; Inter-rater reliability; Simple (philosophy); Medical diagnosis; Medical physics; Clinical Practice; Artificial intelligence; Machine learning; Psychology; Medicine; Statistics; Radiology; Mathematics; Family medicine; Epistemology","score_opus":0.020255133910087788,"score_gpt":0.36900772005413757,"score_spread":0.3487525861440498,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2015679564","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98199874,0.00042866755,0.0018297769,0.007174773,0.00003557588,0.00050197827,5.030697e-7,0.000121291734,0.007908699],"genre_scores_gemma":[0.97533137,0.000056714085,0.02183101,0.00045876706,0.0002120041,0.000044860353,0.000015921127,0.000015283471,0.0020341037],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9987015,0.0003576821,0.00039757352,0.00019202693,0.00018194357,0.00016928281],"domain_scores_gemma":[0.98630506,0.013354117,0.000094450916,0.00016082345,0.000026882572,0.00005863859],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005443239,0.00011225078,0.0003626476,0.0000779218,0.00025161373,0.0000074596783,0.000060536906,0.000059757775,0.000055359706],"category_scores_gemma":[0.07170824,0.00006633386,0.000042982803,0.000053885724,0.00027447962,0.000035360918,0.00001272803,0.0011448696,0.000001019697],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00016289766,0.00009250115,0.009181731,0.00011093934,0.000014571711,0.0000020593743,0.012275799,0.0006424423,0.00011924696,0.0016090379,0.00037482154,0.975414],"study_design_scores_gemma":[0.031671513,0.020242628,0.30797693,0.0916578,0.0016819076,0.0006727469,0.048352703,0.34974372,0.0007721732,0.023949122,0.12178211,0.001496633],"about_ca_topic_score_codex":0.00091816596,"about_ca_topic_score_gemma":0.000007634036,"teacher_disagreement_score":0.9739173,"about_ca_system_score_codex":0.00002413609,"about_ca_system_score_gemma":0.000034537108,"threshold_uncertainty_score":0.93611115},"labels":[],"label_agreement":null},{"id":"W2016332041","doi":"10.1111/ger.12109","title":"Confronting dental uncertainty in old age","year":2014,"lang":"en","type":"review","venue":"Gerodontology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Medicine; Disappointment; Health care; Context (archaeology); Blame; Psychiatry; Social psychology; Psychology","score_opus":0.0660318008420389,"score_gpt":0.4079233803446736,"score_spread":0.3418915795026347,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2016332041","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005414399,0.9922995,0.000018709092,0.000102504884,0.00083142205,0.00055525213,0.00001098482,0.00006440981,0.0055757924],"genre_scores_gemma":[0.00034022995,0.9952168,0.0002521963,0.00044421884,0.00049292,0.00007449615,0.0003711534,0.000046367677,0.0027616303],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9973707,0.00033560066,0.0009969599,0.00060474966,0.00016218603,0.00052979286],"domain_scores_gemma":[0.98413205,0.014778009,0.00032670057,0.00051761646,0.000033345274,0.00021227168],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00064629596,0.00036468112,0.0034870515,0.0001931133,0.000027720329,0.0000138209425,0.00019132912,0.000792246,0.00022120996],"category_scores_gemma":[0.037623193,0.0002843238,0.0005363996,0.0001719272,0.00016689631,0.000011559581,0.00009335995,0.0008121918,0.00041821992],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000019750476,0.0001215194,0.0017400833,0.0030226142,0.000087460336,0.0010532938,0.000034482982,5.379099e-7,5.263872e-8,0.0005524847,0.0032962747,0.9900714],"study_design_scores_gemma":[0.00095257466,0.00015644198,0.00086736213,0.041179672,0.00062741665,0.0005602551,0.000012518715,0.0000085343545,3.4158976e-8,0.000048508,0.9553429,0.00024377447],"about_ca_topic_score_codex":0.00065182307,"about_ca_topic_score_gemma":0.001540235,"teacher_disagreement_score":0.9898277,"about_ca_system_score_codex":0.00019709101,"about_ca_system_score_gemma":0.00032762502,"threshold_uncertainty_score":0.9999609},"labels":[],"label_agreement":null},{"id":"W2017626794","doi":"10.1097/00001888-200110001-00005","title":"Does “Shortness of Breath” = “Dyspnea”?","year":2001,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; MEDLINE; Intensive care medicine","score_opus":0.036198370005158846,"score_gpt":0.38146535231633544,"score_spread":0.3452669823111766,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2017626794","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96239066,0.0014553216,0.00047293617,0.018959679,0.0007592119,0.0002784741,0.0000039436954,0.00009687775,0.015582883],"genre_scores_gemma":[0.9831644,0.0043541207,0.00009334148,0.0027401203,0.0013666523,0.000012886095,0.00003632275,0.000022802453,0.008209334],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9981003,0.00003897025,0.00071806624,0.0003103698,0.0005217679,0.00031051884],"domain_scores_gemma":[0.99406314,0.004831024,0.000177454,0.00040719053,0.00014683224,0.00037437736],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0007173782,0.00017442348,0.00068667746,0.00012186204,0.00003085335,0.0000010952418,0.00016250474,0.0002952434,0.0014091799],"category_scores_gemma":[0.031465966,0.00008622172,0.000078695964,0.0003613494,0.00041382623,0.000037150327,0.000045784935,0.000595568,0.000055281947],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005390339,0.00010888853,0.89036745,0.00011574448,0.00014413962,0.00032896877,0.0016610883,0.000002990382,0.005345194,0.0024820687,0.03968596,0.05921847],"study_design_scores_gemma":[0.018802946,0.0032751726,0.75691164,0.025980989,0.0020899167,0.0018920063,0.0036477456,0.0004544321,0.009326154,0.012959417,0.16383119,0.00082838885],"about_ca_topic_score_codex":0.00020573022,"about_ca_topic_score_gemma":0.0000035607643,"teacher_disagreement_score":0.13345581,"about_ca_system_score_codex":0.000028885255,"about_ca_system_score_gemma":0.00008460218,"threshold_uncertainty_score":0.9995037},"labels":[],"label_agreement":null},{"id":"W2017687106","doi":"10.3822/ijtmb.v1i1.2","title":"Clinical Reasoning in Massage Therapy","year":2008,"lang":"en","type":"article","venue":"International Journal of Therapeutic Massage & Bodywork Research Education & Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Massage Therapy Foundation","keywords":"Massage; Guideline; Manual therapy; Process (computing); Psychology; Medicine; Physical therapy; Alternative medicine; Psychotherapist; Computer science; Pathology","score_opus":0.18699094005891395,"score_gpt":0.5509558062693878,"score_spread":0.36396486621047386,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2017687106","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8543814,0.011794826,0.0014540356,0.093955636,0.005047076,0.0007767545,0.0000041464496,0.000041770127,0.032544397],"genre_scores_gemma":[0.95608395,0.025819585,0.007797561,0.004958671,0.0025959634,0.00002646871,0.000013554939,0.000040193594,0.0026640561],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99326336,0.0016973253,0.0014955561,0.00035870573,0.002748407,0.00043667023],"domain_scores_gemma":[0.9606445,0.03256202,0.00094676897,0.0005130582,0.0048235687,0.000510116],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.012815593,0.00019282404,0.0005003059,0.00073754386,0.00013699348,0.00011884344,0.00067819236,0.00023093047,0.0007529929],"category_scores_gemma":[0.10939718,0.00016004579,0.00033035153,0.0006681834,0.00033439152,0.0005904583,0.00009204242,0.0024377657,0.00010473794],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008328874,0.009826061,0.52403075,0.000010546566,0.00093031354,0.0019515746,0.001565296,0.00002894574,0.0004050395,0.002303577,0.020903545,0.42971545],"study_design_scores_gemma":[0.007016451,0.0018448994,0.59252137,0.002791332,0.00010766692,0.005618713,0.003288052,0.00022647243,0.00021631425,0.0031220575,0.38293883,0.00030786544],"about_ca_topic_score_codex":0.0001836933,"about_ca_topic_score_gemma":0.000005834177,"teacher_disagreement_score":0.4294076,"about_ca_system_score_codex":0.000450593,"about_ca_system_score_gemma":0.0040374654,"threshold_uncertainty_score":0.9998636},"labels":[],"label_agreement":null},{"id":"W2018185664","doi":"10.1051/pmed/20090049","title":"Les approches qualitatives : une invitation à l’innovation et à la découverte dans le champ de la recherche en pédagogie des sciences de la santé","year":2009,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre Hospitalier de l’Université de Montréal; Université de Sherbrooke","funders":"","keywords":"Humanities; Sociology; Philosophy","score_opus":0.34926789689378557,"score_gpt":0.5603003224418694,"score_spread":0.21103242554808388,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2018185664","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7733892,0.011478682,0.0910513,0.01695919,0.00020294714,0.00036198067,0.000036090078,0.0002609751,0.10625962],"genre_scores_gemma":[0.8833814,0.008795973,0.09210191,0.0030805727,0.00057334243,0.00006374526,0.00007174155,0.000055759443,0.011875516],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9836553,0.013375474,0.0008174228,0.00067613507,0.0006147349,0.0008609568],"domain_scores_gemma":[0.8081372,0.19058499,0.000320465,0.00035989267,0.00022318258,0.0003742861],"candidate_categories":["metaresearch","metaepi_narrow","sts"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.032984417,0.00039769505,0.0005794443,0.0002707841,0.00054141297,0.00022746103,0.00039879876,0.0012190458,0.00009255026],"category_scores_gemma":[0.32902175,0.0003676598,0.00017853388,0.0016959005,0.0056428122,0.00030414868,0.00008317717,0.001555211,0.00003617509],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00016968609,0.0024649191,0.015699267,0.0002236213,0.000112048154,0.00009742544,0.13014142,0.0003399208,0.01511805,0.08470833,0.005501423,0.7454239],"study_design_scores_gemma":[0.0032598146,0.0014706749,0.41198388,0.0058636563,0.00021016263,0.00015319363,0.106302395,0.0031362933,0.00806713,0.44086522,0.017919825,0.00076775666],"about_ca_topic_score_codex":0.0012894956,"about_ca_topic_score_gemma":0.00006674856,"teacher_disagreement_score":0.74465615,"about_ca_system_score_codex":0.00066436373,"about_ca_system_score_gemma":0.0030676164,"threshold_uncertainty_score":0.9998775},"labels":[],"label_agreement":null},{"id":"W2018252278","doi":"10.1179/2042618613y.0000000051","title":"Establishing assessment criteria for clinical reasoning in orthopedic manual physical therapy: a consensus-building study","year":2013,"lang":"en","type":"article","venue":"Journal of Manual & Manipulative Therapy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Memorial University of Newfoundland; University of Toronto","funders":"","keywords":"Medicine; Manual therapy; Physical therapy; Orthopedic surgery; Medical physics; Physical medicine and rehabilitation; Medical education; Alternative medicine; Pathology; Surgery","score_opus":0.15302037715609118,"score_gpt":0.5132806462181101,"score_spread":0.3602602690620189,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2018252278","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9942338,0.0005162227,0.0008331684,0.0015437857,0.0006884548,0.0019057799,0.000005323345,0.000032798765,0.00024065482],"genre_scores_gemma":[0.98688376,0.0005318515,0.009509241,0.0009073318,0.0018951899,0.000071474955,0.000006787331,0.000065682114,0.0001286698],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9955247,0.0008718249,0.0018067801,0.00047640767,0.000789336,0.0005309474],"domain_scores_gemma":[0.98007184,0.01735549,0.0010098907,0.00036095505,0.0007932568,0.00040856478],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0047884434,0.00038679026,0.0014874713,0.00029034587,0.00011930972,0.00016725982,0.00027062927,0.00017079251,0.00017742829],"category_scores_gemma":[0.010239614,0.00027403628,0.00056372315,0.00028859146,0.00015899836,0.00032218295,0.00007047208,0.0010985471,0.000007650545],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0047272397,0.0098089,0.8161539,0.000025118767,0.00092730636,0.00039407166,0.0040217414,0.000086402215,0.0013740342,0.00045856184,0.003804391,0.15821832],"study_design_scores_gemma":[0.018416656,0.011699933,0.9587396,0.0011713728,0.00013601084,0.00012540461,0.0037253133,0.0025656961,0.000086146436,0.002208173,0.00080309366,0.00032264693],"about_ca_topic_score_codex":0.00013890259,"about_ca_topic_score_gemma":0.000010112258,"teacher_disagreement_score":0.15789567,"about_ca_system_score_codex":0.00016280617,"about_ca_system_score_gemma":0.00030245975,"threshold_uncertainty_score":0.9999712},"labels":[],"label_agreement":null},{"id":"W2018488739","doi":"10.1007/s10459-012-9347-2","title":"We’re talking about feedback … aren’t we?","year":2012,"lang":"en","type":"editorial","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Computer science; Medical education; Mathematics education; Psychology; Medicine","score_opus":0.025454981100574363,"score_gpt":0.44203735255482085,"score_spread":0.4165823714542465,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2018488739","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0007466731,0.19302095,0.00003254316,0.008001616,0.7926831,0.0005975117,0.000021201562,0.000055021777,0.0048413263],"genre_scores_gemma":[0.012421491,0.21926472,0.00466331,0.0011425292,0.7595127,0.00020563183,0.00036076346,0.000040092676,0.002388764],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99604064,0.00014423224,0.0010210731,0.0007798975,0.0011919279,0.0008222123],"domain_scores_gemma":[0.9824215,0.015455343,0.00088928285,0.00045936712,0.00021949236,0.00055499363],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0022717994,0.00030695533,0.00070832303,0.00033230794,0.00027102104,0.00006305834,0.000364271,0.00043710324,0.00018993046],"category_scores_gemma":[0.040871102,0.00025986537,0.00010072803,0.0008456745,0.0004407275,0.00046126224,0.00006029732,0.00097047066,0.000112646885],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000028662318,0.0003971376,0.009488434,0.00084921106,0.0000033813349,9.127668e-7,0.00091792655,0.00003425984,2.324955e-7,0.00019054383,0.67540914,0.31268016],"study_design_scores_gemma":[0.00038446515,0.00036598128,0.0023903176,0.021897916,0.00002238963,0.0000043813425,0.0011544649,0.000027554717,0.0000024715916,0.0014145086,0.9720952,0.00024038185],"about_ca_topic_score_codex":0.00078347843,"about_ca_topic_score_gemma":0.00043793095,"teacher_disagreement_score":0.31243977,"about_ca_system_score_codex":0.00061429414,"about_ca_system_score_gemma":0.0075211376,"threshold_uncertainty_score":0.99998534},"labels":[],"label_agreement":null},{"id":"W2018724911","doi":"10.1007/s00415-011-6200-0","title":"Challenges of clinical trial design when there is lack of clinical equipoise: use of a response-conditional crossover design","year":2011,"lang":"en","type":"article","venue":"Journal of Neurology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; University Health Network","funders":"Grifols; CSL Behring; Teva Pharmaceutical Industries","keywords":"Clinical trial; Crossover study; Clinical equipoise; Medicine; Clinical study design; Placebo; Randomized controlled trial; Research design; Intensive care medicine; Alternative medicine; Internal medicine; Statistics; Pathology","score_opus":0.6868988774160596,"score_gpt":0.5154992481651863,"score_spread":0.17139962925087326,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2018724911","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99249315,0.0005239758,0.0017738999,0.0036615427,0.0009770017,0.00036205884,0.000024312225,0.0000070734695,0.00017696706],"genre_scores_gemma":[0.98702586,0.0011734485,0.009088185,0.0019147056,0.0006691087,0.0000032853732,0.0000010231296,0.000028580367,0.00009581883],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.9910946,0.0034573802,0.0042961454,0.00031367186,0.00057649764,0.00026174524],"domain_scores_gemma":[0.8943724,0.10066304,0.0033291595,0.00046763464,0.0008399342,0.0003278495],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.009715561,0.00018901883,0.0020060726,0.00017861236,0.0000190134,0.000003950589,0.0002908562,0.00062591105,0.000717441],"category_scores_gemma":[0.1728444,0.00013836232,0.00096416706,0.000091764676,0.0008761973,0.00007804046,0.000077479664,0.00096428854,0.000011318148],"study_design_candidate":"randomized_trial","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.92317665,0.0052517373,0.04889634,0.000041544994,0.00070683786,0.00069136964,0.00044196553,0.000045941015,0.00027237323,0.00035927622,0.013786563,0.006329403],"study_design_scores_gemma":[0.07957363,0.09313235,0.814508,0.00055925717,0.00079559017,0.0005460606,0.000030937455,0.00017355163,0.0010681493,0.008171117,0.0012912457,0.00015008039],"about_ca_topic_score_codex":0.000013649002,"about_ca_topic_score_gemma":0.0000015923288,"teacher_disagreement_score":0.843603,"about_ca_system_score_codex":0.0000064900846,"about_ca_system_score_gemma":0.0009609067,"threshold_uncertainty_score":0.8341231},"labels":[],"label_agreement":null},{"id":"W2019650430","doi":"10.1007/s11017-009-9106-4","title":"Clinical intuition versus statistics: different modes of tacit knowledge in clinical epidemiology and evidence-based medicine","year":2009,"lang":"en","type":"article","venue":"Metamedicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":46,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Tacit knowledge; Intuition; Philosophy of medicine; Epistemology; Clinical epidemiology; Evidence-based medicine; Epidemiology; Context (archaeology); Experiential learning; Empiricism; Medicine; Psychology; Alternative medicine; Philosophy; Pathology; History; Mathematics education","score_opus":0.30877949219722556,"score_gpt":0.5277881040710384,"score_spread":0.21900861187381282,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2019650430","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94450647,0.01865153,0.018204158,0.01621771,0.0011011844,0.000575594,0.000013812217,0.00005600308,0.0006735142],"genre_scores_gemma":[0.98141325,0.009348541,0.0049006217,0.0031502962,0.0009896172,0.000013351172,0.00007303,0.000016124324,0.00009514477],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9938207,0.0014127203,0.0033373998,0.00066723797,0.00034597685,0.0004159418],"domain_scores_gemma":[0.84850967,0.14931212,0.00065373105,0.0005844207,0.0002815505,0.00065847544],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.009754888,0.00029712627,0.0027895789,0.00027444484,0.000029493385,0.0000025048462,0.00013734931,0.00035860846,0.00022151278],"category_scores_gemma":[0.3673453,0.000192673,0.00020383745,0.00029619443,0.0014223895,0.00004959855,0.000041923984,0.0007770508,0.000014595443],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0057211504,0.0018881696,0.5473154,0.00022031981,0.00014046587,0.00009465997,0.00009542545,0.000009400172,0.00015609711,0.0064927093,0.006183397,0.4316828],"study_design_scores_gemma":[0.018513126,0.015303868,0.94228905,0.011996125,0.0008288897,0.0000114874165,0.00009022803,0.0062294123,0.00004399208,0.00423405,0.0002919205,0.0001678261],"about_ca_topic_score_codex":0.00009167438,"about_ca_topic_score_gemma":0.000047696576,"teacher_disagreement_score":0.43151498,"about_ca_system_score_codex":0.000059230282,"about_ca_system_score_gemma":0.00020393704,"threshold_uncertainty_score":0.7856977},"labels":[],"label_agreement":null},{"id":"W2019963249","doi":"10.1097/acm.0000000000000121","title":"Deciding About Fast and Slow Decisions","year":2013,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":104,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Viewpoints; Interdependence; Reductionism; Fidelity; Computer science; Process (computing); Reflexivity; Management science; Psychology; Data science; Epistemology; Cognitive psychology; Sociology","score_opus":0.06039191995923042,"score_gpt":0.36823804114981973,"score_spread":0.3078461211905893,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2019963249","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.01593838,0.014411624,0.0005924635,0.95695347,0.0015697747,0.00077857653,0.00001935938,0.00017665542,0.009559673],"genre_scores_gemma":[0.0073342263,0.022832563,0.00044130845,0.90676546,0.0204453,0.00008185731,0.0005496542,0.00012223831,0.041427366],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99608564,0.000098974946,0.0011452476,0.00087671616,0.0010527715,0.00074065913],"domain_scores_gemma":[0.96133006,0.036749333,0.00036160037,0.0007134191,0.00019952055,0.000646092],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.00081082573,0.0005499015,0.0015323127,0.00035083157,0.00011687492,0.000014784745,0.0003145713,0.0026513585,0.0014243737],"category_scores_gemma":[0.13670436,0.00037212166,0.00013608504,0.00028506224,0.00063740544,0.000059661004,0.00017719885,0.0072630714,0.0005497327],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00001835695,0.000005252218,0.0024544862,0.000064357926,0.00009009731,0.0004916802,0.00032919928,1.2948091e-7,0.00005281839,0.000054992113,0.958231,0.03820766],"study_design_scores_gemma":[0.0028453884,0.0004233913,0.0065808953,0.032814592,0.00097520225,0.00096922915,0.00016984729,0.00010926686,0.00001759093,0.0019212236,0.9527532,0.00042016656],"about_ca_topic_score_codex":0.0001674251,"about_ca_topic_score_gemma":7.382772e-7,"teacher_disagreement_score":0.13589352,"about_ca_system_score_codex":0.00008599785,"about_ca_system_score_gemma":0.00018315601,"threshold_uncertainty_score":0.9998731},"labels":[],"label_agreement":null},{"id":"W2020412748","doi":"10.1186/1472-6920-7-16","title":"Problem formulation by medical students: an observation study","year":2007,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":28,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"University of Ottawa","keywords":"Schema (genetic algorithms); Medical diagnosis; Computer science; Mathematics education; Management science; Medicine; Psychology; Information retrieval; Pathology","score_opus":0.041779125202034036,"score_gpt":0.4474402489073909,"score_spread":0.4056611237053569,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2020412748","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9874971,0.00012159703,0.007304955,0.0024024814,0.00068855844,0.0009558667,0.0000010710015,0.00013046259,0.0008979282],"genre_scores_gemma":[0.9896513,0.000038253467,0.0028281582,0.0050193397,0.0009562485,0.000085447944,0.0005505783,0.00002517991,0.00084547384],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9949261,0.00014378355,0.0008154372,0.00041643347,0.0033877713,0.0003104679],"domain_scores_gemma":[0.9930017,0.004477802,0.00018405975,0.00042234393,0.00032728212,0.0015867893],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.005198554,0.00016097052,0.0002904525,0.00009925764,0.000093415416,0.00004292363,0.000288691,0.00033396494,0.001391802],"category_scores_gemma":[0.10036773,0.00013094273,0.000065423716,0.0003419157,0.0000592784,0.00016683804,0.00005532708,0.0003683595,0.000082365965],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00016304801,0.0112227015,0.835939,0.000033301054,0.000020813926,0.000007603309,0.0005206155,4.0468214e-7,0.000014645488,0.00037114817,0.011602196,0.14010455],"study_design_scores_gemma":[0.0035364772,0.0011688892,0.98835754,0.0011060019,0.00011674314,0.00003398369,0.0015264952,0.0006581271,0.00004078304,0.00037556028,0.0029055625,0.00017385215],"about_ca_topic_score_codex":0.00036465446,"about_ca_topic_score_gemma":0.00058427016,"teacher_disagreement_score":0.15241854,"about_ca_system_score_codex":0.00015434138,"about_ca_system_score_gemma":0.0026819394,"threshold_uncertainty_score":0.9995211},"labels":[],"label_agreement":null},{"id":"W2021139511","doi":"10.1007/s10459-014-9572-y","title":"ECG rhythm analysis with expert and learner-generated schemas in novice learners","year":2014,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":26,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Occupational Cancer Research Centre; Western University","funders":"Health Research Board","keywords":"Schema (genetic algorithms); Cognitive load; Cognition; Computer science; Dreyfus model of skill acquisition; Knowledge acquisition; Artificial intelligence; Natural language processing; Medicine; Machine learning","score_opus":0.02036380449442638,"score_gpt":0.408718305073841,"score_spread":0.3883545005794146,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2021139511","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98142326,0.007331872,0.0009001119,0.008470807,0.00017215802,0.00021324537,3.2877364e-7,0.000020386802,0.001467836],"genre_scores_gemma":[0.9753219,0.004227603,0.016852077,0.0033368468,0.00008080573,0.000039225848,0.000013575661,0.0000049911164,0.0001230038],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9986251,0.00010340239,0.00034149105,0.00041643737,0.0002417949,0.00027178647],"domain_scores_gemma":[0.9978532,0.0016014852,0.00014528057,0.00015817532,0.000057246853,0.00018457706],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012331152,0.000094349256,0.00029199597,0.00034996626,0.00009415776,0.000019388812,0.000054568438,0.00004130468,0.000018474151],"category_scores_gemma":[0.0052085537,0.00007041655,0.000019097177,0.0021467698,0.0002597551,0.00020465223,0.000012044714,0.00015007962,0.0000040052696],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000040589402,0.00020635828,0.80045396,0.000044589044,0.0000054340458,8.49885e-7,0.00081001443,0.004628242,0.00000888861,0.000786439,0.000042300995,0.1929723],"study_design_scores_gemma":[0.0029289539,0.0028262122,0.8529898,0.0047087707,0.000082353436,0.00006795522,0.010365767,0.09756136,0.00026622968,0.00092695066,0.026683846,0.0005918131],"about_ca_topic_score_codex":0.0022293255,"about_ca_topic_score_gemma":0.0022805908,"teacher_disagreement_score":0.19238049,"about_ca_system_score_codex":0.0000868965,"about_ca_system_score_gemma":0.0006882119,"threshold_uncertainty_score":0.6235501},"labels":[],"label_agreement":null},{"id":"W2021473674","doi":"10.1007/s10459-009-9188-9","title":"Diagnostic error in medical education: where wrongs can make rights","year":2009,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":69,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Variety (cybernetics); Perspective (graphical); Heuristic; Psychology; Term (time); Error analysis; Computer science; Medical education; Medicine; Artificial intelligence","score_opus":0.01748751405864477,"score_gpt":0.435543275102612,"score_spread":0.4180557610439672,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2021473674","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7643511,0.044362396,0.000051629002,0.17202225,0.0038263493,0.001111153,0.0000023203156,0.00006373748,0.014209024],"genre_scores_gemma":[0.97678274,0.005840976,0.0026080348,0.013852101,0.00043672413,0.0000997177,0.000024353065,0.000005111378,0.00035023192],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99761987,0.00011261075,0.00070539984,0.00048479208,0.0006312031,0.00044612723],"domain_scores_gemma":[0.9945518,0.004343525,0.00020504474,0.00025568926,0.00009402894,0.00054991886],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0013014604,0.00013784949,0.00031165517,0.00027064793,0.00014461484,0.000023494307,0.0002197692,0.00009900516,0.00017666975],"category_scores_gemma":[0.028327836,0.00011325556,0.0000351681,0.001098155,0.0002631483,0.00016064773,0.000013858238,0.00030587832,0.000020020663],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000017192273,0.0015089898,0.3679416,0.00007867691,6.3051283e-7,0.0000063413854,0.0006719838,0.00006820642,2.1340733e-7,0.0074101533,0.0015045974,0.62079144],"study_design_scores_gemma":[0.00081536424,0.0009408247,0.92050576,0.011723854,0.0000068936774,0.00009353434,0.0011600602,0.00047869104,0.0000062174686,0.029376,0.03467034,0.00022246363],"about_ca_topic_score_codex":0.0029211063,"about_ca_topic_score_gemma":0.013381808,"teacher_disagreement_score":0.620569,"about_ca_system_score_codex":0.00032685982,"about_ca_system_score_gemma":0.008633694,"threshold_uncertainty_score":0.99698645},"labels":[],"label_agreement":null},{"id":"W2022141589","doi":"10.1097/acm.0b013e3182178c3c","title":"How Much Practice Is Enough? Using Learning Curves to Assess the Deliberate Practice of Radiograph Interpretation","year":2011,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":119,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hospital for Sick Children; Professional Engineers Ontario; University of Toronto","funders":"","keywords":"Learning curve; Radiography; Medicine; Test (biology); Formative assessment; Clinical Practice; Sensitivity (control systems); Interpretation (philosophy); Psychology; Physical therapy; Radiology; Mathematics education; Computer science","score_opus":0.14307328520385965,"score_gpt":0.42830590486549347,"score_spread":0.2852326196616338,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2022141589","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.48745307,0.0373386,0.07871414,0.37099406,0.0016812597,0.0023935428,0.00000834801,0.00024673488,0.021170257],"genre_scores_gemma":[0.94003266,0.009754169,0.007932211,0.040604506,0.00050243834,0.000023846926,0.000015859549,0.000038064445,0.0010962405],"study_design_codex":"qualitative","study_design_gemma":"not_applicable","domain_scores_codex":[0.99742,0.0004913413,0.0006825231,0.00036953206,0.0007213534,0.00031522845],"domain_scores_gemma":[0.9734633,0.024394976,0.0006865021,0.00038701782,0.0007695505,0.00029864992],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0034080439,0.00021117763,0.0005467466,0.00017226652,0.00009635495,0.000009075755,0.00023714265,0.00021495376,0.00016894024],"category_scores_gemma":[0.3260173,0.00013285152,0.00008704409,0.00076650415,0.00026046103,0.0003551551,0.00007695429,0.0013624738,0.000014353315],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0132305585,0.0011904241,0.08737829,0.0021935424,0.0045713577,0.0007681894,0.35118988,0.00015453686,0.059179753,0.011230042,0.25349838,0.21541505],"study_design_scores_gemma":[0.019929606,0.01707589,0.029727833,0.1601353,0.027752645,0.007547654,0.2583936,0.01578778,0.042918764,0.0036909163,0.4145079,0.0025321164],"about_ca_topic_score_codex":0.00041275754,"about_ca_topic_score_gemma":7.5845946e-7,"teacher_disagreement_score":0.4525796,"about_ca_system_score_codex":0.000042024432,"about_ca_system_score_gemma":0.00013253487,"threshold_uncertainty_score":0.67965996},"labels":[],"label_agreement":null},{"id":"W2022212199","doi":"10.1007/s10459-011-9290-7","title":"Factors influencing responsiveness to feedback: on the interplay between fear, confidence, and reasoning processes","year":2011,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":403,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University; University of Calgary; University of British Columbia","funders":"","keywords":"Formative assessment; Psychology; Perception; Competence (human resources); Cognition; Perspective (graphical); Psychological intervention; Social psychology; Cognitive psychology; Computer science; Pedagogy","score_opus":0.0591990345861372,"score_gpt":0.4380884293424296,"score_spread":0.3788893947562924,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2022212199","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99188244,0.0018174634,0.000084872816,0.0030929104,0.00035166088,0.00047086622,0.0000018071865,0.000021302725,0.0022766439],"genre_scores_gemma":[0.9950332,0.000717903,0.0013341121,0.002714664,0.00007607057,0.000049816357,0.0000024472422,0.000005795464,0.00006599465],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9986376,0.0001074543,0.00035840826,0.00035521967,0.00026141698,0.00027986476],"domain_scores_gemma":[0.99024534,0.009002764,0.00018514376,0.00018424785,0.00013871814,0.00024379356],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0014359077,0.00011163625,0.00022067911,0.00018417511,0.00024472762,0.000031385385,0.00015936422,0.00003737774,0.000024038687],"category_scores_gemma":[0.043309722,0.00006864742,0.000015509408,0.00084782625,0.00026751508,0.00023492536,0.000033509037,0.0001710422,0.000007995989],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005944956,0.00006975093,0.9628626,0.00009445385,0.0000017382732,3.8685724e-7,0.007482242,0.000015905376,0.0000031361358,0.0025305883,0.00009516599,0.026784576],"study_design_scores_gemma":[0.00012113441,0.00086729816,0.9778577,0.0073546916,0.000005643311,0.000003895753,0.009612251,0.000009503187,0.00043120448,0.0024376004,0.0012009372,0.00009814948],"about_ca_topic_score_codex":0.0008150752,"about_ca_topic_score_gemma":0.00018699952,"teacher_disagreement_score":0.041873813,"about_ca_system_score_codex":0.000096262236,"about_ca_system_score_gemma":0.00158093,"threshold_uncertainty_score":0.96474886},"labels":[],"label_agreement":null},{"id":"W2023458254","doi":"10.1186/1472-6920-14-93","title":"Prognosis: the “missing link” within the CanMEDS competency framework","year":2014,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University; University of Toronto; William Osler Health System","funders":"","keywords":"Citation; Medicine; Curriculum; Medical education; Core competency; Psychology; Computer science; World Wide Web","score_opus":0.024980016562068514,"score_gpt":0.3665883091143833,"score_spread":0.3416082925523148,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2023458254","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.31807962,0.0023479722,0.029970897,0.62969214,0.00627973,0.001184134,0.0000013481588,0.00022910227,0.012215064],"genre_scores_gemma":[0.9578394,0.000060845872,0.0076531554,0.030468386,0.003382181,0.00013335008,0.00002088039,0.000018546418,0.00042320671],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.998088,0.00029572973,0.0004111721,0.00023693492,0.0007570011,0.00021118097],"domain_scores_gemma":[0.9508834,0.046573203,0.00028882833,0.0010651997,0.00027218222,0.0009171718],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002070459,0.00012118901,0.00021582049,0.000025940679,0.00022124461,0.00004629263,0.0002538219,0.00020802072,0.00043932747],"category_scores_gemma":[0.35424563,0.000057315698,0.000099678175,0.00024386523,0.00031832213,0.00002257857,0.000041119336,0.00068496476,0.00011618378],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007071357,0.0010413552,0.19769503,0.000120241384,0.000050002807,0.000002158838,0.0023509911,0.00001209481,0.0000059816007,0.07689452,0.033757824,0.68799907],"study_design_scores_gemma":[0.0019682173,0.0006997815,0.78648984,0.024271881,0.0007649262,0.00026479573,0.0031113524,0.02647587,0.000177044,0.09447939,0.060775705,0.0005211715],"about_ca_topic_score_codex":0.00016334171,"about_ca_topic_score_gemma":0.00006247909,"teacher_disagreement_score":0.6874779,"about_ca_system_score_codex":0.00003903769,"about_ca_system_score_gemma":0.00406906,"threshold_uncertainty_score":0.72183377},"labels":[],"label_agreement":null},{"id":"W2023928525","doi":"10.2147/ijgm.s38805","title":"Pivot and cluster strategy: a preventive measure against diagnostic errors","year":2012,"lang":"en","type":"article","venue":"International Journal of General Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":28,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Faculty of Medicine, Dalhousie University; Dalhousie University","keywords":"Debiasing; Medical diagnosis; Medicine; Process (computing); Cognition; Cluster (spacecraft); Set (abstract data type); Recall; Variety (cybernetics); Computer science; Artificial intelligence; Cognitive psychology; Psychiatry; Pathology; Psychology","score_opus":0.0316435731566111,"score_gpt":0.36162353541443887,"score_spread":0.3299799622578278,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2023928525","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9768052,0.004908112,0.0008836501,0.012480892,0.0022328144,0.00014235821,0.0000059786994,0.0000111703985,0.0025297971],"genre_scores_gemma":[0.98866856,0.0010799092,0.0011860956,0.003252867,0.004983685,0.0000035214798,0.000012668798,0.00001659449,0.00079609366],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99812144,0.0000842367,0.00064885634,0.000119408425,0.00081011566,0.00021596276],"domain_scores_gemma":[0.99489653,0.0035211924,0.00035605716,0.000099871395,0.0006450514,0.00048130067],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0008116221,0.00015027162,0.00039495973,0.00018143137,0.000026365562,0.000013064754,0.00013691736,0.00008551769,0.0001766007],"category_scores_gemma":[0.03417493,0.00009851248,0.00011246137,0.00008200187,0.00016091236,0.00014986447,0.00004340102,0.0003352535,0.000011658173],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018566132,0.0015425211,0.6502763,0.000047635192,0.0018636665,0.0013367708,0.0019724155,0.00019055029,0.0048181806,0.0010136429,0.08111531,0.2539664],"study_design_scores_gemma":[0.013085801,0.002241569,0.96171373,0.008026729,0.00075119844,0.004319174,0.000531062,0.00035862287,0.0006518823,0.0008167244,0.0072261225,0.00027735406],"about_ca_topic_score_codex":0.000035906014,"about_ca_topic_score_gemma":0.0000034409204,"teacher_disagreement_score":0.31143746,"about_ca_system_score_codex":0.00008083215,"about_ca_system_score_gemma":0.00010747938,"threshold_uncertainty_score":0.97396064},"labels":[],"label_agreement":null},{"id":"W2024540370","doi":"10.1186/1472-6920-12-100","title":"Constructing a question bank based on script concordance approach as a novel assessment methodology in surgical education","year":2012,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Concordance; Test (biology); Item bank; Scale (ratio); Construct validity; Psychology; Medical education; Medicine; Medical physics; Psychometrics; Clinical psychology; Item response theory","score_opus":0.11016383214940852,"score_gpt":0.47635237728047625,"score_spread":0.36618854513106774,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2024540370","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8557527,0.0005556747,0.068957694,0.0058732163,0.004907635,0.001013783,0.0000032780426,0.00010797946,0.06282803],"genre_scores_gemma":[0.8426963,0.000022374108,0.15125297,0.0039663226,0.001200116,0.0002778602,0.00027819324,0.000018597055,0.00028727544],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99726176,0.0006843047,0.00060076884,0.00039151753,0.00069144525,0.000370196],"domain_scores_gemma":[0.9831623,0.015222627,0.00022762314,0.00036334715,0.00018655934,0.0008375806],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0038719948,0.00017867882,0.00042878298,0.00020319303,0.000048427875,0.000016397553,0.00010006476,0.00032302862,0.0005297469],"category_scores_gemma":[0.14264336,0.00015438434,0.00009275923,0.00035598167,0.00018085295,0.00008230325,0.000022581777,0.00058847497,0.000040413463],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003539072,0.01118642,0.65835285,0.00019661289,0.000014133145,0.0000030363938,0.00024020908,0.000016590944,0.00004057481,0.09170571,0.0017987185,0.23609123],"study_design_scores_gemma":[0.01166136,0.0011985797,0.90683335,0.013442465,0.00033822365,0.0016674228,0.0069500622,0.04626455,0.0003234929,0.0019453939,0.0084966365,0.00087847555],"about_ca_topic_score_codex":0.00059853925,"about_ca_topic_score_gemma":0.000031279407,"teacher_disagreement_score":0.24848048,"about_ca_system_score_codex":0.00046566533,"about_ca_system_score_gemma":0.010963676,"threshold_uncertainty_score":0.9946433},"labels":[],"label_agreement":null},{"id":"W2024888162","doi":"10.1007/s10459-013-9482-4","title":"Effect of clinically discriminating, evidence-based checklist items on the reliability of scores from an Internal Medicine residency OSCE","year":2014,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":24,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Checklist; Generalizability theory; Reliability (semiconductor); Medicine; Inter-rater reliability; Summative assessment; Psychology; Rating scale; Formative assessment","score_opus":0.039599022990304565,"score_gpt":0.4776864808113299,"score_spread":0.4380874578210254,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2024888162","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9769995,0.0022239275,0.00026572245,0.017918542,0.0009576842,0.0005105884,0.0000024378742,0.000012165786,0.0011094144],"genre_scores_gemma":[0.99485,0.0011474441,0.0017668774,0.0018809304,0.0002761767,0.00004149422,0.000009079292,0.0000053478493,0.000022660543],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99641037,0.0010300558,0.0012845001,0.00044111058,0.0006315229,0.00020246161],"domain_scores_gemma":[0.8942602,0.103509896,0.0011546728,0.00064055005,0.00022824096,0.00020640291],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.012940875,0.0001223129,0.00047862576,0.00011161787,0.00009186001,0.000008007843,0.00035151336,0.00005235669,0.000055610497],"category_scores_gemma":[0.3114926,0.00006602081,0.000058084654,0.0004403516,0.0012974042,0.00021077674,0.000022707152,0.00024264342,0.0000020061934],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004413135,0.0003901651,0.8835863,0.00033861314,0.000001102037,1.1546509e-7,0.00047565147,0.0002720801,0.000045048535,0.0014775376,0.0001646307,0.11280745],"study_design_scores_gemma":[0.0009349946,0.02322137,0.9064419,0.051036462,0.000033285065,0.0000012401447,0.00060200266,0.0052702534,0.001711078,0.010456171,0.000178576,0.00011263716],"about_ca_topic_score_codex":0.008817298,"about_ca_topic_score_gemma":0.00033933067,"teacher_disagreement_score":0.29855174,"about_ca_system_score_codex":0.00007831547,"about_ca_system_score_gemma":0.000990919,"threshold_uncertainty_score":0.99778306},"labels":[],"label_agreement":null},{"id":"W2025246226","doi":"10.3109/0142159x.2014.993951","title":"Twelve tips to support the development of clinical reasoning skills using virtual patient cases","year":2014,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":103,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Curriculum; Medical education; Virtual patient; Core competency; Computer science; Psychology; Medicine; Engineering ethics; Pedagogy; Engineering","score_opus":0.07409517469194055,"score_gpt":0.41842319110046283,"score_spread":0.3443280164085223,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2025246226","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98991805,0.000031213487,0.0040538255,0.0012423298,0.0005711882,0.00028655617,0.0000016234701,0.000048592374,0.0038465888],"genre_scores_gemma":[0.9893826,0.0000064415444,0.006456487,0.0025622481,0.00067087676,0.000017570112,0.000017401182,0.000029728028,0.0008566494],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99620134,0.00033275405,0.0014231191,0.00040308907,0.0012444813,0.00039522233],"domain_scores_gemma":[0.9875217,0.010145177,0.00029567842,0.00063015963,0.00017687351,0.0012303999],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004481894,0.00019444719,0.0006946221,0.000053178683,0.00010364631,0.000011621123,0.00021494742,0.00028012818,0.0022007648],"category_scores_gemma":[0.24525365,0.000116431474,0.00021937562,0.0001880203,0.0002719904,0.000021455313,0.00021328872,0.0006367613,0.00017235427],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00019465288,0.0015010834,0.27978966,0.0000149297875,0.00013184373,0.000104654224,0.0021863154,0.000007962271,0.000011078665,0.00016087116,0.012192073,0.7037049],"study_design_scores_gemma":[0.013490915,0.012080399,0.50767004,0.01442911,0.0011367998,0.0013893492,0.003647946,0.016072692,0.0017013503,0.00018002419,0.42658073,0.0016206284],"about_ca_topic_score_codex":0.000055087265,"about_ca_topic_score_gemma":0.000025468724,"teacher_disagreement_score":0.70208424,"about_ca_system_score_codex":0.000053310636,"about_ca_system_score_gemma":0.00077305466,"threshold_uncertainty_score":0.99871135},"labels":[],"label_agreement":null},{"id":"W2026556764","doi":"10.7202/1017392ar","title":"Évaluations psychosociales","year":2013,"lang":"fr","type":"article","venue":"Nouvelles pratiques sociales","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"Ministère de l’Emploi et de la Solidarité Sociale (Québec)","funders":"","keywords":"Humanities; Valuation (finance); Political science; Sociology; Philosophy; Economics","score_opus":0.0777071653486057,"score_gpt":0.3920793143208907,"score_spread":0.31437214897228505,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2026556764","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.66140455,0.061181467,0.0008837578,0.17333035,0.008620464,0.0023592645,0.00016757747,0.000462805,0.09158975],"genre_scores_gemma":[0.9096649,0.013621685,0.01029341,0.0047017722,0.007856793,0.00033348534,0.00019470621,0.000105513725,0.053227715],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99660313,0.0005606291,0.0008945876,0.00053468265,0.00057833246,0.0008286411],"domain_scores_gemma":[0.9838693,0.014189457,0.0003242036,0.00042920065,0.000658274,0.00052955945],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00081469107,0.00040336754,0.00069994683,0.000100410696,0.00043398477,0.00020237225,0.00022760268,0.00082215917,0.012049818],"category_scores_gemma":[0.037287492,0.0004005986,0.0003958889,0.00037744912,0.001334086,0.0002721657,0.00008802543,0.00063852087,0.008552005],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003051229,0.0020195781,0.021232635,0.00018466053,0.000504317,0.00003315813,0.020857312,0.000007641625,0.00032726885,0.27373785,0.47598818,0.20507689],"study_design_scores_gemma":[0.00570433,0.0018444529,0.5045979,0.015931632,0.0020668618,0.00007146993,0.04072204,0.0015010857,0.0013024462,0.102421306,0.32158408,0.0022524255],"about_ca_topic_score_codex":0.002989457,"about_ca_topic_score_gemma":0.0006240034,"teacher_disagreement_score":0.48336524,"about_ca_system_score_codex":0.00040820305,"about_ca_system_score_gemma":0.0068262727,"threshold_uncertainty_score":0.9998446},"labels":[],"label_agreement":null},{"id":"W2027576789","doi":"10.1016/j.eururo.2014.11.025","title":"Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD): The TRIPOD Statement","year":2015,"lang":"en","type":"article","venue":"European Urology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":444,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Medical Research Council; University of North Carolina at Chapel Hill; Cleveland Clinic; University College London; Albert-Ludwigs-Universität Freiburg; Universiteit Maastricht; ZonMw; University of Oxford; National Institute for Health and Care Research; Cancer Research UK; University of Ottawa; Vanderbilt University; Memorial Sloan-Kettering Cancer Center","keywords":"Tripod (photography); Medicine; Checklist; Statement (logic); Psychology","score_opus":0.239752564895561,"score_gpt":0.38943571751661027,"score_spread":0.14968315262104928,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2027576789","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.92975444,0.00070047914,0.050397266,0.011279725,0.0008219432,0.0032775851,0.00042210947,0.00023229091,0.0031141725],"genre_scores_gemma":[0.9930025,0.000087272085,0.004869513,0.0009989733,0.00021897328,0.0002709874,0.0001358691,0.000040453488,0.00037548144],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9968627,0.00027258348,0.0018772117,0.00035926406,0.0003083103,0.00031994662],"domain_scores_gemma":[0.9938812,0.0042268354,0.0010951804,0.00032716096,0.00025548023,0.00021410546],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004277196,0.00015748647,0.00045489575,0.00006593817,0.000069183814,0.000012573366,0.0001516535,0.00007238132,0.00003044008],"category_scores_gemma":[0.079502285,0.00009480972,0.00014771734,0.0001491529,0.000119673125,0.000035526857,0.000062669125,0.00020592373,0.000007935475],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.011738452,0.007667954,0.58969903,0.00032006874,0.0016658999,0.00048387793,0.017102703,0.0501877,0.00025706712,0.0005838615,0.17589226,0.14440113],"study_design_scores_gemma":[0.048403192,0.029015174,0.35035825,0.0017164784,0.0070146364,0.00048186097,0.0016548993,0.46932995,0.0027878413,0.0014026146,0.086890504,0.00094460667],"about_ca_topic_score_codex":0.00003063783,"about_ca_topic_score_gemma":0.000008196785,"teacher_disagreement_score":0.41914225,"about_ca_system_score_codex":0.00003368757,"about_ca_system_score_gemma":0.00024845902,"threshold_uncertainty_score":0.92825145},"labels":[],"label_agreement":null},{"id":"W2027925451","doi":"10.1207/s15328015tlm1801_6","title":"The Cognitive Validity of the Script Concordance Test: A Processing Time Study","year":2005,"lang":"en","type":"article","venue":"Teaching and Learning in Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":47,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Concordance; Test (biology); Psychology; Affect (linguistics); Scripting language; Cognition; Information processing; Variance (accounting); Repeated measures design; Medicine; Cognitive psychology; Computer science; Statistics; Psychiatry; Mathematics; Communication","score_opus":0.03275952108721707,"score_gpt":0.3638653382226661,"score_spread":0.331105817135449,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2027925451","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98791015,0.0011176695,0.00006561218,0.006715091,0.00006618469,0.000355824,4.934278e-7,0.000046554578,0.0037224474],"genre_scores_gemma":[0.99601483,0.00004419295,0.00013549913,0.00065246725,0.0003057871,0.000011750084,0.0000017856191,0.000013471711,0.0028201938],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99812764,0.00068740017,0.00040605056,0.00022253035,0.00035089394,0.00020546188],"domain_scores_gemma":[0.9745324,0.024923269,0.00020441382,0.00017095798,0.00008296211,0.000085946376],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0051770494,0.00012641947,0.0003711242,0.000040564973,0.00048347676,0.000013422708,0.00010150097,0.000048230075,0.000018910518],"category_scores_gemma":[0.20892732,0.0000616749,0.000036092842,0.00016007842,0.00041789614,0.000031446787,0.000056728477,0.0016300745,0.000005218131],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010651396,0.00041754698,0.7130371,0.000024584746,0.000019725658,0.000010563982,0.009962251,0.000027904567,0.0001733951,0.000017871387,0.00053280644,0.27566978],"study_design_scores_gemma":[0.004820385,0.0019950673,0.9624124,0.008785371,0.00024444252,0.00005772279,0.011264501,0.0069782143,0.000030731127,0.00005835617,0.0032336647,0.00011911903],"about_ca_topic_score_codex":0.0002468688,"about_ca_topic_score_gemma":0.00003168783,"teacher_disagreement_score":0.27555066,"about_ca_system_score_codex":0.000028322778,"about_ca_system_score_gemma":0.00008141059,"threshold_uncertainty_score":0.7977362},"labels":[],"label_agreement":null},{"id":"W2028063503","doi":"10.1016/j.accpm.2015.01.003","title":"Assessing clinical reasoning in anesthesiology: Making the case for the Script Concordance Test","year":2015,"lang":"en","type":"letter","venue":"Anaesthesia Critical Care & Pain Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Medicine; Concordance; Liver transplantation; Anesthesiology; Observational study; Transplantation; Prospective cohort study; Internal medicine; Emergency medicine; Anesthesia","score_opus":0.10615459275248437,"score_gpt":0.44542678322509704,"score_spread":0.3392721904726127,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2028063503","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0074498937,0.02872672,0.0012992688,0.95908314,0.0008233253,0.0018022365,0.00001380693,0.00014955227,0.00065206114],"genre_scores_gemma":[0.32577378,0.0003098419,0.0013679999,0.6623591,0.009481729,0.00030330804,0.00016811035,0.00013235619,0.00010377393],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99187994,0.0025533899,0.0020391315,0.0012834959,0.000920325,0.001323739],"domain_scores_gemma":[0.65578765,0.33945042,0.00051830395,0.0021472096,0.001578027,0.0005184035],"candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.016289316,0.000782197,0.002196078,0.00018258356,0.00048715246,0.00014369043,0.0007219931,0.0016274643,0.00007522413],"category_scores_gemma":[0.55868316,0.00041497956,0.0005385297,0.0005295572,0.0035422593,0.00010231151,0.00009900248,0.0055027166,0.000019571797],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012160568,0.00010528665,0.041212745,0.00043716817,0.000052661293,0.05172059,0.00045121263,0.0000015649449,4.146364e-7,0.00033026404,0.8214622,0.08410427],"study_design_scores_gemma":[0.0040187016,0.005584524,0.014757764,0.023464296,0.0025520804,0.008828828,0.013748055,0.002192103,9.297051e-7,0.0013912709,0.92262244,0.0008389967],"about_ca_topic_score_codex":0.00070336874,"about_ca_topic_score_gemma":0.00015054972,"teacher_disagreement_score":0.5423938,"about_ca_system_score_codex":0.00031633137,"about_ca_system_score_gemma":0.0010891445,"threshold_uncertainty_score":0.9998302},"labels":[],"label_agreement":null},{"id":"W2029342902","doi":"10.1111/j.1365-2929.2004.01972.x","title":"What every teacher needs to know about clinical reasoning","year":2004,"lang":"en","type":"review","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":908,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Context (archaeology); Strengths and weaknesses; Psychology; Unconscious mind; sort; Medical education; Computer science; Medicine; Social psychology","score_opus":0.04688526559608029,"score_gpt":0.47342796339540694,"score_spread":0.42654269779932663,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2029342902","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00012897469,0.97996503,0.00008085856,0.00657978,0.009695646,0.0011342394,0.0000029777095,0.00014355108,0.0022689647],"genre_scores_gemma":[0.000016466343,0.96974176,0.0009459187,0.008686396,0.0057583163,0.00032347813,0.0005675124,0.000107473854,0.0138527],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99507624,0.00039798857,0.0019446459,0.0008071417,0.0012221229,0.000551846],"domain_scores_gemma":[0.9869873,0.008807425,0.00051927066,0.0011507459,0.00028532906,0.0022499566],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0027705382,0.0005246281,0.0025857494,0.00038311636,0.00008160802,0.00012455917,0.0004017481,0.0016235579,0.0024301987],"category_scores_gemma":[0.29183444,0.00039433365,0.0010123856,0.00092224154,0.00020897697,0.00012209163,0.00014700435,0.002096336,0.0024848403],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000015936615,0.0013890805,0.00019902697,0.0022539604,0.00009614429,0.000023889,0.000119029675,7.3560884e-8,2.2113171e-9,0.00012935202,0.11757476,0.87819874],"study_design_scores_gemma":[0.0003676735,0.00019072424,0.00028467248,0.33808866,0.0009913057,0.00015880057,0.0001005421,0.0000022712054,2.9196592e-8,0.000028901684,0.6595371,0.00024935958],"about_ca_topic_score_codex":0.000094405725,"about_ca_topic_score_gemma":0.000006763954,"teacher_disagreement_score":0.87794936,"about_ca_system_score_codex":0.00053723017,"about_ca_system_score_gemma":0.01973514,"threshold_uncertainty_score":0.99985087},"labels":[],"label_agreement":null},{"id":"W2029692488","doi":"10.1503/cmaj.080470","title":"Diagnosing the aging physician","year":2008,"lang":"en","type":"article","venue":"Canadian Medical Association Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Ceiling (cloud); Medicine; Engineering","score_opus":0.01585334830532616,"score_gpt":0.28857410441690395,"score_spread":0.2727207561115778,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2029692488","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7378133,0.00066232844,0.00009168383,0.24160297,0.0014088892,0.00012905982,0.0000060505645,0.000036462778,0.01824926],"genre_scores_gemma":[0.9136035,0.00051767833,0.000076416836,0.08085369,0.004092994,0.0000038985168,0.0000054237935,0.000014206457,0.00083218655],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99781215,0.00011196428,0.0003400778,0.00011449298,0.0011903852,0.00043095506],"domain_scores_gemma":[0.993138,0.0047044843,0.00014618608,0.00011712998,0.00026735492,0.0016268324],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012973562,0.00008924064,0.00021448352,0.000071645096,0.00059391075,0.00004207606,0.00013359074,0.00019622083,0.0010814613],"category_scores_gemma":[0.060697272,0.000060047838,0.00015116473,0.00022201493,0.00008156244,0.000051233328,0.000010921419,0.0011169257,0.00028243757],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000054934553,0.000066858775,0.42468137,0.0000023282173,0.00011441749,0.0024691648,0.0005125875,0.0000045003844,0.0000033493568,0.00018683668,0.4722496,0.0997035],"study_design_scores_gemma":[0.0014970145,0.00008381792,0.5583398,0.0007855967,0.00007703558,0.003014955,0.0003443114,0.00027268147,0.000023765362,0.00022262151,0.43519035,0.00014806309],"about_ca_topic_score_codex":0.0016343021,"about_ca_topic_score_gemma":0.0022258435,"teacher_disagreement_score":0.1757902,"about_ca_system_score_codex":0.0005377004,"about_ca_system_score_gemma":0.0028838925,"threshold_uncertainty_score":0.9998317},"labels":[],"label_agreement":null},{"id":"W2029775674","doi":"10.1111/j.1365-2923.2004.01730.x","title":"Limiting the playing field: does restricting the number of possible diagnoses reduce errors due to diagnosis‐specific feature identification?","year":2004,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; University of Toronto; University Health Network","funders":"","keywords":"Medical diagnosis; Salience (neuroscience); Identification (biology); Test (biology); Feature (linguistics); Diagnostic test; Interpretation (philosophy); Limiting; Psychology; Medicine; Computer science; Cognitive psychology; Pediatrics; Pathology; Engineering","score_opus":0.019980005346858717,"score_gpt":0.37221999842102677,"score_spread":0.35223999307416803,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2029775674","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.72873837,0.00047541363,0.00017822806,0.26831558,0.0011649712,0.0003820195,0.0000021888804,0.000036718735,0.0007064798],"genre_scores_gemma":[0.9887879,0.0005703933,0.0009941091,0.0078001134,0.0011334689,0.00029219955,0.000030091516,0.00001970505,0.00037200525],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9980094,0.00010697701,0.0005378179,0.0003291787,0.00075232086,0.00026432838],"domain_scores_gemma":[0.9810499,0.01742736,0.00027818247,0.0006715459,0.00028579324,0.00028721004],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010804655,0.00014502848,0.0002522654,0.00006284575,0.00036926655,0.000058379024,0.00032124863,0.00018236684,0.00033455863],"category_scores_gemma":[0.17044221,0.00007434281,0.00011835626,0.00062340766,0.00014175495,0.00006395363,0.00007194195,0.00056254334,0.00006924329],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013957039,0.0024671722,0.50236255,0.00012025962,0.000099826495,0.000046145808,0.009046316,0.000067253066,0.0003233485,0.0051506143,0.18064629,0.29953066],"study_design_scores_gemma":[0.0011884556,0.00018122375,0.9554188,0.007741436,0.000222672,0.00028170442,0.008633876,0.000038630296,0.011169259,0.004209825,0.010619643,0.00029444377],"about_ca_topic_score_codex":0.0005408023,"about_ca_topic_score_gemma":0.000052437576,"teacher_disagreement_score":0.4530563,"about_ca_system_score_codex":0.00010653235,"about_ca_system_score_gemma":0.0009353383,"threshold_uncertainty_score":0.8365455},"labels":[],"label_agreement":null},{"id":"W2029865094","doi":"10.1111/j.1365-2929.2006.02556.x","title":"Speed kills? Speed, accuracy, encapsulations and causal understanding","year":2006,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":43,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; The Wilson Centre","funders":"","keywords":"Psychology; Cognitive psychology; Task (project management); Social psychology","score_opus":0.036560183914786014,"score_gpt":0.37378518345363615,"score_spread":0.33722499953885016,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2029865094","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94781667,0.00044802314,0.0013702137,0.03306724,0.0009177197,0.00026653614,0.0000030647443,0.000088046916,0.016022477],"genre_scores_gemma":[0.99406946,0.00009668017,0.0004853739,0.002696744,0.00094859366,0.0000038548924,0.00016915963,0.000013182991,0.0015169441],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.998775,0.000035058743,0.00031141343,0.00022973621,0.00045724396,0.00019155034],"domain_scores_gemma":[0.9960454,0.003315438,0.00007648141,0.00017845968,0.000071472714,0.00031271108],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00035019894,0.00010215619,0.00018797613,0.000083257946,0.00009334593,0.000031651736,0.00004630331,0.00015656032,0.0009738145],"category_scores_gemma":[0.05270568,0.00008611321,0.000045495104,0.00018924879,0.00015690151,0.00005988243,0.000023858693,0.00022344697,0.00006185034],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014250448,0.003891847,0.43267852,0.00012558265,0.00008178147,0.00007635673,0.00080837746,0.000021203905,0.0006022755,0.0726754,0.4119028,0.07699337],"study_design_scores_gemma":[0.0038926078,0.00030150553,0.9366504,0.0029890027,0.0003420386,0.000454082,0.0014615718,0.005157536,0.0002133833,0.027721535,0.020383693,0.00043266945],"about_ca_topic_score_codex":0.0006281341,"about_ca_topic_score_gemma":0.00004724888,"teacher_disagreement_score":0.5039719,"about_ca_system_score_codex":0.000135769,"about_ca_system_score_gemma":0.0009261201,"threshold_uncertainty_score":0.99993944},"labels":[],"label_agreement":null},{"id":"W2030418672","doi":"10.1111/j.1553-2712.2000.tb00467.x","title":"The Cognitive Imperative Thinking about How We Think","year":2000,"lang":"en","type":"review","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":232,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen Elizabeth II Health Sciences Centre","funders":"","keywords":"Heuristics; Cognition; Medicine; Process (computing); Variety (cybernetics); Domain (mathematical analysis); Thinking processes; Clinical Practice; Cognitive psychology; Cognitive science; Psychology; Computer science; Artificial intelligence; Psychiatry; Nursing; Statistical thinking; Mathematics education","score_opus":0.10177299736029838,"score_gpt":0.4490807261252031,"score_spread":0.34730772876490473,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2030418672","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000028408045,0.9688108,0.000018051065,0.011677808,0.0033241345,0.0020327847,0.00005232552,0.00020421947,0.01385145],"genre_scores_gemma":[0.000045745874,0.97014207,0.000019694196,0.00073571445,0.006283304,0.00026739587,0.0004676997,0.00015366946,0.021884734],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9934807,0.00056095916,0.002536004,0.001077657,0.0013575405,0.0009871459],"domain_scores_gemma":[0.96879923,0.028076863,0.0012477086,0.00084067305,0.00033857793,0.00069694564],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0024390987,0.0010483244,0.0035304087,0.00022900167,0.00061212794,0.000005780783,0.0007916687,0.0015478247,0.0028862462],"category_scores_gemma":[0.08476639,0.00052951457,0.00095841254,0.0009966417,0.00078779156,0.000091685964,0.00016350354,0.0054912204,0.00050587195],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005048384,0.000043439897,0.0000681814,0.0017600971,0.0007212331,0.00006228572,0.0014493458,3.6937536e-8,2.264349e-7,0.0005337593,0.14640497,0.8489059],"study_design_scores_gemma":[0.00082452846,0.00040244363,0.00003995588,0.19326228,0.0054308316,0.00015702302,0.00054438587,0.0000044702706,5.4326694e-7,0.0010913953,0.7978245,0.00041762405],"about_ca_topic_score_codex":0.000046349633,"about_ca_topic_score_gemma":0.000005987259,"teacher_disagreement_score":0.84848833,"about_ca_system_score_codex":0.00015905431,"about_ca_system_score_gemma":0.0007503556,"threshold_uncertainty_score":0.99974835},"labels":[],"label_agreement":null},{"id":"W2030482510","doi":"10.1207/s15328015tlm1404_10","title":"Expert-Novice Differences in Memory: A Reformulation","year":2002,"lang":"en","type":"article","venue":"Teaching and Learning in Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":36,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Recall; Task (project management); Free recall; Cued recall; Cognitive psychology; Medical diagnosis; Psychology; Computer science; Medicine","score_opus":0.0501885370964179,"score_gpt":0.34740952279602033,"score_spread":0.2972209856996024,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2030482510","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97646016,0.0018864276,0.00006492068,0.008667025,0.000080228805,0.00011277139,8.40047e-8,0.000080054895,0.012648301],"genre_scores_gemma":[0.99565244,0.00040453102,0.0006053545,0.0010821862,0.00025145378,0.0000088374145,0.000007664389,0.000012124631,0.0019754302],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99866414,0.00029803775,0.0003419306,0.00025498625,0.00022310902,0.00021778356],"domain_scores_gemma":[0.99465156,0.005017777,0.0000699069,0.00012512255,0.000015309419,0.00012029866],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0016863653,0.0001238895,0.00040944392,0.0002089862,0.000099869336,0.000008169985,0.000043860902,0.0000985924,0.00021295878],"category_scores_gemma":[0.04716527,0.00008615023,0.000021652011,0.00015630518,0.000088149136,0.000044159588,0.000020248528,0.0014199296,0.000011763484],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007174808,0.00019536122,0.577347,0.00003226379,0.000009493797,0.00009480965,0.022013197,0.000051719424,0.0001792486,0.00012392887,0.0005435532,0.39933768],"study_design_scores_gemma":[0.0055330466,0.0009901247,0.9308174,0.007635378,0.000030049472,0.00007502109,0.0047720424,0.04664677,0.000004719408,0.00011093659,0.003192247,0.00019225714],"about_ca_topic_score_codex":0.0021958235,"about_ca_topic_score_gemma":0.000048731985,"teacher_disagreement_score":0.39914542,"about_ca_system_score_codex":0.000045145756,"about_ca_system_score_gemma":0.000009048345,"threshold_uncertainty_score":0.96086085},"labels":[],"label_agreement":null},{"id":"W2030522222","doi":"10.1007/s10459-014-9533-5","title":"“I can do better”: exploring purposeful improvement in daily clinical work","year":2014,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":49,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Hospital for Sick Children; The Wilson Centre","funders":"Hospital for Sick Children","keywords":"Context (archaeology); Health care; Psychology; Cognition; Audit; Excellence; Medical education; Applied psychology; Medicine","score_opus":0.06327009384031394,"score_gpt":0.4473135027526643,"score_spread":0.38404340891235034,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2030522222","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97897744,0.002634949,0.000090743655,0.013666865,0.002850738,0.000456852,5.7295256e-7,0.000023872442,0.00129797],"genre_scores_gemma":[0.98249227,0.00444116,0.0062079295,0.0061007296,0.00050422887,0.00018691069,0.000007281854,0.0000070555207,0.000052424544],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9976816,0.00014172946,0.00092150155,0.0005056514,0.00032858254,0.00042092882],"domain_scores_gemma":[0.99478734,0.0043600653,0.0002495825,0.0002711205,0.000048283546,0.00028360417],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0036723255,0.00010875628,0.00032329568,0.00019006377,0.000094491836,0.000027200958,0.00016442465,0.000049951333,0.000020243528],"category_scores_gemma":[0.014952025,0.00009186125,0.000041707135,0.0008412826,0.00022493048,0.00023564369,0.00003300215,0.0002808323,0.00001414185],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00001273821,0.00019812686,0.51219016,0.000027689575,3.9593462e-7,2.9710014e-7,0.0002334276,0.00005865201,5.4110257e-7,0.00044728443,0.000063708525,0.486767],"study_design_scores_gemma":[0.0009060519,0.001283707,0.9703375,0.004358803,0.0000044267194,0.0000029088137,0.0012661574,0.000229948,0.00001973596,0.007277007,0.014159188,0.00015458446],"about_ca_topic_score_codex":0.0004458416,"about_ca_topic_score_gemma":0.0003588691,"teacher_disagreement_score":0.4866124,"about_ca_system_score_codex":0.00019371699,"about_ca_system_score_gemma":0.000932061,"threshold_uncertainty_score":0.99334544},"labels":[],"label_agreement":null},{"id":"W2031055662","doi":"10.1111/medu.12260","title":"The roles of deliberate practice and innate ability in developing expertise: evidence and implications","year":2013,"lang":"en","type":"review","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":63,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Hamilton Health Sciences","funders":"","keywords":"Task (project management); Context (archaeology); Cognition; Cognitive psychology; Working memory; Psychology; Narrative; Neuroscience","score_opus":0.07873834061544066,"score_gpt":0.4838831313418563,"score_spread":0.4051447907264156,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2031055662","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0021843535,0.96604687,0.000043950266,0.030818742,0.00013230913,0.00067890954,0.0000010702117,0.000009302519,0.00008448576],"genre_scores_gemma":[0.0005628688,0.9966635,0.00092415995,0.0012351595,0.00009190877,0.00042415765,0.00001875025,0.000011255216,0.0000682482],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99828804,0.00027186485,0.00075031986,0.0002884643,0.0002607691,0.00014052137],"domain_scores_gemma":[0.95591336,0.042474173,0.00046673702,0.00046986717,0.0004013344,0.00027454746],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0015588055,0.00014004453,0.000592442,0.000058892307,0.00006156469,0.00002927879,0.00010489797,0.00023171403,0.000027234886],"category_scores_gemma":[0.34156635,0.00008284598,0.00004731776,0.00023930277,0.0002712908,0.00008234977,0.000074899195,0.00036394433,0.000006348233],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000004768205,0.00011973945,0.0003214,0.0013248627,0.000015789825,4.7371725e-7,0.00018443298,6.5479155e-9,1.288837e-7,0.0003938705,0.000962435,0.9966721],"study_design_scores_gemma":[0.00018812227,0.000056848865,0.010038989,0.10722627,0.00033104443,0.00015393888,0.00017330537,0.000026566126,7.3996745e-7,0.00029540516,0.8813736,0.0001351706],"about_ca_topic_score_codex":0.00019289291,"about_ca_topic_score_gemma":0.000043796266,"teacher_disagreement_score":0.9965369,"about_ca_system_score_codex":0.00010086778,"about_ca_system_score_gemma":0.0037653567,"threshold_uncertainty_score":0.6679581},"labels":[],"label_agreement":null},{"id":"W2031140420","doi":"10.1007/s10459-013-9485-1","title":"Estimation of post-test probabilities by residents: Bayesian reasoning versus heuristics?","year":2014,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; University of Calgary","funders":"","keywords":"Heuristics; Bayesian probability; Medical diagnosis; Bayesian statistics; Frequentist probability; Heuristic; Empirical probability; Bayes' theorem; Posterior probability; Statistics; Prior probability; Bayesian inference; Computer science; Econometrics; Artificial intelligence; Mathematics; Medicine","score_opus":0.014827767298013128,"score_gpt":0.4000064596297196,"score_spread":0.3851786923317065,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2031140420","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96452457,0.008944705,0.0051374556,0.01075425,0.0025314381,0.0009503217,0.000019969299,0.000073126444,0.007064152],"genre_scores_gemma":[0.9779712,0.0011034359,0.020137005,0.00051318534,0.00009040756,0.000035141813,0.000046617308,0.0000060577777,0.00009697464],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99846315,0.00009060684,0.00056505966,0.0002959735,0.00034183302,0.00024339579],"domain_scores_gemma":[0.9885067,0.010488488,0.00040340453,0.00024267411,0.00017135781,0.00018739779],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0013934831,0.00009024901,0.00023564539,0.0001215415,0.000114706985,0.000017257908,0.00011034914,0.00004218094,0.000020434873],"category_scores_gemma":[0.09986109,0.000079986916,0.000026199672,0.00043327516,0.00032151624,0.00026710477,0.000015233148,0.00010833159,0.0000053902163],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020338448,0.0010797376,0.34769878,0.00086003693,0.000003237891,2.7124855e-7,0.0012510986,0.00262543,0.000024160905,0.015657222,0.0015993346,0.6289973],"study_design_scores_gemma":[0.006522909,0.019890105,0.7396394,0.0376464,0.00008892651,0.00004347255,0.009842791,0.11528019,0.001770323,0.057731256,0.01056256,0.0009816882],"about_ca_topic_score_codex":0.0011742797,"about_ca_topic_score_gemma":0.00014437018,"teacher_disagreement_score":0.62801564,"about_ca_system_score_codex":0.00013620582,"about_ca_system_score_gemma":0.0009411718,"threshold_uncertainty_score":0.90772116},"labels":[],"label_agreement":null},{"id":"W2032181056","doi":"10.1177/0272989x09344747","title":"Measuring Acceptability of Clinical Decision Rules: Validation of the Ottawa Acceptability of Decision Rules Instrument (OADRI) in Four Countries","year":2009,"lang":"en","type":"article","venue":"Medical Decision Making","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":73,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Medical Council of Canada; Ottawa Hospital; Canadian Institutes of Health Research; University of Ottawa","funders":"","keywords":"Decision aids; Clinical decision making; Decision rule; Decision analysis; Management science; Medicine; Computer science; Family medicine; Artificial intelligence; Alternative medicine; Engineering; Statistics; Mathematics; Pathology","score_opus":0.06783647977659336,"score_gpt":0.4054685763401585,"score_spread":0.33763209656356513,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2032181056","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98062295,0.00024308667,0.016123265,0.0005941265,0.00087978895,0.00068352494,0.00002697733,0.00002992977,0.00079638336],"genre_scores_gemma":[0.9721666,0.0003022846,0.027048014,0.00029915426,0.00014576426,0.00000807219,0.0000071386035,0.000019053421,0.000003927128],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.98901045,0.0006312877,0.004986063,0.0007821938,0.0041634566,0.0004265714],"domain_scores_gemma":[0.92278403,0.07248025,0.0015538076,0.0018196048,0.00089659996,0.0004657149],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.014071313,0.0003208511,0.0017228537,0.00026422038,0.00009118182,0.000023875364,0.00081593514,0.0006451127,0.0024394037],"category_scores_gemma":[0.43772322,0.00020677455,0.0007418536,0.00068295933,0.0006746461,0.00016506703,0.000435643,0.00082884816,0.000016464817],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011239129,0.00072467764,0.40425235,0.00004311574,0.000015903353,0.000011064098,0.00009319016,0.000068687885,0.000018187324,0.00015212163,0.00010694023,0.59338987],"study_design_scores_gemma":[0.003930331,0.0005504686,0.9644116,0.006650706,0.000097311095,0.000020423731,0.00010013392,0.0014673845,0.0017544299,0.020593768,0.00026402826,0.00015940184],"about_ca_topic_score_codex":0.00008045142,"about_ca_topic_score_gemma":0.000066764755,"teacher_disagreement_score":0.5932304,"about_ca_system_score_codex":0.00016549215,"about_ca_system_score_gemma":0.0007256152,"threshold_uncertainty_score":0.9984725},"labels":[],"label_agreement":null},{"id":"W2033065001","doi":"10.1016/j.math.2014.10.001","title":"Sensibility of a new instrument to assess clinical reasoning in post-graduate orthopaedic manual physical therapy education","year":2014,"lang":"en","type":"article","venue":"Manual Therapy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Memorial University of Newfoundland; University Health Network; The Wilson Centre; Toronto Rehabilitation Institute; University of Toronto","funders":"","keywords":"Medicine; Sensibility; Certification; Competence (human resources); Medical education; Documentation; Workload; Rating scale; Context (archaeology); Test (biology); Manual therapy; Physical therapy; Applied psychology; Psychology; Alternative medicine; Pathology; Social psychology; Computer science","score_opus":0.11433543714019971,"score_gpt":0.4442714618331841,"score_spread":0.3299360246929844,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2033065001","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9927756,0.00011167057,0.00021188612,0.004544891,0.00055133493,0.0008119989,0.0000054114475,0.00006269163,0.0009244712],"genre_scores_gemma":[0.99116653,0.0003495129,0.0024788792,0.004597822,0.0009062701,0.000031317933,0.000042891024,0.000035069563,0.0003917265],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99718463,0.00042228738,0.0008486368,0.00061324844,0.00056276115,0.00036840627],"domain_scores_gemma":[0.9941313,0.004188719,0.00021461086,0.00070069777,0.00025673432,0.0005078994],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0019650974,0.0002643603,0.0008082211,0.0001419534,0.000037055386,0.000022693877,0.00016417565,0.00015913796,0.00006460812],"category_scores_gemma":[0.009473894,0.00020507156,0.00031671423,0.00034468953,0.00011723587,0.000061948565,0.000053363627,0.00046513806,0.000056084093],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018978028,0.002715913,0.15760495,0.000009407178,0.00006150159,0.0000054556563,0.0009884969,0.0000073498613,0.0006178362,0.00050102803,0.0011958409,0.8343944],"study_design_scores_gemma":[0.005802875,0.0038984597,0.981291,0.00074527663,0.000028026734,0.000018654244,0.0003541162,0.0002777952,0.0021888362,0.0022425612,0.0029042908,0.00024810864],"about_ca_topic_score_codex":0.00077736215,"about_ca_topic_score_gemma":0.00005078781,"teacher_disagreement_score":0.8341463,"about_ca_system_score_codex":0.000069828384,"about_ca_system_score_gemma":0.0006312008,"threshold_uncertainty_score":0.9988697},"labels":[],"label_agreement":null},{"id":"W2033127427","doi":"10.1007/s11251-013-9275-4","title":"Metacognitive scaffolds improve self-judgments of accuracy in a medical intelligent tutoring system","year":2013,"lang":"en","type":"article","venue":"Instructional Science","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":57,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"U.S. National Library of Medicine; Drexel University","keywords":"Metacognition; Dermatopathology; Test (biology); Correctness; Self-confidence; Computer science; Educational psychology; Cognitive psychology; Psychology; Artificial intelligence; Cognition; Developmental psychology; Medicine; Pathology; Social psychology; Psychiatry; Algorithm","score_opus":0.017300568566595242,"score_gpt":0.3305150644929815,"score_spread":0.3132144959263863,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2033127427","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99445724,0.00003941201,0.000094393006,0.00031378647,0.0007964297,0.00029817753,0.00000346737,0.000040026058,0.0039570453],"genre_scores_gemma":[0.9964763,0.000023713776,0.0031513285,0.00011708335,0.0001485921,0.00004275472,0.0000019134143,0.0000050164444,0.000033300963],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99769557,0.000024941492,0.00046730723,0.00032470212,0.0012225307,0.00026492396],"domain_scores_gemma":[0.99681515,0.0021513791,0.00013354585,0.00016351532,0.000383429,0.0003529643],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0008143719,0.000096744945,0.00024705727,0.00018011406,0.00006111526,0.000023841312,0.00019568233,0.00007675117,0.00025134857],"category_scores_gemma":[0.020282619,0.00007449447,0.0000658963,0.0006806952,0.0004286248,0.00023414736,0.000114721865,0.00013473902,0.000077743076],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001301995,0.001335865,0.21586832,0.00022162685,0.00014021352,0.000121041216,0.00086163613,0.00003392934,0.025485253,0.021267865,0.00010174717,0.7344323],"study_design_scores_gemma":[0.0018740771,0.00032787537,0.9550079,0.0034809774,0.000058588263,0.00041223076,0.0010335053,0.0076012136,0.029150628,0.0007116374,0.00013744306,0.00020390429],"about_ca_topic_score_codex":0.00060946384,"about_ca_topic_score_gemma":0.000005555462,"teacher_disagreement_score":0.7391396,"about_ca_system_score_codex":0.00025526443,"about_ca_system_score_gemma":0.0006297086,"threshold_uncertainty_score":0.98796993},"labels":[],"label_agreement":null},{"id":"W2033266691","doi":"10.1007/s10459-009-9179-x","title":"Dual processing and diagnostic errors","year":2009,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":244,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Categorization; Unconscious mind; Cognitive psychology; Process (computing); Dual (grammatical number); Cognition; Dual process theory (moral psychology); Computer science; Psychology; Matching (statistics); Cognitive science; Artificial intelligence; Medicine","score_opus":0.025178317890108687,"score_gpt":0.44325530177301314,"score_spread":0.4180769838829045,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2033266691","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.917977,0.050777562,0.00012382063,0.026720852,0.0005303868,0.00037954169,5.581317e-7,0.000038363243,0.003451967],"genre_scores_gemma":[0.9820163,0.006993836,0.0058144536,0.004936064,0.00013356602,0.000015826365,0.000004217739,0.0000024606288,0.000083254956],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9989893,0.000030237614,0.00029292173,0.0002657133,0.00019102202,0.00023079746],"domain_scores_gemma":[0.9971052,0.0024792221,0.00011521548,0.00007932114,0.000040582247,0.00018044842],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00065173744,0.000068176945,0.00015420248,0.00010876304,0.00013946866,0.000020900052,0.00004145995,0.000028148808,0.000008875238],"category_scores_gemma":[0.018462997,0.000055360906,0.0000118351945,0.00045669326,0.00018315541,0.0002590468,0.0000066360194,0.00010329512,0.0000039752335],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00000794115,0.00020504299,0.17420103,0.000054693406,1.866431e-7,0.0000011303134,0.00035888635,0.000055521836,0.00000222282,0.0011727677,0.00014256129,0.823798],"study_design_scores_gemma":[0.00042896834,0.0011384725,0.97471285,0.0045897886,0.000006124855,0.00006770374,0.0011348834,0.0007558322,0.000019076859,0.012268676,0.0047565456,0.00012107731],"about_ca_topic_score_codex":0.00008700757,"about_ca_topic_score_gemma":0.000036011777,"teacher_disagreement_score":0.82367694,"about_ca_system_score_codex":0.000057660483,"about_ca_system_score_gemma":0.0009589821,"threshold_uncertainty_score":0.9898049},"labels":[],"label_agreement":null},{"id":"W2034005877","doi":"10.1111/j.1365-2923.2008.03096.x","title":"The glass is a little full – of something: revisiting the issue of content specificity of problem solving","year":2008,"lang":"en","type":"letter","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Citation; Content (measure theory); Library science; Psychology; Medicine; Classics; Computer science; Mathematics; History","score_opus":0.04414088458902961,"score_gpt":0.34660841948312215,"score_spread":0.30246753489409256,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2034005877","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.021768177,0.0046080463,0.00013231764,0.9689166,0.0010232733,0.00060401094,0.000014270071,0.000014424217,0.0029188588],"genre_scores_gemma":[0.06795899,0.012298868,0.0035112922,0.87482697,0.02617807,0.00014678824,0.00044336668,0.00013331778,0.014502364],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99660283,0.00023746255,0.0012490568,0.0002781135,0.00137716,0.00025539772],"domain_scores_gemma":[0.97542846,0.021819837,0.001148452,0.0007074391,0.00078160217,0.00011420288],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0019871544,0.00018288929,0.00072265445,0.000069912145,0.00009420304,0.000009908081,0.00035540524,0.00052981084,0.00043026038],"category_scores_gemma":[0.12876484,0.00009826964,0.00029322898,0.00024294383,0.0006910461,0.000019368917,0.00007811789,0.0014984948,0.000012143669],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000044987133,0.000214608,0.0011384841,0.00062611065,0.00010086796,0.0000069403154,0.0015115738,1.7572003e-7,0.00005612505,0.00007400164,0.93213075,0.06409537],"study_design_scores_gemma":[0.0006944958,0.0003290225,0.004457158,0.018155517,0.00038179042,0.00009059028,0.0011007021,0.00012752914,0.00091572397,0.00033631647,0.9732436,0.00016758015],"about_ca_topic_score_codex":0.00058388995,"about_ca_topic_score_gemma":0.0000033642248,"teacher_disagreement_score":0.1267777,"about_ca_system_score_codex":0.00006720576,"about_ca_system_score_gemma":0.002101737,"threshold_uncertainty_score":0.87857395},"labels":[],"label_agreement":null},{"id":"W2034351127","doi":"10.1097/00001888-200310001-00025","title":"Can the Strength of Candidates Be Discriminated Based on Ability to Circumvent the Biasing Effect of Prose? Implications for Evaluation and Education","year":2003,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"","keywords":"Terminology; Medical diagnosis; Aptitude; Reliability (semiconductor); Psychology; Medical terminology; Medical education; Linguistics; Clinical psychology; Medicine; Developmental psychology; Nursing; Pathology","score_opus":0.04863532202192831,"score_gpt":0.42829845623293566,"score_spread":0.37966313421100734,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2034351127","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9461275,0.00039165598,0.00010959746,0.05069146,0.00011310443,0.0022446653,0.000020439096,0.0000075139446,0.00029407372],"genre_scores_gemma":[0.9979991,0.000047921923,0.000037048878,0.001385588,0.00007870649,0.00032290327,0.00008291301,0.000008292549,0.00003753984],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99877214,0.00025223248,0.000393775,0.00019530197,0.00026675066,0.00011976858],"domain_scores_gemma":[0.9801478,0.018997166,0.00018096565,0.00034374002,0.00022765699,0.0001027208],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0032917678,0.00009828551,0.00026504803,0.00005790427,0.000069955866,0.0000016423439,0.00007767002,0.000076576405,0.000017807886],"category_scores_gemma":[0.13500896,0.00004591948,0.000034455003,0.00022800283,0.00021350065,0.000009617029,0.000008336773,0.00018664516,1.7680998e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010178069,0.00034764473,0.67365146,0.00073407486,0.000099499026,1.1087184e-7,0.008249507,0.00016115591,0.016728364,0.0072482843,0.0133827245,0.27837935],"study_design_scores_gemma":[0.0069187847,0.0043944633,0.94030255,0.005701747,0.0020821423,0.000012178886,0.0015537445,0.0034845034,0.029707868,0.0048759216,0.0008120711,0.00015403684],"about_ca_topic_score_codex":0.0002556248,"about_ca_topic_score_gemma":0.000012902621,"teacher_disagreement_score":0.2782253,"about_ca_system_score_codex":0.000060718456,"about_ca_system_score_gemma":0.00030486565,"threshold_uncertainty_score":0.87227726},"labels":[],"label_agreement":null},{"id":"W2034960795","doi":"10.1016/j.annemergmed.2014.05.003","title":"How and When Do Expert Emergency Physicians Generate and Evaluate Diagnostic Hypotheses? A Qualitative Study Using Head-Mounted Video Cued-Recall Interviews","year":2014,"lang":"en","type":"article","venue":"Annals of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":91,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Université de Sherbrooke","funders":"Université de Strasbourg; Université de Sherbrooke; Université Catholique de Louvain","keywords":"Medical diagnosis; Medicine; Perspective (graphical); Context (archaeology); Specialty; Emergency department; Recall; Medical emergency; Process (computing); Qualitative research; Cognition; Medical education; Family medicine; Psychology; Nursing; Computer science; Psychiatry; Artificial intelligence; Cognitive psychology","score_opus":0.2741228230215195,"score_gpt":0.5153491247533556,"score_spread":0.24122630173183607,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2034960795","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96409047,0.021298995,0.00077347056,0.011228896,0.0011466623,0.00096185773,0.000017425431,0.000045810015,0.00043642998],"genre_scores_gemma":[0.97834873,0.018469093,0.00029634486,0.00096696295,0.0009703976,0.000079431615,0.000025508842,0.00006241676,0.0007811334],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.99564123,0.0009530732,0.0013753709,0.00077969464,0.0007320871,0.00051851716],"domain_scores_gemma":[0.99157315,0.005800835,0.0006207306,0.0006568817,0.0008715054,0.00047688856],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0029394417,0.00048099156,0.0014300019,0.00022171024,0.00012716056,0.000011378956,0.00017190057,0.00012818164,0.00070565596],"category_scores_gemma":[0.08139386,0.00035291366,0.00019573797,0.00039515743,0.0003010297,0.00015902394,0.00012935439,0.00028195968,0.0000070828264],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017999725,0.007626128,0.115557455,0.001272604,0.0033229687,0.000111649664,0.45197144,0.000031995976,0.026832353,0.0008422044,0.16335426,0.22727697],"study_design_scores_gemma":[0.041071657,0.07197463,0.40097648,0.053790916,0.008778635,0.00013197628,0.30538428,0.029508036,0.005219641,0.041948132,0.035828922,0.0053866925],"about_ca_topic_score_codex":0.001418024,"about_ca_topic_score_gemma":0.00021249973,"teacher_disagreement_score":0.28541905,"about_ca_system_score_codex":0.000016745242,"about_ca_system_score_gemma":0.00004048356,"threshold_uncertainty_score":0.9998923},"labels":[],"label_agreement":null},{"id":"W2034963030","doi":"10.1016/j.pec.2013.07.007","title":"Integrating clinical communication with clinical reasoning and the broader medical curriculum","year":2013,"lang":"en","type":"article","venue":"Patient Education and Counseling","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Curriculum; Enthusiasm; Communication skills; Salient; Psychology; Medical education; Computer science; Engineering ethics; Management science; Medicine; Artificial intelligence; Pedagogy; Social psychology; Engineering","score_opus":0.014253309928620339,"score_gpt":0.3532111490894975,"score_spread":0.3389578391608772,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2034963030","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9822297,0.0046269903,0.0005652935,0.009362567,0.00032479156,0.0005523854,4.2874208e-7,0.00004567473,0.002292146],"genre_scores_gemma":[0.9804924,0.0062985267,0.0061514815,0.006536487,0.0002573248,0.00011214018,0.00002892349,0.0000150806345,0.00010759618],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9979015,0.00041296668,0.0008237001,0.00030923838,0.00037925507,0.00017332743],"domain_scores_gemma":[0.9870887,0.011202119,0.00032225758,0.00041922773,0.00049973564,0.00046796174],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0020483467,0.00013890823,0.000363577,0.000033558907,0.00022564629,0.000097303455,0.00008981882,0.00015672819,0.00013846415],"category_scores_gemma":[0.028217254,0.00007634031,0.00007386565,0.00012409627,0.0006110318,0.00008775863,0.000050514645,0.0006573975,0.000025350999],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000072684605,0.00048708296,0.2961281,0.0000114113245,0.000039826296,6.717725e-7,0.0013881141,3.1391852e-7,1.3780895e-7,0.001619423,0.0031910331,0.69706124],"study_design_scores_gemma":[0.024062267,0.0034229662,0.6302965,0.029981164,0.001545429,0.0010764623,0.0938668,0.16779652,0.0000103948205,0.0060544633,0.040405646,0.0014813364],"about_ca_topic_score_codex":0.0005875058,"about_ca_topic_score_gemma":0.000030649353,"teacher_disagreement_score":0.6955799,"about_ca_system_score_codex":0.000014980442,"about_ca_system_score_gemma":0.00038899816,"threshold_uncertainty_score":0.9799685},"labels":[],"label_agreement":null},{"id":"W2035103815","doi":"10.1051/pmed/2011110","title":"Aborder le raisonnement clinique du point de vue pédagogique","year":2011,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Philosophy; Humanities","score_opus":0.05850622795046805,"score_gpt":0.3481958211945626,"score_spread":0.28968959324409455,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2035103815","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3588027,0.08942426,0.06086147,0.132234,0.014324871,0.0030051465,0.0002790858,0.0012833363,0.33978513],"genre_scores_gemma":[0.78132397,0.03773826,0.029523103,0.026822852,0.0078670615,0.0004069066,0.00013476158,0.000324473,0.11585862],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9944849,0.00066207256,0.0014533008,0.0010252221,0.00060881994,0.0017656613],"domain_scores_gemma":[0.98638815,0.009975166,0.00034609818,0.001155499,0.0003219737,0.0018131349],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0032427032,0.0006247221,0.0010964217,0.00015442201,0.00025957703,0.00006440232,0.0005461587,0.0010263205,0.017773448],"category_scores_gemma":[0.05656672,0.0006048089,0.000614799,0.00040159945,0.0010228462,0.00016235492,0.00033526213,0.0014055684,0.002464582],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012211226,0.018502627,0.1227406,0.0006258161,0.0009540861,0.004463828,0.03122936,0.000037237394,0.00043334113,0.03297071,0.3710528,0.41576847],"study_design_scores_gemma":[0.013067238,0.004076656,0.117357396,0.011248202,0.0013371387,0.00068792433,0.010866316,0.0020956993,0.008153466,0.021852335,0.80728006,0.0019775715],"about_ca_topic_score_codex":0.0027369175,"about_ca_topic_score_gemma":0.00021524593,"teacher_disagreement_score":0.43622726,"about_ca_system_score_codex":0.000326162,"about_ca_system_score_gemma":0.002313004,"threshold_uncertainty_score":0.99964035},"labels":[],"label_agreement":null},{"id":"W2035386357","doi":"10.3402/meo.v15i0.4846","title":"The Dreyfus model of clinical problem-solving skills acquisition: a critical perspective","year":2010,"lang":"en","type":"review","venue":"Medical Education Online","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":178,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"School of Medicine, University of Alabama at Birmingham; Covis Pharma; McGill University","keywords":"Intuition; Dreyfus model of skill acquisition; Knowledge acquisition; Perspective (graphical); Cognition; Psychology; Cognitive science; Coherence (philosophical gambling strategy); Computer science; Engineering ethics; Epistemology; Knowledge management; Artificial intelligence; Neuroscience","score_opus":0.0669307478922167,"score_gpt":0.5193285895816815,"score_spread":0.4523978416894648,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2035386357","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00010095384,0.97633433,0.00033384297,0.01655223,0.0025605175,0.0010439222,0.000107558575,0.00007521879,0.0028914448],"genre_scores_gemma":[0.000105730636,0.98203707,0.008962254,0.0020090702,0.004269139,0.00015145812,0.0005102876,0.00007406361,0.0018809317],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9939534,0.00044198293,0.0029435079,0.00077728723,0.0014100026,0.00047378594],"domain_scores_gemma":[0.8862463,0.10309759,0.0018030774,0.002417937,0.0033725642,0.0030625265],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.003185184,0.00045151447,0.002654215,0.000139714,0.00018141253,0.000039570143,0.0005515091,0.0015178033,0.00079436926],"category_scores_gemma":[0.4412313,0.00026526942,0.0012653295,0.00042119753,0.0014327033,0.00004627048,0.00017665939,0.0033469931,0.0000826825],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000027349874,0.007834619,0.00003790018,0.001414391,0.00012366459,0.000009631899,0.00009549462,2.4318248e-7,3.7612867e-8,0.008364548,0.007322975,0.9747692],"study_design_scores_gemma":[0.001011683,0.00045663008,0.00019177565,0.09815965,0.0026075167,0.00030344643,0.00028960637,0.0012522923,2.0972024e-7,0.0045352536,0.8907685,0.00042343524],"about_ca_topic_score_codex":0.000042075848,"about_ca_topic_score_gemma":0.0000245088,"teacher_disagreement_score":0.9743457,"about_ca_system_score_codex":0.00014419918,"about_ca_system_score_gemma":0.021900298,"threshold_uncertainty_score":0.99998},"labels":[],"label_agreement":null},{"id":"W2036284976","doi":"10.1111/j.1365-2753.2006.00583.x","title":"Beyond ‘evidence’. Commentary on Tonelli (2006), Integrating evidence into clinical practice: an alternative to evidence‐based approaches. <i> Journal of Evaluation in Clinical Practice</i> 12, 248–256","year":2006,"lang":"en","type":"letter","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"","keywords":"Empirical evidence; Clinical Practice; Evidence-based medicine; Foundation (evidence); Experiential knowledge; Medicine; Epistemology; Alternative medicine; Psychology; Knowledge management; Computer science; Family medicine; Political science; Philosophy; Pathology","score_opus":0.359836839627367,"score_gpt":0.5695918066179319,"score_spread":0.20975496699056484,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2036284976","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.03735958,0.010928653,0.0045206402,0.9294675,0.011753269,0.004686642,0.000021243837,0.000029379145,0.0012331108],"genre_scores_gemma":[0.07034777,0.010211366,0.14473216,0.74229485,0.03185345,0.00014616248,0.00014059432,0.00020256234,0.0000711113],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.8765026,0.07240951,0.030522162,0.0025843133,0.016658373,0.001323067],"domain_scores_gemma":[0.20123817,0.73773277,0.03462769,0.0022957942,0.02261208,0.0014934846],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["metaresearch","research_integrity"],"category_scores_codex":[0.3695336,0.0012510134,0.0049213823,0.0016272907,0.00022216162,0.00055662956,0.0015361321,0.0032499898,0.0004063076],"category_scores_gemma":[0.963036,0.0010419465,0.0021773914,0.0017432552,0.0008783618,0.0086580375,0.00034535138,0.026619043,0.00011398298],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.026551023,0.013729493,0.027157638,0.00022338427,0.001063396,0.0047634915,0.0006793176,0.009358053,0.0000059727026,0.000021055323,0.7009134,0.21553376],"study_design_scores_gemma":[0.054215655,0.11831338,0.092329,0.24379116,0.047549326,0.0055024954,0.0072132195,0.097932845,0.000059882525,0.0063359854,0.32297546,0.0037816004],"about_ca_topic_score_codex":0.0008671648,"about_ca_topic_score_gemma":0.00022320577,"teacher_disagreement_score":0.6752644,"about_ca_system_score_codex":0.0034269926,"about_ca_system_score_gemma":0.013140715,"threshold_uncertainty_score":0.9992031},"labels":[],"label_agreement":null},{"id":"W2037476011","doi":"10.1016/j.ijid.2012.07.003","title":"Advances in clinical education: a model for infectious disease training for mid-level practitioners in Uganda","year":2012,"lang":"en","type":"article","venue":"International Journal of Infectious Diseases","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":20,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Manitoba","funders":"International Society for Infectious Diseases","keywords":"Economic shortage; Medicine; Medical education; Infectious disease (medical specialty); Context (archaeology); Developing country; Health care; Human resources; Professional development; Capacity building; Disease; Nursing; Economic growth; Political science","score_opus":0.06647867385617187,"score_gpt":0.44769485791025243,"score_spread":0.38121618405408053,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2037476011","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9592653,0.003336314,0.029465305,0.0020324613,0.004213098,0.00068238063,0.0001778621,0.000029191306,0.00079807517],"genre_scores_gemma":[0.9942275,0.001152768,0.0010711827,0.0012843622,0.0019552652,0.00015480908,0.00007324537,0.00002180408,0.00005902693],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99824387,0.000056092398,0.0009660303,0.00016645253,0.00032627786,0.00024129602],"domain_scores_gemma":[0.99214983,0.0060364297,0.0005142699,0.00009911089,0.0007292322,0.00047114003],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0006940047,0.00014072901,0.00038298368,0.00038144615,0.000035824207,0.000034279667,0.00010746341,0.00007712457,0.000023704066],"category_scores_gemma":[0.052976284,0.00012555803,0.00039617097,0.00012591694,0.000070012546,0.00057123357,0.000017931767,0.0002170347,0.000002396765],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0020642595,0.004493333,0.90668714,0.00004852034,0.00009925647,0.000016492364,0.00031171113,0.0036432254,0.0000050449194,0.0010819524,0.00094250275,0.08060654],"study_design_scores_gemma":[0.017909443,0.00089494564,0.93410355,0.003836637,0.0004994612,0.0003154547,0.00049384957,0.009672925,0.000012567584,0.027206507,0.0046835355,0.0003711188],"about_ca_topic_score_codex":0.000012048222,"about_ca_topic_score_gemma":0.00004012404,"teacher_disagreement_score":0.08023542,"about_ca_system_score_codex":0.00027125905,"about_ca_system_score_gemma":0.0014730064,"threshold_uncertainty_score":0.9550009},"labels":[],"label_agreement":null},{"id":"W2038485424","doi":"10.1186/1748-717x-4-7","title":"The script concordance test in radiation oncology: validation study of a new tool to assess clinical reasoning","year":2009,"lang":"en","type":"article","venue":"Radiation Oncology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":62,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre Hospitalier de l’Université de Montréal; Hôpital Notre-Dame; Université de Montréal","funders":"Université de Montréal","keywords":"Concordance; Medicine; Radiation oncology; Cronbach's alpha; Test (biology); Medical physics; Radiation therapy; Reliability (semiconductor); Internal medicine; Oncology; Clinical psychology; Psychometrics","score_opus":0.07729824185787898,"score_gpt":0.4654903088375124,"score_spread":0.38819206697963343,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2038485424","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98493814,0.00025827234,0.00046550433,0.0099371085,0.00085062307,0.0013635741,0.0000033696144,0.00004934957,0.0021340444],"genre_scores_gemma":[0.99401104,0.00045284687,0.001712873,0.0027083925,0.00064439647,0.000055523607,0.000033885,0.000013174642,0.00036784576],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99706703,0.0005538576,0.0013577152,0.00040399557,0.00031453633,0.00030286828],"domain_scores_gemma":[0.97607714,0.022452137,0.0005990595,0.0004063336,0.00020447849,0.0002608702],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002959216,0.00015182981,0.00067722885,0.00014978423,0.000096549076,0.000023833165,0.00016176734,0.00028989068,0.000031842006],"category_scores_gemma":[0.12399282,0.0001234581,0.00009427665,0.00058074534,0.000059629918,0.000092718285,0.00002862765,0.00042910667,0.000044067758],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005137172,0.0020717606,0.3821359,0.0000021241872,0.00002380075,0.000026704667,0.0007467945,0.00025326497,0.000106660766,0.00036305308,0.008939806,0.60481644],"study_design_scores_gemma":[0.008116179,0.012587567,0.9413019,0.00015471016,0.00010681013,0.000018131657,0.0005408894,0.0013328756,0.00012768588,0.00021129848,0.035380263,0.000121720346],"about_ca_topic_score_codex":0.00022084643,"about_ca_topic_score_gemma":0.00022891216,"teacher_disagreement_score":0.60469466,"about_ca_system_score_codex":0.0005776981,"about_ca_system_score_gemma":0.0013213587,"threshold_uncertainty_score":0.88338614},"labels":[],"label_agreement":null},{"id":"W2038901171","doi":"10.1007/s10459-013-9451-y","title":"Working memory and mental workload","year":2013,"lang":"en","type":"editorial","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Workload; Psychology; Medical education; Computer science; Applied psychology; Medicine; Operating system","score_opus":0.01932216784707647,"score_gpt":0.4264930847209031,"score_spread":0.4071709168738266,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2038901171","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00451213,0.12748447,0.0000059131717,0.0033600735,0.8595004,0.00067538954,0.0000027422575,0.000030916286,0.0044279266],"genre_scores_gemma":[0.016672008,0.23751508,0.00782084,0.00200361,0.73228925,0.00030607375,0.00019051363,0.000032543598,0.0031700684],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99773,0.00007955793,0.00057969015,0.0005842571,0.0006388806,0.0003875789],"domain_scores_gemma":[0.989726,0.009260647,0.0004054194,0.00021184898,0.0000926341,0.00030341552],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0012858363,0.00017786407,0.00042579428,0.00021211334,0.00019009842,0.000059348105,0.00014689537,0.00022888547,0.000054568773],"category_scores_gemma":[0.020662501,0.0001491642,0.00003431995,0.00044733012,0.0003889943,0.00021701558,0.000045702556,0.00054370804,0.000024183082],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000143237585,0.00013258748,0.011104146,0.00016533321,0.0000017509219,5.0066853e-7,0.0002795299,0.000005556531,1.6991797e-7,0.000042981777,0.5387103,0.4495428],"study_design_scores_gemma":[0.00044193424,0.00040992437,0.002842598,0.019281955,0.000009526718,0.0000061135192,0.0008302961,0.00006762297,0.0000013875053,0.0012118199,0.9747107,0.00018610907],"about_ca_topic_score_codex":0.00086700014,"about_ca_topic_score_gemma":0.00020160552,"teacher_disagreement_score":0.4493567,"about_ca_system_score_codex":0.0002746304,"about_ca_system_score_gemma":0.003253108,"threshold_uncertainty_score":0.98758686},"labels":[],"label_agreement":null},{"id":"W2039069397","doi":"10.1007/s10459-009-9169-z","title":"Is case-specificity content-specificity? An analysis of data from extended-matching questions","year":2009,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":22,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Variance (accounting); Matching (statistics); Reliability (semiconductor); Test (biology); Psychology; Statistics; Computer science; Natural language processing; Mathematics","score_opus":0.1362837500955101,"score_gpt":0.49016003742326275,"score_spread":0.35387628732775267,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2039069397","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97853005,0.009444482,0.004846912,0.0055065453,0.00052168715,0.00030804976,0.00015145335,0.0000335487,0.00065728684],"genre_scores_gemma":[0.97198486,0.003846048,0.021427024,0.002227065,0.00015894105,0.0000066975595,0.00031532344,0.0000035812334,0.000030443856],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9977789,0.00013267032,0.00079519546,0.0006488331,0.00038231688,0.00026210613],"domain_scores_gemma":[0.99562144,0.0028240033,0.0003914176,0.0007537619,0.0001550577,0.0002542989],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0014773603,0.0001123709,0.00045066187,0.00038105587,0.00019063757,0.000030520096,0.0003153676,0.000051215447,0.0001330798],"category_scores_gemma":[0.0042782566,0.000097125834,0.00005536929,0.0016141252,0.00022516862,0.00083979213,0.000029534252,0.00016870508,0.0000045167803],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000092548566,0.0033302645,0.36572352,0.000042960226,0.00003470488,0.000020916517,0.0033297455,0.0019031474,0.00006488148,0.008982408,0.0003811577,0.61609375],"study_design_scores_gemma":[0.00036309243,0.00071092934,0.9574421,0.001311106,0.00018320777,0.000036441063,0.009994551,0.018295854,0.000043770448,0.010082771,0.001357466,0.00017873175],"about_ca_topic_score_codex":0.011612651,"about_ca_topic_score_gemma":0.004135353,"teacher_disagreement_score":0.615915,"about_ca_system_score_codex":0.00012053893,"about_ca_system_score_gemma":0.00074007467,"threshold_uncertainty_score":0.9949691},"labels":[],"label_agreement":null},{"id":"W2039461472","doi":"10.1197/j.aem.2005.11.080","title":"Will a New Clinical Decision Rule Be Widely Used? the Case of the Canadian C‐Spine Rule","year":2006,"lang":"en","type":"article","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":46,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Ottawa","funders":"","keywords":"Medicine; Clinical prediction rule; Decision rule; Logistic regression; Rule-based system; Association rule learning; Statistics; Data mining; Artificial intelligence; Internal medicine","score_opus":0.06780318696104719,"score_gpt":0.4149795650677673,"score_spread":0.3471763781067201,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2039461472","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9082253,0.0078112604,0.00012569819,0.07490951,0.0031916657,0.00063091423,0.000020459605,0.00004580876,0.005039401],"genre_scores_gemma":[0.98273027,0.0028911098,0.00023581204,0.0040656016,0.0034588198,0.000014312569,0.00003479169,0.000043147647,0.0065261303],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9957066,0.00018452927,0.0023722383,0.00047961867,0.0006929676,0.0005640251],"domain_scores_gemma":[0.9922297,0.005159653,0.0005343569,0.0010026309,0.0002579007,0.0008157383],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00272628,0.0002894077,0.00075245503,0.00012285185,0.0002745799,0.0000012371697,0.00050830934,0.00053398206,0.0023268727],"category_scores_gemma":[0.07306449,0.00014514946,0.00034327977,0.0008111178,0.0006124558,0.000075620876,0.000112404065,0.0015696273,0.000039016428],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007769777,0.00011429788,0.43894082,0.000020869915,0.00008719588,0.00027317647,0.0003293298,0.000022135171,0.0001682949,0.0025132212,0.5411273,0.016325613],"study_design_scores_gemma":[0.0064637396,0.0008470276,0.73500544,0.003853116,0.0014638412,0.0016289508,0.00045237187,0.0005392174,0.0001994039,0.024755925,0.2243466,0.0004443588],"about_ca_topic_score_codex":0.31435508,"about_ca_topic_score_gemma":0.13181989,"teacher_disagreement_score":0.31678075,"about_ca_system_score_codex":0.00010794098,"about_ca_system_score_gemma":0.00105156,"threshold_uncertainty_score":0.99858516},"labels":[],"label_agreement":null},{"id":"W2039630921","doi":"10.1111/acem.12034","title":"Assessing Diagnostic Reasoning: A Consensus Statement Summarizing Theory, Practice, and Future Needs","year":2012,"lang":"en","type":"article","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":81,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; University of British Columbia; McMaster University","funders":"","keywords":"Medicine; Engineering ethics; Management science","score_opus":0.053664258048644754,"score_gpt":0.4191965756890384,"score_spread":0.3655323176403936,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2039630921","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.82807374,0.09828158,0.00041844076,0.040046576,0.0049714716,0.000768702,0.000006715214,0.000219747,0.027213054],"genre_scores_gemma":[0.93199694,0.055038817,0.0012851036,0.0026615183,0.007742832,0.000052265077,0.00006960616,0.000067078734,0.001085839],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.9962345,0.00037038402,0.0014301649,0.00040056466,0.00072115334,0.00084328203],"domain_scores_gemma":[0.97330743,0.024307353,0.0006203678,0.00043773148,0.00026163078,0.0010655051],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0050890977,0.00038034617,0.00069446216,0.00023503492,0.00019970346,0.000003743652,0.00012855594,0.00039218584,0.0011196176],"category_scores_gemma":[0.20880385,0.00028142796,0.00009097624,0.0006185818,0.00031685174,0.00031712966,0.00011439457,0.0013309459,0.0000498392],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00032711905,0.00035911246,0.85105956,0.0001769634,0.00037274818,0.00007886548,0.005997693,0.000001069556,0.0008789498,0.022646045,0.08526073,0.03284116],"study_design_scores_gemma":[0.0062629865,0.0011544127,0.46475467,0.007865545,0.0040560123,0.0011612314,0.04603799,0.00013271924,0.00021325043,0.0023674234,0.46504188,0.0009518857],"about_ca_topic_score_codex":0.000100893885,"about_ca_topic_score_gemma":9.920442e-7,"teacher_disagreement_score":0.3863049,"about_ca_system_score_codex":0.0000915437,"about_ca_system_score_gemma":0.00013597745,"threshold_uncertainty_score":0.99996376},"labels":[],"label_agreement":null},{"id":"W2039790015","doi":"10.1037/a0026539","title":"Influence of familiar features on diagnosis: Instantiated features in an applied setting.","year":2012,"lang":"en","type":"article","venue":"Journal of Experimental Psychology Applied","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Hamilton Health Sciences","funders":"","keywords":"Categorization; Medical diagnosis; Categorical variable; Context (archaeology); Feature (linguistics); Consistency (knowledge bases); Matching (statistics); Cognitive psychology; Psychology; Task (project management); Unconscious mind; Process (computing); Computer science; Artificial intelligence; Natural language processing; Machine learning; Medicine; Mathematics; Statistics","score_opus":0.02334006699865672,"score_gpt":0.39004164604782776,"score_spread":0.36670157904917106,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2039790015","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.988128,0.0015024457,0.000003417885,0.0002730693,0.000263826,0.00031067585,0.0000048792176,0.000018365412,0.00949532],"genre_scores_gemma":[0.9948071,0.00013730211,0.00095745985,0.0037483526,0.00026287267,0.00003813828,0.000009933295,0.000028439266,0.000010359871],"study_design_codex":"bench_or_experimental","study_design_gemma":"observational","domain_scores_codex":[0.99807537,0.00007540164,0.00083356246,0.000250524,0.000379433,0.0003857092],"domain_scores_gemma":[0.9975137,0.0012484948,0.0005052397,0.00033519283,0.00006983899,0.0003275519],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.000652774,0.00023946981,0.0007130587,0.00028481716,0.00003738366,0.000008102327,0.00021549582,0.00032008335,0.000051846488],"category_scores_gemma":[0.001008896,0.00018854455,0.00013159959,0.00025841955,0.00021835926,0.00009135184,0.00003549238,0.0007327038,0.000015952864],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.02032535,0.019493468,0.29222894,0.000042793738,0.00035015604,0.00019311799,0.0052865674,0.00030046355,0.6309644,0.0074048038,0.011823308,0.011586652],"study_design_scores_gemma":[0.005853351,0.0017588639,0.827033,0.0005061341,0.000058421956,0.00015902995,0.0010249778,5.4462646e-7,0.16313544,0.00014107124,0.00016105172,0.0001681001],"about_ca_topic_score_codex":0.000007752158,"about_ca_topic_score_gemma":0.0000017157184,"teacher_disagreement_score":0.5348041,"about_ca_system_score_codex":0.00006764152,"about_ca_system_score_gemma":0.00005048129,"threshold_uncertainty_score":0.76886237},"labels":[],"label_agreement":null},{"id":"W2041600754","doi":"10.1097/00001888-200610001-00031","title":"The Value of Basic Science in Clinical Diagnosis","year":2006,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":162,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Recall; Test (biology); Epidemiology; Clinical epidemiology; Set (abstract data type); Basic science; Value (mathematics); Disease; Medicine; Medical science; Medical education; Psychology; Cognitive psychology; Computer science; Pathology","score_opus":0.04617998667261823,"score_gpt":0.42059427639383085,"score_spread":0.3744142897212126,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2041600754","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9680763,0.0029469635,0.00003859695,0.01971545,0.0006993154,0.0002721552,0.0000011593443,0.000025223591,0.008224859],"genre_scores_gemma":[0.991804,0.004520349,0.00006733411,0.0016979648,0.00092754175,0.000026250531,0.0000043199125,0.0000108809245,0.0009413344],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99711394,0.00009936185,0.0013216712,0.0003425041,0.00074094936,0.000381555],"domain_scores_gemma":[0.9673979,0.03158231,0.00024549474,0.00042650348,0.0001335412,0.00021423896],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.006425022,0.00012180324,0.0005080392,0.00013872923,0.00007933253,0.0000027895348,0.00034603974,0.00019328669,0.00009239862],"category_scores_gemma":[0.15617688,0.000067219604,0.00007481782,0.00086689263,0.0026114245,0.000040424944,0.00008187082,0.0008762944,0.000037916165],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007687185,0.0000617965,0.9478852,0.00001168637,0.000006471151,0.000021159232,0.000105082705,0.000007690689,0.00023173523,0.009139418,0.022915687,0.019537196],"study_design_scores_gemma":[0.0026499773,0.00045495405,0.9795549,0.0026011616,0.000086625296,0.000024479714,0.00017437022,0.00037659306,0.00057284907,0.0047505307,0.00867643,0.00007714626],"about_ca_topic_score_codex":0.0004679131,"about_ca_topic_score_gemma":0.0000073701253,"teacher_disagreement_score":0.14975186,"about_ca_system_score_codex":0.000053955962,"about_ca_system_score_gemma":0.0003276126,"threshold_uncertainty_score":0.9621903},"labels":[],"label_agreement":null},{"id":"W2042018541","doi":"10.1097/acm.0b013e31827fc58d","title":"Developing a Unified List of Physicians’ Reasoning Tasks During Clinical Encounters","year":2013,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":53,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Vocabulary; Framing (construction); Task (project management); Psychology; Analytic reasoning; Medical education; Computer science; Deductive reasoning; Medicine; Artificial intelligence; Linguistics","score_opus":0.058771497486580386,"score_gpt":0.3968433081135903,"score_spread":0.3380718106270099,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2042018541","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9813638,0.0008757353,0.00091603847,0.0114182,0.0006711653,0.00046991077,0.0000022167346,0.0001076336,0.004175323],"genre_scores_gemma":[0.98817027,0.0016922296,0.0014063154,0.00572908,0.0014211886,0.000031699925,0.00003462981,0.000037229936,0.0014773665],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99689454,0.00011337422,0.0014869545,0.0004712011,0.00057095865,0.00046296258],"domain_scores_gemma":[0.99186784,0.0066001355,0.00049698044,0.00045298866,0.0002779515,0.00030411556],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0011850842,0.00024865384,0.0010151702,0.00014604468,0.000068080626,0.000004695949,0.0002469729,0.0003915684,0.00041210614],"category_scores_gemma":[0.05491451,0.00018962623,0.00014388788,0.00042080038,0.000586801,0.000101987695,0.0000979426,0.0011843572,0.00009819747],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00066910585,0.00015842667,0.79913527,0.00069044373,0.00067593326,0.00018534884,0.00539041,0.000010413816,0.025765238,0.007994704,0.077707596,0.08161712],"study_design_scores_gemma":[0.010486816,0.0010185199,0.92898834,0.042704377,0.00061615324,0.00019886477,0.004077556,0.0005692674,0.004181691,0.0023784374,0.0042223884,0.0005575942],"about_ca_topic_score_codex":0.00046429184,"about_ca_topic_score_gemma":0.0000028768823,"teacher_disagreement_score":0.12985307,"about_ca_system_score_codex":0.00008409937,"about_ca_system_score_gemma":0.00019279386,"threshold_uncertainty_score":0.9530463},"labels":[],"label_agreement":null},{"id":"W2042618268","doi":"10.1007/s10459-007-9080-4","title":"Is analytic information processing a feature of expertise in medicine?","year":2007,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary; Foothills Medical Centre","funders":"","keywords":"Information processing; Panacea (medicine); Computer science; Context (archaeology); Flexibility (engineering); Statement (logic); Feature (linguistics); Variety (cybernetics); Data processing; Domain (mathematical analysis); Problem statement; Task (project management); Data science; Artificial intelligence; Psychology; Management science; Cognitive psychology; Medicine; Database; Mathematics","score_opus":0.02961447998334938,"score_gpt":0.4748150851829358,"score_spread":0.44520060519958643,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2042618268","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.90241224,0.04692897,0.0025335627,0.034481447,0.0010542171,0.0008447148,0.0000013679727,0.000028855055,0.011714622],"genre_scores_gemma":[0.98981136,0.0017533496,0.004293231,0.003997793,0.00007672684,0.000011748245,0.000007764979,0.0000020534285,0.000045957015],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99866146,0.000021377946,0.0006145967,0.00015407312,0.0003251956,0.00022327759],"domain_scores_gemma":[0.9984303,0.0009393568,0.00030442863,0.00010422897,0.0001048646,0.00011681626],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0018684256,0.00006603644,0.00022514763,0.00043327897,0.000046566238,0.000005859862,0.00007587471,0.00004639762,0.00001866062],"category_scores_gemma":[0.009258339,0.000050271166,0.000016737977,0.0014072023,0.00022390673,0.0005411917,0.000007277272,0.0001251314,0.0000023067576],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000034352142,0.00011916093,0.3441618,0.00024573464,3.7210694e-7,3.2775935e-7,0.0054141725,0.000060387396,0.0000044912813,0.0002523617,0.00042593948,0.6492809],"study_design_scores_gemma":[0.0010683925,0.00075877167,0.9593243,0.014926216,0.0000068518775,0.000019331379,0.010680373,0.003999774,0.00015882839,0.002189956,0.0067450497,0.00012217903],"about_ca_topic_score_codex":0.0007316928,"about_ca_topic_score_gemma":0.00028107743,"teacher_disagreement_score":0.6491587,"about_ca_system_score_codex":0.00016624454,"about_ca_system_score_gemma":0.0010526004,"threshold_uncertainty_score":0.9990871},"labels":[],"label_agreement":null},{"id":"W2043181152","doi":"10.1007/s10459-014-9576-7","title":"Does medical students’ diagnostic performance improve by observing examples of self-explanation provided by peers or experts?","year":2014,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":41,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Sherbrooke","funders":"","keywords":"Set (abstract data type); Task (project management); Active listening; Control (management); Medical education; Peer group; Psychology; Significant difference; Mathematics education; Overtime; Computer science; Medicine; Social psychology; Artificial intelligence","score_opus":0.01638464066975847,"score_gpt":0.39661374845831115,"score_spread":0.3802291077885527,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2043181152","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9882927,0.0032518506,0.00025107106,0.0058280965,0.0014795093,0.0006484699,0.0000052309783,0.00004461159,0.0001984877],"genre_scores_gemma":[0.98166686,0.013106982,0.0023302394,0.002324455,0.00016527825,0.00017182257,0.0000501383,0.000008834808,0.00017537657],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99716944,0.00013439824,0.00074844615,0.00043251654,0.0011709592,0.00034425623],"domain_scores_gemma":[0.9863911,0.012529129,0.0004180604,0.00022331333,0.00012722367,0.00031116162],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0024197318,0.00013350505,0.00033865176,0.000107025015,0.00014388267,0.000030476329,0.00036443476,0.000078547324,0.00006477805],"category_scores_gemma":[0.055081505,0.00007881655,0.000027446682,0.0004507422,0.00020398798,0.00039963462,0.00004639601,0.0001406641,0.000003044717],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004502515,0.0009772347,0.8218505,0.00050326093,0.000003879199,2.5942543e-7,0.0012965888,0.000027138418,0.000045572513,0.00012988338,0.0032578234,0.17186283],"study_design_scores_gemma":[0.009142572,0.014290853,0.73150194,0.0378831,0.00011761514,0.00004628772,0.025301736,0.028501742,0.0070138406,0.0019383298,0.14273186,0.0015301491],"about_ca_topic_score_codex":0.00096507574,"about_ca_topic_score_gemma":0.00019067114,"teacher_disagreement_score":0.17033269,"about_ca_system_score_codex":0.0001594839,"about_ca_system_score_gemma":0.001097559,"threshold_uncertainty_score":0.95287794},"labels":[],"label_agreement":null},{"id":"W2043600616","doi":"10.1097/00001888-200310001-00026","title":"The Privileged Status of Prestigious Terminology: Impact of ???Medicalese??? on Clinical Judgments","year":2003,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Seriousness; Terminology; Medical terminology; Disease; Psychology; Medical psychology; Clinical psychology; Family medicine; Medical education; Medicine; Social psychology; MEDLINE; Pathology; Nursing","score_opus":0.09667777054244836,"score_gpt":0.4720948811430062,"score_spread":0.3754171106005578,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2043600616","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98672885,0.0031898962,0.000052092666,0.0018126062,0.0007146014,0.00052754366,0.000011907745,0.000034369856,0.0069281277],"genre_scores_gemma":[0.9860732,0.01173399,0.00007377728,0.0008655429,0.00035747362,0.000015864543,0.000019822026,0.000023572842,0.0008367556],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9961149,0.00037181858,0.001726746,0.00036938817,0.00078627304,0.000630869],"domain_scores_gemma":[0.9654334,0.032301486,0.0006398086,0.00070548616,0.00017002618,0.0007497735],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0025479638,0.00023246038,0.0010176765,0.00010493886,0.000055851084,0.0000012778246,0.00026461532,0.00049373996,0.00024141604],"category_scores_gemma":[0.2800921,0.00011777293,0.0002457213,0.0002696953,0.001841921,0.000019316418,0.000048932416,0.0013783997,0.000024524506],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007876772,0.00023540927,0.90101427,0.000039965744,0.00038123078,0.00004469215,0.00036558157,0.0000041523112,0.00050474383,0.0021395278,0.06426344,0.030219328],"study_design_scores_gemma":[0.013638875,0.013663346,0.94141513,0.0060232333,0.00071490504,0.00014712321,0.0003174028,0.00008385774,0.0009851339,0.0038149552,0.01897071,0.00022532455],"about_ca_topic_score_codex":0.00010895488,"about_ca_topic_score_gemma":0.0000017069144,"teacher_disagreement_score":0.2775441,"about_ca_system_score_codex":0.00007374989,"about_ca_system_score_gemma":0.00067341723,"threshold_uncertainty_score":0.725972},"labels":[],"label_agreement":null},{"id":"W2044089605","doi":"10.1051/pmed/2012009","title":"Aborder le raisonnement clinique du point de vue pédagogique","year":2012,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Humanities; Philosophy","score_opus":0.039356311083504326,"score_gpt":0.36381620913121865,"score_spread":0.3244598980477143,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2044089605","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3982676,0.20325738,0.039880488,0.23486461,0.017563542,0.0022304354,0.00022610986,0.0008738465,0.10283601],"genre_scores_gemma":[0.8442542,0.027145552,0.010718736,0.022635687,0.01601852,0.00027630504,0.0001402672,0.00022126322,0.07858948],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9939491,0.00076280936,0.0013598906,0.0007429188,0.00070184085,0.0024834105],"domain_scores_gemma":[0.97879654,0.016951386,0.00032422383,0.0010361396,0.00024884575,0.00264288],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.004763875,0.0006192605,0.001081337,0.00014070707,0.0002868684,0.00008299341,0.0004205311,0.0010297544,0.009857202],"category_scores_gemma":[0.07338615,0.00059604633,0.0005973869,0.00040923874,0.0008038801,0.00026800347,0.00033123075,0.0014482298,0.0030281793],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004271191,0.0137556745,0.22814932,0.00046647113,0.0005690587,0.00058021746,0.012068621,0.000069218026,0.00047969216,0.019369483,0.43802917,0.28603595],"study_design_scores_gemma":[0.005734955,0.00094307546,0.07072619,0.00467687,0.00066591153,0.0003728625,0.004562546,0.0008373698,0.0024607095,0.0022457237,0.9058205,0.000953298],"about_ca_topic_score_codex":0.0009767693,"about_ca_topic_score_gemma":0.00006615445,"teacher_disagreement_score":0.46779132,"about_ca_system_score_codex":0.00046045092,"about_ca_system_score_gemma":0.0017280233,"threshold_uncertainty_score":0.9996491},"labels":[],"label_agreement":null},{"id":"W2046004104","doi":"10.1007/s11606-008-0623-z","title":"Tips for Teachers of Evidence-based Medicine: Clinical Prediction Rules (CPRs) and Estimating Pretest Probability","year":2008,"lang":"en","type":"article","venue":"Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":29,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Columbia College","funders":"","keywords":"Medicine; Test (biology); Clinical Practice; Medical education; Interpretation (philosophy); Field (mathematics); Medical physics; Family medicine; Computer science","score_opus":0.2599864484475516,"score_gpt":0.449641863592394,"score_spread":0.1896554151448424,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2046004104","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9639041,0.002645112,0.019588688,0.011614435,0.0015721109,0.00041733956,0.000009255817,0.000019515379,0.0002294755],"genre_scores_gemma":[0.9356204,0.0010049032,0.055800818,0.0011111355,0.0058872863,0.000010519288,0.000010629432,0.00002638309,0.00052791677],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9962314,0.00020074584,0.0023513886,0.0002810859,0.00070617173,0.00022925665],"domain_scores_gemma":[0.9810169,0.015852196,0.001340946,0.00025446483,0.0010102912,0.00052522717],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004256903,0.0002173589,0.001189797,0.00018696948,0.00006480099,0.0000046063697,0.00017115263,0.00017709292,0.000088160356],"category_scores_gemma":[0.18243922,0.00012065134,0.00026452224,0.00012294763,0.0011635377,0.00013246847,0.000032265223,0.00049238885,9.216852e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0041521206,0.0006172256,0.86108124,0.00038033474,0.00037856027,0.00015502369,0.0008833556,0.0001716506,0.0034655405,0.0000849776,0.056002222,0.07262776],"study_design_scores_gemma":[0.026075365,0.050152745,0.7799181,0.08093987,0.0023106015,0.004676701,0.00023485099,0.047595654,0.002166705,0.002569259,0.0030176556,0.00034253314],"about_ca_topic_score_codex":0.0001836782,"about_ca_topic_score_gemma":0.0000035027292,"teacher_disagreement_score":0.17818232,"about_ca_system_score_codex":0.00008647337,"about_ca_system_score_gemma":0.00032554328,"threshold_uncertainty_score":0.82444745},"labels":[],"label_agreement":null},{"id":"W2046626105","doi":"10.1097/acm.0b013e318253acbd","title":"The Relationship Between Response Time and Diagnostic Accuracy","year":2012,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":137,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"Canadian Institutes of Health Research","keywords":"Diagnostic accuracy; Competence (human resources); Test (biology); Cognition; Correlation; Unconscious mind; Medicine; Psychology; Clinical psychology; Cognitive psychology; Social psychology; Psychiatry; Internal medicine","score_opus":0.08749390220777292,"score_gpt":0.4114050314457388,"score_spread":0.3239111292379659,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2046626105","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93604666,0.011385182,0.00004882509,0.051068146,0.00023344514,0.00032484272,0.000003909439,0.00007350253,0.00081547874],"genre_scores_gemma":[0.99061555,0.0010939351,0.00003655791,0.0017060213,0.0020502997,0.000022155662,0.00003037415,0.000021345248,0.0044237366],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982247,0.00026968386,0.00051267346,0.00019251776,0.00037582708,0.00042464404],"domain_scores_gemma":[0.61499,0.3840677,0.00011953481,0.00027745325,0.000041588784,0.00050372945],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004239299,0.00015262073,0.00033741086,0.00006687664,0.00021623484,0.0000053965955,0.00011912439,0.00025252238,0.000111773625],"category_scores_gemma":[0.75605583,0.00008582494,0.000036053254,0.00022539205,0.00047898173,0.00008517912,0.00006739464,0.0009269862,0.00039077262],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00022866634,0.000007054112,0.9599482,0.000007332822,0.000026747028,0.0000054815914,0.0008078531,6.691253e-8,0.00009060703,0.0009262032,0.032569963,0.0053818477],"study_design_scores_gemma":[0.0013891389,0.00023645596,0.9723484,0.0013882186,0.00033446014,0.00009151546,0.00016892649,0.00000870997,0.000032618635,0.0014172924,0.022490805,0.00009344101],"about_ca_topic_score_codex":0.000016748048,"about_ca_topic_score_gemma":2.0516225e-7,"teacher_disagreement_score":0.7518165,"about_ca_system_score_codex":0.000039555172,"about_ca_system_score_gemma":0.00007032826,"threshold_uncertainty_score":0.5022722},"labels":[],"label_agreement":null},{"id":"W2047540418","doi":"10.1007/s10459-006-9057-8","title":"The effect of question format and task difficulty on reasoning strategies and diagnostic performance in Internal Medicine residents","year":2007,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":30,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Calgary General Hospital; Foothills Medical Centre; University of Calgary","funders":"","keywords":"Task (project management); Odds; Association (psychology); Verbal reasoning; Think aloud protocol; Deductive reasoning; Logical reasoning; Matching (statistics); Inductive reasoning; Analytic reasoning; MEDLINE; Cognitive psychology; Psychology; Logistic regression; Computer science; Mathematics education; Cognition; Artificial intelligence; Medicine; Machine learning; Pathology","score_opus":0.010506507166991651,"score_gpt":0.4108181503220709,"score_spread":0.4003116431550793,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2047540418","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98347783,0.013850589,0.00004963439,0.001099997,0.0004041995,0.00031687444,2.5790646e-7,0.0000065044114,0.0007940878],"genre_scores_gemma":[0.9775165,0.022084422,0.00017935896,0.00011457555,0.000069260575,0.000017459659,0.0000029063417,0.0000025351367,0.000013037572],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9988191,0.000090611524,0.00044485935,0.00018676517,0.00024666084,0.00021200963],"domain_scores_gemma":[0.98361254,0.015968215,0.0002050691,0.0000863448,0.000031618576,0.0000962322],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0036476708,0.00008038837,0.00019134086,0.00013496206,0.00013019255,0.000017953751,0.000062517836,0.000030081515,9.194417e-7],"category_scores_gemma":[0.019985337,0.000046102836,0.000008055143,0.00029591357,0.00046017696,0.00030508018,0.000015438254,0.00015312275,4.1189992e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013644743,0.000030591003,0.7666723,0.00016130456,5.7248127e-7,5.8019793e-7,0.0006274811,0.00011380604,0.000003605317,0.0011644247,0.000013704023,0.23107517],"study_design_scores_gemma":[0.0005865542,0.0024373126,0.98268497,0.011161999,0.000003783864,0.000022507962,0.0017664838,0.0006600528,0.00006261954,0.00049369223,0.00007949149,0.0000405082],"about_ca_topic_score_codex":0.0013217295,"about_ca_topic_score_gemma":0.0006856922,"teacher_disagreement_score":0.23103467,"about_ca_system_score_codex":0.00007849393,"about_ca_system_score_gemma":0.00017272864,"threshold_uncertainty_score":0.98826975},"labels":[],"label_agreement":null},{"id":"W2049866975","doi":"10.12968/hosp.2004.65.3.12388","title":"The Ottawa rules for ankle sprains","year":2004,"lang":"en","type":"editorial","venue":"Hospital Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Ankle; Medial malleolus; Radiography; Foot (prosody); Emergency department; Lateral malleolus; Ankle sprain; Physical therapy; Orthodontics; Surgery","score_opus":0.011215180777484055,"score_gpt":0.3327398301271613,"score_spread":0.3215246493496772,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2049866975","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0008428705,0.008010645,0.0001723585,0.029335238,0.9537093,0.0018385377,0.00028382754,0.00017331942,0.005633917],"genre_scores_gemma":[0.0016799251,0.0025729907,0.0005571502,0.0005535055,0.98225975,0.00032034784,0.0022516225,0.00012887991,0.009675828],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9964463,0.000036831076,0.0008944573,0.0006698619,0.0012866848,0.0006658258],"domain_scores_gemma":[0.93456846,0.062729076,0.00043851434,0.00096365274,0.00076428737,0.0005359899],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0012557176,0.00049019296,0.0011403917,0.000087008266,0.00027463655,0.0000367686,0.00037569256,0.0009540958,0.00010106838],"category_scores_gemma":[0.38455477,0.00027223595,0.0003706599,0.00015658319,0.0007734355,0.00002976229,0.00007648811,0.0011248029,0.00014076069],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00016554017,0.00022010425,0.00009431215,0.00013932114,0.00016681911,0.00007671317,0.00013790482,3.404608e-7,0.0000016883639,0.0009431223,0.9852117,0.012842455],"study_design_scores_gemma":[0.0064492566,0.0057129893,0.00083207025,0.008229125,0.0007399911,0.0000072864004,0.00012774655,0.000005736621,0.0000064384467,0.0035178133,0.9740698,0.00030176525],"about_ca_topic_score_codex":0.0002555417,"about_ca_topic_score_gemma":0.000042183812,"teacher_disagreement_score":0.38329905,"about_ca_system_score_codex":0.00023152196,"about_ca_system_score_gemma":0.0010260035,"threshold_uncertainty_score":0.999973},"labels":[],"label_agreement":null},{"id":"W2051990295","doi":"10.5737/1181912x133157163","title":"L’impact de l’administration d’iode radioactif depuis la perspective des patients : les répercussions pour la pratique","year":2003,"lang":"fr","type":"article","venue":"Canadian Oncology Nursing Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"Sunnybrook Health Science Centre","funders":"","keywords":"Humanities; Gynecology; Medicine; Philosophy","score_opus":0.026283949201975965,"score_gpt":0.3906844828250672,"score_spread":0.3644005336230912,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2051990295","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.70811594,0.014835628,0.004257778,0.0490114,0.0038483317,0.00046933405,0.00012124561,0.000038930586,0.21930145],"genre_scores_gemma":[0.9881785,0.0020469595,0.005960271,0.00069045054,0.0010395868,0.000008547189,0.000016888333,0.000057303616,0.0020014716],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99514645,0.0025328447,0.0005620792,0.00034384575,0.00022768245,0.0011870958],"domain_scores_gemma":[0.9838052,0.0112671815,0.00031413985,0.0002184922,0.0006744798,0.0037204989],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0017706065,0.0003313231,0.0005385417,0.0003554265,0.0008632249,0.00016609451,0.00014710898,0.001051957,0.0013480695],"category_scores_gemma":[0.088019475,0.00030629052,0.0002942914,0.00028388252,0.001104363,0.00017169239,0.000005051603,0.002070369,0.000059791866],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.00033039218,0.0036075616,0.32330617,0.000020575424,0.00053446775,0.0057404563,0.020388382,0.00015389323,0.000088017136,0.02048499,0.09543016,0.5299149],"study_design_scores_gemma":[0.0064186077,0.0047314093,0.8459723,0.016959852,0.0014058508,0.03193087,0.020561371,0.00044552848,0.00012790115,0.015664143,0.05513602,0.0006461405],"about_ca_topic_score_codex":0.012769595,"about_ca_topic_score_gemma":0.022187505,"teacher_disagreement_score":0.5292688,"about_ca_system_score_codex":0.026217159,"about_ca_system_score_gemma":0.05109415,"threshold_uncertainty_score":0.9999389},"labels":[],"label_agreement":null},{"id":"W2053033682","doi":"10.1007/s10459-014-9518-4","title":"Reexamining our bias against heuristics","year":2014,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":48,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; University of British Columbia; University of Calgary","funders":"","keywords":"Heuristics; Social heuristics; Computer science; Cognitive bias; Cognition; Artificial intelligence; Machine learning; Data science; Psychology","score_opus":0.059348928753375274,"score_gpt":0.45470787648397154,"score_spread":0.39535894773059627,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2053033682","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9203392,0.009675314,0.002043769,0.023964472,0.005654456,0.0005836561,0.000002195829,0.00010461707,0.037632305],"genre_scores_gemma":[0.9747669,0.003361646,0.014989749,0.006169531,0.0004369565,0.000022015445,0.000011054177,0.0000059726617,0.00023616981],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9985599,0.000101234546,0.0004495051,0.00031031988,0.00028316464,0.00029588264],"domain_scores_gemma":[0.99507546,0.004166039,0.00023352967,0.00020876917,0.000078637175,0.00023753711],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0020508857,0.00008330098,0.00021159044,0.00014123315,0.00014361745,0.000020319569,0.00010877089,0.0000385471,0.0000062566482],"category_scores_gemma":[0.059341654,0.00006941717,0.000024493964,0.0005489449,0.00010444974,0.00014484284,0.000016928192,0.00014911302,0.000027582944],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00000870504,0.00019853032,0.40700075,0.00008326198,5.9480516e-7,4.5896647e-7,0.00023089048,0.00048901595,0.0000021436836,0.002776143,0.00096239,0.5882471],"study_design_scores_gemma":[0.0012512255,0.0027010918,0.777935,0.015633237,0.0000164326,0.0000430028,0.0060871043,0.015162551,0.000082603394,0.012175287,0.16843793,0.0004745429],"about_ca_topic_score_codex":0.00013582227,"about_ca_topic_score_gemma":0.000060567665,"teacher_disagreement_score":0.58777255,"about_ca_system_score_codex":0.00009601834,"about_ca_system_score_gemma":0.0009746559,"threshold_uncertainty_score":0.94858193},"labels":[],"label_agreement":null},{"id":"W2053487491","doi":"10.1097/00001888-200210001-00022","title":"Believing Is Seeing","year":2002,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":129,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University Health Network","funders":"","keywords":"Context (archaeology); Identification (biology); Presentation (obstetrics); Perception; Illusion; Cognitive psychology; Feature (linguistics); Psychology; Representation (politics); Disease; Medicine; Pathology; Radiology; Linguistics","score_opus":0.08190847257802468,"score_gpt":0.3761184170139818,"score_spread":0.2942099444359571,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2053487491","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7312368,0.025341997,0.000599508,0.12347438,0.0010353308,0.00046122674,0.000003466353,0.0003741494,0.11747314],"genre_scores_gemma":[0.9452647,0.0052195895,0.00017148355,0.029139956,0.0017367468,0.000009340379,0.000009335272,0.000026565542,0.018422268],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99828696,0.00002766631,0.0005174771,0.00033190165,0.00047700587,0.0003590015],"domain_scores_gemma":[0.99478245,0.0042815176,0.00010685121,0.00033261743,0.000068191745,0.0004284026],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0005156949,0.00016506795,0.00045741236,0.00010284714,0.000067357054,0.0000025591853,0.00012562478,0.00025882362,0.006870193],"category_scores_gemma":[0.039704856,0.00011874233,0.00006658869,0.00028733825,0.0001848043,0.000043432134,0.000045632925,0.00087522535,0.0008356152],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000036084526,0.000038648213,0.12862623,0.00006306394,0.000082999904,0.00018601322,0.0053607053,0.0000010868224,0.001587432,0.00088802783,0.77701044,0.08611928],"study_design_scores_gemma":[0.020581536,0.002860305,0.11346702,0.043100517,0.0018503825,0.0022420494,0.003010416,0.010265362,0.0031415795,0.0049836226,0.7932189,0.0012782812],"about_ca_topic_score_codex":0.000036601163,"about_ca_topic_score_gemma":2.3989782e-7,"teacher_disagreement_score":0.21402791,"about_ca_system_score_codex":0.00004228336,"about_ca_system_score_gemma":0.000018462759,"threshold_uncertainty_score":0.99994236},"labels":[],"label_agreement":null},{"id":"W2054627871","doi":"10.1016/j.amjsurg.2011.11.006","title":"The script concordance test as a measure of clinical reasoning: a national validation study","year":2012,"lang":"en","type":"article","venue":"The American Journal of Surgery","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":71,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia; University of Ottawa; Western University; Queen's University; McMaster University; University of Alberta; Université de Montréal; Memorial University of Newfoundland; University of Calgary; McGill University","funders":"","keywords":"Concordance; Cronbach's alpha; Test (biology); Reliability (semiconductor); Measure (data warehouse); Medicine; Medical physics; Psychology; Clinical psychology; Psychometrics; Data mining; Internal medicine; Computer science","score_opus":0.10857744296837835,"score_gpt":0.4208620363548373,"score_spread":0.312284593386459,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2054627871","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9947141,0.0007067644,0.000041967796,0.002999598,0.0007272745,0.00013012136,0.0000027449407,0.000007889534,0.0006695787],"genre_scores_gemma":[0.99757737,0.0003109306,0.00014850423,0.0007279091,0.0011015616,0.0000044446438,0.0000011165823,0.000014189431,0.000113976035],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99669623,0.00061286805,0.0013250378,0.00009703398,0.0010371333,0.00023168548],"domain_scores_gemma":[0.86898243,0.12581778,0.003082316,0.00039090237,0.0013528259,0.0003737496],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.014168198,0.00011043155,0.0007486513,0.000066579516,0.00011099466,0.000018482411,0.00017253823,0.000027597378,0.000022905377],"category_scores_gemma":[0.25954235,0.00005511891,0.0003840884,0.00036140456,0.00061255693,0.00008652189,0.000032107735,0.00045417968,0.000016346892],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008713523,0.0010529043,0.9523399,0.0000020527755,0.00020380892,0.000013573657,0.00023758197,0.0000032764663,0.000041866035,0.00006514324,0.005102615,0.040065944],"study_design_scores_gemma":[0.000677503,0.0009288355,0.9938265,0.0005931979,0.00023542675,0.00037374077,0.0019538095,0.0000135082,0.00009539622,0.0001149373,0.001120737,0.000066405366],"about_ca_topic_score_codex":0.00005906995,"about_ca_topic_score_gemma":8.4846073e-7,"teacher_disagreement_score":0.24537416,"about_ca_system_score_codex":0.000040655912,"about_ca_system_score_gemma":0.00076987426,"threshold_uncertainty_score":0.74669486},"labels":[],"label_agreement":null},{"id":"W2054837185","doi":"10.1111/medu.12302","title":"Test‐enhanced learning and its effect on comprehension and diagnostic accuracy","year":2014,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":79,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Memorial University of Newfoundland; The Wilson Centre; University of Toronto","funders":"","keywords":"Test (biology); Comprehension; Psychology; Educational measurement; Computer science; Medicine; Curriculum; Pedagogy; Biology","score_opus":0.009704307280209934,"score_gpt":0.3583462339113984,"score_spread":0.3486419266311885,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2054837185","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99102014,0.0006396182,0.00008436534,0.005772832,0.00032760162,0.00024355682,2.5805517e-7,0.000056434314,0.0018552094],"genre_scores_gemma":[0.9950945,0.0005423735,0.00006440981,0.003595456,0.00036185049,0.00003417271,0.000029755474,0.000012381288,0.00026511328],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99895144,0.000121733035,0.00019088252,0.00026222147,0.00032449586,0.00014923127],"domain_scores_gemma":[0.9062041,0.093037,0.000079550395,0.00013223106,0.00008357173,0.00046354032],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0005054049,0.0001111231,0.00024013095,0.000048559337,0.000078873156,0.000016691822,0.000030676212,0.00012859398,0.0001655977],"category_scores_gemma":[0.6486655,0.00008239048,0.000025291625,0.000081571874,0.000068328314,0.000026798423,0.000029869452,0.00033700207,0.000078113684],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007101214,0.0004983216,0.06489488,0.000113604445,0.000008946924,0.000003935281,0.00014339758,9.695962e-7,0.00065715273,0.00023082856,0.0028608025,0.9305161],"study_design_scores_gemma":[0.0048343083,0.0061623706,0.9424399,0.011951902,0.00022227061,0.00012534246,0.00006249571,0.0067128795,0.003539034,0.00039361362,0.02323541,0.0003204697],"about_ca_topic_score_codex":0.000017632368,"about_ca_topic_score_gemma":0.0000014084823,"teacher_disagreement_score":0.9301957,"about_ca_system_score_codex":0.000018222821,"about_ca_system_score_gemma":0.00015271944,"threshold_uncertainty_score":0.35429394},"labels":[],"label_agreement":null},{"id":"W2055373802","doi":"10.1111/j.1365-2929.2004.02036.x","title":"The value of basic science in clinical diagnosis: creating coherence among signs and symptoms","year":2004,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":218,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medical diagnosis; Mnemonic; Recall; Disease; Test (biology); Psychology; Lesion; Value (mathematics); Clinical diagnosis; Medicine; Physical therapy; Audiology; Cognitive psychology; Physical medicine and rehabilitation; Clinical psychology; Pathology; Psychiatry; Computer science; Machine learning","score_opus":0.017899991308198502,"score_gpt":0.38844421463035667,"score_spread":0.37054422332215814,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2055373802","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9916932,0.0005004065,0.000054540356,0.005813873,0.00048473774,0.00023237205,3.9574783e-7,0.000013774942,0.0012066734],"genre_scores_gemma":[0.99772954,0.00041357605,0.0002857908,0.0012647843,0.0001725823,0.00006959973,0.0000026787634,0.0000057240673,0.000055744295],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982689,0.00007366992,0.0005873715,0.00026195883,0.0006043257,0.00020376843],"domain_scores_gemma":[0.9867835,0.012221728,0.00016778205,0.00028746703,0.00015882093,0.00038068646],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002952776,0.00007535416,0.00022092994,0.00005880843,0.00010560764,0.00002088337,0.00015708353,0.00010866694,0.000055245444],"category_scores_gemma":[0.28464535,0.000048881382,0.00004340549,0.0004014687,0.0016266761,0.00005734825,0.000053941872,0.00031739802,0.000008317322],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000016145028,0.00064912846,0.837713,0.000012729228,0.0000042125116,0.0000029316234,0.0002982054,0.0000051390907,0.000009642148,0.0014668179,0.00017612566,0.15964592],"study_design_scores_gemma":[0.00097320415,0.00040063114,0.9929956,0.0035825826,0.000030435616,0.000010126185,0.00032230484,0.00031656685,0.00013417665,0.0010536144,0.00012519675,0.000055526325],"about_ca_topic_score_codex":0.00059533934,"about_ca_topic_score_gemma":0.00004264165,"teacher_disagreement_score":0.28169256,"about_ca_system_score_codex":0.00006813643,"about_ca_system_score_gemma":0.0032674884,"threshold_uncertainty_score":0.7213804},"labels":[],"label_agreement":null},{"id":"W2055476838","doi":"10.1097/acm.0b013e3181ace703","title":"A Universal Model of Diagnostic Reasoning","year":2009,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":970,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University; Queen Elizabeth II Health Sciences Centre","funders":"","keywords":"Management science; Variety (cybernetics); Process (computing); Computer science; Heuristic; Schematic; Psychology; Risk analysis (engineering); Artificial intelligence; Medicine","score_opus":0.04037542057166423,"score_gpt":0.35484391496390244,"score_spread":0.31446849439223823,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2055476838","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95104116,0.0025565112,0.0062371576,0.017463686,0.00019584857,0.00035200882,0.0000054258667,0.0001381877,0.022009987],"genre_scores_gemma":[0.9910243,0.0021202732,0.0008245173,0.0038990197,0.0004121528,0.000002925475,0.000024541989,0.00001376693,0.00167849],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9983995,0.000031526157,0.0005510029,0.00027599485,0.00044949158,0.00029251582],"domain_scores_gemma":[0.9893796,0.009639597,0.00017861556,0.00031333862,0.00012079517,0.00036805344],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0005661828,0.00016250048,0.0005780392,0.0001301661,0.00003142701,9.952696e-7,0.00014464588,0.000248713,0.00014895145],"category_scores_gemma":[0.13476777,0.0001218136,0.00007389183,0.00027652932,0.00022856389,0.00004609622,0.00002503614,0.000642069,0.00002154028],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0026790455,0.00072005333,0.17618883,0.00033246836,0.00048776937,0.0008950624,0.018703835,0.0035492375,0.061470274,0.20949677,0.2986266,0.22685008],"study_design_scores_gemma":[0.0695269,0.021932708,0.32860202,0.12908952,0.0073787733,0.0015623198,0.007520066,0.27787635,0.015753062,0.12835753,0.00970098,0.0026997882],"about_ca_topic_score_codex":0.000039635524,"about_ca_topic_score_gemma":3.6092084e-7,"teacher_disagreement_score":0.28892562,"about_ca_system_score_codex":0.000045772624,"about_ca_system_score_gemma":0.0001567223,"threshold_uncertainty_score":0.87252045},"labels":[],"label_agreement":null},{"id":"W2055818802","doi":"10.1136/bmjqs-2013-002293.72","title":"041 Making Recommendations About Diagnostic Tests And Strategies: What Do Experts Say?","year":2013,"lang":"en","type":"article","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Data science; Medical physics; Medical education; Management science; Computer science; Engineering","score_opus":0.06567332707782009,"score_gpt":0.4446444415799868,"score_spread":0.3789711145021667,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2055818802","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8750278,0.006629285,0.0023086886,0.0923912,0.0014764486,0.0029845631,0.00006267736,0.00041594033,0.018703409],"genre_scores_gemma":[0.98392004,0.007697646,0.0020323223,0.005205363,0.00041058118,0.00020673957,0.00012490898,0.000030941603,0.00037144808],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9973947,0.00026682758,0.0011310092,0.00047714493,0.0003323464,0.000397975],"domain_scores_gemma":[0.9473979,0.05102056,0.00026145714,0.0006671769,0.00024542044,0.0004075046],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012879088,0.00023653045,0.00051967875,0.00006387439,0.00018984475,0.0003417772,0.00010617762,0.0001784523,0.0014955318],"category_scores_gemma":[0.075546086,0.00019582371,0.00012714475,0.0001863687,0.00017093463,0.0005528139,0.00010415326,0.00025128468,0.00022577867],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018928229,0.0008898311,0.12850861,0.0002282779,0.00016816227,0.000067039844,0.004364921,0.000016774256,0.00015658424,0.011274163,0.058032725,0.7961036],"study_design_scores_gemma":[0.0026309004,0.0002534216,0.9382041,0.010628915,0.00013740802,0.000090724476,0.021661004,0.00012337745,0.000022920623,0.006768619,0.018948987,0.000529638],"about_ca_topic_score_codex":0.00061244366,"about_ca_topic_score_gemma":0.00009958849,"teacher_disagreement_score":0.8096955,"about_ca_system_score_codex":0.000077455756,"about_ca_system_score_gemma":0.0002541732,"threshold_uncertainty_score":0.99941725},"labels":[],"label_agreement":null},{"id":"W2056263664","doi":"10.1007/s10459-013-9490-4","title":"Are we willing to change our impression of first impressions?","year":2014,"lang":"en","type":"letter","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary; Health Sciences Centre","funders":"","keywords":"Impression; Precedent; Impression formation; Psychology; Dentistry; Medicine; Perception; Computer science; Social perception; World Wide Web; Political science","score_opus":0.09067901045991401,"score_gpt":0.4699526615831845,"score_spread":0.3792736511232705,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2056263664","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004889618,0.025896303,0.000110620465,0.96408784,0.0037049425,0.001025791,0.000011928344,0.000030285828,0.00024266317],"genre_scores_gemma":[0.117525265,0.028975153,0.006217234,0.83985513,0.0064913016,0.0004649039,0.00009179018,0.00003451732,0.00034471974],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99727255,0.00012563741,0.0008366771,0.00064381724,0.000623039,0.00049830426],"domain_scores_gemma":[0.99474406,0.0032531118,0.001156942,0.0004260239,0.0001632762,0.00025657998],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010792546,0.00021783881,0.0006604444,0.00036286915,0.00021999372,0.000024139268,0.00029924407,0.00028921242,0.000021823409],"category_scores_gemma":[0.0126217585,0.00016766116,0.00008024512,0.0007081728,0.00013436015,0.00019170025,0.00006102991,0.0007049808,0.000014418352],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000290563,0.00024747092,0.1128423,0.0024316937,0.0000029553714,0.00000514285,0.0009079586,0.00015194138,0.0000020870605,0.00001498663,0.73822856,0.14513582],"study_design_scores_gemma":[0.00024852334,0.0005133757,0.014496964,0.06308511,0.000013685106,0.000012563949,0.00066369714,0.00020498889,0.000021499469,0.0005219298,0.92001945,0.00019819719],"about_ca_topic_score_codex":0.0005988823,"about_ca_topic_score_gemma":0.00014595196,"teacher_disagreement_score":0.18179089,"about_ca_system_score_codex":0.0001696994,"about_ca_system_score_gemma":0.00083621644,"threshold_uncertainty_score":0.99569535},"labels":[],"label_agreement":null},{"id":"W2056826270","doi":"10.1007/s10459-011-9279-2","title":"Optimization of answer keys for script concordance testing: should we exclude deviant panelists, deviant responses, or neither?","year":2011,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":34,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Université de Montréal","funders":"","keywords":"Concordance; Reliability (semiconductor); Outlier; Correlation; Test (biology); Statistics; Psychology; Type I and type II errors; Concordance correlation coefficient; Social psychology; Clinical psychology; Medicine; Power (physics); Mathematics","score_opus":0.19052185520261392,"score_gpt":0.4576104457242404,"score_spread":0.2670885905216265,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2056826270","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.62349415,0.12665465,0.18363301,0.030006822,0.013346653,0.010645906,0.00011737192,0.00037873338,0.011722688],"genre_scores_gemma":[0.60983545,0.008316662,0.37811244,0.0027970443,0.00019986104,0.00022377423,0.00002081866,0.000018556066,0.00047537385],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9979549,0.0001133711,0.0008271554,0.00043501414,0.0003007453,0.00036881331],"domain_scores_gemma":[0.992793,0.0058702105,0.00053466245,0.00026339584,0.00032208065,0.00021665655],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001739339,0.00014070368,0.00036514952,0.00018111938,0.00014635299,0.000012876336,0.00018775385,0.000067060144,0.00005952101],"category_scores_gemma":[0.039084304,0.0001041917,0.000040545936,0.0009255607,0.00036366432,0.00029215773,0.000019393463,0.000119107055,0.0000025970332],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018049543,0.0017660209,0.5081591,0.0012215863,0.000008871423,0.0000044379367,0.0034159087,0.00472021,0.00007064585,0.00408996,0.0013974505,0.47334084],"study_design_scores_gemma":[0.007972113,0.022291314,0.7085378,0.056268133,0.00017267006,0.000289765,0.015768584,0.079740755,0.0032381406,0.018147472,0.08594899,0.0016242502],"about_ca_topic_score_codex":0.00077543285,"about_ca_topic_score_gemma":0.00024885265,"teacher_disagreement_score":0.47171658,"about_ca_system_score_codex":0.00013186829,"about_ca_system_score_gemma":0.0032097842,"threshold_uncertainty_score":0.9690099},"labels":[],"label_agreement":null},{"id":"W2056983349","doi":"10.1007/s10459-011-9338-8","title":"Script concordance test item response process: The argument for probability versus typicality","year":2011,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; McGill University","funders":"","keywords":"Concordance; Test (biology); Argument (complex analysis); Psychology; Statistics; Computer science; Mathematics; Medicine","score_opus":0.08500634466955383,"score_gpt":0.45516326432591964,"score_spread":0.3701569196563658,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2056983349","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97510296,0.0046992055,0.00029872134,0.014241815,0.0021181423,0.002041994,0.000007712617,0.00004072291,0.0014487497],"genre_scores_gemma":[0.98808825,0.00047375672,0.00859775,0.0020995152,0.000118653035,0.0005224678,0.0000046774708,0.0000045886736,0.00009031741],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99842227,0.00013788458,0.00048635568,0.00039411488,0.00025046588,0.00030892002],"domain_scores_gemma":[0.9782645,0.020662334,0.0002957001,0.00034914506,0.00022553705,0.00020274604],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00380654,0.00009257704,0.00019061218,0.000044972414,0.00025068535,0.000014076555,0.00018808345,0.000036275473,0.000025546711],"category_scores_gemma":[0.09865811,0.000059651575,0.00003652367,0.00048817525,0.00058344984,0.00018972217,0.00001526909,0.00012975726,0.0000064324827],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004281002,0.0023854414,0.845408,0.00048261945,0.000003724941,4.1443872e-7,0.0038269847,0.000040142884,0.000008098119,0.006844318,0.0007077318,0.13601153],"study_design_scores_gemma":[0.001996649,0.0054277983,0.9150135,0.0023004087,0.000019361709,0.000008263402,0.004585445,0.0007042013,0.0003238777,0.050205555,0.019211264,0.00020366251],"about_ca_topic_score_codex":0.00019128883,"about_ca_topic_score_gemma":0.00016038414,"teacher_disagreement_score":0.13580786,"about_ca_system_score_codex":0.0002031604,"about_ca_system_score_gemma":0.0025301906,"threshold_uncertainty_score":0.9089343},"labels":[],"label_agreement":null},{"id":"W2057438748","doi":"10.1192/pb.25.6.234","title":"The use of the Royal College of Psychiatrists' trainee's log book: a cross-sectional survey of trainees and trainers","year":2001,"lang":"en","type":"article","venue":"Psychiatric Bulletin","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Formative assessment; Quarter (Canadian coin); Medical education; Medicine; Psychology; Cross-sectional study; Family medicine; Pedagogy; History","score_opus":0.0415995336747196,"score_gpt":0.3142388364177072,"score_spread":0.27263930274298764,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2057438748","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9891302,0.0031575235,0.00008936208,0.0049756197,0.00068156,0.0005056089,0.00047654618,0.000018243345,0.00096534495],"genre_scores_gemma":[0.995269,0.00061368174,0.00088261435,0.00043674186,0.000163964,0.0000123764285,0.000015413269,0.000024016032,0.0025821938],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9977557,0.0002091153,0.0009955415,0.0002951153,0.00047533287,0.0002691814],"domain_scores_gemma":[0.9883161,0.010332988,0.0004588523,0.00044743033,0.00030834152,0.00013626534],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0012274699,0.00019151201,0.00048399862,0.00008676849,0.00012416727,0.000011758377,0.00020954464,0.00016860892,0.00024710194],"category_scores_gemma":[0.015457278,0.0001125097,0.00030067973,0.000544853,0.0006883421,0.000013247306,0.000039374947,0.00027656206,0.000002812534],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011962609,0.00053215807,0.93731236,0.00005109319,0.00011464003,0.0000022391682,0.000079678524,0.0001068199,0.000010954523,0.00045939442,0.059173718,0.00096067006],"study_design_scores_gemma":[0.0029584502,0.00043381352,0.9732345,0.000100265934,0.00009429893,0.000048744125,0.000052030722,0.00019289048,0.000009595817,0.000097166645,0.022674114,0.00010413561],"about_ca_topic_score_codex":0.00087814766,"about_ca_topic_score_gemma":0.00013083847,"teacher_disagreement_score":0.036499605,"about_ca_system_score_codex":0.000015064733,"about_ca_system_score_gemma":0.00021594664,"threshold_uncertainty_score":0.99283594},"labels":[],"label_agreement":null},{"id":"W2058085429","doi":"10.1051/pmed/2011109","title":"Aborder le raisonnement clinique du point de vue pédagogique","year":2011,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":26,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Philosophy; Humanities","score_opus":0.05850622795046805,"score_gpt":0.3481958211945626,"score_spread":0.28968959324409455,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2058085429","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3588027,0.08942426,0.06086147,0.132234,0.014324871,0.0030051465,0.0002790858,0.0012833363,0.33978513],"genre_scores_gemma":[0.78132397,0.03773826,0.029523103,0.026822852,0.0078670615,0.0004069066,0.00013476158,0.000324473,0.11585862],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9944849,0.00066207256,0.0014533008,0.0010252221,0.00060881994,0.0017656613],"domain_scores_gemma":[0.98638815,0.009975166,0.00034609818,0.001155499,0.0003219737,0.0018131349],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0032427032,0.0006247221,0.0010964217,0.00015442201,0.00025957703,0.00006440232,0.0005461587,0.0010263205,0.017773448],"category_scores_gemma":[0.05656672,0.0006048089,0.000614799,0.00040159945,0.0010228462,0.00016235492,0.00033526213,0.0014055684,0.002464582],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012211226,0.018502627,0.1227406,0.0006258161,0.0009540861,0.004463828,0.03122936,0.000037237394,0.00043334113,0.03297071,0.3710528,0.41576847],"study_design_scores_gemma":[0.013067238,0.004076656,0.117357396,0.011248202,0.0013371387,0.00068792433,0.010866316,0.0020956993,0.008153466,0.021852335,0.80728006,0.0019775715],"about_ca_topic_score_codex":0.0027369175,"about_ca_topic_score_gemma":0.00021524593,"teacher_disagreement_score":0.43622726,"about_ca_system_score_codex":0.000326162,"about_ca_system_score_gemma":0.002313004,"threshold_uncertainty_score":0.99964035},"labels":[],"label_agreement":null},{"id":"W2058166804","doi":"10.1080/10401334.2010.488197","title":"Assessment in the Context of Uncertainty Using the Script Concordance Test: More Meaning for Scores","year":2010,"lang":"en","type":"article","venue":"Teaching and Learning in Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":71,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Université de Montréal","funders":"","keywords":"Concordance; Raw score; Standard deviation; Test (biology); Medicine; Meaning (existential); Context (archaeology); Set (abstract data type); Metric (unit); Statistics; Raw data; Psychology; Mathematics; Computer science; Operations management","score_opus":0.039063791596324136,"score_gpt":0.3966457841668997,"score_spread":0.3575819925705756,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2058166804","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98857987,0.0003514531,0.000551591,0.008972607,0.00018718804,0.00040164767,0.0000013160546,0.000021143303,0.0009331871],"genre_scores_gemma":[0.9952644,0.000031552787,0.0022248288,0.0020573793,0.00028228352,0.000021746926,0.000010907653,0.000014221636,0.00009266391],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9984741,0.00033068276,0.00044835723,0.00022459777,0.000290657,0.00023159687],"domain_scores_gemma":[0.96462244,0.034825247,0.00018858897,0.0002377103,0.00006367712,0.00006235927],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0074102413,0.0001353767,0.00043841943,0.000086406704,0.00023194238,0.000013414403,0.00013811926,0.00007736501,0.000012008523],"category_scores_gemma":[0.12228941,0.00006868086,0.000045332345,0.00013905797,0.00043544444,0.00002727716,0.000028632434,0.0026185778,1.7110604e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001049639,0.00016735126,0.91628116,0.00008685424,0.00002039668,0.00002260063,0.024417711,0.00075088814,0.005602226,0.0036661727,0.0003861644,0.048493512],"study_design_scores_gemma":[0.0073148096,0.001570294,0.7874128,0.008923863,0.0002068241,0.0001436101,0.068386436,0.1215053,0.000034993776,0.000584113,0.0037130944,0.00020383969],"about_ca_topic_score_codex":0.0032880278,"about_ca_topic_score_gemma":0.00036633565,"teacher_disagreement_score":0.12886833,"about_ca_system_score_codex":0.000027321947,"about_ca_system_score_gemma":0.0001076545,"threshold_uncertainty_score":0.9996824},"labels":[],"label_agreement":null},{"id":"W2058207300","doi":"10.1007/s10459-011-9305-4","title":"Self-monitoring and its relationship to medical knowledge","year":2011,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":51,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Advancing Health Outcomes; University of British Columbia; Medical Council of Canada","funders":"","keywords":"Summative assessment; Context (archaeology); Test (biology); Psychology; Moment (physics); Applied psychology; Medical education; Formative assessment; Social psychology; Medicine; Mathematics education","score_opus":0.07766023875395874,"score_gpt":0.478431212360064,"score_spread":0.40077097360610525,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2058207300","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93872714,0.03298306,0.000071029026,0.006294706,0.0022600016,0.00048147715,4.2845178e-7,0.0000624943,0.019119656],"genre_scores_gemma":[0.981566,0.003899736,0.0134878615,0.0006430487,0.00022961764,0.00004938807,0.0000010298531,0.0000037960313,0.00011949442],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99892265,0.0000581332,0.00031481087,0.00025789515,0.00023465624,0.00021186759],"domain_scores_gemma":[0.9961803,0.0029775554,0.000076033,0.00009861773,0.00006929631,0.00059818185],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0014102367,0.000061035265,0.00013317491,0.00014877979,0.0001309763,0.0000074894565,0.00008402667,0.000047300837,0.00004044729],"category_scores_gemma":[0.032377597,0.000051360505,0.000011627648,0.0005740668,0.00006812179,0.0001802297,0.000024624314,0.00013893362,0.000037477876],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000009401211,0.00029067462,0.90415955,0.00010321492,6.1553254e-7,6.025465e-7,0.0029936763,0.000001591234,3.8427726e-7,0.008036762,0.00006636157,0.084337175],"study_design_scores_gemma":[0.0002565847,0.0004914251,0.9858244,0.003198599,0.000004534416,0.000023990664,0.000919018,0.0002296603,0.000046657984,0.0027616713,0.0061562927,0.000087159824],"about_ca_topic_score_codex":0.00008534143,"about_ca_topic_score_gemma":0.000049428298,"teacher_disagreement_score":0.08425001,"about_ca_system_score_codex":0.00008384822,"about_ca_system_score_gemma":0.00169862,"threshold_uncertainty_score":0.9757731},"labels":[],"label_agreement":null},{"id":"W2058722633","doi":"10.1136/bmj.39349.461678.43","title":"Clinical Epidemiology: A Basic Science for Clinical Medicine","year":2007,"lang":"en","type":"article","venue":"BMJ","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Epidemiology; Clinical epidemiology; Test (biology); Reading (process); Medicine; Medical education; Family medicine; Pathology; Political science","score_opus":0.29605739528349667,"score_gpt":0.592554439511773,"score_spread":0.29649704422827633,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2058722633","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7868468,0.0014750885,0.03486633,0.11508093,0.008192228,0.0025154531,0.000011455601,0.0002910971,0.050720595],"genre_scores_gemma":[0.89847875,0.00061984453,0.04332733,0.04250371,0.011724261,0.00004734303,0.000030139356,0.00003539671,0.0032332444],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99525785,0.0002060903,0.0027445871,0.000734488,0.00035995516,0.0006970053],"domain_scores_gemma":[0.6700097,0.325694,0.0007463728,0.0012005427,0.00064897817,0.001700462],"candidate_categories":["metaresearch","sts"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.07397497,0.00015904967,0.001186355,0.00012264817,0.00012244817,0.0000043247223,0.00024487163,0.00031654738,0.00018995444],"category_scores_gemma":[0.7962698,0.000107813496,0.00040638557,0.00038140535,0.0027192489,0.000037531525,0.00008381514,0.00050011376,0.00025765406],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005555288,0.00030248417,0.65362227,0.0000120695795,0.00003119886,0.0000396831,0.00001595596,3.006899e-7,0.000008986645,0.002616465,0.19503964,0.14775543],"study_design_scores_gemma":[0.0046212296,0.002351545,0.9058695,0.00089863705,0.000167618,0.00008188754,0.000044395965,0.00036224123,0.000015416459,0.0023695386,0.0830866,0.00013140978],"about_ca_topic_score_codex":0.000022538057,"about_ca_topic_score_gemma":0.000005042094,"teacher_disagreement_score":0.7222948,"about_ca_system_score_codex":0.000044051147,"about_ca_system_score_gemma":0.00056254904,"threshold_uncertainty_score":0.99999475},"labels":[],"label_agreement":null},{"id":"W2059180750","doi":"10.7326/0003-4819-145-1-200607040-00017","title":"Clinical Sensibility and Barriers to Knowledge Translation","year":2006,"lang":"en","type":"letter","venue":"Annals of Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Ottawa","funders":"","keywords":"Medicine; Knowledge translation; Sensibility; Clinical Practice; Family medicine; Law","score_opus":0.13602159211921613,"score_gpt":0.45816323737655756,"score_spread":0.32214164525734146,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2059180750","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.12672044,0.0031068744,0.00026209708,0.85560554,0.0012900911,0.0005414599,0.000057664336,0.000049695944,0.012366114],"genre_scores_gemma":[0.13233368,0.0004860472,0.0004755683,0.85410535,0.0086225085,0.000009993679,0.00018764366,0.000056734836,0.00372248],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9967161,0.0002149903,0.001583489,0.000630257,0.00050704356,0.0003481397],"domain_scores_gemma":[0.9821715,0.015682569,0.0003339692,0.0005228944,0.00058713683,0.00070192595],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0017206393,0.00033493867,0.0015110564,0.00024407556,0.000027125547,0.00000669338,0.00014286354,0.00076819165,0.00038105348],"category_scores_gemma":[0.06618746,0.00024259534,0.00032472383,0.00017045489,0.00058187573,0.00002457107,0.00005157557,0.0015907502,0.00003972711],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00027471117,0.000072719005,0.021179438,0.00021427276,0.00016108072,0.00028968474,0.00018574834,2.429615e-7,0.000022393,0.000005319981,0.9285304,0.049063977],"study_design_scores_gemma":[0.0029486583,0.00459598,0.07506397,0.022403924,0.00059726083,0.000117436146,0.000058367717,0.0000840598,0.00019194515,0.0006633753,0.8929244,0.00035056873],"about_ca_topic_score_codex":0.00058425666,"about_ca_topic_score_gemma":0.00003363419,"teacher_disagreement_score":0.06446683,"about_ca_system_score_codex":0.000016674561,"about_ca_system_score_gemma":0.00019519868,"threshold_uncertainty_score":0.9892751},"labels":[],"label_agreement":null},{"id":"W2060109649","doi":"10.1503/cmaj.100270","title":"To err is human -- and let's not forget it","year":2010,"lang":"en","type":"article","venue":"Canadian Medical Association Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":34,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Criticism; Enlightenment; Line (geometry); Philosophy; Computer science; Classics; Medicine; Psychoanalysis; Literature; Theology; Psychology; History; Art; Mathematics","score_opus":0.014601131781008232,"score_gpt":0.3218230424472156,"score_spread":0.30722191066620735,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2060109649","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7150002,0.000011663419,0.000017037151,0.27617687,0.0006776319,0.000081281934,0.000017030157,0.000011332796,0.008006924],"genre_scores_gemma":[0.86306417,0.00004644779,0.00030213819,0.13193725,0.0022670228,0.0000037020775,0.000007779967,0.000014105781,0.002357387],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.998108,0.000037091067,0.0003670315,0.00016193787,0.00092321244,0.00040268476],"domain_scores_gemma":[0.9935893,0.0015148922,0.0001034193,0.00011506518,0.00028466113,0.0043927087],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0015926415,0.00009608261,0.00024400893,0.00013063708,0.00023415044,0.00007456584,0.00010151591,0.00039869166,0.003990303],"category_scores_gemma":[0.07373281,0.000083318686,0.00008138927,0.000126999,0.000041745785,0.000041120533,0.000016700618,0.0013160625,0.00017571145],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000009290571,0.000037508373,0.20245928,0.0000033579477,0.000058766407,0.0003445151,0.00038344105,7.846946e-8,0.00017457388,0.00024114018,0.72598803,0.07030004],"study_design_scores_gemma":[0.0012174292,0.00014198701,0.31909156,0.00022311744,0.0000458515,0.0004426602,0.00010632727,0.000032006687,0.00005466241,0.00014817285,0.67838246,0.00011375196],"about_ca_topic_score_codex":0.002532306,"about_ca_topic_score_gemma":0.041841913,"teacher_disagreement_score":0.14806393,"about_ca_system_score_codex":0.00019138346,"about_ca_system_score_gemma":0.0013664071,"threshold_uncertainty_score":0.99692017},"labels":[],"label_agreement":null},{"id":"W2060526915","doi":"10.1177/0163278702250082","title":"A Validity Study Of Expert Judgment Procedures For Setting Cutoff Scores On High-Stakes Credentialing Examinations Using Cluster Analysis","year":2003,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Calgary; Royal College of Physicians and Surgeons of Canada; Scanimetrics (Canada)","funders":"","keywords":"Cutoff; Licensure; Credentialing; Cluster (spacecraft); Categorization; Clinical judgment; Test (biology); Psychology; Medical physics; Statistics; Medicine; Computer science; Artificial intelligence; Medical education; Mathematics","score_opus":0.2742225804547531,"score_gpt":0.5179446169931036,"score_spread":0.24372203653835045,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2060526915","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9783287,0.00019811811,0.0087679885,0.0069435136,0.00048399958,0.005147293,0.000017513592,0.000035475834,0.000077372446],"genre_scores_gemma":[0.9937349,0.000025176263,0.0034789406,0.001614707,0.00013196885,0.00075611833,0.000072178555,0.00001966764,0.00016636522],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99578625,0.0017981962,0.0008665644,0.00035197526,0.00090345804,0.00029353437],"domain_scores_gemma":[0.9874585,0.010586137,0.0006238053,0.00048564732,0.00069109246,0.00015478549],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0070394226,0.00015146012,0.00044764768,0.0002477428,0.0009197303,0.00001972801,0.000087987595,0.000077446515,0.00012378248],"category_scores_gemma":[0.05547954,0.000098250006,0.00015053112,0.0006421781,0.00004608755,0.000046799803,0.000029208488,0.00020544008,0.0000023396394],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0031570233,0.049150117,0.43742993,0.0019520753,0.0058909515,0.000006628532,0.25155956,0.15551436,0.0020425657,0.007061516,0.041261807,0.044973455],"study_design_scores_gemma":[0.021459045,0.0074613597,0.64308345,0.011591378,0.010697711,0.000008987932,0.09268834,0.20718692,0.002023793,0.002765286,0.00033710757,0.00069663534],"about_ca_topic_score_codex":0.000330589,"about_ca_topic_score_gemma":0.00027511377,"teacher_disagreement_score":0.20565349,"about_ca_system_score_codex":0.00021693457,"about_ca_system_score_gemma":0.0012384411,"threshold_uncertainty_score":0.95247656},"labels":[],"label_agreement":null},{"id":"W2060674387","doi":"10.1176/appi.ajp.162.3.451","title":"The Physician’s Bias","year":2005,"lang":"en","type":"article","venue":"American Journal of Psychiatry","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Saskatchewan","funders":"","keywords":"Psychology; Bit (key); Family medicine; Medicine; Computer science","score_opus":0.015805819074024313,"score_gpt":0.32843091409607816,"score_spread":0.3126250950220538,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2060674387","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.89198595,0.0020186277,0.00015870627,0.09988443,0.0010907394,0.00004911521,0.000001060247,0.000012952238,0.0047984305],"genre_scores_gemma":[0.97753316,0.00046879545,0.00897475,0.009584373,0.0031104886,8.4024504e-7,4.0580593e-7,0.000014095161,0.00031309732],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99900776,0.000047838097,0.0004190937,0.000075132135,0.00027719667,0.00017297646],"domain_scores_gemma":[0.9965499,0.0024701436,0.00046838346,0.00019729142,0.00012197565,0.00019229097],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00040908143,0.00008054478,0.00026827154,0.000037948193,0.000073862706,0.000017780403,0.00012750884,0.000018653862,0.000025125997],"category_scores_gemma":[0.0033195848,0.000044466182,0.0002426743,0.00018858422,0.00023493616,0.000033811913,0.000011083036,0.00030360903,0.00006688761],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00035366582,0.0004016361,0.13518552,0.0000022654287,0.00016033131,0.000011705713,0.00009619731,0.000021303891,0.000023935534,0.0015798935,0.06480357,0.79735994],"study_design_scores_gemma":[0.0045693605,0.0073295548,0.6486356,0.0022641222,0.00067616557,0.0013592624,0.0040700505,0.00013284254,0.0001737127,0.0040673064,0.3263778,0.0003441622],"about_ca_topic_score_codex":0.000010101489,"about_ca_topic_score_gemma":0.000010612709,"teacher_disagreement_score":0.7970158,"about_ca_system_score_codex":0.00002552737,"about_ca_system_score_gemma":0.0002864904,"threshold_uncertainty_score":0.3974093},"labels":[],"label_agreement":null},{"id":"W2061155967","doi":"10.1111/j.1365-2923.2007.02994.x","title":"Initial diagnostic hypotheses bias analytic information processing in non‐visual domains","year":2008,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Medical diagnosis; Confidence interval; Selection (genetic algorithm); Debiasing; Statistical hypothesis testing; Information processing; Alternative hypothesis; Medicine; Psychology; Cognitive psychology; Statistics; Mathematics; Social psychology; Artificial intelligence; Computer science; Radiology; Internal medicine","score_opus":0.03441244357384741,"score_gpt":0.386644998755461,"score_spread":0.35223255518161356,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2061155967","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9907453,0.00016167882,0.00057252945,0.004145207,0.0003899771,0.00028801124,0.000001169606,0.000051073042,0.0036450736],"genre_scores_gemma":[0.9918111,0.00020731361,0.0003274713,0.006839192,0.00047940877,0.00007377907,0.00013534217,0.000010899512,0.00011547888],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99826986,0.000052999956,0.000570555,0.00017107905,0.0006908229,0.00024471624],"domain_scores_gemma":[0.993909,0.0052128597,0.00015001546,0.0001737062,0.00018004127,0.00037436737],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0004986106,0.0001252313,0.00027747298,0.0002456017,0.00007963153,0.000021625723,0.00007951905,0.00018303229,0.000379517],"category_scores_gemma":[0.27340636,0.00010403325,0.00006222167,0.00048719026,0.00018005715,0.00023423269,0.00002371647,0.00032951386,0.00021003441],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012769792,0.0025762995,0.43517578,0.00013750503,0.000013425607,0.00010694123,0.0026702743,0.0000125704,0.000005364156,0.000067548244,0.014459821,0.54464674],"study_design_scores_gemma":[0.0019261377,0.00027063733,0.98663986,0.0030907004,0.000059287348,0.0002560938,0.00048259876,0.0041927467,0.000054802753,0.00016651934,0.002695517,0.0001650722],"about_ca_topic_score_codex":0.0002766014,"about_ca_topic_score_gemma":0.000030298555,"teacher_disagreement_score":0.5514641,"about_ca_system_score_codex":0.00012400933,"about_ca_system_score_gemma":0.0031093303,"threshold_uncertainty_score":0.73271406},"labels":[],"label_agreement":null},{"id":"W2061672965","doi":"10.1097/00001888-200011000-00001","title":"Clinical Problem Solving and the Clinical Presentation Curriculum","year":2000,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Problem statement; Presentation (obstetrics); Schema (genetic algorithms); Curriculum; Scheme (mathematics); Subject (documents); Computer science; Medical diagnosis; Statement (logic); Mathematics education; Psychology; Medicine; Management science; Epistemology; Mathematics; Pedagogy; Information retrieval","score_opus":0.06221151098634382,"score_gpt":0.43240956507963435,"score_spread":0.37019805409329054,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2061672965","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.012308592,0.010338593,0.000080923295,0.96171474,0.002018528,0.0018202565,0.000008435572,0.00015808328,0.011551873],"genre_scores_gemma":[0.004310998,0.07045894,0.0003312019,0.85731786,0.051224228,0.000092172,0.0005310887,0.000096063784,0.015637442],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99238026,0.0012876244,0.0033784665,0.0011281437,0.0011525414,0.0006729472],"domain_scores_gemma":[0.9321447,0.06550689,0.0008856375,0.0008197432,0.00018303558,0.000460021],"candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.007456559,0.0005310031,0.00235168,0.00013872128,0.00014247501,0.000017883924,0.00043539528,0.003280233,0.0008081474],"category_scores_gemma":[0.09940854,0.00028014742,0.0004219861,0.00027359254,0.0033075572,0.0000619206,0.0001652535,0.01404945,0.00018966326],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00023951594,0.000027353528,0.03651105,0.00011853404,0.00022534639,0.00026782718,0.00030509636,1.8607402e-7,2.0218442e-7,0.00007841169,0.91915154,0.04307491],"study_design_scores_gemma":[0.01615585,0.00084231026,0.010999359,0.011985903,0.0027825327,0.0005219941,0.00020119337,0.00031458645,3.462827e-7,0.004110665,0.95171744,0.0003678065],"about_ca_topic_score_codex":0.00024886066,"about_ca_topic_score_gemma":0.0000018150016,"teacher_disagreement_score":0.10439685,"about_ca_system_score_codex":0.000038677015,"about_ca_system_score_gemma":0.00032513053,"threshold_uncertainty_score":0.9999651},"labels":[],"label_agreement":null},{"id":"W2061723709","doi":"10.1197/j.aem.2003.11.017","title":"A Case Study in Medical Error: The Use of the Portfolio Entry","year":2004,"lang":"en","type":"article","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Accreditation; Medicine; Graduate medical education; Portfolio; Medical education; Reflection (computer programming); Lifelong learning; Pedagogy; Psychology","score_opus":0.14090005720329743,"score_gpt":0.4292455880024372,"score_spread":0.2883455307991397,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2061723709","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97520953,0.0016719593,0.000021059019,0.020207305,0.0013326025,0.0008885912,0.0000027775434,0.000029607656,0.0006365619],"genre_scores_gemma":[0.99389356,0.0027613433,0.00001592676,0.0020319095,0.000598664,0.00004253021,0.00000407146,0.000022005415,0.00062996824],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9963001,0.00018856932,0.001592048,0.0003423633,0.0012186932,0.00035823556],"domain_scores_gemma":[0.9952453,0.003267196,0.00031991483,0.0007339565,0.00011060955,0.00032305717],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0017889161,0.00021082662,0.00054422265,0.00010323474,0.00007885413,3.5782008e-7,0.00036132245,0.00029436706,0.0016420428],"category_scores_gemma":[0.08802404,0.00009748402,0.00013996215,0.00092771486,0.0004444349,0.00004891768,0.00016406193,0.0016255666,0.000013587763],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000084239044,0.000665646,0.97414356,0.000022509807,0.00008113891,0.002590749,0.0037620764,0.00005791341,0.000036365673,0.00028089565,0.01677438,0.0015005164],"study_design_scores_gemma":[0.0141198775,0.0017360144,0.94714874,0.0067909816,0.001103201,0.010391598,0.010401317,0.0002723148,0.000089144385,0.0012360091,0.0063639637,0.00034684385],"about_ca_topic_score_codex":0.00589604,"about_ca_topic_score_gemma":0.0005033223,"teacher_disagreement_score":0.08623513,"about_ca_system_score_codex":0.00006351505,"about_ca_system_score_gemma":0.0003660991,"threshold_uncertainty_score":0.9992706},"labels":[],"label_agreement":null},{"id":"W2061805713","doi":"10.1007/s40263-013-0098-2","title":"Comment on: “Publication Bias, with a Focus on Psychiatry: Causes and Solutions”","year":2013,"lang":"en","type":"letter","venue":"CNS Drugs","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Jewish General Hospital","funders":"","keywords":"Psychopharmacology; Focus (optics); Psychiatry; Clinical psychiatry; Psychology; Medicine","score_opus":0.04215796801755811,"score_gpt":0.30514262517499013,"score_spread":0.262984657157432,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2061805713","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0029621192,0.0003307441,0.000018038396,0.9858459,0.0003895306,0.00061326247,0.000053598567,0.00010845133,0.009678316],"genre_scores_gemma":[0.007863365,0.00020933076,0.00045235213,0.9838442,0.0025837368,0.00020639549,0.0006583426,0.000058152895,0.0041241026],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99836487,0.00007286053,0.00030901452,0.00048345607,0.0004258287,0.00034396994],"domain_scores_gemma":[0.99432814,0.0044624317,0.00019849619,0.00067542563,0.00015442235,0.00018111218],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00019699761,0.00028892478,0.00044364744,0.00016057998,0.00011272779,0.00007681616,0.00009705492,0.0004408809,0.00020717463],"category_scores_gemma":[0.003927691,0.00019739196,0.00008662257,0.00014026274,0.00015155063,0.00002712917,0.00003688827,0.0011432264,0.0004638355],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000046691308,0.00022701961,0.0011235476,0.000037719015,0.00010129095,0.000016807999,0.000060186198,2.576139e-7,1.1902415e-7,0.001157191,0.9927221,0.004507086],"study_design_scores_gemma":[0.0017621996,0.0011956547,0.0048461915,0.003526533,0.0003088038,0.000032814343,0.000038884176,0.000017786884,0.0000064137607,0.0006544497,0.98734117,0.00026909815],"about_ca_topic_score_codex":0.00024483012,"about_ca_topic_score_gemma":0.0000073338715,"teacher_disagreement_score":0.0055542137,"about_ca_system_score_codex":0.00008784041,"about_ca_system_score_gemma":0.00021418586,"threshold_uncertainty_score":0.80494106},"labels":[],"label_agreement":null},{"id":"W2062618525","doi":"10.1186/1472-6947-8-18","title":"Script Concordance Tests: Guidelines for Construction","year":2008,"lang":"en","type":"article","venue":"BMC Medical Informatics and Decision Making","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":253,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Université du Québec à Montréal","funders":"","keywords":"Concordance; Context (archaeology); Computer science; Test (biology); Task (project management); Process (computing); Health informatics; Quality (philosophy); Natural language processing; Information retrieval; Data science; Artificial intelligence; Medicine; Pathology","score_opus":0.1298384051838863,"score_gpt":0.43343385655049155,"score_spread":0.30359545136660526,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2062618525","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.31273347,0.00034734403,0.68494743,0.00017267583,0.00044176495,0.00024101257,0.0000074477175,0.00005148164,0.0010573643],"genre_scores_gemma":[0.08486252,0.0008140043,0.9064167,0.007445702,0.00036207176,0.000020210116,0.000014787159,0.0000111194795,0.00005285245],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9979927,0.000009315787,0.0010122698,0.00012593425,0.0006559169,0.00020388317],"domain_scores_gemma":[0.98391026,0.014830086,0.00019446286,0.00020698147,0.00051397714,0.0003442013],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00074573496,0.00012405074,0.0003636674,0.000078313635,0.00015894565,0.000026724032,0.00007801577,0.00016519651,0.00010055819],"category_scores_gemma":[0.13916983,0.000087786226,0.00007380947,0.00013949804,0.00024406229,0.00008728746,0.00006642341,0.00015140406,0.000018841769],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00027525454,0.000053038228,0.013741148,0.00011538482,0.000014291818,0.00002017845,0.00009855603,0.000027405176,0.0000031929667,0.0015956467,0.058591783,0.9254641],"study_design_scores_gemma":[0.012785857,0.0011282193,0.015971089,0.01905524,0.00016017909,0.005600078,0.0015126547,0.8637027,0.00007772635,0.01485239,0.06457726,0.00057659],"about_ca_topic_score_codex":0.0000030147555,"about_ca_topic_score_gemma":0.0000037735724,"teacher_disagreement_score":0.92488754,"about_ca_system_score_codex":0.000018794894,"about_ca_system_score_gemma":0.00039851584,"threshold_uncertainty_score":0.86808133},"labels":[],"label_agreement":null},{"id":"W2062689150","doi":"10.1197/j.aem.2007.06.025","title":"Cognitive and Social Issues in Emergency Medicine Knowledge Translation: A Research Agenda","year":2007,"lang":"en","type":"article","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":26,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; University of Alberta; University of Ottawa","funders":"","keywords":"Knowledge translation; Medicine; CLARITY; Psychological intervention; Cognition; Relevance (law); Social cognitive theory; Autonomy; Process (computing); Emergency department; Medical education; Knowledge management; Psychology; Nursing; Social psychology; Psychiatry; Computer science","score_opus":0.30558295383956746,"score_gpt":0.5499008717546797,"score_spread":0.24431791791511226,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2062689150","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.79547286,0.09674856,0.00021290865,0.039738975,0.0023012839,0.0010527294,0.000009269001,0.0001188538,0.06434453],"genre_scores_gemma":[0.92423445,0.062261906,0.000042811567,0.00029707648,0.0055056927,0.000044960674,0.0001150959,0.000053164797,0.0074448166],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9946234,0.00029674135,0.0023147,0.0007531245,0.0010745155,0.0009374997],"domain_scores_gemma":[0.99113214,0.0072573354,0.00024786755,0.00023903893,0.0006436971,0.00047990977],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.008978592,0.00035166324,0.0009089528,0.00075798563,0.00023088136,4.737108e-7,0.00019985455,0.0005984786,0.007044384],"category_scores_gemma":[0.04639642,0.0002743767,0.00009953144,0.0018557515,0.0009266006,0.00010157991,0.000078363635,0.0021669841,0.000099268815],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014049769,0.0005498894,0.58793235,0.00057528587,0.0002117249,0.00016766896,0.06592777,1.6738151e-7,0.004401615,0.002464374,0.26406503,0.07229914],"study_design_scores_gemma":[0.012183556,0.0027281488,0.85873485,0.008686112,0.00060542393,0.00007871766,0.019285446,0.00012509129,0.000325518,0.008299766,0.088267565,0.00067978946],"about_ca_topic_score_codex":0.0007862983,"about_ca_topic_score_gemma":0.00031063717,"teacher_disagreement_score":0.27080253,"about_ca_system_score_codex":0.000086650994,"about_ca_system_score_gemma":0.00012057247,"threshold_uncertainty_score":0.99997085},"labels":[],"label_agreement":null},{"id":"W2062996846","doi":"10.1080/10401334.2013.842911","title":"Diagnostic Reasoning: Where We've Been, Where We’re Going","year":2013,"lang":"en","type":"review","venue":"Teaching and Learning in Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":93,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Focus (optics); Component (thermodynamics); Cognition; Process (computing); Dual (grammatical number); Computer science; Unconscious mind; Cognitive science; Cognitive psychology; Dual process theory (moral psychology); Psychology; Neuroscience","score_opus":0.051422806397288946,"score_gpt":0.3870666365911368,"score_spread":0.33564383019384786,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2062996846","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005531451,0.98477775,0.00008540012,0.005279549,0.0005948233,0.0012089952,0.00000433372,0.0004040132,0.0070919897],"genre_scores_gemma":[0.00047510047,0.9790257,0.0015323593,0.00029304845,0.0021732335,0.00018018516,0.00019600245,0.00022491564,0.01589941],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9929678,0.0020989045,0.0018053097,0.0013100009,0.00081848726,0.0009995005],"domain_scores_gemma":[0.9226473,0.0747795,0.0008702567,0.00077351235,0.00009888563,0.00083053566],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0050244797,0.0010828592,0.004755925,0.0005878108,0.00049555843,0.00009070242,0.0003485814,0.0010291246,0.00084047357],"category_scores_gemma":[0.22460274,0.00076360966,0.0004222478,0.00046867202,0.00040291078,0.00010716325,0.00022790024,0.010809221,0.0003525969],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000018181083,0.00014497107,0.0036852704,0.016801478,0.00018363554,0.0004887719,0.002704099,0.0000096057365,3.375331e-7,0.00010672715,0.026822599,0.94903433],"study_design_scores_gemma":[0.0012592939,0.00061704207,0.00024513176,0.40913373,0.0013209025,0.000324915,0.0012079899,0.00015312278,1.8644918e-8,0.000024810439,0.58528626,0.00042679568],"about_ca_topic_score_codex":0.0010141032,"about_ca_topic_score_gemma":0.00008158542,"teacher_disagreement_score":0.9486075,"about_ca_system_score_codex":0.00025467202,"about_ca_system_score_gemma":0.00037758573,"threshold_uncertainty_score":0.9994815},"labels":[],"label_agreement":null},{"id":"W2063251236","doi":"10.1111/j.1365-2923.2009.03516.x","title":"Assessing ethical problem solving by reasoning rather than decision making","year":2009,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":31,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Calgary","funders":"","keywords":"Context (archaeology); Psychology; Reliability (semiconductor); Quality (philosophy); Construct (python library); Construct validity; Medical education; Ethical decision; Think aloud protocol; Ethical issues; Applied psychology; Social psychology; Psychometrics; Clinical psychology; Medicine; Computer science; Engineering ethics","score_opus":0.020320173022115248,"score_gpt":0.42244109972899363,"score_spread":0.4021209267068784,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2063251236","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8865002,0.0013735353,0.028589755,0.06390518,0.0009614159,0.00039537827,7.84134e-7,0.00023465746,0.018039089],"genre_scores_gemma":[0.94519925,0.00007484599,0.027239498,0.026457956,0.00064440496,0.000012647371,0.000045243123,0.000020605106,0.00030552506],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99769515,0.00010743604,0.00048747702,0.00039494396,0.0009778494,0.0003371184],"domain_scores_gemma":[0.99409086,0.004865443,0.0001452498,0.00030466443,0.00016705165,0.00042671815],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0015731641,0.00015917339,0.00030777062,0.000076246455,0.00014700864,0.00011635964,0.0001286824,0.0005424684,0.0006758746],"category_scores_gemma":[0.13711466,0.00012725398,0.000102006015,0.0002640387,0.00008314072,0.0001153523,0.000028962962,0.0010725285,0.00006482358],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000056589,0.00081827014,0.010533671,0.000014692595,0.000010544424,0.000015447671,0.000368876,0.0000011994517,0.00019648673,0.0006395403,0.09511559,0.8922291],"study_design_scores_gemma":[0.0071942927,0.0018838156,0.76047593,0.11762551,0.00063731754,0.0011215914,0.0027128444,0.015428245,0.00083524385,0.03130034,0.059167244,0.001617633],"about_ca_topic_score_codex":0.000031138472,"about_ca_topic_score_gemma":0.0000033741233,"teacher_disagreement_score":0.89061147,"about_ca_system_score_codex":0.00012697579,"about_ca_system_score_gemma":0.0014732083,"threshold_uncertainty_score":0.8701538},"labels":[],"label_agreement":null},{"id":"W2065106049","doi":"10.3138/jvme.0112-13r","title":"Use of the Script Concordance Approach to Evaluate Clinical Reasoning in Food-Ruminant Practitioners","year":2012,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Université de Montréal; Cegep de Saint Hyacinthe","funders":"","keywords":"Concordance; Cronbach's alpha; Reliability (semiconductor); Competence (human resources); Test (biology); Representativeness heuristic; Medicine; Medical physics; Statistics; Psychology; Mathematics; Social psychology; Psychometrics; Internal medicine; Biology","score_opus":0.21133108918490742,"score_gpt":0.47420726378255545,"score_spread":0.262876174597648,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2065106049","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9902668,0.0005320953,0.00030180547,0.0054013887,0.0028471753,0.00023338941,0.0000012479331,0.0000039360234,0.00041213652],"genre_scores_gemma":[0.98813486,0.00024050321,0.008327318,0.0022870547,0.00087643904,0.000011888548,0.0000035133435,0.000012048962,0.00010635576],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99724317,0.00044041575,0.0011267356,0.00012974368,0.0008392553,0.0002206649],"domain_scores_gemma":[0.9945744,0.0034167308,0.0006854774,0.00027868242,0.00034259775,0.0007020894],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.003987059,0.000102819846,0.00042323003,0.00010902328,0.000031200198,0.000009842004,0.00015952908,0.00014349898,0.000058296093],"category_scores_gemma":[0.18606466,0.000064680265,0.00018876424,0.00031061875,0.0001551436,0.00022216898,0.00006216615,0.0006401663,0.0000054506554],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0025196115,0.014062779,0.77950585,0.00016312054,0.0001632747,0.000030100191,0.002071915,0.000042808802,0.00061462296,0.00088128936,0.03968633,0.1602583],"study_design_scores_gemma":[0.0014213282,0.0026420634,0.96678305,0.0059934035,0.00011844574,0.0019296667,0.0005825379,0.00026828633,0.00007220341,0.0000410657,0.020053376,0.00009456904],"about_ca_topic_score_codex":0.000036175996,"about_ca_topic_score_gemma":0.0000014225652,"teacher_disagreement_score":0.1872772,"about_ca_system_score_codex":0.00012000832,"about_ca_system_score_gemma":0.0018571484,"threshold_uncertainty_score":0.8207915},"labels":[],"label_agreement":null},{"id":"W2065667673","doi":"10.1097/00001888-200604000-00017","title":"Giving Learners the Best of Both Worlds: Do Clinical Teachers Need to Guard Against Teaching Pattern Recognition to Novices?","year":2006,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":138,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"McMaster University","keywords":"Guard (computer science); Medical education; Mathematics education; Psychology; Computer science; Medicine","score_opus":0.05732929057871691,"score_gpt":0.395317213593867,"score_spread":0.3379879230151501,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2065667673","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9404625,0.0004501509,0.0016058291,0.047056623,0.00061065756,0.00071830355,0.000008971376,0.0000875621,0.00899935],"genre_scores_gemma":[0.9694513,0.00029182024,0.00054683775,0.024238115,0.0032728338,0.000035244793,0.00006717496,0.000049260474,0.002047381],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9962252,0.00032680476,0.0015992654,0.00058006536,0.0007580113,0.0005106726],"domain_scores_gemma":[0.9847331,0.013658871,0.000409235,0.0005352812,0.000117972035,0.00054554036],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005094106,0.0002846145,0.0008402621,0.00024131298,0.00011315253,0.000010487067,0.00034598625,0.00036480528,0.00016620664],"category_scores_gemma":[0.07938016,0.00018324505,0.00016060343,0.00046448276,0.00027177567,0.000057189507,0.0001222178,0.0022198432,0.000230977],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001847427,0.00012454594,0.40216008,0.00004395793,0.00008403869,0.00003183287,0.004281066,0.0000651174,0.00128404,0.000030701725,0.07554323,0.5161666],"study_design_scores_gemma":[0.022888487,0.008701949,0.6140837,0.08274152,0.0028084582,0.00020761536,0.06393292,0.0011202826,0.0006906249,0.0014414574,0.19940269,0.001980314],"about_ca_topic_score_codex":0.0016379905,"about_ca_topic_score_gemma":0.000068842484,"teacher_disagreement_score":0.5141863,"about_ca_system_score_codex":0.00009105948,"about_ca_system_score_gemma":0.00010661851,"threshold_uncertainty_score":0.9644238},"labels":[],"label_agreement":null},{"id":"W2065926090","doi":"10.7439/ijbar.v5i9.852","title":"Impact of SNAPPS on Clinical Reasoning Skills of Surgery Residents in Outpatient Setting","year":2014,"lang":"en","type":"article","venue":"International Journal of Biomedical and Advance Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Preceptor; Set (abstract data type); Medical education; Medicine; Ambulatory; Clinical Practice; Psychology; Ideal (ethics); Nursing; Computer science; Surgery","score_opus":0.06922869747648246,"score_gpt":0.5058130271908385,"score_spread":0.43658432971435607,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2065926090","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9975197,0.00024309484,0.00019888264,0.0014348088,0.0003292211,0.00006689164,0.000010082334,0.0000020594891,0.00019527548],"genre_scores_gemma":[0.9970883,0.0013341992,0.0010296018,0.00006514555,0.0004279695,0.0000011382107,0.000008190529,0.000007970873,0.000037495323],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99621034,0.0003447172,0.0013456034,0.00016877266,0.0016954693,0.00023510851],"domain_scores_gemma":[0.9673459,0.030597724,0.00052273314,0.00013380688,0.0010243199,0.00037553237],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0069369734,0.00008522852,0.00054575305,0.0006010071,0.000017227254,0.000009949089,0.00018489455,0.000114874456,0.000053974967],"category_scores_gemma":[0.15954816,0.00005709022,0.0002620128,0.00025875217,0.00044433668,0.00006459438,0.00009362661,0.00072484836,0.0000028388952],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010863683,0.0015200358,0.76118875,0.000025205072,0.0001075717,0.00012836626,0.00006792292,0.000014619572,0.00028005446,0.00009691569,0.0008153825,0.2346688],"study_design_scores_gemma":[0.0021941103,0.0026518132,0.9839229,0.00936234,0.000010685494,0.00008286047,0.000050489332,0.00026989478,0.00027077645,0.0008388758,0.00029767974,0.000047547095],"about_ca_topic_score_codex":0.00008046814,"about_ca_topic_score_gemma":0.0000019504546,"teacher_disagreement_score":0.23462127,"about_ca_system_score_codex":0.00008855295,"about_ca_system_score_gemma":0.0003234902,"threshold_uncertainty_score":0.8475313},"labels":[],"label_agreement":null},{"id":"W2066687418","doi":"10.1080/0142159021000061413","title":"Was a breach of examination security unfair in an objective structured clinical examination? A critical incident","year":2003,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Canadian Institutes of Health Research","funders":"","keywords":"Objective structured clinical examination; Medical education; Psychology; Class (philosophy); Significant difference; Perception; Multidisciplinary approach; Mathematics education; Social psychology; Medicine; Applied psychology; Computer science; Political science; Law","score_opus":0.035138645558295545,"score_gpt":0.41178968024487717,"score_spread":0.3766510346865816,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2066687418","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9927677,0.00015516953,0.00044069317,0.000547048,0.00033407836,0.00033658778,0.000003220125,0.000053441137,0.0053620855],"genre_scores_gemma":[0.99868345,0.00003240637,0.00047496494,0.00029300596,0.0002482417,0.000032624925,0.000042411768,0.000024221645,0.00016868475],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9951177,0.001773283,0.0010652065,0.00053508446,0.0011634546,0.00034525074],"domain_scores_gemma":[0.99079794,0.007593975,0.00014703049,0.00044982127,0.00024351885,0.00076770666],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.006132442,0.00019623556,0.0006858689,0.0001482217,0.000034532102,0.000012558047,0.00014946064,0.0006165822,0.0019380813],"category_scores_gemma":[0.30987182,0.00016147578,0.00015547789,0.00035945643,0.0005544799,0.00009767161,0.000047616075,0.0010939819,0.000018053504],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001035483,0.0039581684,0.93197256,0.00004155269,0.000038651513,0.00022172472,0.0020321452,5.746965e-7,0.000018183324,0.0031372253,0.00024087429,0.05823478],"study_design_scores_gemma":[0.0037109042,0.0008986392,0.98762727,0.000685856,0.000096313095,0.00008298169,0.0008198723,0.002603525,0.00013385926,0.0030664557,0.00010266276,0.00017163948],"about_ca_topic_score_codex":0.00013425904,"about_ca_topic_score_gemma":0.0002000288,"teacher_disagreement_score":0.3037394,"about_ca_system_score_codex":0.00014060529,"about_ca_system_score_gemma":0.00044999807,"threshold_uncertainty_score":0.99897426},"labels":[],"label_agreement":null},{"id":"W2068598833","doi":"10.1097/acm.0b013e31829a3c32","title":"Is Clinical Cognition Binary or Continuous?","year":2013,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":24,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Dual (grammatical number); Dual process theory (moral psychology); Unconscious mind; Process (computing); Cognition; Variety (cybernetics); Computer science; Task (project management); Cognitive psychology; Cognitive science; Psychology; Artificial intelligence; Neuroscience","score_opus":0.12907157905463087,"score_gpt":0.4400291950577749,"score_spread":0.310957616003144,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2068598833","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.01123309,0.002915631,0.00005338031,0.96962976,0.0032999932,0.001335014,0.000051512103,0.00028651863,0.0111950785],"genre_scores_gemma":[0.0035166277,0.011396654,0.00011174316,0.8355589,0.04320178,0.00011133694,0.0017809784,0.00012770902,0.10419428],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99400854,0.0003041137,0.0024260937,0.0011520555,0.0012753191,0.00083388964],"domain_scores_gemma":[0.96272653,0.03442925,0.00084758137,0.00091063115,0.00043742917,0.0006485938],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity","insufficient_payload"],"category_scores_codex":[0.0016458351,0.0006682874,0.0025219172,0.00028884812,0.00007672073,0.0000105195195,0.00041048418,0.005267313,0.016706329],"category_scores_gemma":[0.11520272,0.00041104603,0.0003716912,0.00035496464,0.001082824,0.00006970385,0.00014061129,0.011962309,0.0040039346],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00023000341,0.00003552042,0.0039690277,0.00021385428,0.00029500146,0.0019551988,0.00017810393,1.0729786e-8,0.000016278404,0.000007707692,0.9836627,0.009436574],"study_design_scores_gemma":[0.006197168,0.0027150423,0.0037863073,0.022660065,0.0025033127,0.00069227215,0.0001484924,0.00004152909,0.000015360363,0.0007184792,0.96005327,0.0004687194],"about_ca_topic_score_codex":0.00010697323,"about_ca_topic_score_gemma":3.8870812e-7,"teacher_disagreement_score":0.13407089,"about_ca_system_score_codex":0.000073371906,"about_ca_system_score_gemma":0.00048052185,"threshold_uncertainty_score":0.9998341},"labels":[],"label_agreement":null},{"id":"W2069598955","doi":"10.1016/j.annemergmed.2008.07.024","title":"Assessing Clinical Reasoning in Pediatric Emergency Medicine: Validity Evidence for a Script Concordance Test","year":2008,"lang":"en","type":"article","venue":"Annals of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":82,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Centre Hospitalier Universitaire Sainte-Justine","funders":"","keywords":"Concordance; Medicine; Test (biology); Pediatric emergency medicine; Cronbach's alpha; Context (archaeology); Emergency department; Family medicine; Psychometrics; Clinical psychology; Emergency physician; Psychiatry; Internal medicine","score_opus":0.5577419357216217,"score_gpt":0.5586072172472664,"score_spread":0.00086528152564469,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2069598955","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95071995,0.022060005,0.0010804816,0.013379245,0.007094341,0.0010028987,0.000012402804,0.000097141135,0.004553552],"genre_scores_gemma":[0.8614398,0.13003597,0.00071271387,0.0006267258,0.0053437366,0.00008056632,0.000046658723,0.000056447876,0.0016573631],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9926431,0.0002940191,0.004232711,0.0009571383,0.0010535534,0.0008194443],"domain_scores_gemma":[0.9679029,0.02680088,0.0015381989,0.0010295693,0.0018649865,0.00086345786],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0064510023,0.00046238917,0.0018579705,0.0003828671,0.00017344182,0.0000019491401,0.00039677683,0.00031850135,0.0022821166],"category_scores_gemma":[0.53201777,0.00036016968,0.00051376584,0.0015815304,0.0005366694,0.0003453113,0.0000866089,0.00068170665,0.000021155745],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002715056,0.0006166611,0.7798984,0.00031693306,0.000038149934,0.00004843926,0.00033734654,0.0000035460748,0.0001987505,0.00009409894,0.21701872,0.0011574344],"study_design_scores_gemma":[0.0033682743,0.0047400356,0.9793924,0.0071815685,0.0004522064,0.000034247285,0.00029362473,0.000517552,0.000115824594,0.00092678936,0.0026111465,0.00036630392],"about_ca_topic_score_codex":0.00058608514,"about_ca_topic_score_gemma":0.00006040316,"teacher_disagreement_score":0.52556676,"about_ca_system_score_codex":0.00003370684,"about_ca_system_score_gemma":0.0004908993,"threshold_uncertainty_score":0.999885},"labels":[],"label_agreement":null},{"id":"W2069741947","doi":"10.1080/10401334.2011.536897","title":"The Effectiveness of Cognitive Forcing Strategies to Decrease Diagnostic Error: An Exploratory Study","year":2011,"lang":"en","type":"article","venue":"Teaching and Learning in Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":81,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hamilton Health Sciences; McMaster University","funders":"","keywords":"Cognition; Forcing (mathematics); Exploratory research; Psychology; Applied psychology; Medicine; Medical education; Psychiatry; Mathematics; Sociology","score_opus":0.0502170614760472,"score_gpt":0.37506440541895936,"score_spread":0.32484734394291215,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2069741947","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99573284,0.0005096941,0.0006575476,0.000116867486,0.00010835509,0.0006612684,4.955087e-7,0.00006873393,0.0021442238],"genre_scores_gemma":[0.99940974,0.000039357692,0.0001591835,0.00015293146,0.00009493584,0.00007545635,0.000008575829,0.000023236127,0.000036606605],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9961669,0.002585236,0.00039543377,0.00032238243,0.00027370063,0.00025638295],"domain_scores_gemma":[0.9147127,0.08444581,0.00010846638,0.00021472978,0.00009736889,0.00042089797],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.010616367,0.00017331068,0.00050089386,0.0001423448,0.00028591673,0.000013271068,0.00009012108,0.000059604623,0.00002305278],"category_scores_gemma":[0.38421816,0.00010815933,0.000031396765,0.00015881908,0.00022947407,0.00008154138,0.000053673473,0.0010813203,0.0000031841405],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002291537,0.0010970853,0.8423762,0.00009062105,0.00008628059,0.00019938851,0.10297977,0.00009391195,0.00015859288,0.00059528276,0.000020994152,0.050010365],"study_design_scores_gemma":[0.0033775235,0.007416991,0.80344564,0.008712501,0.00020627781,0.00001749584,0.17618914,0.0001632486,0.000029209752,0.00030061876,0.000022532964,0.00011884386],"about_ca_topic_score_codex":0.0018503048,"about_ca_topic_score_gemma":0.00024128178,"teacher_disagreement_score":0.3736018,"about_ca_system_score_codex":0.000024453644,"about_ca_system_score_gemma":0.00012779653,"threshold_uncertainty_score":0.6209688},"labels":[],"label_agreement":null},{"id":"W2070184193","doi":"10.1007/s10459-007-9072-4","title":"Is it science? A study of the attitudes of medical trainees and physicians toward qualitative and quantitative research","year":2007,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; St. Michael's Hospital","funders":"Canadian Institutes of Health Research; Woodrow Wilson International Center for Scholars","keywords":"Qualitative research; Categorization; Construct (python library); Psychology; Qualitative property; Coding (social sciences); Preference; Medical education; Qualitative analysis; Applied psychology; Medicine; Computer science; Sociology; Social science","score_opus":0.23511332382078065,"score_gpt":0.6459115192397108,"score_spread":0.41079819541893015,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2070184193","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98058367,0.004341254,0.000039564868,0.013971089,0.00018057293,0.00046387655,0.0000014343804,0.000002154534,0.00041636586],"genre_scores_gemma":[0.99573404,0.0021252336,0.0015243894,0.00056177535,0.000025640558,0.00001224009,2.4346468e-7,0.0000018794457,0.0000145482445],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9974855,0.00021305055,0.00045957713,0.00028384515,0.001323365,0.0002346601],"domain_scores_gemma":[0.986194,0.01296789,0.00022229478,0.00013952394,0.00032810943,0.0001481789],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.01452693,0.00005411019,0.00021771752,0.00023549284,0.00019097458,0.000009663376,0.00018662281,0.000025696718,0.000006508243],"category_scores_gemma":[0.03335316,0.000034499382,0.0000132577525,0.0016489528,0.0037467247,0.00016701425,0.00006115058,0.000175371,2.0614128e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005705848,0.0018110201,0.5271752,0.00030029306,0.0000040386403,3.480115e-7,0.39054573,0.000005421127,0.000032136544,0.011351197,0.000069998845,0.06864757],"study_design_scores_gemma":[0.00033209202,0.0016816236,0.5938446,0.002399523,0.0000030781557,0.0000022888905,0.3993453,0.000095234136,0.00010990235,0.0021207011,0.00003532035,0.000030359379],"about_ca_topic_score_codex":0.0018677603,"about_ca_topic_score_gemma":0.0012161711,"teacher_disagreement_score":0.06861721,"about_ca_system_score_codex":0.0000496947,"about_ca_system_score_gemma":0.0018252683,"threshold_uncertainty_score":0.9989645},"labels":[],"label_agreement":null},{"id":"W2070415144","doi":"10.1097/acm.0b013e31827b2941","title":"Clinical Decision Making","year":2013,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University; Hamilton Health Sciences","funders":"","keywords":"Assertion; Interpretation (philosophy); Reading (process); Dual (grammatical number); Process (computing); Computer science; Point (geometry); Psychology; Perspective (graphical); Inverse; Cognitive psychology; Epistemology; Calculus (dental); Artificial intelligence; Mathematics; Medicine; Philosophy","score_opus":0.08370502696823086,"score_gpt":0.45619187130802774,"score_spread":0.3724868443397969,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2070415144","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008091379,0.005556064,0.0010643153,0.96433264,0.0055048172,0.00094789383,0.000009540866,0.00028950788,0.014203861],"genre_scores_gemma":[0.0060924944,0.0063590007,0.001083177,0.907995,0.059136234,0.00005297248,0.00038026366,0.0001321767,0.018768666],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9938708,0.00021936798,0.0024821234,0.0011131286,0.0015032666,0.0008113093],"domain_scores_gemma":[0.93194795,0.065486945,0.00069902226,0.0011295477,0.0002744492,0.00046209467],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity","insufficient_payload"],"category_scores_codex":[0.002184013,0.000590456,0.0022084084,0.00030970984,0.0000706135,0.000009912942,0.00053188705,0.005123003,0.0056590526],"category_scores_gemma":[0.20810415,0.0003912934,0.00037986148,0.000317238,0.00072749116,0.00005167896,0.0002003332,0.014235272,0.0030206856],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006234037,0.000010747255,0.008431442,0.00008442017,0.00010849786,0.0009753354,0.00006726354,1.00323994e-7,0.0000019858114,0.000024950445,0.89247465,0.097758256],"study_design_scores_gemma":[0.0027590806,0.0007267936,0.007956507,0.04260048,0.00094255735,0.0005018272,0.000037740534,0.00007969892,0.0000010232646,0.0039267093,0.9401024,0.0003652052],"about_ca_topic_score_codex":0.000045761393,"about_ca_topic_score_gemma":3.3417237e-7,"teacher_disagreement_score":0.20592013,"about_ca_system_score_codex":0.00011213695,"about_ca_system_score_gemma":0.0003043104,"threshold_uncertainty_score":0.9998539},"labels":[],"label_agreement":null},{"id":"W2070631084","doi":"10.1111/acem.12440","title":"Reframing Diagnostic Error: Maybe It's Content, and Not Process, That Leads to Error","year":2014,"lang":"en","type":"letter","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"McMaster University","keywords":"Medicine; Cognitive reframing; Process (computing); Content (measure theory); Social psychology","score_opus":0.14389297520555716,"score_gpt":0.4115606087348453,"score_spread":0.26766763352928813,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2070631084","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.066137135,0.008148496,0.00026087297,0.9162751,0.004176723,0.0016102277,0.0000493307,0.00025920666,0.0030829448],"genre_scores_gemma":[0.14720966,0.015709994,0.000099087694,0.78808767,0.01775742,0.0003270365,0.0008054185,0.0002500123,0.029753737],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9926138,0.00021092065,0.0023577786,0.001694133,0.0017672181,0.0013561511],"domain_scores_gemma":[0.980732,0.015759015,0.0009117061,0.0010975694,0.00043807397,0.0010615839],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0017062089,0.0010783141,0.0023746332,0.00046948434,0.00018563082,0.000003546354,0.0006441869,0.002690436,0.0021993925],"category_scores_gemma":[0.18824737,0.00081711716,0.0002702841,0.00054439396,0.000542311,0.00010329032,0.00020981662,0.007039272,0.0004407392],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000107026994,0.000032121407,0.06452961,0.0008746956,0.00014226334,0.00033681074,0.0010236136,0.0000013434764,0.0001635821,0.000018156486,0.9315275,0.0012432773],"study_design_scores_gemma":[0.003069454,0.0016388524,0.0203723,0.02935907,0.0024042681,0.000387659,0.0006265772,0.00007010579,0.00014967311,0.0002690915,0.94047654,0.0011764045],"about_ca_topic_score_codex":0.00028053817,"about_ca_topic_score_gemma":0.000011313502,"teacher_disagreement_score":0.18654115,"about_ca_system_score_codex":0.00013246597,"about_ca_system_score_gemma":0.0002143189,"threshold_uncertainty_score":0.999428},"labels":[],"label_agreement":null},{"id":"W2071700609","doi":"10.1007/s11606-009-0953-5","title":"Taking the Stress out of Morning Report: An Analytic Approach to the Differential Diagnosis","year":2009,"lang":"en","type":"article","venue":"Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Mount Sinai Hospital; University Health Network","funders":"","keywords":"Medicine; Facilitator; Session (web analytics); Differential diagnosis; Medical education; Pathology; Social psychology; Psychology; Computer science","score_opus":0.06309059494213207,"score_gpt":0.38957670569767344,"score_spread":0.32648611075554135,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2071700609","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9657105,0.0006733146,0.0072063575,0.02066833,0.0013395572,0.00020091205,0.0000033956478,0.000009182644,0.004188458],"genre_scores_gemma":[0.9904769,0.0001708812,0.0008556043,0.002483237,0.0050935824,0.000004,0.0000070156084,0.00001436845,0.00089443225],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9972129,0.00014230232,0.001291247,0.00020053974,0.00090833777,0.0002446958],"domain_scores_gemma":[0.9957892,0.0017313544,0.001235231,0.00044574385,0.000431796,0.00036672738],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0013219175,0.00018741642,0.0007333904,0.00016143305,0.00007053083,0.00002496321,0.0004380076,0.00007054753,0.00017656857],"category_scores_gemma":[0.02041635,0.00007970119,0.00026901564,0.00020372852,0.00014716861,0.00005798247,0.000048344926,0.0006393103,0.0000021772503],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00485384,0.005850535,0.42454755,0.00013742488,0.0025408247,0.004762743,0.018067097,0.0071647293,0.021265013,0.0012120613,0.1370246,0.37257358],"study_design_scores_gemma":[0.0039151814,0.008393084,0.9595426,0.008631519,0.0018283451,0.004096801,0.0014088816,0.0059483587,0.0031668027,0.00030029708,0.002536184,0.00023200219],"about_ca_topic_score_codex":0.00025927552,"about_ca_topic_score_gemma":0.000017638133,"teacher_disagreement_score":0.534995,"about_ca_system_score_codex":0.000053703643,"about_ca_system_score_gemma":0.00007448384,"threshold_uncertainty_score":0.9878351},"labels":[],"label_agreement":null},{"id":"W2072426503","doi":"10.1111/j.1553-2712.2008.00035.x","title":"International Survey of Emergency Physicians’ Priorities for Clinical Decision Rules","year":2008,"lang":"en","type":"article","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":115,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Ottawa Public Health; University of Ottawa","funders":"","keywords":"Medicine; Emergency medicine; Chest pain; Lumbar puncture; Emergency department; Weakness; Pediatrics; Family medicine; Surgery; Internal medicine","score_opus":0.2057009485899969,"score_gpt":0.4904963981012219,"score_spread":0.28479544951122504,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2072426503","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9754832,0.006337303,0.002040862,0.0017646402,0.010855732,0.00048381163,0.00010629449,0.00006435769,0.002863806],"genre_scores_gemma":[0.795704,0.1961243,0.0012669074,0.0002828606,0.0033406143,0.000039915867,0.00034884395,0.000043499298,0.002849056],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99564713,0.00011002848,0.0026841436,0.00048086845,0.0007446268,0.0003331707],"domain_scores_gemma":[0.9890301,0.008781413,0.00062861573,0.00039867937,0.00091129355,0.00024986741],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0023259984,0.00023680188,0.00089462363,0.0001463662,0.00008522823,1.7696667e-7,0.00035281517,0.00035691357,0.0019591826],"category_scores_gemma":[0.15040757,0.00018617137,0.00031391592,0.00032952544,0.00031402422,0.00007701257,0.00007634802,0.00057863444,0.000039820847],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00047885437,0.00015608799,0.77150893,0.000032025924,0.00012391772,0.0000039595834,0.00020230185,9.484253e-7,0.00016646174,0.00017581445,0.21219265,0.014958045],"study_design_scores_gemma":[0.0026355989,0.0008201386,0.9748994,0.0012814165,0.00017615,0.000014223216,0.000067682064,0.00020779845,0.00007568058,0.001238235,0.018411504,0.0001721811],"about_ca_topic_score_codex":0.00031233678,"about_ca_topic_score_gemma":0.000014868435,"teacher_disagreement_score":0.20339046,"about_ca_system_score_codex":0.000028895729,"about_ca_system_score_gemma":0.00012573008,"threshold_uncertainty_score":0.99895316},"labels":[],"label_agreement":null},{"id":"W2072443975","doi":"10.1097/00001888-200010001-00031","title":"The Effects of Examiner Background, Station Organization, and Time of Exam on OSCE Scores Assessing Undergraduate Medical Studentsʼ Physical Examination Skills","year":2000,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":22,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Foothills Medical Centre","funders":"","keywords":"Objective structured clinical examination; Reliability (semiconductor); Medical education; Psychology; Quality (philosophy); Physical examination; Educational measurement; Medicine; Medical physics; Curriculum; Surgery; Pedagogy","score_opus":0.012001898993091908,"score_gpt":0.347376737098911,"score_spread":0.3353748381058191,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2072443975","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9955787,0.00049497414,0.00035349317,0.0022991842,0.00014593905,0.0003662735,0.0000015128205,0.00003317981,0.00072674826],"genre_scores_gemma":[0.99325514,0.0038496777,0.000034287757,0.00062860834,0.00041766174,0.000008166162,0.00007201177,0.000026851309,0.0017075783],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.997365,0.0002781269,0.0006443724,0.00029152117,0.0012143526,0.00020665354],"domain_scores_gemma":[0.96880597,0.030229757,0.00026713658,0.00024685607,0.0002331853,0.00021710731],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0012413466,0.00017919387,0.0005149299,0.00008690161,0.0000858689,0.0000073688398,0.00015356566,0.00020019275,0.00016098551],"category_scores_gemma":[0.0657064,0.00010781258,0.000029581253,0.00043642317,0.00060234644,0.00008353504,0.000035615045,0.00045199253,0.000022920025],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00039664577,0.0016049308,0.153718,0.0008794142,0.0005560235,0.000098663484,0.0114459945,0.000034321438,0.03304079,0.010285737,0.013961523,0.773978],"study_design_scores_gemma":[0.0073514297,0.0023753948,0.96733516,0.011610866,0.00061014114,0.000063147825,0.0006362857,0.00092285563,0.0069365427,0.001708826,0.00023848767,0.0002108428],"about_ca_topic_score_codex":0.000030288638,"about_ca_topic_score_gemma":7.582384e-7,"teacher_disagreement_score":0.81361717,"about_ca_system_score_codex":0.00004122415,"about_ca_system_score_gemma":0.000108058346,"threshold_uncertainty_score":0.9421635},"labels":[],"label_agreement":null},{"id":"W2072760764","doi":"10.1515/dx-2013-0022","title":"How might mathematics education be used to improve diagnostic reasoning?","year":2014,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Mathematics education; Computer science; Psychology","score_opus":0.02228194299328585,"score_gpt":0.31961017257558677,"score_spread":0.29732822958230093,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2072760764","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9409352,0.0003104619,0.0013580633,0.052069765,0.000947482,0.0010905099,0.00002864688,0.00022361721,0.0030362469],"genre_scores_gemma":[0.9758015,0.0001481992,0.01081869,0.008580477,0.0009231629,0.001057924,0.00008896924,0.00007157133,0.0025094822],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99809736,0.00006439348,0.00038676278,0.0005304712,0.00047973444,0.0004412564],"domain_scores_gemma":[0.9703481,0.027719611,0.00015642057,0.0008091406,0.00024546325,0.0007212719],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00046781043,0.00028106698,0.0005465481,0.00015046893,0.000098140605,0.00013589894,0.0001825507,0.00019045836,0.00016642814],"category_scores_gemma":[0.26000574,0.00023984077,0.00020771561,0.00034310634,0.00006637929,0.00008536175,0.000077067554,0.00023161889,0.0003080452],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000042201365,0.0044872616,0.5889327,0.00024647033,0.00014633987,0.000029208477,0.0016337724,0.000009565234,0.00040287155,0.009810479,0.23283204,0.16142711],"study_design_scores_gemma":[0.005396527,0.003208748,0.5658202,0.013674165,0.0017702654,0.00009079977,0.0013157302,0.0011018652,0.027753623,0.0042791604,0.3738543,0.0017345764],"about_ca_topic_score_codex":0.000121502366,"about_ca_topic_score_gemma":0.000039046783,"teacher_disagreement_score":0.2595379,"about_ca_system_score_codex":0.00009874655,"about_ca_system_score_gemma":0.0001833638,"threshold_uncertainty_score":0.9780423},"labels":[],"label_agreement":null},{"id":"W2072875194","doi":"10.1186/1472-6920-13-97","title":"Evaluation of standardized doctor's orders as an educational tool for undergraduate medical students: a prospective cohort study","year":2013,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Toronto Western Hospital; St. Michael's Hospital","funders":"University of Toronto","keywords":"Medical education; Medicine; Educational measurement; Cohort study; Cohort; Psychology; Curriculum; Pedagogy; Internal medicine","score_opus":0.030168919834142383,"score_gpt":0.4609439180495793,"score_spread":0.4307749982154369,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2072875194","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9826973,0.00008731023,0.0012915335,0.0074078026,0.0010545552,0.0065236357,0.0000055040596,0.000041588708,0.0008907861],"genre_scores_gemma":[0.9913423,0.000034220662,0.0017788104,0.0010944111,0.0007259136,0.004133138,0.0002982792,0.000027682849,0.0005652659],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99062693,0.00055274693,0.0008429348,0.000545185,0.0071739308,0.00025824463],"domain_scores_gemma":[0.9824127,0.0103659965,0.0003785625,0.0006945978,0.004938201,0.0012099664],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.008823237,0.00019206922,0.0005013527,0.00012757424,0.00009240774,0.000052620588,0.000297531,0.00022385948,0.007872311],"category_scores_gemma":[0.3892727,0.00015431202,0.00012957116,0.0002812947,0.0001978466,0.00015319885,0.00005173121,0.00024307422,0.00009083414],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002584528,0.012914653,0.9390553,0.000054423086,0.00026236533,4.0267327e-7,0.0004604495,0.0000031705038,0.0000038077203,0.0018319407,0.010640202,0.034514792],"study_design_scores_gemma":[0.009032464,0.0014229805,0.9692554,0.0009371909,0.0009906399,0.000017630337,0.0012810283,0.0016781266,0.000018043742,0.015038698,0.00015215078,0.00017563907],"about_ca_topic_score_codex":0.0011158217,"about_ca_topic_score_gemma":0.00033696985,"teacher_disagreement_score":0.38044944,"about_ca_system_score_codex":0.0004235011,"about_ca_system_score_gemma":0.026518457,"threshold_uncertainty_score":0.9930346},"labels":[],"label_agreement":null},{"id":"W2073265108","doi":"10.1111/j.1741-6787.2011.00237.x","title":"Recognizing and Responding to Uncertainty: A Grounded Theory of Nurses’ Uncertainty","year":2012,"lang":"en","type":"article","venue":"Worldviews on Evidence-Based Nursing","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":46,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"St. Lawrence College; University of Toronto; University of Victoria; University of Alberta; Ministry of Health and Long Term Care","funders":"Canadian Institutes of Health Research","keywords":"Grounded theory; Psychology; Uncertainty theory; Epistemology; Sociology; Social psychology; Mathematics; Qualitative research; Statistics; Social science; Philosophy","score_opus":0.09049349089817829,"score_gpt":0.3958179681374201,"score_spread":0.30532447723924183,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2073265108","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.975024,0.0074635316,0.0014316808,0.00834919,0.0012491753,0.0011090817,0.00000919724,0.00011276882,0.0052514137],"genre_scores_gemma":[0.99101526,0.00018611082,0.003504757,0.0039846026,0.00040489083,0.00004688881,0.0000119519245,0.000044748922,0.00080078625],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9969496,0.0005490462,0.0007347053,0.0005151481,0.0005613377,0.0006901206],"domain_scores_gemma":[0.97159714,0.026644262,0.00027696576,0.00058696646,0.00017055299,0.00072413986],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.003706161,0.00032175542,0.00078691065,0.0007986047,0.00016135752,0.00004014924,0.00014482297,0.0001416619,0.00017341609],"category_scores_gemma":[0.062971294,0.00026312124,0.00021894486,0.0013220815,0.00030187613,0.00019575532,0.000026822629,0.0003951922,0.00006616754],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0070061767,0.00093617616,0.022883309,0.00015493226,0.000060271384,0.000022127486,0.002636985,0.00030515555,0.0021646009,0.009121052,0.002648229,0.952061],"study_design_scores_gemma":[0.011922074,0.009731564,0.2163191,0.69149554,0.0029113935,0.00024025631,0.008728728,0.002556481,0.016165804,0.012542025,0.024675785,0.0027112523],"about_ca_topic_score_codex":0.000057369718,"about_ca_topic_score_gemma":0.0000070163865,"teacher_disagreement_score":0.94934976,"about_ca_system_score_codex":0.0004013989,"about_ca_system_score_gemma":0.00022662577,"threshold_uncertainty_score":0.9999821},"labels":[],"label_agreement":null},{"id":"W2073773377","doi":"10.1007/s10459-008-9120-8","title":"Script concordance testing: more cases or more questions?","year":2008,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":66,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Concordance; Reliability (semiconductor); Variance (accounting); Nested set model; Statistics; Variance components; Test (biology); Computer science; Mathematics; Medicine; Data mining; Internal medicine","score_opus":0.0905657237337662,"score_gpt":0.4756656871024329,"score_spread":0.3850999633686667,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2073773377","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9566485,0.02235373,0.0001478815,0.016085263,0.0017938203,0.0006055338,0.000004986657,0.000098671066,0.0022615958],"genre_scores_gemma":[0.95365894,0.006691803,0.03070287,0.0076543605,0.00031016447,0.000102407976,0.000013205696,0.000008365758,0.0008579015],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9983318,0.00005687336,0.0005128377,0.00040483294,0.00034510755,0.00034856176],"domain_scores_gemma":[0.9926722,0.006371744,0.00023997831,0.00023489115,0.00018907378,0.0002921139],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00059449096,0.000117133735,0.00026405483,0.00014803399,0.0003534445,0.000013261149,0.0001294073,0.00004150812,0.00003655527],"category_scores_gemma":[0.061228402,0.00008804929,0.000026376023,0.0012172592,0.0007810014,0.0003248256,0.000018126719,0.00017303629,0.000016129337],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000057881385,0.0005367247,0.8698336,0.00012584068,0.000001163384,0.00003431254,0.00082547613,0.000435483,0.000004879251,0.0006628986,0.002731446,0.12475029],"study_design_scores_gemma":[0.00083241635,0.0020241512,0.9505214,0.007749124,0.000010611149,0.0018294636,0.0034019533,0.001728214,0.00005406358,0.0012183932,0.030352117,0.0002780979],"about_ca_topic_score_codex":0.0019924606,"about_ca_topic_score_gemma":0.0002060785,"teacher_disagreement_score":0.124472186,"about_ca_system_score_codex":0.00017360128,"about_ca_system_score_gemma":0.00430003,"threshold_uncertainty_score":0.9466793},"labels":[],"label_agreement":null},{"id":"W2075034827","doi":"10.1051/pmed/2012014","title":"Aborder le raisonnement clinique du point de vue pédagogique","year":2012,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Humanities; Philosophy","score_opus":0.039356311083504326,"score_gpt":0.36381620913121865,"score_spread":0.3244598980477143,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2075034827","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3982676,0.20325738,0.039880488,0.23486461,0.017563542,0.0022304354,0.00022610986,0.0008738465,0.10283601],"genre_scores_gemma":[0.8442542,0.027145552,0.010718736,0.022635687,0.01601852,0.00027630504,0.0001402672,0.00022126322,0.07858948],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9939491,0.00076280936,0.0013598906,0.0007429188,0.00070184085,0.0024834105],"domain_scores_gemma":[0.97879654,0.016951386,0.00032422383,0.0010361396,0.00024884575,0.00264288],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.004763875,0.0006192605,0.001081337,0.00014070707,0.0002868684,0.00008299341,0.0004205311,0.0010297544,0.009857202],"category_scores_gemma":[0.07338615,0.00059604633,0.0005973869,0.00040923874,0.0008038801,0.00026800347,0.00033123075,0.0014482298,0.0030281793],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004271191,0.0137556745,0.22814932,0.00046647113,0.0005690587,0.00058021746,0.012068621,0.000069218026,0.00047969216,0.019369483,0.43802917,0.28603595],"study_design_scores_gemma":[0.005734955,0.00094307546,0.07072619,0.00467687,0.00066591153,0.0003728625,0.004562546,0.0008373698,0.0024607095,0.0022457237,0.9058205,0.000953298],"about_ca_topic_score_codex":0.0009767693,"about_ca_topic_score_gemma":0.00006615445,"teacher_disagreement_score":0.46779132,"about_ca_system_score_codex":0.00046045092,"about_ca_system_score_gemma":0.0017280233,"threshold_uncertainty_score":0.9996491},"labels":[],"label_agreement":null},{"id":"W2075571348","doi":"10.1111/j.2042-3306.2007.tb01009.x","title":"Does clinical experience make up for failure to keep up to date?","year":2007,"lang":"en","type":"article","venue":"Equine Veterinary Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Citation; Library science; Medicine; Computer science","score_opus":0.10198547911648846,"score_gpt":0.46964326644486704,"score_spread":0.36765778732837856,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2075571348","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96729124,0.000065568616,0.017147554,0.009172602,0.0054758554,0.0004790241,0.000016133697,0.000059348058,0.00029266323],"genre_scores_gemma":[0.95959365,0.000058670164,0.025867725,0.0058777127,0.0036947234,0.000030750914,0.000013604223,0.000044475113,0.0048186956],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99728554,0.000059619393,0.0010823375,0.0004517049,0.00042529785,0.000695495],"domain_scores_gemma":[0.9931676,0.0043984875,0.00016331895,0.00043861792,0.00026797762,0.001564046],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002152108,0.00023384886,0.00055381906,0.00013930991,0.00017555318,0.000078098696,0.00029429953,0.00014295669,0.0004763556],"category_scores_gemma":[0.028808191,0.00014492763,0.0003149879,0.00024296205,0.00008326349,0.000078079385,0.00019474643,0.00050529063,0.0001602491],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.026182322,0.0015353011,0.0881505,0.00011599848,0.00036282625,0.008372428,0.006305399,0.000028765427,0.045240827,0.00024645712,0.16550395,0.6579552],"study_design_scores_gemma":[0.008693214,0.05382589,0.07416078,0.0031597828,0.00017538574,0.009427438,0.0032023075,0.00007114547,0.0018384531,0.00034087623,0.8443467,0.0007580291],"about_ca_topic_score_codex":0.0000056066256,"about_ca_topic_score_gemma":0.000008093116,"teacher_disagreement_score":0.6788428,"about_ca_system_score_codex":0.000054289183,"about_ca_system_score_gemma":0.00012368317,"threshold_uncertainty_score":0.97937256},"labels":[],"label_agreement":null},{"id":"W2075844793","doi":"10.1161/strokeaha.114.005462","title":"Decision Making in Acute Stroke Care","year":2014,"lang":"en","type":"review","venue":"Stroke","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":28,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Michael's Hospital","funders":"Heart and Stroke Foundation of Canada","keywords":"Medicine; Stroke (engine); Neurology; Acute stroke; Unit (ring theory); Health care; Family medicine; Medical emergency; Emergency department; Psychiatry; Psychology; Law","score_opus":0.04137110674556419,"score_gpt":0.4267353987337427,"score_spread":0.3853642919881785,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2075844793","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00017614164,0.98966175,0.00016457165,0.000020584923,0.0008182073,0.00064179185,0.00020376091,0.00006233372,0.00825086],"genre_scores_gemma":[0.0007076371,0.9942776,0.0021335178,0.00014954864,0.00053384126,0.00006328953,0.00021195247,0.000071847804,0.001850752],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99755603,0.00009908816,0.00089377636,0.0005821551,0.00044402247,0.0004249463],"domain_scores_gemma":[0.9876435,0.011087695,0.00027340665,0.0007119925,0.00007970103,0.00020370555],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00033331313,0.00040298485,0.002540738,0.00029855323,0.000036883488,0.000032174456,0.00022838592,0.00058226235,0.0001646197],"category_scores_gemma":[0.016850578,0.000298328,0.00083503214,0.00020259844,0.00005237084,0.000019623349,0.00012527437,0.0009017823,0.00043581496],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000027091268,0.0000403683,0.00048677603,0.0021674777,0.00009891249,0.00022990999,0.000029109944,3.4797645e-7,4.7943246e-8,0.000034185137,0.0009951573,0.9958906],"study_design_scores_gemma":[0.0007608487,0.00020457798,0.00033973658,0.17871034,0.0014864305,0.00007841599,0.000023707009,0.000003972033,2.2982378e-7,0.000017751881,0.81814206,0.00023190743],"about_ca_topic_score_codex":0.000016127382,"about_ca_topic_score_gemma":0.000028413418,"teacher_disagreement_score":0.9956587,"about_ca_system_score_codex":0.0002222323,"about_ca_system_score_gemma":0.00037289422,"threshold_uncertainty_score":0.9999469},"labels":[],"label_agreement":null},{"id":"W2076578613","doi":"10.1097/00001888-200010001-00041","title":"The Epistemology of Clinical Reasoning","year":2000,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":71,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Health Sciences Centre","funders":"","keywords":"Centrality; Medical diagnosis; Process (computing); Generalization; Psychology; Focus (optics); Computer science; Cognitive science; Epistemology; Medicine","score_opus":0.0688304392211007,"score_gpt":0.4475453438192548,"score_spread":0.3787149045981541,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2076578613","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.91184705,0.005101431,0.000052941286,0.02692708,0.00066726067,0.00025176126,0.0000014460262,0.000057235586,0.055093765],"genre_scores_gemma":[0.9617275,0.013299881,0.00014676552,0.0038004296,0.0016497747,0.00001017315,0.000012627851,0.000017551018,0.019335276],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99738187,0.00020025407,0.0014345336,0.00028146058,0.0003896786,0.00031220194],"domain_scores_gemma":[0.969748,0.029101662,0.00025600323,0.00047415463,0.000099257035,0.00032092855],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0036269412,0.00013016586,0.0007406154,0.000033544286,0.00008185451,0.0000012467242,0.00021770479,0.00035329198,0.0014350845],"category_scores_gemma":[0.11754965,0.00006878597,0.00012408549,0.0002097745,0.0010954984,0.00001797199,0.000028729955,0.0010613247,0.0001526359],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010646669,0.000051469397,0.2735051,0.00003129084,0.00015658206,0.00008354243,0.0005709285,0.0000028804436,0.00008852599,0.008431261,0.11355959,0.6024542],"study_design_scores_gemma":[0.0107702045,0.0031662555,0.54009223,0.008344704,0.0007585486,0.00086516835,0.0007833101,0.00073260453,0.0002089321,0.0068042944,0.4271989,0.00027485754],"about_ca_topic_score_codex":0.000088370594,"about_ca_topic_score_gemma":0.000001728868,"teacher_disagreement_score":0.6021793,"about_ca_system_score_codex":0.000013275485,"about_ca_system_score_gemma":0.00014041166,"threshold_uncertainty_score":0.99947774},"labels":[],"label_agreement":null},{"id":"W2077210274","doi":"10.1007/s40037-014-0126-z","title":"Conscious versus unconscious thinking in the medical domain: the deliberation-without-attention effect examined","year":2014,"lang":"en","type":"article","venue":"Perspectives on Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Unconscious mind; Deliberation; Psychology; Task (project management); Cognitive psychology; Clinical psychology; Psychoanalysis","score_opus":0.012316799011007922,"score_gpt":0.3550769179588404,"score_spread":0.3427601189478325,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2077210274","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.88868594,0.00041631883,0.000931956,0.09744034,0.0016183616,0.0007149051,6.742535e-7,0.000073790216,0.010117686],"genre_scores_gemma":[0.9886494,0.00016685763,0.00015675147,0.009242259,0.001429022,0.00018043019,0.000045787365,0.000019028505,0.000110439614],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99577165,0.0010435506,0.00044650145,0.00046575433,0.0019698695,0.00030266066],"domain_scores_gemma":[0.94243044,0.055181716,0.00026082958,0.0010695509,0.0003429072,0.00071457407],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005550156,0.00020500676,0.00032510827,0.000116395655,0.00029106773,0.00007608096,0.00037070352,0.0002782598,0.00046794498],"category_scores_gemma":[0.22314169,0.00010707023,0.00014182975,0.00036333705,0.00050570624,0.00004935448,0.00003698534,0.00096015667,0.00011630265],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0050565167,0.010697381,0.2143943,0.00015106668,0.00041239566,0.00015369325,0.06696431,0.000023562992,0.000057331505,0.17026779,0.031865835,0.49995583],"study_design_scores_gemma":[0.020411875,0.0049040304,0.9034122,0.00395654,0.00041024922,0.0005185184,0.040399596,0.0040750974,0.00003367274,0.015432664,0.0058143116,0.00063120795],"about_ca_topic_score_codex":0.00017400067,"about_ca_topic_score_gemma":0.00022380764,"teacher_disagreement_score":0.68901795,"about_ca_system_score_codex":0.00020589714,"about_ca_system_score_gemma":0.0010641256,"threshold_uncertainty_score":0.78340214},"labels":[],"label_agreement":null},{"id":"W2078048424","doi":"10.1353/pbm.0.0088","title":"Theory-Based Medicine and the Role of Evidence: Why the Emperor Needs New Clothes, Again","year":2009,"lang":"en","type":"article","venue":"Perspectives in biology and medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":33,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Evidence-based medicine; Randomized controlled trial; Scientific evidence; Emperor; Empirical evidence; Alternative medicine; Epistemology; Psychology; MEDLINE; Core (optical fiber); Positive economics; Medicine; Law; Philosophy; Political science; Computer science; Economics; History; Surgery","score_opus":0.05446294345501092,"score_gpt":0.41028063367959394,"score_spread":0.35581769022458304,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2078048424","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.39926136,0.20033924,0.00040787004,0.39327276,0.00016026257,0.00072770944,0.000002351966,0.00003497228,0.005793481],"genre_scores_gemma":[0.97696877,0.011555398,0.00010801752,0.010546168,0.0005353891,0.000009937274,0.0000040410528,0.0000068837057,0.00026542408],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9985779,0.0004055333,0.0003818472,0.00027247594,0.00014676804,0.00021547916],"domain_scores_gemma":[0.9718792,0.027440326,0.00012839801,0.00032143254,0.00008689175,0.00014376688],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.002696793,0.00017346443,0.0006656981,0.00015822078,0.00008507523,0.0000025773277,0.00013206023,0.00014492327,0.00015064156],"category_scores_gemma":[0.066574484,0.0000638229,0.000052924315,0.00032641395,0.0038843118,0.000021218879,0.000027177728,0.00036972278,9.143034e-7],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0060139173,0.00027027365,0.0922057,0.000031226595,0.00017306903,0.000016917464,0.053171873,0.0000023161704,0.003707164,0.7133953,0.0034250119,0.1275872],"study_design_scores_gemma":[0.022069419,0.006706874,0.31308103,0.00850366,0.0007162741,0.00012856998,0.06879141,0.0003345203,0.00040643595,0.5732748,0.0057476396,0.00023936834],"about_ca_topic_score_codex":0.00056915137,"about_ca_topic_score_gemma":0.000019750703,"teacher_disagreement_score":0.5777074,"about_ca_system_score_codex":0.000023218183,"about_ca_system_score_gemma":0.00009674074,"threshold_uncertainty_score":0.99882656},"labels":[],"label_agreement":null},{"id":"W2078213917","doi":"10.1097/00001888-200002000-00020","title":"Scripts and Medical Diagnostic Knowledge","year":2000,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":446,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Scripting language; Categorization; Computer science; Task (project management); Cognition; Competence (human resources); Medical knowledge; Artificial intelligence; Natural language processing; Psychology; Medical education; Medicine; Social psychology; Programming language","score_opus":0.0343281786136696,"score_gpt":0.3798155505099628,"score_spread":0.3454873718962932,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2078213917","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8844816,0.02556012,0.00005549963,0.03430721,0.00056589395,0.00035052278,0.000002651462,0.00018410987,0.05449235],"genre_scores_gemma":[0.9465497,0.026040684,0.00004649054,0.01146978,0.0020150275,0.000026410306,0.00003267841,0.000029555189,0.013789642],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9978277,0.000077844365,0.0006110671,0.000435229,0.00064119115,0.0004069438],"domain_scores_gemma":[0.9768801,0.02145875,0.00005631499,0.0003061587,0.00006636991,0.0012323571],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0010811563,0.00021717777,0.00060219364,0.00009522468,0.000072680035,0.0000035242379,0.00015521544,0.0004726587,0.012849303],"category_scores_gemma":[0.1623851,0.00014978409,0.000047729765,0.00026895793,0.0006811444,0.00004397963,0.00004975348,0.0010975688,0.00070860545],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001805009,0.000103230326,0.082662985,0.00009351908,0.00008062433,0.0005101938,0.0020106568,3.1814605e-7,0.000089685156,0.0019779205,0.31014046,0.6021499],"study_design_scores_gemma":[0.019730406,0.0025306297,0.39477006,0.035878558,0.0011024931,0.004220891,0.00063309446,0.0012320083,0.00018157138,0.0061721266,0.53277934,0.0007688205],"about_ca_topic_score_codex":0.00008810981,"about_ca_topic_score_gemma":0.000004329332,"teacher_disagreement_score":0.60138106,"about_ca_system_score_codex":0.000031427335,"about_ca_system_score_gemma":0.00017416335,"threshold_uncertainty_score":0.9880531},"labels":[],"label_agreement":null},{"id":"W2080490561","doi":"10.1067/mem.2003.22","title":"Cognitive forcing strategies in clinical decisionmaking","year":2003,"lang":"en","type":"article","venue":"Annals of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":433,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University; Dartmouth General Hospital; Capital District Health Authority","funders":"Agency for Healthcare Research and Quality","keywords":"Debiasing; Forcing (mathematics); Cognition; Medicine; Medical diagnosis; Metacognition; MEDLINE; Cognitive psychology; Intensive care medicine; Psychiatry; Psychology; Cognitive science; Pathology","score_opus":0.30342654719454465,"score_gpt":0.5395058199087454,"score_spread":0.2360792727142007,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2080490561","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9279414,0.0039163823,0.00087513716,0.0015884314,0.0013172461,0.00026866383,0.0000026932594,0.00002363778,0.06406645],"genre_scores_gemma":[0.99260813,0.005757351,0.0002279096,0.0006405592,0.0002892144,0.00000920011,0.000012847088,0.000017897084,0.0004369114],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9970444,0.00016656591,0.0016914767,0.00033851183,0.0004222494,0.00033679965],"domain_scores_gemma":[0.9888864,0.009711956,0.00032981148,0.00029269533,0.0004953184,0.0002838298],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0028031163,0.00016898462,0.00075727364,0.00018374683,0.000030796677,0.0000016025168,0.000079996025,0.00014851043,0.0026642997],"category_scores_gemma":[0.19493416,0.00012560243,0.00022504381,0.00051944726,0.00017270647,0.00007615824,0.00002118165,0.00033119755,0.000026381285],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007000607,0.0013194825,0.91767836,0.00011822009,0.00020743471,0.0001683774,0.00096122216,0.000019840785,0.00020199011,0.015562631,0.02430253,0.038759843],"study_design_scores_gemma":[0.0063187536,0.003635938,0.9378305,0.014772903,0.0002604406,0.000026017406,0.0060798544,0.0001320337,0.0008471437,0.025308302,0.0044529843,0.0003351367],"about_ca_topic_score_codex":0.00012354547,"about_ca_topic_score_gemma":0.000045806875,"teacher_disagreement_score":0.19213104,"about_ca_system_score_codex":0.0000057103794,"about_ca_system_score_gemma":0.00018056105,"threshold_uncertainty_score":0.9982474},"labels":[],"label_agreement":null},{"id":"W2080492880","doi":"10.1111/medu.12253","title":"Students' self‐explanations while solving unfamiliar cases: the role of biomedical knowledge","year":2013,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":49,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Sherbrooke","funders":"","keywords":"Psychology; Inference; Cognitive psychology; Computer science; Artificial intelligence","score_opus":0.012261243455259335,"score_gpt":0.3597735406992749,"score_spread":0.34751229724401556,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2080492880","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97278965,0.0018035495,0.00010514215,0.014922205,0.0013020084,0.0006077378,0.000004608708,0.0000867035,0.00837842],"genre_scores_gemma":[0.99567324,0.0002163853,0.00041911428,0.0020977552,0.0005698211,0.00016536281,0.00009296986,0.000013609629,0.00075176486],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9982105,0.00009614877,0.00045875434,0.00020705773,0.000818836,0.0002087265],"domain_scores_gemma":[0.9927083,0.005847922,0.00012935975,0.00043107755,0.00038837706,0.0004949264],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0005469462,0.000107308,0.00023194909,0.00008891027,0.000093564486,0.000021275071,0.00024899223,0.00016701056,0.00243213],"category_scores_gemma":[0.05884868,0.000067753244,0.00008825591,0.00032977652,0.00021776691,0.000050744893,0.00007436186,0.00029546666,0.00040482465],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000023910772,0.013076896,0.1736666,0.00009290768,0.00018598104,0.000008836795,0.0052685617,2.890506e-7,0.00022124001,0.0025398037,0.4374526,0.36746237],"study_design_scores_gemma":[0.0027110488,0.00086278626,0.6482858,0.004150687,0.0005577727,0.0003104726,0.0107498225,0.0046348344,0.00036504123,0.0023704763,0.3246346,0.00036664543],"about_ca_topic_score_codex":0.00075322075,"about_ca_topic_score_gemma":0.000026769461,"teacher_disagreement_score":0.4746192,"about_ca_system_score_codex":0.0000693939,"about_ca_system_score_gemma":0.0017822988,"threshold_uncertainty_score":0.9984798},"labels":[],"label_agreement":null},{"id":"W2080818078","doi":"10.1207/s15328015tlm1204_5","title":"The Script Concordance Test: A Tool to Assess the Reflective Clinician","year":2000,"lang":"en","type":"article","venue":"Teaching and Learning in Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":371,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Concordance; Test (biology); Scripting language; Face validity; Construct (python library); Reliability (semiconductor); Computer science; Construct validity; Elaboration; Psychology; Test script; Test validity; Medical education; Psychometrics; Medicine; Clinical psychology; Test case; Machine learning","score_opus":0.045362071416878334,"score_gpt":0.40393339590576177,"score_spread":0.35857132448888346,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2080818078","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.91845244,0.0010707759,0.00012343704,0.047642052,0.00021803394,0.000360263,5.629575e-7,0.00010549063,0.032026913],"genre_scores_gemma":[0.9714492,0.000427058,0.0006588607,0.010585748,0.0006892861,0.000035127658,0.0000043141977,0.000022328613,0.016128054],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99781704,0.0006831376,0.000473176,0.00033764838,0.0003321463,0.00035684038],"domain_scores_gemma":[0.9532883,0.046058536,0.00008882081,0.000338084,0.000051266863,0.00017496766],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0070378142,0.00016601525,0.00040219308,0.000052444077,0.0007710144,0.000037828107,0.00015537042,0.000080509744,0.00010308609],"category_scores_gemma":[0.21056923,0.00008188607,0.000046478366,0.0002329931,0.00030864924,0.000029162966,0.00003478888,0.0025977623,0.00004679701],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00031933657,0.00010219158,0.13164423,0.000010586145,0.000027974627,0.000057680827,0.006398117,0.00012288788,0.00010819105,0.00038241464,0.009569797,0.8512566],"study_design_scores_gemma":[0.0028299226,0.0025830152,0.6363804,0.004604984,0.000114341055,0.00012835248,0.0045506964,0.0022375875,0.000006754681,0.00042993264,0.34592637,0.00020768846],"about_ca_topic_score_codex":0.0006728537,"about_ca_topic_score_gemma":0.00007648646,"teacher_disagreement_score":0.8510489,"about_ca_system_score_codex":0.000049642473,"about_ca_system_score_gemma":0.0000769013,"threshold_uncertainty_score":0.9997033},"labels":[],"label_agreement":null},{"id":"W2081042077","doi":"10.1097/acm.0b013e31827b258d","title":"Clinical Decision Making","year":2013,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medical diagnosis; Cognition; Psychology; Cognitive psychology; Interpretation (philosophy); Computer science; Medicine; Psychiatry","score_opus":0.08370502696823086,"score_gpt":0.45619187130802774,"score_spread":0.3724868443397969,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2081042077","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008091379,0.005556064,0.0010643153,0.96433264,0.0055048172,0.00094789383,0.000009540866,0.00028950788,0.014203861],"genre_scores_gemma":[0.0060924944,0.0063590007,0.001083177,0.907995,0.059136234,0.00005297248,0.00038026366,0.0001321767,0.018768666],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9938708,0.00021936798,0.0024821234,0.0011131286,0.0015032666,0.0008113093],"domain_scores_gemma":[0.93194795,0.065486945,0.00069902226,0.0011295477,0.0002744492,0.00046209467],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity","insufficient_payload"],"category_scores_codex":[0.002184013,0.000590456,0.0022084084,0.00030970984,0.0000706135,0.000009912942,0.00053188705,0.005123003,0.0056590526],"category_scores_gemma":[0.20810415,0.0003912934,0.00037986148,0.000317238,0.00072749116,0.00005167896,0.0002003332,0.014235272,0.0030206856],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006234037,0.000010747255,0.008431442,0.00008442017,0.00010849786,0.0009753354,0.00006726354,1.00323994e-7,0.0000019858114,0.000024950445,0.89247465,0.097758256],"study_design_scores_gemma":[0.0027590806,0.0007267936,0.007956507,0.04260048,0.00094255735,0.0005018272,0.000037740534,0.00007969892,0.0000010232646,0.0039267093,0.9401024,0.0003652052],"about_ca_topic_score_codex":0.000045761393,"about_ca_topic_score_gemma":3.3417237e-7,"teacher_disagreement_score":0.20592013,"about_ca_system_score_codex":0.00011213695,"about_ca_system_score_gemma":0.0003043104,"threshold_uncertainty_score":0.9998539},"labels":[],"label_agreement":null},{"id":"W2081711362","doi":"10.1111/j.1365-2923.2009.03610.x","title":"Reading means more than deciphering the words on the page","year":2010,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Biostatistics; Citation; Reading (process); Library science; Computer science; Medicine; Philosophy; Epidemiology; Linguistics; Pathology","score_opus":0.01780742874395911,"score_gpt":0.3629917723502269,"score_spread":0.34518434360626776,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2081711362","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8166537,0.000034611847,0.000106582156,0.16976336,0.0016411827,0.00018987022,5.3083386e-7,0.00004250641,0.011567618],"genre_scores_gemma":[0.9752589,0.000043571734,0.00018437378,0.021538623,0.0014054658,0.0000628302,0.0000136754215,0.000014290837,0.0014783066],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9988102,0.000052261414,0.00020416132,0.00017551548,0.0005751522,0.00018275678],"domain_scores_gemma":[0.98976743,0.009345866,0.00006132654,0.0005228562,0.00006372263,0.00023879274],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011244077,0.00009149119,0.00012452423,0.000022022265,0.00014682325,0.000026231466,0.00018901053,0.0001273898,0.0014042584],"category_scores_gemma":[0.16213441,0.000042320855,0.00007263654,0.00016518921,0.00019470403,0.000017659011,0.000029945142,0.00077811675,0.00012474791],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012071158,0.0017116104,0.080161974,0.000026912167,0.00007718437,0.00002315232,0.0066791014,0.0000027757292,0.0006523242,0.026239078,0.22453736,0.6597678],"study_design_scores_gemma":[0.001634761,0.0004131414,0.7364774,0.0056402823,0.00030867895,0.00024454715,0.006923885,0.0044438154,0.0013152044,0.0040435493,0.23813042,0.0004243697],"about_ca_topic_score_codex":0.00013656766,"about_ca_topic_score_gemma":0.000046525933,"teacher_disagreement_score":0.6593434,"about_ca_system_score_codex":0.000026538988,"about_ca_system_score_gemma":0.00060853275,"threshold_uncertainty_score":0.9995086},"labels":[],"label_agreement":null},{"id":"W2082751133","doi":"10.1111/medu.12665","title":"The impact of adopting EHRs: how losing connectivity affects clinical reasoning","year":2015,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":44,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Innovation, Science and Economic Development Canada; Carleton University; Children's Hospital of Eastern Ontario; University of Ottawa","funders":"Centre Hospitalier pour Enfants de l'est de l'Ontario","keywords":"Grounded theory; Chart; Psychology; Workload; Data collection; Health care; Credibility; Medical education; Medicine; Computer science; Qualitative research; Sociology","score_opus":0.05185101892665846,"score_gpt":0.4524582180607421,"score_spread":0.40060719913408366,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2082751133","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9846701,0.0008325687,0.00052161637,0.010610399,0.0011478154,0.00023373657,9.3534675e-7,0.000044980334,0.0019378484],"genre_scores_gemma":[0.9971117,0.00011996283,0.0008910758,0.0005840116,0.00105087,0.000014308244,0.000019499736,0.000015557218,0.00019302366],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9981461,0.00031811302,0.00037832386,0.00024088452,0.00065452413,0.00026201812],"domain_scores_gemma":[0.9728141,0.02477466,0.0003887091,0.00052722794,0.00050269556,0.0009926427],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00444797,0.000121747755,0.00037530647,0.000038955055,0.00008523661,0.00003443753,0.00012340638,0.00021111104,0.000040737967],"category_scores_gemma":[0.58722335,0.000072054856,0.0002290659,0.00023198748,0.00029951747,0.000060462102,0.00004993007,0.00048917986,0.000014716257],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021494215,0.0011463129,0.5077993,0.000018797227,0.00008009413,0.000008318163,0.00028002236,0.000005540511,0.0000100955795,0.00031449363,0.06857357,0.4215485],"study_design_scores_gemma":[0.003600522,0.0021249903,0.9732716,0.0063049607,0.00022061646,0.00018875801,0.0014537028,0.0075251386,0.00012793807,0.00051942247,0.004424091,0.0002382218],"about_ca_topic_score_codex":0.0003124954,"about_ca_topic_score_gemma":0.00001509676,"teacher_disagreement_score":0.58277535,"about_ca_system_score_codex":0.00014463106,"about_ca_system_score_gemma":0.0047365297,"threshold_uncertainty_score":0.84024},"labels":[],"label_agreement":null},{"id":"W2082790809","doi":"10.1007/s11013-014-9408-5","title":"Clinician Reasoning in the Use of Cultural Formulation to Resolve Uncertainty in the Diagnosis of Psychosis","year":2014,"lang":"en","type":"article","venue":"Culture Medicine and Psychiatry","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":36,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Jewish General Hospital","funders":"","keywords":"Psychosis; Psychology; Thematic analysis; Competence (human resources); Clinical psychology; Psychiatry; Psychotherapist; Qualitative research; Social psychology; Sociology","score_opus":0.07395641424119966,"score_gpt":0.3870609372211324,"score_spread":0.31310452297993274,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2082790809","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9563168,0.0012119496,0.000073737574,0.03992278,0.00034087856,0.0005643182,0.000006452752,0.000008150925,0.0015548781],"genre_scores_gemma":[0.9847835,0.0006315964,0.00163273,0.012329361,0.0004957438,0.00004896071,0.00003938149,0.000007671006,0.00003105821],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9983586,0.00022013744,0.0006597444,0.00023322011,0.00034637112,0.00018191036],"domain_scores_gemma":[0.99447125,0.004769689,0.00019286956,0.0003771171,0.000104950195,0.00008414114],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0013609886,0.0001482541,0.00045232338,0.00009522944,0.000044587166,0.000009693045,0.0001503014,0.00013841558,0.00001679834],"category_scores_gemma":[0.01797314,0.00006496947,0.00009281619,0.00059151836,0.00012788894,0.000050454775,0.000015368967,0.00037201692,0.000001279445],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002856815,0.00032682452,0.8803758,0.000076410426,0.000021498963,0.0000018352209,0.020197265,0.00007258133,0.000011858097,0.0040467395,0.084134966,0.010448549],"study_design_scores_gemma":[0.0034103668,0.0018906355,0.936347,0.007974846,0.0002157343,0.000021092075,0.01281659,0.00035973277,0.000006595702,0.001305312,0.03552071,0.00013136257],"about_ca_topic_score_codex":0.0010224553,"about_ca_topic_score_gemma":0.0011105331,"teacher_disagreement_score":0.055971228,"about_ca_system_score_codex":0.000009862269,"about_ca_system_score_gemma":0.000025752104,"threshold_uncertainty_score":0.99029887},"labels":[],"label_agreement":null},{"id":"W2084061419","doi":"10.1016/s0302-2838(02)00053-2","title":"Assessment of Clinical Reasoning Competence in Urology with the Script Concordance Test: an Exploratory Study Across Two Sites from Different Countries","year":2002,"lang":"en","type":"article","venue":"European Urology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":29,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Jewish General Hospital; McGill University; Université de Montréal","funders":"McGill University","keywords":"Concordance; Cronbach's alpha; Medicine; Competence (human resources); Test (biology); Urology; Significant difference; Internal medicine; Clinical psychology; Psychometrics; Psychology; Social psychology","score_opus":0.06177579569147867,"score_gpt":0.3867001813956141,"score_spread":0.3249243857041354,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2084061419","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9955992,0.0005438664,0.00017372041,0.0013643196,0.00023998019,0.0004156495,0.000029882165,0.00006245307,0.0015709109],"genre_scores_gemma":[0.9949452,0.00011661919,0.00027703386,0.004228508,0.0002916084,0.000020499781,0.00002197104,0.0000404912,0.00005804709],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99564844,0.0020350842,0.00092203857,0.0006570435,0.0003090222,0.00042838117],"domain_scores_gemma":[0.97871614,0.01975074,0.0003969394,0.0007943914,0.00014231874,0.00019945584],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0019271037,0.00024918263,0.00088023423,0.00004232002,0.00010063907,0.000023507058,0.0003553602,0.00006983899,0.00016664308],"category_scores_gemma":[0.0112066595,0.0001544871,0.00007629158,0.00014341103,0.0009933464,0.000060322855,0.0001936273,0.00081027555,0.000049808164],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00031453895,0.0022215273,0.9926072,0.0000035263154,0.00006758892,0.00076624274,0.0022129312,0.000075658645,0.000073104435,0.00011784165,0.00028970098,0.0012501357],"study_design_scores_gemma":[0.0057456386,0.008749871,0.9802637,0.00008023161,0.00014066359,0.000032946908,0.0013512657,0.002649249,0.0000068842896,0.000015857388,0.0008183074,0.00014536726],"about_ca_topic_score_codex":0.00008368436,"about_ca_topic_score_gemma":0.0012022783,"teacher_disagreement_score":0.017715657,"about_ca_system_score_codex":0.000019901594,"about_ca_system_score_gemma":0.000058959413,"threshold_uncertainty_score":0.99712235},"labels":[],"label_agreement":null},{"id":"W2086172038","doi":"10.3138/jvme.0212-017r","title":"Development and Preliminary Evaluation of Student-Authored Electronic Cases","year":2012,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Curriculum; Medical education; Quality (philosophy); Process (computing); Medicine; Psychology; Computer science; Pedagogy","score_opus":0.15042779786429225,"score_gpt":0.4872539525156984,"score_spread":0.3368261546514062,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2086172038","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9918879,0.006284228,0.0000369639,0.00088910904,0.00060697703,0.00014298836,2.222313e-7,0.0000037420941,0.00014784528],"genre_scores_gemma":[0.9973563,0.00032131118,0.001620817,0.00020611672,0.0004391537,0.000010020902,0.000007803217,0.000007309211,0.000031189884],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9980001,0.00013468537,0.0005699351,0.00007319263,0.0010578261,0.00016425448],"domain_scores_gemma":[0.99746346,0.0012970308,0.0003283614,0.00009336998,0.00039322875,0.00042452334],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00309129,0.000076255805,0.00023828053,0.000104406,0.000026853904,0.0000043344085,0.000057284407,0.00008921961,0.00024245802],"category_scores_gemma":[0.026563905,0.000057921083,0.00005209358,0.00008564476,0.000057767964,0.00009597298,0.00002867537,0.0002154034,0.0000028186223],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009854231,0.0049224496,0.08768385,0.000104699124,0.00016340327,0.00003296412,0.0032598937,0.0000013587111,0.0006040493,0.00015997789,0.0017587618,0.90032315],"study_design_scores_gemma":[0.0019298844,0.0036869848,0.9790819,0.0023699475,0.00045989643,0.00633915,0.0013328855,0.0000789853,0.0003632073,0.000060553506,0.004216243,0.000080348516],"about_ca_topic_score_codex":0.0000038175153,"about_ca_topic_score_gemma":4.2232512e-7,"teacher_disagreement_score":0.9002428,"about_ca_system_score_codex":0.00014871304,"about_ca_system_score_gemma":0.0030408907,"threshold_uncertainty_score":0.98163575},"labels":[],"label_agreement":null},{"id":"W2086487466","doi":"10.1097/acm.0b013e31826735fc","title":"Renowned Physicians’ Perceptions of Expert Diagnostic Practice","year":2012,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":72,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"SickKids Foundation; University of Toronto; Hospital for Sick Children","funders":"","keywords":"Excellence; Grounded theory; Thematic analysis; Medical education; Constructivist grounded theory; Psychology; Context (archaeology); Possession (linguistics); Perception; Qualitative research; Medicine; Sociology","score_opus":0.04510240496446501,"score_gpt":0.41393838488857954,"score_spread":0.36883597992411454,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2086487466","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8558487,0.023266086,0.0023250866,0.079769515,0.0021087048,0.00087007077,0.000011929854,0.00022091657,0.03557897],"genre_scores_gemma":[0.97525984,0.008067137,0.00077697565,0.011136078,0.0033398443,0.000039951006,0.000053940432,0.00002976997,0.0012964642],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9979802,0.000115323695,0.00067058706,0.00023505054,0.0005662163,0.0004326639],"domain_scores_gemma":[0.96457326,0.03412507,0.0002688688,0.00040271255,0.0002018583,0.00042822232],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010473734,0.00018436243,0.00059474714,0.00010624327,0.000051866988,0.0000012275067,0.00013218641,0.00026135065,0.00086764625],"category_scores_gemma":[0.25848472,0.0001341937,0.00009005741,0.0003401374,0.00039624784,0.0001647958,0.000048316335,0.00071812916,0.00020198166],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000533112,0.00092290086,0.2986559,0.00015292555,0.00036136142,0.000052764157,0.03477317,0.0000037452128,0.02436532,0.0060591022,0.5439273,0.09019237],"study_design_scores_gemma":[0.0119001735,0.0029536674,0.643246,0.02441322,0.0031830254,0.00094090484,0.035600316,0.00010540315,0.0040308684,0.0010327848,0.27176237,0.000831306],"about_ca_topic_score_codex":0.00013580233,"about_ca_topic_score_gemma":4.422833e-7,"teacher_disagreement_score":0.34459007,"about_ca_system_score_codex":0.00005172475,"about_ca_system_score_gemma":0.00006517445,"threshold_uncertainty_score":0.9500118},"labels":[],"label_agreement":null},{"id":"W2086779023","doi":"10.1109/icalt.2012.44","title":"A Realistic Digital Deteriorating Patient to Foster Emergency Decision-Making Skills in Medical Students","year":2012,"lang":"en","type":"article","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Computer science; Medical education; Multimedia; Medical emergency; Medicine","score_opus":0.01932761266577334,"score_gpt":0.3948900385673911,"score_spread":0.37556242590161776,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2086779023","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99058,0.000077278266,0.0021137998,0.0005758885,0.000643248,0.00034271568,0.0000059142926,0.000046699028,0.005614422],"genre_scores_gemma":[0.99415356,0.00001519032,0.0021350668,0.0032681425,0.0002300805,0.000031385618,0.00000972738,0.000021758318,0.00013506215],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.997238,0.000031367414,0.0007582807,0.0002715639,0.001219777,0.0004810424],"domain_scores_gemma":[0.99277294,0.0059731165,0.000076133714,0.00030639657,0.000078886864,0.0007925591],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0005888253,0.00015785481,0.00033835327,0.00011656957,0.00003506801,0.000045308145,0.0001717389,0.00012366094,0.0015983264],"category_scores_gemma":[0.12662771,0.00011574638,0.00009239894,0.00027916176,0.000021531278,0.000116323434,0.00034051033,0.00019182524,0.0003594601],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000062886116,0.0008266434,0.6533227,0.000005436027,0.000011099182,0.00009319523,0.0006571645,0.000001946186,0.0000035984722,0.000015569356,0.0028213363,0.34217843],"study_design_scores_gemma":[0.0016359125,0.0007290516,0.98778236,0.006990707,0.00004723318,0.00012085416,0.000538168,0.00018906705,0.000022666449,0.00035141903,0.0012618962,0.00033064449],"about_ca_topic_score_codex":0.000022044149,"about_ca_topic_score_gemma":0.00004638651,"teacher_disagreement_score":0.3418478,"about_ca_system_score_codex":0.00007899664,"about_ca_system_score_gemma":0.00005760215,"threshold_uncertainty_score":0.99931437},"labels":[],"label_agreement":null},{"id":"W2086878092","doi":"10.1007/s10459-011-9306-3","title":"Critically ill patients and end-of-life decision-making: the senior medical resident experience","year":2011,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":29,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hôpital Maisonneuve-Rosemont; Université de Montréal","funders":"","keywords":"Feeling; Narrative; Distress; Psychology; Qualitative research; Narrative inquiry; Medical education; Medicine; Nursing; Social psychology; Psychotherapist; Sociology","score_opus":0.036224876312886074,"score_gpt":0.4512923448838473,"score_spread":0.4150674685709612,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2086878092","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9835232,0.0064869714,0.0003925939,0.0063027632,0.0010613056,0.00032980004,0.0000011987287,0.0000108378445,0.0018913157],"genre_scores_gemma":[0.97643846,0.0031653869,0.008739271,0.011540977,0.00006386448,0.000037468326,0.0000011302087,0.0000034985014,0.000009922075],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99811983,0.00008538978,0.0005940177,0.00029063498,0.0006720125,0.00023813313],"domain_scores_gemma":[0.9885419,0.010241644,0.00020222654,0.00021680271,0.0004357633,0.00036166832],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0013027461,0.000075521595,0.00019414203,0.00008373069,0.00014104444,0.000009739222,0.00018988673,0.000045533667,0.00019143287],"category_scores_gemma":[0.18131751,0.00004694508,0.000022284657,0.0003707422,0.00096412946,0.00017531111,0.000049396192,0.0001323648,0.0000037012826],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000055311488,0.0003226412,0.67155004,0.000043395474,7.6274887e-7,7.312427e-7,0.0036927313,0.0000017729743,1.8413866e-7,0.0032402922,0.00032128108,0.32077086],"study_design_scores_gemma":[0.00040426617,0.00065370486,0.97725356,0.0037411943,0.0000047390276,0.0000117106865,0.0059357146,0.00022578728,0.000013653651,0.00990593,0.0017641548,0.0000856038],"about_ca_topic_score_codex":0.00043739766,"about_ca_topic_score_gemma":0.00013455012,"teacher_disagreement_score":0.32068527,"about_ca_system_score_codex":0.000041614203,"about_ca_system_score_gemma":0.0017225108,"threshold_uncertainty_score":0.8255786},"labels":[],"label_agreement":null},{"id":"W2087170012","doi":"10.1007/s10459-008-9149-8","title":"Qualitative differences in knowledge structure are associated with diagnostic performance in medical students","year":2008,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":27,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Foothills Medical Centre; University of Calgary","funders":"","keywords":"Medical knowledge; Computer science; Knowledge management; Medicine; Medical education","score_opus":0.04658494499910382,"score_gpt":0.4827754880675091,"score_spread":0.43619054306840527,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2087170012","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9904124,0.0063300948,0.0000061341293,0.0020155008,0.00044096497,0.0004434039,0.0000031170048,0.000014905036,0.0003334943],"genre_scores_gemma":[0.9908637,0.007927977,0.0002525231,0.0007290503,0.000060922957,0.00008344463,0.000016741285,0.0000056765666,0.00006000649],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99736434,0.0003267631,0.0006115136,0.00040905402,0.00089122367,0.00039708088],"domain_scores_gemma":[0.984166,0.015061534,0.00029761766,0.00013158424,0.000099293626,0.000243946],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0018543764,0.00013693082,0.0004244134,0.00029680075,0.00012242202,0.000012526209,0.00026912725,0.00008957393,0.00003909305],"category_scores_gemma":[0.049480747,0.00009845557,0.000016049327,0.001533189,0.00054893934,0.00026407244,0.00003195174,0.00038855852,0.0000043986065],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000028266088,0.0009048671,0.97776735,0.000067467554,0.000001480055,0.000008622103,0.011764529,0.00008257644,5.981916e-8,0.00006794329,0.00002781558,0.009279041],"study_design_scores_gemma":[0.0010309088,0.0006176987,0.97949094,0.01109524,0.0000021878127,0.000014731678,0.006756724,0.00046150558,0.0000021969108,0.00037678785,0.000045010882,0.00010605167],"about_ca_topic_score_codex":0.00050901395,"about_ca_topic_score_gemma":0.0071833734,"teacher_disagreement_score":0.047626372,"about_ca_system_score_codex":0.00033917496,"about_ca_system_score_gemma":0.0025095765,"threshold_uncertainty_score":0.9585259},"labels":[],"label_agreement":null},{"id":"W2087850911","doi":"10.1007/s11606-007-0330-1","title":"Tips for Teachers of Evidence-based Medicine: Making Sense of Diagnostic Test Results Using Likelihood Ratios","year":2007,"lang":"en","type":"article","venue":"Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Columbia College","funders":"","keywords":"Odds; Test (biology); Diagnostic odds ratio; Odds ratio; Medicine; Diagnostic test; Clinical Practice; Disease; Medical education; Cognitive psychology; Psychology; Family medicine; Meta-analysis; Pathology; Pediatrics","score_opus":0.15611129767273688,"score_gpt":0.44121468680699966,"score_spread":0.2851033891342628,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2087850911","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9347114,0.0043415884,0.04696659,0.01113273,0.0017262311,0.00048713153,0.000025950781,0.000014146644,0.00059424114],"genre_scores_gemma":[0.9809265,0.0004246816,0.012774843,0.001476985,0.004051468,0.0000017537909,0.000008855306,0.000035084347,0.00029982737],"study_design_codex":"bench_or_experimental","study_design_gemma":"systematic_review","domain_scores_codex":[0.9952274,0.00010746105,0.0029742497,0.00024134813,0.0010738352,0.00037571808],"domain_scores_gemma":[0.90647864,0.0883188,0.0026381256,0.00033770673,0.0017610586,0.00046569205],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005924711,0.0002660244,0.001293,0.00050896464,0.000039185805,0.0000051153047,0.00020642125,0.00016960976,0.000081593884],"category_scores_gemma":[0.4688669,0.00015770488,0.00032524538,0.00033825892,0.0004908435,0.000102008125,0.000032424436,0.000513113,9.1344947e-7],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.031342566,0.0023243304,0.27383608,0.0011671898,0.0012763939,0.0031238203,0.0045426153,0.0012214229,0.45825875,0.00013374703,0.09363302,0.12914006],"study_design_scores_gemma":[0.06690664,0.06838434,0.23630285,0.43463248,0.0069617033,0.0066482555,0.0030275974,0.014194416,0.15591133,0.0009755563,0.005203984,0.0008508583],"about_ca_topic_score_codex":0.0003620857,"about_ca_topic_score_gemma":0.000025965766,"teacher_disagreement_score":0.4629422,"about_ca_system_score_codex":0.00016993818,"about_ca_system_score_gemma":0.00038632235,"threshold_uncertainty_score":0.6431019},"labels":[],"label_agreement":null},{"id":"W2088720687","doi":"10.1097/acm.0000000000000432","title":"In Reply to Norman et al and to Ilgen et al","year":2014,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Victoria General Hospital; Nova Scotia Health Authority; Capital District Health Authority","funders":"","keywords":"Certainty; Respondent; Consistency (knowledge bases); Psychology; Point (geometry); Computer science; Cognitive psychology; Epistemology; Artificial intelligence; Philosophy; Mathematics","score_opus":0.045871171497257994,"score_gpt":0.4069495223790373,"score_spread":0.3610783508817793,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2088720687","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.018920278,0.0006308011,0.00013823555,0.9746344,0.00084461214,0.00079360674,0.000014345696,0.00007132611,0.003952389],"genre_scores_gemma":[0.0027744544,0.0012126006,0.0002598635,0.98578036,0.0033700073,0.00007963725,0.00028833916,0.00008510102,0.0061496655],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9962152,0.00021474434,0.0010396803,0.0009669119,0.00091490045,0.00064851873],"domain_scores_gemma":[0.9866837,0.011610743,0.0001919456,0.0006850067,0.000094324445,0.00073423743],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0029276807,0.0004524567,0.0013486098,0.00045826996,0.000026791706,0.0000074918485,0.00029197204,0.00082614395,0.00033917985],"category_scores_gemma":[0.075508654,0.00033983766,0.00006939441,0.00034470798,0.00015952764,0.000043663906,0.00020162316,0.0056910724,0.00038025086],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011715362,0.00000867734,0.003563395,0.000045430465,0.000035189343,0.0006475317,0.0011254707,0.0000010318889,0.00008435814,0.000045707886,0.99315387,0.0011721699],"study_design_scores_gemma":[0.0023578028,0.0010937339,0.0038165932,0.009641975,0.00013676123,0.000363053,0.00004829523,0.000014103754,0.000014306184,0.0001870718,0.9820062,0.00032014004],"about_ca_topic_score_codex":0.000894537,"about_ca_topic_score_gemma":0.000017032296,"teacher_disagreement_score":0.07258098,"about_ca_system_score_codex":0.000093506234,"about_ca_system_score_gemma":0.00017090335,"threshold_uncertainty_score":0.99990535},"labels":[],"label_agreement":null},{"id":"W2089837266","doi":"10.1023/b:ahse.0000027436.17220.9c","title":"In Pursuit of Expertise. Toward an Educational Model for Expertise Development","year":2004,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":64,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Atlantic School of Theology; Nova Scotia Health Authority","funders":"","keywords":"Metacognition; Flexibility (engineering); Relation (database); Nurse education; Psychology; Computer science; Process management; Medical education; Mathematics education; Knowledge management; Medicine; Cognition; Engineering; Management","score_opus":0.06976774425162986,"score_gpt":0.4809545600942789,"score_spread":0.41118681584264904,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2089837266","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93856186,0.01169016,0.009389301,0.03764241,0.0011780896,0.0011003764,0.000003143434,0.000014592287,0.00042005145],"genre_scores_gemma":[0.8945949,0.0009655113,0.101490736,0.0023456735,0.00009428547,0.0003853977,0.000020555319,0.0000052693886,0.0000976547],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9984972,0.000020821339,0.0006036095,0.00031645517,0.00030183507,0.00026008394],"domain_scores_gemma":[0.9987409,0.0007001906,0.00016614322,0.00012557497,0.00009295451,0.00017425076],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00084002584,0.00008305307,0.00020998972,0.00019884446,0.00006810981,0.000009645847,0.00014840341,0.000039521598,0.000015823613],"category_scores_gemma":[0.0046482794,0.00007394313,0.000023490422,0.00036318053,0.00014691378,0.00038649893,0.00001244884,0.00006375236,0.0000020163802],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00028205855,0.007571493,0.112447634,0.0007039867,0.000004041826,6.085558e-7,0.054399014,0.06677913,0.00015611431,0.032315873,0.0006734825,0.7246666],"study_design_scores_gemma":[0.01147654,0.0052026333,0.5351791,0.02493438,0.000021941767,0.0000616354,0.016722158,0.090142205,0.0064596892,0.29889318,0.009573486,0.001333048],"about_ca_topic_score_codex":0.00037081444,"about_ca_topic_score_gemma":0.00041569787,"teacher_disagreement_score":0.72333354,"about_ca_system_score_codex":0.0004636052,"about_ca_system_score_gemma":0.0069069057,"threshold_uncertainty_score":0.99872303},"labels":[],"label_agreement":null},{"id":"W2092264731","doi":"10.1097/00001888-200510001-00007","title":"Does the Medical College Admission Test Predict Clinical Reasoning Skills? A Longitudinal Study Employing the Medical Council of Canada Clinical Reasoning Examination","year":2005,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":22,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Calgary","funders":"","keywords":"Verbal reasoning; Predictive validity; Test (biology); Entrance exam; Stepwise regression; Variance (accounting); Clinical judgment; Medical school; Psychology; Educational measurement; Medicine; Clinical psychology; Medical education; Internal medicine; Cognition; Curriculum; Psychiatry; Pedagogy; Medical physics","score_opus":0.07387066750429407,"score_gpt":0.4058286142227868,"score_spread":0.33195794671849277,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2092264731","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9171663,0.0015825795,0.00024182355,0.07372964,0.0024941203,0.0015785518,0.000025838539,0.00014199816,0.0030391852],"genre_scores_gemma":[0.9768582,0.0032076307,0.00016637752,0.0075902417,0.009011328,0.00009188978,0.000035245444,0.000073479794,0.0029656352],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9813768,0.0019471162,0.0048485654,0.001228865,0.009573387,0.0010252235],"domain_scores_gemma":[0.8394392,0.15270223,0.0016992637,0.0017493122,0.0014485004,0.0029614803],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.050301444,0.00059023337,0.0019470671,0.00009973162,0.00056247244,0.000014469241,0.001376372,0.0012073371,0.0022685835],"category_scores_gemma":[0.71086,0.0002475649,0.0003001183,0.00086719444,0.0019901558,0.00012762361,0.00050594617,0.005793068,0.000019690806],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000357396,0.0008854599,0.8654412,0.000045076737,0.00032228406,0.0006557263,0.0014156963,0.0000043034515,0.0000029719654,0.00019879501,0.08323671,0.047434356],"study_design_scores_gemma":[0.011393002,0.0026840956,0.9146271,0.012729642,0.0012573308,0.0009384121,0.0057261153,0.0073236334,0.000014358535,0.000060075945,0.042867534,0.00037870803],"about_ca_topic_score_codex":0.018013239,"about_ca_topic_score_gemma":0.032916076,"teacher_disagreement_score":0.6605586,"about_ca_system_score_codex":0.0007411227,"about_ca_system_score_gemma":0.014749644,"threshold_uncertainty_score":0.9999977},"labels":[],"label_agreement":null},{"id":"W2093350345","doi":"10.1515/cam.2006.008","title":"What healthcare students do with what they don't know: The socializing power of 'uncertainty' in the case presentation","year":2006,"lang":"en","type":"article","venue":"Communication & Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Waterloo","funders":"","keywords":"Rhetoric; Competence (human resources); Criticism; Health care; Presentation (obstetrics); Psychology; Identity (music); Empathy; Social psychology; Medicine; Political science; Law; Aesthetics","score_opus":0.043053316205234456,"score_gpt":0.41392380119897254,"score_spread":0.3708704849937381,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2093350345","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.82046926,0.024785517,0.000019377148,0.1528777,0.00016855272,0.0008949514,0.000001679157,0.000021113518,0.0007618663],"genre_scores_gemma":[0.9869522,0.009782448,0.00004996488,0.0027734747,0.0001217811,0.000080893136,0.00010134622,0.000014772089,0.0001231601],"study_design_codex":"observational","study_design_gemma":"qualitative","domain_scores_codex":[0.9977622,0.00073017646,0.00056178827,0.0001772137,0.00058483734,0.0001838107],"domain_scores_gemma":[0.98746246,0.0106740175,0.0003009607,0.0012042553,0.00029685954,0.00006144861],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002228385,0.0001361006,0.00033056675,0.00007141395,0.00019105761,0.000081534105,0.00046788523,0.00008029903,0.000049852748],"category_scores_gemma":[0.0036202196,0.000065034954,0.000051895582,0.0003313214,0.00041848273,0.00021378169,0.00007236248,0.0004054119,0.0000037879577],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019699417,0.0043627787,0.43965632,0.00036144888,0.00037870422,0.00086221594,0.29477942,0.0008907635,0.00016463158,0.024429386,0.03149765,0.20064676],"study_design_scores_gemma":[0.011220471,0.001377035,0.36023423,0.027261756,0.00048649946,0.00090197515,0.58524865,0.0003200345,0.00003938746,0.0055314833,0.0070879282,0.0002905322],"about_ca_topic_score_codex":0.007838351,"about_ca_topic_score_gemma":0.005925699,"teacher_disagreement_score":0.29046923,"about_ca_system_score_codex":0.000070995804,"about_ca_system_score_gemma":0.000093057315,"threshold_uncertainty_score":0.99876857},"labels":[],"label_agreement":null},{"id":"W2093549930","doi":"10.1007/s10459-014-9516-6","title":"Predictable chaos: a review of the effects of emotions on attention, memory and decision making","year":2014,"lang":"en","type":"review","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":268,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; McMaster University; University of Toronto; University Health Network","funders":"","keywords":"Recall; Cognition; Psychology; Interpersonal communication; Field (mathematics); Psychological intervention; Cognitive psychology; Applied psychology; Social psychology","score_opus":0.02960308155730451,"score_gpt":0.4784855391185926,"score_spread":0.4488824575612881,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2093549930","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00032497226,0.99568397,0.000068366855,0.00036968858,0.0015696853,0.0015444371,0.000003177862,0.000005755311,0.00042992347],"genre_scores_gemma":[0.00068322214,0.99699634,0.0013181104,0.00075485057,0.00009791656,0.00010484413,0.000008161021,0.000007047549,0.000029513714],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9978377,0.00023461693,0.0010290785,0.0003300282,0.00040308255,0.00016547457],"domain_scores_gemma":[0.984145,0.0140597485,0.0012304517,0.0003703121,0.000111568355,0.00008291162],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002034878,0.00013588333,0.0010415377,0.00017716212,0.000099515455,0.000004805537,0.00017707398,0.00007062677,0.0000074598383],"category_scores_gemma":[0.039045773,0.000079304846,0.0001374332,0.0009005632,0.0002904568,0.00007060482,0.000039961124,0.00019450045,0.0000016879794],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000001549014,0.000115994095,0.00067854207,0.16457556,0.0000019844504,4.2289674e-8,0.000017672633,0.0000014857817,1.1294632e-8,0.00027136016,0.00023171293,0.83410406],"study_design_scores_gemma":[0.000106807805,0.00038340726,0.003386806,0.91936225,0.00010008524,0.000012398961,0.000017799546,0.000010175043,1.4360911e-7,0.00068389747,0.0758822,0.000054052412],"about_ca_topic_score_codex":0.000029751169,"about_ca_topic_score_gemma":0.000008749784,"teacher_disagreement_score":0.83405006,"about_ca_system_score_codex":0.00009023981,"about_ca_system_score_gemma":0.0015977579,"threshold_uncertainty_score":0.96904874},"labels":[],"label_agreement":null},{"id":"W2093844513","doi":"10.1080/0142159021000012599","title":"Stability of clinical reasoning assessment results with the Script Concordance test across two different linguistic, cultural and learning environments","year":2002,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":45,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McGill University; Jewish General Hospital; Université de Montréal","funders":"","keywords":"Concordance; Cronbach's alpha; Test (biology); Construct validity; Construct (python library); Reliability (semiconductor); Psychology; Variance (accounting); Medicine; Clinical psychology; Psychometrics; Computer science","score_opus":0.07608468968488996,"score_gpt":0.41621979171105977,"score_spread":0.3401351020261698,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2093844513","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99036485,0.00045621113,0.00022962417,0.0031131625,0.00015333158,0.00026546698,0.000008086446,0.000039459996,0.00536979],"genre_scores_gemma":[0.99570304,0.0003875418,0.0004906743,0.0003629828,0.00045186307,0.000019272602,0.000020029283,0.000019963329,0.0025446366],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9967325,0.00047400256,0.0008808606,0.00051499094,0.0010163481,0.00038130305],"domain_scores_gemma":[0.9862903,0.012178859,0.00035634573,0.00047074308,0.00007827554,0.00062549225],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0031791136,0.00021790148,0.0006891403,0.00000963451,0.00016697343,0.000026049842,0.00015499684,0.00017117351,0.00060814916],"category_scores_gemma":[0.19270842,0.00011020338,0.00011826371,0.00008304094,0.0013125117,0.000026976979,0.00014891385,0.0017136483,0.000013516538],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021271482,0.0011533973,0.9655356,0.000018513381,0.00007453831,0.00006711221,0.0012762686,0.0000014312074,0.000021507098,0.00002266802,0.0012646978,0.030351568],"study_design_scores_gemma":[0.010026916,0.0017580573,0.9539303,0.0016525009,0.00019702775,0.00005419246,0.0011950234,0.0185436,0.00004980129,0.000009259589,0.012366294,0.00021702792],"about_ca_topic_score_codex":0.00011412403,"about_ca_topic_score_gemma":0.0000311306,"teacher_disagreement_score":0.1895293,"about_ca_system_score_codex":0.000052419375,"about_ca_system_score_gemma":0.00005344329,"threshold_uncertainty_score":0.81409174},"labels":[],"label_agreement":null},{"id":"W2094776956","doi":"10.1016/j.accpm.2014.11.001","title":"The Script Concordance Test in anesthesiology: Validation of a new tool for assessing clinical reasoning","year":2015,"lang":"en","type":"article","venue":"Anaesthesia Critical Care & Pain Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":28,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; University of Toronto","funders":"","keywords":"Anesthesiology; Concordance; Medicine; Test (biology); Observational study; Cronbach's alpha; Pain medicine; Family medicine; Internal medicine; Anesthesia; Psychometrics; Clinical psychology","score_opus":0.0793877531549422,"score_gpt":0.4282176168772304,"score_spread":0.3488298637222882,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2094776956","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9388115,0.007333724,0.005846211,0.046170216,0.00032131176,0.0008899941,0.00000216651,0.000083251696,0.00054163445],"genre_scores_gemma":[0.9837718,0.0002080688,0.0072267223,0.0077674645,0.00081452046,0.000051540454,0.00004823723,0.000034743607,0.00007688464],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99660456,0.00067258993,0.0012781104,0.00046114286,0.0004968075,0.00048678243],"domain_scores_gemma":[0.89970195,0.09723222,0.00027900407,0.00077844126,0.001212617,0.0007957729],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.008900617,0.00022713812,0.00090499024,0.00008427827,0.000100338846,0.000026144256,0.00020736318,0.00026381804,0.000013154534],"category_scores_gemma":[0.59187126,0.0001483406,0.00016902339,0.00033490176,0.00097632105,0.00008485522,0.000025105328,0.00043301,0.0000053201547],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00039010472,0.00019019567,0.8161466,0.000099385914,0.000012807241,0.0001297521,0.00055590447,0.0000036029107,0.000029438039,0.0033552798,0.0125852395,0.1665017],"study_design_scores_gemma":[0.017631477,0.021300528,0.85524255,0.018451829,0.0009537909,0.00019477967,0.019884404,0.0027140102,0.00050872605,0.01219712,0.0501562,0.0007645959],"about_ca_topic_score_codex":0.00029781784,"about_ca_topic_score_gemma":0.00003119024,"teacher_disagreement_score":0.5829706,"about_ca_system_score_codex":0.000104925566,"about_ca_system_score_gemma":0.0008848309,"threshold_uncertainty_score":0.6049154},"labels":[],"label_agreement":null},{"id":"W2094827510","doi":"10.1016/j.amjsurg.2006.10.012","title":"Is the script concordance test a valid instrument for assessment of intraoperative decision-making skills?","year":2007,"lang":"en","type":"article","venue":"The American Journal of Surgery","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":73,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Royal Victoria Hospital; McGill University","funders":"","keywords":"Concordance; Test (biology); Cronbach's alpha; Certification; Medicine; Gold standard (test); Medical physics; Face validity; Reliability (semiconductor); Multiple choice; Physical therapy; Medical education; Psychology; Clinical psychology; Psychometrics; Significant difference; Radiology; Internal medicine","score_opus":0.05061703554908804,"score_gpt":0.40552725139746276,"score_spread":0.35491021584837473,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2094827510","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9786779,0.0002542608,0.015235678,0.00498328,0.0004573797,0.00018785223,0.00001205398,0.00000491708,0.00018665058],"genre_scores_gemma":[0.98621863,0.00039218817,0.008586715,0.004421942,0.0003352189,0.0000045043785,7.079014e-7,0.000015527576,0.000024541276],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9980184,0.00008721653,0.0010016438,0.0001212259,0.0005177546,0.00025376046],"domain_scores_gemma":[0.8635587,0.13399231,0.0014207709,0.0003158662,0.00056700537,0.00014530115],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0044426653,0.00013058037,0.0007335972,0.000102498874,0.00010022028,0.000020613117,0.00018325415,0.000022706163,0.000043519645],"category_scores_gemma":[0.04050735,0.00006284109,0.0004002203,0.00036746621,0.0005577931,0.00005332469,0.000036906436,0.00030565378,0.0000014770168],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008513578,0.0005853092,0.4486672,0.000011461771,0.00027871068,0.00006706883,0.0007344904,0.000034631328,0.00047763067,0.00010605518,0.0068748537,0.5413112],"study_design_scores_gemma":[0.00094280136,0.0015494145,0.98337823,0.005405225,0.00031200857,0.0005039819,0.0034490312,0.00021119688,0.0018130002,0.00096857647,0.0013269531,0.00013958864],"about_ca_topic_score_codex":0.000022760689,"about_ca_topic_score_gemma":0.0000016363634,"teacher_disagreement_score":0.5411716,"about_ca_system_score_codex":0.00008867809,"about_ca_system_score_gemma":0.0004806544,"threshold_uncertainty_score":0.96757483},"labels":[],"label_agreement":null},{"id":"W2095996801","doi":"10.1136/ebm.10.3.66-a","title":"Does clinical experience make up for failure to keep up to date?","year":2005,"lang":"en","type":"editorial","venue":"Evidence-Based Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Computer science; Psychology","score_opus":0.08451577484482581,"score_gpt":0.46034501818996654,"score_spread":0.37582924334514073,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2095996801","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0043682936,0.0016457083,0.0037267436,0.12281383,0.862999,0.0037622964,0.0002458948,0.00031113124,0.000127078],"genre_scores_gemma":[0.004886279,0.0014959327,0.006549252,0.021168455,0.94149476,0.0012713663,0.00093621627,0.0001948827,0.022002837],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99027693,0.00027653607,0.0027873437,0.0023920715,0.0029839436,0.001283154],"domain_scores_gemma":[0.87328154,0.119400956,0.0005934904,0.0021588807,0.0016975924,0.0028675483],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004696036,0.0010285908,0.003135684,0.00041098046,0.00019009481,0.00007224253,0.0010780295,0.0015815187,0.0011890976],"category_scores_gemma":[0.67847437,0.00060520024,0.00069881766,0.0007913526,0.00060780597,0.00011758272,0.00023814455,0.0019619022,0.0005339741],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0040754764,0.00019823994,0.0010097655,0.0003366328,0.00010172687,0.00013254897,0.0006311005,0.0000130812505,0.000050997885,0.000010542909,0.94250244,0.050937433],"study_design_scores_gemma":[0.00689577,0.00653924,0.00060747674,0.062609576,0.00092070573,0.0000064190685,0.0004835006,0.000054398384,0.00021522303,0.00003153207,0.9209149,0.00072124787],"about_ca_topic_score_codex":0.00026978814,"about_ca_topic_score_gemma":0.0002827011,"teacher_disagreement_score":0.67377836,"about_ca_system_score_codex":0.00031530362,"about_ca_system_score_gemma":0.0027031344,"threshold_uncertainty_score":0.999724},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2096268278","doi":"10.1186/1748-5908-2-18","title":"Do physician outcome judgments and judgment biases contribute to inappropriate use of treatments? Study protocol","year":2007,"lang":"en","type":"article","venue":"Implementation Science","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Ottawa Hospital; University of Ottawa","funders":"","keywords":"Medicine; Health services research; Health informatics; Protocol (science); Health administration; Outcome (game theory); Public health; Quality of Life Research; Family medicine; Alternative medicine; Nursing; Pathology","score_opus":0.22007442544235273,"score_gpt":0.5435438645489481,"score_spread":0.32346943910659537,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2096268278","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9526487,0.000001328646,0.00037875332,0.0002174086,0.0000620177,0.046577513,0.00002705161,0.000018621296,0.00006863319],"genre_scores_gemma":[0.99093926,7.89905e-7,0.0015471312,0.00085006736,0.00002502648,0.0065660793,0.000007254691,0.000006594487,0.000057812515],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99794394,0.000028850733,0.0006112628,0.000369657,0.00073302013,0.00031327302],"domain_scores_gemma":[0.99707186,0.0019336204,0.00020625142,0.00025798232,0.00023981948,0.0002904613],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0014236333,0.000119916906,0.0002589405,0.00018759181,0.00011877298,0.00006998117,0.00008642934,0.000016429212,0.000040659055],"category_scores_gemma":[0.007163013,0.00009020489,0.00003476913,0.0006024279,0.00013323026,0.00020221472,0.00010018745,0.00004600221,0.0000100046245],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00024735852,0.0009125727,0.93082076,0.0000075582184,0.000029089708,0.000012278084,0.0007173688,0.0000038410326,0.0035942905,0.00006214956,0.0001009261,0.063491784],"study_design_scores_gemma":[0.0061589866,0.0021887368,0.9723875,0.00014107567,0.00007707074,0.0000023610319,0.001875353,0.0000107201995,0.016734643,0.00002466379,0.00031039878,0.00008847704],"about_ca_topic_score_codex":0.00032588953,"about_ca_topic_score_gemma":0.000046645993,"teacher_disagreement_score":0.06340331,"about_ca_system_score_codex":0.00011784173,"about_ca_system_score_gemma":0.000103879865,"threshold_uncertainty_score":0.85753125},"labels":[],"label_agreement":null},{"id":"W2096673765","doi":"10.1186/1472-6920-13-166","title":"Construction and utilization of a script concordance test as an assessment tool for dcem3 (5th year) medical students in rheumatology","year":2013,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":24,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Concordance; Cronbach's alpha; Test (biology); Medicine; Rheumatology; Internal medicine; Context (archaeology); Medical education; Family medicine; Consistency (knowledge bases); Clinical psychology; Artificial intelligence; Computer science; Psychometrics","score_opus":0.030486382302422527,"score_gpt":0.4513157390124501,"score_spread":0.4208293567100276,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2096673765","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9918815,0.000105358224,0.004346279,0.0021170867,0.0006087043,0.0007346158,0.0000023916218,0.000020689173,0.00018337597],"genre_scores_gemma":[0.98833215,0.00031116355,0.009855191,0.0009487272,0.00009145062,0.00024192249,0.00012084952,0.0000110204655,0.00008755026],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9981411,0.00010082677,0.00056637055,0.0002655929,0.0007675105,0.00015860562],"domain_scores_gemma":[0.992738,0.006144489,0.00017797503,0.00020374762,0.00026785673,0.00046791494],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.000951308,0.00009507821,0.0003252119,0.000085060215,0.000026886935,0.000018059076,0.0001109524,0.00023261122,0.0009659536],"category_scores_gemma":[0.14503244,0.000080879145,0.00003205052,0.00013396527,0.0002532979,0.000082380815,0.000034087974,0.00016888075,0.000012374633],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000046907593,0.0013462229,0.9237244,0.00016055402,0.000008093276,7.287794e-7,0.00012417618,1.2043354e-7,0.00001503162,0.0036612565,0.0017193296,0.06919317],"study_design_scores_gemma":[0.003668134,0.00065982086,0.98508996,0.0027515474,0.00003945351,0.00011796654,0.0009432726,0.00358514,0.00004470263,0.002790951,0.0002119641,0.00009709661],"about_ca_topic_score_codex":0.00057706534,"about_ca_topic_score_gemma":0.000117469834,"teacher_disagreement_score":0.14408112,"about_ca_system_score_codex":0.000057171746,"about_ca_system_score_gemma":0.0035025717,"threshold_uncertainty_score":0.9999473},"labels":[],"label_agreement":null},{"id":"W2097442776","doi":"10.3109/0142159x.2013.774336","title":"Effect of stress on clinical reasoning during simulated ambulatory consultations","year":2013,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":71,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Stress (linguistics); Ambulatory; Psychology; Medicine; Medical education; Clinical psychology; Internal medicine","score_opus":0.021605517415035996,"score_gpt":0.38840991688748505,"score_spread":0.366804399472449,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2097442776","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99353236,0.00011585733,0.000017449429,0.00081208965,0.00024205008,0.00042714138,0.0000024526125,0.00011471895,0.004735896],"genre_scores_gemma":[0.99721205,0.000039773677,0.000059879687,0.0002825185,0.0003431145,0.000020957481,0.00002795887,0.000027314956,0.0019864151],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9974981,0.00041868977,0.00074688735,0.0003287444,0.0007201562,0.00028744133],"domain_scores_gemma":[0.9833776,0.015080254,0.00016748332,0.00042924148,0.00013923268,0.0008062176],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001385496,0.00017134991,0.0006200046,0.00006520183,0.000051480336,0.000009532447,0.00011638738,0.00038068564,0.005467044],"category_scores_gemma":[0.21133295,0.000118361335,0.0002201857,0.00016368703,0.00034068633,0.000030061212,0.00005028817,0.00083291106,0.00043826093],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003431107,0.00097592274,0.96475726,0.00008524562,0.0002175096,0.00009990682,0.00010826445,0.0000377903,0.00006324964,0.00006246159,0.0030057265,0.030243544],"study_design_scores_gemma":[0.012606351,0.0033686229,0.9598042,0.0070661292,0.0004037646,0.00003174504,0.00007183585,0.014816165,0.0012601594,0.000036517613,0.00026463988,0.00026990566],"about_ca_topic_score_codex":0.00012856063,"about_ca_topic_score_gemma":0.00000136688,"teacher_disagreement_score":0.20994745,"about_ca_system_score_codex":0.000032847467,"about_ca_system_score_gemma":0.00012543224,"threshold_uncertainty_score":0.9954421},"labels":[],"label_agreement":null},{"id":"W2098566000","doi":"","title":"What do GPs need to know? The use of knowledge in general practice consultations.","year":2000,"lang":"en","type":"article","venue":"PubMed","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Scarborough Hospital","funders":"","keywords":"General practice; Primary care; Medicine; Global Positioning System; Need to know; Medical education; Knowledge management; Nursing; Data science; Computer science; Family medicine","score_opus":0.056310657923380025,"score_gpt":0.34156169299748695,"score_spread":0.28525103507410693,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2098566000","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97142816,0.001595443,0.000009704681,0.019931788,0.00030745502,0.0010364054,0.0000060420057,0.000025729625,0.0056593055],"genre_scores_gemma":[0.9805278,0.0021318665,0.0009157004,0.004485962,0.00018942279,0.00068998366,0.000012691136,0.000015652487,0.011030926],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99897087,0.00012645427,0.00033313807,0.00017373993,0.0001586476,0.0002371768],"domain_scores_gemma":[0.9818652,0.017369896,0.00006264697,0.00030844542,0.00018829596,0.00020545922],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00055994536,0.00008353079,0.00021215914,0.00006665018,0.000026898972,0.000053430747,0.00007384905,0.00006329497,0.0001513225],"category_scores_gemma":[0.07470894,0.000056939793,0.000053912452,0.00044936815,0.0000757028,0.00018671756,0.00002178179,0.00015785184,0.00011443404],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004517233,0.0006640375,0.00890905,0.000008270458,0.00003810966,0.000021373295,0.0013233188,0.00021345101,0.000008949738,0.00031490583,0.03715343,0.9508934],"study_design_scores_gemma":[0.0022640896,0.00007689707,0.5922672,0.00045257123,0.00013485117,0.000031704556,0.0005940526,0.000200598,0.00015830553,0.000042457556,0.40366244,0.00011481263],"about_ca_topic_score_codex":0.0001774249,"about_ca_topic_score_gemma":0.000057571717,"teacher_disagreement_score":0.95077854,"about_ca_system_score_codex":0.000038048453,"about_ca_system_score_gemma":0.000076576376,"threshold_uncertainty_score":0.9330852},"labels":[],"label_agreement":null},{"id":"W2098712941","doi":"10.1111/j.1365-2929.2005.02258.x","title":"Heuristics and biases − a biased perspective on clinical reasoning","year":2005,"lang":"en","type":"letter","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":103,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Heuristics; Citation; Perspective (graphical); Library science; Computer science; Artificial intelligence","score_opus":0.05384115042139082,"score_gpt":0.44556151165089736,"score_spread":0.39172036122950654,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2098712941","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.016875243,0.001451746,0.000034338547,0.9700969,0.0022189338,0.00043137622,0.000020485799,0.00012013162,0.008750885],"genre_scores_gemma":[0.008216823,0.0017272392,0.0010979666,0.9557577,0.027320001,0.000059830683,0.00077830144,0.0000695082,0.0049726437],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9966187,0.00027260673,0.0007983158,0.0008026715,0.001128151,0.00037955036],"domain_scores_gemma":[0.96690047,0.031042682,0.00033697454,0.0006398174,0.00042790343,0.00065215956],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0009659294,0.000337041,0.0008112113,0.00019316791,0.00008519623,0.00004432606,0.00014378474,0.001608151,0.0010619821],"category_scores_gemma":[0.5318599,0.00027296736,0.00023509022,0.00018340892,0.00044848266,0.000024073095,0.00005059955,0.0036562234,0.00021669618],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007919983,0.00089207274,0.0028861286,0.000038321043,0.000072609495,0.0002308663,0.00009522757,1.3493819e-7,5.8900937e-8,0.00015032422,0.9260147,0.069540314],"study_design_scores_gemma":[0.0017787192,0.0012509762,0.016836777,0.0153634045,0.00059640856,0.00029486112,0.00020881352,0.00032438006,0.0000019067693,0.00026977697,0.9626889,0.00038507828],"about_ca_topic_score_codex":0.00024835448,"about_ca_topic_score_gemma":0.000009472104,"teacher_disagreement_score":0.530894,"about_ca_system_score_codex":0.00026915787,"about_ca_system_score_gemma":0.0040574884,"threshold_uncertainty_score":0.9999722},"labels":[],"label_agreement":null},{"id":"W2100328560","doi":"10.1136/emermed-2013-202421","title":"How do emergency physicians make discharge decisions?","year":2013,"lang":"en","type":"article","venue":"Emergency Medicine Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":113,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"Society for Academic Emergency Medicine","keywords":"Medicine; Medical emergency; Intensive care medicine; Emergency medicine","score_opus":0.04732266687148116,"score_gpt":0.36156449172046246,"score_spread":0.3142418248489813,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2100328560","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8084459,0.01078427,0.005316473,0.095440246,0.017415047,0.00081398094,0.00001091535,0.00015197023,0.06162125],"genre_scores_gemma":[0.7220059,0.02169095,0.00068423053,0.0003776786,0.011623745,0.00005904206,0.00004136806,0.00010208198,0.24341503],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9962026,0.000117987736,0.0012932982,0.00044079262,0.0012024317,0.0007429183],"domain_scores_gemma":[0.99593365,0.0010931499,0.000510762,0.000625637,0.00077793124,0.0010588639],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0007709443,0.00038322614,0.00079995126,0.00025384853,0.00030901004,0.000028882863,0.00031641574,0.00016844938,0.1420606],"category_scores_gemma":[0.039709013,0.00024990056,0.0004628953,0.00063807284,0.00013280596,0.00022794816,0.00007148077,0.0010057031,0.00065709575],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000037676436,0.00038137415,0.057161815,0.000013991535,0.0001419012,0.000015250057,0.00042413958,0.0000015251628,0.0015260286,0.0001972276,0.89980686,0.04029219],"study_design_scores_gemma":[0.005302988,0.0023640802,0.68283236,0.005000149,0.0008131965,0.0009875486,0.003079698,0.0002598036,0.00013143643,0.016701085,0.28177455,0.00075307355],"about_ca_topic_score_codex":0.00007237426,"about_ca_topic_score_gemma":0.000011446585,"teacher_disagreement_score":0.62567055,"about_ca_system_score_codex":0.00006957146,"about_ca_system_score_gemma":0.00009141318,"threshold_uncertainty_score":0.99999535},"labels":[],"label_agreement":null},{"id":"W2101532867","doi":"","title":"Aborder l’éthique en supervision clinique Une approche en 3 temps","year":2013,"lang":"fr","type":"article","venue":"Europe PMC (PubMed Central)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Humanities; Philosophy","score_opus":0.01846445234964524,"score_gpt":0.2830936237484193,"score_spread":0.26462917139877407,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2101532867","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8310742,0.018608293,0.0017556383,0.08537625,0.012212679,0.005683611,0.00009451277,0.0005506806,0.04464417],"genre_scores_gemma":[0.85179603,0.042417023,0.0056166756,0.011136353,0.010356596,0.00081563636,0.0002734694,0.00038503742,0.07720321],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99182326,0.0017564071,0.0014991973,0.0012367858,0.0007765739,0.0029077649],"domain_scores_gemma":[0.97562903,0.019634707,0.00029972984,0.0012824803,0.0005595256,0.0025945231],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0028287673,0.00074180035,0.0010649755,0.00014438976,0.00014842283,0.00019373791,0.0006412255,0.0010242712,0.003326637],"category_scores_gemma":[0.115817055,0.00064095925,0.00038650056,0.0010484225,0.00042119462,0.00032806201,0.00032466775,0.0023099107,0.007307809],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00031465915,0.003470253,0.13894843,0.00059532264,0.00050309376,0.00092865754,0.022989154,0.0001147595,0.0009892603,0.0065825093,0.17504512,0.6495188],"study_design_scores_gemma":[0.0031560997,0.0004282772,0.46969572,0.0037562686,0.0003785896,0.00012226558,0.000759976,0.00039502105,0.00095489156,0.00074335444,0.5189064,0.00070311147],"about_ca_topic_score_codex":0.0016868534,"about_ca_topic_score_gemma":0.00003235096,"teacher_disagreement_score":0.6488157,"about_ca_system_score_codex":0.0003433884,"about_ca_system_score_gemma":0.0006586934,"threshold_uncertainty_score":0.9999918},"labels":[],"label_agreement":null},{"id":"W2101607015","doi":"10.3138/jvme.35.3.375","title":"A Developmental Approach to Evaluating Competence in Clinical Reasoning","year":2008,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":25,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Competence (human resources); Medical education; Objectivity (philosophy); Core competency; Psychology; Accountability; Concordance; Medicine; Social psychology","score_opus":0.22590127705781737,"score_gpt":0.4977869703053811,"score_spread":0.27188569324756373,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2101607015","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99442416,0.00023610095,0.0005738076,0.0014660428,0.000675156,0.00016197604,2.3962212e-7,0.000009061883,0.00245344],"genre_scores_gemma":[0.90042275,0.00017880385,0.09615945,0.0024699015,0.0006435523,0.000009949246,0.000004512831,0.000010778854,0.000100289486],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99722487,0.00021558661,0.0012551179,0.00019402827,0.000912066,0.00019832797],"domain_scores_gemma":[0.9959757,0.0024932341,0.00030469982,0.00010583436,0.00026810987,0.00085245894],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0031535004,0.000106879554,0.00045039188,0.00017786991,0.00005029059,0.0000072202383,0.00016036822,0.00013733139,0.00012977114],"category_scores_gemma":[0.13898328,0.0000858795,0.00012999952,0.0003062127,0.000108138694,0.00007872727,0.00005678539,0.00063656067,0.000024171712],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0020594315,0.009048083,0.59277004,0.00007608145,0.00007569655,0.001040384,0.002635977,0.000029255974,0.00042985147,0.00012226588,0.010159534,0.38155338],"study_design_scores_gemma":[0.0022916335,0.00321012,0.96497715,0.0047120648,0.000032742104,0.021308262,0.0006893374,0.001042595,0.000009610009,0.000032301257,0.0015752643,0.00011891177],"about_ca_topic_score_codex":0.000023326336,"about_ca_topic_score_gemma":9.1384885e-7,"teacher_disagreement_score":0.38143447,"about_ca_system_score_codex":0.00017105295,"about_ca_system_score_gemma":0.004792798,"threshold_uncertainty_score":0.86826944},"labels":[],"label_agreement":null},{"id":"W2103425045","doi":"10.1111/1467-9280.00225","title":"On the Difficulty of Noticing Obvious Features in Patient Appearance","year":2000,"lang":"en","type":"article","venue":"Psychological Science","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":91,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Psychology; Key (lock); Cognitive psychology; Facilitation; Feature (linguistics); Social psychology; Linguistics; Computer science; Neuroscience","score_opus":0.027547008236860588,"score_gpt":0.3698635853433076,"score_spread":0.34231657710644703,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2103425045","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97980297,0.000072008515,0.000010317422,0.0024165206,0.00007716947,0.00016640036,9.880774e-7,0.000016048276,0.017437559],"genre_scores_gemma":[0.99674946,0.000038560996,0.00013721809,0.0029552153,0.00001594476,0.0000072266307,2.4022788e-7,0.0000019890226,0.000094161405],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9988534,0.000034382,0.00019615612,0.00029790102,0.0003851004,0.00023309981],"domain_scores_gemma":[0.9976001,0.001966765,0.000038617916,0.00026460734,0.000030440353,0.00009947523],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0004438455,0.00006940584,0.00015326824,0.000027359903,0.000059778202,0.000010505084,0.00018963472,0.00004667697,0.00032877456],"category_scores_gemma":[0.011708417,0.000032002834,0.000046745485,0.00054068473,0.0006180082,0.000015965727,0.000013347025,0.00022361091,0.000066616565],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012309134,0.0042462875,0.027321124,0.00000894714,0.0000059948297,0.00018244635,0.0010670568,0.001038895,0.006138856,0.00455028,0.002052979,0.95215625],"study_design_scores_gemma":[0.00043177258,0.0007752471,0.99617165,0.0005571176,0.000003483368,0.000021999509,0.000033581073,0.00011651193,0.0007295755,0.0010568317,0.0000477282,0.00005452692],"about_ca_topic_score_codex":0.000028430924,"about_ca_topic_score_gemma":0.0000022355005,"teacher_disagreement_score":0.9688505,"about_ca_system_score_codex":0.00002096061,"about_ca_system_score_gemma":0.000013774376,"threshold_uncertainty_score":0.99661636},"labels":[],"label_agreement":null},{"id":"W2104095965","doi":"10.1111/j.1365-2929.2007.02688.x","title":"The benefits of flexibility: the pedagogical value of instructions to adopt multifaceted diagnostic reasoning strategies","year":2007,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":118,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medical diagnosis; Novelty; Task (project management); Psychology; Feeling; Flexibility (engineering); Test (biology); Cognitive psychology; Computer science; Artificial intelligence; Social psychology; Medicine; Radiology","score_opus":0.04647642659629579,"score_gpt":0.41587508682985275,"score_spread":0.36939866023355694,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2104095965","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9837343,0.00078351045,0.0011119284,0.011720458,0.00093836634,0.0004274883,0.0000036974918,0.000027818474,0.0012524335],"genre_scores_gemma":[0.9968013,0.00020513433,0.0014457077,0.0009671358,0.00034077026,0.000033311633,0.000018578497,0.0000101175565,0.0001779163],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9980577,0.00009342701,0.00064946903,0.00021158425,0.00074964063,0.00023818272],"domain_scores_gemma":[0.9652069,0.033084024,0.00022151487,0.0006041381,0.00049486523,0.0003885298],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002024927,0.0001093073,0.00025358936,0.00005496338,0.00014423062,0.00001530248,0.00019534751,0.00014705876,0.00011632623],"category_scores_gemma":[0.32654965,0.000058925743,0.000101066995,0.0004299351,0.0004253038,0.000033097756,0.000054118234,0.00033085083,0.000010789632],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00033529632,0.0018523417,0.06778702,0.000075556905,0.00008522023,0.000003061695,0.005162085,0.00033701878,0.0001044295,0.059713047,0.004365773,0.8601791],"study_design_scores_gemma":[0.0007416902,0.000475695,0.9748363,0.002408655,0.00016452702,0.000050450235,0.013557129,0.0004464474,0.0008090973,0.0011130426,0.005287993,0.00010897344],"about_ca_topic_score_codex":0.00051776745,"about_ca_topic_score_gemma":0.00014534612,"teacher_disagreement_score":0.9070493,"about_ca_system_score_codex":0.000048978545,"about_ca_system_score_gemma":0.0015477147,"threshold_uncertainty_score":0.6791231},"labels":[],"label_agreement":null},{"id":"W2104889028","doi":"10.1136/bmjqs-2012-001537","title":"Do you have to re-examine to reconsider your diagnosis? Checklists and cardiac exam","year":2013,"lang":"en","type":"article","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":34,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; University Health Network","funders":"","keywords":"Medicine; Medical education; Medical physics; Data science; Computer science","score_opus":0.10233417627568472,"score_gpt":0.43024670547710586,"score_spread":0.3279125292014211,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2104889028","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9315262,0.0005275545,0.00011905814,0.05762745,0.00050886814,0.0021985376,0.00010728069,0.00012441211,0.0072606537],"genre_scores_gemma":[0.9780091,0.00039546003,0.0030625905,0.010120478,0.00088880304,0.00058379024,0.000076308,0.000051322422,0.0068121664],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.996643,0.00032739824,0.0011765979,0.0007940524,0.00051577564,0.00054319296],"domain_scores_gemma":[0.9835303,0.01356135,0.00016376738,0.0010413249,0.00035712853,0.0013461278],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0030171005,0.0003113842,0.00096402346,0.00009138538,0.0001110523,0.00008090334,0.00012851684,0.00022877606,0.0011114599],"category_scores_gemma":[0.11833203,0.00026205642,0.00021357008,0.00023057919,0.000110097106,0.000092401926,0.00022294959,0.00030655082,0.001092702],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00040205056,0.00036267182,0.66332006,0.00008873136,0.00014791058,0.000039900948,0.002051613,0.0000068772883,0.00019942895,0.00039744042,0.21263723,0.12034607],"study_design_scores_gemma":[0.0015316781,0.0003796767,0.9549753,0.0011169402,0.00012122687,0.000011945415,0.0014868323,0.0000067129663,0.00027228455,0.0003377598,0.03936258,0.00039708018],"about_ca_topic_score_codex":0.002479327,"about_ca_topic_score_gemma":0.000100181824,"teacher_disagreement_score":0.2916552,"about_ca_system_score_codex":0.0001223701,"about_ca_system_score_gemma":0.0001582659,"threshold_uncertainty_score":0.9999832},"labels":[],"label_agreement":null},{"id":"W2105400978","doi":"10.1051/pmed:2000005","title":"Pourquoi « se centrer sur le participant »en formation médicale continue ?","year":2000,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval","funders":"","keywords":"Humanities; Philosophy; Political science","score_opus":0.09339803299109975,"score_gpt":0.35676318886641306,"score_spread":0.2633651558753133,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2105400978","genre_codex":"empirical","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.81234497,0.06491677,0.0007762528,0.091643736,0.0025257464,0.0010741383,0.00024026947,0.00042608954,0.026052019],"genre_scores_gemma":[0.48253763,0.016829815,0.0015595688,0.0066717574,0.0036014009,0.00010525309,0.00031578206,0.00012787184,0.48825094],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9947422,0.0005037657,0.0013992945,0.0007418227,0.00088905304,0.0017238315],"domain_scores_gemma":[0.9839594,0.013207104,0.00022338257,0.0008569039,0.00018360886,0.0015696007],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0013650648,0.00050234457,0.0010500223,0.00009739883,0.000247987,0.00013228774,0.0003613767,0.00068817753,0.025219802],"category_scores_gemma":[0.036213297,0.00047218427,0.0004756349,0.00041722006,0.0006650786,0.0003598099,0.00012479503,0.00094781467,0.010341353],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004244566,0.0049215807,0.010700885,0.00021957056,0.00021757712,0.0007074779,0.0022532989,0.00019551424,0.00024813446,0.0006367141,0.25757346,0.7219013],"study_design_scores_gemma":[0.008437367,0.0006074583,0.026862158,0.0056022657,0.0009321218,0.00021046738,0.0010315423,0.011544162,0.0017109337,0.00090088096,0.9414194,0.0007412004],"about_ca_topic_score_codex":0.0011166205,"about_ca_topic_score_gemma":0.00007190309,"teacher_disagreement_score":0.7211601,"about_ca_system_score_codex":0.000194373,"about_ca_system_score_gemma":0.000862666,"threshold_uncertainty_score":0.99977297},"labels":[],"label_agreement":null},{"id":"W2106288357","doi":"10.1111/j.1365-2923.2011.04211.x","title":"How to construct and implement script concordance tests: insights from a systematic review","year":2012,"lang":"en","type":"review","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":143,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université de Montréal","funders":"","keywords":"Concordance; MEDLINE; Construct (python library); PsycINFO; Construct validity; Competence (human resources); Medical education; Formative assessment; Psychology; Computer science; Medicine; Psychometrics; Social psychology; Mathematics education; Clinical psychology","score_opus":0.04894979007000593,"score_gpt":0.41865530109327226,"score_spread":0.3697055110232663,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2106288357","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000021753694,0.98953146,0.00004461463,0.005278987,0.0013001759,0.0035528583,0.00002658035,0.00004079245,0.00020276642],"genre_scores_gemma":[0.00006919874,0.9835894,0.0006130695,0.0126190465,0.00092891057,0.0011477914,0.0005630743,0.00003600095,0.00043352414],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99696356,0.00031922935,0.0011667606,0.0004955251,0.00077571324,0.00027918894],"domain_scores_gemma":[0.98908186,0.008121716,0.0005880732,0.000760652,0.00020074795,0.0012469522],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00064761005,0.00037328684,0.0031956078,0.000100722566,0.000047071135,0.000049437465,0.00018939082,0.00030102758,0.00034957906],"category_scores_gemma":[0.17437409,0.00023676838,0.00023961188,0.00030095124,0.00009999447,0.000046575125,0.00008974847,0.00046996778,0.00016413268],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000013806081,0.00017008737,0.000025198897,0.45385364,0.00010339424,0.0000051922543,0.000039647537,6.335857e-11,2.5342654e-8,0.00022575117,0.020502528,0.5250732],"study_design_scores_gemma":[0.00012699133,0.00005011231,0.00002488406,0.5401536,0.0038301153,0.000117234704,0.00003418231,2.665738e-7,4.6501683e-8,0.000026617861,0.4555075,0.00012843736],"about_ca_topic_score_codex":0.00010683463,"about_ca_topic_score_gemma":0.000017243521,"teacher_disagreement_score":0.5249447,"about_ca_system_score_codex":0.00020892434,"about_ca_system_score_gemma":0.0031371615,"threshold_uncertainty_score":0.96551347},"labels":[],"label_agreement":null},{"id":"W2106660946","doi":"10.1002/j.0022-0337.2011.75.3.tb05045.x","title":"Evolution of Clinical Reasoning in Dental Education","year":2011,"lang":"en","type":"article","venue":"Journal of Dental Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":51,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Psychosocial; Curriculum; Competence (human resources); Dental education; Psychology; Evidence-based dentistry; Oral health; Medical education; Medicine; Engineering ethics; Pedagogy; Dentistry; Alternative medicine; Social psychology; Psychotherapist","score_opus":0.04329727987987227,"score_gpt":0.4176527625913759,"score_spread":0.37435548271150365,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2106660946","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9890173,0.0009257052,0.000064273525,0.00008287079,0.0031872953,0.00013070126,9.107891e-7,0.0000044131916,0.006586495],"genre_scores_gemma":[0.9939351,0.00014329857,0.0046889107,0.00012574141,0.0007041686,0.00000353278,0.000007806079,0.000010065714,0.00038139007],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.998205,0.00009373149,0.0011784195,0.00011943581,0.00029383515,0.000109578956],"domain_scores_gemma":[0.997926,0.0006145775,0.00081643125,0.00013915582,0.0003051557,0.00019870026],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00096967747,0.00008154999,0.0003476306,0.0002533316,0.000016833264,0.000006130338,0.00008750005,0.00011847997,0.00021592015],"category_scores_gemma":[0.016257472,0.000071284936,0.00020042913,0.00023436004,0.00006718369,0.00016539708,0.000016516678,0.0003148127,0.000028340544],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003749868,0.0040389826,0.94647735,0.000019080813,0.000021343769,0.000005152784,0.00040345383,8.8179536e-7,0.0001934032,0.00054317067,0.0016651469,0.04625703],"study_design_scores_gemma":[0.0011088831,0.0005942921,0.98863685,0.002553781,0.00011778974,0.00047829407,0.005243804,0.000026400046,0.0005939865,0.00044456866,0.00014044055,0.000060928127],"about_ca_topic_score_codex":0.000477531,"about_ca_topic_score_gemma":0.000040707204,"teacher_disagreement_score":0.046196103,"about_ca_system_score_codex":0.00023546204,"about_ca_system_score_gemma":0.0023009602,"threshold_uncertainty_score":0.992029},"labels":[],"label_agreement":null},{"id":"W2106671991","doi":"10.1161/circulationaha.114.014508","title":"Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD)","year":2015,"lang":"en","type":"article","venue":"Circulation","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":771,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Medical Research Council; University of North Carolina at Chapel Hill; Nederlandse Organisatie voor Wetenschappelijk Onderzoek; Cleveland Clinic; University College London; Albert-Ludwigs-Universität Freiburg; Universiteit Maastricht; ZonMw; University of Oxford; National Institute for Health and Care Research; Cancer Research UK; University of Ottawa; Vanderbilt University; Memorial Sloan-Kettering Cancer Center","keywords":"Tripod (photography); Medicine; Checklist; Statement (logic); Predictive modelling; Computer science; Machine learning; Psychology; Cognitive psychology","score_opus":0.2694120548083633,"score_gpt":0.40068665296584277,"score_spread":0.1312745981574795,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2106671991","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8915281,0.00010168756,0.10640227,0.00036817847,0.00017752085,0.0010764564,0.00007783038,0.00008406759,0.00018386751],"genre_scores_gemma":[0.9875511,0.000013295496,0.01145928,0.00006229598,0.00013647329,0.00039470854,0.0003228118,0.000017586579,0.000042472602],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99807495,0.000021633663,0.0011811319,0.00023794571,0.00032984483,0.00015449384],"domain_scores_gemma":[0.99694586,0.0015912247,0.00075861806,0.00016564764,0.0003715639,0.00016711545],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0011108342,0.0000944209,0.00033244648,0.00007160796,0.000037302863,0.000010590881,0.000036392077,0.00013083883,0.00001280621],"category_scores_gemma":[0.074053586,0.00007639811,0.000118126416,0.0001721259,0.000029086532,0.00007765261,0.000009484806,0.0000749027,0.0000013377506],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00033068884,0.0005663758,0.946954,0.0001028938,0.00007809041,0.0000029296837,0.0010724068,0.030322012,0.00017063956,0.000079733574,0.0010687311,0.019251494],"study_design_scores_gemma":[0.0031316087,0.00027935058,0.4152936,0.0010496642,0.00049353077,0.000012392856,0.00010875738,0.57753104,0.0012459755,0.00069701084,0.00006819883,0.000088890105],"about_ca_topic_score_codex":0.000043276297,"about_ca_topic_score_gemma":0.000004990063,"teacher_disagreement_score":0.547209,"about_ca_system_score_codex":0.00006707956,"about_ca_system_score_gemma":0.00027931962,"threshold_uncertainty_score":0.93374604},"labels":[],"label_agreement":null},{"id":"W2109326207","doi":"10.1016/j.ijporl.2012.02.060","title":"Evidence based curriculum design—Use of a surgical database to identify non-perceived needs in educating otolaryngologists on pediatric conditions","year":2012,"lang":"en","type":"article","venue":"International Journal of Pediatric Otorhinolaryngology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Stollery Children's Hospital; University of Alberta Hospital","funders":"","keywords":"Medicine; Otorhinolaryngology; Medical diagnosis; Pediatrics; Retrospective cohort study; Pediatric surgery; Asthma; Disease; Pediatric gastroenterology; General surgery; Surgery; Internal medicine; Pathology","score_opus":0.0902893777163538,"score_gpt":0.41030009115488797,"score_spread":0.32001071343853416,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2109326207","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99363184,0.00075724226,0.00092148205,0.0017719257,0.0024160442,0.00040107264,0.000036931717,0.00001587716,0.000047606714],"genre_scores_gemma":[0.9897804,0.0011508516,0.0049826084,0.0009417212,0.003034676,0.000027450182,0.000031115986,0.000024659317,0.000026486556],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99658436,0.0004197919,0.0014134805,0.00025702547,0.0008683321,0.00045704085],"domain_scores_gemma":[0.97818565,0.019140845,0.0008439681,0.00030516883,0.0009537432,0.00057061855],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0017864751,0.00024683896,0.0006590675,0.0019841,0.000038720416,0.000033612254,0.00048475992,0.0002191405,0.00027438253],"category_scores_gemma":[0.04896167,0.00021146022,0.00029699475,0.0008722484,0.000063484214,0.0005714322,0.00012384658,0.00075418106,0.00007381402],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006724531,0.0022648538,0.9887425,0.00003414639,0.00006645238,0.000632182,0.0002174529,0.0005311581,0.00088696775,0.00011081635,0.005649388,0.00019166677],"study_design_scores_gemma":[0.0026106872,0.0010848067,0.9933066,0.0008670086,0.00033820973,0.000846115,0.00006124197,0.00033264328,0.00021539861,0.000022863363,0.00011600494,0.00019845564],"about_ca_topic_score_codex":0.00020500581,"about_ca_topic_score_gemma":0.0000033463543,"teacher_disagreement_score":0.047175195,"about_ca_system_score_codex":0.00024287864,"about_ca_system_score_gemma":0.0007253477,"threshold_uncertainty_score":0.95904934},"labels":[],"label_agreement":null},{"id":"W2109630225","doi":"10.1006/jbin.2001.1021","title":"Analysis of Complex Decision-Making Processes in Health Care: Cognitive Approaches to Health Informatics","year":2001,"lang":"en","type":"article","venue":"Journal of Biomedical Informatics","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":108,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"York University","funders":"","keywords":"Health informatics; Computer science; R-CAST; Health care; Cognition; Variety (cybernetics); Clinical decision support system; Decision engineering; Decision support system; Management science; Knowledge management; Context (archaeology); Data science; Business decision mapping; Decision analysis; Artificial intelligence; Psychology","score_opus":0.1092909583761866,"score_gpt":0.41118709112473917,"score_spread":0.30189613274855254,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2109630225","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.82373476,0.0011475908,0.16747537,0.005645639,0.00020138145,0.000650716,0.00008869574,0.000021999476,0.0010338587],"genre_scores_gemma":[0.9388465,0.0009147707,0.05470278,0.005365201,0.00008135999,0.0000028745326,0.00007292012,0.000009210685,0.000004354518],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99386305,0.000055149965,0.0044286563,0.000066268876,0.0012000547,0.00038683967],"domain_scores_gemma":[0.9886873,0.0075396975,0.0022569112,0.00018283978,0.0005532835,0.0007799691],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0022358182,0.00017222464,0.0014194541,0.0015726282,0.00005671122,0.000028699724,0.00020182732,0.00012408459,0.0000606433],"category_scores_gemma":[0.032437164,0.00012370659,0.00023708298,0.0035494622,0.00016892322,0.00017724407,0.000079958765,0.0004312336,0.0000070849264],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001085089,0.0014690704,0.048522994,0.0022967842,0.0011241545,0.000049079274,0.0814409,0.0018063636,4.3181996e-7,0.000064215055,0.0035732586,0.85856766],"study_design_scores_gemma":[0.021393996,0.024300162,0.3927372,0.11267308,0.0033817927,0.0017598189,0.30580565,0.09457855,0.000031942232,0.00042718364,0.04165052,0.0012600947],"about_ca_topic_score_codex":0.0000323692,"about_ca_topic_score_gemma":0.000050572547,"teacher_disagreement_score":0.85730755,"about_ca_system_score_codex":0.0002528811,"about_ca_system_score_gemma":0.0018862969,"threshold_uncertainty_score":0.975713},"labels":[],"label_agreement":null},{"id":"W2109781953","doi":"10.1007/s10459-012-9427-3","title":"Script concordance testing in continuing professional development: local or international reference panels?","year":2012,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"College of Family Physicians of Canada; Université de Montréal","funders":"","keywords":"Cronbach's alpha; Generalizability theory; Concordance; Statistics; Psychology; Descriptive statistics; Medicine; Mathematics","score_opus":0.08091195238689003,"score_gpt":0.4649335106945761,"score_spread":0.3840215583076861,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2109781953","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9816826,0.0052887104,0.00055347336,0.0022279706,0.0054026316,0.00042803973,0.0000013197103,0.000032089527,0.004383149],"genre_scores_gemma":[0.9720237,0.0004761891,0.025339333,0.0012804,0.00021590177,0.000089096866,0.000015701671,0.0000048000056,0.00055486785],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982608,0.00007101762,0.0006136062,0.00027910277,0.0003599267,0.00041554257],"domain_scores_gemma":[0.995967,0.0033710946,0.00025213844,0.00009302945,0.0001138553,0.00020288795],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00176012,0.00009562887,0.00019937895,0.00016525343,0.000097883596,0.00001552513,0.00014790251,0.000045124212,0.00006152105],"category_scores_gemma":[0.018576056,0.00007272424,0.000010294337,0.00065586023,0.00021508323,0.00055877375,0.000037713457,0.00023234752,0.000018201717],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000032116215,0.00027698113,0.58431417,0.000042791966,4.471292e-7,6.594272e-7,0.0011270728,0.000040229574,0.0000055301043,0.0005675045,0.000048784907,0.4135437],"study_design_scores_gemma":[0.00057141733,0.00015324722,0.9576448,0.009509255,0.0000015158113,0.000035304052,0.0040943744,0.0008285988,0.000073655465,0.0003963811,0.026562657,0.00012876393],"about_ca_topic_score_codex":0.00026251766,"about_ca_topic_score_gemma":0.00021094948,"teacher_disagreement_score":0.41341493,"about_ca_system_score_codex":0.00031692997,"about_ca_system_score_gemma":0.0021417043,"threshold_uncertainty_score":0.9896909},"labels":[],"label_agreement":null},{"id":"W2109878483","doi":"10.5858/2003-127-1489-iosbbg","title":"Interpretation of Skin Biopsies by General Pathologists: Diagnostic Discrepancy Rate Measured by Blinded Review","year":2003,"lang":"en","type":"article","venue":"Archives of Pathology & Laboratory Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":51,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Calgary Laboratory Services; University of Calgary","funders":"","keywords":"Medicine; Blinded study; Interpretation (philosophy); Pathology; Dermatology; Medical physics; Radiology; Surgery; Computer science","score_opus":0.01878408377095505,"score_gpt":0.3282379413903853,"score_spread":0.3094538576194303,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2109878483","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8541284,0.13586281,0.0024526208,0.0036049436,0.00045694332,0.00086658064,0.00036950808,0.00007292415,0.0021852525],"genre_scores_gemma":[0.9555586,0.038121756,0.0023251968,0.0032126463,0.000079796846,0.00007768087,0.00037915734,0.000038482445,0.0002067007],"study_design_codex":"bench_or_experimental","study_design_gemma":"observational","domain_scores_codex":[0.9964137,0.0011416443,0.0012440594,0.0005219868,0.0003316068,0.00034702464],"domain_scores_gemma":[0.98476094,0.013314287,0.0006851366,0.0006782762,0.00029290875,0.0002684712],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001541975,0.00031257106,0.0013710062,0.00015628495,0.000043662465,0.0000022192262,0.00018408395,0.0001589979,0.00016126403],"category_scores_gemma":[0.15672171,0.00023028995,0.0001590053,0.0004407394,0.0018210963,0.00004959351,0.00004064846,0.0003490704,0.000008672015],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001776094,0.003883193,0.17403471,0.004564793,0.0007499382,0.0016091305,0.004887857,0.000010782897,0.6391496,0.005517781,0.11213365,0.05168247],"study_design_scores_gemma":[0.045893505,0.022944752,0.46122438,0.14902103,0.008463502,0.0011847841,0.0025696587,0.00036407213,0.26183087,0.013058183,0.030704515,0.00274075],"about_ca_topic_score_codex":0.000026898093,"about_ca_topic_score_gemma":0.0000057722614,"teacher_disagreement_score":0.3773187,"about_ca_system_score_codex":0.000020212157,"about_ca_system_score_gemma":0.00023511052,"threshold_uncertainty_score":0.93909514},"labels":[],"label_agreement":null},{"id":"W2110061847","doi":"10.7202/1035703ar","title":"Comment améliorer la réussite à l’examen d’anatomie en 1 année de bachelier en médecine ? Étude de l’impact d’une activité intégrée dans un dispositif hybride de formation","year":2014,"lang":"fr","type":"article","venue":"Revue internationale des technologies en pédagogie universitaire","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Humanities; Political science; Philosophy","score_opus":0.015932293912879033,"score_gpt":0.2911145444365629,"score_spread":0.2751822505236839,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2110061847","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9013979,0.0012076768,0.04642609,0.046840765,0.00023872942,0.00040887293,0.00015691279,0.0006248996,0.0026981425],"genre_scores_gemma":[0.9810081,0.0075105703,0.006825729,0.0003706083,0.00029446412,0.000029881168,0.00020047063,0.00006886767,0.0036912917],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99689895,0.0005544614,0.00058335916,0.00057112554,0.00041678775,0.00097531534],"domain_scores_gemma":[0.9826265,0.015762182,0.00036289002,0.00070644356,0.00022123852,0.00032076472],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.001503532,0.00051905855,0.00066319446,0.0005186238,0.00021423846,0.00010689239,0.0010394083,0.0008030104,0.00011930011],"category_scores_gemma":[0.025470993,0.0005279671,0.0004098431,0.00052550924,0.00064904854,0.0005979355,0.0006923419,0.0011469971,0.00009110097],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00034780012,0.0010186795,0.31494734,0.00024262478,0.00089433434,0.001112421,0.0070206355,0.0029488243,0.0022056345,0.008647949,0.002732232,0.65788156],"study_design_scores_gemma":[0.00812483,0.002107168,0.74979603,0.017721077,0.0015586374,0.0031138787,0.044429787,0.07604568,0.024190797,0.014726804,0.056668,0.0015173049],"about_ca_topic_score_codex":0.0005976525,"about_ca_topic_score_gemma":0.00015735505,"teacher_disagreement_score":0.65636426,"about_ca_system_score_codex":0.006343398,"about_ca_system_score_gemma":0.00032930818,"threshold_uncertainty_score":0.9997172},"labels":[],"label_agreement":null},{"id":"W2110201643","doi":"10.3138/cbmh.26.1.107","title":"The Epistemological Aspects in the Historical Works of Don Bates; For Don, from one unremitting Kuhnian to another","year":2009,"lang":"en","type":"article","venue":"Canadian Journal of Health History","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"BATES; Soul; Epistemology; Hacker; Philosophy; Closing (real estate); Object (grammar); Sociology; Computer science; Law","score_opus":0.0685371328478442,"score_gpt":0.31818950542147606,"score_spread":0.24965237257363188,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2110201643","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.44399405,0.15630536,0.0017087869,0.38817433,0.0034732642,0.0014256325,0.000018481318,0.00001665593,0.0048834546],"genre_scores_gemma":[0.97593147,0.00019652757,0.0012710252,0.0217829,0.0005070906,0.000005094173,0.00000254931,0.000012737773,0.0002906273],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99799734,0.0002485197,0.0009394654,0.00015050184,0.00025143026,0.0004127627],"domain_scores_gemma":[0.9913646,0.007083418,0.00046787082,0.00025804728,0.00016322779,0.00066285796],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0027663747,0.00010910857,0.00053401,0.00015196324,0.00014473466,0.000009213794,0.0002180727,0.0001238861,0.000063233],"category_scores_gemma":[0.020028831,0.00006735589,0.00015449709,0.00014979286,0.00009300954,0.000023721332,0.0000036670808,0.00052704307,0.0000022996876],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009180288,0.00047316198,0.023064496,0.000050865754,0.00006428678,0.0004898058,0.007906264,0.00004415606,0.00007220457,0.0017656664,0.62016284,0.34498823],"study_design_scores_gemma":[0.0018373764,0.0028469868,0.23024867,0.003078572,0.000047510854,0.00011222726,0.00052086945,0.000016460715,0.000006131901,0.0013815459,0.7597714,0.00013225246],"about_ca_topic_score_codex":0.017598739,"about_ca_topic_score_gemma":0.020444963,"teacher_disagreement_score":0.5319374,"about_ca_system_score_codex":0.0039227284,"about_ca_system_score_gemma":0.0032608367,"threshold_uncertainty_score":0.99990106},"labels":[],"label_agreement":null},{"id":"W2110273965","doi":"10.3138/jvme.38.4.360","title":"An Instructional Model for Training Competence in Solving Clinical Problems","year":2011,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Premise; Competence (human resources); Computer science; Openness to experience; Instructional design; Mathematics education; Psychology; Medical education; Medicine","score_opus":0.33385320046281663,"score_gpt":0.4837545079428948,"score_spread":0.14990130748007818,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2110273965","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9867377,0.00011774577,0.010385357,0.0010330525,0.0012794902,0.00016008885,0.0000011780132,0.000009534366,0.00027584893],"genre_scores_gemma":[0.9224563,0.00013042586,0.07565202,0.0009419753,0.0007601947,0.000014016351,0.000008488749,0.000011614432,0.000024944178],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99813706,0.00008410349,0.0010458521,0.00015982169,0.00040108038,0.00017208027],"domain_scores_gemma":[0.9971846,0.0014908558,0.00034611157,0.0001290614,0.00027668694,0.00057268556],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0021797097,0.000089680754,0.00036169574,0.00013437611,0.000032589276,0.000007884336,0.00014575938,0.00016775567,0.00014651142],"category_scores_gemma":[0.030374829,0.00007263034,0.00014597284,0.000098505436,0.0001282159,0.00017557475,0.000014857996,0.0004711346,0.0000016050076],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0025240083,0.008053577,0.09307493,0.00014244663,0.000075823635,0.00012524448,0.006224726,0.00016640352,0.00034384927,0.001469875,0.001061346,0.88673776],"study_design_scores_gemma":[0.008713629,0.015299303,0.65767413,0.010816135,0.00018502427,0.007536648,0.0020023999,0.2853603,0.000018892797,0.010836493,0.0012243415,0.00033270736],"about_ca_topic_score_codex":0.000012066432,"about_ca_topic_score_gemma":0.000004534627,"teacher_disagreement_score":0.88640505,"about_ca_system_score_codex":0.00005769226,"about_ca_system_score_gemma":0.0031408197,"threshold_uncertainty_score":0.97779274},"labels":[],"label_agreement":null},{"id":"W2112312328","doi":"10.1111/j.1365-2753.2011.01729.x","title":"Tacit clues and the science of clinical judgement [a commentary on Henry<i>et al</i>.]","year":2011,"lang":"en","type":"letter","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Tacit knowledge; Judgement; Bioethics; Citation; Relevance (law); Medicine; Epistemology; Psychology; Philosophy; Library science; Computer science; Law","score_opus":0.19451317305968552,"score_gpt":0.5448406020447776,"score_spread":0.35032742898509206,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2112312328","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.02211931,0.0009009388,0.000054777225,0.96804595,0.0034158025,0.0008520018,0.00000818228,0.0000055028067,0.004597524],"genre_scores_gemma":[0.03706365,0.0065905126,0.0033193293,0.9500324,0.0028790287,0.000013643653,0.000015769661,0.000026300208,0.000059357328],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9824527,0.006288363,0.0066334154,0.0005726586,0.0037360012,0.00031685786],"domain_scores_gemma":[0.5845276,0.40063047,0.009412191,0.0010762017,0.0039893426,0.00036419052],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.13906835,0.00026931273,0.0017483983,0.00029770806,0.00008105284,0.000059846734,0.00042165068,0.0005050871,0.00022059686],"category_scores_gemma":[0.6932193,0.0001609496,0.00063864625,0.00035191842,0.0021648167,0.00037387057,0.0001976557,0.0068347733,0.000020168965],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004959538,0.0015846884,0.0043830485,0.000028356846,0.0003087924,0.00020840316,0.00013563842,0.000016727092,5.569482e-7,0.00012991144,0.9704991,0.01774523],"study_design_scores_gemma":[0.031046832,0.007807987,0.045212805,0.006846843,0.004589265,0.00041325024,0.00029958258,0.0012684613,0.000011391926,0.004867677,0.89731425,0.0003216846],"about_ca_topic_score_codex":0.00009480343,"about_ca_topic_score_gemma":0.0000036529784,"teacher_disagreement_score":0.55415094,"about_ca_system_score_codex":0.000118395445,"about_ca_system_score_gemma":0.0016300473,"threshold_uncertainty_score":0.9954565},"labels":[],"label_agreement":null},{"id":"W2113256441","doi":"10.1016/j.bpobgyn.2010.04.001","title":"The assessment of professional competence: building blocks for theory development","year":2010,"lang":"en","type":"review","venue":"Best Practice & Research Clinical Obstetrics & Gynaecology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":342,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre","funders":"","keywords":"Counterintuitive; Competence (human resources); Medicine; Engineering ethics; Miller; Technology assessment; Psychology; Epistemology; Engineering; Social psychology; Political science","score_opus":0.24399870384519537,"score_gpt":0.6044930499658941,"score_spread":0.36049434612069875,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2113256441","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00012869037,0.97407854,0.0005056491,0.0017730932,0.010686987,0.006050621,0.000049430604,0.000044875575,0.0066821426],"genre_scores_gemma":[0.00017149864,0.952221,0.041327767,0.0002664163,0.00067824684,0.0015433603,0.00013976078,0.000106780455,0.0035451846],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.98348504,0.0064602783,0.0047835424,0.0013553911,0.0023973074,0.0015184304],"domain_scores_gemma":[0.14884701,0.8420657,0.0023728814,0.0014768977,0.004383108,0.00085440563],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["metaresearch","research_integrity"],"category_scores_codex":[0.05895921,0.0005616958,0.0035516904,0.0005856966,0.0009826157,0.000119416916,0.0014745344,0.0024111622,0.00040374335],"category_scores_gemma":[0.915594,0.00034066616,0.0012002625,0.0015676974,0.0020296024,0.00010933495,0.0012500267,0.010232087,0.00009865233],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014002663,0.0053813388,0.0005858403,0.001724777,0.0006466659,0.00016387747,0.000021906022,1.864288e-7,1.4912504e-7,0.0116066,0.0023348767,0.97739375],"study_design_scores_gemma":[0.001890941,0.0031984528,0.0013194254,0.004737593,0.0017388738,0.000085412416,0.00023047846,0.000029084105,6.4204625e-7,0.0008443129,0.985608,0.00031677872],"about_ca_topic_score_codex":0.000007042298,"about_ca_topic_score_gemma":0.000008047938,"teacher_disagreement_score":0.98327315,"about_ca_system_score_codex":0.0006741668,"about_ca_system_score_gemma":0.022546137,"threshold_uncertainty_score":0.9999045},"labels":[],"label_agreement":null},{"id":"W2113289599","doi":"10.4212/cjhp.v65i4.1155","title":"Diagnostic Reasoning by Hospital Pharmacists: Assessment of Attitudes, Knowledge, and Skills","year":2012,"lang":"en","type":"article","venue":"The Canadian Journal of Hospital Pharmacy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"Vancouver General Hospital","funders":"","keywords":"Pharmacy; Descriptive statistics; Family medicine; Observational study; Medicine; Medical education; Pharmacist; Psychology; Pathology","score_opus":0.015513288996438224,"score_gpt":0.36253302035376733,"score_spread":0.3470197313573291,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2113289599","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9715056,0.021534422,0.000036317426,0.0028560052,0.0016455911,0.00033068086,0.000056173976,0.000009154053,0.0020261004],"genre_scores_gemma":[0.99690646,0.0012824773,0.00041291156,0.00035887453,0.0007815225,0.000007078534,0.000012621041,0.000033569995,0.00020449328],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99820596,0.00010634616,0.0006543034,0.00014927276,0.00034369781,0.0005404318],"domain_scores_gemma":[0.9926989,0.004577538,0.00044345693,0.00021331546,0.00036744235,0.0016993232],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0011348556,0.00022387134,0.00050922093,0.00012937568,0.00017798816,0.000049629507,0.00024726152,0.000058568345,0.00027633703],"category_scores_gemma":[0.011924539,0.00015708676,0.00019464485,0.00020617893,0.00038964997,0.00023980511,0.000040712916,0.0006303263,0.000013668656],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00001961376,0.00093729777,0.9233742,0.000042541673,0.00030814108,0.00013505248,0.001049708,0.0000033401743,0.000074434116,0.00026425358,0.06440714,0.009384311],"study_design_scores_gemma":[0.0067150374,0.001773309,0.96043795,0.0024260883,0.0012098795,0.00072735496,0.00019096224,0.00008290579,0.0015159127,0.00014014127,0.024320174,0.00046028206],"about_ca_topic_score_codex":0.0013252939,"about_ca_topic_score_gemma":0.00014280187,"teacher_disagreement_score":0.04008697,"about_ca_system_score_codex":0.00022503981,"about_ca_system_score_gemma":0.0011706114,"threshold_uncertainty_score":0.99639845},"labels":[],"label_agreement":null},{"id":"W2114027476","doi":"10.1111/tct.12062","title":"Pro forma: impact on communication skills?","year":2013,"lang":"en","type":"article","venue":"The Clinical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Wilcoxon signed-rank test; Communication skills; Interpersonal communication; Medical education; Test (biology); Medical history; Psychology; Social skills; Set (abstract data type); Objective structured clinical examination; Family medicine; Medicine; Social psychology; Computer science; Pedagogy; Psychiatry; Surgery","score_opus":0.07006527274153086,"score_gpt":0.4451004306309185,"score_spread":0.3750351578893876,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2114027476","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94711924,0.0001488552,0.000022885544,0.018640637,0.00014290378,0.00071149727,0.000001867536,0.00012871667,0.0330834],"genre_scores_gemma":[0.9848822,0.00021211397,0.0004450685,0.0048614615,0.0004136036,0.00007292019,0.0000270275,0.000025578453,0.009060034],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982627,0.00032891706,0.00065002276,0.0002243681,0.00025557625,0.00027838783],"domain_scores_gemma":[0.9788561,0.019236716,0.00017486536,0.0013029411,0.00011903793,0.0003103329],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0018821127,0.0001597047,0.00040067793,0.000023169752,0.00010053522,0.00003739598,0.00028194662,0.00021056895,0.002016457],"category_scores_gemma":[0.044463158,0.00007569946,0.0004067975,0.00011544003,0.00029314402,0.00006118118,0.000099794466,0.0009820047,0.0039351857],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003241149,0.0047865477,0.34991384,0.000007797507,0.00024983488,0.0000054999514,0.0003194558,0.0000098745695,0.000009478784,0.00062836474,0.32019502,0.32355016],"study_design_scores_gemma":[0.0018438766,0.0010673546,0.9860763,0.00044220136,0.00011513898,0.000013814355,0.000062931045,0.0005116007,0.000016968515,0.0036295995,0.0060949354,0.00012531287],"about_ca_topic_score_codex":0.00014809512,"about_ca_topic_score_gemma":0.000002253072,"teacher_disagreement_score":0.63616246,"about_ca_system_score_codex":0.000046504632,"about_ca_system_score_gemma":0.000079461795,"threshold_uncertainty_score":0.9988958},"labels":[],"label_agreement":null},{"id":"W2114789700","doi":"10.1177/0022146509361193","title":"Physician Cognitive Processing as a Source of Diagnostic and Treatment Disparities in Coronary Heart Disease","year":2010,"lang":"en","type":"article","venue":"Journal of Health and Social Behavior","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":32,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"National Heart, Lung, and Blood Institute","keywords":"Medicine; Vignette; Cognition; Health and Retirement Study; Health care; Medical prescription; Disease; Coronary heart disease; Family medicine; Clinical psychology; Gerontology; Psychology; Psychiatry; Social psychology; Internal medicine; Nursing","score_opus":0.030788298440967626,"score_gpt":0.38717246911487013,"score_spread":0.3563841706739025,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2114789700","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99355274,0.0018181044,0.0000028184136,0.004296266,0.0000714074,0.00023627418,0.00000941176,0.0000035612527,0.000009407355],"genre_scores_gemma":[0.99796,0.00031073784,0.00006357041,0.0014433638,0.00016852042,0.000012684963,0.00000622818,0.000008517895,0.000026395635],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99911547,0.000056139557,0.00040993324,0.000091120906,0.0001653896,0.00016197276],"domain_scores_gemma":[0.997161,0.0021306619,0.00024135673,0.00003035322,0.00009551961,0.00034114782],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00023409647,0.000086942826,0.00044406514,0.000056701567,0.00009997498,0.000011726342,0.000016182863,0.00005398141,0.0000048585603],"category_scores_gemma":[0.0044367514,0.00006712701,0.00006331427,0.000046956116,0.00018620642,0.000047876783,0.00001002271,0.00023697363,3.928587e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00033233248,0.00080654887,0.7627637,0.00005876085,0.000005752106,0.000070290764,0.0015416148,2.5467662e-8,0.000019716952,0.000016139147,0.00002250199,0.23436266],"study_design_scores_gemma":[0.0024627713,0.002140974,0.99226135,0.0012103457,0.00017691246,0.00006907983,0.0014327744,0.0000026593543,0.000009735749,0.00011120208,0.000072129915,0.000050035993],"about_ca_topic_score_codex":0.0003006461,"about_ca_topic_score_gemma":0.0000447502,"teacher_disagreement_score":0.23431262,"about_ca_system_score_codex":0.00003073531,"about_ca_system_score_gemma":0.0007001134,"threshold_uncertainty_score":0.5311526},"labels":[],"label_agreement":null},{"id":"W2115868070","doi":"10.1111/j.1464-5491.2001.001-2.x","title":"How to pass on your expertise: a four‐point approach to training physical examination skills in the out‐patient setting","year":2001,"lang":"en","type":"article","venue":"Diabetic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Jewish General Hospital","funders":"","keywords":"Medicine; Citation; Point (geometry); General hospital; General practice; Library science; Medical education; Family medicine; Computer science","score_opus":0.06657608194352754,"score_gpt":0.338125201071328,"score_spread":0.27154911912780044,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2115868070","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9156218,0.000060036688,0.0024705227,0.06896206,0.00027384682,0.0011644068,0.000003056378,0.00005637534,0.011387906],"genre_scores_gemma":[0.97852564,0.000011088503,0.00088588777,0.018806351,0.00095779187,0.00032239323,0.000034231147,0.000033632507,0.0004229746],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9971926,0.0002769385,0.00044960805,0.00055451575,0.0009571028,0.00056919164],"domain_scores_gemma":[0.9922624,0.006525212,0.00010165956,0.00052833697,0.00010504052,0.00047733233],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0014330802,0.00027624107,0.0006359156,0.00023739993,0.00007289861,0.000032827094,0.000189183,0.00009388067,0.00002013066],"category_scores_gemma":[0.06653016,0.0001670032,0.000092731665,0.0005799543,0.000085327076,0.000038117432,0.0000462009,0.0003934011,0.00005017866],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010844405,0.0016626826,0.0027071827,0.000032192802,0.000033144966,0.00019895202,0.13008243,0.00012950771,0.0008362793,0.00024064064,0.010262451,0.8537061],"study_design_scores_gemma":[0.017183613,0.017620128,0.7982642,0.021949131,0.0005724898,0.00031486247,0.11012467,0.012993276,0.0022250914,0.0009923109,0.016247865,0.0015123411],"about_ca_topic_score_codex":0.000025788742,"about_ca_topic_score_gemma":0.0000043788295,"teacher_disagreement_score":0.8521938,"about_ca_system_score_codex":0.00013295359,"about_ca_system_score_gemma":0.00004573355,"threshold_uncertainty_score":0.9413329},"labels":[],"label_agreement":null},{"id":"W2116670704","doi":"10.36834/cmej.36594","title":"Consistency in diagnostic suggestions does not influence the tendency to accept them","year":2012,"lang":"en","type":"article","venue":"Canadian Medical Education Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Consistency (knowledge bases); Computer science; Artificial intelligence","score_opus":0.02127964150353493,"score_gpt":0.3438206985595111,"score_spread":0.32254105705597613,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2116670704","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7467507,0.0004741679,0.000016619557,0.24653971,0.0020560725,0.00026320587,0.000007657706,0.000012109272,0.0038797322],"genre_scores_gemma":[0.91436106,0.00026514474,0.00026298297,0.083163634,0.0012755666,0.000052120515,0.000006208099,0.000012183384,0.000601098],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9980559,0.000128312,0.0005369558,0.00014784125,0.000583855,0.00054714835],"domain_scores_gemma":[0.9519143,0.021732595,0.0002112015,0.0006744082,0.0006936464,0.024773858],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00149877,0.000121808254,0.00020544014,0.00019033122,0.00022630792,0.00005349547,0.00024943816,0.00016180244,0.014728619],"category_scores_gemma":[0.6998404,0.00007045955,0.00007683508,0.0003827982,0.00018601824,0.00010661853,0.0000205302,0.00093557726,0.0004388564],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000017318605,0.0005243899,0.7489611,0.000011093488,0.000026156482,0.000100149766,0.0025212679,0.000006460756,0.000007188299,0.0039581535,0.091856346,0.15201038],"study_design_scores_gemma":[0.0004065153,0.00005019558,0.9435044,0.0014268897,0.000054395336,0.0013813575,0.001515819,0.0000075712423,0.000012674764,0.0005095703,0.05100663,0.00012398865],"about_ca_topic_score_codex":0.009565355,"about_ca_topic_score_gemma":0.014283946,"teacher_disagreement_score":0.69834167,"about_ca_system_score_codex":0.00033456393,"about_ca_system_score_gemma":0.024117153,"threshold_uncertainty_score":0.99703},"labels":[],"label_agreement":null},{"id":"W2118651356","doi":"10.1186/cc8858","title":"Patient safety and acute care medicine: lessons for the future, insights from the past","year":2010,"lang":"en","type":"review","venue":"Critical Care","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":52,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta Hospital; Alberta Hospital Edmonton","funders":"","keywords":"Yearbook; Medicine; Acute medicine; Intensive care; Medical emergency; Emergency medicine; Library science; Intensive care medicine; Computer science","score_opus":0.044530495957130926,"score_gpt":0.41131737956671355,"score_spread":0.3667868836095826,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2118651356","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000126428795,0.9747733,0.0000672803,0.019074004,0.0022993002,0.0014549133,0.0014590385,0.00004124726,0.0008183097],"genre_scores_gemma":[0.001194755,0.98853827,0.0001563951,0.0022954363,0.0058683776,0.00025955704,0.0015876305,0.00006452258,0.000035052024],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9978886,0.00014605712,0.00064462837,0.0005780136,0.00040639506,0.00033627023],"domain_scores_gemma":[0.9343163,0.06382855,0.000099592704,0.00086451083,0.0005191073,0.0003719634],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00011437856,0.00043447234,0.0015501309,0.000030425099,0.00041491102,0.00004908564,0.00026714933,0.0006611685,0.00007004087],"category_scores_gemma":[0.026124144,0.00017570019,0.000520118,0.00013586148,0.00081159483,0.000019783129,0.00015860128,0.001335873,0.000016631633],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003841653,0.000021161082,0.0000146102075,0.0022124636,0.000118875025,0.0000424878,0.0011220824,8.365766e-9,3.601522e-8,0.0048397314,0.0068284166,0.9847617],"study_design_scores_gemma":[0.0005993964,0.00036824684,0.00031801907,0.023807414,0.00883455,0.00004358255,0.005678487,0.0000010657737,1.881241e-7,0.00022792147,0.9599126,0.00020852416],"about_ca_topic_score_codex":0.000049903203,"about_ca_topic_score_gemma":0.00008297356,"teacher_disagreement_score":0.98455316,"about_ca_system_score_codex":0.00006697512,"about_ca_system_score_gemma":0.0002772017,"threshold_uncertainty_score":0.9820792},"labels":[],"label_agreement":null},{"id":"W2118786685","doi":"10.1111/j.1365-2929.2006.02541.x","title":"Assessment of clinical reasoning in the context of uncertainty: the effect of variability within the reference panel","year":2006,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":68,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Concordance; Context (archaeology); Cronbach's alpha; Test (biology); Statistics; Psychology; Medicine; Clinical psychology; Mathematics; Psychometrics; Internal medicine","score_opus":0.03606216007835677,"score_gpt":0.43600450767934396,"score_spread":0.3999423476009872,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2118786685","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9861162,0.000114429,0.00011745203,0.009241363,0.00035593312,0.00050316873,0.0000033516756,0.0000052100468,0.0035429334],"genre_scores_gemma":[0.99851125,0.000028267035,0.00013104021,0.0010220223,0.00018079416,0.000055366392,0.0000321089,0.000004740671,0.000034408375],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9958044,0.0018945016,0.0011796366,0.00019225241,0.0008093898,0.000119780816],"domain_scores_gemma":[0.86721766,0.13030177,0.0009523037,0.0010919191,0.0003302941,0.000106054584],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.018902544,0.000092268696,0.00047680936,0.000022848575,0.00003313888,0.000004394352,0.00028266408,0.0001529524,0.00009031202],"category_scores_gemma":[0.2596491,0.000036361966,0.00012366587,0.0002591247,0.0006727099,0.000014079618,0.000034716788,0.0005949952,0.0000011399812],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001772281,0.0013158311,0.924513,0.000093231734,0.000028164497,0.0000010989218,0.00054331106,0.00002878718,0.000019358034,0.012713074,0.0023839779,0.058182955],"study_design_scores_gemma":[0.0010958257,0.0006249322,0.9917542,0.0015257372,0.000140002,0.000007637822,0.0009509469,0.002703846,0.00009079675,0.00078304,0.00028515855,0.00003791409],"about_ca_topic_score_codex":0.004256749,"about_ca_topic_score_gemma":0.0001838802,"teacher_disagreement_score":0.24074656,"about_ca_system_score_codex":0.00003193709,"about_ca_system_score_gemma":0.0017041878,"threshold_uncertainty_score":0.74658716},"labels":[],"label_agreement":null},{"id":"W2120230938","doi":"10.1111/j.1365-2929.2005.02127.x","title":"Research in clinical reasoning: past history and current trends","year":2005,"lang":"en","type":"review","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":991,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Representation (politics); Process (computing); Cognitive science; Mental representation; Scripting language; Cognition; Cognitive psychology; Psychology; Computer science; Semantic memory; Epistemology","score_opus":0.21266783297139527,"score_gpt":0.5752715931910062,"score_spread":0.36260376021961094,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2120230938","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000081560655,0.9874129,0.0000019525123,0.003272233,0.002672849,0.00035445593,0.0000021261676,0.000034417666,0.0061674733],"genre_scores_gemma":[0.00004393653,0.990522,0.00015954806,0.0005946321,0.0051406976,0.00021446656,0.00034934885,0.000043587166,0.0029317793],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9951147,0.0009559004,0.0015392817,0.00077136786,0.0011702493,0.00044851413],"domain_scores_gemma":[0.9819322,0.015771404,0.00030343764,0.00066533923,0.00019061551,0.0011369552],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0064047165,0.0002778994,0.0016805499,0.00075271714,0.000037675356,0.000015294876,0.00021230063,0.00089138723,0.0012856562],"category_scores_gemma":[0.10631657,0.000212731,0.00033176417,0.00061155634,0.00055608083,0.00003187035,0.00011910059,0.0034388998,0.0001981696],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000010543674,0.0014547291,0.00074254395,0.0011930853,0.000015027664,0.000015879219,0.000038465485,2.2755997e-9,2.0059555e-10,0.00015288623,0.19983439,0.79654247],"study_design_scores_gemma":[0.0004961495,0.00014647763,0.0036791645,0.11585927,0.00024475358,0.000103164486,0.00001813269,0.000012174107,8.8584634e-10,0.00001573167,0.87927353,0.0001514234],"about_ca_topic_score_codex":0.00008731748,"about_ca_topic_score_gemma":0.000025266138,"teacher_disagreement_score":0.796391,"about_ca_system_score_codex":0.0009859323,"about_ca_system_score_gemma":0.011003255,"threshold_uncertainty_score":0.9996273},"labels":[],"label_agreement":null},{"id":"W2120554597","doi":"10.1016/j.amjmed.2008.02.001","title":"Overconfidence in Clinical Decision Making","year":2008,"lang":"en","type":"letter","venue":"The American Journal of Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":260,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Dalhousie University","funders":"","keywords":"Medicine; Emergency department; Presentation (obstetrics); Overconfidence effect; Variety (cybernetics); Process (computing); Family medicine; Psychiatry; Artificial intelligence; Surgery; Social psychology; Psychology","score_opus":0.05037540079234066,"score_gpt":0.42119622746827673,"score_spread":0.3708208266759361,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2120554597","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.17343634,0.0024859817,0.00044754776,0.8212203,0.0014685094,0.00018529373,0.000002165158,0.00001245003,0.00074140663],"genre_scores_gemma":[0.16820425,0.01026423,0.0013377261,0.80248255,0.01742699,0.0000019199454,0.0000050491185,0.00004839392,0.0002289256],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9947983,0.00054100854,0.0025107604,0.0002960064,0.0014550332,0.00039889736],"domain_scores_gemma":[0.92908275,0.066930085,0.0027226014,0.00069516536,0.00036210578,0.00020730874],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.003993379,0.00030415322,0.0023805667,0.00037038326,0.000044274748,0.000007753849,0.00057431235,0.0002689215,0.00020666968],"category_scores_gemma":[0.11439437,0.00015421852,0.00046781308,0.00056728773,0.002466909,0.000035388173,0.00007289862,0.005846618,0.000036537847],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009148921,0.00006541887,0.010897888,0.000010865579,0.000093559865,0.01007214,0.00014157212,0.0000038678004,0.0000017644969,0.00000206964,0.7766528,0.20114313],"study_design_scores_gemma":[0.009459741,0.0149953365,0.33953553,0.088778764,0.0012994756,0.030570881,0.000524878,0.00007592431,0.0000020243886,0.001763716,0.5125283,0.0004654272],"about_ca_topic_score_codex":0.00010362779,"about_ca_topic_score_gemma":0.0000032478495,"teacher_disagreement_score":0.32863766,"about_ca_system_score_codex":0.00011929648,"about_ca_system_score_gemma":0.00056876364,"threshold_uncertainty_score":0.9964469},"labels":[],"label_agreement":null},{"id":"W2120656821","doi":"10.1111/medu.12623","title":"Self‐explanation in learning clinical reasoning: the added value of examples and prompts","year":2015,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":66,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Sherbrooke","funders":"","keywords":"Value (mathematics); Psychology; Cognitive psychology; Medical education; Medicine; Computer science; Machine learning","score_opus":0.04485477790797955,"score_gpt":0.41248423158030256,"score_spread":0.367629453672323,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2120656821","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9863515,0.00067281106,0.00008681882,0.010066843,0.0003765233,0.00025627195,3.1553554e-7,0.0000315782,0.0021573638],"genre_scores_gemma":[0.9962392,0.00029662612,0.0018527887,0.0010303864,0.00032490498,0.000028090742,0.00004153461,0.000008471247,0.00017803503],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9984314,0.0002890807,0.00048242838,0.00018958106,0.00048477127,0.00012275498],"domain_scores_gemma":[0.9935427,0.005676651,0.00015581207,0.00016821086,0.00016499628,0.000291658],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0031569726,0.000071406575,0.00023737551,0.000056351593,0.000025875017,0.0000085947095,0.000067989145,0.0001534709,0.000048025864],"category_scores_gemma":[0.25545555,0.00004691573,0.000035986275,0.00018013432,0.0001372775,0.00003284565,0.000031014384,0.00039133942,0.000009239919],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011784595,0.0013332716,0.8174526,0.000043583274,0.000024479104,0.0000050371327,0.002606036,0.000004407286,0.000003894812,0.002858261,0.008798824,0.16675174],"study_design_scores_gemma":[0.0030863562,0.0007351891,0.9547555,0.0025115449,0.00011454395,0.00007228515,0.0025537787,0.013592373,0.000045546178,0.0010358144,0.021375664,0.00012142466],"about_ca_topic_score_codex":0.00024336785,"about_ca_topic_score_gemma":0.000024135517,"teacher_disagreement_score":0.25229856,"about_ca_system_score_codex":0.00003835289,"about_ca_system_score_gemma":0.0014996406,"threshold_uncertainty_score":0.75081605},"labels":[],"label_agreement":null},{"id":"W2122398545","doi":"10.1016/j.amjsurg.2012.09.002","title":"Assessing clinical judgment using the Script Concordance test: the importance of using specialty-specific experts to develop the scoring key","year":2012,"lang":"en","type":"article","venue":"The American Journal of Surgery","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; McGill University","funders":"","keywords":"Concordance; Test (biology); Specialty; Reliability (semiconductor); Key (lock); Medicine; Scoring system; Cronbach's alpha; Medical education; Psychology; Medical physics; Computer science; Family medicine; Clinical psychology; Psychometrics; Surgery; Internal medicine","score_opus":0.21937843267960663,"score_gpt":0.4300570789545134,"score_spread":0.21067864627490676,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2122398545","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98591053,0.0031541516,0.0022126222,0.0064344686,0.0019672601,0.00020168521,0.0000018294712,0.000008593027,0.00010887836],"genre_scores_gemma":[0.98750365,0.00076149305,0.0031262152,0.0051745186,0.0033800148,0.0000030084132,3.5703025e-7,0.00003473201,0.000015986547],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99607,0.00065332063,0.0017624282,0.00018220936,0.00080120657,0.0005307884],"domain_scores_gemma":[0.93540514,0.060175195,0.0026478455,0.000811973,0.00061826594,0.00034156986],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.010399016,0.00021937465,0.00095945713,0.00007722584,0.0003852536,0.00008971617,0.00044456674,0.000037220783,0.000034369255],"category_scores_gemma":[0.03985114,0.00008896083,0.00039098575,0.0010335562,0.0012081057,0.00018521,0.00014850928,0.00065718096,0.0000036830777],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002471776,0.00020279778,0.9518226,0.00000499204,0.00011903241,0.00003917031,0.0011756405,0.00011705335,0.0010990866,0.00003445932,0.0028315492,0.042306475],"study_design_scores_gemma":[0.0003727822,0.00014529168,0.9788015,0.0030897683,0.00029164072,0.0011242074,0.010922642,0.00016334737,0.0008185538,0.000057569672,0.0040087346,0.00020395442],"about_ca_topic_score_codex":0.00006863825,"about_ca_topic_score_gemma":0.0000027138817,"teacher_disagreement_score":0.060664896,"about_ca_system_score_codex":0.0001483524,"about_ca_system_score_gemma":0.0007909929,"threshold_uncertainty_score":0.9682366},"labels":[],"label_agreement":null},{"id":"W2122612736","doi":"10.1111/j.1365-2923.2007.02924.x","title":"Scripts and clinical reasoning","year":2007,"lang":"en","type":"review","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":439,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Espace pour la vie","funders":"","keywords":"Scripting language; Meaning (existential); Cognition; Context (archaeology); Psychology; Computer science; Cognitive psychology; Cognitive science; Psychotherapist; Psychiatry","score_opus":0.09509583332470403,"score_gpt":0.5232068158400778,"score_spread":0.4281109825153737,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2122612736","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000066746354,0.99055207,0.00006035619,0.00086089474,0.0025471607,0.0004967402,0.0000022768843,0.0000714545,0.005342289],"genre_scores_gemma":[0.000008563355,0.9892156,0.0012007602,0.004422841,0.0026870614,0.000038649287,0.00030556903,0.000046146364,0.002074784],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9969958,0.00018849353,0.0012607228,0.0005672089,0.0006835472,0.00030423256],"domain_scores_gemma":[0.98585856,0.0119008375,0.00036563037,0.0005179044,0.00014903872,0.0012080221],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0026565017,0.00028336706,0.0016990878,0.00016568688,0.000055790177,0.000026623335,0.00013143387,0.0011251027,0.0006170376],"category_scores_gemma":[0.27751198,0.00020438367,0.00042349356,0.00028887615,0.000291287,0.000025166659,0.00007654256,0.0013272769,0.00024744455],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000007367378,0.0005598506,0.000634447,0.0017660022,0.00006907492,0.00003252974,0.000010485724,2.7728658e-10,4.525863e-10,0.00019808447,0.045357876,0.9513643],"study_design_scores_gemma":[0.0003273482,0.00013474161,0.0009549604,0.12262836,0.0012273036,0.00047530042,0.0000129729615,0.00000637364,9.196052e-9,0.000027323527,0.87404734,0.00015793969],"about_ca_topic_score_codex":0.000061954204,"about_ca_topic_score_gemma":0.000006295986,"teacher_disagreement_score":0.9512063,"about_ca_system_score_codex":0.000099501485,"about_ca_system_score_gemma":0.0063917446,"threshold_uncertainty_score":0.9992411},"labels":[],"label_agreement":null},{"id":"W2123203085","doi":"10.1002/chp.1340230303","title":"Stemming the tide: Cognitive aging theories and their implications for continuing education in the health professions","year":2003,"lang":"en","type":"article","venue":"Journal of Continuing Education in the Health Professions","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":33,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Cognition; Psychology; Health care; MEDLINE; Continuing medical education; Medical education; Population; Continuing education; Gerontology; Applied psychology; Medicine; Psychiatry; Political science; Environmental health","score_opus":0.04509581087301431,"score_gpt":0.46262531198446977,"score_spread":0.4175295011114555,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2123203085","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.75588256,0.007865072,0.0020393317,0.22937961,0.0013399455,0.003072392,0.000013071123,0.000012299156,0.0003957367],"genre_scores_gemma":[0.9857686,0.0009232999,0.0008225298,0.011150769,0.00040932946,0.00038045197,0.000020598529,0.000021318616,0.00050310174],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.99558514,0.002279569,0.001291113,0.00021186071,0.00020900198,0.00042328765],"domain_scores_gemma":[0.9619917,0.03503131,0.0016458462,0.00042609888,0.00069607684,0.00020893515],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0105530955,0.00018406332,0.00048151726,0.0002301174,0.00073089724,0.000073939584,0.00025337105,0.000091497146,0.000005497167],"category_scores_gemma":[0.030027868,0.00008668727,0.00011725403,0.000516924,0.00014797524,0.0001459599,0.000025284227,0.0010684278,9.070687e-7],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014690029,0.0016815399,0.13453372,0.00027292143,0.000030769876,4.6795222e-7,0.26928386,0.0000021660949,0.000031897653,0.04371516,0.013680094,0.5366205],"study_design_scores_gemma":[0.0013986088,0.00037554305,0.3358872,0.025951102,0.000061106024,0.00020756968,0.6109593,0.000011894069,0.000013277352,0.020339178,0.004678218,0.00011703968],"about_ca_topic_score_codex":0.00009523286,"about_ca_topic_score_gemma":0.00015279582,"teacher_disagreement_score":0.5365035,"about_ca_system_score_codex":0.00015779733,"about_ca_system_score_gemma":0.0049178684,"threshold_uncertainty_score":0.9781426},"labels":[],"label_agreement":null},{"id":"W2124322671","doi":"10.1111/j.1532-5415.2010.03136.x","title":"The Script Concordance Test as a Measure of Clinical Reasoning Skills in Geriatric Urinary Incontinence","year":2010,"lang":"en","type":"article","venue":"Journal of the American Geriatrics Society","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":28,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Geriatrics; Medicine; Concordance; Cronbach's alpha; Test (biology); Urinary incontinence; Family medicine; Gerontology; Psychometrics; Internal medicine; Clinical psychology; Urology; Psychiatry","score_opus":0.013723196616024698,"score_gpt":0.34402942573161494,"score_spread":0.3303062291155902,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2124322671","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9916061,0.0010279558,0.00006147416,0.004817723,0.0017705217,0.00024089353,0.000006465683,0.000012100682,0.00045674347],"genre_scores_gemma":[0.9863998,0.0036457966,0.007132292,0.0012734313,0.0012451103,0.000004370121,4.9291623e-7,0.000027996615,0.00027067153],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9962543,0.00028258265,0.0019304636,0.00025675728,0.00088175375,0.00039413868],"domain_scores_gemma":[0.961949,0.030808078,0.0048680576,0.0009360931,0.0010339052,0.00040489735],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.006117411,0.00021390062,0.0010766673,0.00005819189,0.0001697295,0.000039346796,0.0007392053,0.00015699059,0.000013056757],"category_scores_gemma":[0.17498821,0.000118595475,0.0012309995,0.0014381037,0.0010904818,0.00008415122,0.0002082802,0.0021665033,0.0000050571753],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00040953394,0.0010712981,0.9585322,0.000016723578,0.00011854477,0.0000502166,0.0003511402,0.000012918918,0.0014937134,0.000060578608,0.009509788,0.028373307],"study_design_scores_gemma":[0.0026697188,0.0014638382,0.9887308,0.0005802498,0.0002743277,0.00063195865,0.0008252135,0.00019765327,0.000189511,0.0003453731,0.0039283205,0.00016305884],"about_ca_topic_score_codex":0.00024705412,"about_ca_topic_score_gemma":0.000019049874,"teacher_disagreement_score":0.1688708,"about_ca_system_score_codex":0.0000749338,"about_ca_system_score_gemma":0.001394837,"threshold_uncertainty_score":0.94124997},"labels":[],"label_agreement":null},{"id":"W2124446425","doi":"10.1111/medu.12670","title":"When I say… cognitive debiasing","year":2015,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":31,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Debiasing; Nova scotia; Suite; Queen (butterfly); Citation; Library science; Psychology; Computer science; History; Social psychology; Archaeology","score_opus":0.04335094787769664,"score_gpt":0.40803961145868417,"score_spread":0.3646886635809875,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2124446425","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8746924,0.0010178998,0.0018375868,0.06929443,0.0018830713,0.0002810043,0.000001503499,0.000115927185,0.050876193],"genre_scores_gemma":[0.966614,0.00002956755,0.001192039,0.0284936,0.0009234075,0.000025893305,0.00010936832,0.0000121986,0.002599903],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9988993,0.000048251684,0.00020123294,0.00016842199,0.00053665956,0.00014614157],"domain_scores_gemma":[0.99620205,0.002591705,0.00004945179,0.00013701117,0.00028882825,0.0007309615],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0005370046,0.00006970531,0.00015705687,0.000043126747,0.000024972991,0.000012010436,0.000046158948,0.00012035693,0.00088888087],"category_scores_gemma":[0.2962928,0.000055598855,0.000042561052,0.00009683197,0.0000914472,0.000032843243,0.000019882094,0.00019726335,0.00041690882],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010623923,0.0011979587,0.047486737,0.000012920371,0.000022430517,0.000021034939,0.001289887,9.258463e-8,0.0000031662123,0.0003206391,0.32841328,0.62112564],"study_design_scores_gemma":[0.020413272,0.002849116,0.51430863,0.031195892,0.0012964988,0.0015655032,0.012832198,0.002587914,0.0011213619,0.029923424,0.3807389,0.0011672612],"about_ca_topic_score_codex":0.00017493733,"about_ca_topic_score_gemma":0.000009090807,"teacher_disagreement_score":0.61995834,"about_ca_system_score_codex":0.000066832596,"about_ca_system_score_gemma":0.003041721,"threshold_uncertainty_score":0.9732622},"labels":[],"label_agreement":null},{"id":"W2124535217","doi":"10.1017/s1481803500011829","title":"Does outcome feedback make you a better emergency physician? A systematic review and research framework proposal","year":2009,"lang":"en","type":"review","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute of Population and Public Health; University of Ottawa; Montfort Hospital; Ontario Centre of Excellence for Child and Youth Mental Health; Children's Hospital of Eastern Ontario","funders":"","keywords":"CINAHL; PsycINFO; Medicine; MEDLINE; Emergency department; Outcome (game theory); Medical emergency; Cochrane Library; Intervention (counseling); Health care; Randomized controlled trial; Emergency medicine; Family medicine; Psychological intervention; Nursing; Surgery","score_opus":0.17371109972071197,"score_gpt":0.4836649375808137,"score_spread":0.3099538378601017,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2124535217","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000014189811,0.97431827,0.000018584098,0.015457054,0.006101286,0.0029562272,0.000044884604,0.000014425278,0.0010750637],"genre_scores_gemma":[0.000015800486,0.98937696,0.00026275724,0.0014308878,0.004427277,0.00007558853,0.000045033412,0.000111798574,0.0042539234],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9902892,0.0009922119,0.0056310287,0.0006431227,0.001351765,0.0010926406],"domain_scores_gemma":[0.9892519,0.002783431,0.0021334528,0.00113123,0.0016236764,0.0030762781],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.006050045,0.0008093487,0.00832542,0.0010584488,0.00023027469,0.000016676764,0.00070700207,0.00063348515,0.0033331418],"category_scores_gemma":[0.107347846,0.00039528336,0.0011790198,0.0017816176,0.00034571526,0.00007489388,0.000045828012,0.0033606181,0.00011499629],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000010817995,0.00011110251,0.0005194985,0.8009622,0.0009822214,0.001011129,0.00024117023,2.2163533e-8,9.215893e-8,0.00027437345,0.09717811,0.09870925],"study_design_scores_gemma":[0.0002939021,0.0008011107,0.000094556184,0.6883934,0.013853223,0.00022521534,0.00007982177,3.0803696e-7,2.2609004e-8,0.0012213053,0.2947228,0.0003143206],"about_ca_topic_score_codex":0.00072256906,"about_ca_topic_score_gemma":0.00076607295,"teacher_disagreement_score":0.19754471,"about_ca_system_score_codex":0.00036228765,"about_ca_system_score_gemma":0.0030318622,"threshold_uncertainty_score":0.9998499},"labels":[],"label_agreement":null},{"id":"W2125025318","doi":"10.1177/1049732314553594","title":"Seeing in Different Ways","year":2014,"lang":"en","type":"article","venue":"Qualitative Health Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":88,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"Canadian Institutes of Health Research","keywords":"Context (archaeology); Domain (mathematical analysis); Process (computing); Visual methods; Value (mathematics); Psychology; Computer science; Data science; Cognitive science","score_opus":0.5723700287326392,"score_gpt":0.6481113801961634,"score_spread":0.07574135146352423,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2125025318","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9373594,0.0005555326,0.0017676065,0.04781162,0.00010850474,0.00066643965,0.000005906188,0.000046753685,0.011678221],"genre_scores_gemma":[0.99592686,0.00016848241,0.0010298566,0.0017977268,0.000194268,0.000071750634,0.000022790826,0.000017622186,0.0007706423],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9947643,0.0028126277,0.00049602694,0.0003406093,0.00084056077,0.00074583833],"domain_scores_gemma":[0.8958942,0.10245675,0.00008009486,0.00041982284,0.00035500957,0.0007940803],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.013938574,0.0001016483,0.00045288415,0.00027898402,0.00013072649,0.00001558573,0.00008642988,0.000078007564,0.00012469495],"category_scores_gemma":[0.14996336,0.00007339693,0.000058212405,0.00045948508,0.00020076105,0.000023689629,0.00006186795,0.0008873178,0.00025290213],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014470507,0.0035968183,0.10426934,0.0016009819,0.00007299528,0.00009499297,0.170826,0.000010328363,0.00025330085,0.24159412,0.0666624,0.40957168],"study_design_scores_gemma":[0.009825929,0.009330907,0.8258813,0.012561567,0.000010118767,0.000015169735,0.03524429,0.0030287108,0.00037564785,0.075015225,0.028292175,0.00041897842],"about_ca_topic_score_codex":0.0009794565,"about_ca_topic_score_gemma":0.00012617407,"teacher_disagreement_score":0.7216119,"about_ca_system_score_codex":0.00028853372,"about_ca_system_score_gemma":0.00032199477,"threshold_uncertainty_score":0.85719687},"labels":[],"label_agreement":null},{"id":"W2125511086","doi":"10.12927/hcq.2012.22841","title":"Perspectives on Diagnostic Failure and Patient Safety","year":2012,"lang":"en","type":"review","venue":"Healthcare Quarterly","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":50,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Patient safety; Best practice; Medicine; Medical emergency; Health care; Political science","score_opus":0.04378638011307922,"score_gpt":0.37935787678953603,"score_spread":0.3355714966764568,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2125511086","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00010473502,0.9936156,0.000004492567,0.003645734,0.00043149685,0.001509257,0.0001495606,0.0001056191,0.00043344838],"genre_scores_gemma":[0.0035995233,0.99432814,0.00017974628,0.0006031595,0.0007570989,0.00017685298,0.00018866424,0.00008382003,0.000083006744],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99711895,0.0003341964,0.0008647862,0.0007078916,0.00039148354,0.000582714],"domain_scores_gemma":[0.97819394,0.019805029,0.0003292825,0.00064280105,0.00013668527,0.0008922693],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00025181408,0.00052549277,0.0020773516,0.00017086754,0.00012445224,0.000030538737,0.00008423297,0.00058277464,0.00011637277],"category_scores_gemma":[0.0073872767,0.0003782975,0.0004311177,0.00020111454,0.000117500895,0.00003672694,0.000019876197,0.0010252506,0.00039024925],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00002829109,0.00026978122,0.000070801645,0.005863504,0.00008200024,0.0001046137,0.0021050326,4.9812177e-9,1.8899977e-9,0.0010782442,0.0010214075,0.9893763],"study_design_scores_gemma":[0.00041171905,0.004562551,0.0005834591,0.08369633,0.00085801043,0.00034717482,0.002061772,2.0412713e-7,1.2178574e-8,0.000040944626,0.907108,0.00032982763],"about_ca_topic_score_codex":0.00014394033,"about_ca_topic_score_gemma":0.00003772472,"teacher_disagreement_score":0.9890465,"about_ca_system_score_codex":0.0002676407,"about_ca_system_score_gemma":0.00050931366,"threshold_uncertainty_score":0.9998669},"labels":[],"label_agreement":null},{"id":"W2125611482","doi":"10.1007/s10459-014-9517-5","title":"The Bias in researching cognitive bias","year":2014,"lang":"en","type":"editorial","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Cognitive bias; Psychology; Cognition; Epistemology; Psychoanalysis; Philosophy; Psychiatry","score_opus":0.07944655223948864,"score_gpt":0.5212954788938702,"score_spread":0.4418489266543815,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2125611482","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.011320535,0.039893307,0.00003604418,0.009491783,0.93217283,0.001202668,0.0000147899755,0.000030388626,0.005837649],"genre_scores_gemma":[0.08671021,0.15069208,0.0012547402,0.0020987454,0.75627387,0.00058086956,0.00033648437,0.00004505496,0.0020079403],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9960305,0.0006263865,0.00088795053,0.00061575265,0.0012179984,0.00062140037],"domain_scores_gemma":[0.63311714,0.3635218,0.0013358576,0.0006893472,0.000726235,0.00060959184],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01168183,0.00017440053,0.0004447393,0.0003611338,0.00037630566,0.000091109956,0.00032920184,0.0002124162,0.000010210261],"category_scores_gemma":[0.44671527,0.00011565026,0.000054109783,0.001063184,0.0007030775,0.0001769635,0.000056457757,0.001255386,0.000030895535],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013169812,0.0003207484,0.039811146,0.00034913173,0.0000028951183,0.0000023273444,0.0007205841,0.00007569329,6.3892045e-8,0.0005420323,0.41733402,0.5407097],"study_design_scores_gemma":[0.0006899163,0.00088894326,0.01015606,0.025639914,0.000007645132,0.0000027645888,0.0017479897,0.000283956,0.0000015400759,0.005883527,0.95452154,0.00017621153],"about_ca_topic_score_codex":0.0032055213,"about_ca_topic_score_gemma":0.0039994707,"teacher_disagreement_score":0.5405334,"about_ca_system_score_codex":0.0003785707,"about_ca_system_score_gemma":0.009511022,"threshold_uncertainty_score":0.9961041},"labels":[],"label_agreement":null},{"id":"W2125983962","doi":"10.3109/09593985.2012.676941","title":"Physiotherapy as bricolage: Theorizing expert practice","year":2012,"lang":"en","type":"article","venue":"Physiotherapy Theory and Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":73,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Situated; Bricolage; Clinical Practice; Tacit knowledge; Relation (database); Epistemology; Psychology; Medicine; Physical therapy; Knowledge management; Computer science; Artificial intelligence","score_opus":0.024126141750067123,"score_gpt":0.4379625102722246,"score_spread":0.4138363685221575,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2125983962","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4511507,0.11453762,0.013064212,0.05560004,0.0033080976,0.0029580994,0.0000224806,0.0010867065,0.35827208],"genre_scores_gemma":[0.88749456,0.016389903,0.0065829074,0.08465804,0.0024628488,0.00006631421,0.000013277841,0.00010523112,0.002226937],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.9967162,0.0015675244,0.00034073295,0.000412924,0.00041998157,0.0005426502],"domain_scores_gemma":[0.91576123,0.08277109,0.0002891395,0.00053795765,0.00019161204,0.00044898016],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0041001355,0.00031998925,0.000453768,0.00006501614,0.00026689185,0.00009754862,0.00011943301,0.0001636066,0.000840686],"category_scores_gemma":[0.06861407,0.00025292463,0.00015793605,0.00027256086,0.00020481434,0.0013439108,0.000042826683,0.0004490971,0.00022415005],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.04812938,0.008293371,0.00070206355,0.00009249953,0.0014777028,0.00010731814,0.020496152,0.00000423553,0.1613212,0.41279477,0.007839433,0.3387419],"study_design_scores_gemma":[0.0070861513,0.0021553498,0.004805001,0.0010617669,0.0005669517,0.0020304185,0.010113889,0.000024335855,0.013340497,0.02428174,0.93356234,0.0009715694],"about_ca_topic_score_codex":0.00013763791,"about_ca_topic_score_gemma":2.625114e-7,"teacher_disagreement_score":0.9257229,"about_ca_system_score_codex":0.00004609526,"about_ca_system_score_gemma":0.00010045088,"threshold_uncertainty_score":0.9999923},"labels":[],"label_agreement":null},{"id":"W2126257109","doi":"10.1503/cmaj.1031666","title":"Tips for learners of evidence-based medicine: 5. The effect of spectrum of disease on the performance of diagnostic tests","year":2005,"lang":"en","type":"review","venue":"Canadian Medical Association Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":52,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Institute of Nutrition, Metabolism and Diabetes","funders":"University of California, San Francisco; McMaster University; U.S. Department of Veterans Affairs","keywords":"Diagnostic test; Test (biology); Medicine; Disease; Diagnostic accuracy; Medical physics; Intensive care medicine; Evidence-based medicine; Clinical Practice; Alternative medicine; Pathology; Physical therapy; Pediatrics; Radiology","score_opus":0.05723612649017828,"score_gpt":0.37447914724015235,"score_spread":0.31724302074997407,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2126257109","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.016799644,0.9404493,0.000010370995,0.040233683,0.00064866466,0.0015505456,0.00014687586,0.000005550333,0.00015538943],"genre_scores_gemma":[0.0964731,0.9011906,0.000009699074,0.0009132254,0.0012107027,0.000046335226,0.000023064902,0.000028291846,0.000105009254],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.995265,0.0007413699,0.001575393,0.00017981138,0.0018880485,0.0003503749],"domain_scores_gemma":[0.81266046,0.18404578,0.001808011,0.00030496993,0.00032811696,0.0008526827],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0075728726,0.00025370464,0.0017846392,0.00023585372,0.00008702273,0.0000057858715,0.00044702005,0.00033952916,0.0007074457],"category_scores_gemma":[0.5699195,0.00010941363,0.0007498839,0.00042696646,0.00041020752,0.000025944046,0.000014188594,0.001111617,0.00000715486],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020105633,0.00009813083,0.025200788,0.0056588524,0.0005027426,0.00003138556,0.00006108647,0.00002879194,1.8213088e-7,0.00008421282,0.026495142,0.94163764],"study_design_scores_gemma":[0.0035336018,0.0071703373,0.014382951,0.57180846,0.0069212522,0.00008034289,0.000062520616,0.00044196125,0.00002744848,0.000034254204,0.39524558,0.00029126648],"about_ca_topic_score_codex":0.0004274987,"about_ca_topic_score_gemma":0.00034653867,"teacher_disagreement_score":0.94134635,"about_ca_system_score_codex":0.00044748886,"about_ca_system_score_gemma":0.006981184,"threshold_uncertainty_score":0.9986483},"labels":[],"label_agreement":null},{"id":"W2127204700","doi":"10.1111/j.1365-2923.2007.02911.x","title":"Science is fundamental: the role of biomedical knowledge in clinical reasoning","year":2007,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":268,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; University of Toronto","funders":"","keywords":"Medical knowledge; Knowledge base; Psychology; Engineering ethics; Knowledge management; Computer science; Medical education; Medicine; Artificial intelligence","score_opus":0.021740025221144533,"score_gpt":0.4489767582466433,"score_spread":0.42723673302549875,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2127204700","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9768386,0.0012043526,0.00009467493,0.007118377,0.0010646874,0.00020259304,8.430066e-7,0.000020907817,0.013455004],"genre_scores_gemma":[0.9944814,0.00012771404,0.0006154207,0.0039171195,0.00055809517,0.000008102189,0.000011145901,0.000008933815,0.00027208362],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99749875,0.00006793184,0.0008541812,0.0003082531,0.00094713474,0.0003237442],"domain_scores_gemma":[0.99096566,0.0074736914,0.00018261392,0.0004604595,0.00023589851,0.00068167597],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.007994602,0.000091686365,0.00027370692,0.00023733768,0.000067942776,0.000010424063,0.00026805437,0.00018675705,0.0005479628],"category_scores_gemma":[0.12134803,0.000059158276,0.00009704097,0.0012374327,0.001658445,0.00004217126,0.00008998892,0.00050453725,0.0000685801],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007642326,0.0033311574,0.35320398,0.0000132186005,0.000009119413,0.0000050395092,0.0010223163,2.0068539e-8,0.00012455892,0.002734964,0.0055067483,0.63397247],"study_design_scores_gemma":[0.001242763,0.00044374235,0.92685,0.002376984,0.000044717406,0.00006038516,0.0027988257,0.0007863536,0.0007247215,0.000633815,0.06393726,0.000100423975],"about_ca_topic_score_codex":0.00020370736,"about_ca_topic_score_gemma":0.000031174233,"teacher_disagreement_score":0.63387203,"about_ca_system_score_codex":0.00011502475,"about_ca_system_score_gemma":0.0057316786,"threshold_uncertainty_score":0.9999049},"labels":[],"label_agreement":null},{"id":"W2128894270","doi":"10.1186/1472-6947-5-18","title":"Online clinical reasoning assessment with the Script Concordance test: a feasibility study","year":2005,"lang":"en","type":"article","venue":"BMC Medical Informatics and Decision Making","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":31,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Test (biology); Concordance; Context (archaeology); Medicine; Health informatics; Web application; Summative assessment; Certification; Medical physics; Computer science; Medical education; World Wide Web; Formative assessment; Psychology; Pathology; Public health; Internal medicine","score_opus":0.08053502247139271,"score_gpt":0.46259963379427743,"score_spread":0.3820646113228847,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2128894270","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8726479,0.00014330194,0.12479991,0.00048027345,0.0001292807,0.00065062166,0.0000063039324,0.000057121542,0.0010853221],"genre_scores_gemma":[0.7535797,0.000101970116,0.24149728,0.0044393274,0.00031423845,0.000014428228,0.00000661836,0.000012046601,0.00003442434],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9962333,0.0001144591,0.0014766267,0.000276172,0.0015766347,0.00032284707],"domain_scores_gemma":[0.95032877,0.047796678,0.0003886437,0.0006731781,0.0002349976,0.000577708],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004963698,0.00021672006,0.0006343738,0.000055871504,0.00021659175,0.000112978596,0.00023477475,0.00016223159,0.00015776877],"category_scores_gemma":[0.06336119,0.00010953748,0.00011793224,0.00025133218,0.00033560852,0.00012258264,0.00025726395,0.0008439587,0.000020228148],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002953535,0.0017620405,0.57945234,0.000025491208,0.000030945932,0.000031879685,0.0002564284,0.00004526394,7.236034e-8,0.000113497765,0.002185173,0.4158015],"study_design_scores_gemma":[0.0065176226,0.0019060617,0.6284088,0.0038940911,0.00015862116,0.00020892899,0.0033647052,0.35383376,2.6254554e-7,0.00013684006,0.001390846,0.00017950413],"about_ca_topic_score_codex":0.00000810013,"about_ca_topic_score_gemma":0.00019905304,"teacher_disagreement_score":0.415622,"about_ca_system_score_codex":0.000058944144,"about_ca_system_score_gemma":0.00070225634,"threshold_uncertainty_score":0.9445285},"labels":[],"label_agreement":null},{"id":"W2129887897","doi":"10.1037/0278-7393.32.6.1416","title":"Producing biased diagnoses with unambiguous stimuli: The importance of feature instantiations.","year":2006,"lang":"en","type":"article","venue":"Journal of Experimental Psychology Learning Memory and Cognition","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Feature (linguistics); Medical diagnosis; Heuristic; Psychology; Limit (mathematics); Computer science; Cognitive psychology; Artificial intelligence; Natural language processing; Mathematics; Linguistics; Medicine","score_opus":0.02582839977103753,"score_gpt":0.3513290148609789,"score_spread":0.3255006150899414,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2129887897","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9923221,0.0032764303,0.00011055205,0.0016123109,0.00013521152,0.000156142,0.0000014135765,0.000010814833,0.0023750502],"genre_scores_gemma":[0.9981507,0.0001424971,0.0005062203,0.00077748194,0.00023788886,0.0000075016073,0.000014917569,0.000010774975,0.000151984],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.999097,0.000115022835,0.00035126237,0.00014301983,0.00017990518,0.00011381959],"domain_scores_gemma":[0.9980583,0.0011148984,0.00050024415,0.00009590272,0.00018062063,0.000050000897],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00037926668,0.000097236836,0.00025678583,0.000070094946,0.00012601366,0.000008729316,0.000039568968,0.000076895485,0.000051844603],"category_scores_gemma":[0.0019191243,0.000059381648,0.00006066439,0.00013365952,0.00019985523,0.00006869255,0.00000961631,0.00042827323,0.0000014990607],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.005604438,0.002458631,0.849039,0.00005790531,0.00033465546,0.0004456288,0.0019731463,0.00034714618,0.12215933,0.00009978212,0.006275008,0.011205295],"study_design_scores_gemma":[0.0075589586,0.0044088475,0.94415706,0.0019257801,0.0003452046,0.0016872411,0.0025316838,0.000028129234,0.0366283,0.00021815937,0.00036326676,0.00014734562],"about_ca_topic_score_codex":0.00000806764,"about_ca_topic_score_gemma":0.000002934641,"teacher_disagreement_score":0.095118046,"about_ca_system_score_codex":0.000014458351,"about_ca_system_score_gemma":0.000050924267,"threshold_uncertainty_score":0.24215132},"labels":[],"label_agreement":null},{"id":"W2130011931","doi":"10.2310/8000.2013.130860","title":"Ineffectiveness of cognitive forcing strategies to reduce biases in diagnostic reasoning: acontrolled trial","year":2014,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":109,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Medical diagnosis; Cognitive bias; Satisficing; Cognition; Psychological intervention; Test (biology); Physical therapy; Clinical psychology; Psychiatry; Pathology","score_opus":0.08481384238368306,"score_gpt":0.40469010399264094,"score_spread":0.3198762616089579,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2130011931","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9883741,0.0015054726,0.00076685275,0.0019091355,0.0038272713,0.00070014934,0.0000084903195,0.000005648903,0.00290286],"genre_scores_gemma":[0.9977306,0.00020982532,0.000073336436,0.0001786542,0.0016487772,0.000022097618,0.000010131127,0.000024077632,0.00010247569],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9972434,0.00024950717,0.0014929536,0.00021645402,0.00035833017,0.00043933964],"domain_scores_gemma":[0.9632419,0.033011112,0.00072716817,0.00025118355,0.0010177627,0.0017508728],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0023867597,0.00020897633,0.0011312749,0.00070101686,0.00005147557,0.00000521548,0.00015782345,0.000117971176,0.0007452581],"category_scores_gemma":[0.5927861,0.00015893154,0.00018811676,0.00069864717,0.00013809674,0.00008971186,0.000010427376,0.00036640916,0.000004915458],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.12307981,0.002100299,0.6958953,0.0007365085,0.0019760039,0.002967144,0.02225042,0.0055205915,0.0059348843,0.004406714,0.04730987,0.08782244],"study_design_scores_gemma":[0.23384231,0.03691526,0.6127683,0.09815749,0.0021254986,0.00016087345,0.009122286,0.0005933242,0.001321218,0.0021724834,0.0020836678,0.00073730183],"about_ca_topic_score_codex":0.009389728,"about_ca_topic_score_gemma":0.012178229,"teacher_disagreement_score":0.5903993,"about_ca_system_score_codex":0.00011675551,"about_ca_system_score_gemma":0.0017161836,"threshold_uncertainty_score":0.9972068},"labels":[],"label_agreement":null},{"id":"W2130902245","doi":"10.1542/hpeds.2012-0058","title":"When and How Pediatric History and Physical Diagnosis Are Taught in Medical School: A Survey of Pediatric Clerkship Directors","year":2013,"lang":"en","type":"article","venue":"Hospital Pediatrics","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Preceptor; Medicine; Curriculum; Medical education; Demographics; Physical exam; Medical school; Family medicine; Clinical clerkship; Medical history; Surgery; Psychology; Pedagogy","score_opus":0.023674360833917874,"score_gpt":0.27232839729168967,"score_spread":0.2486540364577718,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2130902245","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9777612,0.010997039,0.0000056515923,0.010077579,0.00039850053,0.000563354,0.00003033747,0.00006853559,0.00009778505],"genre_scores_gemma":[0.98439443,0.011753041,0.00019879466,0.0012434381,0.0021036249,0.00010878253,0.000035171943,0.00005223188,0.0001104843],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9971065,0.00021875996,0.0005349044,0.00066461123,0.00094146567,0.0005337403],"domain_scores_gemma":[0.98272777,0.01500491,0.00036401517,0.00037344938,0.00033730126,0.0011925773],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0008457619,0.00034957466,0.00089091517,0.0004224081,0.000037958576,0.000038609516,0.00018253867,0.00046112438,0.0002353932],"category_scores_gemma":[0.12816094,0.00030535972,0.00013250585,0.0007045874,0.0002177944,0.00020601846,0.00020118392,0.00089353387,0.00003527648],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000022938475,0.00065917405,0.9000777,0.00017970339,0.00001170895,0.00009384168,0.00022621237,1.7451248e-7,2.5731828e-7,0.0000070113074,0.096378565,0.0023427026],"study_design_scores_gemma":[0.0018874115,0.0005656599,0.9963459,0.00007319569,0.00029010623,0.000008669862,0.00007109341,0.00010222833,0.0000017997464,0.00012978645,0.0001959596,0.00032821074],"about_ca_topic_score_codex":0.0014272284,"about_ca_topic_score_gemma":0.00007893201,"teacher_disagreement_score":0.12731518,"about_ca_system_score_codex":0.00014238115,"about_ca_system_score_gemma":0.00049459626,"threshold_uncertainty_score":0.99993986},"labels":[],"label_agreement":null},{"id":"W2131996163","doi":"10.36834/cmej.36531","title":"The Reliability and Validity of a Paediatric Script Concordance Test with Medical Students, Paediatric Residents and Experienced Paediatricians","year":2010,"lang":"en","type":"article","venue":"Canadian Medical Education Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Université de Montréal; University of Calgary","funders":"Royal College of Physicians and Surgeons of Canada","keywords":"Concordance; Test (biology); Grading (engineering); Reliability (semiconductor); Medicine; Multiple choice; Medical education; Family medicine; Psychology; Pediatrics; Significant difference; Internal medicine","score_opus":0.009798068903059506,"score_gpt":0.3212783576356686,"score_spread":0.3114802887326091,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2131996163","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9621051,0.0017551976,0.000018731516,0.034078453,0.0015180305,0.00026905848,0.000008394942,0.00001172343,0.00023532305],"genre_scores_gemma":[0.9871163,0.0042314813,0.00035604744,0.006500172,0.001598631,0.000026492817,0.00000576027,0.00001715755,0.00014800878],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9962957,0.0001690889,0.00070547214,0.00032283305,0.0020673021,0.00043959584],"domain_scores_gemma":[0.9621365,0.018448483,0.00050366006,0.00062449643,0.0010726341,0.017214244],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0039576073,0.00017517823,0.00035883844,0.00018050404,0.0004067881,0.00011761922,0.0004134638,0.00031675477,0.0011242946],"category_scores_gemma":[0.5463667,0.00010949727,0.00006457268,0.00047084098,0.0010014703,0.00008670689,0.000062075895,0.0015287515,0.000004775037],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004885594,0.00028478677,0.91645193,0.000022679309,0.000016428536,0.00009406854,0.0004397578,1.0148912e-7,0.0000026081618,0.000080467886,0.025083987,0.057474345],"study_design_scores_gemma":[0.0021769258,0.0005851912,0.98226315,0.0005521825,0.00015458382,0.0019485231,0.0014174905,0.000095128606,0.000009343182,0.0003386907,0.010280202,0.00017858729],"about_ca_topic_score_codex":0.004834711,"about_ca_topic_score_gemma":0.014620302,"teacher_disagreement_score":0.54240906,"about_ca_system_score_codex":0.00010001578,"about_ca_system_score_gemma":0.029830065,"threshold_uncertainty_score":0.9997888},"labels":[],"label_agreement":null},{"id":"W2133780453","doi":"10.1177/1553350606291042","title":"A Novel Method of Assessing Clinical Reasoning in Surgical Residents","year":2006,"lang":"en","type":"review","venue":"Surgical Innovation","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":41,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Montreal General Hospital","funders":"","keywords":"Summative assessment; Concordance; Formative assessment; Medicine; Test (biology); Medical education; Clinical judgment; Medical physics; Psychology; Mathematics education","score_opus":0.19508254442248776,"score_gpt":0.5425451436749902,"score_spread":0.34746259925250245,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2133780453","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0016585501,0.97809505,0.0029372785,0.00012044265,0.0007379308,0.0011277738,0.000041512965,0.00010535141,0.0151761025],"genre_scores_gemma":[0.00028660338,0.9697877,0.026424794,0.00005323802,0.0010867871,0.000061812694,0.0013609255,0.00009430525,0.00084383436],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9924774,0.0006877351,0.004757136,0.0008492183,0.0007726537,0.0004558678],"domain_scores_gemma":[0.9646915,0.032260675,0.0017389972,0.0006114492,0.00053948123,0.00015785018],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0069093835,0.00046247806,0.004068449,0.0008155972,0.000040538438,0.000049187824,0.00019462232,0.001327335,0.00009504971],"category_scores_gemma":[0.043177027,0.00036173957,0.00079759187,0.003779564,0.00018014744,0.00010213447,0.00013721063,0.0016984416,0.000025483509],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000816158,0.001025208,0.0034385931,0.0022361518,0.00009386327,0.000680872,0.0000037336863,0.0000025376064,4.834116e-7,0.004544831,0.00035148291,0.9875406],"study_design_scores_gemma":[0.0044113854,0.00022363696,0.0075482437,0.14465182,0.00085642136,0.0014513697,0.000007283629,0.00014275053,0.000001958227,0.00021286255,0.8399953,0.00049693347],"about_ca_topic_score_codex":0.00030839813,"about_ca_topic_score_gemma":0.000006040764,"teacher_disagreement_score":0.9870437,"about_ca_system_score_codex":0.0001873465,"about_ca_system_score_gemma":0.00095695,"threshold_uncertainty_score":0.9999691},"labels":[],"label_agreement":null},{"id":"W2135415725","doi":"10.1111/j.1553-2712.2011.01182.x","title":"Adjusting Our Lens: Can Developmental Differences in Diagnostic Reasoning Be Harnessed to Improve Health Professional and Trainee Assessment?","year":2011,"lang":"en","type":"article","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":28,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"National Center for Advancing Translational Sciences; National Center for Research Resources","keywords":"Vignette; Medicine; Reliability (semiconductor); Test (biology); Medical education; Social psychology; Psychology","score_opus":0.12721178930794225,"score_gpt":0.4117235342177596,"score_spread":0.28451174490981734,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2135415725","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9687026,0.0019010233,0.000105058636,0.025289344,0.0016539817,0.00094478694,0.00001797308,0.0000846124,0.0013006093],"genre_scores_gemma":[0.9879996,0.004739877,0.0022073507,0.0034726413,0.00056823203,0.00018438447,0.000062077204,0.00003896065,0.0007268606],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9962698,0.00014834221,0.0015828864,0.0006901396,0.0005907201,0.0007180681],"domain_scores_gemma":[0.99568677,0.0029146601,0.0003696414,0.00019894655,0.000106701904,0.0007232642],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0017669272,0.00034968083,0.0008580975,0.00025649756,0.00014342483,9.92217e-7,0.00018944162,0.00025322763,0.00039674283],"category_scores_gemma":[0.076111235,0.00026280322,0.000051318628,0.0005215379,0.00009028736,0.00008986641,0.00013228717,0.0011772271,0.000006508018],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000084284446,0.00017394716,0.97158366,0.00014776259,0.000044476536,0.000040488132,0.008697719,2.182191e-7,0.00077041227,0.00036401866,0.008088633,0.010004385],"study_design_scores_gemma":[0.002361071,0.00092475675,0.9778331,0.00839622,0.000088406916,0.000043043183,0.0094129825,0.000117788855,0.00007376795,0.00022774632,0.000236465,0.00028464853],"about_ca_topic_score_codex":0.0018767146,"about_ca_topic_score_gemma":0.0002288556,"teacher_disagreement_score":0.07434431,"about_ca_system_score_codex":0.0001737915,"about_ca_system_score_gemma":0.0006354968,"threshold_uncertainty_score":0.9999824},"labels":[],"label_agreement":null},{"id":"W2135428545","doi":"10.1111/j.1532-5415.2012.04152.x","title":"Assessment of Undergraduate Clinical Reasoning in Geriatric Medicine: Application of a Script Concordance Test","year":2012,"lang":"en","type":"article","venue":"Journal of the American Geriatrics Society","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":25,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; McGill University","funders":"","keywords":"Medicine; Concordance; Cronbach's alpha; Competence (human resources); Geriatrics; Test (biology); Medical education; Construct validity; Educational measurement; MEDLINE; Psychometrics; Family medicine; Clinical psychology; Psychology; Internal medicine; Psychiatry","score_opus":0.023912079243470226,"score_gpt":0.38966012628006963,"score_spread":0.3657480470365994,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2135428545","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9880083,0.0013286081,0.0051846774,0.004185514,0.0006468097,0.0002592948,0.0000066238185,0.000007516238,0.00037261497],"genre_scores_gemma":[0.97716665,0.0033037981,0.018038802,0.0006267002,0.00080436043,0.0000039915626,0.0000016405368,0.000019028403,0.000035040386],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9968062,0.0002285859,0.0018277287,0.00015574711,0.00070081087,0.0002809175],"domain_scores_gemma":[0.9845908,0.009859881,0.0043874127,0.0004385773,0.00047633226,0.00024699265],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004117231,0.00015376594,0.0011359567,0.000087835484,0.000037623304,0.0000043417485,0.00029171465,0.00008725793,0.0000058871187],"category_scores_gemma":[0.026115518,0.000098403,0.00066904677,0.0014610989,0.00064196816,0.00007400852,0.000092489965,0.0006572672,7.82836e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000069778194,0.00077534834,0.9869017,0.00004552639,0.00009667947,0.00000148357,0.00018219216,0.000044177665,0.0013808539,0.0002600456,0.003654251,0.00658799],"study_design_scores_gemma":[0.0021403239,0.0009571281,0.99352527,0.00055476226,0.00043151912,0.00006196933,0.00055860047,0.000882266,0.000086883716,0.00024773247,0.00046243152,0.000091130525],"about_ca_topic_score_codex":0.00023969174,"about_ca_topic_score_gemma":0.0000022353233,"teacher_disagreement_score":0.021998288,"about_ca_system_score_codex":0.00019067508,"about_ca_system_score_gemma":0.00052968547,"threshold_uncertainty_score":0.9820879},"labels":[],"label_agreement":null},{"id":"W2136263013","doi":"10.1111/j.1365-2923.2007.02912.x","title":"Cognitive metaphors of expertise and knowledge: prospects and limitations for medical education","year":2007,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":147,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre","funders":"","keywords":"Cognition; Scope (computer science); Context (archaeology); Process (computing); Resource (disambiguation); Frame (networking); Knowledge management; Cognitive science; Engineering ethics; Computer science; Psychology; Management science; Epistemology; Engineering","score_opus":0.030825513160696063,"score_gpt":0.421177705394035,"score_spread":0.39035219223333895,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2136263013","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97282386,0.0067332657,0.0025575424,0.0120073985,0.0009628635,0.00080451695,0.0000030217075,0.00003088758,0.0040766457],"genre_scores_gemma":[0.9933845,0.00084569474,0.001701312,0.0028578602,0.00044420912,0.0001511427,0.00010039621,0.000013527676,0.0005014058],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99875015,0.000034932807,0.00041083244,0.00024311364,0.00039891404,0.0001620692],"domain_scores_gemma":[0.9863707,0.011924561,0.00013153983,0.0001377131,0.0005974364,0.00083806243],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001099124,0.00009918906,0.00025001328,0.000118435244,0.00005650485,0.000008298839,0.000040499745,0.00020274457,0.00011633341],"category_scores_gemma":[0.3114985,0.00008168725,0.000046472705,0.00018051713,0.00030682184,0.000042107997,0.00002311277,0.00016051703,0.000004219121],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014393861,0.002450629,0.014854253,0.00012856572,0.00003508569,0.0000012196142,0.0023809432,5.4638156e-9,0.000027229891,0.002492596,0.009390244,0.9680953],"study_design_scores_gemma":[0.009252169,0.0021923243,0.92144114,0.01584155,0.0011458391,0.0003007657,0.012191304,0.0015221008,0.002903277,0.0057575167,0.026907576,0.0005444117],"about_ca_topic_score_codex":0.000056175602,"about_ca_topic_score_gemma":0.000103134036,"teacher_disagreement_score":0.9675509,"about_ca_system_score_codex":0.000037020596,"about_ca_system_score_gemma":0.0041008717,"threshold_uncertainty_score":0.7274771},"labels":[],"label_agreement":null},{"id":"W2136755565","doi":"10.1186/s13012-014-0088-x","title":"Theory of planned behaviour can help understand processes underlying the use of two emergency medicine diagnostic imaging rules","year":2014,"lang":"en","type":"article","venue":"Implementation Science","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":20,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Ottawa Hospital; University of Ottawa","funders":"Ontario Ministry of Research and Innovation; Canadian Institutes of Health Research; University of Alberta; London Health Sciences Centre","keywords":"Medicine; Baseline (sea); Knowledge translation; Randomized controlled trial; Emergency department; Health administration; Health services research; Health informatics; Inefficiency; Family medicine; Public health; Nursing; Knowledge management; Computer science; Surgery","score_opus":0.21514577715961664,"score_gpt":0.48407317388483895,"score_spread":0.2689273967252223,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2136755565","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98472244,0.0001234414,0.011796925,0.0025031515,0.00023334159,0.0003476696,0.00003190197,0.00001904365,0.0002220756],"genre_scores_gemma":[0.9989726,0.00011162242,0.0005106663,0.00028401284,0.000037180496,0.000011898927,0.000024743887,0.000007921558,0.000039364368],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99820757,0.00008591232,0.000585769,0.00023598108,0.0006536869,0.00023106186],"domain_scores_gemma":[0.9833109,0.0154679315,0.00036785405,0.00027190123,0.00046085092,0.00012061205],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0017041741,0.00010064066,0.00021248714,0.00013143435,0.00017198396,0.000014829664,0.00018241578,0.000012895451,0.00034363574],"category_scores_gemma":[0.06154638,0.0000636521,0.0000347909,0.00058974937,0.0008125649,0.00015915316,0.000049826,0.00006860646,0.0000016126935],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005250642,0.000095616284,0.9650732,0.000072869596,0.000016397613,0.0000019437505,0.0033844067,0.000043828368,0.0040777912,0.009172406,0.0009263207,0.017082704],"study_design_scores_gemma":[0.002849756,0.00066178926,0.9515731,0.0012612233,0.00032800168,0.000019033088,0.02025545,0.00036441913,0.008534758,0.013883215,0.00010438549,0.00016487156],"about_ca_topic_score_codex":0.00075671717,"about_ca_topic_score_gemma":0.0003373522,"teacher_disagreement_score":0.059842207,"about_ca_system_score_codex":0.000038442857,"about_ca_system_score_gemma":0.00030565867,"threshold_uncertainty_score":0.9463586},"labels":[],"label_agreement":null},{"id":"W2136955507","doi":"10.1192/apt.bp.107.005298","title":"How psychiatrists think","year":2009,"lang":"en","type":"article","venue":"Advances in Psychiatric Treatment","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":56,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Trinity College","funders":"","keywords":"Heuristics; Cognition; Psychology; Clinical decision making; Cognitive bias; Cognitive psychology; Computer science; Management science; Medicine; Psychiatry; Intensive care medicine","score_opus":0.01216235191121721,"score_gpt":0.3447781912980376,"score_spread":0.3326158393868204,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2136955507","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6782823,0.1712763,0.0009525112,0.07286468,0.0040334044,0.0020105056,0.000024882094,0.00051909697,0.070036314],"genre_scores_gemma":[0.9630987,0.017427623,0.015956812,0.0012287444,0.00053329766,0.000028336191,0.000029018425,0.000016166638,0.0016812925],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9985703,0.00003094566,0.00032999896,0.00045704824,0.00024564774,0.00036605788],"domain_scores_gemma":[0.9983808,0.00080134405,0.00010978851,0.00046074542,0.000032053562,0.00021527903],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00009116583,0.00024760116,0.00043219232,0.00017302607,0.000055271466,0.000027831964,0.00008768874,0.00010927675,0.00002908114],"category_scores_gemma":[0.0009843645,0.00017675388,0.00020643209,0.000509012,0.000037107387,0.00013688962,0.000006802452,0.0001511334,0.000049326256],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00043365467,0.003349053,0.22051696,0.000012438237,0.00003495937,0.00013893622,0.00015584311,0.000035897137,0.0000045115285,0.0041457023,0.0016034287,0.7695686],"study_design_scores_gemma":[0.024177698,0.014572197,0.7519907,0.00075053045,0.0005439297,0.0002847142,0.00035115957,0.00013628058,0.000121344834,0.054635413,0.15164307,0.0007930083],"about_ca_topic_score_codex":0.000017534956,"about_ca_topic_score_gemma":0.000041531515,"teacher_disagreement_score":0.7687756,"about_ca_system_score_codex":0.00017616751,"about_ca_system_score_gemma":0.00009925135,"threshold_uncertainty_score":0.7207814},"labels":[],"label_agreement":null},{"id":"W2137126279","doi":"10.1111/j.1365-2923.2012.04242.x","title":"Clinical reasoning processes: unravelling complexity through graphical representation","year":2012,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":136,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Bausch Health (Canada); Centre de Santé et de Services Sociaux Cavendish; McGill University; Université de Montréal; Montreal General Hospital","funders":"","keywords":"GRASP; Computer science; Process (computing); Curriculum; Action (physics); Artificial intelligence; Management science; Psychology; Data science; Software engineering","score_opus":0.09182582099801148,"score_gpt":0.46844557967056105,"score_spread":0.3766197586725496,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2137126279","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95183635,0.0017874751,0.012500764,0.017054548,0.0031537714,0.0005209369,0.0000024475667,0.00022215051,0.012921534],"genre_scores_gemma":[0.97168845,0.0008141823,0.0133622885,0.010341231,0.0031287358,0.000056133296,0.00024963697,0.000025301253,0.00033403956],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9973525,0.0001989094,0.0007913856,0.00037575426,0.00087212527,0.00040934162],"domain_scores_gemma":[0.9915707,0.0064536934,0.00023461923,0.00045968895,0.0003941214,0.000887143],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0017199551,0.0001563701,0.00041339578,0.000058051643,0.00012121273,0.000023108603,0.00012476931,0.00034023408,0.0009725773],"category_scores_gemma":[0.24159378,0.00012872832,0.00015400944,0.00056343304,0.00039131683,0.00019161521,0.00004647677,0.00064579124,0.00017935215],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012234224,0.003727507,0.89323163,0.00010525566,0.000051275285,0.0000055274295,0.00088433747,4.974976e-7,0.000003465612,0.007347841,0.027194636,0.06732568],"study_design_scores_gemma":[0.002400955,0.000354963,0.94273305,0.0048222253,0.00036148264,0.00029269414,0.0011348234,0.00079123466,0.00016238421,0.0060591167,0.040498983,0.00038806014],"about_ca_topic_score_codex":0.00021112064,"about_ca_topic_score_gemma":0.000010833113,"teacher_disagreement_score":0.23987383,"about_ca_system_score_codex":0.000057322653,"about_ca_system_score_gemma":0.0018468952,"threshold_uncertainty_score":0.9999407},"labels":[],"label_agreement":null},{"id":"W2137901319","doi":"10.1051/pmed/2010006","title":"Le test de concordance de script comme outil d’enseignement et d’apprentissage : un projet-pilote pour les étudiants de première année de médecine","year":2010,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université de Montréal","funders":"Université de Montréal","keywords":"Humanities; Psychology; Art","score_opus":0.05535437903520025,"score_gpt":0.35213624011383626,"score_spread":0.296781861078636,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2137901319","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8914004,0.013181774,0.00978028,0.07152932,0.0016521716,0.0010686674,0.0004095348,0.0003389601,0.010638918],"genre_scores_gemma":[0.9470118,0.0032896732,0.024135558,0.0036290723,0.0019290183,0.00017779149,0.00009543975,0.00015317605,0.01957847],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9943134,0.000589547,0.0011021915,0.00095531176,0.00069457357,0.0023449836],"domain_scores_gemma":[0.98253644,0.013725715,0.00039255698,0.0011377885,0.00026986154,0.001937652],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0036818471,0.0006638715,0.0010560382,0.00012129139,0.0003997578,0.00014659119,0.0007344722,0.00084714935,0.0014036368],"category_scores_gemma":[0.087158255,0.0006740286,0.00036318268,0.00037672542,0.0010796987,0.00015308501,0.0003949167,0.0024338625,0.00029604038],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021081477,0.006438315,0.8163087,0.00028771284,0.00018772222,0.0010670485,0.0013281438,0.000030272682,0.014559954,0.001753939,0.06925946,0.08856793],"study_design_scores_gemma":[0.021120429,0.0021484338,0.78064126,0.011139356,0.0013890297,0.0012704479,0.008547159,0.00850107,0.017082388,0.008240701,0.13831519,0.0016045463],"about_ca_topic_score_codex":0.0009759187,"about_ca_topic_score_gemma":0.0007033499,"teacher_disagreement_score":0.08696338,"about_ca_system_score_codex":0.0004876156,"about_ca_system_score_gemma":0.0052373935,"threshold_uncertainty_score":0.99986756},"labels":[],"label_agreement":null},{"id":"W2137929031","doi":"10.1080/01421590802572783","title":"Is a web-based concordance test feasible to assess therapeutic decision-making skills in a French context?","year":2009,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Concordance; Context (archaeology); Curriculum; Test (biology); Medical education; Matching (statistics); Subject (documents); Educational measurement; Clinical decision making; Rank (graph theory); Psychology; Medicine; Computer science; Pedagogy; Family medicine; World Wide Web; Internal medicine; Pathology","score_opus":0.04391080003756575,"score_gpt":0.40044352230581826,"score_spread":0.3565327222682525,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2137929031","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9622604,0.0008838362,0.0046732766,0.023639679,0.00027925172,0.0005958713,0.000005458426,0.00015215125,0.007510046],"genre_scores_gemma":[0.94547606,0.00003154518,0.001607359,0.05142442,0.00025880942,0.00003996906,0.0000056271238,0.000025479,0.0011307537],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99702364,0.00011010398,0.0006504443,0.0005569802,0.0011205374,0.00053829176],"domain_scores_gemma":[0.98472106,0.013747327,0.00009234341,0.00054563297,0.000115403614,0.0007782351],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0013117708,0.00025343333,0.0006742829,0.00016241553,0.000048658017,0.000037625825,0.0002812601,0.00033040866,0.004657522],"category_scores_gemma":[0.1352419,0.00019383049,0.00017179592,0.0006065663,0.00012586245,0.00003958158,0.000040038634,0.00078713236,0.00032273162],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000098867626,0.0020947685,0.53087014,0.000008360348,0.000015782798,0.00040463515,0.00020880776,0.0000040550403,0.00004119521,0.00003979406,0.015883595,0.45033],"study_design_scores_gemma":[0.0092917085,0.0016265247,0.93954605,0.01474482,0.00009040986,0.00006598045,0.00010678954,0.013582248,0.00010280797,0.001773936,0.018614102,0.00045464825],"about_ca_topic_score_codex":0.00009615545,"about_ca_topic_score_gemma":0.00011111837,"teacher_disagreement_score":0.44987532,"about_ca_system_score_codex":0.00017051908,"about_ca_system_score_gemma":0.0007903638,"threshold_uncertainty_score":0.99625236},"labels":[],"label_agreement":null},{"id":"W2138081764","doi":"10.1097/acm.0000000000000566","title":"In Reply to Webster","year":2015,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Victoria General Hospital; Nova Scotia Health Authority; Capital District Health Authority","funders":"","keywords":"Reflexivity; Aside; Unconscious mind; Process (computing); Point (geometry); Cognitive psychology; Psychology; Computer science; Reflection (computer programming); Epistemology; Cognitive science; Psychoanalysis; Sociology","score_opus":0.07705099555766232,"score_gpt":0.40415766348082777,"score_spread":0.32710666792316545,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2138081764","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0030625602,0.0019178486,0.000022376642,0.9744953,0.0020858648,0.0007340209,0.000010508955,0.00008692243,0.017584626],"genre_scores_gemma":[0.00093246624,0.00046047306,0.00008666784,0.8894899,0.024199504,0.00007156067,0.00029251815,0.00008175195,0.084385164],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99596274,0.000116087336,0.0011907181,0.00082987134,0.0012009088,0.00069969567],"domain_scores_gemma":[0.9900528,0.00801704,0.00021677857,0.00087817595,0.00020037449,0.00063482975],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity","insufficient_payload"],"category_scores_codex":[0.0016623393,0.00043197107,0.0014821038,0.00053487706,0.000016064369,0.0000037783843,0.0003492362,0.0018616996,0.0012492777],"category_scores_gemma":[0.12736632,0.00030941112,0.00009424948,0.0005330725,0.00019457607,0.000040551862,0.00012779105,0.007293839,0.0017061407],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000089489076,0.0000076195947,0.001620567,0.000077776946,0.000030557643,0.0027680085,0.0007525542,3.9461636e-7,0.000014031064,0.0000111082645,0.9936552,0.0009726775],"study_design_scores_gemma":[0.0027602655,0.000668925,0.00061560964,0.013340076,0.0002302613,0.0003443752,0.00008208386,0.0000058920978,0.000011528754,0.0010663472,0.9805927,0.00028192915],"about_ca_topic_score_codex":0.00040267807,"about_ca_topic_score_gemma":0.0000026942828,"teacher_disagreement_score":0.12570398,"about_ca_system_score_codex":0.0002701291,"about_ca_system_score_gemma":0.00040362947,"threshold_uncertainty_score":0.9999358},"labels":[],"label_agreement":null},{"id":"W2138947939","doi":"10.1111/j.1365-2834.2011.01369.x","title":"Patient experiences of uncertainty - a synthesis to guide nursing practice and research","year":2012,"lang":"en","type":"review","venue":"Journal of Nursing Management","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":55,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Movement Disorders","funders":"Helse Vest","keywords":"CINAHL; Feeling; Nursing; Psychological intervention; Qualitative research; Nursing management; Stressor; Medicine; Coping (psychology); Intervention (counseling); Health care; Nursing research; Nursing care; Psychology; Clinical psychology; Social psychology","score_opus":0.19267863454356896,"score_gpt":0.5324925817544487,"score_spread":0.33981394721087976,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2138947939","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0004093021,0.9876333,0.00015824256,0.00113811,0.00084579113,0.0007242989,0.000002047971,0.000006540159,0.0090824],"genre_scores_gemma":[0.0021567976,0.9879866,0.009171104,0.00008675647,0.00038090255,0.000039773688,0.0000013106342,0.000031591917,0.00014513636],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9961484,0.00052871235,0.0013999022,0.0002569711,0.0012664468,0.00039956084],"domain_scores_gemma":[0.98255605,0.015003041,0.0010523472,0.00036459335,0.0005503953,0.00047358274],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0035688523,0.00023039222,0.0014915119,0.00069707714,0.000087069006,0.00004856506,0.00019070347,0.00013812874,0.00004285556],"category_scores_gemma":[0.036145765,0.00016017875,0.00032366268,0.000526142,0.0002542626,0.00009656901,0.00007082987,0.000535164,0.00000857962],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015247289,0.0011383484,0.0000012423575,0.0017253453,0.00029554305,0.00014279714,0.0061435015,0.0000033517724,1.3975735e-7,0.00021127994,0.00783969,0.9823463],"study_design_scores_gemma":[0.00015106254,0.0007917967,0.00001624559,0.35898724,0.0037278298,0.00051731436,0.017166186,0.000002184903,0.0000016877444,0.00004994896,0.6184772,0.00011127713],"about_ca_topic_score_codex":0.000007908444,"about_ca_topic_score_gemma":7.20342e-8,"teacher_disagreement_score":0.982235,"about_ca_system_score_codex":0.00070926483,"about_ca_system_score_gemma":0.00022869384,"threshold_uncertainty_score":0.9719732},"labels":[],"label_agreement":null},{"id":"W2139919326","doi":"","title":"Efficacy of Teaching Clinical Clerks and Residents How to Fill Out the Form 1 of the Mental Health Act Using an e-Learning Module","year":2009,"lang":"en","type":"article","venue":"The Electronic Journal of e-Learning","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Mental health; Test (biology); Grading (engineering); Intervention (counseling); Intraclass correlation; Psychology; Medical education; Medicine; Clinical psychology; Psychiatry; Psychometrics","score_opus":0.03337663441103699,"score_gpt":0.3866972832914595,"score_spread":0.3533206488804225,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2139919326","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9851858,0.0011559986,0.0004909923,0.012538837,0.00014581926,0.00030833585,5.8711055e-7,0.000011904751,0.00016171894],"genre_scores_gemma":[0.9978117,0.0006774479,0.0002785033,0.0005892103,0.00036429055,4.8987795e-7,0.0000017565715,0.000023886427,0.00025271333],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9963515,0.0012715505,0.0009684682,0.00020043613,0.0006347942,0.000573296],"domain_scores_gemma":[0.9929149,0.005067264,0.0013388565,0.00030268094,0.00014347279,0.0002327892],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.008242756,0.00018033359,0.00065977644,0.00008133589,0.0005145859,0.000041153184,0.0003921254,0.00009708435,0.0000126701025],"category_scores_gemma":[0.032149605,0.00009077185,0.0003208252,0.00017141798,0.00014700576,0.00011214237,0.00010729725,0.003274679,8.4906435e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.007640284,0.002758412,0.44944766,0.00009669595,0.0014206693,0.000022894956,0.03694701,0.020657316,0.009371501,0.0034962222,0.0018568592,0.46628448],"study_design_scores_gemma":[0.010813792,0.027367221,0.9218855,0.008172872,0.00084444234,0.0010088703,0.00916239,0.013861451,0.0012970438,0.0023453715,0.0028123057,0.00042875073],"about_ca_topic_score_codex":0.00006815614,"about_ca_topic_score_gemma":0.000019497402,"teacher_disagreement_score":0.47243783,"about_ca_system_score_codex":0.00016946104,"about_ca_system_score_gemma":0.0005576233,"threshold_uncertainty_score":0.9990248},"labels":[],"label_agreement":null},{"id":"W2140353941","doi":"10.1002/chp.69","title":"The difficulty with experience: Does practice increase susceptibility to premature closure?","year":2006,"lang":"en","type":"article","venue":"Journal of Continuing Education in the Health Professions","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":70,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medical diagnosis; Confounding; Cognition; Test (biology); Psychology; Closure (psychology); Medicine; Clinical psychology; Psychiatry","score_opus":0.014355155020215987,"score_gpt":0.41655146399222204,"score_spread":0.40219630897200603,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2140353941","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.85319567,0.0007412443,0.000058076566,0.14413004,0.0007794201,0.00069371983,0.0000014807512,0.000008631077,0.0003917173],"genre_scores_gemma":[0.99137616,0.000097608274,0.0020813635,0.0044362266,0.00067785423,0.00004423502,0.0000037842342,0.000008884039,0.0012738591],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9975059,0.0008254194,0.0007999012,0.00013458852,0.0004858515,0.0002483383],"domain_scores_gemma":[0.982269,0.015437473,0.00088250457,0.00039413272,0.00078403513,0.00023287177],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0038441333,0.00010232945,0.00026693306,0.00006437,0.00030190375,0.000047850906,0.00018975981,0.0000698783,0.0000107972355],"category_scores_gemma":[0.04758381,0.0000378044,0.00006460811,0.0003273635,0.0000819416,0.000105957246,0.000021095322,0.0007649371,0.0000028772536],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0042540617,0.007139411,0.6144918,0.00018410596,0.000053576117,0.000035078207,0.104724675,0.000035035322,0.00020935913,0.0025132713,0.20768711,0.0586725],"study_design_scores_gemma":[0.0009956534,0.00058874907,0.88291514,0.0074500577,0.000071339855,0.00024456572,0.084801,0.000003327018,0.000028065026,0.0005199274,0.022304598,0.00007756937],"about_ca_topic_score_codex":0.00038230722,"about_ca_topic_score_gemma":0.00039136014,"teacher_disagreement_score":0.26842332,"about_ca_system_score_codex":0.00013610121,"about_ca_system_score_gemma":0.001747991,"threshold_uncertainty_score":0.9604388},"labels":[],"label_agreement":null},{"id":"W2140529495","doi":"10.1111/j.1365-2923.2011.03933.x","title":"The influence of medical students’ self‐explanations on diagnostic performance","year":2011,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":77,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Sherbrooke","funders":"","keywords":"Context (archaeology); Psychology; Variety (cybernetics); Medical education; Mathematics education; Medicine; Computer science; Artificial intelligence","score_opus":0.018654223296594146,"score_gpt":0.37154308177206374,"score_spread":0.3528888584754696,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2140529495","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98605025,0.00019200466,0.000014496913,0.010330206,0.00053417357,0.0002201111,8.7906767e-7,0.000042161097,0.0026157384],"genre_scores_gemma":[0.9924593,0.0022726445,0.00009803231,0.004665135,0.00019833355,0.00008705762,0.000014964377,0.000009155106,0.00019535806],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9968302,0.00006665576,0.00042881793,0.00017413018,0.0023181054,0.0001820595],"domain_scores_gemma":[0.9867173,0.01213482,0.00012338308,0.00037483108,0.00021602142,0.00043366264],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0014401727,0.00009297963,0.0001725103,0.000053690594,0.00010531201,0.000011240913,0.00045346358,0.00016895619,0.00067261735],"category_scores_gemma":[0.228554,0.000057313628,0.00005261662,0.00020463835,0.00022629225,0.00003467893,0.00006184359,0.00036081672,0.00015364018],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012736476,0.007209866,0.79495335,0.00007776379,0.00009912097,0.000020028394,0.0024833395,0.000005813591,0.0000014895624,0.0061205523,0.036829907,0.15207137],"study_design_scores_gemma":[0.00068729464,0.00043701974,0.989864,0.0024148861,0.00006633325,0.00003192974,0.00017804962,0.00023977044,0.00006315625,0.00014798596,0.00580134,0.00006823825],"about_ca_topic_score_codex":0.0001825293,"about_ca_topic_score_gemma":0.000014508438,"teacher_disagreement_score":0.22711383,"about_ca_system_score_codex":0.000044218672,"about_ca_system_score_gemma":0.0017589382,"threshold_uncertainty_score":0.7779442},"labels":[],"label_agreement":null},{"id":"W2140898426","doi":"10.3109/0142159x.2013.847910","title":"Twelve tips on engaging learners in checking health care decisions","year":2013,"lang":"en","type":"review","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; University Health Network","funders":"Universiteit Maastricht","keywords":"Medical education; Psychology; MEDLINE; Health care; Medicine; Nursing; Political science","score_opus":0.15851986478513946,"score_gpt":0.4793713135161756,"score_spread":0.32085144873103616,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2140898426","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00022921343,0.9854937,0.000030256246,0.0020453501,0.000503268,0.00096310466,0.0000038645207,0.00010897366,0.010622239],"genre_scores_gemma":[0.0003352192,0.9938436,0.00022377998,0.0016898788,0.00089282496,0.00016652573,0.00025126192,0.000104917795,0.0024919638],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99527836,0.0006849765,0.0012811292,0.0007689656,0.0012830291,0.00070352707],"domain_scores_gemma":[0.98605734,0.011583969,0.0002995416,0.00070920977,0.00004920142,0.0013007219],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0021627184,0.00048278028,0.0027472554,0.00030715807,0.00008992896,0.000028376491,0.00032365505,0.00092635956,0.0032796785],"category_scores_gemma":[0.17161009,0.00032340075,0.0006368744,0.00043058995,0.00015224298,0.000021332826,0.00014546938,0.004365846,0.0012507114],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000060974708,0.00024134359,0.00041523942,0.001127278,0.000054893535,0.00019104824,0.00059244176,1.5333508e-7,9.958004e-10,0.00003278006,0.018342061,0.97899663],"study_design_scores_gemma":[0.00080177595,0.00018334811,0.00004466437,0.18802638,0.00016676607,0.00008726511,0.0004452113,0.000029541545,1.325347e-8,0.000011141635,0.8099712,0.00023270164],"about_ca_topic_score_codex":0.0003242037,"about_ca_topic_score_gemma":0.000033460812,"teacher_disagreement_score":0.97876394,"about_ca_system_score_codex":0.0006675661,"about_ca_system_score_gemma":0.0020757203,"threshold_uncertainty_score":0.9999218},"labels":[],"label_agreement":null},{"id":"W2141932098","doi":"10.1080/01421590802512979","title":"Teaching clinical thinking to first-year medical students","year":2009,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":51,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McGill University","funders":"Fondation de l'Hôpital Général de Montréal; McGill University; Max Bell Foundation","keywords":"Modalities; Active listening; Medical education; Inference; Clinical Practice; Psychology; Critical thinking; Medicine; Mathematics education; Computer science; Artificial intelligence; Nursing; Sociology","score_opus":0.038839555648362126,"score_gpt":0.44133814570655855,"score_spread":0.40249859005819644,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2141932098","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8236295,0.00014585796,0.0017600188,0.14664198,0.00090960745,0.0004334983,9.527387e-7,0.0003863588,0.026092233],"genre_scores_gemma":[0.93468946,0.000113113965,0.0025953196,0.052565698,0.0037931479,0.000017485381,0.000021557842,0.000043809072,0.006160412],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99187744,0.0003515543,0.0011404494,0.000710987,0.0052562915,0.00066329236],"domain_scores_gemma":[0.99104536,0.004541256,0.0001080996,0.00068521505,0.00008021545,0.003539861],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.010104797,0.0002826306,0.00084307126,0.000101702084,0.00018192435,0.0000638272,0.0009569185,0.00081201707,0.008933586],"category_scores_gemma":[0.285363,0.00020781465,0.00036214798,0.00020660834,0.00017722642,0.000047043457,0.00033445848,0.0027028064,0.0013352872],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021917591,0.002700801,0.6495689,0.000012333744,0.00010422703,0.0010729076,0.0005310135,3.8443162e-7,5.283811e-7,0.0010801399,0.11931102,0.22539853],"study_design_scores_gemma":[0.01189388,0.003534257,0.77760255,0.0091106435,0.00033807845,0.00035095873,0.00031485272,0.0013416319,0.0000054068337,0.00365563,0.19108547,0.0007666346],"about_ca_topic_score_codex":0.000105643216,"about_ca_topic_score_gemma":0.000025161347,"teacher_disagreement_score":0.27525818,"about_ca_system_score_codex":0.00008872528,"about_ca_system_score_gemma":0.00036275858,"threshold_uncertainty_score":0.999598},"labels":[],"label_agreement":null},{"id":"W2143195179","doi":"10.1503/cmaj.1031667","title":"Tips for learners of evidence-based medicine: 2. Measures of precision (confidence intervals)","year":2004,"lang":"en","type":"article","venue":"Canadian Medical Association Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":77,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Confidence interval; Computer science; Absolute risk reduction; Statistics; Series (stratigraphy); Reduction (mathematics); Medicine; Medical physics; Mathematics","score_opus":0.11687635632388078,"score_gpt":0.3764734606212213,"score_spread":0.2595971042973405,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2143195179","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.66853535,0.003751649,0.030562088,0.29233333,0.002346616,0.00090512045,0.00005098473,0.00003590352,0.0014789182],"genre_scores_gemma":[0.993309,0.00048614407,0.0008736185,0.0044786087,0.00067689363,0.0000067154756,0.000006551544,0.000013026605,0.00014944025],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9967308,0.0001282755,0.00091713434,0.00015907039,0.0017626302,0.00030206935],"domain_scores_gemma":[0.98567516,0.010979698,0.0005541221,0.00014813486,0.0012873719,0.0013555398],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.005052602,0.00010974351,0.00050822244,0.000232784,0.00006740821,0.000010420999,0.0001953078,0.00033576097,0.0009455438],"category_scores_gemma":[0.40603217,0.000075744865,0.00022185165,0.00025316884,0.00016212615,0.00007206513,0.000007975399,0.00051860494,0.00001059134],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012037482,0.00055342936,0.2522531,0.0003572932,0.00073642586,0.00036347445,0.0025818844,0.00064436195,0.0012527182,0.0027473383,0.15995412,0.5773521],"study_design_scores_gemma":[0.06440653,0.016904388,0.4605424,0.26054317,0.0025865242,0.0015903328,0.006270047,0.0031906618,0.016643653,0.024461119,0.14138484,0.001476339],"about_ca_topic_score_codex":0.00399814,"about_ca_topic_score_gemma":0.0034686928,"teacher_disagreement_score":0.5758758,"about_ca_system_score_codex":0.00058282405,"about_ca_system_score_gemma":0.0061694193,"threshold_uncertainty_score":0.99996775},"labels":[],"label_agreement":null},{"id":"W2143428226","doi":"10.1007/s10459-009-9182-2","title":"Clinical cognition and diagnostic error: applications of a dual process model of reasoning","year":2009,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":477,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Cognition; Schema (genetic algorithms); Variety (cybernetics); Dual process theory (moral psychology); Process (computing); Medical diagnosis; Computer science; Dual (grammatical number); Cognitive science; Management science; Cognitive psychology; Psychology; Artificial intelligence; Medicine; Machine learning; Neuroscience","score_opus":0.04411648419866486,"score_gpt":0.4876456073800888,"score_spread":0.4435291231814239,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2143428226","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9811469,0.009716259,0.0036687409,0.0025590295,0.00014865938,0.00085098314,0.0000072111807,0.000018114386,0.0018841053],"genre_scores_gemma":[0.97795033,0.0063917018,0.014834964,0.0006663835,0.000064407664,0.00006320035,0.000013561841,0.0000027316376,0.000012739475],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99855816,0.00004862487,0.000736358,0.00026881543,0.00023181034,0.00015622075],"domain_scores_gemma":[0.9946256,0.0045096087,0.0004229542,0.00012536727,0.00017307371,0.00014333936],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0012259422,0.000067390814,0.0002792802,0.00012563162,0.00006441371,0.000004344695,0.000059824168,0.000047768914,0.0000036611611],"category_scores_gemma":[0.018487751,0.000058968784,0.000027745416,0.00050393114,0.0003696734,0.0001729617,0.000007894544,0.00011381577,6.780609e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000056966343,0.0015205039,0.3828866,0.00037530114,0.0000023539865,1.898323e-7,0.0007158722,0.0021998629,0.00001971632,0.0074580843,0.00004102298,0.6047235],"study_design_scores_gemma":[0.0022911094,0.0043838173,0.7766656,0.016808107,0.00009649816,0.00003722595,0.0038666285,0.0537159,0.0003667078,0.14116494,0.00027909363,0.0003243088],"about_ca_topic_score_codex":0.000040751398,"about_ca_topic_score_gemma":0.000020207406,"teacher_disagreement_score":0.6043992,"about_ca_system_score_codex":0.00002160868,"about_ca_system_score_gemma":0.0011827996,"threshold_uncertainty_score":0.98977995},"labels":[],"label_agreement":null},{"id":"W2143440110","doi":"10.1111/j.1466-769x.2004.00175.x","title":"Medical and nursing clinical decision making: a comparative epistemological analysis","year":2004,"lang":"en","type":"article","venue":"Nursing Philosophy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":36,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Montreal Children's Hospital; McGill University; Children's Hospital of Eastern Ontario","funders":"","keywords":"Judgement; Epistemology; Task (project management); Psychology; Realism; Section (typography); Engineering ethics; Nursing; Medicine; Computer science; Philosophy","score_opus":0.10655134412989602,"score_gpt":0.4713283994366047,"score_spread":0.3647770553067087,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2143440110","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9372397,0.0011612672,0.016157994,0.022700652,0.0006004879,0.00029771094,0.0000060246202,0.00016476357,0.02167142],"genre_scores_gemma":[0.9836042,0.00007920875,0.013620463,0.0019128262,0.0007234496,0.000005348199,0.00002389284,0.000016518006,0.000014123958],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99714744,0.00015513295,0.0008589745,0.00068557484,0.0007930046,0.0003598959],"domain_scores_gemma":[0.9902409,0.00835065,0.00016675316,0.00039547004,0.00012788009,0.0007184023],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.000887604,0.00024249138,0.0010913488,0.0002446907,0.00016364247,0.000038814884,0.00013468134,0.00044655288,0.00014025792],"category_scores_gemma":[0.017312458,0.00018870989,0.0004348022,0.00073124765,0.001233731,0.000047922094,0.00003138261,0.00068493263,0.000059681042],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.006043655,0.015360973,0.22528914,0.000056025336,0.0021295822,0.0027864357,0.002903031,0.00036013563,0.000017935892,0.101627216,0.0028164494,0.64060944],"study_design_scores_gemma":[0.007054268,0.0020958113,0.6181787,0.023783624,0.0029309008,0.0008405655,0.00014925361,0.0022679728,0.000019230618,0.34212285,0.00011939212,0.00043742734],"about_ca_topic_score_codex":0.000006084497,"about_ca_topic_score_gemma":0.0000019400654,"teacher_disagreement_score":0.640172,"about_ca_system_score_codex":0.00017129614,"about_ca_system_score_gemma":0.0002042892,"threshold_uncertainty_score":0.9909651},"labels":[],"label_agreement":null},{"id":"W2144038358","doi":"10.36834/cmej.36521","title":"Does Applying Biomedical Knowledge Improve Diagnostic Performance When Solving Electrolyte Problems?","year":2010,"lang":"en","type":"article","venue":"Canadian Medical Education Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Odds ratio; Odds; Logistic regression; Medicine; Medical education; Internal medicine","score_opus":0.00792698013410816,"score_gpt":0.2946518359373324,"score_spread":0.28672485580322427,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2144038358","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9016753,0.00097498397,0.00020046647,0.076234095,0.012608625,0.0008097386,0.000009237368,0.00008920781,0.0073983422],"genre_scores_gemma":[0.9767241,0.00049653684,0.0014162941,0.0123960115,0.0066799303,0.00013865448,0.000045593024,0.000045074918,0.0020577891],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.997266,0.00006615893,0.0006981787,0.0003533682,0.0007876846,0.00082859845],"domain_scores_gemma":[0.97687364,0.004052222,0.00021191551,0.00040781137,0.0005846051,0.017869832],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0016140145,0.0002266744,0.000371099,0.00038586036,0.00040880192,0.00012294333,0.00033352344,0.00046688365,0.018008221],"category_scores_gemma":[0.33517522,0.00014445411,0.00014288316,0.0002757863,0.00034017285,0.00013758233,0.000030307816,0.002781575,0.00027634823],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000024129342,0.0007209219,0.022234453,0.00011026156,0.00007825139,0.00012680863,0.0011996032,6.297073e-7,0.00061684765,0.00030673743,0.112572916,0.86200845],"study_design_scores_gemma":[0.0033411752,0.00078515615,0.033617657,0.006482587,0.00030657012,0.0048408485,0.0006308342,0.004274615,0.00029263762,0.0016923904,0.94302535,0.00071015017],"about_ca_topic_score_codex":0.0024575621,"about_ca_topic_score_gemma":0.007534217,"teacher_disagreement_score":0.86129826,"about_ca_system_score_codex":0.0002856493,"about_ca_system_score_gemma":0.054767,"threshold_uncertainty_score":0.99951905},"labels":[],"label_agreement":null},{"id":"W2144895093","doi":"10.1136/bmjopen-2013-002887","title":"Influence of societal and practice contexts on health professionals’ clinical reasoning: a scoping study protocol","year":2013,"lang":"en","type":"article","venue":"BMJ Open","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval; Health and Social Services Centre University Institute of Geriatrics of Sherbrooke; Centre de réadaptation Lethbridge-Layton-Mackay; Centre de Santé et de Services Sociaux de la Vieille-Capitale; Université de Sherbrooke","funders":"","keywords":"Dissemination; Medicine; Context (archaeology); Thematic analysis; Health care; Grey literature; Psychological intervention; Multidisciplinary approach; Knowledge management; Public relations; Medical education; Qualitative research; Nursing; MEDLINE; Political science; Computer science; Sociology; Social science","score_opus":0.18837704226062132,"score_gpt":0.6113300730530619,"score_spread":0.4229530307924405,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2144895093","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5677545,0.000011647678,0.00000313614,0.012158326,0.000040054703,0.41773424,0.0000016661842,0.000020087069,0.0022763582],"genre_scores_gemma":[0.6612854,0.0000076044194,0.0037955544,0.015898772,0.0001583064,0.31797376,0.0000047105464,0.00002846344,0.0008474312],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9971592,0.00076036464,0.0010066075,0.00040038055,0.00044877428,0.0002246897],"domain_scores_gemma":[0.97759396,0.020458223,0.00074453076,0.00048763858,0.00035792717,0.00035774286],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0054869056,0.00013573618,0.00069834443,0.00002657126,0.00010168228,0.000056552806,0.00018405734,0.000094126044,0.00007141298],"category_scores_gemma":[0.17147279,0.00009472294,0.00006283078,0.00011531612,0.00013142046,0.00022451355,0.00038411928,0.00030399038,0.000058346133],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.009023874,0.01820817,0.8457515,0.0010903593,0.00035772662,0.0000966769,0.0036852127,0.000019495588,0.000026497584,0.0006812939,0.060186278,0.06087291],"study_design_scores_gemma":[0.011709626,0.009321853,0.9325949,0.04328254,0.000051450083,0.000057331108,0.0021887461,0.00005672009,0.000014776382,0.000113071634,0.0004447661,0.00016425243],"about_ca_topic_score_codex":0.00066061725,"about_ca_topic_score_gemma":0.00002117325,"teacher_disagreement_score":0.16598588,"about_ca_system_score_codex":0.000026486498,"about_ca_system_score_gemma":0.0011993402,"threshold_uncertainty_score":0.83550626},"labels":[],"label_agreement":null},{"id":"W2145349784","doi":"10.1097/00001888-200512000-00012","title":"Viewpoint:","year":2005,"lang":"en","type":"editorial","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":45,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hospital for Sick Children; University of Toronto","funders":"","keywords":"Active listening; Interview; Medical education; Medical history; Psychology; Narrative; Data collection; Physical examination; Medicine; Radiology","score_opus":0.029084905488237358,"score_gpt":0.4027778598111686,"score_spread":0.37369295432293126,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2145349784","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00016063599,0.020840919,0.000061894796,0.036922697,0.9190581,0.0005220292,0.000042408887,0.0002712508,0.022120088],"genre_scores_gemma":[0.00021522187,0.03653172,0.00010377317,0.008246064,0.9037021,0.00004564131,0.0013758669,0.00011096788,0.049668647],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99499524,0.0000856158,0.0013789061,0.000861039,0.0020318346,0.00064734736],"domain_scores_gemma":[0.96274936,0.03468296,0.0004794228,0.0008400074,0.0003772294,0.00087101763],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity","insufficient_payload"],"category_scores_codex":[0.0016428964,0.00057398557,0.0018757281,0.00025614907,0.00006298799,0.0000049561027,0.00040960655,0.003597174,0.0022550945],"category_scores_gemma":[0.34308434,0.00040479805,0.00022000121,0.00032127462,0.00047107108,0.000040330313,0.00012474364,0.006696224,0.0013821207],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012485408,0.000020622934,0.00005969835,0.00013120107,0.00010557274,0.000117091986,0.00024790937,2.6123055e-7,0.000016575115,0.000109145134,0.9824141,0.016652962],"study_design_scores_gemma":[0.0037332026,0.0006475666,0.00010653654,0.02102489,0.0011212346,0.00005914037,0.000058626745,0.000014152757,0.0000142271365,0.00061013934,0.97228056,0.0003297484],"about_ca_topic_score_codex":0.00010580044,"about_ca_topic_score_gemma":0.0000032648472,"teacher_disagreement_score":0.34144142,"about_ca_system_score_codex":0.00022462309,"about_ca_system_score_gemma":0.00079341885,"threshold_uncertainty_score":0.9998404},"labels":[],"label_agreement":null},{"id":"W2147558794","doi":"10.1136/bmjqs-2012-001713","title":"Cognitive debiasing 2: impediments to and strategies for change","year":2013,"lang":"en","type":"article","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":402,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Debiasing; Variety (cybernetics); Schema (genetic algorithms); Cognition; Psychological intervention; Computer science; Cognitive psychology; Psychology; Knowledge management; Social psychology; Artificial intelligence; Machine learning","score_opus":0.16991509695484305,"score_gpt":0.48729092231251037,"score_spread":0.3173758253576673,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2147558794","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9824295,0.00015612351,0.0034466302,0.009130998,0.0001641336,0.0026236717,0.00008034603,0.00006269528,0.001905883],"genre_scores_gemma":[0.9909794,0.000046961242,0.0017402687,0.006118172,0.0003723133,0.00040092072,0.00006806054,0.000014996645,0.00025891574],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9987271,0.00006702593,0.0004947914,0.0002711969,0.00017918486,0.0002607103],"domain_scores_gemma":[0.98895276,0.010285505,0.00008221075,0.00014814523,0.00021781425,0.0003135821],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0008408898,0.0001226151,0.0003258751,0.000030836367,0.00008229371,0.00004385629,0.00003501212,0.00008399902,0.000115487244],"category_scores_gemma":[0.025482208,0.000099824734,0.000074084914,0.00007479046,0.000061901366,0.00011159541,0.000048784437,0.000085809195,0.00008932877],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0031821593,0.00088106154,0.14345002,0.0005421533,0.00036335638,0.000022763532,0.011230008,0.0000012016666,0.001161773,0.008177402,0.016967792,0.81402034],"study_design_scores_gemma":[0.004065098,0.0006916695,0.97898823,0.0022044757,0.00015051811,0.000013799591,0.00852554,0.00009296593,0.00020863149,0.00371999,0.0010632679,0.00027582253],"about_ca_topic_score_codex":0.0010470558,"about_ca_topic_score_gemma":0.000037474645,"teacher_disagreement_score":0.8355382,"about_ca_system_score_codex":0.000027898575,"about_ca_system_score_gemma":0.00010849341,"threshold_uncertainty_score":0.9827266},"labels":[],"label_agreement":null},{"id":"W2147782722","doi":"10.1111/j.1365-2923.2009.03507.x","title":"Diagnostic error and clinical reasoning","year":2009,"lang":"en","type":"review","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":491,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Dual process theory (moral psychology); Cognition; Commit; Context (archaeology); Analytic reasoning; Cognitive psychology; Psychology; Debiasing; Process (computing); Dual (grammatical number); Confirmation bias; Psychology of reasoning; Thinking processes; Computer science; Cognitive science; Verbal reasoning; Artificial intelligence; Reasoning system; Social psychology; Mathematics education","score_opus":0.05800360363798306,"score_gpt":0.48941264769180265,"score_spread":0.4314090440538196,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2147782722","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00012465443,0.9924343,0.000025492216,0.0027005034,0.0015478634,0.00074879604,0.0000037074656,0.00009202455,0.0023226587],"genre_scores_gemma":[0.000027644013,0.9903013,0.0008233161,0.0045182183,0.0025629534,0.00009304095,0.00042385058,0.00004630947,0.0012033528],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9966777,0.0003089802,0.0013397904,0.0006752611,0.0006701426,0.0003281092],"domain_scores_gemma":[0.9543124,0.04282225,0.00051625003,0.00071990024,0.00015934501,0.0014698597],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0015655543,0.00035881184,0.002090597,0.00014115388,0.000066381406,0.000034393368,0.00016646282,0.0010770478,0.0005799967],"category_scores_gemma":[0.5559118,0.0002604308,0.00050816336,0.00028444285,0.00025349975,0.00003133365,0.00006632943,0.0013603697,0.00028519053],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000054010507,0.00088141335,0.00073097897,0.0014229381,0.00006513832,0.0000494558,0.000011533775,2.3087905e-9,2.6485525e-10,0.00014445763,0.045099646,0.95158905],"study_design_scores_gemma":[0.00038990643,0.0002739282,0.0033025062,0.15943858,0.0020848953,0.0005177685,0.000010118935,0.000012910078,4.7814814e-9,0.000071531,0.8336917,0.00020614959],"about_ca_topic_score_codex":0.00004086963,"about_ca_topic_score_gemma":0.0000054565076,"teacher_disagreement_score":0.9513829,"about_ca_system_score_codex":0.000104450904,"about_ca_system_score_gemma":0.0067378343,"threshold_uncertainty_score":0.9999848},"labels":[],"label_agreement":null},{"id":"W2148188287","doi":"10.1177/000992280404300807","title":"Student Evaluation Practices in Pediatric Clerkships: A Survey of the Medical Schools in the United States and Canada","year":2004,"lang":"en","type":"article","venue":"Clinical Pediatrics","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Clinical clerkship; Curriculum; Medical education; Consistency (knowledge bases); Medical school; Educational measurement; Ambulatory; Family medicine; Pediatrics; Psychology; Surgery; Pedagogy","score_opus":0.131127595154078,"score_gpt":0.46566928855620204,"score_spread":0.33454169340212403,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2148188287","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.987472,0.0009355059,0.0000045299935,0.010454782,0.00040719783,0.0006406432,0.000015346639,0.0000062550075,0.000063753105],"genre_scores_gemma":[0.98789454,0.008884788,0.00004305677,0.0027355775,0.00034334545,0.000020831263,0.00006275158,0.000009993779,0.0000050951826],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9945945,0.0017039361,0.0013424612,0.00033301744,0.0017691894,0.000256885],"domain_scores_gemma":[0.8770051,0.12015607,0.0011185348,0.0007006615,0.0006204999,0.00039915784],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.018036999,0.0001331938,0.0004098524,0.000120374665,0.00004232596,0.000022659202,0.00034563887,0.00023664889,0.000017850401],"category_scores_gemma":[0.57726395,0.00007522815,0.00006592008,0.0014464569,0.00012676467,0.000042370622,0.00012373668,0.0012665922,0.000002716937],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009387117,0.0007755451,0.9964443,0.000029124323,0.0000078072335,0.00003996304,0.00012485974,0.0002404901,1.752406e-8,0.000018546742,0.0013668005,0.00085870177],"study_design_scores_gemma":[0.0040668854,0.00022399488,0.99399775,0.00014764245,0.00022161141,0.000008120383,0.0005230431,0.0003881758,2.2515398e-7,0.00028022396,0.000060997507,0.00008134787],"about_ca_topic_score_codex":0.5189646,"about_ca_topic_score_gemma":0.7285882,"teacher_disagreement_score":0.55922693,"about_ca_system_score_codex":0.00013723168,"about_ca_system_score_gemma":0.004612485,"threshold_uncertainty_score":0.81823504},"labels":[],"label_agreement":null},{"id":"W2150519824","doi":"10.1080/01421590601032427","title":"Composition of the panel of reference for concordance tests: Do teaching functions have an impact on examinees’ ranks and absolute scores?","year":2007,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":29,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Concordance; Ranking (information retrieval); Test (biology); Medicine; Ambiguity; Family medicine; Psychology; Statistics; Mathematics; Computer science; Internal medicine; Artificial intelligence","score_opus":0.0828956857860335,"score_gpt":0.4049771565153758,"score_spread":0.3220814707293423,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2150519824","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99379784,0.00016088942,0.0023928469,0.000499575,0.00011683533,0.00030568632,0.000018415329,0.000021328971,0.0026865688],"genre_scores_gemma":[0.9987511,0.000019457506,0.00031127693,0.0002554558,0.00019442363,0.00000945838,0.000033027278,0.0000135181035,0.00041228582],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9986538,0.00009474638,0.00041179635,0.0002082951,0.00043598295,0.00019534482],"domain_scores_gemma":[0.99373466,0.005404695,0.00016137713,0.00032685994,0.00008874968,0.0002836812],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0017695901,0.00011927173,0.00033559784,0.000044175296,0.00006611992,0.0000057856673,0.000106365624,0.00017303805,0.0001621518],"category_scores_gemma":[0.025729923,0.00006830018,0.000099396486,0.00006610834,0.00033271278,0.000028753531,0.00003180441,0.0005202383,0.0000014857178],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0023976853,0.0032121278,0.817408,0.00016589377,0.00020850751,0.000024187715,0.0015110539,0.00002597747,0.0032522294,0.0013621558,0.005004468,0.16542771],"study_design_scores_gemma":[0.0031578743,0.0017762346,0.98859495,0.0028096905,0.00014439617,0.00003915202,0.00030965157,0.0022337937,0.00028440933,0.00024096417,0.00030567174,0.0001032102],"about_ca_topic_score_codex":0.00028617374,"about_ca_topic_score_gemma":0.000030605894,"teacher_disagreement_score":0.17118695,"about_ca_system_score_codex":0.000040725394,"about_ca_system_score_gemma":0.000092477625,"threshold_uncertainty_score":0.9824768},"labels":[],"label_agreement":null},{"id":"W2152496920","doi":"10.1007/s10459-006-9054-y","title":"The role of biomedical knowledge in diagnosis of difficult clinical cases","year":2007,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":162,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; McMaster University; University of Toronto","funders":"","keywords":"Terminology; MEDLINE; Medicine; Clinical diagnosis; Intensive care medicine","score_opus":0.03628295317185566,"score_gpt":0.49633138522047443,"score_spread":0.4600484320486188,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2152496920","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9530143,0.042776108,0.000018234758,0.0017337058,0.0009552244,0.0002736587,0.0000017421166,0.000005631856,0.0012213952],"genre_scores_gemma":[0.97885865,0.019593481,0.0011752676,0.00017937568,0.00013983685,0.00002717217,0.000004205375,0.0000027965648,0.00001919488],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9979519,0.00011747928,0.0011653268,0.00022202823,0.0002760673,0.00026716143],"domain_scores_gemma":[0.95965993,0.039394755,0.00040620638,0.00019177755,0.00014537883,0.00020193719],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0050623533,0.00006413047,0.00028512272,0.00015915497,0.00006628035,0.000003585268,0.00014986287,0.000059666996,0.000009230072],"category_scores_gemma":[0.047934704,0.00004060629,0.000043918633,0.0011689685,0.0009877685,0.00007336237,0.000024303505,0.00015949471,0.0000024809724],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000027908623,0.0009188975,0.58900106,0.00003230494,6.25068e-7,4.2877858e-7,0.00020644732,0.0000034682034,0.0000015131847,0.0012146132,0.000069625545,0.4085231],"study_design_scores_gemma":[0.00044991294,0.0011097207,0.9696234,0.002576055,0.0000047225444,0.000008970962,0.0046824804,0.00013263538,0.00018504965,0.0018410421,0.019332662,0.000053376818],"about_ca_topic_score_codex":0.0008409501,"about_ca_topic_score_gemma":0.0017143793,"teacher_disagreement_score":0.40846974,"about_ca_system_score_codex":0.00006903648,"about_ca_system_score_gemma":0.0015025457,"threshold_uncertainty_score":0.960085},"labels":[],"label_agreement":null},{"id":"W2154102994","doi":"10.1093/intqhc/mzm062","title":"Development of medical checklists for improved quality of patient care","year":2007,"lang":"en","type":"review","venue":"International Journal for Quality in Health Care","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":378,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; Sunnybrook Health Science Centre","funders":"","keywords":"Checklist; MEDLINE; Quality (philosophy); Evidence-based medicine; Systematic review; Medical literature; Task (project management); Medical education; Inclusion (mineral); Medicine; Computer science; Psychology; Alternative medicine; Pathology","score_opus":0.22430809677527153,"score_gpt":0.604112524744309,"score_spread":0.37980442796903746,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2154102994","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0021142566,0.9897677,0.001179987,0.0008981142,0.0029060002,0.0020157425,0.001063253,0.000011001386,0.000043969205],"genre_scores_gemma":[0.001542192,0.97705275,0.017244518,0.000800927,0.00079269684,0.00018614325,0.0023030413,0.000051405532,0.00002630157],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9912104,0.00029042093,0.006078069,0.0003459546,0.0017002375,0.0003749104],"domain_scores_gemma":[0.97915727,0.014035771,0.003364846,0.00025114522,0.0026849236,0.0005060523],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0062811016,0.00029157405,0.002592256,0.00034983968,0.00006300015,0.000014346628,0.00047171133,0.0005440017,0.000038985756],"category_scores_gemma":[0.06294465,0.00023185955,0.0011067201,0.00013606547,0.00008262567,0.000026517813,0.00009669078,0.0007122487,9.225777e-7],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004481952,0.00029618252,0.00013507798,0.040850807,0.00025010583,0.000005029626,0.0036394277,4.5226804e-7,1.5986566e-7,0.00052517373,0.00022028072,0.9536291],"study_design_scores_gemma":[0.0033161636,0.0006523251,0.0003521678,0.11667189,0.000096370095,0.000051129082,0.003319072,0.0000034775708,0.000013376475,0.0000714728,0.87522167,0.00023091506],"about_ca_topic_score_codex":0.0003691035,"about_ca_topic_score_gemma":0.00044242703,"teacher_disagreement_score":0.95339817,"about_ca_system_score_codex":0.0017185012,"about_ca_system_score_gemma":0.00874108,"threshold_uncertainty_score":0.99687845},"labels":[],"label_agreement":null},{"id":"W2155284013","doi":"10.3109/0142159x.2013.760036","title":"Script concordance testing: From theory to practice: AMEE Guide No. 75","year":2013,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":190,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; McGill University","funders":"","keywords":"Concordance; Competence (human resources); Medical education; Psychology; Test (biology); Variety (cybernetics); Medicine; Applied psychology; Management science; Computer science; Social psychology; Artificial intelligence","score_opus":0.03335734072977935,"score_gpt":0.36915876320728613,"score_spread":0.3358014224775068,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2155284013","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.582793,0.0012421423,0.0092700925,0.057974543,0.0028454775,0.0012484482,0.000009090505,0.0006060692,0.34401113],"genre_scores_gemma":[0.5113795,0.00012701703,0.108333066,0.15414238,0.008660712,0.0004605853,0.000104342726,0.00018885834,0.21660358],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9971939,0.00026518825,0.0005798086,0.000500883,0.001008289,0.0004519395],"domain_scores_gemma":[0.959431,0.03567024,0.00022139745,0.0011367521,0.0008527678,0.0026878295],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0016413954,0.00021279318,0.00048075782,0.000039908864,0.000059299386,0.000042001782,0.00026065216,0.00027622998,0.021927929],"category_scores_gemma":[0.6780049,0.00015454367,0.00010193214,0.00021575134,0.00023744299,0.00009070096,0.00014322644,0.00080957764,0.013109857],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001403879,0.000771646,0.067957655,0.000009645696,0.00012462893,0.0002734477,0.0003066287,4.927101e-7,0.00020282064,0.00019524826,0.7252444,0.20477301],"study_design_scores_gemma":[0.0034186523,0.00091830466,0.06537399,0.0031860936,0.00032879447,0.000097958175,0.0006643305,0.0019008713,0.00007914047,0.0028778906,0.9207127,0.0004412386],"about_ca_topic_score_codex":0.0017235326,"about_ca_topic_score_gemma":0.0000059247714,"teacher_disagreement_score":0.6763635,"about_ca_system_score_codex":0.000077464036,"about_ca_system_score_gemma":0.0003210803,"threshold_uncertainty_score":0.98765856},"labels":[],"label_agreement":null},{"id":"W2155301421","doi":"10.1016/j.jclinepi.2008.08.014","title":"A pragmatic resolution","year":2009,"lang":"en","type":"article","venue":"Journal of Clinical Epidemiology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute for Clinical Evaluative Sciences; University of Toronto; Sunnybrook Hospital; University of Victoria; Canadian College of Naturopathic Medicine","funders":"","keywords":"Scopus; Resolution (logic); Psychology; Medicine; MEDLINE; Computer science; Political science; Artificial intelligence; Law","score_opus":0.21719798640587784,"score_gpt":0.5402549566568264,"score_spread":0.3230569702509486,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2155301421","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.81910753,0.0022328775,0.014689482,0.15298225,0.0020428377,0.00021950965,0.0000017240044,0.000037484268,0.008686283],"genre_scores_gemma":[0.9315495,0.0014882986,0.03556929,0.028730422,0.00234743,6.6959564e-7,0.0000035401572,0.000007342145,0.00030349471],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9926478,0.0019016468,0.0047968985,0.00017566522,0.00019583598,0.0002821446],"domain_scores_gemma":[0.70872647,0.28571123,0.0037237054,0.00042032366,0.0004909548,0.0009273435],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.026034504,0.00010801378,0.0023508172,0.00008317752,0.000025524942,0.0000027847802,0.000120481156,0.00036212508,0.00012494987],"category_scores_gemma":[0.8499952,0.00006904891,0.00089091586,0.00011507489,0.00016482749,0.000044472017,0.00001465242,0.0010824421,0.000071996285],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0028117963,0.0022739933,0.45916063,0.000025809692,0.00023781417,0.0005019253,0.00003976533,0.00017620802,0.00003564822,0.015177219,0.13054244,0.38901678],"study_design_scores_gemma":[0.0030874684,0.0063786777,0.9247086,0.0012066986,0.00022327229,0.0009394012,0.000011230057,0.0010370774,0.00000292042,0.04927791,0.013053897,0.00007282287],"about_ca_topic_score_codex":0.0000036275273,"about_ca_topic_score_gemma":5.4405416e-7,"teacher_disagreement_score":0.8239607,"about_ca_system_score_codex":0.000026319185,"about_ca_system_score_gemma":0.0002903794,"threshold_uncertainty_score":0.9023092},"labels":[],"label_agreement":null},{"id":"W2155597260","doi":"10.1016/s0004-9514(08)70005-9","title":"Expert therapists use specific clinical reasoning processes in the assessment and management of patients with shoulder pain: a qualitative study","year":2008,"lang":"en","type":"article","venue":"Australian Journal of Physiotherapy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":62,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Canadian Patient Safety Institute","keywords":"Context (archaeology); Delphi method; Narrative; Clinical judgment; Psychology; Deductive reasoning; Medicine; Computer science; Artificial intelligence; Medical physics","score_opus":0.11749879996270743,"score_gpt":0.4762836684424098,"score_spread":0.35878486847970237,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2155597260","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9981694,0.00022277066,0.00014578253,0.0008219797,0.000046646288,0.00056954694,0.0000025513762,0.0000039750244,0.000017300006],"genre_scores_gemma":[0.9927868,0.0024820939,0.0041244137,0.00041960267,0.00009266075,0.000011820379,0.0000028348834,0.000018117204,0.000061649305],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99759096,0.00066480844,0.0007603449,0.0001952637,0.00062107947,0.00016757225],"domain_scores_gemma":[0.9949528,0.003947129,0.00046615335,0.0002249765,0.00030195215,0.00010699627],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0016621809,0.00016206439,0.00051657786,0.00007090959,0.00004793966,0.000022759374,0.00013357452,0.000044777644,0.000032091233],"category_scores_gemma":[0.00046871437,0.000086391185,0.00009399155,0.00027222832,0.00020679247,0.00012522656,0.000009666614,0.00029524812,2.8113902e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0015380436,0.0063868365,0.9426137,0.00006140462,0.00048121138,0.00027628595,0.040076204,0.0000090252015,0.000013790127,0.000024668818,0.001751966,0.0067668674],"study_design_scores_gemma":[0.007452025,0.0058635953,0.96291703,0.0021384917,0.000040594976,0.000016615757,0.021099731,0.000001459206,0.000013959586,0.000029054587,0.00032893434,0.00009853479],"about_ca_topic_score_codex":0.000022859285,"about_ca_topic_score_gemma":0.0000041526782,"teacher_disagreement_score":0.020303307,"about_ca_system_score_codex":0.00003014427,"about_ca_system_score_gemma":0.00009887762,"threshold_uncertainty_score":0.352293},"labels":[],"label_agreement":null},{"id":"W2156137524","doi":"10.1177/0734282914562212","title":"What Comes Before Report Writing? Attending to Clinical Reasoning and Thinking Errors in School Psychology","year":2014,"lang":"en","type":"article","venue":"Journal of Psychoeducational Assessment","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":25,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Psychology; Heuristics; Clinical judgment; Deductive reasoning; Cognition; School psychology; Logical reasoning; Process (computing); Psychology of reasoning; Critical thinking; Cognitive psychology; Applied psychology; Verbal reasoning; Mathematics education; Artificial intelligence; Computer science","score_opus":0.04265540508168172,"score_gpt":0.4968014186378206,"score_spread":0.4541460135561389,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2156137524","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.92274284,0.0002145503,0.0011437031,0.07112176,0.0033791682,0.00017827762,6.268777e-7,0.000009983948,0.0012091054],"genre_scores_gemma":[0.93114555,0.00025821803,0.058522966,0.008075907,0.0017654803,0.000011227217,0.000011383008,0.000019694004,0.00018959267],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9969478,0.00019186303,0.0016637936,0.00034726353,0.00059330807,0.00025596903],"domain_scores_gemma":[0.9942797,0.0037359304,0.00073811173,0.00028609406,0.00036234685,0.00059786],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0056760833,0.00015560332,0.00058583514,0.0003148087,0.00007620071,0.000102278005,0.00015016229,0.00013266804,0.00012703739],"category_scores_gemma":[0.020329578,0.00012794806,0.00016756552,0.00024072354,0.0000740598,0.0002305596,0.000044334814,0.00078710023,0.000010555335],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009894126,0.00065133336,0.968897,0.000012856546,0.00006631,0.00009045905,0.00020047156,0.000014601283,0.000016491422,0.0013583117,0.006899445,0.02169377],"study_design_scores_gemma":[0.0017553295,0.00085088727,0.9812795,0.0042212927,0.00006749391,0.0017965069,0.0011006037,0.000068368216,0.0000016299194,0.004425746,0.0043223393,0.000110302295],"about_ca_topic_score_codex":0.000010175072,"about_ca_topic_score_gemma":0.000015467951,"teacher_disagreement_score":0.06304585,"about_ca_system_score_codex":0.00015566271,"about_ca_system_score_gemma":0.0005147712,"threshold_uncertainty_score":0.9879226},"labels":[],"label_agreement":null},{"id":"W2157202661","doi":"10.1111/j.1365-2929.2005.02092.x","title":"Assessment in the context of uncertainty: how many members are needed on the panel of reference of a script concordance test?","year":2005,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":151,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Concordance; Cronbach's alpha; Statistics; Context (archaeology); Sample size determination; Reliability (semiconductor); Panel analysis; Test (biology); Panel data; Correlation; Sample (material); Medicine; Kappa; Mathematics; Psychometrics; Internal medicine","score_opus":0.05041126147914685,"score_gpt":0.3757654797721233,"score_spread":0.32535421829297645,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2157202661","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.83577096,0.00017229767,0.000017700135,0.16078275,0.00014591146,0.000296978,0.0000067815035,0.000004756522,0.0028018516],"genre_scores_gemma":[0.98420817,0.00014042592,0.000118505595,0.015177789,0.00009822917,0.000048468573,0.000017685397,0.0000050736576,0.0001856536],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99851567,0.00013754149,0.00040829007,0.00013950073,0.000681856,0.00011716403],"domain_scores_gemma":[0.9859961,0.012955512,0.00034399956,0.00038358214,0.00021781916,0.00010297236],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0011373485,0.00008381788,0.00030536452,0.00004418546,0.000017426026,0.0000041897747,0.00019365561,0.00008867334,0.00024631282],"category_scores_gemma":[0.11667757,0.000044000466,0.000057138313,0.00021820999,0.0002997652,0.000018311612,0.000016187201,0.00033036913,0.000002607042],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00053762697,0.014903116,0.21620592,0.0004599239,0.00012795483,0.000011222057,0.012239727,0.000048244732,0.0013640854,0.043077108,0.330138,0.38088706],"study_design_scores_gemma":[0.004108624,0.0016860971,0.864007,0.015876772,0.00020643478,0.000022794204,0.088886224,0.004264582,0.0021957713,0.00069386244,0.01784739,0.00020447644],"about_ca_topic_score_codex":0.0011142511,"about_ca_topic_score_gemma":0.00020761976,"teacher_disagreement_score":0.64780104,"about_ca_system_score_codex":0.000060942097,"about_ca_system_score_gemma":0.0013373811,"threshold_uncertainty_score":0.89076304},"labels":[],"label_agreement":null},{"id":"W2157225736","doi":"10.1111/medu.12635","title":"Influences of OSCE design on students’ diagnostic reasoning","year":2015,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":55,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval","funders":"","keywords":"Objective structured clinical examination; Task (project management); Context (archaeology); Cognition; Psychology; Applied psychology; Simulated patient; Perception; Medical education; Medicine; Psychiatry","score_opus":0.04751099940083737,"score_gpt":0.4220051603214645,"score_spread":0.37449416092062715,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2157225736","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9915273,0.00049644103,0.00062200736,0.004250243,0.0008257246,0.00027445817,6.911854e-7,0.000041450927,0.001961646],"genre_scores_gemma":[0.9929767,0.00010863222,0.0016768787,0.004480917,0.00033097324,0.000047341855,0.000019147217,0.000010779467,0.00034861043],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99803734,0.00011331671,0.00034572152,0.00019801524,0.0011512553,0.00015433875],"domain_scores_gemma":[0.98533255,0.01330784,0.00014106663,0.00030031402,0.0002840967,0.0006341369],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0012021934,0.00009449207,0.00024474316,0.00007127935,0.000022499566,0.000011025173,0.00015006037,0.00012529061,0.0002091145],"category_scores_gemma":[0.5120169,0.00007205748,0.000046647678,0.00020041651,0.00012664984,0.00003132225,0.000028885102,0.00020588562,0.00012618207],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00028420181,0.0052144253,0.7564945,0.000042980748,0.000055431814,0.000029938661,0.0011604468,0.000054428947,0.000014115292,0.00074896024,0.103944235,0.1319563],"study_design_scores_gemma":[0.0027068893,0.001796776,0.9790039,0.0094441,0.00019192854,0.00005231329,0.00074304605,0.0002969775,0.00062356284,0.0010723108,0.0038805252,0.00018765914],"about_ca_topic_score_codex":0.00015371492,"about_ca_topic_score_gemma":0.0000022306956,"teacher_disagreement_score":0.51081467,"about_ca_system_score_codex":0.00006628648,"about_ca_system_score_gemma":0.0021751828,"threshold_uncertainty_score":0.4920936},"labels":[],"label_agreement":null},{"id":"W2157525579","doi":"10.13096/rfre.v1n1.8","title":"Influence des ressources disponibles sur le raisonnement clinique : quelles preuves scientifiques ? Résultats d'une recension élargie","year":2015,"lang":"fr","type":"article","venue":"DOAJ (DOAJ: Directory of Open Access Journals)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Sherbrooke","funders":"","keywords":"Humanities; Political science; Philosophy","score_opus":0.37581163932329137,"score_gpt":0.5747159361881231,"score_spread":0.19890429686483174,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2157525579","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.749526,0.2413578,0.0002745156,0.0032619273,0.002158636,0.0009719666,0.00012249162,0.00009046885,0.0022361628],"genre_scores_gemma":[0.78400034,0.20999816,0.0015284675,0.00072460104,0.00070603384,0.000069290334,0.000055462286,0.000110606714,0.0028070076],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99092406,0.0018341937,0.002762545,0.0012866836,0.0019091389,0.0012833955],"domain_scores_gemma":[0.97560656,0.016370583,0.0020327556,0.0011827791,0.0028735907,0.0019337086],"candidate_categories":["metaresearch","metaepi_narrow","scholarly_communication","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.011070443,0.0007946236,0.002057706,0.00087938225,0.00077927124,0.0020328313,0.0027497862,0.0005181867,0.0019441325],"category_scores_gemma":[0.12810025,0.0006914808,0.00047930848,0.002242981,0.0024064127,0.003185399,0.0016036701,0.0011442875,0.00009928624],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018415477,0.0036662556,0.74965626,0.00080249674,0.0005285667,0.00041945753,0.004369993,0.0032685543,0.017009282,0.00028037463,0.1327201,0.08543712],"study_design_scores_gemma":[0.0040325336,0.0002086915,0.762611,0.063194476,0.00061099435,0.00007226752,0.0012537059,0.00041558544,0.068472244,0.020195762,0.07790338,0.0010294082],"about_ca_topic_score_codex":0.02351058,"about_ca_topic_score_gemma":0.0012332386,"teacher_disagreement_score":0.1170298,"about_ca_system_score_codex":0.0003794589,"about_ca_system_score_gemma":0.0018775534,"threshold_uncertainty_score":0.9995536},"labels":[],"label_agreement":null},{"id":"W2158021615","doi":"10.1023/b:ahse.0000038209.65714.d4","title":"How Can I Know What I Don't Know? Poor Self Assessment in a Well-Defined Domain","year":2004,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":300,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"University of Toronto","keywords":"Psychology; Test (biology); Status quo; Argument (complex analysis); Self-assessment; Norm (philosophy); Computer science; Social psychology; Mathematics education; Medicine; Epistemology","score_opus":0.015334665470981213,"score_gpt":0.4038341377515762,"score_spread":0.388499472280595,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2158021615","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.73694164,0.03124947,0.0002538872,0.22542262,0.0029049045,0.0011161345,0.0000022449003,0.00006828997,0.0020408109],"genre_scores_gemma":[0.91512495,0.03938557,0.039865423,0.0050085564,0.00022461178,0.00016952051,0.000023827062,0.000009886506,0.00018764952],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99787366,0.00011224007,0.00052440754,0.0005290986,0.0004528179,0.00050777255],"domain_scores_gemma":[0.99722445,0.0019125703,0.00024636337,0.00023309169,0.000081091595,0.00030244977],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0016115819,0.00015184375,0.00033560293,0.00029269795,0.00014893143,0.00011276904,0.000171735,0.00006993886,0.00002163142],"category_scores_gemma":[0.0030197501,0.00012653242,0.000041435927,0.0012224095,0.00020782185,0.00068483845,0.000025747346,0.00028992194,0.00000909387],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006087289,0.0033460183,0.43574685,0.00038992747,0.0000053548947,0.00001302628,0.0044210893,0.0015882727,0.000021109643,0.022691857,0.00047617036,0.53123945],"study_design_scores_gemma":[0.0085600745,0.0045918524,0.64248556,0.037860077,0.00003057211,0.00012885101,0.059832074,0.00079184916,0.00018113671,0.0870561,0.15756981,0.00091201504],"about_ca_topic_score_codex":0.0019935376,"about_ca_topic_score_gemma":0.0042134104,"teacher_disagreement_score":0.53032744,"about_ca_system_score_codex":0.0010124153,"about_ca_system_score_gemma":0.005725585,"threshold_uncertainty_score":0.999911},"labels":[],"label_agreement":null},{"id":"W2158828425","doi":"10.1111/j.1365-2923.2010.03863.x","title":"Script concordance testing: a review of published validity evidence","year":2011,"lang":"en","type":"review","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":170,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Centre de Santé et de Services Sociaux Cavendish; McGill University","funders":"","keywords":"Concordance; Psychology; External validity; Construct validity; Competence (human resources); PsycINFO; Content validity; Validity; Clinical psychology; Internal validity; MEDLINE; Social psychology; Applied psychology; Psychometrics; Medicine; Statistics; Mathematics","score_opus":0.22723637268626784,"score_gpt":0.47941607765817446,"score_spread":0.2521797049719066,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2158828425","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000004035141,0.9902592,0.00003548689,0.0019789168,0.001832844,0.0011063572,0.0000062400213,0.000058443286,0.004718492],"genre_scores_gemma":[0.0000057178077,0.9903753,0.0011863924,0.006196987,0.0008012505,0.0002914553,0.00021107093,0.00003638123,0.0008954693],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9961343,0.00039518723,0.0016881783,0.00058699114,0.0009252177,0.00027016233],"domain_scores_gemma":[0.94362557,0.046631712,0.002935354,0.0026868302,0.0024471488,0.0016733868],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0032460135,0.00032566985,0.0025739982,0.00011961824,0.000030361729,0.000014074718,0.00043640137,0.0004704903,0.0032747998],"category_scores_gemma":[0.8236926,0.00022779714,0.0005078718,0.00086519937,0.0002765864,0.00009118969,0.000096428084,0.0008902242,0.00017744898],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000003275338,0.000454291,0.0001430519,0.11870245,0.000024720983,0.00000478391,0.000008832647,2.0245378e-10,9.2916705e-9,0.00006941958,0.15849586,0.7220933],"study_design_scores_gemma":[0.000070180206,0.00012369992,0.00009973403,0.499073,0.0010095261,0.00010873001,0.0000015055749,7.211507e-7,1.4121206e-7,0.000041307634,0.49937508,0.00009640254],"about_ca_topic_score_codex":0.0002236591,"about_ca_topic_score_gemma":0.0000018985668,"teacher_disagreement_score":0.8204466,"about_ca_system_score_codex":0.00014635081,"about_ca_system_score_gemma":0.01876049,"threshold_uncertainty_score":0.9976363},"labels":[],"label_agreement":null},{"id":"W2160065345","doi":"10.1186/1472-6920-12-29","title":"Summative assessment of 5thyear medical students’ clinical reasoning by script concordance test: requirements and challenges","year":2012,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":54,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Summative assessment; Concordance; Multidisciplinary approach; Test (biology); Medical education; Curriculum; Educational measurement; Medicine; Academic year; Formative assessment; Mathematics education; Psychology; Internal medicine; Pedagogy","score_opus":0.0947932579973595,"score_gpt":0.48898982492034304,"score_spread":0.39419656692298355,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2160065345","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97197175,0.0134326555,0.000768687,0.0075072506,0.001780719,0.00049673574,0.000007355665,0.000053202377,0.0039816634],"genre_scores_gemma":[0.97720695,0.014182545,0.004513259,0.002448724,0.0010871514,0.00007042629,0.000075708915,0.000023645442,0.00039159978],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99552983,0.00035872223,0.0010009465,0.00037233846,0.0023820503,0.00035610082],"domain_scores_gemma":[0.979668,0.017244415,0.00041371427,0.00042215135,0.0002569871,0.0019947353],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005863891,0.0001777683,0.0005995808,0.000055155815,0.000053904354,0.000014988429,0.00024735605,0.0003484145,0.0007833898],"category_scores_gemma":[0.26442748,0.00014324437,0.00009784739,0.00012457199,0.00045932853,0.000110496105,0.00015693589,0.00049736426,0.000022127444],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003901588,0.0043893014,0.916778,0.000168901,0.0000538683,0.0000022225881,0.00025723575,1.3407612e-8,0.00000679359,0.0012171364,0.017461,0.05962647],"study_design_scores_gemma":[0.0026036324,0.00060353265,0.9831194,0.005967491,0.0001524052,0.00003840585,0.0010345706,0.00030008817,0.000031036474,0.00007697831,0.00591943,0.00015300846],"about_ca_topic_score_codex":0.000067451576,"about_ca_topic_score_gemma":0.000017292889,"teacher_disagreement_score":0.25856358,"about_ca_system_score_codex":0.0000837287,"about_ca_system_score_gemma":0.0028070372,"threshold_uncertainty_score":0.8577569},"labels":[],"label_agreement":null},{"id":"W2160357909","doi":"10.1111/j.1365-2923.2007.02923.x","title":"Teaching from the clinical reasoning literature: combined reasoning strategies help novice diagnosticians overcome misleading information","year":2007,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":213,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia; University of Toronto; McMaster University","funders":"","keywords":"Presentation (obstetrics); Psychology; Task (project management); Test (biology); Cognitive psychology; Confirmation bias; Analytic reasoning; Artificial intelligence; Computer science; Social psychology; Deductive reasoning; Medicine","score_opus":0.01432967744878596,"score_gpt":0.37418293863360114,"score_spread":0.3598532611848152,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2160357909","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9421144,0.0011768683,0.011506758,0.029150922,0.003452834,0.0006081,0.00001372816,0.00025815234,0.011718221],"genre_scores_gemma":[0.9622015,0.0002686108,0.0050981655,0.027967116,0.0032803959,0.000030562533,0.0008917931,0.000029039225,0.00023285937],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99655235,0.0002857258,0.001231782,0.00037330366,0.0010665129,0.0004903303],"domain_scores_gemma":[0.9586671,0.03887582,0.00051274267,0.00074711064,0.0004227226,0.000774525],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005204145,0.00026115982,0.00046579112,0.00011013132,0.00034824896,0.000264334,0.000300359,0.00048782318,0.00024406507],"category_scores_gemma":[0.29246688,0.00018525257,0.00020207852,0.00041321162,0.00020848926,0.0005505345,0.00008273276,0.0018188204,0.00014814633],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00052292185,0.0015276213,0.31815132,0.00007328206,0.00014908671,0.000064058775,0.008098763,0.0000063611024,0.00001895185,0.013262354,0.060538527,0.59758675],"study_design_scores_gemma":[0.003557266,0.00056252524,0.90232605,0.018756708,0.00040500314,0.00014083783,0.010236231,0.0025927601,0.000042145704,0.0016699497,0.059223343,0.0004871685],"about_ca_topic_score_codex":0.0018231838,"about_ca_topic_score_gemma":0.00016560145,"teacher_disagreement_score":0.59709954,"about_ca_system_score_codex":0.0001487572,"about_ca_system_score_gemma":0.0017752093,"threshold_uncertainty_score":0.7901971},"labels":[],"label_agreement":null},{"id":"W2160510017","doi":"10.1111/j.1365-2753.2009.01232.x","title":"‘We hold these truths to be self‐evident’: deconstructing ‘evidence‐based’ medical practice","year":2009,"lang":"en","type":"article","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":35,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University","funders":"","keywords":"Rigour; Evidence-based medicine; Epistemology; Scientific evidence; Intuition; Evidence-based practice; Politics; Medicine; Alternative medicine; Law; Philosophy; Political science","score_opus":0.2380824231568183,"score_gpt":0.5719931694614765,"score_spread":0.3339107463046582,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2160510017","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.34339848,0.0027334027,0.0020842312,0.64544773,0.0019281978,0.0007966129,0.000001547554,0.00004283612,0.0035669387],"genre_scores_gemma":[0.828747,0.0033295448,0.07535808,0.09063557,0.0018719869,0.000009592826,0.0000021067592,0.000021756503,0.000024368454],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9863468,0.0037811804,0.0043329056,0.0005093051,0.0046163974,0.00041345775],"domain_scores_gemma":[0.42625883,0.56086713,0.004416984,0.000942141,0.0055645155,0.0019504047],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.07605309,0.0002503541,0.0009727845,0.00035352365,0.00007343721,0.00012436725,0.00035543682,0.00048168146,0.00075547874],"category_scores_gemma":[0.9586885,0.00019957715,0.00046533998,0.0008423219,0.00010894635,0.0015420333,0.000051585524,0.002485262,0.00009442647],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.012807745,0.012414592,0.047733482,0.000066110886,0.00052681234,0.0036841761,0.0018811489,0.0019650399,0.00008459642,0.0010821797,0.028105257,0.88964885],"study_design_scores_gemma":[0.06168475,0.04711959,0.43979248,0.0775374,0.018682947,0.026844269,0.015785558,0.08910742,0.00038008788,0.009739003,0.21111311,0.0022133652],"about_ca_topic_score_codex":0.00002486796,"about_ca_topic_score_gemma":0.0000072041066,"teacher_disagreement_score":0.8874355,"about_ca_system_score_codex":0.00035998647,"about_ca_system_score_gemma":0.0043415925,"threshold_uncertainty_score":0.99981606},"labels":[],"label_agreement":null},{"id":"W2161126681","doi":"10.1111/j.1365-2753.2012.01896.x","title":"Reason and value: making reasoning fit for practice","year":2012,"lang":"en","type":"editorial","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":20,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Guelph; Dalhousie University; De Veber; University of Toronto","funders":"Wellcome Trust","keywords":"Value (mathematics); Psychology; Computer science; Medicine; Machine learning","score_opus":0.12477927970058793,"score_gpt":0.559299578907684,"score_spread":0.43452029920709606,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2161126681","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0017467781,0.02283295,0.0017477467,0.019074276,0.94436127,0.0020719308,0.000041023548,0.000030011292,0.008094034],"genre_scores_gemma":[0.007990759,0.010318941,0.1174546,0.0024750151,0.8612135,0.000058268506,0.000086274944,0.00011564737,0.00028697908],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9880345,0.0030068422,0.004396002,0.0005855952,0.0034960206,0.0004810313],"domain_scores_gemma":[0.14623064,0.82954633,0.011213447,0.000757711,0.011509424,0.0007424259],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["metaresearch","research_integrity"],"category_scores_codex":[0.09415234,0.00037380712,0.0016294422,0.00028843476,0.00010473526,0.00014404906,0.00019288194,0.0017755021,0.00009922371],"category_scores_gemma":[0.98036236,0.00032030893,0.00054815656,0.00029628677,0.00015065864,0.0011329872,0.00010413034,0.004541475,0.000018304449],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008553794,0.0018982793,0.0012886856,0.00016388981,0.0005525764,0.00014028011,0.00018230782,0.00006994591,8.009661e-7,0.00019298229,0.90358,0.08337645],"study_design_scores_gemma":[0.010106111,0.002682176,0.0035331638,0.011038552,0.010076513,0.00070518244,0.00043009827,0.0020368833,8.1006095e-7,0.0006722692,0.95840204,0.00031621678],"about_ca_topic_score_codex":0.000038680962,"about_ca_topic_score_gemma":0.000004750302,"teacher_disagreement_score":0.88621,"about_ca_system_score_codex":0.00047110592,"about_ca_system_score_gemma":0.003039422,"threshold_uncertainty_score":0.9999249},"labels":[],"label_agreement":null},{"id":"W2161954364","doi":"10.12927/hcq.2010.21817","title":"Skill Development and Capacity Building Program for Best Practices in Rehabilitation","year":2010,"lang":"en","type":"article","venue":"Healthcare Quarterly","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Rehabilitation Institute","funders":"","keywords":"Best practice; Rehabilitation; Capacity building; Medical education; Process management; Nursing; Medicine; Psychology; Business; Physical therapy; Political science","score_opus":0.06292552256220073,"score_gpt":0.4260083712909386,"score_spread":0.36308284872873786,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2161954364","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9837609,0.00009013804,0.00006367356,0.014436214,0.0002595222,0.00123622,0.0000042401134,0.000058281636,0.00009078725],"genre_scores_gemma":[0.84945196,0.00000547555,0.14972061,0.0002874625,0.00012765198,0.00034746595,0.000016606507,0.000011686317,0.000031085514],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9988461,0.000046565092,0.0003905998,0.0003070364,0.00014915902,0.00026050836],"domain_scores_gemma":[0.9918686,0.007353076,0.00019244905,0.00017371324,0.00017581988,0.00023631948],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0008349434,0.000102851685,0.00023361482,0.00008000762,0.00007622682,0.000027406732,0.00003417674,0.00014979388,0.0000047613553],"category_scores_gemma":[0.022027528,0.00009092283,0.000034104833,0.0001069809,0.00006453593,0.00008169546,0.0000046213886,0.00032831822,0.000005695034],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000116397845,0.0005364632,0.12350609,0.00032825663,0.000007195028,0.0000053829644,0.00526417,5.049409e-8,0.00015244272,0.0014560008,0.00007369139,0.8685539],"study_design_scores_gemma":[0.003584287,0.011347688,0.9430048,0.0025894134,0.000045942998,0.00008121731,0.0044573965,0.0002822902,0.00022217429,0.002725382,0.031356238,0.00030315673],"about_ca_topic_score_codex":0.0011360096,"about_ca_topic_score_gemma":0.004400524,"teacher_disagreement_score":0.8682507,"about_ca_system_score_codex":0.000046948073,"about_ca_system_score_gemma":0.00027366268,"threshold_uncertainty_score":0.98621035},"labels":[],"label_agreement":null},{"id":"W2162089719","doi":"","title":"Problems for clinical judgement: 1. Eliciting an insightful history of present illness.","year":2001,"lang":"en","type":"article","venue":"PubMed","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":35,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Recall; Comprehension; Judgement; Forgetting; Expression (computer science); Cognitive psychology; Clinical judgement; Psychology; Computer science; Medicine; Epistemology","score_opus":0.15754180292416167,"score_gpt":0.3662018001953982,"score_spread":0.20865999727123655,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2162089719","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96597946,0.0025142166,0.0006409521,0.0021656845,0.0015650954,0.0028117122,0.000009563646,0.0001323126,0.024181025],"genre_scores_gemma":[0.99103904,0.0005257932,0.0004265275,0.0009800232,0.00094570604,0.001536237,0.000048770984,0.000030259884,0.0044676512],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99820787,0.00006738908,0.0007610284,0.0003363522,0.00026044462,0.00036692477],"domain_scores_gemma":[0.9944022,0.0044198567,0.00023173283,0.0003706767,0.00019641033,0.00037908374],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0012945882,0.00011949318,0.00044291132,0.00006723324,0.000026740265,0.000006117021,0.00010920377,0.00012866533,0.00005570361],"category_scores_gemma":[0.024438513,0.00009917592,0.00019796894,0.00008830652,0.00014055704,0.000058454014,0.00004103651,0.00015433466,0.0000063160987],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003949911,0.0014989062,0.08973443,0.000091429654,0.000096502474,0.000016522152,0.00029297284,0.000012554104,0.000009024575,0.00028241088,0.03471162,0.87285864],"study_design_scores_gemma":[0.004407184,0.00047385873,0.4135542,0.00026575808,0.0002267456,0.000014663746,0.00007688484,0.00053999416,0.00007030939,0.0005415908,0.5796415,0.0001872787],"about_ca_topic_score_codex":0.00007475769,"about_ca_topic_score_gemma":0.000009259402,"teacher_disagreement_score":0.87267137,"about_ca_system_score_codex":0.00012854597,"about_ca_system_score_gemma":0.00011779799,"threshold_uncertainty_score":0.9837791},"labels":[],"label_agreement":null},{"id":"W2162360374","doi":"10.1136/bmjqs-2012-001712","title":"Cognitive debiasing 1: origins of bias and theory of debiasing","year":2013,"lang":"en","type":"article","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":590,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Debiasing; Heuristics; Variety (cybernetics); Cognition; Cognitive bias; Computer science; Cognitive psychology; Process (computing); Dual process theory (moral psychology); Psychology; Artificial intelligence; Social psychology","score_opus":0.13518600231153577,"score_gpt":0.4334345885807633,"score_spread":0.2982485862692275,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2162360374","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9920935,0.0003255382,0.0024041617,0.0008248419,0.00008589521,0.00053563464,0.000037274793,0.000029734418,0.0036634083],"genre_scores_gemma":[0.9976167,0.00014360079,0.0014656471,0.0005227928,0.00008224274,0.000009253997,0.000027178423,0.000014844505,0.00011770627],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99773127,0.00043666142,0.0010912418,0.00023693532,0.0003001778,0.00020369326],"domain_scores_gemma":[0.9468778,0.051700745,0.00044206242,0.00030494342,0.00044913305,0.00022533185],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0028786988,0.00013500375,0.00061563123,0.000053865042,0.000047739733,0.000008601054,0.00005088281,0.000116726915,0.0002261123],"category_scores_gemma":[0.114949055,0.00010575906,0.00014317557,0.00015241196,0.00037158074,0.00006497437,0.0000569283,0.00016307227,0.000019900084],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019487869,0.0008623293,0.5390735,0.0005158346,0.00034451272,0.000016960888,0.003215394,0.000002935445,0.0029671856,0.031098863,0.00035866007,0.41959503],"study_design_scores_gemma":[0.0029167747,0.00028460307,0.9769007,0.005004157,0.00025346244,0.000028494818,0.0036222807,0.00007271852,0.0049061845,0.0057055443,0.00012468504,0.00018041822],"about_ca_topic_score_codex":0.0014107173,"about_ca_topic_score_gemma":0.000016534657,"teacher_disagreement_score":0.43782717,"about_ca_system_score_codex":0.000043954307,"about_ca_system_score_gemma":0.00020774656,"threshold_uncertainty_score":0.8925061},"labels":[],"label_agreement":null},{"id":"W2163361764","doi":"10.1046/j.1365-2923.2001.01090.x","title":"A schematic approach to diagnosing and resolving lecturalgia","year":2001,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":63,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Calgary","funders":"","keywords":"Schematic; Judgement; Presentation (obstetrics); License; Expert opinion; Subcategory; Point (geometry); Curriculum; Psychology; Medical education; Clinical judgement; Computer science; Medicine; Pedagogy; Family medicine; Epistemology","score_opus":0.02049714229768936,"score_gpt":0.3604846469119443,"score_spread":0.339987504614255,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2163361764","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9600258,0.000857962,0.002644582,0.025746634,0.0002653281,0.00028914725,1.5993741e-7,0.00005817868,0.010112233],"genre_scores_gemma":[0.9773919,0.00015560957,0.008328936,0.013015863,0.00038814795,0.0000674997,0.000020547723,0.000012274295,0.00061922317],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9989844,0.000031032443,0.00022194188,0.00022583592,0.00035992434,0.00017684897],"domain_scores_gemma":[0.9971189,0.0021254253,0.000033246506,0.00017352424,0.000064252956,0.00048463885],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00040583033,0.00008337352,0.00018880425,0.00007057426,0.00005758919,0.000023815075,0.000049974107,0.000098226286,0.00020553621],"category_scores_gemma":[0.14391887,0.000064773274,0.000031053827,0.00023866954,0.000055174554,0.000030434416,0.000030072824,0.00018658607,0.000049426457],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009166376,0.002387804,0.13083173,0.00021021554,0.000036648034,0.000024182036,0.0023295216,0.0000026271716,0.000112856316,0.001186865,0.10581781,0.7569681],"study_design_scores_gemma":[0.0038348865,0.00079177576,0.8998378,0.020526472,0.00039794325,0.002528024,0.0018026391,0.009802471,0.00031564935,0.002221252,0.057200145,0.000740953],"about_ca_topic_score_codex":0.000084555584,"about_ca_topic_score_gemma":0.0000028821466,"teacher_disagreement_score":0.7690061,"about_ca_system_score_codex":0.000045692603,"about_ca_system_score_gemma":0.0004558472,"threshold_uncertainty_score":0.8632923},"labels":[],"label_agreement":null},{"id":"W2164105380","doi":"10.1046/j.1365-2923.2001.00911.x","title":"Measurement of clinical reflective capacity early in training as a predictor of clinical reasoning performance at the end of residency: an experimental study on the script concordance test","year":2001,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":128,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université de Sherbrooke; Université de Montréal; Université Laval","funders":"","keywords":"Concordance; Test (biology); Pearson product-moment correlation coefficient; Objective structured clinical examination; Psychology; Predictive validity; Medicine; Medical education; Cohort; Clinical psychology; Statistics; Mathematics; Pathology; Internal medicine","score_opus":0.15579555954710375,"score_gpt":0.4619869817728976,"score_spread":0.3061914222257939,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2164105380","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9937863,0.00025171638,0.0000053411495,0.0023019386,0.0006756028,0.0008249465,0.000003289815,0.000012414856,0.0021384882],"genre_scores_gemma":[0.9983648,0.00010215428,0.00011568413,0.0008144689,0.00040660304,0.0000963739,0.000004696977,0.000015106246,0.000080129175],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9946324,0.00094998244,0.002023929,0.00041702532,0.0017446623,0.00023200671],"domain_scores_gemma":[0.9863505,0.0115723945,0.00076223974,0.0006246259,0.00042253075,0.00026772477],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.011588489,0.00015993715,0.00076542696,0.00006150213,0.000075921664,0.000005484713,0.00027728215,0.00019008805,0.00026025312],"category_scores_gemma":[0.27803487,0.00009404339,0.00016182382,0.00032744207,0.0010080028,0.00006293034,0.000060167004,0.00076806266,0.0000058198884],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012958416,0.010302049,0.9634563,0.000016109518,0.000052715863,0.000005053632,0.008065216,0.0000013645617,0.0002065469,0.00011824789,0.00048752935,0.015993051],"study_design_scores_gemma":[0.0026139123,0.009195168,0.9750518,0.00360511,0.00008484991,0.000015258239,0.0075109806,0.00026834977,0.0014444812,0.000059536094,0.00007202685,0.00007851787],"about_ca_topic_score_codex":0.0012275755,"about_ca_topic_score_gemma":0.00039343795,"teacher_disagreement_score":0.26644638,"about_ca_system_score_codex":0.00012967554,"about_ca_system_score_gemma":0.003346395,"threshold_uncertainty_score":0.72804654},"labels":[],"label_agreement":null},{"id":"W2164610426","doi":"10.1177/0163278704267043","title":"Standardized Assessment of Reasoning in Contexts of Uncertainty","year":2004,"lang":"en","type":"article","venue":"Evaluation & the Health Professions","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":153,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Psychology; Standardized test; Medical physics; Computer science; Natural language processing; Medical education; Mathematics education; Medicine","score_opus":0.10049369076300795,"score_gpt":0.5254536042443122,"score_spread":0.42495991348130424,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2164610426","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.962121,0.00061406294,0.00156059,0.032174155,0.0003010174,0.0019755373,0.000020527254,0.000017357397,0.0012157416],"genre_scores_gemma":[0.99612856,0.00012622627,0.0026727933,0.00082539377,0.00003233995,0.00010565741,0.000037178623,0.0000080583395,0.00006377699],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99740607,0.0007308756,0.0007802877,0.00015161566,0.00074062345,0.00019052559],"domain_scores_gemma":[0.99427545,0.0043633427,0.00046404012,0.00032793777,0.00045542998,0.00011378423],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0074756294,0.00007780342,0.0004125875,0.000075530246,0.00010772293,0.0000019472368,0.00006725996,0.000067645684,0.00015286703],"category_scores_gemma":[0.028282883,0.000047618647,0.00007337005,0.000303967,0.000095081145,0.000023368011,0.000032758624,0.0002966629,0.000002428985],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0038262962,0.00851727,0.43907312,0.0014289764,0.00041217226,0.000014692361,0.04592658,0.06241105,0.0021128557,0.10596884,0.0076221153,0.32268602],"study_design_scores_gemma":[0.016326826,0.0011704166,0.93878645,0.02097092,0.00013762034,0.0000047734375,0.0036450343,0.00737332,0.00021507601,0.011135766,0.0001309449,0.0001028629],"about_ca_topic_score_codex":0.0005610876,"about_ca_topic_score_gemma":0.00016840313,"teacher_disagreement_score":0.4997133,"about_ca_system_score_codex":0.00035192372,"about_ca_system_score_gemma":0.004706956,"threshold_uncertainty_score":0.9799023},"labels":[],"label_agreement":null},{"id":"W2165166892","doi":"10.1002/chp.20076","title":"Poorly performing physicians: Does the script concordance test detect bad clinical reasoning? *","year":2010,"lang":"en","type":"article","venue":"Journal of Continuing Education in the Health Professions","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":33,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université de Montréal; Centrale des Syndicats du Québec","funders":"","keywords":"Concordance; Cronbach's alpha; Intraclass correlation; Kappa; Competence (human resources); Medicine; Cohen's kappa; Clinical psychology; Family medicine; Psychology; Psychometrics; Internal medicine; Statistics; Social psychology","score_opus":0.024501722991541876,"score_gpt":0.43560330200620423,"score_spread":0.4111015790146624,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2165166892","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.91421115,0.0007852523,0.0000789821,0.07736787,0.0065579573,0.000527932,0.000002224955,0.00001385555,0.000454748],"genre_scores_gemma":[0.9815134,0.00047730567,0.002755949,0.011514375,0.002602013,0.000025218376,0.0000023938455,0.000017867978,0.0010915177],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9970255,0.0006296451,0.0014302261,0.0001689854,0.00041272884,0.0003329369],"domain_scores_gemma":[0.966179,0.030461757,0.0018433451,0.0005946991,0.0006499464,0.0002712178],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.007857332,0.00013682201,0.00048955315,0.000100499805,0.00033393435,0.00004418949,0.00036026078,0.0001285725,0.00003561304],"category_scores_gemma":[0.10261358,0.00006181655,0.00020184956,0.00030872188,0.00020250815,0.00012290067,0.000032340966,0.0026647076,0.000010016996],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012874007,0.0010300345,0.46159127,0.000043872857,0.00001708678,0.000006318641,0.009081708,5.3781946e-7,0.00020284962,0.000404328,0.029293133,0.49820012],"study_design_scores_gemma":[0.0014486518,0.0007803056,0.9377668,0.01600819,0.00009188162,0.00027338808,0.02217564,0.00008775622,0.00006333384,0.0011704545,0.020017797,0.000115790965],"about_ca_topic_score_codex":0.00009055178,"about_ca_topic_score_gemma":0.0001492753,"teacher_disagreement_score":0.4980843,"about_ca_system_score_codex":0.000059377668,"about_ca_system_score_gemma":0.002539309,"threshold_uncertainty_score":0.9996362},"labels":[],"label_agreement":null},{"id":"W2167003269","doi":"10.1136/emermed-2011-200468","title":"Experiential and rational decision making: a survey to determine how emergency physicians make clinical decisions","year":2011,"lang":"en","type":"article","venue":"Emergency Medicine Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":30,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Ottawa Hospital; Dalhousie University; University of Ottawa","funders":"","keywords":"Medicine; Experiential learning; Descriptive statistics; Likert scale; Family medicine; Psychology; Statistics","score_opus":0.18799926331613917,"score_gpt":0.4473193372139271,"score_spread":0.2593200738977879,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2167003269","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9487096,0.0009249076,0.03257077,0.0024123602,0.010784609,0.00036876314,0.000019101359,0.000043717366,0.0041661635],"genre_scores_gemma":[0.9796356,0.0030548414,0.00599014,0.0003130993,0.004103213,0.000020233974,0.000029925455,0.000055717028,0.0067972136],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99561834,0.00031448365,0.0018642928,0.0005889491,0.0010393804,0.0005745535],"domain_scores_gemma":[0.9931959,0.0037830877,0.00049238926,0.0005052376,0.0008345531,0.0011888334],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0031614862,0.00034148924,0.0008509492,0.00030267655,0.0002965785,0.000012779905,0.0002619807,0.00019890365,0.028958896],"category_scores_gemma":[0.13561064,0.00024937655,0.00031634863,0.0005865514,0.00017600956,0.00011121849,0.00014147931,0.0007047075,0.000056087472],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012576102,0.0007100966,0.47020376,0.0000059247286,0.00013258094,0.000061778715,0.0019180819,0.0000012051976,0.00014948261,0.00004908279,0.44001374,0.08549664],"study_design_scores_gemma":[0.0025048698,0.0024994568,0.97891855,0.0014889803,0.00024473225,0.00044114888,0.0004976723,0.00017525883,0.000017147397,0.0010175154,0.011914014,0.0002806784],"about_ca_topic_score_codex":0.00004317225,"about_ca_topic_score_gemma":0.00023246485,"teacher_disagreement_score":0.50871474,"about_ca_system_score_codex":0.00004346733,"about_ca_system_score_gemma":0.000160208,"threshold_uncertainty_score":0.9999958},"labels":[],"label_agreement":null},{"id":"W2168809148","doi":"10.1007/s10459-006-9055-x","title":"It all make sense: biomedical knowledge, causal connections and memory in the novice diagnostician","year":2007,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":142,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; McMaster University; University of Toronto","funders":"","keywords":"Casual; Causal model; Psychology; Medical knowledge; Causal analysis; Cognitive psychology; Computer science; Medical education; Medicine; Pathology","score_opus":0.032253403750275335,"score_gpt":0.45921377999152907,"score_spread":0.4269603762412537,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2168809148","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.871635,0.020096816,0.0003612521,0.09017543,0.002713145,0.00087691203,0.0000037603434,0.00003174808,0.014105939],"genre_scores_gemma":[0.97909987,0.004429312,0.001526821,0.014424991,0.00037646564,0.000041699896,0.000014616114,0.000004666128,0.000081561135],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9983829,0.000121357254,0.00050605356,0.00032544913,0.00028313708,0.00038107406],"domain_scores_gemma":[0.97856855,0.020811813,0.00013141845,0.00017224431,0.000065371285,0.00025058098],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.003971033,0.000092802984,0.00018718922,0.00025023037,0.00016353582,0.000020030251,0.0000927335,0.000063077125,0.000018308698],"category_scores_gemma":[0.02308201,0.000065481836,0.000018291586,0.0010415209,0.0005400027,0.00010790411,0.000019635436,0.0002576823,0.000012160774],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000059869948,0.0025165286,0.27958855,0.00024475675,0.0000039529314,0.000029363338,0.013591266,0.000055777175,0.000009857806,0.0060444484,0.0044118585,0.6934438],"study_design_scores_gemma":[0.0010370148,0.0010388397,0.80621296,0.0029121845,0.000017280618,0.00027987608,0.020313643,0.00051996333,0.00001560626,0.0023694995,0.16509421,0.00018893796],"about_ca_topic_score_codex":0.0011763613,"about_ca_topic_score_gemma":0.009975774,"teacher_disagreement_score":0.6932548,"about_ca_system_score_codex":0.00011676319,"about_ca_system_score_gemma":0.0010502043,"threshold_uncertainty_score":0.985147},"labels":[],"label_agreement":null},{"id":"W2169217275","doi":"10.1111/j.1365-2923.2010.03705.x","title":"How clinical features are presented matters to weaker diagnosticians","year":2010,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":30,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Ottawa; Medical Council of Canada; University of British Columbia","funders":"","keywords":"Terminology; Medical diagnosis; Test (biology); Context (archaeology); Medical terminology; Psychology; Unified Medical Language System; Language assessment; Sample (material); Medical education; Medicine; Family medicine; Linguistics; Computer science; Artificial intelligence; Nursing; Mathematics education; Radiology","score_opus":0.015691738964612397,"score_gpt":0.3921015967209331,"score_spread":0.3764098577563207,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2169217275","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.47772625,0.00007061328,0.00019614874,0.51657224,0.0042336676,0.00034330558,0.000004281742,0.0000809641,0.00077255204],"genre_scores_gemma":[0.82284015,0.00008314413,0.0018961456,0.1632621,0.0040329527,0.00012983289,0.000107346255,0.000035493085,0.0076128263],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9980724,0.00008187216,0.0003894638,0.000417059,0.000729285,0.00030995687],"domain_scores_gemma":[0.9888885,0.00869815,0.00013392202,0.00060602045,0.00027357184,0.0013998442],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00066904747,0.00015586153,0.0003645861,0.00008983652,0.000060844286,0.000059857677,0.00017932644,0.0003815832,0.00079816533],"category_scores_gemma":[0.38397437,0.00011984503,0.00015005213,0.00023222655,0.0001922313,0.000044636792,0.000057677855,0.00097249483,0.00031074512],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000052508898,0.0012695327,0.17693584,0.000014940718,0.000022299278,0.000025654499,0.00007654943,9.900272e-8,0.000036761892,0.00014286708,0.7305755,0.09084742],"study_design_scores_gemma":[0.0007160886,0.00016142266,0.7283203,0.0009652087,0.00008510652,0.00006883911,0.00021182171,0.000024108807,0.00007610648,0.00009029559,0.26915666,0.00012405992],"about_ca_topic_score_codex":0.00013340385,"about_ca_topic_score_gemma":0.00022493227,"teacher_disagreement_score":0.55138445,"about_ca_system_score_codex":0.000025533545,"about_ca_system_score_gemma":0.0009750303,"threshold_uncertainty_score":0.87393504},"labels":[],"label_agreement":null},{"id":"W2169807821","doi":"10.1111/j.1365-2929.2005.02339.x","title":"A practical guide to assessing clinical decision‐making skills using the key features approach","year":2005,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":136,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Key (lock); Computer science; Clinical decision making; Reliability (semiconductor); Test (biology); Quality (philosophy); Management science; Medicine; Engineering","score_opus":0.0577737618307707,"score_gpt":0.529533529534393,"score_spread":0.4717597677036223,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2169807821","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.63000005,0.0006044459,0.21801423,0.13415621,0.0028832473,0.00096259365,0.0000020296457,0.00016186544,0.0132153295],"genre_scores_gemma":[0.54005,0.00005914514,0.40156516,0.053737897,0.004028606,0.000035651567,0.000020857764,0.000029400338,0.00047325567],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99663043,0.00027826396,0.0009398124,0.0005102843,0.0012592649,0.00038193652],"domain_scores_gemma":[0.97361284,0.02428624,0.00022484761,0.00071551756,0.00034486788,0.00081570575],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.003653579,0.00018492891,0.00041644214,0.00009602867,0.00020788511,0.00012841138,0.00021900258,0.0003571216,0.00045218156],"category_scores_gemma":[0.45730647,0.00011509456,0.00019636005,0.00041271618,0.00020620106,0.00013105446,0.0001234748,0.0009409229,0.00010666856],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007715465,0.0028794042,0.013013347,0.00001030568,0.00003622028,0.000012233965,0.00039496573,0.000039223763,0.0000045931893,0.00048437642,0.33250263,0.65054554],"study_design_scores_gemma":[0.0031362236,0.00042160205,0.4723516,0.0107642505,0.0007814956,0.002625636,0.0021078703,0.026838461,0.000047615125,0.0005596875,0.47969455,0.00067100057],"about_ca_topic_score_codex":0.00005584036,"about_ca_topic_score_gemma":0.000014280032,"teacher_disagreement_score":0.64987457,"about_ca_system_score_codex":0.00018238272,"about_ca_system_score_gemma":0.004053945,"threshold_uncertainty_score":0.71915245},"labels":[],"label_agreement":null},{"id":"W2169966878","doi":"10.1186/1472-6920-4-23","title":"The impact of two multiple-choice question formats on the problem-solving strategies used by novices and experts","year":2004,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":73,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Foothills Medical Centre; Health Sciences Centre; University of Calgary","funders":"","keywords":"Multiple choice; Think aloud protocol; Computer science; Psychometrics; Matching (statistics); Pencil (optics); Test (biology); Educational measurement; Mathematics education; Management science; Psychology; Curriculum; Mathematics; Statistics; Human–computer interaction; Clinical psychology; Pedagogy","score_opus":0.029454014674175348,"score_gpt":0.4096948106071412,"score_spread":0.38024079593296584,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2169966878","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99227,0.00068363117,0.00040786024,0.0053710346,0.000147428,0.00037823513,0.000001908116,0.000026964073,0.000712938],"genre_scores_gemma":[0.99862146,0.00022074023,0.0003589869,0.00044863948,0.00020764611,0.000055708544,0.00002862581,0.000009200963,0.000048976708],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99884206,0.00009862115,0.00031450263,0.00015161702,0.0004259025,0.00016730645],"domain_scores_gemma":[0.9748759,0.024414413,0.0001568618,0.00022274535,0.00009929324,0.00023078383],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0007773749,0.00010837337,0.00015652715,0.000025270261,0.00013262776,0.000049304774,0.00009557858,0.000086567146,0.00004582879],"category_scores_gemma":[0.07496894,0.00004977972,0.00006541811,0.0000993351,0.00019391305,0.00008688868,0.000018184233,0.00017218049,0.0000046291757],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010054223,0.007204382,0.7033776,0.00044089637,0.00027206272,0.0000040101063,0.018968385,0.0006213786,0.0036479833,0.03647156,0.03837162,0.18961471],"study_design_scores_gemma":[0.0109740095,0.0028070938,0.91848874,0.01884188,0.0002259945,0.000114276634,0.019817775,0.0061595025,0.0019895849,0.019137828,0.0009298509,0.0005134582],"about_ca_topic_score_codex":0.0033637362,"about_ca_topic_score_gemma":0.00054428016,"teacher_disagreement_score":0.21511117,"about_ca_system_score_codex":0.000072716015,"about_ca_system_score_gemma":0.0020401985,"threshold_uncertainty_score":0.932823},"labels":[],"label_agreement":null},{"id":"W2170132509","doi":"10.2310/8000.2013.131053","title":"ED cognition: any decision by anyone at any time","year":2014,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":84,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Medicine; Action (physics); Cognition; Content (measure theory); Psychiatry","score_opus":0.042757781780123655,"score_gpt":0.35005351537818363,"score_spread":0.30729573359805995,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2170132509","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9141453,0.012143793,0.0011312108,0.021886727,0.010493143,0.00032289643,0.00005393999,0.00002899559,0.039794005],"genre_scores_gemma":[0.9849702,0.0012228681,0.00034370655,0.0019874703,0.003454793,0.0000025606764,0.0000883633,0.00003590227,0.007894116],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99767184,0.000078264384,0.0010799537,0.00021485132,0.00051234773,0.0004427477],"domain_scores_gemma":[0.9939961,0.0019751112,0.00037461377,0.00029381787,0.0006453988,0.0027149795],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012238438,0.00019993912,0.00065885455,0.00025240914,0.00014389792,0.000004298074,0.00017267487,0.0001549596,0.033930197],"category_scores_gemma":[0.05478733,0.00014859647,0.00018037365,0.0003307896,0.00017454197,0.000062745974,0.000012738222,0.00034567222,0.00039480973],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001455436,0.00005413308,0.016809545,0.00001468169,0.00008043264,0.00015861071,0.00010071288,0.0000020345453,0.001712756,0.00004047552,0.9740383,0.0068427413],"study_design_scores_gemma":[0.008795548,0.0048479186,0.078139,0.007081425,0.0010973373,0.00062686583,0.000106747284,0.00026124757,0.00042888813,0.0015533756,0.8966389,0.0004227902],"about_ca_topic_score_codex":0.0010984441,"about_ca_topic_score_gemma":0.0012225689,"teacher_disagreement_score":0.07739948,"about_ca_system_score_codex":0.00014937276,"about_ca_system_score_gemma":0.0004921157,"threshold_uncertainty_score":0.9669529},"labels":[],"label_agreement":null},{"id":"W2170282272","doi":"10.3109/0142159x.2012.733041","title":"Clinical reasoning difficulties: A taxonomy for clinical teachers","year":2012,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":95,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Montréal; Université de Montréal; College of Family Physicians of Canada","funders":"","keywords":"Cornerstone; Clinical Practice; Psychology; Medical education; Construct (python library); Process (computing); Medicine; Computer science; Nursing","score_opus":0.13241689454923747,"score_gpt":0.44706824760079317,"score_spread":0.31465135305155567,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2170282272","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9661835,0.002257548,0.0066696214,0.0050814124,0.0030182416,0.0014395926,0.0000068742843,0.00039691746,0.014946305],"genre_scores_gemma":[0.9457417,0.00072327803,0.016653363,0.0056761247,0.0137744425,0.0005680982,0.0001316342,0.0001138145,0.016617548],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.9953156,0.00048485296,0.0017193577,0.0006349634,0.0008558966,0.0009893316],"domain_scores_gemma":[0.98029864,0.015375743,0.00030583565,0.0007409768,0.00014274071,0.0031360735],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.010111548,0.00032603665,0.0011726836,0.000060454815,0.00011724452,0.00002428669,0.0002776929,0.0010512625,0.0032058447],"category_scores_gemma":[0.31295305,0.00023675189,0.0009992808,0.00016286621,0.0006773054,0.000094544856,0.00013646946,0.0016721672,0.00043030735],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00017834005,0.002173648,0.6868272,0.0000150329515,0.00017409727,0.000015891486,0.00015112874,6.583579e-8,2.8339124e-7,0.00017425811,0.17205162,0.13823847],"study_design_scores_gemma":[0.0077603827,0.0007859403,0.39894357,0.0012062049,0.00048146537,0.000083877974,0.00044825717,0.0010123079,0.000002519793,0.000034553326,0.58886963,0.00037125693],"about_ca_topic_score_codex":0.000031617787,"about_ca_topic_score_gemma":0.0000034577151,"teacher_disagreement_score":0.41681802,"about_ca_system_score_codex":0.00008461752,"about_ca_system_score_gemma":0.00041876838,"threshold_uncertainty_score":0.99770534},"labels":[],"label_agreement":null},{"id":"W2171297463","doi":"10.1016/s0140-6736(14)60376-7","title":"Bias in assessing trainees' clinical competence: the influence of assessors' recent experiences of other performances on present assessment scores","year":2014,"lang":"en","type":"article","venue":"The Lancet","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia; Dalhousie University","funders":"","keywords":"Competence (human resources); Confidence interval; Psychology; Medicine; Clinical psychology; Social psychology","score_opus":0.1333152110856842,"score_gpt":0.4544746419643681,"score_spread":0.3211594308786839,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2171297463","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9907352,0.00013286523,0.00003895184,0.0043667024,0.00020185989,0.0002662002,0.0000031437164,0.00001612656,0.004238959],"genre_scores_gemma":[0.99710387,0.00044898601,0.0006497258,0.0012877736,0.00043579083,0.000038569153,0.000001993979,0.000008926556,0.000024384215],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9977116,0.00047138787,0.0007345564,0.00021134462,0.00063019193,0.00024091202],"domain_scores_gemma":[0.9808875,0.017881772,0.00037642443,0.00063975534,0.00015082244,0.000063766136],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0044239075,0.00012679945,0.00067187136,0.000041047297,0.000054648848,0.000024503634,0.00039763426,0.000070502305,0.00006640108],"category_scores_gemma":[0.020473406,0.000058348105,0.00009538542,0.00024394359,0.00062587386,0.000058205103,0.00006296244,0.00039813964,0.0000039059105],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003173084,0.0006926495,0.96220994,0.00007399283,0.000050539984,0.0000028957966,0.0053168745,0.002577803,0.00018092869,0.0016301014,0.0012134231,0.025733571],"study_design_scores_gemma":[0.0015251099,0.00065703865,0.98625135,0.0024527556,0.00003537592,0.0000034020827,0.0030497292,0.0036258379,0.00056187983,0.0002718756,0.0014854387,0.00008019192],"about_ca_topic_score_codex":0.00009449428,"about_ca_topic_score_gemma":0.000013997219,"teacher_disagreement_score":0.025653379,"about_ca_system_score_codex":0.00002138034,"about_ca_system_score_gemma":0.00017189313,"threshold_uncertainty_score":0.9877776},"labels":[],"label_agreement":null},{"id":"W2171379382","doi":"10.1186/1756-0500-4-418","title":"An innovative method to assess clinical reasoning skills: Clinical reasoning tests in the second national medical science Olympiad in Iran","year":2011,"lang":"en","type":"article","venue":"BMC Research Notes","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":36,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Sante Montreal","funders":"Islamic Azad University; Humanities Research Group, University of Windsor; Islamic Azad University, Mashhad Branch; Shahid Beheshti University of Medical Sciences; Babol University of Medical Sciences; Isfahan University of Medical Sciences; Shiraz University; Zahedan University of Medical Sciences; Mashhad University of Medical Sciences; Tehran University of Medical Sciences and Health Services; Shiraz University of Medical Sciences; Ministry of Health and Medical Education","keywords":"Olympiad; Test (biology); Cronbach's alpha; Reliability (semiconductor); Concordance; Multiple choice; Measure (data warehouse); Medicine; Medical education; Mathematics education; Psychology; Computer science; Statistics; Mathematics; Data mining; Psychometrics; Clinical psychology; Significant difference","score_opus":0.5392440816124454,"score_gpt":0.6267709878968422,"score_spread":0.08752690628439674,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2171379382","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9841401,0.00007158892,0.0022541059,0.0015428419,0.0002789583,0.0009558862,0.000012837366,0.000052946623,0.010690754],"genre_scores_gemma":[0.9008555,0.00003634497,0.09559366,0.002466645,0.00074943714,0.00014318983,0.000019310291,0.000033670774,0.00010228879],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.98484844,0.005537191,0.002089779,0.001482556,0.004705006,0.0013370385],"domain_scores_gemma":[0.5138725,0.47472906,0.00057304127,0.0026269343,0.0049185855,0.0032799065],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.1396298,0.00027425148,0.0008588262,0.0010300311,0.00028569115,0.0001537086,0.0014886581,0.0004956244,0.0005985236],"category_scores_gemma":[0.7898128,0.00019065198,0.00016025378,0.005906721,0.0019604114,0.00035232858,0.00048497104,0.0032964505,0.0001328296],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00046245847,0.0026724006,0.9567804,0.00001949791,0.000016965903,0.00026695643,0.0018503803,0.0000041206035,0.000066749155,0.004829901,0.0004445513,0.032585636],"study_design_scores_gemma":[0.0026306333,0.0015345233,0.9875455,0.0022079297,0.000008671944,0.000077518824,0.0009310866,0.002600049,0.0003155684,0.0015705156,0.00035356672,0.00022446453],"about_ca_topic_score_codex":0.0007449688,"about_ca_topic_score_gemma":0.0021520674,"teacher_disagreement_score":0.650183,"about_ca_system_score_codex":0.00024435116,"about_ca_system_score_gemma":0.0055976505,"threshold_uncertainty_score":0.999003},"labels":[],"label_agreement":null},{"id":"W2171823291","doi":"10.22329/il.v26i1.430","title":"Applied Epistemology and Argumentation in Epidemiology","year":2008,"lang":"en","type":"article","venue":"Informal Logic","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Capilano University","funders":"","keywords":"Argumentation theory; Epistemology; Argument (complex analysis); Epidemiology; Social epistemology; Causal reasoning; Sociology; Causal inference; Philosophy; Psychology; Medicine; Cognition","score_opus":0.07565957776028776,"score_gpt":0.35967247807376496,"score_spread":0.2840129003134772,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2171823291","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95788586,0.00014322173,0.00017187801,0.00091557007,0.00006616692,0.00014964108,8.919805e-7,0.00002625644,0.040640537],"genre_scores_gemma":[0.9920864,0.0002247925,0.001740059,0.0056495834,0.000043102777,0.00001777087,0.000029386036,0.0000028907677,0.00020601987],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99929094,0.000023510813,0.0003385865,0.0001088495,0.000044832293,0.00019329373],"domain_scores_gemma":[0.9958364,0.003910351,0.00006322675,0.00008494365,0.000014156416,0.00009091848],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0003637583,0.000064924396,0.00026710893,0.00005923288,0.000027917977,0.0000010130175,0.000023803657,0.00011303508,0.000060932416],"category_scores_gemma":[0.009809688,0.00004970511,0.000024582869,0.000066954344,0.00015002409,0.000027889919,0.000023941835,0.00013699556,0.000089579815],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00022051274,0.00006902605,0.9637548,0.0000137787665,0.0000083341365,0.00009497484,0.0002425603,0.00002977924,0.000027032867,0.01993842,0.00036809474,0.015232683],"study_design_scores_gemma":[0.0015510811,0.00027965844,0.9944592,0.000042763942,0.000008256314,0.0004841285,0.00004319876,0.00014856772,0.000026789327,0.0022318044,0.0006686355,0.00005590762],"about_ca_topic_score_codex":0.0000548427,"about_ca_topic_score_gemma":0.0000057258085,"teacher_disagreement_score":0.040434517,"about_ca_system_score_codex":0.000015377305,"about_ca_system_score_gemma":0.000033539145,"threshold_uncertainty_score":0.9985311},"labels":[],"label_agreement":null},{"id":"W2172071598","doi":"10.1177/014107680309600621","title":"Consulting with NLP: Neuro-linguistic Programming in the Medical Consultation","year":2003,"lang":"en","type":"article","venue":"Journal of the Royal Society of Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Computer science; Anger; Perception; Artificial intelligence; Psychology; Social psychology","score_opus":0.021117650612453166,"score_gpt":0.3248067701731596,"score_spread":0.30368911956070643,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2172071598","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9449441,0.0014750074,0.0013356608,0.04979204,0.00056743936,0.00043395598,8.864458e-7,0.000009761208,0.0014411609],"genre_scores_gemma":[0.9924137,0.00011100413,0.0028589733,0.004114348,0.00039428292,0.0000024767596,9.560696e-7,0.000010574794,0.00009369248],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99743354,0.00022886752,0.00078328664,0.00010485887,0.0012563185,0.00019312275],"domain_scores_gemma":[0.98011947,0.01823172,0.00074115494,0.00021183552,0.0005117718,0.00018407444],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0041245096,0.000116660114,0.0004593323,0.000026067197,0.00007256454,0.0000063515104,0.0002050359,0.00009889815,0.00006617099],"category_scores_gemma":[0.21305035,0.000047494836,0.00018442661,0.0003294374,0.0006736049,0.00001312374,0.000015824942,0.0008038974,7.1213617e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019948836,0.0034711664,0.7988935,0.0008452881,0.0011181048,0.0012578553,0.04734951,0.0010918982,0.00042620974,0.002785527,0.051432285,0.08933379],"study_design_scores_gemma":[0.10464237,0.01291216,0.66885024,0.076164566,0.0042416146,0.009833118,0.059589423,0.01297408,0.00084858935,0.0018980646,0.04732244,0.00072331703],"about_ca_topic_score_codex":0.000071255956,"about_ca_topic_score_gemma":0.0000063836565,"teacher_disagreement_score":0.20892584,"about_ca_system_score_codex":0.00003693368,"about_ca_system_score_gemma":0.00039766528,"threshold_uncertainty_score":0.79357845},"labels":[],"label_agreement":null},{"id":"W2175164517","doi":"10.1136/bmjqs-2015-004150","title":"Patient safety climate strength: a concept that requires more attention","year":2015,"lang":"en","type":"article","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":47,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Saint Mary's University; York University","funders":"Direction Générale de l’offre de Soins; York University","keywords":"Medicine; Accreditation; Proxy (statistics); Intraclass correlation; Health care; Statistics; Mathematics; Medical education; Political science","score_opus":0.1055800043241647,"score_gpt":0.4305763613240075,"score_spread":0.3249963569998428,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2175164517","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96541417,0.0006892825,0.00084876065,0.017832745,0.0012060348,0.0013469236,0.00047726088,0.00036408516,0.011820713],"genre_scores_gemma":[0.9946382,0.00027540757,0.0009869752,0.0022700373,0.00039533517,0.00003540953,0.00088232534,0.000031928976,0.0004843534],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99617577,0.00040962078,0.0014688334,0.00052011956,0.0009132844,0.0005123999],"domain_scores_gemma":[0.9923874,0.0054914197,0.0004162452,0.00073995796,0.00030389684,0.0006610422],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0025428103,0.0002626077,0.0006862791,0.00004688274,0.00012880121,0.000029173169,0.000113138514,0.00021806761,0.0002199791],"category_scores_gemma":[0.037865896,0.0002106512,0.00031243148,0.00018350591,0.00025258053,0.00011730341,0.00015920898,0.00031353973,0.00021020226],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.012230395,0.0031520976,0.52068794,0.0003075933,0.00046361922,0.00039612706,0.009080296,0.00016236436,0.00020474289,0.0130663095,0.08265084,0.35759768],"study_design_scores_gemma":[0.012729696,0.0013941982,0.9306288,0.0038943782,0.00039460053,0.00009606301,0.012435975,0.00047277022,0.00041095857,0.00087532424,0.035836972,0.00083028275],"about_ca_topic_score_codex":0.00045454822,"about_ca_topic_score_gemma":0.000035415385,"teacher_disagreement_score":0.40994084,"about_ca_system_score_codex":0.00022638206,"about_ca_system_score_gemma":0.00033994735,"threshold_uncertainty_score":0.97023857},"labels":[],"label_agreement":null},{"id":"W2175528350","doi":"10.1007/s10459-015-9651-8","title":"Age-related decline and diagnostic performance of more and less prevalent clinical cases","year":2015,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Sherbrooke","funders":"","keywords":"Medicine; Analysis of variance; Cognition; Gerontology; Post-hoc analysis; Cognitive decline; Age groups; Demography; Pediatrics; Clinical psychology; Dementia; Internal medicine; Psychiatry; Disease","score_opus":0.07620710697587592,"score_gpt":0.4830294257940997,"score_spread":0.4068223188182238,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2175528350","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95939237,0.0356441,0.000006599461,0.0037567604,0.0006291655,0.00032327222,0.0000016484117,0.000010588469,0.00023551754],"genre_scores_gemma":[0.9498498,0.047011055,0.0022683106,0.00071502366,0.000072608294,0.000019744224,0.0000074270815,0.0000036453769,0.00005238314],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9985321,0.00008881994,0.00067468785,0.00028534365,0.00024151614,0.00017749521],"domain_scores_gemma":[0.9898781,0.009343806,0.00024924055,0.00012619769,0.0000955169,0.00030712265],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0021098468,0.00007815979,0.0002899888,0.00008864779,0.000057777404,0.00000905672,0.000062504674,0.000049176964,0.0000034518648],"category_scores_gemma":[0.043655194,0.000060771294,0.00001660348,0.00030764064,0.00081702875,0.00017599977,0.000040852985,0.00013033202,0.0000011795428],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000032556898,0.00023268793,0.8075294,0.00016670865,0.0000010808179,0.0000019893996,0.00037119584,0.00011386261,1.9243909e-7,0.00018931802,0.00007640963,0.19128461],"study_design_scores_gemma":[0.0008930029,0.0020883428,0.98794746,0.0035983406,0.000016026825,0.00008891201,0.0014257975,0.0019700644,0.000008359221,0.0010201748,0.0008628298,0.000080696],"about_ca_topic_score_codex":0.00038556723,"about_ca_topic_score_gemma":0.000083773775,"teacher_disagreement_score":0.19120392,"about_ca_system_score_codex":0.00003699694,"about_ca_system_score_gemma":0.0009450486,"threshold_uncertainty_score":0.96440053},"labels":[],"label_agreement":null},{"id":"W2181091525","doi":"10.1016/j.ijmedinf.2015.09.004","title":"The EHR and building the patient’s story: A qualitative investigation of how EHR use obstructs a vital clinical activity","year":2015,"lang":"en","type":"article","venue":"International Journal of Medical Informatics","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":94,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University; Wilfrid Laurier University; Children's Hospital of Eastern Ontario; University of Ottawa","funders":"","keywords":"Documentation; Narrative; Credibility; Medicine; Thematic analysis; Grounded theory; Medical education; Qualitative research; Patient safety; Health care; Psychology; Nursing; Computer science","score_opus":0.13661941746239561,"score_gpt":0.44907234169053734,"score_spread":0.31245292422814175,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2181091525","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96398985,0.00013722554,0.0019793422,0.032470357,0.0012550405,0.0001060116,0.000010120269,0.000005646164,0.000046389745],"genre_scores_gemma":[0.9952106,0.00031578788,0.0027908816,0.001305172,0.00034521116,0.0000017943903,0.0000036132717,0.0000061686214,0.000020796622],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99619657,0.0003589022,0.0009683732,0.000055943146,0.0022943695,0.00012587354],"domain_scores_gemma":[0.9669189,0.029673813,0.0013510203,0.00015048504,0.0013742119,0.0005315565],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0056140423,0.00010080822,0.00031184228,0.000070779366,0.000058200516,0.00009471042,0.00031880292,0.00014135362,0.000005346467],"category_scores_gemma":[0.20783924,0.000050803414,0.00013760888,0.00009283751,0.000758414,0.00037087945,0.00014638112,0.0008156792,0.0000016711552],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0036596155,0.0007841804,0.11985945,0.00009650096,0.0026749154,0.00025838762,0.111637995,0.00006432141,0.000121782956,0.014301569,0.07945324,0.66708803],"study_design_scores_gemma":[0.057922766,0.01694733,0.41816676,0.026039686,0.002011308,0.009487737,0.24205509,0.097712584,0.004005527,0.03688309,0.08730922,0.0014588981],"about_ca_topic_score_codex":0.000022786176,"about_ca_topic_score_gemma":0.0000058699598,"teacher_disagreement_score":0.66562915,"about_ca_system_score_codex":0.00007844408,"about_ca_system_score_gemma":0.00072343636,"threshold_uncertainty_score":0.7988335},"labels":[],"label_agreement":null},{"id":"W2181879244","doi":"","title":"DETERMINING CLIENT EFFORT: UNDERSTANDING THERAPISTS' PRACTICES","year":2013,"lang":"en","type":"dissertation","venue":"Library and Archives Canada (Government of Canada)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Psychology; Psychotherapist; Medicine","score_opus":0.01585954056201986,"score_gpt":0.23679139401839508,"score_spread":0.22093185345637523,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2181879244","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5530504,0.00085668336,0.000015194968,0.002101482,0.0007223837,0.0003958539,0.00006238643,0.000017738072,0.44277787],"genre_scores_gemma":[0.96052825,0.0008108628,0.00050760567,0.0012436698,0.00013255537,0.00001760248,0.00014348472,0.00004264966,0.036573313],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9972645,0.000035855337,0.00045119852,0.00033279546,0.0016189431,0.00029672764],"domain_scores_gemma":[0.9948224,0.003784187,0.000762596,0.00023396894,0.0000010325216,0.00039580412],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.000017133318,0.0002617064,0.0004850301,0.0000311102,0.00014095326,0.00003628559,0.00012378838,0.00008696623,0.0001595585],"category_scores_gemma":[0.00026452733,0.00022670696,0.000061316765,0.000050121715,0.000048762347,0.00016234828,0.000039490802,0.00031856753,1.08928955e-8],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.01047467,0.0008181267,0.55360216,0.0070655523,0.0032241074,0.0027421708,0.002656567,0.000056118184,0.022829078,0.20407131,0.0226728,0.1697873],"study_design_scores_gemma":[0.0047071483,0.0015061253,0.7890674,0.022303797,0.0012143233,0.000114637405,0.068051726,0.0034261793,0.07128818,0.0057101008,0.030644063,0.0019663237],"about_ca_topic_score_codex":0.002615257,"about_ca_topic_score_gemma":0.018264694,"teacher_disagreement_score":0.40747786,"about_ca_system_score_codex":0.000019130302,"about_ca_system_score_gemma":0.0029118957,"threshold_uncertainty_score":0.9996494},"labels":[],"label_agreement":null},{"id":"W2184682380","doi":"10.22605/rrh2858","title":"Deepening the quality of clinical reasoning and decision-making in rural hospital nursing practice","year":2014,"lang":"en","type":"article","venue":"Rural and Remote Health","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":27,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Lethbridge","funders":"","keywords":"Nursing; Feeling; Rural area; Acute care; Quality (philosophy); Psychology; Medicine; Health care; Social psychology; Political science","score_opus":0.03542301424859339,"score_gpt":0.4684438070776939,"score_spread":0.4330207928291005,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2184682380","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9807354,0.002149454,0.0030172442,0.01288671,0.0003019805,0.00025676002,0.0000011517538,0.000020186562,0.0006310784],"genre_scores_gemma":[0.974209,0.0017816497,0.022068342,0.0016736303,0.00024441458,4.9762036e-7,0.0000027457168,0.000010798006,0.000008892648],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9977233,0.000440457,0.0010303361,0.00022418522,0.0002902519,0.00029147608],"domain_scores_gemma":[0.9610902,0.03776467,0.00053085724,0.00025295783,0.000094578085,0.00026670576],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005651274,0.00012191211,0.0006301914,0.000041902338,0.00015119785,0.000022313186,0.000058326434,0.000119741046,0.000003911976],"category_scores_gemma":[0.099288724,0.000074396456,0.00008264273,0.0001433281,0.0002472174,0.00007544062,0.000052201573,0.0005014526,0.0000016533745],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00038684066,0.00011496845,0.06316056,0.00003762511,0.000010674142,0.0000040289638,0.00075002003,0.0000025366421,0.000002091042,0.0010775111,0.000065576554,0.93438756],"study_design_scores_gemma":[0.0020206794,0.00087715976,0.96477956,0.015804952,0.000039048748,0.00012188309,0.0028618576,0.007908427,0.0000012896642,0.0052674697,0.00020034514,0.0001173215],"about_ca_topic_score_codex":0.001129596,"about_ca_topic_score_gemma":0.000048578382,"teacher_disagreement_score":0.93427026,"about_ca_system_score_codex":0.00003367804,"about_ca_system_score_gemma":0.00014424443,"threshold_uncertainty_score":0.9082984},"labels":[],"label_agreement":null},{"id":"W2185484677","doi":"10.4300/jgme-d-15-00438.1","title":"Thinking on the Front Line: Understanding Resident Decision Making in Clinical Practice","year":2015,"lang":"en","type":"article","venue":"Journal of Graduate Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Competence (human resources); Grounded theory; Medical diagnosis; Psychology; Cognition; Curiosity; Medical education; Qualitative research; MEDLINE; Front line; Medicine; Social psychology; Pathology; Psychiatry","score_opus":0.23804562332717558,"score_gpt":0.5029804777038022,"score_spread":0.26493485437662656,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2185484677","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7163249,0.0009679195,0.011302022,0.26067257,0.0056636245,0.00027476548,2.8347975e-7,0.000014915451,0.004778986],"genre_scores_gemma":[0.97724307,0.0006656571,0.0052690245,0.014930843,0.001811057,0.000002960396,0.0000017250829,0.000014425223,0.000061237515],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99543095,0.00046093273,0.0013987353,0.00016835773,0.0023538887,0.000187154],"domain_scores_gemma":[0.9381884,0.058444336,0.0012263648,0.000390783,0.00085541315,0.00089473266],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01621126,0.000112479465,0.0003797471,0.00017392659,0.00006281845,0.000055505367,0.00022874362,0.00017204856,0.000093148854],"category_scores_gemma":[0.6429026,0.000063963605,0.00016144555,0.00022024919,0.00010741938,0.00012483109,0.000047817797,0.0014461405,0.00003845249],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.006214402,0.006106043,0.069166236,0.000033198237,0.00019687499,0.0010698114,0.0070487144,0.00015168692,0.0000016893165,0.019590368,0.5160999,0.37432107],"study_design_scores_gemma":[0.01658035,0.009211848,0.36672306,0.10193841,0.0011363848,0.0049796696,0.061541494,0.0063315067,0.000028707012,0.4058766,0.024989085,0.00066286424],"about_ca_topic_score_codex":0.000036353285,"about_ca_topic_score_gemma":0.000017894145,"teacher_disagreement_score":0.6266914,"about_ca_system_score_codex":0.0005411535,"about_ca_system_score_gemma":0.0036898893,"threshold_uncertainty_score":0.6545705},"labels":[],"label_agreement":null},{"id":"W2185684651","doi":"10.1017/s1481803500011787","title":"“I never hear what happens, even if they die”: a survey of emergency physicians about outcome feedback","year":2009,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":30,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; Montfort Hospital; Children's Hospital of Eastern Ontario","funders":"Canadian Institutes of Health Research; University of Ottawa","keywords":"Outcome (game theory); Medicine; Perception; Family medicine; Psychology","score_opus":0.11405558863211727,"score_gpt":0.39610939907999976,"score_spread":0.28205381044788247,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2185684651","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9271134,0.02187816,0.00013903594,0.027655035,0.018033497,0.00047560185,0.00007514678,0.000019951189,0.004610182],"genre_scores_gemma":[0.9839819,0.0105605675,0.00008388273,0.0014688335,0.0016468433,0.000001677361,0.00005265198,0.00004240985,0.0021612325],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99557346,0.0002149934,0.0024605978,0.0003147255,0.0007277804,0.0007084662],"domain_scores_gemma":[0.99444884,0.0009073571,0.0009827527,0.0005692953,0.0013542387,0.0017374882],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0019857343,0.00036150008,0.0014117921,0.00039304173,0.00011754333,0.00000786939,0.00039803644,0.00020068126,0.0058828206],"category_scores_gemma":[0.031508755,0.000274423,0.00044488345,0.0005902994,0.00013740464,0.00035036175,0.000013499906,0.00067469757,0.00005076979],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00040095107,0.00044931928,0.7248756,0.00010894566,0.00048350895,0.0005739272,0.00305257,0.00007050565,0.00052471133,0.0002639423,0.21550302,0.053692937],"study_design_scores_gemma":[0.0020056297,0.0021947178,0.9776855,0.0040114266,0.00041245812,0.000038141414,0.0007044894,0.00003434905,0.000029897761,0.00058261264,0.012041726,0.0002590421],"about_ca_topic_score_codex":0.03279118,"about_ca_topic_score_gemma":0.032349274,"teacher_disagreement_score":0.25280985,"about_ca_system_score_codex":0.00013976259,"about_ca_system_score_gemma":0.0008683649,"threshold_uncertainty_score":0.9999708},"labels":[],"label_agreement":null},{"id":"W2185800458","doi":"10.1111/j.1365-2923.2010.03748.x","title":"Exploring how students think: a new method combining think‐aloud and concept mapping protocols","year":2010,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":56,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; University of Toronto; University Health Network","funders":"","keywords":"Think aloud protocol; Competence (human resources); Deductive reasoning; Logical reasoning; Inductive reasoning; Psychology; Reliability (semiconductor); Curriculum; Mathematics education; Medical education; Computer science; Social psychology; Pedagogy; Medicine; Artificial intelligence","score_opus":0.10299951226535854,"score_gpt":0.4558883495143263,"score_spread":0.35288883724896775,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2185800458","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9494311,0.00011182702,0.00367879,0.039228927,0.0017829278,0.004883211,9.2722127e-7,0.00016328605,0.0007189693],"genre_scores_gemma":[0.895259,0.00012068394,0.07665681,0.016850194,0.0041128336,0.0027448575,0.000092710376,0.00007509403,0.004087809],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99803233,0.00009581815,0.00033626638,0.0003718873,0.00089643104,0.00026725075],"domain_scores_gemma":[0.994686,0.0038863067,0.00013731257,0.00035343313,0.000122965,0.0008139759],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0013582798,0.00015949803,0.00035926496,0.00008070755,0.00010725562,0.00010017061,0.00015466311,0.0001995965,0.0003850789],"category_scores_gemma":[0.110276125,0.00012906238,0.00006379427,0.00018872587,0.00011691736,0.00014589689,0.00009801915,0.00086902245,0.000023091065],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000084481246,0.0017269724,0.06938421,0.00012437906,0.000099140605,0.00003902481,0.031895947,3.546428e-7,0.0005323855,0.0021403828,0.0466802,0.84729254],"study_design_scores_gemma":[0.016956035,0.0014465711,0.8302091,0.022873508,0.00041682043,0.0008299996,0.02334544,0.0018334152,0.0018410196,0.005988219,0.093147874,0.0011120192],"about_ca_topic_score_codex":0.00018975954,"about_ca_topic_score_gemma":0.000010568623,"teacher_disagreement_score":0.8461805,"about_ca_system_score_codex":0.00003343983,"about_ca_system_score_gemma":0.0014407752,"threshold_uncertainty_score":0.8972184},"labels":[],"label_agreement":null},{"id":"W2186245659","doi":"10.1007/s10459-015-9648-3","title":"Hemispheric activation differences in novice and expert clinicians during clinical decision making","year":2015,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":29,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria; University of Calgary","funders":"University of Calgary","keywords":"Medical diagnosis; Clinical decision making; Functional magnetic resonance imaging; Computer science; Process (computing); Medicine; Psychology; Medical physics; Radiology; Family medicine","score_opus":0.075337043306456,"score_gpt":0.5067941155769315,"score_spread":0.43145707227047553,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2186245659","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98740476,0.007769684,0.00042520204,0.0025669583,0.0009320944,0.00029892958,2.631861e-7,0.000021130576,0.000580996],"genre_scores_gemma":[0.97982395,0.0071255695,0.01107455,0.0016712766,0.00024060569,0.00003544698,0.0000022644526,0.000005449282,0.000020888861],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99796253,0.00012563645,0.00085110555,0.0004560416,0.0003301354,0.0002745308],"domain_scores_gemma":[0.9938088,0.0054740044,0.00025047205,0.00015794324,0.000077750134,0.00023101122],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002206334,0.00010154916,0.00032121927,0.00014199254,0.00009119575,0.000026829162,0.00010372506,0.000078254976,0.0000086772],"category_scores_gemma":[0.029178191,0.00008344055,0.000024245313,0.00070196413,0.00019068972,0.00037788547,0.00003434928,0.00022147245,0.0000040627997],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008981457,0.00024532015,0.7191768,0.000028247667,4.4579687e-7,6.4100914e-7,0.0005176092,0.00008803906,0.0000012416301,0.000078797304,0.000025217336,0.27974778],"study_design_scores_gemma":[0.0010541398,0.00052385946,0.97913545,0.0065697352,0.0000021432686,0.000011179703,0.0042621903,0.004600102,0.0000055468995,0.0032482683,0.0004675313,0.000119878336],"about_ca_topic_score_codex":0.0004873104,"about_ca_topic_score_gemma":0.0003416561,"teacher_disagreement_score":0.27962792,"about_ca_system_score_codex":0.00016900463,"about_ca_system_score_gemma":0.0011582421,"threshold_uncertainty_score":0.97899944},"labels":[],"label_agreement":null},{"id":"W2186562134","doi":"10.1002/j.0022-0337.2013.77.6.tb05527.x","title":"Integration of Basic Sciences and Clinical Sciences in Oral Radiology Education for Dental Students","year":2013,"lang":"en","type":"article","venue":"Journal of Dental Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":84,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; Women's College Hospital; University of Toronto","funders":"","keywords":"Medical education; Basic research; Biomedical sciences; Medicine; Medical physics; Dental education; Radiology; Computer science; Pathology","score_opus":0.05410580931274067,"score_gpt":0.4855085360068777,"score_spread":0.43140272669413704,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2186562134","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9965891,0.00051611214,0.00005497266,0.000715446,0.0015779162,0.00031035632,7.8699384e-7,0.000001737614,0.00023354044],"genre_scores_gemma":[0.99373865,0.00020176929,0.0052936585,0.0003690608,0.0002639527,0.000011242213,0.00000484995,0.000003315932,0.00011352636],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9987659,0.00007824041,0.00067315233,0.00013136966,0.00025375822,0.00009759153],"domain_scores_gemma":[0.99751455,0.0017100131,0.0004195711,0.0000491125,0.00021397384,0.000092784154],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0014013499,0.000062633226,0.00025137071,0.00024298923,0.000045692956,0.000036271602,0.00011296323,0.00006516159,0.000043808206],"category_scores_gemma":[0.0106803095,0.0000456031,0.00008005404,0.00020539298,0.0003487228,0.00023629764,0.000017294093,0.00012261744,0.00000455233],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000058386868,0.0014107971,0.91831404,0.0000100509415,0.000010800377,2.2171704e-7,0.00013394405,0.0000011172513,0.00070857094,0.000223017,0.0020570802,0.07707199],"study_design_scores_gemma":[0.0010258403,0.0018411876,0.989812,0.0008092208,0.000060577702,0.000211815,0.0046745236,0.000117779906,0.00034517402,0.001020894,0.000036708294,0.000044234326],"about_ca_topic_score_codex":0.000195757,"about_ca_topic_score_gemma":0.000042727468,"teacher_disagreement_score":0.07702775,"about_ca_system_score_codex":0.00006972101,"about_ca_system_score_gemma":0.0010368854,"threshold_uncertainty_score":0.9976531},"labels":[],"label_agreement":null},{"id":"W2192468482","doi":"10.1007/978-3-319-01671-9_2","title":"The Necessary Forms of the Knowledge","year":2013,"lang":"en","type":"book-chapter","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Computer science","score_opus":0.028581007729731954,"score_gpt":0.3105442766222076,"score_spread":0.2819632688924757,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2192468482","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00039825978,0.002364974,0.0000062304093,0.002513186,0.00043376232,0.00039168508,0.000006005889,0.000025545167,0.99386036],"genre_scores_gemma":[0.018126326,0.0010935507,0.000035702236,0.00057776814,0.00026116517,0.0000100666175,0.00000906855,0.00003031934,0.979856],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.9990632,0.000008241108,0.00038275102,0.00016388441,0.0002332148,0.00014873593],"domain_scores_gemma":[0.98895484,0.00989292,0.00017463,0.00071162864,0.00016649248,0.00009948821],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0002155072,0.00016757328,0.00035490893,0.000018644205,0.000076574244,0.000011587372,0.00022834638,0.00026458825,0.0018969802],"category_scores_gemma":[0.0071323807,0.000059801245,0.00033210384,0.000019895055,0.00029244053,0.000011054309,0.00014731595,0.00039789037,0.0008019031],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000034763034,0.00007319776,0.00042547463,0.000050803643,0.00019192405,0.0000033670995,0.000050118808,1.5932604e-7,0.0000036514055,0.550874,0.32575303,0.122539535],"study_design_scores_gemma":[0.00073950033,0.00020780276,0.005060999,0.0043540336,0.00034453615,0.000022132965,0.000030838026,0.000025363048,0.000069338166,0.05724899,0.93172777,0.00016871736],"about_ca_topic_score_codex":0.000027500955,"about_ca_topic_score_gemma":0.000034899396,"teacher_disagreement_score":0.60597473,"about_ca_system_score_codex":0.00002157609,"about_ca_system_score_gemma":0.00018892411,"threshold_uncertainty_score":0.9999761},"labels":[],"label_agreement":null},{"id":"W2210225355","doi":"10.36834/cmej.36653","title":"Internal Medicine residents use heuristics to estimate disease probability","year":2015,"lang":"en","type":"article","venue":"Canadian Medical Education Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Heuristics; Computer science","score_opus":0.051245388913261324,"score_gpt":0.41364268345326105,"score_spread":0.36239729453999975,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2210225355","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.61456615,0.00066802884,0.00096692436,0.37192148,0.006706256,0.0005333252,0.000018721998,0.000051901934,0.0045672306],"genre_scores_gemma":[0.89011663,0.000065073735,0.0039441655,0.09732596,0.003943197,0.000025731697,0.00006619835,0.00003775212,0.004475311],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99737024,0.00012302959,0.0006243038,0.00026664263,0.001190663,0.00042509314],"domain_scores_gemma":[0.93062407,0.0014800761,0.00014155862,0.00044148054,0.0013319112,0.06598091],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00153616,0.00015586143,0.0003054589,0.00026527137,0.00010800141,0.00007895719,0.00023090321,0.0001300677,0.005759255],"category_scores_gemma":[0.7772539,0.00012100772,0.00007160992,0.0002629398,0.00021196288,0.000092990085,0.000029811481,0.00071266526,0.00022089397],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00016554033,0.0002459419,0.16327329,0.000016922395,0.000024349652,0.0006421206,0.00043188583,0.00001244458,4.356044e-7,0.0004827975,0.80862385,0.026080439],"study_design_scores_gemma":[0.0023017875,0.00054882496,0.8124408,0.00685567,0.0002052983,0.0018740229,0.00027296037,0.00046043078,0.0000016991232,0.0067848098,0.16799623,0.00025742955],"about_ca_topic_score_codex":0.024730446,"about_ca_topic_score_gemma":0.009011457,"teacher_disagreement_score":0.7757178,"about_ca_system_score_codex":0.0007555421,"about_ca_system_score_gemma":0.05898702,"threshold_uncertainty_score":0.9951496},"labels":[],"label_agreement":null},{"id":"W2221865989","doi":"10.7759/cureus.381","title":"Emergency Physicians Think in Spirals","year":2015,"lang":"en","type":"article","venue":"Cureus","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Memorial University of Newfoundland; University of Ontario Institute of Technology","funders":"Memorial University of Newfoundland","keywords":"Medicine; Emergency department; Mnemonic; Presentation (obstetrics); Disposition; Medical education; Medical emergency; Spiral (railway); Nursing; Surgery; Epistemology","score_opus":0.0731026489232073,"score_gpt":0.37875725067145155,"score_spread":0.30565460174824427,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2221865989","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94194084,0.0018690423,0.00004529491,0.002138991,0.0007421808,0.00014420917,0.0000030007088,0.000065610206,0.053050805],"genre_scores_gemma":[0.9958514,0.00020850432,0.0003859652,0.0008621978,0.0003935327,0.0000070320557,0.0000131202805,0.00001209246,0.0022661767],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9992912,0.000020318057,0.00020463923,0.00014996469,0.00016305919,0.00017086627],"domain_scores_gemma":[0.9991335,0.0003800061,0.00003722905,0.00023610127,0.000059465812,0.00015367415],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00021746173,0.000072970266,0.00019495249,0.000021311225,0.000009719584,0.0000035989449,0.000060405346,0.00006604245,0.00021683656],"category_scores_gemma":[0.016935028,0.000060337836,0.0000595757,0.00013469522,0.000022469354,0.000033116154,0.000023719038,0.00012358848,0.00034653267],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013530458,0.0016618992,0.5697144,0.000023205988,0.0000565249,0.0003255744,0.0029593532,0.00003088696,0.000341177,0.0029608163,0.3724711,0.04931972],"study_design_scores_gemma":[0.007049502,0.0016086826,0.91355246,0.0020139965,0.00018820702,0.000041937365,0.0015022372,0.00060781697,0.00095422904,0.022906275,0.049030587,0.0005440782],"about_ca_topic_score_codex":0.0002254623,"about_ca_topic_score_gemma":0.000072155664,"teacher_disagreement_score":0.34383804,"about_ca_system_score_codex":0.000041880405,"about_ca_system_score_gemma":0.00005513734,"threshold_uncertainty_score":0.99134576},"labels":[],"label_agreement":null},{"id":"W2222860611","doi":"10.1007/s10459-015-9649-2","title":"Working memory, reasoning, and expertise in medicine—insights into their relationship using functional neuroimaging","year":2015,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":20,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria; University of Calgary","funders":"University of Calgary","keywords":"Working memory; Functional magnetic resonance imaging; Prefrontal cortex; Psychology; Functional neuroimaging; Neuroimaging; Cognitive psychology; Posterior parietal cortex; Cognition; Neuroscience","score_opus":0.12570736599196689,"score_gpt":0.44776998994320055,"score_spread":0.32206262395123364,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2222860611","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.92553484,0.061391648,0.0010932073,0.007364242,0.0019785718,0.00030486012,7.296533e-8,0.000026792257,0.0023057698],"genre_scores_gemma":[0.9906085,0.0013143067,0.0059794034,0.0017027446,0.00030767886,0.000024967616,0.000006566186,0.0000070875494,0.000048710746],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99840575,0.00014835593,0.0004978981,0.00039409945,0.0003110185,0.0002428505],"domain_scores_gemma":[0.9969,0.0024107154,0.00019719936,0.00014344307,0.00007407403,0.00027453055],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0015581312,0.000108251144,0.00023588823,0.0003483371,0.00017821508,0.000019757854,0.00006752077,0.000041635572,0.000007318595],"category_scores_gemma":[0.019940859,0.00008472532,0.000014202732,0.00087609794,0.000376551,0.00037586543,0.000023808472,0.0002159429,0.0000017315164],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000029092453,0.00009102202,0.87531525,0.000029303224,5.535355e-7,0.0000016520341,0.0066728205,0.0014295595,0.0000071500676,0.0014146286,0.00015694214,0.11485201],"study_design_scores_gemma":[0.0018114636,0.0005188899,0.8656671,0.013023716,0.000010287846,0.000111755515,0.01577191,0.06970754,0.000021330992,0.029760629,0.0033356112,0.0002597635],"about_ca_topic_score_codex":0.0017591589,"about_ca_topic_score_gemma":0.0005732757,"teacher_disagreement_score":0.11459225,"about_ca_system_score_codex":0.0003110771,"about_ca_system_score_gemma":0.001350635,"threshold_uncertainty_score":0.98831457},"labels":[],"label_agreement":null},{"id":"W2224460879","doi":"10.1097/acm.0000000000001024","title":"Clinical Reasoning Tasks and Resident Physicians: What Do They Reason About?","year":2015,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":42,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"MEDLINE; Medical education; Psychology; Medicine; Family medicine; Political science","score_opus":0.07252311214599699,"score_gpt":0.42173369474886546,"score_spread":0.34921058260286847,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2224460879","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.88387865,0.08271614,0.00012208323,0.02363945,0.0018678182,0.00054424006,0.0000032504445,0.00019409841,0.007034274],"genre_scores_gemma":[0.88758487,0.09460717,0.00040708005,0.010370477,0.0044874568,0.000027574843,0.000055094853,0.000055599423,0.0024047063],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99633896,0.00027101854,0.0011848934,0.00075263955,0.00091451383,0.0005379504],"domain_scores_gemma":[0.9885296,0.00900373,0.00037526235,0.0006558737,0.00026358408,0.0011719537],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004416116,0.00032333852,0.0010937905,0.000116917705,0.000078146666,0.000030734966,0.00022260482,0.00058369216,0.00008435074],"category_scores_gemma":[0.12077237,0.00023083412,0.00011130585,0.00024088331,0.0006071526,0.00025835377,0.00014761585,0.0017042372,0.00011668877],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000902727,0.00009293322,0.35515866,0.00007237299,0.00020602168,0.00032504697,0.009016847,0.0000036474567,0.00012342827,0.004296719,0.1907111,0.43909052],"study_design_scores_gemma":[0.03724489,0.0073367693,0.5237928,0.17936692,0.0029662368,0.0014705685,0.05389125,0.00083182927,0.00025560203,0.034130506,0.15716949,0.0015431427],"about_ca_topic_score_codex":0.00029832005,"about_ca_topic_score_gemma":0.000008122092,"teacher_disagreement_score":0.43754736,"about_ca_system_score_codex":0.00008166777,"about_ca_system_score_gemma":0.00022181406,"threshold_uncertainty_score":0.9413142},"labels":[],"label_agreement":null},{"id":"W2228345066","doi":"10.1016/j.jpain.2004.02.438","title":"Electrodiagnostic approaches","year":2004,"lang":"en","type":"article","venue":"Journal of Pain","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Physical therapy; McGill Pain Questionnaire; Visual analogue scale","score_opus":0.04381674303346229,"score_gpt":0.3089877876008999,"score_spread":0.2651710445674376,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2228345066","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9686289,0.001514753,0.013122875,0.014116303,0.0002141098,0.00010427141,6.7523905e-7,0.000018028024,0.0022800972],"genre_scores_gemma":[0.9927325,0.000123814,0.005106678,0.0013794036,0.00053841,0.0000010246524,0.0000013480009,0.000009161836,0.00010765912],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99915653,0.00005839846,0.00033874268,0.00006741255,0.00023396724,0.00014493473],"domain_scores_gemma":[0.99392587,0.0055162995,0.0001655897,0.00010219215,0.00008521663,0.00020483372],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001518403,0.00006660975,0.00025002356,0.00007199788,0.000019504318,0.000009901677,0.000057694426,0.000057005956,0.000050137554],"category_scores_gemma":[0.08692922,0.000046466845,0.0001603858,0.00011245158,0.000036399862,0.000029346305,0.000006936756,0.00026811758,0.000026054968],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00224931,0.00971361,0.34285662,0.00024469054,0.0012113148,0.00856974,0.0021000179,0.0036976887,0.005615931,0.018139208,0.08550558,0.5200963],"study_design_scores_gemma":[0.034790955,0.024860354,0.80061346,0.017654816,0.0018023793,0.009910528,0.0013004423,0.0005303863,0.013777793,0.06848425,0.025346665,0.00092794746],"about_ca_topic_score_codex":0.0000054894704,"about_ca_topic_score_gemma":0.0000015349301,"teacher_disagreement_score":0.5191683,"about_ca_system_score_codex":0.00006554154,"about_ca_system_score_gemma":0.000245515,"threshold_uncertainty_score":0.92076194},"labels":[],"label_agreement":null},{"id":"W2230818777","doi":"10.26443/ijwpc.v1i1.59","title":"Doctor, You Can Be Less Error Prone Right Now","year":2014,"lang":"en","type":"article","venue":"International Journal of Whole Person Care","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Patience; Certainty; Psychology; Social psychology; Medicine; Epistemology; Philosophy","score_opus":0.034609748600180366,"score_gpt":0.3486504110823165,"score_spread":0.31404066248213613,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2230818777","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.951452,0.00040291937,0.00022932267,0.038877197,0.0020951636,0.00009230378,0.00006643953,0.000019008721,0.0067656008],"genre_scores_gemma":[0.99269193,0.000030908723,0.0007275387,0.0020166317,0.0019382547,0.0000026861173,0.000055947086,0.000018344032,0.002517751],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.9984355,0.000043785047,0.00037232446,0.00015675917,0.0008306576,0.00016097332],"domain_scores_gemma":[0.9973438,0.0007942463,0.00026330326,0.00013916574,0.0011988163,0.00026066523],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00026154178,0.00012803171,0.00030614348,0.0001443907,0.000037969286,0.00004857975,0.00025789783,0.00009119217,0.0002556562],"category_scores_gemma":[0.00595351,0.00009643121,0.00025078072,0.000060810984,0.00007297019,0.000068884,0.000028310931,0.00034536404,0.00003494389],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019706336,0.0013813947,0.43317518,0.00015146473,0.0016509637,0.0029664987,0.015509577,0.0003452045,0.0081958715,0.001727703,0.1311977,0.40172783],"study_design_scores_gemma":[0.024195522,0.0041372967,0.31850368,0.011603858,0.0009865046,0.0044422294,0.023579363,0.0011208764,0.012564212,0.00037267018,0.59753966,0.00095409737],"about_ca_topic_score_codex":0.00007359812,"about_ca_topic_score_gemma":0.000077223325,"teacher_disagreement_score":0.46634197,"about_ca_system_score_codex":0.0002018854,"about_ca_system_score_gemma":0.00020032236,"threshold_uncertainty_score":0.71273375},"labels":[],"label_agreement":null},{"id":"W2231471519","doi":"10.1016/j.jbtep.2015.12.004","title":"Self-reported intolerance of uncertainty and behavioural decisions","year":2015,"lang":"en","type":"article","venue":"Journal of Behavior Therapy and Experimental Psychiatry","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":89,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Brock University; University of Regina","funders":"Social Sciences and Humanities Research Council; Social Sciences and Humanities Research Council of Canada; Canadian Institutes of Health Research","keywords":"Psychology; Anxiety; Clinical psychology; Compensation (psychology); Task (project management); Depression (economics); Social psychology; Psychiatry","score_opus":0.08591375941881627,"score_gpt":0.3852059013431724,"score_spread":0.29929214192435616,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2231471519","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9833703,0.015244376,0.0000089814885,0.00038921813,0.0007418729,0.00015448027,0.0000038327257,0.000010600271,0.00007632869],"genre_scores_gemma":[0.9931603,0.0007630484,0.005667726,0.00023625804,0.00013487296,0.000004843918,0.0000029577907,0.000012666486,0.000017293802],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99876606,0.000034475685,0.0006330429,0.0001312587,0.00031302145,0.00012211331],"domain_scores_gemma":[0.99869424,0.00026153147,0.0003591677,0.00014108859,0.0001581441,0.00038580317],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00032645636,0.00012978207,0.00040084438,0.00007820145,0.000040190098,0.000014647715,0.00005687747,0.00009606383,0.000028132628],"category_scores_gemma":[0.0002997483,0.00009306565,0.000127332,0.000083699415,0.00011962034,0.0000811974,0.0000202327,0.00020842487,6.0796856e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0026832803,0.0041219527,0.9645754,0.0000063317466,0.00017665557,0.000101127545,0.0019101644,0.000002947788,0.00699992,0.0000799759,0.0008701891,0.018472055],"study_design_scores_gemma":[0.029273564,0.01275432,0.925868,0.0017085633,0.00087010243,0.005562552,0.008405186,0.00007693008,0.013972117,0.00044219493,0.00065713463,0.00040934858],"about_ca_topic_score_codex":0.000049371643,"about_ca_topic_score_gemma":0.0000025539348,"teacher_disagreement_score":0.038707413,"about_ca_system_score_codex":0.000029565334,"about_ca_system_score_gemma":0.00015173896,"threshold_uncertainty_score":0.37951073},"labels":[],"label_agreement":null},{"id":"W2235258262","doi":"10.1002/j.0022-0337.2012.76.9.tb05366.x","title":"Clinical Reasoning in Dentistry: A Conceptual Framework for Dental Education","year":2012,"lang":"en","type":"article","venue":"Journal of Dental Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":34,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Vignette; Biopsychosocial model; Psychology; Think aloud protocol; Medical education; Scripting language; Deductive reasoning; Logical reasoning; Mathematics education; Medicine; Computer science; Social psychology; Artificial intelligence; Psychotherapist","score_opus":0.04422459238924625,"score_gpt":0.4546769061356679,"score_spread":0.41045231374642166,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2235258262","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9861387,0.0022915641,0.0010619444,0.0003812193,0.009350218,0.00028599994,0.000003409646,0.000007385496,0.0004795835],"genre_scores_gemma":[0.96412104,0.00014329536,0.030101527,0.0008871352,0.0041858684,0.00002130703,0.00003563859,0.000018648609,0.0004855349],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99834526,0.00008551534,0.00090507837,0.00013349624,0.00029300785,0.00023763295],"domain_scores_gemma":[0.9954279,0.0031935875,0.00058290747,0.00014410082,0.00022464564,0.00042685715],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010753709,0.00011343241,0.0003558744,0.00014611435,0.0000406105,0.00002849327,0.00009662361,0.0001971541,0.00016067234],"category_scores_gemma":[0.03795814,0.00010221238,0.00025254092,0.00018222266,0.00009107992,0.00024107264,0.000019156465,0.0004598766,0.000046735793],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00029248843,0.0038020024,0.92551595,0.000019243087,0.000036693073,0.000003172816,0.00050741597,6.433355e-7,0.000051457893,0.0033746238,0.008246062,0.058150224],"study_design_scores_gemma":[0.001985166,0.00060450507,0.9701314,0.004274386,0.0002657326,0.0014437003,0.014865173,0.000015588259,0.00024046106,0.0009469524,0.005077385,0.00014957249],"about_ca_topic_score_codex":0.000037334834,"about_ca_topic_score_gemma":0.000007708815,"teacher_disagreement_score":0.05800065,"about_ca_system_score_codex":0.00026318795,"about_ca_system_score_gemma":0.0013586383,"threshold_uncertainty_score":0.9701455},"labels":[],"label_agreement":null},{"id":"W2255159030","doi":"","title":"Enquête sur certaines pratiques médicales au Québec","year":2013,"lang":"fr","type":"preprint","venue":"RePEc: Research Papers in Economics","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université de Montréal","funders":"","keywords":"Humanities; Political science; Art","score_opus":0.044330830746063414,"score_gpt":0.3677438131318271,"score_spread":0.32341298238576366,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2255159030","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8774014,0.0028047368,0.0000038640574,0.013956158,0.0023388085,0.0019808086,0.00008379735,0.000101437145,0.10132898],"genre_scores_gemma":[0.8285071,0.07844897,0.001048479,0.0007201419,0.003651865,0.0010037853,0.0002307094,0.0001985411,0.08619039],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99217993,0.0010652483,0.0020245959,0.0018290904,0.0006583505,0.0022427917],"domain_scores_gemma":[0.9335455,0.06179731,0.00046591618,0.0020560701,0.00059546746,0.0015397758],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.004692529,0.00075896864,0.0018035786,0.00049244874,0.0003017729,0.00023545626,0.00096470135,0.0015643756,0.0021857242],"category_scores_gemma":[0.16697226,0.0007276813,0.00057708105,0.00021627042,0.0020039086,0.00015212987,0.001416882,0.0042494717,0.0007478645],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.00029199442,0.0013891107,0.35448748,0.000520396,0.0004046024,0.0003133309,0.0014793967,0.0012480762,0.00006476449,0.0026552284,0.002548017,0.6345976],"study_design_scores_gemma":[0.004683093,0.001545646,0.5991601,0.024466882,0.00020332678,0.00013136241,0.0033580705,0.023111988,0.00033415057,0.002709212,0.33847865,0.0018175379],"about_ca_topic_score_codex":0.15187581,"about_ca_topic_score_gemma":0.108402394,"teacher_disagreement_score":0.6327801,"about_ca_system_score_codex":0.0047413344,"about_ca_system_score_gemma":0.015076316,"threshold_uncertainty_score":0.9997318},"labels":[],"label_agreement":null},{"id":"W2262586608","doi":"10.1177/0956797615573759","title":"Publication Bias and the Validity of Evidence","year":2015,"lang":"en","type":"letter","venue":"Psychological Science","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":72,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Baycrest Hospital; York University","funders":"","keywords":"Psychology; Social psychology; Cognitive psychology","score_opus":0.5309093860922428,"score_gpt":0.5046901048042801,"score_spread":0.026219281287962692,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2262586608","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.032611266,0.0003727305,0.00006494379,0.95777005,0.0003583925,0.00037800547,0.000005620984,0.00003407018,0.0084049525],"genre_scores_gemma":[0.34506342,0.0005543457,0.0006673058,0.651494,0.00090370333,0.00003510005,0.000012709099,0.0000066037564,0.001262774],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9975326,0.00020822397,0.00039865525,0.0005694376,0.0010255778,0.00026552664],"domain_scores_gemma":[0.97738457,0.020733373,0.00031801173,0.0007295656,0.00064784876,0.00018663176],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.006810525,0.00012739222,0.0004006891,0.00006530593,0.0000756095,0.000068925394,0.0004962191,0.00030982605,0.00011844824],"category_scores_gemma":[0.37312263,0.00005103867,0.000089537156,0.00062953815,0.0045473156,0.00007527967,0.00014224854,0.0008827345,0.000052453273],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006796638,0.00006938505,0.0060444768,0.000014571853,0.0000039417396,0.000017176862,0.000035900168,6.505682e-8,0.000016306572,0.00023065001,0.98215395,0.011345615],"study_design_scores_gemma":[0.0036045893,0.0019488176,0.49099863,0.0033941956,0.00032048757,0.00043358016,0.000022934855,0.00019284165,0.00007109928,0.018931221,0.47964418,0.00043742658],"about_ca_topic_score_codex":0.00003101731,"about_ca_topic_score_gemma":2.5364253e-7,"teacher_disagreement_score":0.5025098,"about_ca_system_score_codex":0.000033454995,"about_ca_system_score_gemma":0.0001429559,"threshold_uncertainty_score":0.99816173},"labels":[],"label_agreement":null},{"id":"W2262973509","doi":"10.1017/cem.2015.95","title":"A Survey to Determine Decision-Making Styles of Working Paramedics and Student Paramedics","year":2016,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":26,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centennial College; Ottawa Hospital; McMaster University; Dalhousie University","funders":"","keywords":"Experiential learning; Likert scale; Medicine; Preference; Descriptive statistics; Critical thinking; Psychology; Medical education; Mathematics education; Developmental psychology","score_opus":0.16019750304122726,"score_gpt":0.42445584629841426,"score_spread":0.264258343257187,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2262973509","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98405457,0.0051170755,0.0025319175,0.004059007,0.0035583293,0.0002018446,0.000017208347,0.000005380047,0.00045464234],"genre_scores_gemma":[0.99561375,0.001974287,0.00088709994,0.00045378745,0.0008950443,0.0000023990508,0.000002155164,0.000024955982,0.00014649874],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9970757,0.00010213715,0.0015044379,0.00022926374,0.0006700322,0.00041846407],"domain_scores_gemma":[0.9857743,0.010501474,0.0005439693,0.00032823542,0.000699871,0.00215216],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0024989434,0.00020289706,0.00084114337,0.0004639715,0.00006321745,0.000004126317,0.00023712976,0.00012751966,0.001211528],"category_scores_gemma":[0.15345192,0.00011936625,0.000113925606,0.00047192717,0.0002620451,0.000042991785,0.0000310425,0.00025029818,0.0000057690377],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015846432,0.000044532535,0.81154394,0.000018358305,0.00011131398,0.00031817847,0.00064744655,0.0000039860156,0.000071532326,0.000045668385,0.017094377,0.1699422],"study_design_scores_gemma":[0.0023347083,0.002112977,0.962309,0.02307094,0.0002853252,0.00013438982,0.0005253898,0.000020759282,0.000019720446,0.000470923,0.00853729,0.0001786026],"about_ca_topic_score_codex":0.0012725341,"about_ca_topic_score_gemma":0.013446652,"teacher_disagreement_score":0.1697636,"about_ca_system_score_codex":0.00010728762,"about_ca_system_score_gemma":0.0008880523,"threshold_uncertainty_score":0.9997015},"labels":[],"label_agreement":null},{"id":"W2276416864","doi":"10.71781/5666","title":"L’évaluation du raisonnement clinique des résidents en hématologie par l’approche de concordance de script","year":2014,"lang":"fr","type":"dissertation","venue":"Open MIND","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Humanities; Philosophy; Psychology","score_opus":0.09751516460118276,"score_gpt":0.4101905739703057,"score_spread":0.31267540936912297,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2276416864","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9673989,0.0012572004,0.0052579707,0.0010096895,0.0019258254,0.0017991497,0.00003626803,0.00001007089,0.021304943],"genre_scores_gemma":[0.9127168,0.002449664,0.07191913,0.0007948658,0.00110835,0.00043581563,0.001671561,0.00008429794,0.008819516],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9952945,0.0010174345,0.0013261463,0.00091812894,0.0005342888,0.00090950495],"domain_scores_gemma":[0.9804689,0.01709262,0.0007636697,0.0006075121,0.0004613549,0.00060589483],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.006289122,0.000523014,0.0011395316,0.00009797833,0.00031075187,0.00025978836,0.00079152663,0.00094800675,0.0028961054],"category_scores_gemma":[0.17087927,0.00050684076,0.00025960442,0.00020707736,0.00029960362,0.00014866477,0.0001778339,0.0007731018,0.0017493189],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00080680446,0.0009814488,0.45612672,0.0002849268,0.0002932968,0.00019264415,0.011196723,0.00017978158,0.001114129,0.00018763971,0.005767395,0.5228685],"study_design_scores_gemma":[0.011747009,0.0022423216,0.8232765,0.044634417,0.0032042947,0.00022038701,0.012100234,0.011678141,0.047460075,0.0064430907,0.0354664,0.0015271035],"about_ca_topic_score_codex":0.0019431574,"about_ca_topic_score_gemma":0.0013620725,"teacher_disagreement_score":0.5213414,"about_ca_system_score_codex":0.0005599969,"about_ca_system_score_gemma":0.0025763153,"threshold_uncertainty_score":0.99973834},"labels":[],"label_agreement":null},{"id":"W2278007292","doi":"10.1177/1715163515625656","title":"How do community pharmacists make decisions? Results of an exploratory qualitative study in Ontario","year":2016,"lang":"en","type":"article","venue":"Canadian Pharmacists Journal / Revue des Pharmaciens du Canada","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":53,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"University of Toronto","funders":"","keywords":"Exploratory research; Qualitative research; Psychology; Business; Sociology; Social science","score_opus":0.14252054821938984,"score_gpt":0.38790776723136816,"score_spread":0.24538721901197832,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2278007292","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9961665,0.00021895974,0.00002836497,0.0002958257,0.0011779706,0.0007009145,0.00070418976,0.000022000753,0.00068523065],"genre_scores_gemma":[0.99677354,0.00071597163,0.0002136575,0.0008989221,0.00027713634,0.00003281063,0.000048386446,0.00006243873,0.0009771212],"study_design_codex":"design_other","study_design_gemma":"case_report","domain_scores_codex":[0.99417955,0.001919296,0.0014000155,0.0005446786,0.0009120083,0.0010444343],"domain_scores_gemma":[0.9861711,0.0074103815,0.00059384154,0.00076172035,0.0011390792,0.0039238636],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0046148775,0.00048961525,0.0009683749,0.00069804327,0.0005188138,0.00013751275,0.0007465165,0.00010911731,0.00034006103],"category_scores_gemma":[0.019371187,0.00039335966,0.00021318297,0.00073562854,0.00041191909,0.00046571135,0.00008994498,0.0017096475,0.0000072984126],"study_design_candidate":"case_report","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.007134177,0.0039017561,0.05494343,0.000113884416,0.000898821,0.21825592,0.18318594,0.00011124914,0.0036083707,0.00002521988,0.01514689,0.51267433],"study_design_scores_gemma":[0.2248837,0.008327472,0.024456274,0.00580802,0.0026511105,0.64328665,0.029709611,0.00051377167,0.017364306,0.0021281282,0.034906413,0.0059645306],"about_ca_topic_score_codex":0.31305444,"about_ca_topic_score_gemma":0.9541897,"teacher_disagreement_score":0.6411353,"about_ca_system_score_codex":0.004844703,"about_ca_system_score_gemma":0.010216614,"threshold_uncertainty_score":0.9998518},"labels":[],"label_agreement":null},{"id":"W2279635397","doi":"10.1007/s11739-016-1410-1","title":"Do emergency physicians trust their patients?","year":2016,"lang":"en","type":"article","venue":"Internal and Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":24,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Université de Montréal; Université de Sherbrooke","funders":"H2020 European Research Council; Université de Sherbrooke","keywords":"Medicine; Recall; Consistency (knowledge bases); Perspective (graphical); Emergency department; Variety (cybernetics); Cognition; Focus group; Point (geometry); Medical education; Nursing; Psychology; Psychiatry; Cognitive psychology; Computer science","score_opus":0.02955018736687497,"score_gpt":0.33306499909604803,"score_spread":0.30351481172917305,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2279635397","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9680292,0.0014712383,0.0004035214,0.006878647,0.0034897581,0.00016362063,0.0000221982,0.000060484403,0.019481309],"genre_scores_gemma":[0.9734303,0.0077123335,0.000018377208,0.00027644113,0.0015111675,0.000014155567,0.000012782382,0.000024155464,0.017000305],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99832684,0.000030193169,0.00064924784,0.00036852487,0.00031205182,0.00031312933],"domain_scores_gemma":[0.9985805,0.00048115884,0.00014685062,0.00028230192,0.0002065493,0.000302613],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00020330785,0.00022319678,0.00041928634,0.000078807214,0.00005621113,0.0000017184739,0.00011722944,0.00007943142,0.011504762],"category_scores_gemma":[0.0068380707,0.00010873046,0.00013936436,0.00014355162,0.00014846657,0.00006025263,0.000058795704,0.000144173,0.00013310424],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020340811,0.0002823444,0.715684,0.000025279913,0.000100043784,0.000011846042,0.00042702025,3.7221657e-8,0.0012576493,0.00025575334,0.086675234,0.1950774],"study_design_scores_gemma":[0.0040053697,0.0020433785,0.9406653,0.0040312503,0.00018824638,0.000020400626,0.00023085233,0.000022287495,0.0004042973,0.0029403528,0.045180377,0.00026786342],"about_ca_topic_score_codex":0.00016241534,"about_ca_topic_score_gemma":0.000013422616,"teacher_disagreement_score":0.22498135,"about_ca_system_score_codex":0.00002433376,"about_ca_system_score_gemma":0.0000125166325,"threshold_uncertainty_score":0.98939884},"labels":[],"label_agreement":null},{"id":"W2280717170","doi":"10.1007/978-3-319-19773-9_63","title":"Towards Investigating Performance Differences in Clinical Reasoning in a Technology Rich Learning Environment","year":2015,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Correctness; Computer science; Metacognition; Artificial intelligence; Human–computer interaction; Cognition; Psychology; Programming language","score_opus":0.05251419946581889,"score_gpt":0.3267423188506005,"score_spread":0.27422811938478164,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2280717170","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94092804,0.0016618406,0.049943954,0.0018472467,0.00073869026,0.00061034353,0.0000015584031,0.00011637254,0.0041519557],"genre_scores_gemma":[0.9459709,0.00042559791,0.052673377,0.00041456736,0.00027592803,0.000014637259,0.000006693936,0.000027789132,0.000190526],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99631363,0.00007018747,0.0010691695,0.0011562569,0.0007847918,0.000605952],"domain_scores_gemma":[0.995075,0.0037495566,0.00033227756,0.0005127352,0.00008027524,0.00025015185],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0028983557,0.00038444082,0.0010287822,0.00087548804,0.00007325243,0.000050355262,0.000551792,0.0007023291,0.000029268327],"category_scores_gemma":[0.02524528,0.0003237494,0.00008022206,0.00056740915,0.0014117165,0.000083885854,0.0006053017,0.0029732413,0.000031764102],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000017766799,0.00006250639,0.48736212,0.00002368239,0.00000517952,0.00012685868,0.00030225675,0.005174651,0.0000058431574,0.00018833914,0.0000026782855,0.5067281],"study_design_scores_gemma":[0.0038966544,0.0031630266,0.4486432,0.034962613,0.00006245757,0.0002450489,0.000010166265,0.4699042,0.00019611865,0.036998093,0.00056804146,0.0013503474],"about_ca_topic_score_codex":0.000054944274,"about_ca_topic_score_gemma":0.00004410698,"teacher_disagreement_score":0.50537777,"about_ca_system_score_codex":0.00042713236,"about_ca_system_score_gemma":0.00085663877,"threshold_uncertainty_score":0.99992144},"labels":[],"label_agreement":null},{"id":"W2289784479","doi":"10.1071/hc12166","title":"Viewpoint: Diagnosis in primary care: probabilistic reasoning","year":2012,"lang":"en","type":"article","venue":"Journal of Primary Health Care","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia; University of Alberta","funders":"","keywords":"Probabilistic logic; Test (biology); Fagan inspection; Nomogram; Pre- and post-test probability; Computer science; Medical diagnosis; Machine learning; Artificial intelligence; Statistics; Medicine; Mathematics","score_opus":0.023936072754902487,"score_gpt":0.3425943368982359,"score_spread":0.31865826414333337,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2289784479","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.74586874,0.2262771,0.00006314122,0.015473002,0.0019349047,0.0009622054,0.00002786172,0.000046804504,0.009346253],"genre_scores_gemma":[0.95124346,0.007925265,0.010543887,0.02821635,0.0018248191,0.000041837757,0.000095951225,0.00006040974,0.000048019818],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99689627,0.00023349437,0.0013390275,0.00020725827,0.0006706156,0.00065334747],"domain_scores_gemma":[0.99392503,0.0036778455,0.00069854414,0.00031324048,0.00044363915,0.0009416776],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001446182,0.00021792746,0.0010827488,0.00023546016,0.00008215151,0.00001591726,0.00014752123,0.00016359436,0.00006120793],"category_scores_gemma":[0.01423303,0.00017522274,0.0002847064,0.0003355028,0.000070414644,0.00019308476,0.00006941962,0.00089584605,0.000022725304],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00055327377,0.00083630835,0.76245517,0.0032488953,0.00005143847,0.00022794715,0.007919348,0.000025882073,0.000010676498,0.00014896857,0.015703108,0.20881902],"study_design_scores_gemma":[0.0033965134,0.001417576,0.96336526,0.011044175,0.00014290708,0.0005141612,0.001156636,0.0000014878495,0.000023840603,0.00004677821,0.018694814,0.00019582902],"about_ca_topic_score_codex":0.00009487745,"about_ca_topic_score_gemma":0.000020433145,"teacher_disagreement_score":0.21835184,"about_ca_system_score_codex":0.0030157315,"about_ca_system_score_gemma":0.0024497916,"threshold_uncertainty_score":0.99407053},"labels":[],"label_agreement":null},{"id":"W2290537467","doi":"10.1002/j.0022-0337.2016.80.3.tb06090.x","title":"Three Modeling Applications to Promote Automatic Item Generation for Examinations in Dentistry","year":2016,"lang":"en","type":"article","venue":"Journal of Dental Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Computer science; Test (biology); Item bank; Cognition; Domain (mathematical analysis); Item response theory; Information retrieval; Artificial intelligence; Psychology; Psychometrics; Mathematics; Clinical psychology","score_opus":0.056061734017115186,"score_gpt":0.40163017521939237,"score_spread":0.3455684412022772,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2290537467","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.82117826,0.000094299074,0.17636093,0.0014234497,0.00040124464,0.00048402572,0.000004143747,0.000006912436,0.000046748075],"genre_scores_gemma":[0.97558564,0.000017898952,0.02323186,0.00015604419,0.0006065261,0.00013977781,0.000013413643,0.000010201777,0.00023864838],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99914724,0.000010681958,0.00046857886,0.00010531009,0.00017562999,0.0000925417],"domain_scores_gemma":[0.99868625,0.00062438624,0.00016593233,0.0001131988,0.0002750161,0.00013520064],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00032991634,0.0000600549,0.00014238199,0.0001827919,0.00003634924,0.000020107016,0.00005946238,0.000049691436,0.00003638237],"category_scores_gemma":[0.0069878,0.000043714077,0.0000695706,0.00015506176,0.000009726764,0.000112364396,0.000008594421,0.00005969515,0.000026040752],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006257607,0.002189416,0.037247002,0.00007135287,0.000047153746,0.0000045178317,0.00041866416,0.00017809933,0.048443206,0.00093904266,0.0048857746,0.90551317],"study_design_scores_gemma":[0.016421875,0.0034445792,0.6347854,0.02284174,0.001310226,0.0031658728,0.0034363465,0.26996374,0.02280623,0.01594106,0.004791267,0.0010916439],"about_ca_topic_score_codex":0.000010501833,"about_ca_topic_score_gemma":0.00004759071,"teacher_disagreement_score":0.90442157,"about_ca_system_score_codex":0.00022632419,"about_ca_system_score_gemma":0.00036292762,"threshold_uncertainty_score":0.83655536},"labels":[],"label_agreement":null},{"id":"W2297851153","doi":"10.1136/bmjqs-2015-005014","title":"Is bias in the eye of the beholder? A vignette study to assess recognition of cognitive biases in clinical case workups","year":2016,"lang":"en","type":"article","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":134,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"McMaster University","keywords":"Cognitive bias; Vignette; Hindsight bias; Cognition; Confirmation bias; Identification (biology); Medicine; Outcome (game theory); Reporting bias; MEDLINE; Clinical psychology; Psychology; Psychiatry; Cognitive psychology; Social psychology","score_opus":0.4710586535600591,"score_gpt":0.5460120951856193,"score_spread":0.07495344162556022,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2297851153","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98899126,0.000030362504,0.000074320196,0.008063263,0.00014356695,0.0020450011,0.00026281556,0.000010842801,0.0003785843],"genre_scores_gemma":[0.9967915,0.000050480234,0.00006166537,0.0028201318,0.000114797156,0.000074030555,0.000012821043,0.00001186936,0.000062724364],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9934519,0.003014331,0.00235954,0.00038178245,0.00054078887,0.00025166795],"domain_scores_gemma":[0.8171182,0.18097301,0.0006266151,0.00077577727,0.00034997813,0.00015639485],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.014160116,0.00015485502,0.0006941383,0.00007983428,0.000038946724,0.0000069322696,0.00017183309,0.000121371515,0.00010778487],"category_scores_gemma":[0.3476381,0.0000749264,0.0002601502,0.0006132211,0.00024323375,0.00004188707,0.00011794221,0.00031062082,0.000021492535],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017169405,0.0043065557,0.92182475,0.000021947399,0.00006071485,0.00025776858,0.0038590748,9.758868e-7,0.000019230985,0.000016267935,0.00044887714,0.06746691],"study_design_scores_gemma":[0.0054562283,0.0008967434,0.97194964,0.008944882,0.00016546503,0.000035068308,0.011991202,0.000004971143,0.000228902,0.00019019868,0.00001897232,0.00011774136],"about_ca_topic_score_codex":0.0018036463,"about_ca_topic_score_gemma":0.002283633,"teacher_disagreement_score":0.33347797,"about_ca_system_score_codex":0.000043482076,"about_ca_system_score_gemma":0.00027373238,"threshold_uncertainty_score":0.657857},"labels":[],"label_agreement":null},{"id":"W2298314988","doi":"10.7287/peerj.preprints.1859v1","title":"Rushing Roulette – how do learners perform routine tasks under time pressure?","year":2016,"lang":"en","type":"preprint","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Computer science; Adaptation (eye); Task (project management); Process (computing); Resource (disambiguation); Coaching; Variety (cybernetics); Artificial intelligence; Analytics; Machine learning; Human–computer interaction; Data science; Psychology","score_opus":0.027236613081984547,"score_gpt":0.3159742843145293,"score_spread":0.28873767123254473,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2298314988","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.47019342,0.0040237834,0.0059263376,0.23743288,0.002906314,0.0023647642,0.00023032195,0.0015313516,0.27539083],"genre_scores_gemma":[0.8682028,0.00022341745,0.0011986778,0.0033063102,0.0016143359,0.000042773747,0.00021391947,0.00009526894,0.12510246],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9970364,0.00009548867,0.0005141353,0.0009915308,0.0007288483,0.0006335775],"domain_scores_gemma":[0.9919338,0.0058075027,0.0002919984,0.0012519769,0.00018608509,0.00052864186],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00094767107,0.00051939226,0.0010369447,0.0001312153,0.00009001348,0.00019357719,0.00033701034,0.0008998245,0.00493427],"category_scores_gemma":[0.02129506,0.0003364452,0.0004791431,0.00008668585,0.0001742809,0.00007102681,0.000727164,0.0015178056,0.0011027394],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009870712,0.0011612349,0.07712723,0.0007480124,0.003470794,0.00038017044,0.0006894826,0.00065684016,0.00080676045,0.0053876615,0.819041,0.089543745],"study_design_scores_gemma":[0.05422949,0.004161548,0.27239895,0.1431506,0.020166961,0.0009951055,0.0018882152,0.03981688,0.006202275,0.051190536,0.39303252,0.012766912],"about_ca_topic_score_codex":0.00010649988,"about_ca_topic_score_gemma":0.0000031748136,"teacher_disagreement_score":0.42600846,"about_ca_system_score_codex":0.00012403601,"about_ca_system_score_gemma":0.00036615354,"threshold_uncertainty_score":0.99990875},"labels":[],"label_agreement":null},{"id":"W2298620443","doi":"10.1136/bmjqs-2015-005150","title":"Unwanted patients and unwanted diagnostic errors","year":2016,"lang":"en","type":"letter","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; Sunnybrook Health Science Centre; Institute for Clinical Evaluative Sciences; University of Toronto","funders":"Canadian Institutes of Health Research","keywords":"Medicine; Medical emergency; Patient safety; Intensive care medicine; Medical physics; Risk analysis (engineering); Health care","score_opus":0.04366374722465598,"score_gpt":0.38724099526965977,"score_spread":0.3435772480450038,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2298620443","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.04008528,0.00047994795,0.0002462892,0.9497673,0.0016202267,0.0022103244,0.0017548449,0.00031930167,0.003516507],"genre_scores_gemma":[0.036973555,0.0016933334,0.00031965968,0.93770367,0.0064357575,0.0001535235,0.006103507,0.00021319762,0.010403791],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99401814,0.00068857515,0.0022332019,0.0010812504,0.0011292637,0.00084957306],"domain_scores_gemma":[0.9187239,0.078440994,0.0007508008,0.0011093025,0.0005096526,0.00046536615],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.001720229,0.0006555593,0.0016341292,0.00015438157,0.00015950619,0.00003663848,0.00023701345,0.0014253049,0.00054008135],"category_scores_gemma":[0.23154503,0.0004637273,0.00042299202,0.0002265995,0.0004904078,0.00006509155,0.00022406269,0.0017580995,0.00046197706],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00042129017,0.00019177597,0.06500113,0.00034868304,0.00024229453,0.00046640667,0.00008554432,8.1128704e-8,0.0000044580133,0.000026105801,0.9133003,0.019911932],"study_design_scores_gemma":[0.007500927,0.0005301773,0.30975583,0.011837394,0.0007036465,0.00004513688,0.000042603402,0.000005301908,0.000028643633,0.00049382256,0.66814405,0.0009124772],"about_ca_topic_score_codex":0.0001714781,"about_ca_topic_score_gemma":0.000017705794,"teacher_disagreement_score":0.24515626,"about_ca_system_score_codex":0.00020508625,"about_ca_system_score_gemma":0.00051881623,"threshold_uncertainty_score":0.9998711},"labels":[],"label_agreement":null},{"id":"W2299692357","doi":"10.14288/1.0076726","title":"When Experts Disagree: The Art of Medical Decision Making","year":2013,"lang":"en","type":"article","venue":"cIRcle (University of British Columbia)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Computer science","score_opus":0.013616275418466776,"score_gpt":0.24882846063161412,"score_spread":0.23521218521314735,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2299692357","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99280113,0.0002308491,0.00110019,0.002531171,0.00013695467,0.00024466883,0.000012771692,0.000028357721,0.002913896],"genre_scores_gemma":[0.99696404,0.00020616615,0.0010952202,0.00037995685,0.000051805317,6.779944e-7,0.000007643832,0.000008657194,0.0012858608],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99864316,0.000041689927,0.00018701352,0.00022571905,0.0007292598,0.00017318066],"domain_scores_gemma":[0.99625534,0.002919456,0.00011290654,0.00033749902,0.00018508286,0.00018973295],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00034307505,0.00003438642,0.00030379344,0.000021512325,0.0000915807,0.000032209107,0.0002985202,0.00015264523,0.0032507477],"category_scores_gemma":[0.0066173933,0.000074328775,0.0001615926,0.00012965752,0.0004293662,0.00008827935,0.00017685826,0.00014746217,0.00008353784],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000016435395,0.00019761283,0.03981033,0.000019458184,0.00003247589,0.000114952454,0.00021885008,4.8172234e-7,0.000016336135,0.0000030141027,0.14075544,0.81881464],"study_design_scores_gemma":[0.001275515,0.00012377337,0.9909842,0.0035969038,0.000058915746,0.00012749922,0.0014486072,0.00031906745,5.425893e-7,0.001223018,0.0007653356,0.0000766124],"about_ca_topic_score_codex":0.034736536,"about_ca_topic_score_gemma":0.032602802,"teacher_disagreement_score":0.9511739,"about_ca_system_score_codex":0.000025761221,"about_ca_system_score_gemma":0.00010632669,"threshold_uncertainty_score":0.9976604},"labels":[],"label_agreement":null},{"id":"W2301136743","doi":"10.1007/s11606-016-3592-7","title":"What’s in a Label? Is Diagnosis the Start or the End of Clinical Reasoning?","year":2016,"lang":"en","type":"review","venue":"Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":75,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"University of California, San Francisco; University of Washington","keywords":"Medicine","score_opus":0.17333649712810154,"score_gpt":0.49900209630624437,"score_spread":0.32566559917814286,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2301136743","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.001489248,0.963817,0.000019381658,0.027541216,0.006000131,0.0005442059,0.000022968617,0.000005636323,0.00056022406],"genre_scores_gemma":[0.00021902875,0.9768896,0.00018869112,0.004181085,0.008141157,0.00003234386,0.0000049400796,0.000058085443,0.010285073],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9916206,0.0009397286,0.005364794,0.00037088103,0.0012890333,0.00041494364],"domain_scores_gemma":[0.94469833,0.049023267,0.0043335203,0.0008327291,0.00063113094,0.0004810037],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00726743,0.0005031797,0.0047839563,0.0003104903,0.000044769386,0.000035838046,0.0009911001,0.00044770783,0.002001389],"category_scores_gemma":[0.09750149,0.00014178308,0.0013145139,0.0004571216,0.0010509259,0.00012121966,0.00018873397,0.0023283528,0.000033540422],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006228304,0.0003451823,0.0026814786,0.0006278289,0.0008244565,0.0007336418,0.00021428127,9.7501555e-8,3.215644e-7,0.000046451976,0.13479662,0.85910684],"study_design_scores_gemma":[0.0027361088,0.0028977478,0.0009121791,0.30448478,0.0018511222,0.001621479,0.00010784123,0.000007289634,0.0000034586412,0.00008867458,0.68516695,0.00012233696],"about_ca_topic_score_codex":0.00031181704,"about_ca_topic_score_gemma":0.00006814375,"teacher_disagreement_score":0.8589845,"about_ca_system_score_codex":0.00015061231,"about_ca_system_score_gemma":0.001235044,"threshold_uncertainty_score":0.9999733},"labels":[],"label_agreement":null},{"id":"W2307036059","doi":"10.1002/j.0022-0337.2014.78.9.tb05799.x","title":"Dental and Dental Hygiene Students’ Diagnostic Accuracy in Oral Radiology: Effect of Diagnostic Strategy and Instructional Method","year":2014,"lang":"en","type":"article","venue":"Journal of Dental Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":32,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; University of Toronto; Memorial University of Newfoundland","funders":"","keywords":"Diagnostic accuracy; Dental hygiene; Oral hygiene; Interpretation (philosophy); Radiography; Medical physics; Medicine; Psychology; Radiology; Medical education; Computer science; Dentistry","score_opus":0.01200523155713096,"score_gpt":0.38893399865280404,"score_spread":0.3769287670956731,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2307036059","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9970128,0.0014550546,0.00009964545,0.00010184955,0.0009463244,0.00028117324,0.0000049532737,0.0000048768343,0.00009333003],"genre_scores_gemma":[0.99721074,0.00067884504,0.001587707,0.00010990716,0.00033359887,0.000011563426,0.00001903208,0.0000139668455,0.000034621036],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982107,0.000333653,0.0007023633,0.00019951964,0.0003833682,0.00017036055],"domain_scores_gemma":[0.9671211,0.03195811,0.00044637418,0.00010784638,0.00010430366,0.00026226704],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0011893342,0.00016681345,0.0005634709,0.00026249414,0.000038330487,0.000034726527,0.00010378901,0.00012083774,0.00005297939],"category_scores_gemma":[0.082280524,0.0001330886,0.00009877906,0.00016642037,0.00015605603,0.00018824528,0.000057694488,0.00033455904,0.0000046963455],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00030144778,0.0006098156,0.9106993,0.00005713062,0.000060327064,0.000025501206,0.00009130073,0.00001875088,0.0005979104,0.00008707466,0.00018435552,0.08726707],"study_design_scores_gemma":[0.0046083676,0.0024044947,0.98431015,0.0012496936,0.00029724374,0.004928862,0.0006133574,0.000068084875,0.001143352,0.00022253027,0.000048657042,0.000105185434],"about_ca_topic_score_codex":0.00013965748,"about_ca_topic_score_gemma":0.000049001534,"teacher_disagreement_score":0.08716189,"about_ca_system_score_codex":0.000115174174,"about_ca_system_score_gemma":0.0001938281,"threshold_uncertainty_score":0.9254498},"labels":[],"label_agreement":null},{"id":"W2311934776","doi":"10.1007/s10459-016-9671-z","title":"A bridge too far","year":2016,"lang":"en","type":"editorial","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Bridge (graph theory); Computer science; Medicine","score_opus":0.023421164070074673,"score_gpt":0.46810794253996474,"score_spread":0.4446867784698901,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2311934776","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005198738,0.036299396,0.000094533345,0.008029282,0.9480434,0.00055841246,0.000033586985,0.000048178164,0.0063733365],"genre_scores_gemma":[0.0027115392,0.10140252,0.0025650794,0.00364052,0.8850264,0.00020427124,0.0002125282,0.000027914317,0.004209213],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9971199,0.00009710166,0.0007421222,0.00068351364,0.0008657222,0.0004916088],"domain_scores_gemma":[0.9826026,0.015946418,0.00051786465,0.00036758152,0.00023135662,0.00033420065],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0017918275,0.00019499443,0.0005103666,0.00025669945,0.00014385869,0.0000317145,0.00025951592,0.00027551118,0.00006425648],"category_scores_gemma":[0.07782073,0.0001405683,0.000071454815,0.00053425157,0.00037289428,0.00024597062,0.000035924597,0.00045498498,0.00009659849],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000020984953,0.0001729816,0.0036637024,0.0001872517,0.0000013652013,9.923518e-7,0.00007539934,0.0000022879246,1.7169062e-7,0.00030998644,0.7526123,0.24295262],"study_design_scores_gemma":[0.0003852326,0.0005389405,0.004820701,0.007573865,0.000008082366,0.0000040251653,0.00004615519,0.000007493283,0.0000013373179,0.0021880507,0.98427856,0.0001475651],"about_ca_topic_score_codex":0.00086960586,"about_ca_topic_score_gemma":0.00019036344,"teacher_disagreement_score":0.24280505,"about_ca_system_score_codex":0.00045808963,"about_ca_system_score_gemma":0.010828295,"threshold_uncertainty_score":0.9947794},"labels":[],"label_agreement":null},{"id":"W2312287125","doi":"10.1097/acm.0b013e31822a70be","title":"Exploring Residentsʼ Perceptions of Expertise and Expert Development","year":2011,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":35,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"SickKids Foundation","funders":"","keywords":"Conceptualization; Perception; Dominance (genetics); Inclusion (mineral); Process (computing); Medical education; Psychology; Grounded theory; Applied psychology; Computer science; Medicine; Qualitative research; Sociology; Social psychology; Artificial intelligence","score_opus":0.3147470952936216,"score_gpt":0.39526898858928294,"score_spread":0.08052189329566134,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2312287125","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99220425,0.002410756,0.0003726499,0.0010820093,0.0002563096,0.00016981339,4.5588604e-7,0.00005022643,0.003453502],"genre_scores_gemma":[0.98783845,0.008480118,0.0016212501,0.0010019325,0.00031179585,0.000060963073,0.000007453972,0.000017019394,0.00066102244],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99863327,0.000027700034,0.00057247485,0.000247029,0.00031197298,0.00020756938],"domain_scores_gemma":[0.99824387,0.0010405435,0.00009663328,0.0002154224,0.000071931805,0.00033162313],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00040876016,0.00013106027,0.00040186933,0.00012123734,0.00004575594,7.447334e-7,0.00008267419,0.00011803219,0.00065930386],"category_scores_gemma":[0.011332192,0.00009468703,0.000028579616,0.00012791014,0.00027172314,0.00007361183,0.000060862043,0.00031151474,0.00002568096],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00090260623,0.00019766929,0.4291986,0.00017322929,0.00022351951,0.0001417446,0.31063703,2.0622427e-7,0.027330188,0.0015639677,0.019091487,0.21053979],"study_design_scores_gemma":[0.0040942747,0.0009201183,0.94592935,0.010738968,0.00016988063,0.00014150816,0.018401073,0.000022579114,0.01549064,0.0004064302,0.0034144006,0.00027077462],"about_ca_topic_score_codex":0.00021472976,"about_ca_topic_score_gemma":0.0000020827747,"teacher_disagreement_score":0.5167308,"about_ca_system_score_codex":0.000034291872,"about_ca_system_score_gemma":0.00006734245,"threshold_uncertainty_score":0.99699575},"labels":[],"label_agreement":null},{"id":"W2313586819","doi":"10.1097/acm.0b013e31823a91e2","title":"Black Balls and Diagnostic Reasoning","year":2011,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"","keywords":"Medical diagnosis; Complaint; Ball (mathematics); Mathematics education; Medical education; Computer science; Medicine; Psychology; Mathematics; Pathology","score_opus":0.04755836845215343,"score_gpt":0.3394408998282956,"score_spread":0.2918825313761422,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2313586819","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.014177015,0.014414382,0.0001435128,0.9420756,0.001413951,0.0010268036,0.000024224424,0.00028760615,0.026436923],"genre_scores_gemma":[0.014385239,0.024078691,0.00020963859,0.9152356,0.020703748,0.00006029263,0.00063431554,0.0001639242,0.024528557],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9960121,0.00014726126,0.0011082395,0.0010181287,0.00087471533,0.0008395553],"domain_scores_gemma":[0.9620096,0.03600264,0.00045177923,0.00076386717,0.0001638216,0.00060828996],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.001006848,0.00065992115,0.001690449,0.00028853826,0.00007346198,0.000008450351,0.00031679813,0.0028818543,0.0016743274],"category_scores_gemma":[0.19780542,0.0004820643,0.00013815843,0.00025927724,0.001252082,0.000052815325,0.00017424578,0.007983594,0.0004695572],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004067698,0.000009336038,0.008646761,0.00027530445,0.00014915694,0.0019186146,0.00096866343,6.647453e-8,0.0000109589855,0.00017601746,0.9847317,0.0030727596],"study_design_scores_gemma":[0.003511367,0.0010096674,0.010495691,0.040877044,0.002363864,0.0011429801,0.00018585316,0.00003694044,0.000025397514,0.002482237,0.93722254,0.000646429],"about_ca_topic_score_codex":0.00025424684,"about_ca_topic_score_gemma":0.0000011629177,"teacher_disagreement_score":0.19679856,"about_ca_system_score_codex":0.00008279373,"about_ca_system_score_gemma":0.00021242313,"threshold_uncertainty_score":0.99976313},"labels":[],"label_agreement":null},{"id":"W2314875266","doi":"10.1097/acm.0b013e3182a3116f","title":"Understanding Clinical Uncertainty","year":2013,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":68,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"Canadian Institutes of Health Research","keywords":"Novelty; Action (physics); Psychology; Qualitative research; Computer science; Medical education; Management science; Medicine; Social psychology","score_opus":0.29919923060637654,"score_gpt":0.45243530938688775,"score_spread":0.1532360787805112,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2314875266","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8244631,0.0014465861,0.006992123,0.10015538,0.0020879563,0.0010655185,0.0000022374472,0.00036582194,0.06342123],"genre_scores_gemma":[0.97682047,0.0016499072,0.00015683971,0.0148600675,0.002122999,0.000025350322,0.0000323414,0.00002384284,0.004308177],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99771804,0.00009250968,0.00093471754,0.0003895884,0.00046705265,0.0003980669],"domain_scores_gemma":[0.9849327,0.01368383,0.00016471678,0.00037197108,0.00010629746,0.0007404885],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0013906425,0.00017844923,0.0006362438,0.00009332326,0.00006360524,0.000004813552,0.00014747313,0.00041044533,0.003575408],"category_scores_gemma":[0.07537547,0.0001166797,0.00010659673,0.00025586036,0.000549093,0.000060971404,0.000053929354,0.0012478946,0.0009124226],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009865506,0.000046388905,0.27847013,0.000030834948,0.0001142006,0.000052772215,0.000496501,0.0000049459004,0.0002437792,0.010940049,0.69573313,0.0137686],"study_design_scores_gemma":[0.035075963,0.006343059,0.6421132,0.022788426,0.0018216431,0.0009392966,0.012531825,0.00896823,0.00013428135,0.18147334,0.086422436,0.001388288],"about_ca_topic_score_codex":0.00020615701,"about_ca_topic_score_gemma":0.0000014168102,"teacher_disagreement_score":0.6093107,"about_ca_system_score_codex":0.00012074956,"about_ca_system_score_gemma":0.00011882522,"threshold_uncertainty_score":0.9998655},"labels":[],"label_agreement":null},{"id":"W2318340323","doi":"10.1017/s031716710000723x","title":"An Unusual Patient Request","year":2009,"lang":"fr","type":"article","venue":"Canadian Journal of Neurological Sciences / Journal Canadien des Sciences Neurologiques","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Toronto Western Hospital","funders":"","keywords":"Action (physics); Content (measure theory); Computer science; Medicine; Mathematics; Physics","score_opus":0.04191117804428899,"score_gpt":0.3297899233158421,"score_spread":0.2878787452715531,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2318340323","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93796223,0.01785693,0.00003464099,0.030328577,0.0064183343,0.00022729214,0.00003261371,0.00002835729,0.007111024],"genre_scores_gemma":[0.97205937,0.0025078242,0.0028541056,0.020131884,0.0023161718,0.000001203407,0.0000010869807,0.00001864193,0.000109706176],"study_design_codex":"observational","study_design_gemma":"nonrandomized_trial","domain_scores_codex":[0.9901447,0.0023182589,0.0020088986,0.00094627246,0.0015286449,0.00305322],"domain_scores_gemma":[0.9833826,0.005015576,0.0014037762,0.0003360418,0.0010341067,0.008827929],"candidate_categories":["metaresearch","metaepi_narrow","sts","scholarly_communication","research_integrity"],"consensus_categories":["sts"],"category_scores_codex":[0.008094625,0.0006267401,0.0011154447,0.0012026293,0.0032120377,0.0013314676,0.0024808107,0.00060755876,0.0006708991],"category_scores_gemma":[0.04794666,0.00044388563,0.00054086605,0.002060016,0.013446245,0.0016746477,0.000049441373,0.0030522358,0.000030025065],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00037413542,0.0006134161,0.67491394,0.000013555902,0.000044681838,0.07530094,0.0013146157,0.0072919135,0.00006169287,0.0085997,0.0073769786,0.22409442],"study_design_scores_gemma":[0.0008690328,0.52021885,0.31159523,0.0011481326,0.00025997005,0.108228534,0.0006706958,0.002698285,0.00013605099,0.041402385,0.012047978,0.000724867],"about_ca_topic_score_codex":0.0028500308,"about_ca_topic_score_gemma":0.02595725,"teacher_disagreement_score":0.5196054,"about_ca_system_score_codex":0.0006113296,"about_ca_system_score_gemma":0.011348117,"threshold_uncertainty_score":0.9998013},"labels":[],"label_agreement":null},{"id":"W2318619027","doi":"10.1097/acm.0b013e31823c3e86","title":"Black Balls and Diagnostic Reasoning","year":2011,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Calgary","funders":"","keywords":"Information processing; Data processing; Computer science; Statement (logic); Analytic reasoning; Electronic data processing; Artificial intelligence; Cognitive psychology; Psychology; Data science; Epistemology; Deductive reasoning","score_opus":0.060893813664462004,"score_gpt":0.3468235018620387,"score_spread":0.2859296881975767,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2318619027","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94681114,0.0044089095,0.00047187626,0.0044733556,0.00036208596,0.00038899132,0.0000022898626,0.00016006095,0.042921294],"genre_scores_gemma":[0.98475564,0.0061551253,0.0005739468,0.0052189757,0.00080672326,0.000016684413,0.000017777258,0.000030534655,0.0024245738],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9983099,0.000050067338,0.0005177143,0.000403802,0.00034205057,0.00037644187],"domain_scores_gemma":[0.98825127,0.010591445,0.00012806979,0.0003157967,0.000079207595,0.0006341842],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0007906932,0.00020693001,0.0005297455,0.00010391952,0.000050316772,0.0000024623162,0.000116979594,0.00029168313,0.0010906236],"category_scores_gemma":[0.13606359,0.00014751326,0.000043352837,0.00020509362,0.0006309648,0.000053301592,0.00007400017,0.00077778514,0.00018318584],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00035535492,0.00009327825,0.8429016,0.00014811341,0.00018402294,0.00067235256,0.015698288,5.3028276e-7,0.00076896616,0.017251402,0.087193884,0.03473221],"study_design_scores_gemma":[0.008221539,0.0021709597,0.94466203,0.017142884,0.0011719038,0.0009330142,0.0028112791,0.00036346324,0.00091703096,0.009566474,0.01150793,0.0005314702],"about_ca_topic_score_codex":0.00016882848,"about_ca_topic_score_gemma":0.0000016898575,"teacher_disagreement_score":0.13527289,"about_ca_system_score_codex":0.000026298632,"about_ca_system_score_gemma":0.0000620739,"threshold_uncertainty_score":0.9998225},"labels":[],"label_agreement":null},{"id":"W2318752613","doi":"10.1097/mcc.0000000000000146","title":"Default options in the ICU","year":2014,"lang":"en","type":"review","venue":"Current Opinion in Critical Care","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":25,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute of Health Economics","funders":"National Heart, Lung, and Blood Institute","keywords":"Default; Medicine; Harm; Psychological intervention; Health care; Unintended consequences; Quality management; Quality (philosophy); Risk analysis (engineering); MEDLINE; Intensive care medicine; Actuarial science; Business; Nursing; Service (business); Marketing; Economics; Finance; Psychology","score_opus":0.33434805723857647,"score_gpt":0.5737227716038373,"score_spread":0.23937471436526087,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2318752613","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000029154112,0.9929467,0.000059318747,0.0007936072,0.004391681,0.0007249905,0.000079948775,0.000028965593,0.0009718939],"genre_scores_gemma":[0.0005638169,0.99655885,0.0000690361,0.00010872413,0.0014059221,0.00037319213,0.0008843009,0.000028908667,0.0000072476023],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9973872,0.00045908222,0.000868776,0.0004872175,0.000412311,0.00038547174],"domain_scores_gemma":[0.97395813,0.025045149,0.00006421668,0.00059394125,0.00013811547,0.00020045052],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0005794043,0.00030221953,0.0013242922,0.00018867721,0.000044619643,0.000042105923,0.00027110023,0.00036361176,0.0000897007],"category_scores_gemma":[0.071068615,0.00018853154,0.0005067802,0.00044887152,0.00018216342,0.0000217584,0.00007857819,0.0014589636,0.0002438652],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000039604292,0.0002931536,0.00019319961,0.028432736,0.000004665803,0.000009808462,0.0001277296,2.8665463e-7,7.496605e-10,0.009622719,0.0048944256,0.9564173],"study_design_scores_gemma":[0.0003509968,0.00010013871,0.00032833693,0.16936831,0.00018529317,0.000042250755,0.00017278039,0.000006659265,3.540733e-9,0.00023672353,0.8290392,0.00016926244],"about_ca_topic_score_codex":0.000014142668,"about_ca_topic_score_gemma":0.0000045098686,"teacher_disagreement_score":0.95624804,"about_ca_system_score_codex":0.00015466848,"about_ca_system_score_gemma":0.00031980872,"threshold_uncertainty_score":0.9367562},"labels":[],"label_agreement":null},{"id":"W2320000683","doi":"10.1080/08957347.2016.1171768","title":"Evaluating the Psychometric Characteristics of Generated Multiple-Choice Test Items","year":2016,"lang":"en","type":"article","venue":"Applied Measurement in Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":33,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Medical Council of Canada; University of Ottawa; University of Alberta","funders":"","keywords":"Item response theory; Item bank; Item analysis; Licensure; Psychology; Test (biology); Psychometrics; Cognition; Multiple choice; Differential item functioning; Applied psychology; Cognitive psychology; Computer science; Clinical psychology; Statistics; Medicine; Medical education","score_opus":0.1433803510187834,"score_gpt":0.4009891954820371,"score_spread":0.2576088444632537,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2320000683","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9957502,0.0002490041,0.00019366678,0.00143513,0.0005333007,0.00066714745,0.0000042735855,0.000020387199,0.0011468848],"genre_scores_gemma":[0.99823135,0.00011577234,0.0005720544,0.00038724253,0.00034598805,0.00019399846,0.000020027086,0.000014711791,0.000118851196],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99849504,0.00004498626,0.0005079143,0.00022255398,0.0005637342,0.00016579857],"domain_scores_gemma":[0.9879235,0.010806289,0.0002755947,0.00040492343,0.0004985972,0.00009109909],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001702891,0.00010678307,0.00021307752,0.00014680834,0.000036409976,0.000009104926,0.000089508314,0.0000664815,0.000051142266],"category_scores_gemma":[0.13965367,0.00006069112,0.000039180184,0.00057580497,0.000047642352,0.0000193779,0.000012353493,0.000104613486,0.00003333936],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000056151508,0.0010163079,0.480281,0.000018424364,0.000013885147,4.7962367e-8,0.00008952293,0.0000014802366,0.15043533,0.00006678365,0.0005750424,0.36744604],"study_design_scores_gemma":[0.0016364981,0.00019185043,0.9893171,0.0009796123,0.00007346819,0.0000014695346,0.00007776323,0.000075811804,0.0069060638,0.00010181484,0.00055250636,0.00008602907],"about_ca_topic_score_codex":0.00007539517,"about_ca_topic_score_gemma":0.000013831017,"teacher_disagreement_score":0.5090361,"about_ca_system_score_codex":0.0001921021,"about_ca_system_score_gemma":0.0004330451,"threshold_uncertainty_score":0.8675934},"labels":[],"label_agreement":null},{"id":"W2320565199","doi":"10.1097/acm.0b013e31828a3d7f","title":"Clinical Decision Making, Fast and Slow","year":2013,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Dalhousie University","funders":"","keywords":"Premise; Fallacy; Variable (mathematics); Statement (logic); Interpretation (philosophy); Cognitive psychology; Anticipation (artificial intelligence); Computer science; Categorical variable; Categorization; Epistemology; Psychology; Confirmation bias; Social psychology; Artificial intelligence; Machine learning; Philosophy; Mathematics","score_opus":0.055406039043061295,"score_gpt":0.4232631908556358,"score_spread":0.3678571518125745,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2320565199","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.022953153,0.009443054,0.0009235563,0.953981,0.0034615577,0.0009880543,0.000015831962,0.00021339423,0.008020408],"genre_scores_gemma":[0.008390626,0.014137653,0.0010335948,0.92052907,0.034035645,0.00004542746,0.00028572127,0.000115575174,0.021426704],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9949391,0.00018539828,0.0019462279,0.001106395,0.0011375328,0.00068535196],"domain_scores_gemma":[0.9524444,0.045406822,0.0005717611,0.0008458769,0.00020334899,0.0005277927],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0017758048,0.0005671363,0.0020046202,0.00026525176,0.000071934606,0.0000124389435,0.00034808274,0.004169404,0.0019631914],"category_scores_gemma":[0.13214733,0.00037496036,0.00021066098,0.00021706038,0.0010018302,0.000051089115,0.00022447223,0.010349207,0.00076348],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000604663,0.000009576384,0.015520936,0.00009563335,0.00008771399,0.0005681605,0.00009669739,4.8138993e-8,0.0000019719416,0.000021683809,0.8245153,0.15902182],"study_design_scores_gemma":[0.003651577,0.0009895229,0.022479938,0.034420364,0.0009784104,0.0007492896,0.000062605235,0.00014415634,8.7828863e-7,0.0028027522,0.9333302,0.00039030163],"about_ca_topic_score_codex":0.000051962154,"about_ca_topic_score_gemma":5.333063e-7,"teacher_disagreement_score":0.15863152,"about_ca_system_score_codex":0.00006214904,"about_ca_system_score_gemma":0.00019013758,"threshold_uncertainty_score":0.99987024},"labels":[],"label_agreement":null},{"id":"W2321328495","doi":"10.1097/acm.0b013e31828ffb05","title":"In Reply to Petrie and Campbell","year":2013,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University; Hamilton Health Sciences","funders":"","keywords":"Assertion; Fallacy; Aside; Psychology; Computer science; Epistemology; Cognitive psychology; Philosophy; Linguistics","score_opus":0.03958801493273005,"score_gpt":0.3632516371782095,"score_spread":0.32366362224547945,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2321328495","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.021391079,0.003750959,0.000019612724,0.9680704,0.0009322924,0.0009490144,0.0000070801966,0.00006001856,0.004819551],"genre_scores_gemma":[0.0026603013,0.0032261682,0.00009958545,0.929878,0.01210777,0.00008762249,0.00015577853,0.00006606687,0.051718723],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99685687,0.00008552931,0.0009662899,0.0008006979,0.00068338413,0.0006072327],"domain_scores_gemma":[0.9858195,0.012749095,0.00020193002,0.0006037924,0.000104022314,0.00052166387],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.00085651746,0.00039448406,0.0012783658,0.0004761189,0.000025613423,0.000005350337,0.00021280492,0.0015020377,0.0013694694],"category_scores_gemma":[0.095128335,0.00028207048,0.000057153724,0.00041500616,0.00026494308,0.00004186589,0.00012401615,0.005905506,0.00060366595],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000033452652,0.0000054677357,0.003691348,0.00012711663,0.000026501382,0.0010990521,0.00052392343,8.9165795e-8,0.000033170763,0.000019653904,0.99245936,0.0019808435],"study_design_scores_gemma":[0.0021780003,0.00058032555,0.003725978,0.010704882,0.00020314517,0.0003694858,0.000086333675,0.000008939265,0.00001807298,0.0007921908,0.98106706,0.00026556314],"about_ca_topic_score_codex":0.0013100673,"about_ca_topic_score_gemma":0.000002526818,"teacher_disagreement_score":0.09427182,"about_ca_system_score_codex":0.00010776629,"about_ca_system_score_gemma":0.00015407546,"threshold_uncertainty_score":0.99996316},"labels":[],"label_agreement":null},{"id":"W2321394128","doi":"10.1097/ta.0000000000000447","title":"Re","year":2014,"lang":"en","type":"article","venue":"The Journal of Trauma: Injury, Infection, and Critical Care","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"London Health Sciences Centre","funders":"","keywords":"Library science; World Wide Web; Web site; Computer science; The Internet","score_opus":0.022147344822842148,"score_gpt":0.3616504584180069,"score_spread":0.3395031135951648,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2321394128","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9862227,0.0009432712,0.0009933866,0.005604216,0.00062631076,0.000067435525,0.0000038105466,0.000020918558,0.0055179875],"genre_scores_gemma":[0.99742603,0.0003047397,0.00010476721,0.0011969266,0.00090973557,7.9931783e-7,0.0000013124265,0.000010550376,0.00004516405],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9990185,0.00013930353,0.0003558049,0.0000811664,0.00025051864,0.00015472445],"domain_scores_gemma":[0.9941983,0.0048012766,0.00006819807,0.00015280281,0.00053338934,0.00024602065],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0007290736,0.000096153824,0.00026579504,0.000052911495,0.0001493135,0.000025700105,0.00005191429,0.00009260812,0.00012075376],"category_scores_gemma":[0.02335622,0.00005625457,0.00013009422,0.00009038026,0.00022619974,0.000055262128,0.000016212905,0.00040009635,0.000009844594],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0031836338,0.0013420889,0.49222603,0.000830118,0.00030829408,0.000083628,0.006618062,0.000025150128,0.00052083255,0.018009717,0.07983466,0.39701778],"study_design_scores_gemma":[0.004428389,0.013726128,0.91155744,0.0028886418,0.0023991633,0.002013415,0.0045894557,0.00008857649,0.0024838406,0.015501897,0.039920088,0.00040295543],"about_ca_topic_score_codex":0.00005605543,"about_ca_topic_score_gemma":0.0000229001,"teacher_disagreement_score":0.4193314,"about_ca_system_score_codex":0.000026362146,"about_ca_system_score_gemma":0.000044019118,"threshold_uncertainty_score":0.9848705},"labels":[],"label_agreement":null},{"id":"W2321735259","doi":"10.1097/acm.0000000000000415","title":"Reflecting Upon Reflection in Diagnostic Reasoning","year":2014,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of British Columbia","funders":"","keywords":"Surprise; Confirmation bias; Context (archaeology); Assertion; Psychology; Certainty; Cognition; Affect (linguistics); Cognitive psychology; Epistemology; Cognitive bias; Generalizability theory; Ecological validity; Social psychology; Computer science; Psychiatry; Developmental psychology","score_opus":0.0837015185357991,"score_gpt":0.42038679158953446,"score_spread":0.33668527305373536,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2321735259","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.018082805,0.004112736,0.00042131334,0.95208603,0.0023899265,0.0009380843,0.0000031194925,0.0003533869,0.021612579],"genre_scores_gemma":[0.07540653,0.0044997525,0.00026380963,0.8777369,0.033382125,0.00014186358,0.00078336254,0.00019448815,0.0075911195],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99439114,0.0003998443,0.0018716541,0.0011580872,0.0011412866,0.0010379999],"domain_scores_gemma":[0.9361736,0.061825264,0.00073551433,0.0007713021,0.00016680262,0.00032750054],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0029312402,0.0006493048,0.0018298007,0.0007484553,0.00009942393,0.0000105434765,0.0003241092,0.0033430555,0.00042929576],"category_scores_gemma":[0.4388643,0.00052367774,0.00017218932,0.0008511973,0.00022968881,0.000061126935,0.000090707246,0.012933918,0.00025948024],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008242683,0.000012207136,0.010463448,0.00026798045,0.0000574679,0.0014830832,0.0005265688,0.0000057907946,0.00010125855,0.000053452,0.98136705,0.0055792527],"study_design_scores_gemma":[0.0043951506,0.0011651994,0.0030372634,0.07116442,0.0008164924,0.0014658622,0.00014436086,0.00034940123,0.000081974744,0.0017632948,0.9149491,0.00066747604],"about_ca_topic_score_codex":0.00064125686,"about_ca_topic_score_gemma":0.000019647772,"teacher_disagreement_score":0.43593305,"about_ca_system_score_codex":0.0004894694,"about_ca_system_score_gemma":0.00026392864,"threshold_uncertainty_score":0.99972147},"labels":[],"label_agreement":null},{"id":"W2321863106","doi":"10.1097/acm.0000000000000475","title":"Integrating Competencies","year":2014,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Holland Bloorview Kids Rehabilitation Hospital","funders":"","keywords":"Context (archaeology); Audit; Thematic analysis; Grounded theory; Set (abstract data type); Process (computing); Psychology; Medical education; Qualitative research; Medicine; Computer science; Sociology","score_opus":0.04142868484158728,"score_gpt":0.36574634148291946,"score_spread":0.32431765664133216,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2321863106","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.824119,0.0016489342,0.009097487,0.038669117,0.0009792633,0.0003040005,0.000001248128,0.00032931723,0.12485167],"genre_scores_gemma":[0.9824467,0.00031781392,0.0009377715,0.011377994,0.0017043387,0.000009246182,0.000025664125,0.00001696325,0.003163517],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.99863106,0.000054669625,0.0004552107,0.0002442157,0.0003555972,0.00025923838],"domain_scores_gemma":[0.9899645,0.0092563955,0.00010445869,0.00026474588,0.000090186695,0.0003197543],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010178486,0.00013890701,0.00044923354,0.00007771275,0.00005308241,0.0000022485413,0.00011863693,0.00017903691,0.00048455966],"category_scores_gemma":[0.12699804,0.00008780103,0.000049136055,0.00018798988,0.00027954043,0.000026966107,0.00003727528,0.00076487364,0.00017710612],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00023435369,0.00006529136,0.33007383,0.00016090892,0.00013918048,0.00006946384,0.006249894,0.0000057377656,0.01735624,0.13938256,0.2645203,0.24174225],"study_design_scores_gemma":[0.020401888,0.006455838,0.26609123,0.04222623,0.0013285532,0.0013878155,0.0084057655,0.012378398,0.003845201,0.044375677,0.59191567,0.0011877168],"about_ca_topic_score_codex":0.00007942682,"about_ca_topic_score_gemma":0.0000018906557,"teacher_disagreement_score":0.3273954,"about_ca_system_score_codex":0.000026239168,"about_ca_system_score_gemma":0.000045138844,"threshold_uncertainty_score":0.8803556},"labels":[],"label_agreement":null},{"id":"W2322101209","doi":"10.1177/0008417413501466","title":"Book Review: Clinical reasoning in occupational therapy: Controversies in practice","year":2013,"lang":"en","type":"article","venue":"Canadian Journal of Occupational Therapy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Occupational therapy; Clinical Practice; Psychology; Psychotherapist; Medicine; Physical therapy","score_opus":0.11500940107380161,"score_gpt":0.45356679807831757,"score_spread":0.3385573970045159,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2322101209","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5124609,0.43561724,0.00006640844,0.047645498,0.0008815623,0.001206909,0.000030435654,0.000009633707,0.0020814438],"genre_scores_gemma":[0.499805,0.17261657,0.003068593,0.32253143,0.0011213949,0.000052477204,0.00007106674,0.0000469457,0.00068650965],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9974672,0.00032123746,0.0012562564,0.00020115796,0.00043747015,0.00031666554],"domain_scores_gemma":[0.9829449,0.01431676,0.00066313695,0.0002318131,0.0010540532,0.00078933407],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0019000007,0.00016614687,0.0005982767,0.00034744397,0.000051167517,0.000041229363,0.00016304081,0.00014835772,0.0027592059],"category_scores_gemma":[0.05700188,0.00013550904,0.0002286989,0.0003301708,0.00013985779,0.00034308305,0.00000546197,0.0006121631,0.00006317581],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011535557,0.00023483898,0.6813046,0.000008976433,0.00010004365,0.00021984831,0.000103621,0.000025286205,0.000005455433,0.0001937814,0.2972007,0.019449294],"study_design_scores_gemma":[0.0044485573,0.00045516164,0.7089386,0.002684498,0.000018228351,0.0001544165,0.000051921295,0.000046417892,0.000004212171,0.00008164554,0.28300944,0.0001068886],"about_ca_topic_score_codex":0.004275757,"about_ca_topic_score_gemma":0.00091844966,"teacher_disagreement_score":0.27488592,"about_ca_system_score_codex":0.00025169918,"about_ca_system_score_gemma":0.0042847698,"threshold_uncertainty_score":0.99815243},"labels":[],"label_agreement":null},{"id":"W2323428358","doi":"10.15766/mep_2374-8265.8191","title":"PedsCases - A Learning Module of Pediatric Emergency Medicine Cases for Medical Students","year":2010,"lang":"en","type":"article","venue":"MedEdPORTAL","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Pediatric emergency medicine; Medical education; Medicine; Emergency department; Psychology; Emergency physician","score_opus":0.0361851804766008,"score_gpt":0.4114613332064659,"score_spread":0.3752761527298651,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2323428358","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9949261,0.00038996572,0.0001193463,0.0007887298,0.001671161,0.00037167757,0.000013718849,0.00007501902,0.0016442722],"genre_scores_gemma":[0.99331075,0.0003161571,0.000303838,0.00010218958,0.0023882492,0.000054822074,0.00010799041,0.000032080927,0.0033839492],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99715215,0.000029450703,0.0008633825,0.0003343551,0.0012996598,0.00032100515],"domain_scores_gemma":[0.99361855,0.0049977726,0.00027934826,0.00029331303,0.00026696213,0.00054407376],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011858363,0.00018351935,0.0005666311,0.00013638403,0.000079107405,0.000004200393,0.00023183852,0.00020958284,0.0081605455],"category_scores_gemma":[0.16496328,0.0001369476,0.0002065097,0.00024693005,0.00015648874,0.000034408775,0.000079271485,0.00047183325,0.000028551578],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015317311,0.000728339,0.97187483,0.00009961013,0.000071075934,0.00024080157,0.00013907303,0.0000013419088,0.00024614023,0.00016972015,0.023843704,0.002432213],"study_design_scores_gemma":[0.009400554,0.0049899872,0.9708138,0.00065428944,0.00166925,0.0008489124,0.0005609338,0.00063958566,0.0007646338,0.00042610316,0.008726093,0.00050582684],"about_ca_topic_score_codex":0.00020589263,"about_ca_topic_score_gemma":0.00013170339,"teacher_disagreement_score":0.16377744,"about_ca_system_score_codex":0.0000079704905,"about_ca_system_score_gemma":0.00020041832,"threshold_uncertainty_score":0.9927461},"labels":[],"label_agreement":null},{"id":"W2323745045","doi":"10.1097/acm.0000000000000550","title":"Blink or Think","year":2014,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":29,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"HEC Montréal","funders":"","keywords":"Medical diagnosis; Certification; Psychology; Focus (optics); MEDLINE; Cognition; Medicine; Psychiatry; Pathology","score_opus":0.06204253602701446,"score_gpt":0.3997168727205619,"score_spread":0.3376743366935474,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2323745045","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7177768,0.0025864495,0.006159647,0.14874937,0.002374559,0.000798696,0.0000033880308,0.00071796175,0.12083317],"genre_scores_gemma":[0.9401231,0.0012603276,0.00043173798,0.031095877,0.0038271109,0.000011763476,0.000038247967,0.0000310268,0.02318083],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99823475,0.000064774176,0.0005438603,0.00033560823,0.00048276043,0.00033822123],"domain_scores_gemma":[0.98700356,0.011836647,0.00011882383,0.0004278662,0.000089047375,0.00052407937],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0013086665,0.0001744876,0.000553827,0.00010092821,0.000052299107,0.000002511387,0.00016667637,0.0003378625,0.002019886],"category_scores_gemma":[0.15592632,0.000097908196,0.0000572458,0.0002782168,0.00026649286,0.000033237873,0.000047111746,0.0009339479,0.00053183903],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009667508,0.00007516413,0.046854444,0.00012882588,0.00013157369,0.00016387741,0.0031568154,0.0000034972288,0.0015984222,0.017423345,0.76768386,0.16181341],"study_design_scores_gemma":[0.025822686,0.0069292756,0.09210842,0.01916888,0.0014013023,0.0014522668,0.0014964028,0.003509554,0.0013504078,0.022920188,0.8230072,0.00083338976],"about_ca_topic_score_codex":0.00004475916,"about_ca_topic_score_gemma":0.0000015648275,"teacher_disagreement_score":0.22234632,"about_ca_system_score_codex":0.000029005338,"about_ca_system_score_gemma":0.00009547952,"threshold_uncertainty_score":0.9988924},"labels":[],"label_agreement":null},{"id":"W2325153051","doi":"10.1017/s0317167100010799","title":"Sources of Error in Neuropathology Intraoperative Diagnosis","year":2010,"lang":"en","type":"article","venue":"Canadian Journal of Neurological Sciences / Journal Canadien des Sciences Neurologiques","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; University of Toronto; London Health Sciences Centre; Sunnybrook Health Science Centre; Western University","funders":"","keywords":"Medical diagnosis; Neuropathology; Audit; Gold standard (test); Sampling (signal processing); Medicine; Sampling error; Set (abstract data type); Medical physics; Radiology; Statistics; Observational error; Pathology; Computer science; Disease; Accounting","score_opus":0.04626030648891541,"score_gpt":0.3309572529299006,"score_spread":0.2846969464409852,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2325153051","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98907197,0.00057084335,0.0000044694107,0.0070688715,0.0010599503,0.000149206,0.000008117593,0.000009146909,0.0020574385],"genre_scores_gemma":[0.9944042,0.0002950459,0.0018712298,0.003151451,0.000245577,0.0000037012805,1.7995092e-7,0.000011238114,0.00001738496],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9959629,0.0007307745,0.0011995898,0.0004757261,0.00060606847,0.0010249334],"domain_scores_gemma":[0.99218434,0.0044899816,0.00077870645,0.00017154067,0.0005263762,0.001849036],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.0048259352,0.00026731682,0.00076989323,0.0010899489,0.00060622056,0.00017828667,0.0011399678,0.00026104006,0.00032323433],"category_scores_gemma":[0.05922769,0.0001726926,0.00026265447,0.0013291881,0.009576787,0.00039499588,0.000057758065,0.0021784238,0.0000042623437],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009566141,0.00009100214,0.9853873,0.000004621327,0.00000663979,0.009407724,0.0003796701,0.00041109975,0.00012834788,0.00062298425,0.0002308648,0.003234067],"study_design_scores_gemma":[0.0006619718,0.06587216,0.88315845,0.00017643158,0.000052729807,0.039986826,0.00026196832,0.00034710046,0.0005707651,0.007938494,0.00075208023,0.00022101098],"about_ca_topic_score_codex":0.0015792277,"about_ca_topic_score_gemma":0.07150793,"teacher_disagreement_score":0.102228865,"about_ca_system_score_codex":0.00006463182,"about_ca_system_score_gemma":0.0030575113,"threshold_uncertainty_score":0.9931186},"labels":[],"label_agreement":null},{"id":"W2326997006","doi":"10.1097/01.eem.0000432257.88168.ce","title":"Special Report","year":2013,"lang":"en","type":"article","venue":"Emergency Medicine News","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Blame; Psychology; Emergency department; Notice; Medical emergency; Medicine; Medical education; Psychiatry; Political science; Law","score_opus":0.04768831426705673,"score_gpt":0.3741037172722693,"score_spread":0.3264154030052126,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2326997006","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.28434166,0.00043414312,0.00013077224,0.03026586,0.011144615,0.0005547736,9.620203e-7,0.00013782317,0.67298937],"genre_scores_gemma":[0.33731714,0.003235957,0.0017442695,0.0057295463,0.1746308,0.00025156516,0.0003190079,0.00013484679,0.47663686],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9979255,0.000028917959,0.0008790795,0.00036439637,0.00047876438,0.00032332854],"domain_scores_gemma":[0.9977494,0.00075326377,0.00016719234,0.00057665847,0.00023540364,0.00051811035],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00029030323,0.00017706932,0.000459515,0.00008430811,0.00005260893,0.0000024899532,0.000094720606,0.00010757788,0.13850082],"category_scores_gemma":[0.046992023,0.000120701145,0.00014144291,0.0002945251,0.000106919964,0.000054540742,0.000033830824,0.00024631975,0.0025453104],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000016177482,0.00010950283,0.08671355,0.000009918821,0.000027008435,0.00015492317,0.00012231014,2.4059366e-7,0.00020473739,0.00015132068,0.8960389,0.016451415],"study_design_scores_gemma":[0.001608538,0.00061238697,0.33115968,0.00054753304,0.00015483373,0.0003700109,0.00025973454,0.000027445636,0.000056797624,0.0015407458,0.66347635,0.00018596812],"about_ca_topic_score_codex":0.001583286,"about_ca_topic_score_gemma":0.00003972065,"teacher_disagreement_score":0.24444613,"about_ca_system_score_codex":0.000028510081,"about_ca_system_score_gemma":0.00008156779,"threshold_uncertainty_score":0.99823135},"labels":[],"label_agreement":null},{"id":"W2328803679","doi":"10.3928/01484834-20110331-03","title":"Use of a Script Concordance Test to Assess Development of Clinical Reasoning in Nursing Students","year":2011,"lang":"en","type":"article","venue":"Journal of Nursing Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":57,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Collège de Maisonneuve","funders":"","keywords":"Concordance; Cronbach's alpha; Test (biology); Reliability (semiconductor); Bachelor; Psychology; Educational measurement; Medicine; Psychometrics; Clinical psychology; Medical education; Nursing; Curriculum; Pedagogy","score_opus":0.2619656988717902,"score_gpt":0.5232820715785464,"score_spread":0.26131637270675623,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2328803679","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9961601,0.00015581769,0.0008748524,0.00029861505,0.001673163,0.00015773081,6.5394113e-7,0.0000037726895,0.0006753159],"genre_scores_gemma":[0.89165115,0.000027573953,0.10794211,0.00012091659,0.00016855632,0.0000018360893,0.0000016457136,0.000011103539,0.00007511673],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9976193,0.00009130787,0.0015757263,0.00013340893,0.00043683132,0.00014348392],"domain_scores_gemma":[0.99516827,0.0027644548,0.0010025769,0.00016967292,0.0006558107,0.00023922922],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0014307766,0.00009091475,0.000531836,0.00023964848,0.000021223976,0.000011373161,0.00012380275,0.00008324562,0.000013638443],"category_scores_gemma":[0.046767835,0.00008020945,0.00009633787,0.0002791189,0.000091422626,0.000120765,0.000009709304,0.0002693293,0.0000019371507],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000397879,0.0095075695,0.7250189,0.000018009914,0.00002310702,0.0000049143014,0.007765967,0.0000034714315,0.0006660112,0.00003596488,0.001403695,0.25515452],"study_design_scores_gemma":[0.00083542557,0.0006996086,0.9430885,0.050032604,0.00010445885,0.000070616545,0.0020933033,0.00001757367,0.002795509,0.00012307042,0.00007386534,0.00006551444],"about_ca_topic_score_codex":0.000024422803,"about_ca_topic_score_gemma":0.0000029907644,"teacher_disagreement_score":0.25508898,"about_ca_system_score_codex":0.00023104055,"about_ca_system_score_gemma":0.0015462429,"threshold_uncertainty_score":0.96126163},"labels":[],"label_agreement":null},{"id":"W2330528987","doi":"10.1097/acm.0000000000000314","title":"In Reply to Mamede and Schmidt","year":2014,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Deductive reasoning; Inductive reasoning; Epistemology; Reading (process); Psychology; Limiting; Cognitive psychology; Raising (metalworking); Computer science; Cognitive science; Artificial intelligence; Linguistics; Philosophy; Mathematics","score_opus":0.045442038948420714,"score_gpt":0.38145452624515924,"score_spread":0.33601248729673855,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2330528987","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008695541,0.0019227671,0.00006139298,0.98038906,0.00057225925,0.00059854635,0.00000341511,0.000068033245,0.0076890127],"genre_scores_gemma":[0.00576566,0.0017160062,0.00015356988,0.95028746,0.015620016,0.00005749906,0.00013983557,0.00006771953,0.02619221],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9967924,0.000113196365,0.0009673483,0.0008284009,0.0007009351,0.0005977328],"domain_scores_gemma":[0.9864964,0.012092323,0.00019949638,0.00064431917,0.00007515197,0.00049229455],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0013149168,0.00039216457,0.0013709128,0.00045146822,0.000026355621,0.000004672719,0.00022508198,0.0015432814,0.0006296507],"category_scores_gemma":[0.099132635,0.00028569996,0.000060705577,0.00033209872,0.00029730418,0.000028141587,0.000111053254,0.0058923694,0.00038359736],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000057919337,0.0000043018545,0.004179625,0.00014778222,0.0000259861,0.0010670068,0.00031169425,1.5373989e-7,0.000039717324,0.000057132052,0.99205786,0.0020508189],"study_design_scores_gemma":[0.0022764818,0.0006093083,0.0027877982,0.0163218,0.00022262803,0.00043947087,0.000059716167,0.00001312778,0.000020222431,0.0009238836,0.97606295,0.00026262854],"about_ca_topic_score_codex":0.00041936527,"about_ca_topic_score_gemma":0.0000017813123,"teacher_disagreement_score":0.09781772,"about_ca_system_score_codex":0.00009230668,"about_ca_system_score_gemma":0.000120004195,"threshold_uncertainty_score":0.9999595},"labels":[],"label_agreement":null},{"id":"W2332467280","doi":"10.1097/acm.0000000000000510","title":"Safety in Numbers","year":2014,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Calgary","funders":"","keywords":"Polling; Anonymity; Session (web analytics); Psychology; Medical education; Gatekeeping; Medicine; Social psychology; Public relations; Law","score_opus":0.04361032693270577,"score_gpt":0.3699816108433881,"score_spread":0.3263712839106823,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2332467280","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0014231632,0.0015826306,0.000105496256,0.94809836,0.0011795388,0.0004886949,0.0000064340784,0.00011682224,0.04699885],"genre_scores_gemma":[0.006887245,0.0031284357,0.000056706693,0.94508344,0.020033963,0.00003255392,0.00093127,0.00008878671,0.0237576],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.996141,0.00017191538,0.0013145201,0.00074043765,0.0009284116,0.00070370256],"domain_scores_gemma":[0.97999257,0.018591907,0.0003254418,0.0007007552,0.00008143213,0.0003079002],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0015025,0.00045243616,0.0016626599,0.00033005112,0.000030287309,0.0000026131247,0.00032300092,0.0029878228,0.0016621572],"category_scores_gemma":[0.07405457,0.0003302643,0.00014284899,0.00037273852,0.0004328451,0.00002363053,0.00006589687,0.009957216,0.0006084214],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007493434,0.000006201571,0.00750645,0.00014305615,0.000045102133,0.0011245274,0.00026371557,5.984757e-7,0.00000618644,0.000068353234,0.98807776,0.0026831005],"study_design_scores_gemma":[0.0040298947,0.0002968536,0.0027498845,0.018806612,0.00033491486,0.0004035825,0.000052721894,0.000035328652,0.0000044686935,0.0009714627,0.97202224,0.00029202123],"about_ca_topic_score_codex":0.0004134608,"about_ca_topic_score_gemma":0.0000046031355,"teacher_disagreement_score":0.07255207,"about_ca_system_score_codex":0.00020535484,"about_ca_system_score_gemma":0.00021958745,"threshold_uncertainty_score":0.99991494},"labels":[],"label_agreement":null},{"id":"W2333886831","doi":"10.4322/acr.2016.023","title":"Patient cueing, a type of diagnostic error","year":2016,"lang":"en","type":"article","venue":"Autopsy and Case Reports","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Dalhousie University","keywords":"Medicine; Emergency physician; Intensive care medicine; Variety (cybernetics); Diagnostic test; Emergency department; Medical emergency; Pediatrics; Psychiatry; Artificial intelligence","score_opus":0.025720477747552827,"score_gpt":0.328367474021765,"score_spread":0.3026469962742122,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2333886831","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9965318,0.00094730983,0.000031385298,0.00039780734,0.0002520502,0.00014182967,0.000003175748,0.000032172655,0.0016624869],"genre_scores_gemma":[0.99887764,0.00019136892,0.00015424362,0.00017839165,0.000046725087,0.00000772567,0.000002997708,0.000008929982,0.0005319675],"study_design_codex":"case_report","study_design_gemma":"case_report","domain_scores_codex":[0.9991631,0.000017605804,0.00035936423,0.00019917995,0.00013318691,0.000127562],"domain_scores_gemma":[0.9943867,0.004937732,0.00015297518,0.00026297115,0.00009137395,0.00016828916],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00016859735,0.00007981793,0.00022169681,0.00002903289,0.000026480056,0.0000034273858,0.0000067539,0.00006625445,0.00017427636],"category_scores_gemma":[0.051358566,0.000046847304,0.00005440888,0.00007040674,0.00008264145,0.000022865317,0.000027120826,0.00004865725,0.000008497084],"study_design_candidate":"case_report","study_design_consensus":"case_report","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012898078,0.00059375993,0.28355256,0.00007712752,0.00014066244,0.36750424,0.0004523731,8.4211723e-7,0.0011808111,0.00049520796,0.004033014,0.34184045],"study_design_scores_gemma":[0.0030830512,0.0042857598,0.30476123,0.010172003,0.0008780175,0.62810606,0.00028337695,0.000038746177,0.009371891,0.0031892925,0.035265777,0.00056481926],"about_ca_topic_score_codex":0.00009539807,"about_ca_topic_score_gemma":0.0000051570437,"teacher_disagreement_score":0.3412756,"about_ca_system_score_codex":0.000010231485,"about_ca_system_score_gemma":0.0000776953,"threshold_uncertainty_score":0.95663226},"labels":[],"label_agreement":null},{"id":"W2334143500","doi":"10.1097/acm.0b013e31826d5962","title":"Fluency—Not Competency or Expertise—Is Needed to Incorporate Evidence Into Practice","year":2012,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"","keywords":"Fluency; Medical education; MEDLINE; Psychology; Medicine; Mathematics education; Political science","score_opus":0.14602729354125904,"score_gpt":0.4644728442389273,"score_spread":0.3184455506976683,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2334143500","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6382442,0.010288025,0.0011182174,0.3448217,0.0017062114,0.0009887988,0.0000033375181,0.00023835356,0.0025911457],"genre_scores_gemma":[0.8205241,0.0032915126,0.003449662,0.16644509,0.0027911963,0.00008593028,0.000013388395,0.00004470597,0.0033544195],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9963729,0.00019168208,0.0010963382,0.00053505914,0.001058646,0.00074536767],"domain_scores_gemma":[0.97556996,0.021232037,0.00035732804,0.0007054417,0.00037299684,0.0017622609],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.002546384,0.00035038573,0.0008457482,0.00020752412,0.00011620084,0.000007691694,0.00034306213,0.0004042903,0.0015158958],"category_scores_gemma":[0.32914606,0.00022665848,0.00007680627,0.0008498771,0.00028530674,0.00044168998,0.00017712958,0.0011453374,0.0012767446],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0044654254,0.0002562433,0.050123855,0.00020771423,0.0002054597,0.00027269442,0.07489953,0.0000017627258,0.027501768,0.0013765148,0.7899398,0.050749186],"study_design_scores_gemma":[0.022570193,0.015171424,0.105217405,0.106235944,0.005078091,0.0045194943,0.05102092,0.00074873655,0.045378312,0.0031891419,0.63765186,0.0032184739],"about_ca_topic_score_codex":0.0011011079,"about_ca_topic_score_gemma":0.0000042067,"teacher_disagreement_score":0.32659966,"about_ca_system_score_codex":0.00021543639,"about_ca_system_score_gemma":0.00041881457,"threshold_uncertainty_score":0.9995009},"labels":[],"label_agreement":null},{"id":"W2336120766","doi":"10.1016/j.annemergmed.2015.10.016","title":"Why I Chose Emergency Medicine","year":2016,"lang":"en","type":"article","venue":"Annals of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Victoria Hospital","funders":"","keywords":"Brother; Internship; Medicine; Anecdote; Wonder; Medical education; Law; Psychology","score_opus":0.17868612712977858,"score_gpt":0.45650075525528117,"score_spread":0.2778146281255026,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2336120766","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.39058444,0.021485658,0.0013638963,0.44897744,0.013762918,0.0009187745,0.00004614559,0.0002832123,0.12257752],"genre_scores_gemma":[0.84196013,0.10167274,0.000105205945,0.005453939,0.0060635987,0.000056650933,0.000075970245,0.00010800999,0.04450377],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9955285,0.00010199479,0.002104322,0.00060230185,0.0010168594,0.00064601563],"domain_scores_gemma":[0.9940809,0.0023833252,0.00058494846,0.000998871,0.001131135,0.00082084426],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0013802336,0.00039558465,0.0011615976,0.00030494411,0.00007176398,4.31619e-7,0.0003031434,0.00020757025,0.090249196],"category_scores_gemma":[0.09330697,0.00020970525,0.00035205673,0.0008013,0.0004494764,0.00010503756,0.00007587897,0.0002165909,0.00019588337],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00028268003,0.00037063245,0.047043767,0.00010580021,0.00020304743,0.00002055565,0.00029391164,2.432142e-7,0.011157066,0.0013084526,0.92781115,0.011402698],"study_design_scores_gemma":[0.008195239,0.009351102,0.22870634,0.0163016,0.0008836688,0.000035843885,0.0003909178,0.000021383325,0.006008202,0.014462264,0.7148869,0.0007565595],"about_ca_topic_score_codex":0.0006462607,"about_ca_topic_score_gemma":0.000037442675,"teacher_disagreement_score":0.45137566,"about_ca_system_score_codex":0.000015915004,"about_ca_system_score_gemma":0.000075746895,"threshold_uncertainty_score":0.9143305},"labels":[],"label_agreement":null},{"id":"W2339051350","doi":"10.1002/j.0022-0337.2005.69.11.tb04019.x","title":"Knowledge Integration and Reasoning as a Function of Instruction in a Hybrid Medical Curriculum","year":2005,"lang":"en","type":"article","venue":"Journal of Dental Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":22,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; McGill University Health Centre; University of Waterloo","funders":"Social Sciences and Humanities Research Council of Canada; Columbia University","keywords":"Curriculum; Computer science; Cognition; Function (biology); Problem-based learning; Mathematics education; Psychology; Pedagogy","score_opus":0.009568397545983775,"score_gpt":0.34932689142954704,"score_spread":0.3397584938835633,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2339051350","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99594855,0.001085459,0.00017347813,0.0006185761,0.00080645666,0.000078906945,3.1876985e-7,0.000003993895,0.0012842847],"genre_scores_gemma":[0.9979823,0.00031070257,0.00084547943,0.000119054224,0.00060313405,0.0000025766537,0.000007702391,0.0000054046122,0.00012361567],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99904305,0.00004343208,0.00049785885,0.00008117249,0.0002673801,0.00006712706],"domain_scores_gemma":[0.99900156,0.0003104037,0.00028589767,0.00005359294,0.0001995383,0.00014903142],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00047874922,0.00006151255,0.00019189523,0.00021948243,0.000016399732,0.000009302009,0.000030141895,0.0000661922,0.0001408253],"category_scores_gemma":[0.012596696,0.000049660248,0.000058477115,0.00016203299,0.000041874424,0.00016895834,0.00001039163,0.00024924602,0.000010784463],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002635164,0.0015703897,0.13814034,0.000025469435,0.000021252825,0.000005846125,0.00051486917,0.0000031849752,0.0005172949,0.0005344081,0.0014260034,0.8569774],"study_design_scores_gemma":[0.0055972068,0.0018387937,0.96013194,0.01129025,0.00026160336,0.006219028,0.006592861,0.002458732,0.0029690142,0.00065213745,0.0018384461,0.00014997942],"about_ca_topic_score_codex":0.00008136971,"about_ca_topic_score_gemma":0.000053655796,"teacher_disagreement_score":0.85682744,"about_ca_system_score_codex":0.0001366306,"about_ca_system_score_gemma":0.0005632149,"threshold_uncertainty_score":0.9957206},"labels":[],"label_agreement":null},{"id":"W2339830158","doi":"10.19144/1911-1606.11.1.5","title":"Developing, Maintaining, and Teaching Clinical Diagnostic Expertise","year":2016,"lang":"en","type":"article","venue":"Canadian Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"University of Alberta; M.S.I. Foundation","keywords":"Scripting language; Process (computing); Identification (biology); Clinical Practice; Key (lock); Abstraction; Computer science; Medicine; Artificial intelligence; Knowledge management; Nursing","score_opus":0.05445950420105238,"score_gpt":0.38618587382271824,"score_spread":0.33172636962166585,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2339830158","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93237895,0.0026343702,0.0072024744,0.052388612,0.0021548378,0.00010973206,0.0000046911196,0.000011267809,0.0031150626],"genre_scores_gemma":[0.9834265,0.0008101009,0.003105472,0.0064995647,0.0035133802,0.0000021715139,0.0000016876292,0.000027110078,0.0026139813],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9977621,0.00015970152,0.0011603974,0.00023017802,0.0002835428,0.0004041228],"domain_scores_gemma":[0.9887133,0.008059404,0.0003843587,0.00020205148,0.00029601564,0.0023448383],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0016776418,0.0001947263,0.0006847933,0.00029338856,0.00008364619,0.000021783899,0.00018573426,0.00013848163,0.00046705524],"category_scores_gemma":[0.15067278,0.00010681337,0.00014617175,0.00005072695,0.0005588212,0.000091306974,0.00002705631,0.0005330286,0.000018699971],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00017382059,0.000034269437,0.6072344,0.000010761357,0.00019802492,0.0040653297,0.0004201137,5.412155e-7,0.000274127,0.0055373814,0.07095184,0.31109938],"study_design_scores_gemma":[0.013821399,0.0076164585,0.76602167,0.040932693,0.0005517248,0.01388996,0.00045056109,0.00007676903,0.00040712956,0.0042993263,0.1514088,0.00052349054],"about_ca_topic_score_codex":0.0048899963,"about_ca_topic_score_gemma":0.0017997889,"teacher_disagreement_score":0.3105759,"about_ca_system_score_codex":0.00025606298,"about_ca_system_score_gemma":0.0010853164,"threshold_uncertainty_score":0.85648143},"labels":[],"label_agreement":null},{"id":"W2346011204","doi":"10.4236/ce.2016.76084","title":"Development of a Method to Measure Clinical Reasoning in Pediatric Residents: The Pediatric Script Concordance Test","year":2016,"lang":"en","type":"article","venue":"Creative Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Calgary","funders":"Royal College of Physicians and Surgeons of Canada","keywords":"Concordance; Cronbach's alpha; Test (biology); Formative assessment; Reliability (semiconductor); Concordance correlation coefficient; Measure (data warehouse); Psychology; Medical education; Medical physics; Medicine; Clinical psychology; Computer science; Psychometrics; Mathematics education; Statistics; Mathematics; Data mining","score_opus":0.05894085427532954,"score_gpt":0.4255854136084464,"score_spread":0.36664455933311685,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2346011204","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9856279,0.0007684157,0.0069121495,0.0026841273,0.00048190547,0.00064569555,0.00000456766,0.000023929968,0.0028512948],"genre_scores_gemma":[0.9287153,0.0003402451,0.067497864,0.00053908204,0.0010278588,0.00016312522,0.0000068290624,0.000019618972,0.0016900997],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99795604,0.0002807001,0.0008387883,0.0003716208,0.0003387211,0.00021411378],"domain_scores_gemma":[0.97805214,0.020264419,0.0003825496,0.00041311677,0.00061575876,0.00027202815],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0026704632,0.00013952937,0.0003673069,0.00016915733,0.000061779734,0.000010395603,0.0001470239,0.00009917585,0.00006942373],"category_scores_gemma":[0.16781314,0.00008015049,0.00007855617,0.0006986506,0.000056676614,0.000059286944,0.000049184877,0.0001917372,0.00008481513],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000106748405,0.00059156824,0.8931864,0.00001958753,0.000010578791,0.0000011912824,0.0011094214,3.6401244e-7,0.00009111575,0.00014637904,0.005215678,0.09952097],"study_design_scores_gemma":[0.000970798,0.00019647463,0.9953432,0.0011833953,0.00012749799,0.0000051543657,0.0007506007,0.0000029005719,0.0004478198,0.00013857175,0.0007280972,0.00010548826],"about_ca_topic_score_codex":0.0001013609,"about_ca_topic_score_gemma":0.000035691293,"teacher_disagreement_score":0.16514269,"about_ca_system_score_codex":0.00017219558,"about_ca_system_score_gemma":0.0023103482,"threshold_uncertainty_score":0.83919674},"labels":[],"label_agreement":null},{"id":"W2365584262","doi":"","title":"Effects of teaching in actual scene on neuropsychology by consistency check","year":2015,"lang":"en","type":"article","venue":"Academic Journal of Chinese PLA Medical School","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Consistency (knowledge bases); Fluency; Medicine; Medical education; Psychology; Mathematics education; Computer science; Artificial intelligence","score_opus":0.01762115710903169,"score_gpt":0.3682525948284039,"score_spread":0.3506314377193722,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2365584262","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9844075,0.0024521116,0.00007151263,0.0103874635,0.0008036825,0.00016365007,0.0000024929675,0.000013294937,0.0016983319],"genre_scores_gemma":[0.9911866,0.0008129647,0.00019820656,0.00692647,0.00076382194,0.0000027945034,0.0000042287343,0.00002103774,0.00008383817],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99636936,0.00039717625,0.001325997,0.00022138286,0.001350257,0.00033583638],"domain_scores_gemma":[0.98144513,0.015610546,0.00054488704,0.0001769902,0.00012131785,0.0021011091],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0031666583,0.00021270773,0.0009329802,0.00020895238,0.000020464335,0.0000045648176,0.00034198846,0.00056747714,0.000095275376],"category_scores_gemma":[0.5431318,0.00013380394,0.00017507409,0.00022866785,0.00020278849,0.000084019026,0.00005656796,0.004608103,0.000029219551],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0044175414,0.0025227182,0.83487815,0.0001453231,0.00022892351,0.002973551,0.0009953834,0.000012786304,0.007907088,0.000308898,0.100816734,0.04479289],"study_design_scores_gemma":[0.050838936,0.009029377,0.9012331,0.019529995,0.0003989653,0.005810071,0.00027108926,0.0004153314,0.0012899158,0.005681119,0.0049335323,0.00056858035],"about_ca_topic_score_codex":0.000021395817,"about_ca_topic_score_gemma":0.0000011031198,"teacher_disagreement_score":0.53996515,"about_ca_system_score_codex":0.00008361406,"about_ca_system_score_gemma":0.0007353385,"threshold_uncertainty_score":0.9976883},"labels":[],"label_agreement":null},{"id":"W2388924974","doi":"10.1016/j.cjca.2015.07.220","title":"STRESS ELECTROCARDIOGRAPHIC INTERPRETATION SKILLS OF CARDIOLOGY TRAINEES: ARE THEY COMPETENT?","year":2015,"lang":"en","type":"article","venue":"Canadian Journal of Cardiology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Toronto Public Health","funders":"","keywords":"Medicine; Internal medicine; Cardiology; Interpretation (philosophy); Medical physics","score_opus":0.02029702906842485,"score_gpt":0.28471969539988284,"score_spread":0.264422666331458,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2388924974","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97233486,0.003412283,0.0076520066,0.0014025514,0.0018072498,0.00024347958,0.00010402404,0.000014374485,0.01302915],"genre_scores_gemma":[0.998023,0.00022531336,0.00019507752,0.00025901166,0.0012225581,0.0000034338711,0.000026142048,0.000019752573,0.000025708783],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99816734,0.00035051632,0.00069862534,0.00017909388,0.00020266045,0.0004017541],"domain_scores_gemma":[0.99569356,0.0017068076,0.00038322364,0.00027165635,0.00086272595,0.0010819951],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0008764073,0.00016937799,0.0013408313,0.00052766246,0.000030946958,0.000008659357,0.00019678935,0.0002570037,0.0000058375417],"category_scores_gemma":[0.015598273,0.00013814242,0.000640337,0.00020633759,0.00030864566,0.000043537522,0.000011195915,0.00045234605,0.0000060367756],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003440344,0.000011894868,0.9642194,0.00003208735,0.0014886438,0.0010255865,0.00067107024,0.009472188,0.000100804435,0.00024426344,0.014191911,0.008198121],"study_design_scores_gemma":[0.0072411103,0.005301953,0.94543135,0.002595693,0.0017258208,0.007875856,0.0033452886,0.00013127757,0.00023249113,0.0038825346,0.021781253,0.00045535053],"about_ca_topic_score_codex":0.00042793696,"about_ca_topic_score_gemma":0.0006675382,"teacher_disagreement_score":0.025688121,"about_ca_system_score_codex":0.00013356046,"about_ca_system_score_gemma":0.0015513455,"threshold_uncertainty_score":0.9926938},"labels":[],"label_agreement":null},{"id":"W2394742858","doi":"10.47102/annals-acadmedsg.v30n4p353","title":"Clinical Reasoning Learning Sessions","year":2001,"lang":"en","type":"article","venue":"Annals of the Academy of Medicine Singapore","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Jewish General Hospital","funders":"","keywords":"Medicine; Problem-based learning; Curriculum; Medical education; Pedagogy; Psychology","score_opus":0.1673624761501252,"score_gpt":0.48004008652025687,"score_spread":0.31267761037013164,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2394742858","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8603596,0.0019911868,0.000041749514,0.12656409,0.00027252044,0.00021543364,0.0000021932553,0.00004862799,0.010504582],"genre_scores_gemma":[0.9853708,0.0035102286,0.0004405389,0.007392495,0.0008314266,0.0000027629164,0.000007196275,0.000025262329,0.0024192845],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99715465,0.00028094574,0.0012359996,0.00030435977,0.0006907234,0.0003333],"domain_scores_gemma":[0.990235,0.008087274,0.0008121096,0.00033902773,0.00020141095,0.00032519922],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0030594198,0.00018252681,0.0008969348,0.000102877006,0.00011931453,0.0000023346697,0.0003273597,0.000303763,0.00026483228],"category_scores_gemma":[0.12816389,0.00010492383,0.0003949283,0.0005113035,0.0010920163,0.00004680382,0.00015572054,0.0012426357,0.00000924498],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004905969,0.00042557134,0.7964765,0.000080722486,0.00021882531,0.000022540133,0.00038933655,0.00006476895,0.0010278178,0.0016780213,0.085026704,0.11409861],"study_design_scores_gemma":[0.003679155,0.0012838334,0.88976455,0.024191605,0.00049628265,0.00023789574,0.00063829997,0.00042303104,0.0028045438,0.0031124451,0.07316604,0.0002023305],"about_ca_topic_score_codex":0.000054176468,"about_ca_topic_score_gemma":5.4496877e-7,"teacher_disagreement_score":0.12510446,"about_ca_system_score_codex":0.00000742633,"about_ca_system_score_gemma":0.00010701571,"threshold_uncertainty_score":0.87917995},"labels":[],"label_agreement":null},{"id":"W2399287554","doi":"10.1097/acm.0000000000001234","title":"A Big Data and Learning Analytics Approach to Process-Level Feedback in Cognitive Simulations","year":2016,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":47,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hospital for Sick Children; University of Toronto; Professional Engineers Ontario","funders":"","keywords":"Computer science; Process (computing); Visualization; Big data; Artificial intelligence; Learning analytics; Metacognition; Cognition; Machine learning; Data science; Analytics; Human–computer interaction; Orientation (vector space); Data mining; Psychology","score_opus":0.2529543061901321,"score_gpt":0.4380322477492504,"score_spread":0.18507794155911833,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2399287554","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95823467,0.0006316182,0.022509888,0.013729351,0.000119723074,0.0005820576,0.00006217087,0.000060195853,0.0040703244],"genre_scores_gemma":[0.99379516,0.00046953716,0.00025953024,0.0022887513,0.0006752469,0.000012705155,0.00015314885,0.000020803289,0.0023250938],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982333,0.00006639958,0.00049915194,0.0005408161,0.00036086328,0.00029948007],"domain_scores_gemma":[0.9862193,0.012837726,0.0001035567,0.0002974129,0.0001308974,0.0004110964],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010455386,0.00015988562,0.000456673,0.00021727211,0.00004920689,0.0000033073918,0.0001713737,0.00020544426,0.000054195258],"category_scores_gemma":[0.20252559,0.00009549824,0.000013829262,0.0005309952,0.00022765192,0.00006600098,0.0001486083,0.00071518676,0.000024871219],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00054940424,0.0001322277,0.74438477,0.000109271336,0.000112128815,0.00003553033,0.0056164637,0.000078958146,0.00068463065,0.00015312716,0.0058898255,0.24225365],"study_design_scores_gemma":[0.0374219,0.0026559613,0.83653176,0.05524239,0.0016822515,0.00033279695,0.015398021,0.036759045,0.0002345686,0.0033303495,0.00930193,0.0011089992],"about_ca_topic_score_codex":0.000048333,"about_ca_topic_score_gemma":0.000007732865,"teacher_disagreement_score":0.24114464,"about_ca_system_score_codex":0.000033510903,"about_ca_system_score_gemma":0.0001307627,"threshold_uncertainty_score":0.8041919},"labels":[],"label_agreement":null},{"id":"W2401959660","doi":"10.3233/978-1-61499-432-9-73","title":"A Causal Loop Approach to the Study of Diagnostic Errors","year":2014,"lang":"en","type":"article","venue":"Studies in health technology and informatics","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria","funders":"","keywords":"Causal loop diagram; Diagram; Computer science; Loop (graph theory); Causal model; Diagnostic accuracy; Systematic error; System dynamics; Statistics; Medicine; Artificial intelligence; Mathematics; Radiology","score_opus":0.060481799517343234,"score_gpt":0.40011964306132103,"score_spread":0.3396378435439778,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2401959660","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99026555,0.0013132644,0.00039249274,0.005904602,0.00016504197,0.0010361506,0.0000017730965,0.000048388345,0.0008727383],"genre_scores_gemma":[0.99522704,0.0011011477,0.001719541,0.001772038,0.000020730238,0.000119733835,0.0000014739161,0.000005072091,0.000033253444],"study_design_codex":"observational","study_design_gemma":"qualitative","domain_scores_codex":[0.9986828,0.000048942846,0.0007864498,0.00009766632,0.00014022323,0.00024393442],"domain_scores_gemma":[0.99241894,0.006913881,0.00017982177,0.00033342608,0.00008400623,0.00006993107],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0011619966,0.000105047555,0.0005082115,0.00025034224,0.000115730516,0.0000022784495,0.0001068651,0.000093312636,7.0417707e-7],"category_scores_gemma":[0.085261025,0.000063405496,0.00001710147,0.0006037696,0.0003281274,0.000025790308,0.00019249928,0.00032827404,0.0000052418145],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014172986,0.0023963319,0.73821473,0.0018460731,0.00027502235,0.000005531142,0.06595252,0.0004793615,2.3246507e-7,0.02499392,0.00743594,0.15825862],"study_design_scores_gemma":[0.0114416145,0.026104024,0.38132766,0.0073609557,0.00027753742,0.00020680626,0.5442202,0.008133152,0.000019751864,0.0056448504,0.014702471,0.0005609544],"about_ca_topic_score_codex":0.00004606058,"about_ca_topic_score_gemma":0.00005155681,"teacher_disagreement_score":0.4782677,"about_ca_system_score_codex":0.00003072463,"about_ca_system_score_gemma":0.000048609127,"threshold_uncertainty_score":0.9224442},"labels":[],"label_agreement":null},{"id":"W2402201187","doi":"10.1111/1742-6723.12606","title":"Potential role for psychological skills training in emergency medicine: Part 1 ‐ Introduction and background","year":2016,"lang":"en","type":"review","venue":"Emergency Medicine Australasia","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Manitoba","funders":"Dartmouth College","keywords":"Medicine; Training (meteorology); Medical education; Athletes; Relaxation (psychology); Cognition; Physical therapy; Psychiatry","score_opus":0.10475164422324135,"score_gpt":0.4495353064551078,"score_spread":0.34478366223186646,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2402201187","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00096788345,0.97160906,0.00018476661,0.0063310177,0.014938852,0.0024154428,0.000060904924,0.00013096207,0.0033611355],"genre_scores_gemma":[0.0002856243,0.9574446,0.00014872898,0.000048551596,0.023568617,0.00047166788,0.00061662507,0.0001285918,0.017286967],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99240196,0.00028454498,0.003813972,0.0017132895,0.0007100788,0.0010761627],"domain_scores_gemma":[0.9952193,0.0019507969,0.00093972916,0.00085609156,0.00024839162,0.0007856734],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0020187679,0.0010186929,0.003971378,0.00058258855,0.00013033299,0.0000045353227,0.00032400244,0.0010272891,0.02434454],"category_scores_gemma":[0.024415346,0.0006214412,0.00066052214,0.0008239905,0.0005190032,0.000110709196,0.000074036914,0.0010568713,0.00010822191],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009575388,0.00033061916,0.0003030046,0.002884371,0.00025082674,0.000038437505,0.00029728925,2.2298205e-7,0.000009843095,0.00039573747,0.2631637,0.7322302],"study_design_scores_gemma":[0.0029878162,0.0021991793,0.0015334835,0.04206286,0.0023150225,0.00035635362,0.00040709786,0.0000040569994,3.5818383e-7,0.0019030586,0.9456289,0.00060183014],"about_ca_topic_score_codex":0.00005364059,"about_ca_topic_score_gemma":0.0000182479,"teacher_disagreement_score":0.73162836,"about_ca_system_score_codex":0.00013859454,"about_ca_system_score_gemma":0.00019151202,"threshold_uncertainty_score":0.9996237},"labels":[],"label_agreement":null},{"id":"W2402471406","doi":"","title":"Fostering Diagnostic Accuracy in a Medical Intelligent Tutoring System.","year":2013,"lang":"en","type":"article","venue":"AIED Workshops","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Computer science; Intelligent tutoring system; Artificial intelligence; Human–computer interaction","score_opus":0.04059755092363861,"score_gpt":0.34766565679484995,"score_spread":0.30706810587121136,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2402471406","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97861767,0.0011394784,0.0012390097,0.0043698745,0.0010868828,0.00093429524,0.0000012078654,0.00026151523,0.012350039],"genre_scores_gemma":[0.99731576,0.0002866786,0.00039878945,0.00097208866,0.00046921286,0.00022883974,0.00001430953,0.00003407861,0.00028025737],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9978156,0.00006975757,0.0006832823,0.00038940794,0.0005394297,0.0005025213],"domain_scores_gemma":[0.9568671,0.04171242,0.00010967468,0.0005228309,0.00009953253,0.0006884166],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00056191755,0.00020334542,0.0004852501,0.00012759576,0.00004555473,0.000059155078,0.00019132048,0.00025406005,0.00086175534],"category_scores_gemma":[0.15263031,0.00016717639,0.00012232746,0.00035801352,0.00007039096,0.00009559831,0.00034176602,0.0005798668,0.0009807105],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000779974,0.0005012885,0.111692764,0.00019508715,0.000074877105,0.0013905829,0.0006447766,0.00009533183,0.00006172358,0.001030152,0.003899679,0.88033575],"study_design_scores_gemma":[0.00969081,0.0007504979,0.84660345,0.10768358,0.00034236387,0.0013322455,0.0046787537,0.020110574,0.0011616889,0.0015616517,0.004597227,0.0014871698],"about_ca_topic_score_codex":0.00042791458,"about_ca_topic_score_gemma":0.000030548385,"teacher_disagreement_score":0.87884855,"about_ca_system_score_codex":0.00017667993,"about_ca_system_score_gemma":0.00014410113,"threshold_uncertainty_score":0.99979717},"labels":[],"label_agreement":null},{"id":"W2402997040","doi":"10.1177/1049732316650418","title":"Examining the Influence of Context and Professional Culture on Clinical Reasoning Through Rhetorical-Narrative Analysis","year":2016,"lang":"en","type":"article","venue":"Qualitative Health Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":26,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"Royal College of Physicians and Surgeons of Canada","keywords":"Rhetorical question; Narrative; Context (archaeology); Narrative inquiry; Psychology; Sociology; Qualitative research; Epistemology; Social psychology; Linguistics; Social science; History; Philosophy","score_opus":0.5009035667002973,"score_gpt":0.6487979135113933,"score_spread":0.14789434681109592,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2402997040","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96821845,0.0010441715,0.00027041865,0.029183578,0.0000715864,0.0006290404,0.000032974287,0.000018354755,0.00053145183],"genre_scores_gemma":[0.9929454,0.0014264658,0.0010956432,0.0021847456,0.00012642964,0.00007644919,0.000009637674,0.000012887718,0.0021223773],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.9899006,0.0068144156,0.0009965207,0.00055096945,0.0012185744,0.0005189524],"domain_scores_gemma":[0.81433123,0.18273017,0.00043532177,0.0005323948,0.001461842,0.00050901965],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.017459497,0.00015110463,0.00080040883,0.00014305532,0.0004182855,0.000013765324,0.00015187915,0.00016909746,0.00006877867],"category_scores_gemma":[0.23402882,0.000067215224,0.00013680902,0.0011358157,0.0015174141,0.000080656915,0.000116780124,0.0010108141,0.000026330366],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0034701612,0.0013668813,0.37145,0.00020597885,0.0013753871,0.00002838628,0.4590076,0.00001056804,0.00013757923,0.050533567,0.03158641,0.08082749],"study_design_scores_gemma":[0.0030965628,0.008224109,0.8049038,0.0072665024,0.00010105097,0.0000044822523,0.17072749,0.00004237228,0.000048189173,0.0026786348,0.0027521867,0.00015461797],"about_ca_topic_score_codex":0.0003737068,"about_ca_topic_score_gemma":0.000031157288,"teacher_disagreement_score":0.4334538,"about_ca_system_score_codex":0.00014009244,"about_ca_system_score_gemma":0.00076846534,"threshold_uncertainty_score":0.77242327},"labels":[],"label_agreement":null},{"id":"W2404949001","doi":"10.1007/s40037-016-0267-3","title":"Bonfire red titles","year":2016,"lang":"en","type":"editorial","venue":"Perspectives on Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Medical education; Computer science; Data science; Medicine; Library science","score_opus":0.010596888633917428,"score_gpt":0.389574524290382,"score_spread":0.37897763565646453,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2404949001","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005379044,0.0025284374,0.000037488808,0.029814914,0.8770694,0.00036219723,0.000042631877,0.0001575512,0.08944945],"genre_scores_gemma":[0.0034834414,0.004784178,0.00015441277,0.00082704733,0.9587892,0.00009765344,0.0004570357,0.00006657258,0.03134042],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99673456,0.000087585875,0.00037670392,0.0006905089,0.0018308709,0.0002797964],"domain_scores_gemma":[0.9800535,0.017068274,0.00025227992,0.00078988104,0.00079955783,0.001036519],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0005123025,0.00027520806,0.00054053933,0.00016525165,0.000064825465,0.000025368612,0.00019729404,0.0012136421,0.0058273613],"category_scores_gemma":[0.45189193,0.00018746979,0.00022210141,0.00013480817,0.00025325682,0.000027989152,0.00003693921,0.0011496168,0.0009859557],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00017073732,0.0012260763,0.000059514772,0.000038018443,0.00006440412,0.000014861804,0.00046997372,9.8730855e-9,0.0000016637039,0.0006221466,0.95804936,0.03928325],"study_design_scores_gemma":[0.0011022829,0.00044073412,0.0007838109,0.009329002,0.00014885697,0.000007750309,0.00090885203,0.0000035639105,0.0000068945833,0.00062909327,0.9864158,0.00022335055],"about_ca_topic_score_codex":0.00009933642,"about_ca_topic_score_gemma":0.00000780674,"teacher_disagreement_score":0.45137963,"about_ca_system_score_codex":0.0004696432,"about_ca_system_score_gemma":0.0067896615,"threshold_uncertainty_score":0.9997919},"labels":[],"label_agreement":null},{"id":"W2405477931","doi":"10.2310/8000.2014.141491","title":"The arithmetic of gestalt","year":2014,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Daughter; Triage; Medicine; Clothing; Patrolling; Medical emergency; Law","score_opus":0.070938732583288,"score_gpt":0.3708664890610505,"score_spread":0.2999277564777625,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2405477931","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.82264435,0.019714106,0.0012747442,0.073062226,0.021981843,0.00032442342,0.00000835167,0.000011934501,0.06097801],"genre_scores_gemma":[0.99534166,0.0011239034,0.00011782669,0.00029465617,0.0013970265,8.536899e-7,0.0000020520868,0.000012937849,0.0017090916],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9984242,0.000064879525,0.0008806299,0.000080312806,0.00030063788,0.00024934506],"domain_scores_gemma":[0.9950776,0.002657156,0.00038152156,0.00026524861,0.0005767436,0.001041708],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0014882619,0.00009425949,0.0003861799,0.00014169845,0.00008764145,0.0000018481124,0.0001738653,0.00005982742,0.001283483],"category_scores_gemma":[0.088502586,0.00005177816,0.0001394043,0.0002522955,0.0002608967,0.000020270933,0.000005147343,0.0002723521,0.000009843683],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021754745,0.00013178808,0.25237852,0.00012821321,0.00043143192,0.0002418912,0.0014787259,0.000037620135,0.0007710451,0.012517754,0.61462146,0.11704402],"study_design_scores_gemma":[0.003915592,0.005431861,0.32829112,0.0050681382,0.0008647972,0.00029430908,0.00085801695,0.0002630175,0.00023228364,0.007853472,0.64670515,0.00022226159],"about_ca_topic_score_codex":0.0028671734,"about_ca_topic_score_gemma":0.0040783044,"teacher_disagreement_score":0.17269729,"about_ca_system_score_codex":0.000035837827,"about_ca_system_score_gemma":0.00069111196,"threshold_uncertainty_score":0.9996295},"labels":[],"label_agreement":null},{"id":"W2411649786","doi":"10.1007/s10459-016-9692-7","title":"Sequential dependencies in categorical judgments of radiographic images","year":2016,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Categorical variable; Context (archaeology); Context effect; Psychology; Task (project management); Cognitive psychology; Sequence (biology); Social psychology; Computer science; Mathematics; Machine learning","score_opus":0.028753847497368215,"score_gpt":0.41914819886684346,"score_spread":0.39039435136947526,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2411649786","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97800726,0.011664832,0.00044651577,0.0070787757,0.0010826809,0.00033993312,0.0000024692815,0.000014302594,0.0013632323],"genre_scores_gemma":[0.98653173,0.010874949,0.0020557037,0.00035602678,0.000066139604,0.00003187706,0.0000024501799,0.0000030629687,0.00007804939],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99863726,0.00007361789,0.00048521673,0.0002587325,0.00029497105,0.00025022952],"domain_scores_gemma":[0.9974759,0.0020543109,0.00017359846,0.00012614756,0.00005277538,0.00011726168],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010290388,0.000067532164,0.00021366087,0.0003175063,0.000039904768,0.0000051866523,0.00010617422,0.0000359447,0.000034039364],"category_scores_gemma":[0.006223406,0.00004392854,0.000032378703,0.0008090638,0.00035395313,0.00024739795,0.000015611899,0.00007409846,0.000004701663],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000022972108,0.00029383626,0.8621274,0.00005532445,8.234452e-7,0.0000010578857,0.00013105317,0.000019266974,0.00010058269,0.0016024954,0.000106858635,0.13553831],"study_design_scores_gemma":[0.0010288808,0.0006704854,0.9774479,0.0031522322,0.000005351793,0.000021429045,0.0005737955,0.000040225954,0.00077005546,0.01527272,0.000908175,0.0001087233],"about_ca_topic_score_codex":0.0012406118,"about_ca_topic_score_gemma":0.00031030498,"teacher_disagreement_score":0.13542959,"about_ca_system_score_codex":0.00012524381,"about_ca_system_score_gemma":0.0010851995,"threshold_uncertainty_score":0.7450448},"labels":[],"label_agreement":null},{"id":"W2415965036","doi":"10.1007/s40037-016-0276-2","title":"Reducing length of stay and satisfying learner needs","year":2016,"lang":"en","type":"article","venue":"Perspectives on Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Western University","funders":"","keywords":"Psychosocial; Curriculum; Emergency department; Medicine; Medical education; Family medicine; Psychology; Nursing; Psychiatry; Pedagogy","score_opus":0.017443884502700832,"score_gpt":0.3583520926444104,"score_spread":0.3409082081417096,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2415965036","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97485465,0.0009564065,0.0002795956,0.016586855,0.00023471682,0.00012634844,0.0000014290823,0.000030439196,0.0069295713],"genre_scores_gemma":[0.99601513,0.0012436565,0.00067921344,0.00048807575,0.00034995118,0.000011672321,0.0000025965403,0.000011996796,0.0011977305],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9989628,0.000053918437,0.00022359286,0.00022384105,0.0004012006,0.00013462706],"domain_scores_gemma":[0.9955449,0.0037090762,0.00008050164,0.0001933051,0.00013818509,0.00033405094],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00042859142,0.00008590337,0.00020853817,0.00013394826,0.00003414671,0.000005535968,0.000043403958,0.000099084784,0.0006573928],"category_scores_gemma":[0.09963765,0.00005281312,0.00004727052,0.0001494104,0.00024796824,0.000039187904,0.000018554572,0.00016075514,0.000019713092],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002126744,0.0013239143,0.045916438,0.000038338258,0.00007235016,0.000003349819,0.014086915,3.3818606e-7,0.0010848867,0.021289986,0.0026617202,0.9133091],"study_design_scores_gemma":[0.006897872,0.002793101,0.87154067,0.020451814,0.0003343397,0.00015799898,0.084719345,0.00017159426,0.0035548375,0.0047733965,0.004062901,0.00054212287],"about_ca_topic_score_codex":0.00012750593,"about_ca_topic_score_gemma":0.0000017391185,"teacher_disagreement_score":0.912767,"about_ca_system_score_codex":0.00008351478,"about_ca_system_score_gemma":0.0005816146,"threshold_uncertainty_score":0.90794647},"labels":[],"label_agreement":null},{"id":"W2417633983","doi":"10.1007/s11019-016-9714-5","title":"Uncertainty and objectivity in clinical decision making: a clinical case in emergency medicine","year":2016,"lang":"en","type":"article","venue":"Medicine Health Care and Philosophy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":78,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"National Institute for Health and Care Research","keywords":"Philosophy of medicine; Clinical judgment; Objectivity (philosophy); Standardization; Judgement; Intuition; Guideline; Engineering ethics; Medical law; Clinical judgement; Epistemology; Clinical Practice; Evidence-based medicine; Creativity; Psychology; Philosophy of biology; Management science; Medicine; MEDLINE; Philosophy of science; Alternative medicine; Social psychology; Cognitive science; Law; Nursing; Political science; Psychiatry; Intensive care medicine","score_opus":0.09142841511080897,"score_gpt":0.48718065461098403,"score_spread":0.3957522395001751,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2417633983","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.91273683,0.014985636,0.00012087645,0.06725685,0.0014303132,0.00073089625,0.000012734199,0.00004416667,0.0026816996],"genre_scores_gemma":[0.9698769,0.021916915,0.00023511983,0.005814835,0.0020686002,0.000023020017,0.0000115961075,0.00002563981,0.0000273905],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99493665,0.0005932891,0.0026799513,0.00088544603,0.00041994013,0.00048471292],"domain_scores_gemma":[0.9768129,0.021364547,0.00036573195,0.00048287137,0.00013061715,0.00084331783],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0051176203,0.0003007377,0.0015759309,0.00032626453,0.00008877856,0.0000023913426,0.000077926445,0.00033213315,0.0001814126],"category_scores_gemma":[0.06976785,0.00017119572,0.000109418936,0.00045701806,0.000755792,0.000055487497,0.00009774427,0.0007292799,0.0000074125182],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00088536605,0.00020356804,0.5947106,0.00014577075,0.000013358385,0.0010989327,0.0008290066,1.3539399e-7,4.6459064e-7,0.00033854885,0.004237024,0.39753726],"study_design_scores_gemma":[0.021240067,0.011879364,0.904751,0.026548343,0.00010657738,0.0011576277,0.00084605796,0.00007536708,1.8145617e-7,0.027199779,0.005886133,0.00030951615],"about_ca_topic_score_codex":0.0014386957,"about_ca_topic_score_gemma":0.0018215192,"teacher_disagreement_score":0.39722773,"about_ca_system_score_codex":0.00012847499,"about_ca_system_score_gemma":0.00030259875,"threshold_uncertainty_score":0.9380679},"labels":[],"label_agreement":null},{"id":"W2418063830","doi":"10.1016/j.ecns.2016.04.001","title":"Online Virtual Simulation and Diagnostic Reasoning: A Scoping Review","year":2016,"lang":"en","type":"review","venue":"Clinical Simulation in Nursing","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":114,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Regina; Saskatchewan Polytechnic","funders":"","keywords":"Virtual patient; Health care; Variety (cybernetics); Medical education; Instructional simulation; Nursing; Medicine; Virtual reality; Psychology; Computer science; Human–computer interaction","score_opus":0.13989303817114992,"score_gpt":0.5524018586512925,"score_spread":0.4125088204801426,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2418063830","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00021570297,0.99408793,0.0014178155,0.00027156953,0.00074620475,0.002924176,0.000028099737,0.000112046604,0.00019647625],"genre_scores_gemma":[0.0024915345,0.99472725,0.00043655426,0.0005912059,0.0011715873,0.00005165586,0.0003342406,0.000095646865,0.00010032625],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9936896,0.00076990615,0.003581598,0.0010457609,0.00045004752,0.00046306886],"domain_scores_gemma":[0.74358016,0.2533634,0.001375452,0.0008070482,0.00024840323,0.00062555564],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0019613116,0.00054075423,0.0037913406,0.000241165,0.00008801434,0.00004108425,0.00013566317,0.00086288154,0.00015007719],"category_scores_gemma":[0.29481804,0.00040168228,0.0006982897,0.00052582007,0.00033601475,0.00011812786,0.00007444782,0.0009723027,0.00009284704],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000033412,0.00056317856,0.00081494317,0.013603975,0.000036099205,0.00003573445,0.000008807328,0.00076688465,2.0912818e-9,0.000072459516,0.00004475979,0.98401976],"study_design_scores_gemma":[0.0015165235,0.00034819037,0.0012422391,0.9503742,0.0013992524,0.000023071858,0.0000023094035,0.008594465,5.6343703e-9,0.00010772042,0.035987124,0.00040490428],"about_ca_topic_score_codex":0.0000019383308,"about_ca_topic_score_gemma":0.0000012580485,"teacher_disagreement_score":0.98361486,"about_ca_system_score_codex":0.0002550882,"about_ca_system_score_gemma":0.00054489105,"threshold_uncertainty_score":0.9998435},"labels":[],"label_agreement":null},{"id":"W2418237503","doi":"10.1097/smj.0b013e3182621b1a","title":"Commentary on “Propagation of Uncertainty in Bayesian Diagnostic Test Interpretation”","year":2012,"lang":"en","type":"letter","venue":"Southern Medical Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Vancouver Coastal Health Research Institute","funders":"","keywords":"Medicine; Interpretation (philosophy); Test (biology); Bayesian probability; Artificial intelligence","score_opus":0.013616377097411304,"score_gpt":0.3058383872197302,"score_spread":0.2922220101223189,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2418237503","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.016962016,0.00050340197,0.00075667864,0.979181,0.0011731865,0.00038117656,0.00011416249,0.000028675453,0.00089973945],"genre_scores_gemma":[0.31862912,0.00021403636,0.00008355998,0.67253697,0.00790855,0.000015281987,0.00038048605,0.000056902507,0.00017510694],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.996074,0.00034618733,0.0011749639,0.00026382075,0.0016702763,0.00047073863],"domain_scores_gemma":[0.9447902,0.053573743,0.00058596267,0.0003304284,0.00016354237,0.0005561125],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0015173898,0.00029928738,0.000797484,0.00026898386,0.000048681984,0.000020898487,0.0002689333,0.000936753,0.0026657775],"category_scores_gemma":[0.17247833,0.00021541797,0.00029818952,0.00017010339,0.000275529,0.000032291297,0.000048357786,0.005139425,0.00017751144],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001918795,0.00079314236,0.04986294,0.00014193909,0.00012687403,0.002338422,0.0010260699,0.000024641275,0.000002371608,0.0000027191556,0.88552535,0.059963677],"study_design_scores_gemma":[0.042807706,0.0129283145,0.079571776,0.3721018,0.0049118446,0.0073717707,0.0043140524,0.013598672,0.00009176112,0.008228337,0.45030233,0.0037716187],"about_ca_topic_score_codex":0.00015463113,"about_ca_topic_score_gemma":0.00002445105,"teacher_disagreement_score":0.43522298,"about_ca_system_score_codex":0.00022843505,"about_ca_system_score_gemma":0.0004136312,"threshold_uncertainty_score":0.9982459},"labels":[],"label_agreement":null},{"id":"W2419153322","doi":"10.3760/cma.j.issn.2095-4352.2014.10.022","title":"[The experience and thought in Department of Critical Care Medicine, Sunnybrook Health Sciences Centre, University of Toronto].","year":2014,"lang":"en","type":"article","venue":"PubMed","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Family medicine; Gerontology","score_opus":0.0267865465795042,"score_gpt":0.32452974763068937,"score_spread":0.2977432010511852,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2419153322","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97565484,0.0077910754,0.000050680555,0.007352542,0.00017798865,0.00044845126,0.0000060282955,0.000011328092,0.008507059],"genre_scores_gemma":[0.9988812,0.0006474021,0.00019564954,0.00015010149,0.000030419118,0.000005671176,0.000001820397,0.0000020350435,0.00008565838],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99915755,0.00007970952,0.00018554284,0.00015095223,0.00021141369,0.00021483576],"domain_scores_gemma":[0.9954125,0.0041607814,0.000057249483,0.00012932897,0.000059242415,0.000180868],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00062638154,0.000050255432,0.00021614067,0.000013107899,0.00006807071,0.0000024025398,0.000076217264,0.000027040625,0.000014521777],"category_scores_gemma":[0.013930717,0.00003353693,0.000021629106,0.00006808405,0.00094077154,0.000027618027,0.00003475535,0.00004436389,2.1833782e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000160952,0.00020411162,0.60284036,0.00015449278,0.000008789246,0.0000068322215,0.0065672244,0.0000015721599,0.0000029612524,0.011962035,0.0011184108,0.37697223],"study_design_scores_gemma":[0.0008997647,0.00021432,0.9872939,0.0002691522,0.000016391881,0.0000025713448,0.007537977,0.000034088098,0.000060684815,0.000086175314,0.0035480538,0.00003691857],"about_ca_topic_score_codex":0.0028577403,"about_ca_topic_score_gemma":0.002255428,"teacher_disagreement_score":0.3844535,"about_ca_system_score_codex":0.00005743056,"about_ca_system_score_gemma":0.000059478294,"threshold_uncertainty_score":0.99437535},"labels":[],"label_agreement":null},{"id":"W2419267504","doi":"10.1007/s40037-016-0268-2","title":"Exploring cognitive integration of basic science and its effect on diagnostic reasoning in novices","year":2016,"lang":"en","type":"article","venue":"Perspectives on Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":88,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; Toronto General Hospital; Toronto Rehabilitation Institute; University of Toronto","funders":"","keywords":"Basic science; Cognition; Test (biology); Clinical Practice; Psychology; Computer science; Medicine; Cognitive psychology; Physical therapy; Psychiatry","score_opus":0.042602759093186705,"score_gpt":0.38317906047800565,"score_spread":0.3405763013848189,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2419267504","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99296635,0.00044620552,0.000054875574,0.0035450805,0.00025937206,0.0003372526,0.0000018596108,0.00002144893,0.0023675747],"genre_scores_gemma":[0.9983411,0.00097553583,0.000040960404,0.0002935057,0.00017703808,0.00012454338,0.0000028306229,0.0000094209645,0.000035095403],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99831784,0.00009433423,0.00022975053,0.00041581315,0.00076338334,0.00017887774],"domain_scores_gemma":[0.9630671,0.03564384,0.00013285734,0.00018180914,0.0005505614,0.0004238441],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0013832628,0.00011648347,0.00023837683,0.00032122948,0.000051826915,0.000011037621,0.000069453476,0.000061430736,0.00011071144],"category_scores_gemma":[0.6081344,0.00007223954,0.000031421354,0.0004572932,0.00035972137,0.00017169547,0.000021700493,0.00021122432,0.00003492427],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007726918,0.002118031,0.043743204,0.00006359587,0.000021053122,0.000013267634,0.010591645,6.094108e-7,0.0026275215,0.012184713,0.000049398594,0.92781425],"study_design_scores_gemma":[0.0027620567,0.0032532206,0.9238963,0.03879494,0.00006382694,0.000016960987,0.009419995,0.00024390068,0.021201773,0.00017138637,0.000014397719,0.00016120893],"about_ca_topic_score_codex":0.00006118569,"about_ca_topic_score_gemma":0.0000070310516,"teacher_disagreement_score":0.9276531,"about_ca_system_score_codex":0.00024112978,"about_ca_system_score_gemma":0.00085148786,"threshold_uncertainty_score":0.3951665},"labels":[],"label_agreement":null},{"id":"W2424551559","doi":"10.1891/9780826152237.0043","title":"Lower Extremity Edema","year":2018,"lang":"en","type":"article","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières","funders":"","keywords":"Medical diagnosis; Differential diagnosis; Medicine; Heart failure; Physical examination; Complaint; Signs and symptoms; Intensive care medicine; Diagnostic test; Clinical Practice; Edema; Physical therapy; Pediatrics; Surgery; Radiology; Internal medicine; Pathology","score_opus":0.033334627439725616,"score_gpt":0.353254722311281,"score_spread":0.3199200948715554,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2424551559","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.59538555,0.000032092623,0.00068053545,0.0026942948,0.00040135317,0.0000662421,6.29887e-7,0.000099372446,0.4006399],"genre_scores_gemma":[0.96951765,0.000010125447,0.0019891846,0.0044791284,0.0007197206,0.0000019578738,0.0000025791921,0.0000062551285,0.023273405],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9994982,0.0000072915573,0.000115074356,0.00013590716,0.00011494295,0.00012860655],"domain_scores_gemma":[0.9978654,0.0016586445,0.000015540518,0.00021754377,0.00008091393,0.00016199127],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0001323591,0.000050743944,0.00011862751,0.000015001916,0.000026736252,0.0000059956483,0.000030642885,0.00005638361,0.0065715057],"category_scores_gemma":[0.019265886,0.000034260614,0.00005657585,0.00006732511,0.00009958688,0.000012848621,0.00002142371,0.000059806152,0.002115578],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00029462937,0.0008019206,0.3779521,0.000007599417,0.000044626362,0.000104807244,0.00007641513,7.354454e-8,0.00046192112,0.0041533066,0.51467794,0.10142465],"study_design_scores_gemma":[0.003375411,0.0012244271,0.8138481,0.000933072,0.00014201709,0.00012846234,0.000055711476,0.0003952832,0.0040295552,0.0024598935,0.17313765,0.00027042296],"about_ca_topic_score_codex":0.000035417946,"about_ca_topic_score_gemma":0.0000088484,"teacher_disagreement_score":0.435896,"about_ca_system_score_codex":0.000010310095,"about_ca_system_score_gemma":0.00003876165,"threshold_uncertainty_score":0.9986614},"labels":[],"label_agreement":null},{"id":"W2424610854","doi":"10.1080/10401334.2016.1190730","title":"In Brief: Validity of Case Summaries in Written Examinations of Clinical Reasoning","year":2016,"lang":"en","type":"article","venue":"Teaching and Learning in Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; McGill University","funders":"","keywords":"Generalizability theory; Test (biology); Inter-rater reliability; Rubric; Objective structured clinical examination; Construct validity; Medicine; MEDLINE; Curriculum; Psychology; Clinical psychology; Psychometrics; Medical education; Mathematics education; Rating scale; Developmental psychology","score_opus":0.07290377866669441,"score_gpt":0.4191633828742299,"score_spread":0.3462596042075355,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2424610854","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9940107,0.00032574328,0.00021216668,0.0034179168,0.00010600021,0.00012795767,0.0000017804722,0.000019265945,0.0017784833],"genre_scores_gemma":[0.99755496,0.00029167678,0.0013328282,0.0001521518,0.00014018122,0.000005043938,0.000006511489,0.000012302609,0.00050435174],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99694026,0.0011778093,0.0011563338,0.0002830918,0.00023138893,0.00021110303],"domain_scores_gemma":[0.95937276,0.039993763,0.0002677753,0.00019177067,0.000054174372,0.00011972936],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.011770319,0.00011890121,0.00084590504,0.0003087844,0.00004283598,0.0000023084563,0.000055081,0.00015262974,0.00005720991],"category_scores_gemma":[0.3275527,0.00007903238,0.000048425463,0.00021290987,0.0004134365,0.00006008891,0.000045912955,0.0012379586,8.729735e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011546484,0.00023528548,0.913525,0.0000558423,0.000009667776,0.00065110775,0.0025443977,0.000012244638,0.00012729886,0.0005546076,0.0001721366,0.08199694],"study_design_scores_gemma":[0.006348607,0.0009059682,0.9757882,0.012850747,0.000052422387,0.0002525165,0.0019179871,0.00063748076,0.000022967339,0.00018829311,0.000941503,0.00009332965],"about_ca_topic_score_codex":0.0063055567,"about_ca_topic_score_gemma":0.00048094286,"teacher_disagreement_score":0.3157824,"about_ca_system_score_codex":0.00003220725,"about_ca_system_score_gemma":0.00008466674,"threshold_uncertainty_score":0.9532158},"labels":[],"label_agreement":null},{"id":"W2429364556","doi":"10.1097/acm.0000000000000896","title":"Cause and Effect","year":2015,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":57,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Public Health","funders":"","keywords":"Test (biology); Basic science; Psychology; Cognition; Cognitive psychology; Medical education; Medicine; Neuroscience","score_opus":0.07933634874970559,"score_gpt":0.41694266613494957,"score_spread":0.337606317385244,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2429364556","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9690144,0.0061174813,0.0000589735,0.016467338,0.0003890491,0.00023337922,8.0193576e-7,0.0000925646,0.0076260353],"genre_scores_gemma":[0.9902243,0.0009965923,0.000052251595,0.0045866747,0.000991213,0.0000112239395,0.000015876938,0.000015123914,0.0031067436],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9989237,0.000052485855,0.00027210414,0.0002235136,0.0003258848,0.0002022747],"domain_scores_gemma":[0.99336827,0.0055729235,0.0000544182,0.00019544615,0.000058157795,0.0007508132],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0011275865,0.00012301344,0.00040121612,0.00006295311,0.000021015408,0.0000016046706,0.000053302836,0.00019264649,0.00007243452],"category_scores_gemma":[0.09645577,0.00007578906,0.000020663914,0.0001329157,0.00023556016,0.000026108892,0.000039844854,0.0005187819,0.00009700842],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004910614,0.0000179234,0.6404808,0.000066130706,0.00009549039,0.00034941768,0.0021308837,6.019506e-7,0.00061676657,0.0015167714,0.32190257,0.032331593],"study_design_scores_gemma":[0.09501578,0.02705765,0.48368335,0.030049551,0.0050175586,0.008601374,0.003694644,0.001245647,0.003682169,0.016827745,0.32367083,0.001453678],"about_ca_topic_score_codex":0.000076911616,"about_ca_topic_score_gemma":7.6718317e-7,"teacher_disagreement_score":0.15679744,"about_ca_system_score_codex":0.000033715998,"about_ca_system_score_gemma":0.00006566138,"threshold_uncertainty_score":0.91115516},"labels":[],"label_agreement":null},{"id":"W2438721897","doi":"10.3138/jvme.0815-130r1","title":"An Investigation into the Clinical Reasoning Development of Veterinary Students","year":2016,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":25,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"KU Leuven","keywords":"Curriculum; Thematic analysis; Medical education; Documentation; Veterinary education; Psychology; Focus group; Medicine; Veterinary medicine; Qualitative research; Pedagogy; Computer science","score_opus":0.1188468470912927,"score_gpt":0.49392614786876593,"score_spread":0.3750793007774732,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2438721897","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99267066,0.00028192403,0.00047674394,0.004988514,0.0013987719,0.00013217384,3.2990997e-7,0.000010356199,0.000040513554],"genre_scores_gemma":[0.98693484,0.00032955507,0.01077204,0.00095269026,0.00089767127,0.000008954871,0.0000056835133,0.000014338394,0.00008419885],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9967147,0.00038600474,0.0014407893,0.00017557734,0.0011310262,0.00015187124],"domain_scores_gemma":[0.9941647,0.003531191,0.0007929429,0.00032090553,0.0004552094,0.0007350731],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0043161474,0.00012151794,0.0003837263,0.000103493796,0.000075437616,0.000013994006,0.0003562999,0.00016906488,0.00019529807],"category_scores_gemma":[0.04175867,0.00006270364,0.00013895053,0.00014294172,0.00029485303,0.0001644284,0.00007695583,0.00033214432,0.0000129777],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008854227,0.0018686152,0.21265809,0.00004153169,0.000103828344,0.00006508539,0.0028342367,2.504734e-7,0.0044467077,0.00003583462,0.0012212378,0.77583915],"study_design_scores_gemma":[0.0024373608,0.0072069094,0.97054833,0.009198162,0.0001236851,0.001215485,0.0008393735,0.000022983922,0.0004977943,0.0002798667,0.007509856,0.00012022112],"about_ca_topic_score_codex":0.000013784528,"about_ca_topic_score_gemma":0.0000017464782,"teacher_disagreement_score":0.7757189,"about_ca_system_score_codex":0.00011359707,"about_ca_system_score_gemma":0.0033910759,"threshold_uncertainty_score":0.966313},"labels":[],"label_agreement":null},{"id":"W2439832058","doi":"10.1007/978-1-0716-1138-8_3","title":"Definitions of Bias in Clinical Research","year":2021,"lang":"en","type":"book-chapter","venue":"Methods in molecular biology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Memorial University of Newfoundland","funders":"","keywords":"Psychology; Computer science","score_opus":0.4844115607422507,"score_gpt":0.6084546793884437,"score_spread":0.12404311864619294,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2439832058","genre_codex":"other","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0055258153,0.013759327,0.06701967,0.0021423257,0.0009933106,0.0009707213,0.00007752569,0.000033646626,0.90947765],"genre_scores_gemma":[0.0010659865,0.015264544,0.96288854,0.0010401319,0.00019583758,0.000051132698,0.00034755128,0.000082092454,0.019064175],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99433863,0.002933677,0.0014768873,0.00067711197,0.00021455194,0.0003591294],"domain_scores_gemma":[0.93760145,0.06099536,0.00019604847,0.0007565538,0.00030582907,0.00014473489],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.011629525,0.00020462376,0.0013905085,0.0005958383,0.000014263727,0.000003814146,0.0001730546,0.0014085886,0.00045109735],"category_scores_gemma":[0.23741175,0.00018655362,0.0004276107,0.00021353946,0.0007982464,0.0000037548011,0.00024954608,0.0022167398,0.000029832965],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015909951,0.00050662254,0.0143876085,0.00008432709,0.00015733385,0.0019089313,0.00003337697,0.000002323409,0.0024784245,0.8851907,0.00031577624,0.094775446],"study_design_scores_gemma":[0.0036267731,0.0019733105,0.0075830235,0.009695975,0.00021985723,0.00012739708,0.000049071554,0.00003589953,0.0029335895,0.93614674,0.037130933,0.00047741996],"about_ca_topic_score_codex":0.00008359674,"about_ca_topic_score_gemma":0.000047294347,"teacher_disagreement_score":0.8958689,"about_ca_system_score_codex":0.00006452951,"about_ca_system_score_gemma":0.0005771836,"threshold_uncertainty_score":0.99988776},"labels":[],"label_agreement":null},{"id":"W2440226612","doi":"10.1017/cjn.2016.117","title":"P.011 Evaluation of educational needs in neurology in the province of Quebec: a survey-based study","year":2016,"lang":"en","type":"article","venue":"Canadian Journal of Neurological Sciences / Journal Canadien des Sciences Neurologiques","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Neurology; Medicine; Medical school; Clinical neurology; Neurological examination; Medical education; Family medicine; Psychology; Pediatrics; Psychiatry","score_opus":0.07811380428438527,"score_gpt":0.36165734488763895,"score_spread":0.2835435406032537,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2440226612","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9842953,0.00032728116,0.0000048184297,0.01430189,0.00030694206,0.00038250184,0.00000901251,0.0000017872717,0.00037043987],"genre_scores_gemma":[0.9981289,0.000033605535,0.00010274377,0.0016439438,0.00007242062,0.0000060064,2.9404646e-7,0.0000049379437,0.0000071560453],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9927123,0.0038673135,0.0012321792,0.0003256981,0.001219846,0.0006426738],"domain_scores_gemma":[0.9819374,0.015615844,0.00089506607,0.00018258569,0.00080523174,0.0005638556],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.021016672,0.00018525468,0.0005487588,0.0013273399,0.00026105135,0.0000668247,0.0011730074,0.00013372299,0.00016358713],"category_scores_gemma":[0.078898266,0.00009174784,0.00014269348,0.0019096246,0.0038513914,0.00028112502,0.00002734531,0.00067648856,7.479882e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00024487876,0.0002832696,0.9935078,0.0000031228237,0.000004909474,0.0005341573,0.00026343297,0.0012708147,0.000014110124,0.00019475003,0.0000991237,0.0035796443],"study_design_scores_gemma":[0.0010896478,0.043157563,0.95067555,0.00013465501,0.000037757374,0.001095013,0.00018362088,0.00045824013,0.000019216382,0.003043674,0.000020373589,0.00008470543],"about_ca_topic_score_codex":0.020065399,"about_ca_topic_score_gemma":0.6897607,"teacher_disagreement_score":0.6696953,"about_ca_system_score_codex":0.00015865469,"about_ca_system_score_gemma":0.013008089,"threshold_uncertainty_score":0.9988596},"labels":[],"label_agreement":null},{"id":"W2442662700","doi":"10.1503/cmaj.151079","title":"Pareidolia and clinical reasoning: the pattern awakens","year":2015,"lang":"en","type":"article","venue":"Canadian Medical Association Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"University of British Columbia, Okanagan Campus; Kelowna General Hospital; University of British Columbia","funders":"","keywords":"Adventure; Cliff; Jumping; Emergency department; Computer science; Sign (mathematics); Reflection (computer programming); Artificial intelligence; Medicine; History; Nursing; Programming language; Archaeology","score_opus":0.03374419213822955,"score_gpt":0.34436120461411773,"score_spread":0.3106170124758882,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2442662700","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.75842583,0.0009069948,0.0001424428,0.23023604,0.0027220838,0.00017924665,0.000014716428,0.0000330982,0.00733958],"genre_scores_gemma":[0.9586721,0.00041089611,0.00007834682,0.0364534,0.0032842783,0.0000031100096,0.000011701398,0.000016047736,0.0010700985],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99748844,0.00030682038,0.0005751482,0.00016803428,0.0010447038,0.0004168708],"domain_scores_gemma":[0.99019235,0.004894646,0.00021194467,0.0001563333,0.0004957696,0.0040489826],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005806823,0.000105585365,0.0002988576,0.000053765245,0.00017867151,0.000091459835,0.00013338367,0.00040539322,0.0006738912],"category_scores_gemma":[0.20141704,0.00006793622,0.00012717117,0.00013214821,0.00011448084,0.000041037358,0.000024195939,0.0014488947,0.00024139871],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00000674337,0.000027786095,0.64771646,8.074894e-7,0.00007165067,0.0002960078,0.00010151643,2.0515083e-7,3.6504616e-8,0.000041490854,0.26759043,0.084146895],"study_design_scores_gemma":[0.002662309,0.00021627439,0.64506155,0.00041592578,0.00013071552,0.0015544912,0.0005449407,0.00038805904,5.9303073e-7,0.00019813677,0.34871984,0.00010716665],"about_ca_topic_score_codex":0.002838023,"about_ca_topic_score_gemma":0.0075637302,"teacher_disagreement_score":0.20024632,"about_ca_system_score_codex":0.00037043772,"about_ca_system_score_gemma":0.0040426673,"threshold_uncertainty_score":0.8053098},"labels":[],"label_agreement":null},{"id":"W2462464478","doi":"10.82308/20661","title":"The effects of specific support to hypothesis generation on the diagnostic performance of medical students","year":2006,"lang":"en","type":"article","venue":"eScholarship@McGill (McGill)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Medical diagnosis; Test (biology); Curriculum; Disease; Process (computing); Psychology; Medical education; Medicine; Computer science; Pathology","score_opus":0.022723564017313718,"score_gpt":0.28079145339907385,"score_spread":0.2580678893817601,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2462464478","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9940643,0.00014996248,4.4313683e-7,0.000861689,0.0005830508,0.00081991736,0.000045525398,0.000038245176,0.0034369226],"genre_scores_gemma":[0.99704486,0.000907948,0.00009536814,0.0010363243,0.00016029402,0.00010580632,0.000016038042,0.000043479056,0.0005898773],"study_design_codex":"design_other","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.9956851,0.00028472935,0.0008453843,0.00041269755,0.0023645293,0.00040754883],"domain_scores_gemma":[0.9573186,0.041197494,0.0002493115,0.00074263086,0.00021616151,0.00027581726],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0025781912,0.0002625848,0.0004831966,0.00009642223,0.000426315,0.000033022578,0.000692906,0.00021492728,0.00016823152],"category_scores_gemma":[0.09535444,0.0001542011,0.00021427493,0.00039449692,0.00019728337,0.00006503575,0.00018962823,0.0005194277,0.00020452068],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014857695,0.0066732625,0.06203572,0.0005409008,0.000633381,0.00042730704,0.000034357356,0.00027922916,0.06558178,0.12334619,0.003966362,0.7349957],"study_design_scores_gemma":[0.0043409145,0.0048314496,0.42243978,0.0050796503,0.00041654345,0.0000662902,0.000055238623,0.00013472432,0.5304338,0.0017544102,0.029890241,0.0005569398],"about_ca_topic_score_codex":0.00011570796,"about_ca_topic_score_gemma":0.00004671314,"teacher_disagreement_score":0.7344388,"about_ca_system_score_codex":0.00013291757,"about_ca_system_score_gemma":0.000052332252,"threshold_uncertainty_score":0.9122658},"labels":[],"label_agreement":null},{"id":"W2463429832","doi":"10.11622/smedj.2016015","title":"Teaching metacognition in clinical decision-making using a novel mnemonic checklist: an exploratory study","year":2016,"lang":"en","type":"article","venue":"Singapore Medical Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":60,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Dalhousie University","keywords":"Checklist; Mnemonic; Metacognition; Intervention (counseling); Debiasing; Context (archaeology); Exploratory research; Medicine; Cognition; Clinical psychology; Psychology; Medical education; Psychiatry; Cognitive psychology; Social psychology","score_opus":0.08098392058322759,"score_gpt":0.44892453620952333,"score_spread":0.36794061562629576,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2463429832","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96630436,0.00018457956,0.030440353,0.0012906938,0.0011516557,0.00027587483,0.0000033319398,0.00006330758,0.00028586414],"genre_scores_gemma":[0.9858932,0.000073299496,0.010300289,0.0014285924,0.0022210693,0.000006028119,0.0000028936167,0.000051329484,0.000023310298],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99414635,0.00076921127,0.0020701946,0.0005819495,0.0018401382,0.0005921595],"domain_scores_gemma":[0.97918874,0.017953282,0.0005579594,0.00048152567,0.0002857527,0.0015327409],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.013919682,0.0002815246,0.00087715115,0.0003414643,0.00025926373,0.00007887123,0.00030154112,0.00037398541,0.000654955],"category_scores_gemma":[0.21953744,0.00017147142,0.00032468597,0.00026769872,0.00027758026,0.00035470573,0.00014586838,0.002213319,0.000030874875],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00075184513,0.0061909654,0.20444652,0.000008478535,0.00012484597,0.003046971,0.0009292282,0.000003360243,0.0003987736,0.000049077586,0.00015676141,0.78389317],"study_design_scores_gemma":[0.07187949,0.012512426,0.75089186,0.10645644,0.002142118,0.014673243,0.014564856,0.008998665,0.00007316713,0.0144790225,0.0013877932,0.0019409378],"about_ca_topic_score_codex":0.000025280315,"about_ca_topic_score_gemma":0.000109705936,"teacher_disagreement_score":0.78195226,"about_ca_system_score_codex":0.00027232978,"about_ca_system_score_gemma":0.001217574,"threshold_uncertainty_score":0.9615893},"labels":[],"label_agreement":null},{"id":"W2463763854","doi":"","title":"Cognition and expertise: acquisition of medical competence.","year":2000,"lang":"en","type":"article","venue":"PubMed","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":79,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Competence (human resources); Cognition; Health care; Medical education; Medical knowledge; Knowledge management; Psychology; Quality (philosophy); Medicine; Computer science; Social psychology; Political science","score_opus":0.026297136606445535,"score_gpt":0.2986878626768672,"score_spread":0.27239072607042164,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2463763854","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9847448,0.00056390656,0.000038458915,0.0050140754,0.00005581898,0.0002513369,0.000004640718,0.00003203522,0.009294919],"genre_scores_gemma":[0.9974936,0.00070394785,0.00006968814,0.0011722989,0.00011799671,0.00014509718,0.000023857596,0.000004981312,0.0002684783],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99919665,0.000021274109,0.00018752771,0.000119453165,0.00034433618,0.00013073064],"domain_scores_gemma":[0.99867415,0.0009337817,0.000026739905,0.00008524644,0.000033681583,0.0002463826],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0002797356,0.000049527294,0.00016178896,0.000024239995,0.00001680926,0.000005841962,0.000029282864,0.000088279645,0.0014820632],"category_scores_gemma":[0.0046960567,0.000039888953,0.000033811128,0.000063839325,0.000120636614,0.000025866837,0.000010246083,0.00008260745,0.000017059952],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013333604,0.00017496845,0.008109774,0.000015255759,0.0000173078,0.000026695794,0.000053115305,1.6124059e-7,0.00001695673,0.000118194825,0.0010452996,0.9902889],"study_design_scores_gemma":[0.0020881963,0.000078111945,0.99526364,0.00041811043,0.00006866712,0.00009555957,0.00002130975,0.00020463682,0.00075770135,0.0003209952,0.0006251318,0.00005792456],"about_ca_topic_score_codex":0.000036099173,"about_ca_topic_score_gemma":0.0000017333542,"teacher_disagreement_score":0.99023104,"about_ca_system_score_codex":0.000008256827,"about_ca_system_score_gemma":0.00002273054,"threshold_uncertainty_score":0.9994307},"labels":[],"label_agreement":null},{"id":"W2489634693","doi":"10.1057/9781403980939_3","title":"Teaching Clinical Reasoning to Undergraduate Medical Students","year":2005,"lang":"en","type":"book-chapter","venue":"Palgrave Macmillan US eBooks","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Simon Fraser University","funders":"","keywords":"Rationality; Intuition; Psychology; Confirmation bias; Medical education; Medicine; Epistemology; Social psychology; Cognitive science","score_opus":0.034345463756818004,"score_gpt":0.3701233531774222,"score_spread":0.33577788942060416,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2489634693","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.006632107,0.00054237986,0.00022865782,0.003422281,0.0010984436,0.0011905893,0.000052977248,0.0004114081,0.98642117],"genre_scores_gemma":[0.6572329,0.00063221826,0.0035865419,0.020279273,0.006219703,0.000051819763,0.00023340085,0.00039872853,0.31136543],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9921275,0.00018259794,0.0020545041,0.0014852677,0.0032407653,0.00090936426],"domain_scores_gemma":[0.9880864,0.007154426,0.0004958006,0.0013711373,0.00021569383,0.0026765438],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.00321739,0.00093202555,0.0020737373,0.00032996575,0.00021303195,0.00011357121,0.00094948383,0.0015057075,0.00087468635],"category_scores_gemma":[0.02739624,0.00080953096,0.0010240531,0.000027091855,0.0003426061,0.0000022266597,0.00086364546,0.0031995114,0.0017331021],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006830894,0.000082212515,0.0314245,0.00011858967,0.0015083888,0.0035450007,0.00024641398,0.0000018155695,0.000004453224,0.2969689,0.005852033,0.6595646],"study_design_scores_gemma":[0.016945666,0.0053682164,0.047027174,0.06265572,0.0044645243,0.0022607578,0.00006679191,0.00026148348,0.000029276915,0.21564339,0.6403628,0.004914198],"about_ca_topic_score_codex":0.00019750492,"about_ca_topic_score_gemma":0.00018260948,"teacher_disagreement_score":0.67505574,"about_ca_system_score_codex":0.0002311953,"about_ca_system_score_gemma":0.00062860595,"threshold_uncertainty_score":0.99979055},"labels":[],"label_agreement":null},{"id":"W2494177323","doi":"10.1016/j.revmed.2016.06.008","title":"Validation d’une grille d’évaluation des compétences cliniques utilisée par des patients standardisés","year":2016,"lang":"fr","type":"article","venue":"La Revue de Médecine Interne","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa Skills and Simulation Centre; Ottawa Hospital; University of Ottawa","funders":"","keywords":"Humanities; Psychology; Philosophy","score_opus":0.03883608267457688,"score_gpt":0.3448154633587417,"score_spread":0.3059793806841648,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2494177323","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95534396,0.0027503667,0.028486023,0.005947251,0.0018529806,0.0004647509,0.00020712893,0.00012155108,0.004825984],"genre_scores_gemma":[0.97845775,0.0048069176,0.009131636,0.0002426294,0.0015375754,0.000046724712,0.00015700619,0.000059434667,0.005560335],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9964929,0.0006619681,0.0011213344,0.00058777275,0.00048232928,0.0006536476],"domain_scores_gemma":[0.9830085,0.014896611,0.00038745496,0.00049963396,0.0007488735,0.00045892008],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0021599263,0.00036462428,0.00072721555,0.00012883967,0.00013895915,0.0000855785,0.00025169775,0.00034815265,0.0012302793],"category_scores_gemma":[0.06086485,0.00026776083,0.00029825448,0.00030519476,0.0010417968,0.00025749428,0.00013084142,0.00028352143,0.00033563096],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00030987928,0.0007568242,0.59289896,0.00013108163,0.00010234375,0.00004419993,0.0007746634,0.000009071687,0.00023185687,0.0005715395,0.0028938097,0.40127578],"study_design_scores_gemma":[0.0058309943,0.0017512984,0.9244332,0.031170806,0.00048510425,0.00006893447,0.00026414677,0.00038716441,0.003637823,0.025874972,0.005747638,0.00034792343],"about_ca_topic_score_codex":0.0011590633,"about_ca_topic_score_gemma":0.00025221592,"teacher_disagreement_score":0.40092784,"about_ca_system_score_codex":0.00047691804,"about_ca_system_score_gemma":0.00019901864,"threshold_uncertainty_score":0.99997747},"labels":[],"label_agreement":null},{"id":"W2497511312","doi":"10.1007/s10459-016-9701-x","title":"What physicians reason about during admission case review","year":2016,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":31,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; Queen's University","funders":"","keywords":"Task (project management); Context (archaeology); Coding (social sciences); Process (computing); Computer science; MEDLINE; Psychology; Cognitive psychology; Medical education; Medicine","score_opus":0.02652541580868318,"score_gpt":0.45395200392489593,"score_spread":0.42742658811621276,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2497511312","genre_codex":"empirical","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.50310045,0.4353696,0.000054302705,0.057155576,0.002546638,0.00070667313,0.000001752917,0.00004868831,0.001016315],"genre_scores_gemma":[0.45044678,0.5426307,0.0010783265,0.0049681417,0.00022471958,0.00005884515,0.0000032606486,0.0000058633173,0.00058341824],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99850166,0.00008674055,0.0004569432,0.00037922652,0.00026769107,0.00030775365],"domain_scores_gemma":[0.997236,0.0019585327,0.00024972414,0.0002397135,0.00007860161,0.0002373903],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0009513717,0.00009629302,0.00025730976,0.00008616095,0.00017036025,0.000019577521,0.00008550727,0.000031633514,0.0000777214],"category_scores_gemma":[0.012908317,0.000060476803,0.00004167726,0.00052724575,0.00016099033,0.0007929572,0.000017632658,0.00008826476,0.000029927558],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000013522949,0.00020302334,0.022569131,0.0007995167,6.9752923e-7,0.000017016797,0.00015374461,0.0000031780173,0.000018655372,0.00024356438,0.0006733257,0.9753046],"study_design_scores_gemma":[0.0016178592,0.001147788,0.13526051,0.75932163,0.000036522582,0.001938748,0.0026376587,0.0000634609,0.00054054026,0.003194962,0.0937826,0.00045769304],"about_ca_topic_score_codex":0.00015810883,"about_ca_topic_score_gemma":0.00009553654,"teacher_disagreement_score":0.97484696,"about_ca_system_score_codex":0.00018726668,"about_ca_system_score_gemma":0.0008881222,"threshold_uncertainty_score":0.9954064},"labels":[],"label_agreement":null},{"id":"W2498525818","doi":"10.1186/s12909-016-0714-1","title":"Validity of a new assessment rubric for a short-answer test of clinical reasoning","year":2016,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"The Wilson Centre; Memorial University of Newfoundland; University Health Network; University of Toronto","funders":"University of Toronto","keywords":"Rubric; Generalizability theory; Certification; Medical education; Test (biology); Psychology; Competence (human resources); Construct validity; Medicine; Applied psychology; Clinical psychology; Psychometrics; Social psychology; Mathematics education","score_opus":0.13314555267024047,"score_gpt":0.500389192385086,"score_spread":0.36724363971484547,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2498525818","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93103474,0.00011908368,0.06021652,0.0047827857,0.0014127626,0.00058339245,0.000013761307,0.00003284857,0.0018041227],"genre_scores_gemma":[0.9565378,0.00019939711,0.039984357,0.000492127,0.0012899039,0.000051482977,0.000035649966,0.00001639633,0.0013928827],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9976243,0.00011421023,0.001129645,0.00029752543,0.0006469602,0.00018739786],"domain_scores_gemma":[0.90466154,0.09203414,0.0005557585,0.00072767516,0.00066416804,0.001356699],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0026448097,0.0001092537,0.00058424385,0.000059088645,0.000019446808,0.0000033767712,0.00012344465,0.0002400013,0.0006809077],"category_scores_gemma":[0.5261185,0.00006855212,0.00028217808,0.0001464819,0.00017275114,0.000034175773,0.000031807376,0.00014747458,0.000008502613],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011231278,0.0020474708,0.760063,0.00010556713,0.000021020101,5.5871385e-7,0.000032733133,9.923556e-8,0.00008540207,0.0008968125,0.027392082,0.20924295],"study_design_scores_gemma":[0.0030014943,0.0012421866,0.9821338,0.0074336026,0.0002818215,0.000015404517,0.0000831625,0.0003863878,0.00047269466,0.0010377128,0.0037999088,0.00011179833],"about_ca_topic_score_codex":0.000087671375,"about_ca_topic_score_gemma":0.000025976977,"teacher_disagreement_score":0.5234737,"about_ca_system_score_codex":0.000046383106,"about_ca_system_score_gemma":0.010967693,"threshold_uncertainty_score":0.9946392},"labels":[],"label_agreement":null},{"id":"W2508591350","doi":"10.1051/pmed/2016024","title":"Questions de planification clinique : un nouvel outil pour évaluer la capacité des étudiants en médecine à identifier les éléments-clés discriminants d’un diagnostic différentiel","year":2016,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval","funders":"","keywords":"Humanities; Philosophy","score_opus":0.05485033759314088,"score_gpt":0.3562906494361205,"score_spread":0.3014403118429796,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2508591350","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9049204,0.012418209,0.052186728,0.024060676,0.002906342,0.0008888262,0.0006934936,0.00025389442,0.0016714188],"genre_scores_gemma":[0.91705173,0.037258662,0.0046095997,0.00051537284,0.0017159036,0.00025102985,0.00011948123,0.00013626761,0.038341958],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9938257,0.0010920087,0.001503816,0.0011242373,0.0008279688,0.0016262891],"domain_scores_gemma":[0.9371718,0.05948742,0.00045269157,0.0011046887,0.0003418562,0.0014415105],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0028107786,0.00068781874,0.0010015777,0.00020833137,0.00049318396,0.00012131923,0.0005561179,0.0009493645,0.0028832548],"category_scores_gemma":[0.17483772,0.0005188266,0.00045196878,0.00037052055,0.0015459522,0.0003227655,0.00028269188,0.0008376133,0.0016752237],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015334533,0.004979919,0.5860571,0.00017569028,0.00032542215,0.0004778831,0.0011297661,0.0000054954544,0.0018473724,0.0021725283,0.012753733,0.38992172],"study_design_scores_gemma":[0.0067463205,0.00039271306,0.94557625,0.014366445,0.0015932068,0.00022042032,0.0011591389,0.00016532153,0.004787879,0.018339109,0.0059341122,0.0007190868],"about_ca_topic_score_codex":0.00059033936,"about_ca_topic_score_gemma":0.00013637959,"teacher_disagreement_score":0.38920265,"about_ca_system_score_codex":0.0005498951,"about_ca_system_score_gemma":0.00092095794,"threshold_uncertainty_score":0.99972636},"labels":[],"label_agreement":null},{"id":"W2508867272","doi":"10.15694/mep.2016.000060","title":"Introducing recent medical graduates as members of Script Concordance Test expert reference panels: what impact?","year":2016,"lang":"en","type":"article","venue":"MedEdPublish","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Concordance; Context (archaeology); Test (biology); Medical education; Burnout; Panel discussion; Cohort; Psychology; Medicine; Family medicine; Clinical psychology; Pathology","score_opus":0.04672677307640451,"score_gpt":0.36204966930698096,"score_spread":0.31532289623057647,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2508867272","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8982604,0.004333078,0.00024928257,0.0872841,0.0021949387,0.00044999318,0.000028544635,0.00022493122,0.0069747115],"genre_scores_gemma":[0.97253215,0.020506928,0.0007579364,0.0038919612,0.0007690615,0.000042341268,0.000052965013,0.00004262266,0.001404043],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.996751,0.00010814878,0.00078468735,0.0006097295,0.0011985247,0.0005479183],"domain_scores_gemma":[0.97679794,0.019830178,0.00035472328,0.0010470891,0.00092555315,0.0010445092],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0013291322,0.00026644222,0.00072664267,0.0001281372,0.00005430412,0.00012022724,0.00035396306,0.00025636255,0.0060115736],"category_scores_gemma":[0.37026617,0.00015429358,0.00016106175,0.00041613373,0.00044837315,0.00063930335,0.00014367675,0.00034915414,0.00012203527],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00033414044,0.00086878653,0.056066357,0.000042090047,0.000116035466,0.00015092427,0.00073892606,6.258283e-7,0.005261134,0.00031955508,0.14190614,0.7941953],"study_design_scores_gemma":[0.040498286,0.013281561,0.2021311,0.11472162,0.0008570649,0.0018888308,0.010704554,0.00068432896,0.17857578,0.009727762,0.4234012,0.0035278914],"about_ca_topic_score_codex":0.00094595365,"about_ca_topic_score_gemma":0.000055183005,"teacher_disagreement_score":0.7906674,"about_ca_system_score_codex":0.00016636754,"about_ca_system_score_gemma":0.0007928416,"threshold_uncertainty_score":0.99489707},"labels":[],"label_agreement":null},{"id":"W2515681744","doi":"10.1016/j.physio.2015.03.3438","title":"The contribution of conceptual frameworks to knowledge translation interventions in physical therapy","year":2015,"lang":"en","type":"article","venue":"Physiotherapy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Université du Québec à Montréal; Université de Montréal; Centre for Interdisciplinary Research in Rehabilitation","funders":"","keywords":"Knowledge translation; Psychological intervention; Clinical Practice; Process (computing); Conceptual framework; Translation (biology); Medicine; Knowledge management; Psychology; Computer science; Physical therapy; Epistemology; Nursing","score_opus":0.09463807707426404,"score_gpt":0.4513297930683854,"score_spread":0.35669171599412136,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2515681744","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99052334,0.0026412397,0.00318844,0.002537971,0.00017828222,0.0004867108,0.0000038888975,0.000025693203,0.0004144628],"genre_scores_gemma":[0.9988881,0.00025750417,0.00018071845,0.00030357976,0.00022663262,0.000046921505,0.000013830805,0.000011579469,0.000071152186],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99919474,0.000093347044,0.00026287118,0.0001377222,0.00016258752,0.00014875545],"domain_scores_gemma":[0.99577004,0.0036722024,0.000060911912,0.00019628201,0.00017947779,0.00012110024],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0003131014,0.00008838782,0.00026121206,0.000037807276,0.000029022665,0.000009029367,0.00007402981,0.000086925866,0.000013889039],"category_scores_gemma":[0.0028886204,0.00005630197,0.00017932749,0.0002637973,0.00015124705,0.000029519591,0.000006809328,0.0002098541,0.000023820114],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.005853239,0.007647521,0.022419356,0.00002297247,0.00026630744,0.00000445702,0.032789197,0.00042050236,0.025272598,0.026174208,0.0060704877,0.87305915],"study_design_scores_gemma":[0.073754944,0.019282952,0.65962404,0.014737102,0.00018921298,0.000011589816,0.0075035593,0.013604025,0.071963646,0.0679864,0.07006208,0.0012804586],"about_ca_topic_score_codex":0.00005428842,"about_ca_topic_score_gemma":0.00004494361,"teacher_disagreement_score":0.8717787,"about_ca_system_score_codex":0.000041306426,"about_ca_system_score_gemma":0.00008308542,"threshold_uncertainty_score":0.3458157},"labels":[],"label_agreement":null},{"id":"W2517719729","doi":"10.3138/jvme.0216-040r","title":"Development of a New Scale to Measure Ambiguity Tolerance in Veterinary Students","year":2017,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":24,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"National Institute for Health and Care Research","keywords":"Ambiguity; Scale (ratio); Exploratory factor analysis; Reliability (semiconductor); Psychology; Ambiguity tolerance; Role conflict; Veterinary medicine; Medicine; Psychometrics; Clinical psychology; Social psychology; Computer science; Geography","score_opus":0.13657946339967642,"score_gpt":0.4757791502915499,"score_spread":0.33919968689187346,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2517719729","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99402326,0.00029210027,0.0001718462,0.0038805418,0.0011932208,0.00015760162,5.799084e-7,0.0000040307573,0.000276802],"genre_scores_gemma":[0.9870523,0.00007866433,0.011619442,0.0006234344,0.00040866452,0.000006511205,0.0000019689671,0.000009910106,0.00019905745],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.997705,0.00005532215,0.0008512793,0.00014889958,0.0010743181,0.00016520005],"domain_scores_gemma":[0.9978809,0.00032573115,0.00047418947,0.00032176665,0.00023284537,0.00076456566],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0014123091,0.00010906142,0.00043972625,0.00012960052,0.00006368083,0.00002302723,0.00042330305,0.00011816774,0.00015156718],"category_scores_gemma":[0.029932378,0.00008990752,0.00009270474,0.000084219486,0.000063616375,0.00010455802,0.00012805992,0.0003148395,0.000012172085],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002470136,0.003981845,0.15375663,0.00012109588,0.000057085963,0.0003163827,0.0035280897,0.0000026163852,0.003922957,0.000003365633,0.0036289422,0.82821083],"study_design_scores_gemma":[0.0023178428,0.0019501441,0.9723816,0.009973085,0.00003032162,0.0007025119,0.00020691304,0.000011258961,0.00030422103,0.0000292511,0.0120028565,0.000090031426],"about_ca_topic_score_codex":0.00007817777,"about_ca_topic_score_gemma":0.000017792187,"teacher_disagreement_score":0.8281208,"about_ca_system_score_codex":0.00013834334,"about_ca_system_score_gemma":0.0032556572,"threshold_uncertainty_score":0.97823894},"labels":[],"label_agreement":null},{"id":"W2518246739","doi":"","title":"Le test de concordance de script en formation initiale : avantages et inconvénients du test. Revue de la littérature","year":2016,"lang":"fr","type":"article","venue":"HAL (Le Centre pour la Communication Scientifique Directe)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Humanities; Political science; Philosophy","score_opus":0.009380441846921874,"score_gpt":0.26715714774724697,"score_spread":0.2577767059003251,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2518246739","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2895415,0.013689489,0.20328279,0.28748927,0.001052586,0.000986627,0.00055860874,0.00047103816,0.20292808],"genre_scores_gemma":[0.94694585,0.011581175,0.016884914,0.0014748675,0.00024664335,0.00005687686,0.00014608563,0.000053093932,0.022610486],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9917783,0.0061523337,0.00062976725,0.0004699069,0.00029162946,0.00067804166],"domain_scores_gemma":[0.8980446,0.09831224,0.00047956788,0.0012295267,0.001304825,0.0006292334],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.010985417,0.00028172933,0.00039610488,0.00011058819,0.0003005401,0.00022237249,0.0005810251,0.00044935106,0.00018704798],"category_scores_gemma":[0.24931327,0.00026440492,0.00017082957,0.00039726685,0.0006273395,0.0003665835,0.00028654656,0.0006129198,0.00016538629],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010282206,0.0048430716,0.5239598,0.00047868543,0.000100218225,0.00020993176,0.020114433,0.00001887497,0.034732647,0.15716344,0.021406673,0.23686936],"study_design_scores_gemma":[0.009767472,0.000028260418,0.4103998,0.086372316,0.00039720236,0.0012128586,0.0018313692,0.012489064,0.17439663,0.012715027,0.28923205,0.0011579691],"about_ca_topic_score_codex":0.0012181553,"about_ca_topic_score_gemma":0.0010892259,"teacher_disagreement_score":0.65740436,"about_ca_system_score_codex":0.00045356018,"about_ca_system_score_gemma":0.0013746801,"threshold_uncertainty_score":0.9999808},"labels":[],"label_agreement":null},{"id":"W2521312816","doi":"10.1007/s10459-016-9711-8","title":"Inter-rater variability as mutual disagreement: identifying raters’ divergent points of view","year":2016,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":43,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia; University of Northern British Columbia","funders":"National Board of Medical Examiners","keywords":"Psychology; Variation (astronomy); Salient; Variance (accounting); Inter-rater reliability; Cognitive psychology; Point (geometry); Applied psychology; Cognition; Social psychology; Computer science; Rating scale; Artificial intelligence; Developmental psychology; Mathematics","score_opus":0.039062740839399604,"score_gpt":0.4441012497172203,"score_spread":0.4050385088778207,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2521312816","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97978956,0.0030747477,0.0016733997,0.010514548,0.002527499,0.00052282127,0.000004915165,0.000014779195,0.0018777131],"genre_scores_gemma":[0.99219465,0.0045093023,0.0017398386,0.0011803441,0.00008447683,0.000035538465,0.000004929683,0.0000037121433,0.00024722933],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982868,0.00016140162,0.000638695,0.00035977448,0.0002955319,0.00025782365],"domain_scores_gemma":[0.99590474,0.0033094021,0.00027610647,0.00023547233,0.00009568496,0.00017859269],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0021548432,0.00008897278,0.00024392622,0.00010307492,0.000076010154,0.000010431042,0.00013844889,0.00003067942,0.0004561337],"category_scores_gemma":[0.014691321,0.00005494567,0.000041994146,0.00035417682,0.00033533893,0.00033260573,0.000048958023,0.00006211362,0.000061587874],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000399447,0.0004276934,0.7161919,0.00023243779,0.0000026575678,3.4652e-7,0.000738293,0.0000029343157,0.000054697975,0.0021705744,0.00014257434,0.27999595],"study_design_scores_gemma":[0.0021508485,0.0023244456,0.90343493,0.028751658,0.000042015203,0.00002453711,0.002375803,0.00010901969,0.0026808837,0.041272357,0.016462885,0.0003706341],"about_ca_topic_score_codex":0.00036972918,"about_ca_topic_score_gemma":0.00007758413,"teacher_disagreement_score":0.27962533,"about_ca_system_score_codex":0.00018141003,"about_ca_system_score_gemma":0.00075779506,"threshold_uncertainty_score":0.99360836},"labels":[],"label_agreement":null},{"id":"W2521651092","doi":"10.1007/s10459-016-9712-7","title":"A neural marker of medical visual expertise: implications for training","year":2016,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Pulmonologists; Medicine; Audiology; Cognitive psychology; Psychology; Intensive care medicine","score_opus":0.06766844591366125,"score_gpt":0.521980551627644,"score_spread":0.45431210571398273,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2521651092","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7582455,0.010708486,0.009319189,0.2182873,0.001477935,0.0009406455,0.000009840298,0.000037275484,0.00097380846],"genre_scores_gemma":[0.99003977,0.0029405367,0.0045474516,0.0020215171,0.00017435534,0.00019706207,0.0000051153397,0.0000045650177,0.000069630485],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9987147,0.00004118998,0.00047768568,0.00024604518,0.00027971665,0.00024068337],"domain_scores_gemma":[0.9935409,0.0058701094,0.00017369058,0.00011291509,0.00008159307,0.00022080673],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0013293637,0.000059365062,0.00019203436,0.000103046215,0.00007714062,0.0000048530296,0.00012159445,0.000046367386,0.00005854083],"category_scores_gemma":[0.026456889,0.000037822814,0.00003824835,0.0003172457,0.00030477732,0.00015411171,0.000013274314,0.000049156824,0.0000015088897],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00002478815,0.00018609987,0.11332929,0.000044397555,8.2583193e-7,5.2802708e-8,0.00030556694,0.0000029999776,0.000030377065,0.0041255117,0.0002602057,0.8816899],"study_design_scores_gemma":[0.002395221,0.0020320022,0.9472114,0.00917712,0.000012011442,0.000056121196,0.0021148715,0.0028071501,0.00013597668,0.018091157,0.01575456,0.00021244513],"about_ca_topic_score_codex":0.00008646163,"about_ca_topic_score_gemma":0.000057221405,"teacher_disagreement_score":0.8814774,"about_ca_system_score_codex":0.000073562886,"about_ca_system_score_gemma":0.001999272,"threshold_uncertainty_score":0.9817437},"labels":[],"label_agreement":null},{"id":"W2522677081","doi":"10.1136/bmjqs-2016-005817","title":"The evolution of morbidity and mortality conferences","year":2016,"lang":"en","type":"letter","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Accreditation; Patient safety; Medicine; Multidisciplinary approach; Quality management; Medical education; Quality (philosophy); Best practice; Health care; Nursing; Operations management; Engineering","score_opus":0.09623840150231208,"score_gpt":0.42946734366663736,"score_spread":0.3332289421643253,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2522677081","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.04607872,0.0012895138,0.0007211702,0.9384662,0.0009780242,0.0010263011,0.00065917557,0.00007439681,0.010706537],"genre_scores_gemma":[0.80220675,0.0039720535,0.00015208551,0.17539819,0.009936788,0.00009798822,0.00059817574,0.000058241916,0.0075797434],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99636966,0.0006284833,0.0014721968,0.00044460135,0.00074096717,0.00034408984],"domain_scores_gemma":[0.963614,0.034302928,0.0007000209,0.00090899813,0.00034493115,0.00012913509],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004184971,0.00024873138,0.00085119705,0.000030952753,0.0001410368,0.00001991368,0.00017992736,0.0006575458,0.00011059993],"category_scores_gemma":[0.049860746,0.00013281127,0.00026447093,0.000079458965,0.00088867825,0.000030996693,0.00011000068,0.0008394067,0.00002714487],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00031572097,0.00006376266,0.13835986,0.00037594687,0.00031985505,0.000045165412,0.00006208761,8.6446384e-8,0.000019171166,0.004947758,0.8396463,0.015844326],"study_design_scores_gemma":[0.0011600648,0.00015360699,0.67722255,0.0025534136,0.00035633237,0.000017837549,0.000123343,0.0000075361095,0.000022913111,0.008649336,0.30947474,0.0002583259],"about_ca_topic_score_codex":0.003969262,"about_ca_topic_score_gemma":0.00016992429,"teacher_disagreement_score":0.76306796,"about_ca_system_score_codex":0.00014020986,"about_ca_system_score_gemma":0.0010854084,"threshold_uncertainty_score":0.9581427},"labels":[],"label_agreement":null},{"id":"W2524190225","doi":"10.1016/j.cjca.2016.07.162","title":"EXPLAINING CAUSAL LINKS AMONG RISK FACTORS, CARDIOVASCULAR PATHOPHYSIOLOGY, SYMPTOMS AND HEALTH BEHAVIOURS: A RANDOMIZED CONTROLLED TRIAL OF PATIENT EDUCATION STRATEGIES TO ENHANCE 'DEEP' VS. 'SURFACE' LEVEL OF KNOWLEDGE","year":2016,"lang":"en","type":"article","venue":"Canadian Journal of Cardiology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Medicine; Randomized controlled trial; Disease; Rehabilitation; Patient education; Physical therapy; Health education; Intensive care medicine; Public health; Family medicine; Nursing; Internal medicine","score_opus":0.025297621368265508,"score_gpt":0.32438399883671726,"score_spread":0.29908637746845174,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2524190225","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.987382,0.005596585,0.0036812914,0.00038712847,0.0013595289,0.0013388601,0.00007589521,0.000004037695,0.00017465076],"genre_scores_gemma":[0.99825823,0.0009291512,0.00020794653,0.000058304722,0.00046418695,0.00003266073,0.000010109199,0.000017541748,0.000021849084],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9965972,0.0012439125,0.0013885815,0.0002399795,0.0001852435,0.00034509128],"domain_scores_gemma":[0.99140406,0.005787805,0.0010011037,0.0002920495,0.00066180574,0.0008531733],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0027060693,0.00020987332,0.0037162602,0.00027035698,0.000063358944,0.000008605686,0.00010639233,0.00030818232,0.000009137056],"category_scores_gemma":[0.020100672,0.00012908137,0.0010052213,0.00010649377,0.00051938446,0.0000566959,0.000017523746,0.00035856958,6.932384e-7],"study_design_candidate":"randomized_trial","study_design_consensus":"randomized_trial","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.61175025,0.00025060316,0.15201193,0.00011753524,0.010052144,0.00007881965,0.008455422,0.00254303,0.00041111378,0.00045533036,0.0035442524,0.21032959],"study_design_scores_gemma":[0.8576505,0.0112315295,0.119386025,0.004122599,0.0031670777,0.00010104958,0.002302805,0.000013311103,0.0004750178,0.00065810943,0.0005097203,0.00038223327],"about_ca_topic_score_codex":0.007849197,"about_ca_topic_score_gemma":0.0012541042,"teacher_disagreement_score":0.24590029,"about_ca_system_score_codex":0.00019216104,"about_ca_system_score_gemma":0.0047511235,"threshold_uncertainty_score":0.9987576},"labels":[],"label_agreement":null},{"id":"W2524972078","doi":"10.1111/ap.12206","title":"Does Size Really Matter? Contributions to the Debate on Short Versus Long Neuropsychology Assessments","year":2016,"lang":"en","type":"article","venue":"Australian Psychologist","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Royal Ottawa Mental Health Centre","funders":"","keywords":"Neuropsychology; Psychology; Selection (genetic algorithm); Interpretation (philosophy); Variation (astronomy); Test (biology); Engineering ethics; Applied psychology; Computer science; Cognition; Artificial intelligence; Engineering; Psychiatry","score_opus":0.06040745736353241,"score_gpt":0.45739489079904894,"score_spread":0.3969874334355165,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2524972078","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7248735,0.000005073804,0.00044395108,0.2565535,0.0038642415,0.00060983974,0.00021681485,0.00012715001,0.013305972],"genre_scores_gemma":[0.96706146,0.0000379687,0.00019721924,0.01251404,0.00047148747,0.00007983338,0.00002197683,0.000027440285,0.019588562],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9978997,0.00016305744,0.00043340516,0.0006506254,0.00027856312,0.00057465135],"domain_scores_gemma":[0.9916874,0.0066475715,0.000081053,0.0010491264,0.00013305302,0.0004018033],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00044791587,0.00023869066,0.0003308456,0.00005262099,0.00012683574,0.00004101566,0.00033138902,0.00019303607,0.0012505784],"category_scores_gemma":[0.013081352,0.00009873532,0.00015696383,0.00021774168,0.00030837167,0.000036431047,0.000048083755,0.00032608007,0.0032354456],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0036562646,0.000979681,0.35746193,0.000005693697,0.00021915449,0.000694524,0.000029663648,0.0000035485493,0.0015788787,0.00071210414,0.59850866,0.036149904],"study_design_scores_gemma":[0.0030961887,0.0017787965,0.9628407,0.00037771702,0.000104471335,0.00006657938,0.000019350207,2.3450282e-7,0.0002738813,0.00016298078,0.031102704,0.00017642892],"about_ca_topic_score_codex":0.000060669994,"about_ca_topic_score_gemma":0.000058108428,"teacher_disagreement_score":0.60537875,"about_ca_system_score_codex":0.000094673946,"about_ca_system_score_gemma":0.000041560517,"threshold_uncertainty_score":0.9996624},"labels":[],"label_agreement":null},{"id":"W2530421077","doi":"10.1016/j.jamcollsurg.2016.06.389","title":"Training Surgical Residents to Use a Framework to Promote Shared Decision-Making for Patients with Poor Baseline Prognosis Experiencing Surgical Emergencies","year":2016,"lang":"en","type":"article","venue":"Journal of the American College of Surgeons","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Medicine; Baseline (sea); General surgery; Gerontology; Law","score_opus":0.03336230966079773,"score_gpt":0.3407767317342463,"score_spread":0.30741442207344855,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2530421077","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9901789,0.000023446983,0.0029131449,0.005444868,0.0003034256,0.000831506,0.00027755322,0.00001502965,0.000012137071],"genre_scores_gemma":[0.95019925,0.000026220789,0.049045116,0.00038572092,0.0001597332,0.00004224678,0.0000010793201,0.000040978513,0.000099642646],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.996823,0.00014285125,0.0011716221,0.0003196355,0.0010662475,0.00047664388],"domain_scores_gemma":[0.96630985,0.03037559,0.0010047483,0.00040834866,0.0012411341,0.0006603195],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0012260427,0.00023676742,0.0010612849,0.000269338,0.00014572004,0.000022839178,0.00033151417,0.0000677527,0.00010842552],"category_scores_gemma":[0.10741051,0.00011783795,0.00048654844,0.001101574,0.00018341054,0.00013123378,0.00014033774,0.00021138831,0.0000026735952],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.018538166,0.0013018448,0.8838768,0.000022631168,0.00035265432,0.0004904862,0.001486528,0.0000659789,0.00017030863,0.000055321896,0.009479053,0.08416025],"study_design_scores_gemma":[0.00745104,0.010111,0.9131223,0.046386514,0.00058139814,0.0005572104,0.0024262166,0.00008106928,0.0007333083,0.00028679613,0.017625239,0.0006378937],"about_ca_topic_score_codex":0.00004712932,"about_ca_topic_score_gemma":0.0000252411,"teacher_disagreement_score":0.10618447,"about_ca_system_score_codex":0.00012629078,"about_ca_system_score_gemma":0.0004373768,"threshold_uncertainty_score":0.90010816},"labels":[],"label_agreement":null},{"id":"W2530536149","doi":"10.1111/jep.12641","title":"Phronesis: practical wisdom the role of professional practice knowledge in the clinical reasoning of Bobath instructors","year":2016,"lang":"en","type":"article","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":22,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Rehabilitation Institute; University of Toronto","funders":"","keywords":"Phronesis; Psychology; Process (computing); Embodied cognition; Perception; Medical education; Medicine; Artificial intelligence; Computer science","score_opus":0.10937584674844418,"score_gpt":0.54890136740859,"score_spread":0.4395255206601458,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2530536149","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.92594147,0.0021621871,0.00023369053,0.06145037,0.0019935928,0.0009552393,0.0000047096287,0.0000067065994,0.0072520147],"genre_scores_gemma":[0.98551095,0.0017661139,0.009953836,0.0016076835,0.0010769537,0.000020579566,0.0000018600125,0.000021178108,0.000040823703],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9770087,0.01325549,0.0064457287,0.00039948098,0.0025599208,0.00033068075],"domain_scores_gemma":[0.37312704,0.6123567,0.0083432235,0.00092529727,0.0049458607,0.0003018894],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.10652679,0.00020539353,0.0010396928,0.00018526005,0.00006372533,0.000021387956,0.00037380657,0.0004318047,0.0001603944],"category_scores_gemma":[0.8789496,0.000094611845,0.0005030822,0.0006973696,0.00073000113,0.0008890234,0.00012229712,0.002228299,0.00002212806],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.023043891,0.02771398,0.5599145,0.00006931354,0.00078829995,0.00023944888,0.0030121845,0.000057526126,0.0001601368,0.010817333,0.009445156,0.3647382],"study_design_scores_gemma":[0.01663267,0.0043672686,0.9124439,0.007288005,0.002811821,0.0015984665,0.01837199,0.0015343563,0.00009596963,0.0047641126,0.02986578,0.0002256982],"about_ca_topic_score_codex":0.000053118973,"about_ca_topic_score_gemma":0.00003030445,"teacher_disagreement_score":0.7724228,"about_ca_system_score_codex":0.00014246085,"about_ca_system_score_gemma":0.0036279345,"threshold_uncertainty_score":0.9680974},"labels":[],"label_agreement":null},{"id":"W2545846385","doi":"10.1097/acm.0000000000001421","title":"The Causes of Errors in Clinical Reasoning: Cognitive Biases, Knowledge Deficits, and Dual Process Thinking","year":2016,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":555,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Dual process theory (moral psychology); Cognitive psychology; Cognition; Cognitive bias; Process (computing); Psychology; Dual (grammatical number); Representation (politics); Working memory; Component (thermodynamics); Computer science; Debiasing; Cognitive science","score_opus":0.09262940850429977,"score_gpt":0.4358231361784074,"score_spread":0.3431937276741076,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2545846385","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.32494643,0.039248377,0.00002023634,0.62913716,0.0015020061,0.0012703269,0.000059213457,0.00009887585,0.0037173552],"genre_scores_gemma":[0.51192456,0.058684967,0.000039564155,0.39842594,0.023195399,0.00016504293,0.00043976848,0.00021851926,0.006906225],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9946419,0.00051983714,0.0022955604,0.0008737892,0.0009183145,0.0007506087],"domain_scores_gemma":[0.8579361,0.14003865,0.0009388205,0.00040224358,0.00040572422,0.0002784599],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0047627883,0.00051925675,0.0018228949,0.00027587914,0.00012257848,0.000007423544,0.00033502502,0.0022336524,0.00006717918],"category_scores_gemma":[0.39622775,0.00027068006,0.0001544689,0.0004046532,0.0022390343,0.000051891962,0.00017895689,0.0063750846,0.000025132771],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005827862,0.00005492118,0.09286664,0.0004774532,0.00035046582,0.00088851986,0.003385901,1.1046252e-7,0.000009345051,0.00025222133,0.8730379,0.028093763],"study_design_scores_gemma":[0.024320085,0.0042391573,0.06577067,0.5786916,0.004683249,0.0013071428,0.006216683,0.0000943648,0.00019901559,0.0073660417,0.30561218,0.0014997788],"about_ca_topic_score_codex":0.00011454157,"about_ca_topic_score_gemma":0.000020880294,"teacher_disagreement_score":0.57821417,"about_ca_system_score_codex":0.000065853616,"about_ca_system_score_gemma":0.00072923576,"threshold_uncertainty_score":0.99997455},"labels":[],"label_agreement":null},{"id":"W2545952627","doi":"10.1007/s10459-016-9725-2","title":"Measuring physician cognitive load: validity evidence for a physiologic and a psychometric tool","year":2016,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":92,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University; Kingston General Hospital","funders":"","keywords":"Cognitive load; Cognition; Construct validity; Construct (python library); Psychometrics; Clinical psychology; Psychology; Medicine; Scale (ratio); Cognitive psychology; Psychiatry; Computer science","score_opus":0.21536521667046993,"score_gpt":0.49873746150631026,"score_spread":0.28337224483584034,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2545952627","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9757233,0.015336885,0.0017572393,0.0052308855,0.00069643086,0.0008673486,0.0000045114784,0.000025079564,0.00035829406],"genre_scores_gemma":[0.98208326,0.012234266,0.0030550293,0.0021582802,0.00021999303,0.00019036858,0.0000014269813,0.000003818642,0.0000535559],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.998696,0.00006937221,0.00029805733,0.00041262145,0.0002521876,0.00027175384],"domain_scores_gemma":[0.98298216,0.016442271,0.0001938429,0.0001071729,0.00015073305,0.00012383073],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001552152,0.000085594475,0.00021512818,0.00013712778,0.00013934738,0.000014336477,0.000071645896,0.000029899062,0.000005721716],"category_scores_gemma":[0.06636277,0.000053596024,0.000032018852,0.00071028667,0.0002630997,0.00038185503,0.000014598791,0.000057661506,0.0000060142206],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009907177,0.0002096438,0.15851314,0.00014112605,0.0000012470368,8.383529e-8,0.00013758948,0.000002114484,0.00005221109,0.0005009391,0.00007276096,0.84027004],"study_design_scores_gemma":[0.0013669537,0.0023898846,0.94704616,0.021046093,0.000017459244,0.0000070327465,0.0005204387,0.0001620536,0.0002526416,0.025731249,0.0012718911,0.0001881551],"about_ca_topic_score_codex":0.000080335245,"about_ca_topic_score_gemma":0.00002430125,"teacher_disagreement_score":0.84008193,"about_ca_system_score_codex":0.00016795145,"about_ca_system_score_gemma":0.001112201,"threshold_uncertainty_score":0.9415017},"labels":[],"label_agreement":null},{"id":"W2548421507","doi":"10.1186/s12911-016-0377-1","title":"Cognitive biases associated with medical decisions: a systematic review","year":2016,"lang":"en","type":"review","venue":"BMC Medical Informatics and Decision Making","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":961,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Institute for Clinical Evaluative Sciences","funders":"University of Toronto; Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung; Heart and Stroke Foundation of Canada","keywords":"Cognitive bias; Overconfidence effect; Cognition; MEDLINE; Clinical psychology; Cochrane Library; Medicine; Meta-analysis; Personality; Publication bias; Psychology; Systematic review; Psychiatry; Social psychology; Pathology","score_opus":0.1397034793305594,"score_gpt":0.46194588544176884,"score_spread":0.32224240611120947,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2548421507","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000020378617,0.94026685,0.05613848,0.000027148693,0.0001844747,0.0024400805,0.00007507439,0.000101083664,0.00074645964],"genre_scores_gemma":[0.00001401312,0.9898516,0.0055277133,0.0037744266,0.00014472684,0.00028460225,0.00027709265,0.00006525196,0.00006059804],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9893911,0.00039752584,0.0044445046,0.00044084946,0.0047576204,0.0005684284],"domain_scores_gemma":[0.24480706,0.74958533,0.002233699,0.00085144525,0.0006055193,0.0019169619],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.006305424,0.0007178206,0.006364531,0.00036593425,0.00015884281,0.00010861205,0.00045822852,0.001170509,0.0013321486],"category_scores_gemma":[0.93642926,0.0003274071,0.00075321144,0.00071321166,0.00038692294,0.00011257582,0.00044630654,0.0010552731,0.00036911285],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000036829417,0.0001254119,0.000004972946,0.17279387,0.00029574975,0.0002789497,0.000015169722,1.7782853e-8,3.460525e-11,0.00004862229,0.0021116068,0.8242888],"study_design_scores_gemma":[0.0017905507,0.00029189093,0.0000033505473,0.9855096,0.0055761463,0.0021055527,0.00005700712,0.0012326882,3.6496393e-9,0.00012018062,0.0028870208,0.00042601593],"about_ca_topic_score_codex":0.0000012039364,"about_ca_topic_score_gemma":0.00000925084,"teacher_disagreement_score":0.9301238,"about_ca_system_score_codex":0.000115815266,"about_ca_system_score_gemma":0.0033093926,"threshold_uncertainty_score":0.9999178},"labels":[],"label_agreement":null},{"id":"W2549718520","doi":"10.1002/aet2.10010","title":"Cognitive Debiasing Strategies for the Emergency Department","year":2016,"lang":"en","type":"article","venue":"AEM Education and Training","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":36,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Debiasing; Cognitive bias; Confirmation bias; Intuition; Cognition; Pessimism; Psychology; Cognitive psychology; Emergency department; Computer science; Social psychology; Cognitive science; Epistemology; Psychiatry","score_opus":0.08723981673305986,"score_gpt":0.41223259523354905,"score_spread":0.3249927785004892,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2549718520","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9767771,0.0011321758,0.0073322225,0.009190702,0.0006123609,0.00034759552,0.0000074182963,0.000029813951,0.0045706076],"genre_scores_gemma":[0.9975006,0.0002819157,0.00049554696,0.0006445193,0.00024361283,0.00009774085,0.000010925725,0.0000069050343,0.000718211],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99958247,0.000011303546,0.00012472444,0.0001160952,0.00005794214,0.00010746788],"domain_scores_gemma":[0.9950728,0.0046679787,0.00003932652,0.0000605871,0.00008418015,0.00007511896],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0001653272,0.000054448123,0.00008032418,0.000016645834,0.00010367336,0.000014743827,0.00001745851,0.000024738605,0.00010018447],"category_scores_gemma":[0.0058473167,0.000027976217,0.000037453006,0.00003709671,0.00004365107,0.000039301456,0.0000048778616,0.000029796209,0.0000066403563],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000378919,0.000091859634,0.00752947,0.000006769535,0.00003782544,3.1409772e-7,0.0018112026,7.244035e-8,0.00005406876,0.0062339376,0.0013039635,0.98289263],"study_design_scores_gemma":[0.0036164918,0.0005788171,0.90001357,0.0027737198,0.00060103805,0.00008673629,0.053039104,0.00008322624,0.00034710034,0.020890094,0.017711142,0.00025894705],"about_ca_topic_score_codex":0.000009554978,"about_ca_topic_score_gemma":0.0000075548273,"teacher_disagreement_score":0.98263365,"about_ca_system_score_codex":0.00000916849,"about_ca_system_score_gemma":0.00027578673,"threshold_uncertainty_score":0.7000207},"labels":[],"label_agreement":null},{"id":"W2550035039","doi":"10.1016/j.jclinepi.2016.10.004","title":"Four hundred or more participants needed for stable contingency table estimates of clinical prediction rule performance","year":2016,"lang":"en","type":"article","venue":"Journal of Clinical Epidemiology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Contingency table; Table (database); Statistics; Medicine; Mathematics; Computer science; Data mining","score_opus":0.4199366815559962,"score_gpt":0.5455118565174738,"score_spread":0.12557517496147758,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2550035039","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97421724,0.00035161688,0.01277639,0.009366013,0.0026453424,0.000394328,0.00006385687,0.0000297139,0.0001554836],"genre_scores_gemma":[0.9567777,0.0035263242,0.034375705,0.0021300756,0.0023676949,0.000021200272,0.000009084582,0.000034505036,0.0007577038],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.98698825,0.0010947069,0.010429631,0.00042813402,0.00031585217,0.00074345234],"domain_scores_gemma":[0.5920532,0.3979705,0.00687865,0.0005667768,0.001429207,0.0011016917],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.036951676,0.0002281294,0.0046632676,0.00010094136,0.000068520436,0.000004969862,0.00026696234,0.0007014792,0.0002364472],"category_scores_gemma":[0.8405612,0.00012455744,0.0011125408,0.00016118148,0.0006639695,0.00016968796,0.000072675306,0.000691189,0.000019826966],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.010235404,0.0008837653,0.8720736,0.000091232156,0.0003968881,0.00002206398,0.000015593261,0.000047697962,0.00008709266,0.000100237405,0.014223888,0.10182254],"study_design_scores_gemma":[0.014534998,0.011975579,0.95547503,0.0037450532,0.0008877161,0.0001484055,0.00007886036,0.0065412233,0.00033524472,0.0027362,0.00338135,0.00016031678],"about_ca_topic_score_codex":0.000029818237,"about_ca_topic_score_gemma":0.000004745141,"teacher_disagreement_score":0.80360955,"about_ca_system_score_codex":0.000032437492,"about_ca_system_score_gemma":0.0010003769,"threshold_uncertainty_score":0.9916609},"labels":[],"label_agreement":null},{"id":"W2550977815","doi":"10.1007/s10459-016-9736-z","title":"The influence of first impressions on subsequent ratings within an OSCE station","year":2016,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Psychology; Rating scale; Quality (philosophy); Point (geometry); Applied psychology; Social psychology; Scale (ratio); Medical education; Medicine; Developmental psychology","score_opus":0.027231195978238806,"score_gpt":0.43830848110638543,"score_spread":0.4110772851281466,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2550977815","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98716223,0.0009034335,0.00009675841,0.010857286,0.0004660366,0.00030541126,0.0000026382065,0.000012734741,0.00019344872],"genre_scores_gemma":[0.9939945,0.0030401107,0.0019586585,0.00079959625,0.00005037719,0.000056282573,0.000002832264,0.0000032500704,0.00009444185],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9987298,0.00008608582,0.00046724177,0.00022816194,0.00030281587,0.00018589229],"domain_scores_gemma":[0.9920992,0.0070460434,0.0003705484,0.00021639948,0.00012733879,0.000140423],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001446301,0.000062656516,0.00011818767,0.000065761174,0.0002609904,0.000011668382,0.00013002765,0.000024862373,0.000006592888],"category_scores_gemma":[0.023377337,0.000031171094,0.000015407571,0.00033514702,0.0003751367,0.00030580524,0.000010324242,0.00007439998,0.000003898341],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013076387,0.00065821706,0.6319908,0.00009577878,0.0000015424333,3.2361547e-7,0.0024254983,0.0045115515,0.0004064447,0.015035396,0.00019921808,0.3445445],"study_design_scores_gemma":[0.0005820755,0.002157858,0.9721177,0.00923202,0.0000044628323,0.000005179143,0.0016557436,0.0005219004,0.0010155725,0.010140373,0.0024597594,0.00010733986],"about_ca_topic_score_codex":0.0006497129,"about_ca_topic_score_gemma":0.0008569135,"teacher_disagreement_score":0.34443715,"about_ca_system_score_codex":0.00010515432,"about_ca_system_score_gemma":0.0009907058,"threshold_uncertainty_score":0.98484915},"labels":[],"label_agreement":null},{"id":"W2553776533","doi":"10.36834/cmej.36690","title":"Learning-by-Concordance (LbC): introducing undergraduate students to the complexity and uncertainty of clinical practice","year":2016,"lang":"en","type":"article","venue":"Canadian Medical Education Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":53,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Université de Montréal; University of Toronto; Université du Québec à Montréal","funders":"","keywords":"Concordance; Medical education; Thematic analysis; Critical thinking; Psychology; Qualitative research; Mathematics education; Computer science; Medicine; Internal medicine","score_opus":0.03344151846575948,"score_gpt":0.422022675573961,"score_spread":0.3885811571082015,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2553776533","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.43590155,0.00037496779,0.00038639203,0.5611671,0.0013550394,0.00016640493,0.000005411103,0.000007192322,0.00063592975],"genre_scores_gemma":[0.94999933,0.00086440466,0.000837989,0.04560746,0.0014639278,0.000006885955,0.0000052458972,0.000012238609,0.0012025215],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9974402,0.00042108752,0.00076738145,0.00024596215,0.00083758426,0.00028777163],"domain_scores_gemma":[0.9784708,0.010647849,0.0004320129,0.0003392568,0.0008674697,0.009242603],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0042901984,0.000104480176,0.00032929995,0.000077453224,0.000183967,0.000040675823,0.00024283264,0.00012221556,0.0013495248],"category_scores_gemma":[0.66565996,0.000058161764,0.000075586046,0.0001756261,0.00045629148,0.00006453799,0.000037793638,0.00080074405,0.000046852063],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001284863,0.00030595646,0.09839814,0.000008315419,0.000090349706,0.000028936432,0.0003415925,0.0000034102943,0.0000059410054,0.0006520227,0.38233283,0.517704],"study_design_scores_gemma":[0.0027438921,0.0009808737,0.23043373,0.0030558587,0.00017546258,0.0012258836,0.0015709858,0.000084383864,0.00000970373,0.0011221127,0.7584136,0.00018347322],"about_ca_topic_score_codex":0.007275708,"about_ca_topic_score_gemma":0.005523797,"teacher_disagreement_score":0.66136974,"about_ca_system_score_codex":0.00016751196,"about_ca_system_score_gemma":0.014951526,"threshold_uncertainty_score":0.9995634},"labels":[],"label_agreement":null},{"id":"W2555063473","doi":"10.36834/cmej.36777","title":"Systems 1 and 2 thinking processes and cognitive reflection testing in medical students","year":2016,"lang":"en","type":"article","venue":"Canadian Medical Education Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":94,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Intuition; Cognition; Metacognition; Psychology; Test (biology); Mathematics education; Clinical Practice; Cognitive psychology; Computer science; Medicine; Cognitive science; Psychiatry","score_opus":0.031076934763741442,"score_gpt":0.3970032903146228,"score_spread":0.3659263555508813,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2555063473","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9595961,0.002526173,0.00010097084,0.03554054,0.00078318664,0.00017195025,0.0000019572856,0.00001348717,0.001265625],"genre_scores_gemma":[0.98921347,0.00063118263,0.00008292539,0.009059603,0.0007860544,0.000014570655,0.000002516,0.000010252842,0.00019940929],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9981481,0.00008965727,0.00035701692,0.00018379402,0.0009755081,0.0002459212],"domain_scores_gemma":[0.9812246,0.007833274,0.00015479661,0.00008844181,0.0006589745,0.01003992],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0018230653,0.000087113316,0.00019264828,0.00022093912,0.00013234212,0.00009057115,0.000090002315,0.00019673356,0.0007744385],"category_scores_gemma":[0.57934874,0.00005731306,0.00001233129,0.00023488699,0.00015452954,0.0000804912,0.00001985545,0.00044962968,0.0000071211725],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000026498177,0.00014362423,0.62379235,0.00006251448,0.0000270056,0.00031194344,0.0005641058,3.404707e-8,0.0000025704787,0.000087790795,0.0034992087,0.37148237],"study_design_scores_gemma":[0.0062308204,0.000577207,0.87115437,0.08848814,0.00016269193,0.021202521,0.0039259107,0.00015147417,0.000007970834,0.0014112233,0.006355057,0.00033263772],"about_ca_topic_score_codex":0.0043652905,"about_ca_topic_score_gemma":0.0075298855,"teacher_disagreement_score":0.5775257,"about_ca_system_score_codex":0.00016421596,"about_ca_system_score_gemma":0.024479799,"threshold_uncertainty_score":0.9810505},"labels":[],"label_agreement":null},{"id":"W2557391392","doi":"10.1097/ceh.0000000000000121","title":"Assessing Clinical Faculty Understanding of Statistical Terms Used to Measure Treatment Effects and Their Application to Teaching","year":2016,"lang":"en","type":"article","venue":"Journal of Continuing Education in the Health Professions","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"Dalhousie University","keywords":"Measure (data warehouse); Psychology; Medical education; Statistical analysis; Medicine; Medical physics; Statistics; Computer science; Mathematics; Data mining","score_opus":0.15141652157530214,"score_gpt":0.5319494056849279,"score_spread":0.3805328841096257,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2557391392","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.83127886,0.00008025412,0.13019589,0.037167843,0.0004383323,0.0007700306,0.0000045527863,0.000006287299,0.000057981164],"genre_scores_gemma":[0.9927859,0.000032366195,0.005805361,0.0010405856,0.00024038822,0.000026204241,0.0000030685455,0.000010039828,0.000056087105],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9976703,0.00084117823,0.0009292335,0.0001487437,0.00022836527,0.00018218225],"domain_scores_gemma":[0.97920126,0.019571524,0.00056386396,0.00018916372,0.00013124413,0.00034292098],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0041887397,0.00010177739,0.00047295683,0.00015019333,0.000112850576,0.00002115011,0.000082478226,0.00007193576,0.000002370128],"category_scores_gemma":[0.026958138,0.00004571595,0.00006560148,0.00012530644,0.00004348612,0.000074046795,0.000018941322,0.000294777,0.0000015633232],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000179997,0.0009779681,0.18740703,0.00007527825,0.000025206264,0.0000015776328,0.01797369,0.0000010896173,0.0008844266,0.001644893,0.002018056,0.7888108],"study_design_scores_gemma":[0.0028603533,0.002281043,0.934159,0.030797303,0.00008033363,0.00005368647,0.025869172,0.000023353758,0.00011904404,0.002748271,0.00090230856,0.00010616207],"about_ca_topic_score_codex":0.00003321391,"about_ca_topic_score_gemma":0.000019598498,"teacher_disagreement_score":0.78870463,"about_ca_system_score_codex":0.00031700707,"about_ca_system_score_gemma":0.0008372435,"threshold_uncertainty_score":0.9812382},"labels":[],"label_agreement":null},{"id":"W2560014499","doi":"10.1007/s40037-016-0314-0","title":"From problem solving to problem definition: scrutinizing the complex nature of clinical practice","year":2016,"lang":"en","type":"article","venue":"Perspectives on Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":22,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia; Western University","funders":"Canadian Institutes of Health Research","keywords":"Management science; Computer science; Clinical Practice; Epistemology; Mathematics education; Psychology; Medicine; Philosophy; Nursing","score_opus":0.05753997358081065,"score_gpt":0.4430659722756378,"score_spread":0.3855259986948272,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2560014499","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.34617168,0.0013198871,0.0026794388,0.56597376,0.00136503,0.0013004,0.000028824565,0.00013429277,0.081026696],"genre_scores_gemma":[0.93457365,0.00043730793,0.049078364,0.013231894,0.0023658436,0.000063213796,0.000027887123,0.000023257719,0.00019861248],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99712515,0.0003330765,0.00077685295,0.0005342546,0.0009945855,0.00023608354],"domain_scores_gemma":[0.94218135,0.05444778,0.0005352938,0.0007637088,0.0011962131,0.00087563356],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0021488995,0.00015661516,0.0004126225,0.000072237555,0.000102108585,0.000020741798,0.00021810955,0.00031619292,0.0011077357],"category_scores_gemma":[0.38374382,0.000083372026,0.00016699467,0.00028401113,0.00032303843,0.0000840354,0.00006756497,0.00076908583,0.0001613141],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0022874686,0.011834087,0.017805118,0.0000692324,0.00045209104,0.000024909445,0.023573805,0.0000015815893,0.0004928325,0.10872304,0.15365133,0.6810845],"study_design_scores_gemma":[0.009503667,0.004684198,0.66101736,0.033981785,0.0010537581,0.00020675153,0.112901285,0.00012358975,0.00044524297,0.048312772,0.12685579,0.00091381365],"about_ca_topic_score_codex":0.00018822018,"about_ca_topic_score_gemma":0.000009918134,"teacher_disagreement_score":0.6801707,"about_ca_system_score_codex":0.000109971275,"about_ca_system_score_gemma":0.0016335996,"threshold_uncertainty_score":0.9998054},"labels":[],"label_agreement":null},{"id":"W2560797262","doi":"10.1051/pmed/2016031","title":"Optimiser l’usage des cartes conceptuelles dans l’apprentissage par problèmes (APP) dans le cursus préclinique : le point de vue des enseignants","year":2016,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université TÉLUQ; Université du Québec à Montréal; Université de Montréal","funders":"Université de Montréal","keywords":"Humanities; Philosophy; Political science","score_opus":0.059957699086143006,"score_gpt":0.32520936411279516,"score_spread":0.26525166502665215,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2560797262","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.866378,0.047736738,0.048017718,0.010024167,0.0013793544,0.0013039489,0.00042815186,0.000561251,0.02417064],"genre_scores_gemma":[0.86059177,0.015769405,0.014640724,0.00043532075,0.0011556938,0.00029172614,0.00006135344,0.00020069104,0.10685329],"study_design_codex":"design_other","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.9934103,0.0009961793,0.0014154015,0.0012882921,0.0006780992,0.002211733],"domain_scores_gemma":[0.97913206,0.017268,0.00039619132,0.0011109678,0.0003134376,0.0017793565],"candidate_categories":["metaresearch","metaepi_narrow","sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0021074864,0.00080873125,0.0013127815,0.000105826846,0.00076093356,0.00013563955,0.0006858643,0.0008233402,0.0018003734],"category_scores_gemma":[0.043231644,0.000625935,0.0006477185,0.00037581645,0.00717782,0.00039769075,0.00038054102,0.0007007578,0.0004823527],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006404613,0.006222675,0.24231333,0.00091350987,0.0008618919,0.0023508158,0.03511041,0.00013038513,0.04402498,0.0028225416,0.047106113,0.61750287],"study_design_scores_gemma":[0.03424122,0.0057410756,0.28018525,0.08631483,0.0030442078,0.0013536983,0.07323699,0.0017480389,0.37854406,0.07442231,0.056459554,0.0047087567],"about_ca_topic_score_codex":0.001464456,"about_ca_topic_score_gemma":0.0004351816,"teacher_disagreement_score":0.61279416,"about_ca_system_score_codex":0.00041055348,"about_ca_system_score_gemma":0.001944792,"threshold_uncertainty_score":0.9996192},"labels":[],"label_agreement":null},{"id":"W2566299261","doi":"10.1007/s10459-016-9743-0","title":"Examining the effect of self-explanation on cognitive integration of basic and clinical sciences in novices","year":2016,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; University of Toronto; Canada Research Chairs; University Health Network","funders":"","keywords":"Cognition; Task (project management); Basic science; Psychology; Test (biology); Cognitive psychology; Computer science","score_opus":0.04491802409993078,"score_gpt":0.4721473392987002,"score_spread":0.4272293151987694,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2566299261","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9947649,0.0016720133,0.00006949248,0.001829519,0.00036109707,0.0004072689,0.0000015571436,0.000006041121,0.0008880856],"genre_scores_gemma":[0.9949122,0.0040975413,0.0005906593,0.00031418502,0.00004972036,0.000025516754,0.0000018288755,0.000001937096,0.0000063842062],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99842435,0.00031265576,0.0005820158,0.0002686363,0.00027689352,0.00013544725],"domain_scores_gemma":[0.96540344,0.03398911,0.00041358802,0.00008113181,0.000065987326,0.000046721827],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0054926462,0.000066960805,0.00023392617,0.00020403927,0.00006871865,0.0000063762436,0.00008278069,0.00003665193,0.0000049646715],"category_scores_gemma":[0.02741867,0.00003269923,0.00001848823,0.00060983986,0.0007014238,0.00022262857,0.000012193567,0.000087808316,0.0000011096336],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005113725,0.0001020592,0.65418947,0.000046888585,8.4054227e-7,5.9053455e-8,0.0002646268,0.0000064994288,0.000010206756,0.00052862806,0.0000044185804,0.34479517],"study_design_scores_gemma":[0.00088713056,0.005792703,0.9805579,0.009507631,0.000009428045,0.000002529529,0.0014371446,0.00039141145,0.0007403865,0.00059540116,0.000032509866,0.000045776247],"about_ca_topic_score_codex":0.00025450924,"about_ca_topic_score_gemma":0.0002813555,"teacher_disagreement_score":0.3447494,"about_ca_system_score_codex":0.00004635292,"about_ca_system_score_gemma":0.0005034814,"threshold_uncertainty_score":0.9807738},"labels":[],"label_agreement":null},{"id":"W2582439002","doi":"10.3138/jvme.0916-150r1","title":"Comparing Tolerance of Ambiguity in Veterinary and Medical Students","year":2017,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Ambiguity; Ambiguity tolerance; Population; Scale (ratio); Medicine; Veterinary medicine; Medical education; Curriculum; Diversity (politics); Psychology; Computer science; Pedagogy; Environmental health","score_opus":0.11738570796674724,"score_gpt":0.48574053716064886,"score_spread":0.3683548291939016,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2582439002","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9933334,0.00087075325,0.00002503392,0.0042107576,0.0010579617,0.00008471497,4.552933e-7,0.0000036138772,0.00041330606],"genre_scores_gemma":[0.99729097,0.0012958361,0.00041458514,0.0005231048,0.00042737985,0.0000032898242,0.0000018823949,0.000007955504,0.000035013836],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9974965,0.00009681809,0.0007836137,0.00013644363,0.0013414454,0.00014519972],"domain_scores_gemma":[0.9973565,0.001070276,0.0005397913,0.00027161968,0.0001583505,0.0006034906],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0021118212,0.00009806616,0.0005221598,0.00010364226,0.000056872228,0.000021214673,0.00042311483,0.00016028646,0.0001779906],"category_scores_gemma":[0.061103728,0.00007901631,0.00008056854,0.000047876656,0.00028012512,0.00012641794,0.00019753845,0.0004978051,0.000002171275],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009844541,0.0024342975,0.8689585,0.00015616912,0.000038145696,0.0007897727,0.0003116528,6.5454054e-7,0.00028204045,0.000021618767,0.000847187,0.12517555],"study_design_scores_gemma":[0.0028538203,0.0020509239,0.98099494,0.009551177,0.000035299716,0.002890443,0.00010004034,0.00024628048,0.000024102703,0.000091779395,0.001096752,0.000064412765],"about_ca_topic_score_codex":0.00011801448,"about_ca_topic_score_gemma":0.000006458787,"teacher_disagreement_score":0.12511113,"about_ca_system_score_codex":0.000054090193,"about_ca_system_score_gemma":0.0010541998,"threshold_uncertainty_score":0.946805},"labels":[],"label_agreement":null},{"id":"W2583546893","doi":"10.1007/s10459-017-9759-0","title":"Enhancing technical skill learning through interleaved mixed-model observational practice","year":2017,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University Medical Centre; McMaster University","funders":"","keywords":"Observational study; Observational learning; Psychomotor learning; Psychology; Motor learning; Task (project management); Dreyfus model of skill acquisition; Motor skill; Cognitive psychology; Computer science; Applied psychology; Mathematics education; Developmental psychology; Cognition; Mathematics; Statistics; Engineering","score_opus":0.06855834781292847,"score_gpt":0.5012789416311738,"score_spread":0.43272059381824535,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2583546893","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6814487,0.016324412,0.045756463,0.15987994,0.007056084,0.0018808297,0.0000048410525,0.0002665141,0.08738224],"genre_scores_gemma":[0.8713137,0.0034029742,0.120193236,0.0042773117,0.0002468277,0.00006228679,0.000014759787,0.0000079938645,0.00048090873],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9981162,0.000091859736,0.00058472937,0.00044183328,0.00042721373,0.00033818546],"domain_scores_gemma":[0.99296117,0.005690049,0.0006487249,0.0003387521,0.00020497614,0.00015630286],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0022759025,0.000111593785,0.00025170343,0.000073205745,0.00071916514,0.00008303043,0.00026958925,0.00007047862,0.000017810096],"category_scores_gemma":[0.13431491,0.00009745078,0.000043578075,0.00021280405,0.00039651632,0.0014173163,0.00006660222,0.00039486942,0.000019735593],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00023858434,0.0022659407,0.48572424,0.00041363668,0.000009804837,0.0000068102595,0.004377665,0.011981888,0.00021308922,0.07454237,0.0016349931,0.418591],"study_design_scores_gemma":[0.0025347897,0.0027934117,0.7795361,0.021651147,0.00006256971,0.0002445126,0.013247302,0.03197324,0.00070249924,0.06793148,0.07847426,0.00084866973],"about_ca_topic_score_codex":0.0006329696,"about_ca_topic_score_gemma":0.00038453628,"teacher_disagreement_score":0.4177423,"about_ca_system_score_codex":0.0002258461,"about_ca_system_score_gemma":0.0022411235,"threshold_uncertainty_score":0.87297714},"labels":[],"label_agreement":null},{"id":"W2584820851","doi":"10.1111/medu.13221","title":"Should learners reason one step at a time? A randomised trial of two diagnostic scheme designs","year":2017,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Western University","funders":"","keywords":"Medical diagnosis; Auscultation; Scheme (mathematics); Cognition; Medicine; Computer science; Algorithm; Artificial intelligence; Mathematics; Radiology","score_opus":0.06677853953739667,"score_gpt":0.41452785225158706,"score_spread":0.3477493127141904,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2584820851","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97142994,0.00033269895,0.00026157906,0.017293068,0.0009844218,0.001245724,0.0000031328489,0.000050311264,0.00839913],"genre_scores_gemma":[0.9895935,0.00029980816,0.0022795282,0.0014232027,0.0012143021,0.00022079745,0.00011100554,0.000030485606,0.0048273453],"study_design_codex":"design_other","study_design_gemma":"randomized_trial","domain_scores_codex":[0.99770355,0.00015352367,0.00058294146,0.00033679808,0.0009579976,0.00026518287],"domain_scores_gemma":[0.9837829,0.0141832335,0.00041838852,0.00083662843,0.00018641638,0.00059247453],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0015989032,0.00015857037,0.0006079227,0.000069175825,0.00016070095,0.000027608812,0.00024355765,0.00024404589,0.0025896386],"category_scores_gemma":[0.5794691,0.00012960267,0.00020120847,0.00007820829,0.00037029758,0.00005482014,0.00008138051,0.00033000065,0.00027155888],"study_design_candidate":"randomized_trial","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.18172415,0.017932948,0.06124436,0.00028160054,0.00052517484,0.00008119375,0.0009713066,0.0000027089186,0.0017855699,0.0011292181,0.19666156,0.53766024],"study_design_scores_gemma":[0.90646803,0.0034480582,0.056053355,0.013084936,0.0015437495,0.00010234623,0.0001493019,0.0032831088,0.0025463172,0.00059874373,0.012076498,0.0006455617],"about_ca_topic_score_codex":0.00041997235,"about_ca_topic_score_gemma":0.000022106275,"teacher_disagreement_score":0.7247439,"about_ca_system_score_codex":0.00008965707,"about_ca_system_score_gemma":0.0019206085,"threshold_uncertainty_score":0.9983221},"labels":[],"label_agreement":null},{"id":"W2587663074","doi":"10.1186/s12909-017-0871-x","title":"Immersive high fidelity simulation of critically ill patients to study cognitive errors: a pilot study","year":2017,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":38,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"Flinders University","keywords":"Teamwork; Internship; Cognition; Likert scale; Fidelity; Medicine; Odds; Graduate medical education; Health care; Clinical psychology; Psychology; Rating scale; Cognitive skill; Medical education; Psychiatry; Developmental psychology; Internal medicine; Computer science","score_opus":0.06415350339241237,"score_gpt":0.44672165706264316,"score_spread":0.3825681536702308,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2587663074","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9935457,0.000011253591,0.001505673,0.0016356105,0.00077731203,0.0021273727,0.000012065458,0.000017969705,0.00036703938],"genre_scores_gemma":[0.9959094,0.0000022495524,0.00040048172,0.0031036143,0.0002867465,0.00013366471,0.00006718398,0.000015974305,0.000080668266],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.997413,0.00022869918,0.0006515927,0.0004232182,0.0010753673,0.00020816262],"domain_scores_gemma":[0.98012835,0.012554186,0.00036224222,0.0008279223,0.005104956,0.0010223694],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00094011414,0.00015124903,0.0004226335,0.000083832725,0.00016957181,0.00002534314,0.00019757844,0.000075636286,0.0005355054],"category_scores_gemma":[0.68715954,0.00012718912,0.000058721846,0.000104876926,0.00014126684,0.00007964633,0.00011789097,0.00020698833,0.00007636701],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011589256,0.04722349,0.9213951,0.000035841516,0.00007703954,0.0000057734123,0.0019181059,0.000011105664,0.0000035100988,0.000025605777,0.0005074302,0.027638063],"study_design_scores_gemma":[0.005276146,0.010061334,0.9757305,0.0011426258,0.0003447317,7.1014347e-7,0.0068649394,0.00029910298,0.000020598565,0.00013464739,0.0000067113165,0.0001180029],"about_ca_topic_score_codex":0.0014511005,"about_ca_topic_score_gemma":0.0002366364,"teacher_disagreement_score":0.68621945,"about_ca_system_score_codex":0.00006685349,"about_ca_system_score_gemma":0.0013580994,"threshold_uncertainty_score":0.58634084},"labels":[],"label_agreement":null},{"id":"W2588982045","doi":"10.1097/acm.0000000000001559","title":"Transforming Medical Assessment: Integrating Uncertainty Into the Evaluation of Clinical Reasoning in Medical Education","year":2017,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":77,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Royal College of Physicians and Surgeons of Canada; University of Calgary","funders":"","keywords":"Context (archaeology); MEDLINE; Engineering ethics; Management science; Field (mathematics); Psychology; Medical education; Medicine; Knowledge management; Computer science; Political science","score_opus":0.11096823427097736,"score_gpt":0.5498533590944679,"score_spread":0.4388851248234905,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2588982045","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9038992,0.0016327975,0.00059612346,0.07409291,0.0013847123,0.0007264638,6.3463114e-7,0.00002810725,0.017639045],"genre_scores_gemma":[0.9911907,0.002621648,0.00029049895,0.0037059968,0.0018992095,0.00009110673,0.000049003826,0.000020990215,0.00013088733],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9921196,0.00073917885,0.0023405673,0.0004943315,0.0039556483,0.00035068148],"domain_scores_gemma":[0.9819213,0.014882947,0.00095775176,0.00090706215,0.00059779617,0.0007330999],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.040514726,0.00021986052,0.00095910253,0.00013917279,0.00023696008,0.0000103844295,0.0007667702,0.00082403526,0.0012966938],"category_scores_gemma":[0.6309383,0.00012691843,0.00014493744,0.00021963377,0.0012649976,0.00013352475,0.00011819724,0.003040216,0.0000071108097],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000097753145,0.00010951595,0.28516755,0.00004268121,0.000047300633,0.000015426556,0.0025820786,0.0000043233285,0.000025349736,0.0055760723,0.001563549,0.7047684],"study_design_scores_gemma":[0.018597355,0.00091811933,0.72730553,0.0695576,0.0011945989,0.0002547971,0.010483853,0.15549472,0.00008281093,0.011992773,0.0037416043,0.00037623857],"about_ca_topic_score_codex":0.003249082,"about_ca_topic_score_gemma":0.0005374723,"teacher_disagreement_score":0.7043922,"about_ca_system_score_codex":0.00016323474,"about_ca_system_score_gemma":0.006682941,"threshold_uncertainty_score":0.99961627},"labels":[],"label_agreement":null},{"id":"W2589185236","doi":"","title":"Theories of Expertise and Measures of Competence: Cognitive and Interactional Perspectives","year":2008,"lang":"en","type":"article","venue":"eScholarship (California Digital Library)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McGill University","funders":"","keywords":"Psychology; Competence (human resources); Cognition; Medical education; Social psychology; Medicine","score_opus":0.028033659747398534,"score_gpt":0.27644133289294764,"score_spread":0.24840767314554912,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2589185236","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9901795,0.0023040087,0.00006619203,0.00034054427,0.00003110722,0.00014089301,0.0006090368,0.000037425074,0.0062913196],"genre_scores_gemma":[0.9987345,0.00046006558,0.00035639224,0.00013554977,0.000052103507,0.000006186746,0.00007079616,0.0000172007,0.00016723429],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9990497,0.00004022078,0.00030952232,0.00022800118,0.0002465604,0.00012596071],"domain_scores_gemma":[0.9958683,0.0036124485,0.000111014226,0.00009952988,0.00012070267,0.00018797332],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00008261388,0.0001329644,0.00034246582,0.00008167802,0.000051487208,0.00004151437,0.00004812223,0.000071141665,0.00012685025],"category_scores_gemma":[0.014547555,0.00010388931,0.00008121324,0.00012141133,0.0008035226,0.0006009763,0.00008081858,0.00019474061,0.0000125628285],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012106922,0.0004996052,0.9817728,0.000052983243,0.00012321562,0.000045680626,0.000941498,1.596066e-7,0.00034634955,0.0037625027,0.00016213385,0.011082404],"study_design_scores_gemma":[0.0035238347,0.0011237791,0.95964944,0.0034058865,0.00012033193,0.00039785114,0.0039773826,0.00003241613,0.015203177,0.010986485,0.0012466796,0.00033271598],"about_ca_topic_score_codex":0.0000039523156,"about_ca_topic_score_gemma":1.8562316e-7,"teacher_disagreement_score":0.022123314,"about_ca_system_score_codex":0.000006598683,"about_ca_system_score_gemma":0.000094899005,"threshold_uncertainty_score":0.9937533},"labels":[],"label_agreement":null},{"id":"W2590714892","doi":"10.1080/10833196.2017.1289647","title":"Factors influencing physical therapists’ clinical reasoning: qualitative systematic review and meta-synthesis","year":2017,"lang":"en","type":"article","venue":"Physical Therapy Reviews","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":28,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Biopsychosocial model; Psychosocial; Critical appraisal; Medicine; Context (archaeology); Identification (biology); Qualitative research; Systematic review; MEDLINE; Applied psychology; Psychology; Alternative medicine; Psychiatry; Pathology","score_opus":0.32293666965846435,"score_gpt":0.5143538260193206,"score_spread":0.1914171563608562,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2590714892","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.68756944,0.30142012,0.00003209376,0.004125396,0.00013295078,0.0051777447,0.000032430475,0.00014210303,0.0013677137],"genre_scores_gemma":[0.8068512,0.18461238,0.00034828932,0.0051502674,0.0005612462,0.0019194947,0.00001133017,0.00008969435,0.00045608825],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99514955,0.0018651997,0.0013400771,0.0007650359,0.0004732187,0.00040692333],"domain_scores_gemma":[0.97512543,0.02093575,0.0013384963,0.0018983511,0.00017881228,0.00052315183],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0035040488,0.00059963006,0.0065349583,0.000032496537,0.0003335697,0.00012837739,0.00044291565,0.00011442069,0.00011721559],"category_scores_gemma":[0.11041184,0.00030316715,0.0020430782,0.00012155975,0.00053892867,0.00019402274,0.00010910338,0.00057721144,0.00013136958],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012246385,0.021013739,0.017412923,0.62257314,0.062278636,0.00016358969,0.059996836,8.229229e-7,0.00095994805,0.01803211,0.015050166,0.18129347],"study_design_scores_gemma":[0.008020349,0.0065934416,0.059017673,0.6650086,0.2260411,0.00007587999,0.0017469444,0.00029556052,0.0015845281,0.012900223,0.014197815,0.004517902],"about_ca_topic_score_codex":0.000030110075,"about_ca_topic_score_gemma":0.0000010211435,"teacher_disagreement_score":0.17677557,"about_ca_system_score_codex":0.000040403713,"about_ca_system_score_gemma":0.0000726518,"threshold_uncertainty_score":0.99994206},"labels":[],"label_agreement":null},{"id":"W2590799813","doi":"10.1016/j.neurad.2016.12.010","title":"Overdiagnosis, incidental findings and Socrates’ three sieves","year":2017,"lang":"en","type":"editorial","venue":"Journal of Neuroradiology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institut Universitaire de Gériatrie de Montréal; Centre Integre de Sante et de Services Sociaux de Laval; Hôpital Notre-Dame","funders":"","keywords":"Overdiagnosis; SOCRATES; Medicine; Philosophy; Internal medicine; Epistemology","score_opus":0.021851614162881627,"score_gpt":0.34778771883130527,"score_spread":0.32593610466842365,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2590799813","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.37793967,0.0023216477,0.000006803162,0.0017820105,0.61744887,0.00013401933,0.000049855615,0.0000133088,0.0003038257],"genre_scores_gemma":[0.08994245,0.010476982,0.00021050834,0.00045953508,0.89828885,0.000004872621,0.000057229496,0.00007481649,0.00048477232],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99776244,0.000112953145,0.00087869423,0.0003449987,0.00056498154,0.00033590107],"domain_scores_gemma":[0.9696773,0.027918696,0.0012043536,0.0003908285,0.0004042286,0.00040453795],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0007512443,0.00032153152,0.0015818899,0.00021017712,0.00012030281,0.00007552968,0.00034444474,0.0010623254,0.000054952434],"category_scores_gemma":[0.18242696,0.00024633572,0.00037769062,0.00004358821,0.00042013387,0.00007918673,0.00015753777,0.0022308123,0.000018697887],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00036067513,0.00010152887,0.057083,0.000042793145,0.00018779983,0.0024001894,0.000023506218,4.637299e-7,0.00008650472,0.000008797635,0.9383205,0.0013842268],"study_design_scores_gemma":[0.006369282,0.007235345,0.33695635,0.0042698164,0.0015411561,0.0066075847,0.000017177596,0.000009866034,0.00005356536,0.0015930634,0.63490564,0.0004411559],"about_ca_topic_score_codex":0.000022896575,"about_ca_topic_score_gemma":0.000011292677,"teacher_disagreement_score":0.30341488,"about_ca_system_score_codex":0.000080666774,"about_ca_system_score_gemma":0.0007595352,"threshold_uncertainty_score":0.99999887},"labels":[],"label_agreement":null},{"id":"W2591350882","doi":"10.1515/dx-2016-0043","title":"Five things to know about diagnostic error","year":2017,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Sinai Health System; University Health Network; University of Toronto","funders":"National Institute of Arthritis and Musculoskeletal and Skin Diseases","keywords":"Medical diagnosis; Diagnostic test; Need to know; Diagnostic accuracy; Root cause analysis; Medicine; Root cause; Psychology; Intensive care medicine; Computer science; Engineering; Pathology; Forensic engineering; Pediatrics; Operations management; Computer security","score_opus":0.025735293458198954,"score_gpt":0.3592916808582509,"score_spread":0.33355638740005195,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2591350882","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9395513,0.0010753456,0.00007077563,0.042325247,0.00107393,0.00079224946,0.000047425194,0.0001908688,0.0148728825],"genre_scores_gemma":[0.9821294,0.0013453253,0.0014085808,0.010623392,0.00062273693,0.0005047346,0.000028398625,0.000050066756,0.0032873682],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9981331,0.000034600875,0.00037809653,0.0005523348,0.00039869116,0.00050317927],"domain_scores_gemma":[0.9769099,0.020593036,0.00018391371,0.0013268763,0.00020291723,0.0007834058],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.000365847,0.00025208262,0.0005426498,0.000096671676,0.00042932865,0.00017959723,0.00042264425,0.0001818033,0.0007142297],"category_scores_gemma":[0.33138663,0.0002152406,0.0002505622,0.00010167945,0.0001626956,0.00016968064,0.00024454173,0.0003009791,0.0024902436],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006880535,0.00057575776,0.7007232,0.00003000235,0.00007196379,0.00036816392,0.0005993758,0.000008518303,0.000018253286,0.0005302747,0.22084957,0.076156124],"study_design_scores_gemma":[0.0015145534,0.000533369,0.91883105,0.0049001053,0.00025811742,0.00002070695,0.00008300804,0.000053310054,0.0012028657,0.00033478148,0.071952626,0.0003155331],"about_ca_topic_score_codex":0.00087726937,"about_ca_topic_score_gemma":0.000061029452,"teacher_disagreement_score":0.33102077,"about_ca_system_score_codex":0.00007026366,"about_ca_system_score_gemma":0.00010171063,"threshold_uncertainty_score":0.9982864},"labels":[],"label_agreement":null},{"id":"W2591356326","doi":"10.1016/j.amjmed.2017.01.045","title":"How Expert Clinicians Intuitively Recognize a Medical Diagnosis","year":2017,"lang":"en","type":"review","venue":"The American Journal of Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":114,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Experiential knowledge; Diagnostic test; Medical diagnosis; Focus (optics); Process (computing); Expert system; Test (biology); Artificial intelligence; Epistemology; Computer science; Pathology; Pediatrics","score_opus":0.20080716380146102,"score_gpt":0.49406034116621156,"score_spread":0.29325317736475054,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2591356326","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00009842727,0.89324653,0.00010697427,0.10405042,0.0014555426,0.0004530165,0.000013646586,0.00002406892,0.00055137504],"genre_scores_gemma":[0.0002865566,0.9825853,0.00027455925,0.005553424,0.01017773,0.000055146273,0.000017740293,0.00010590879,0.0009436403],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9936212,0.00097136316,0.0023350671,0.00044034174,0.0020799683,0.00055208977],"domain_scores_gemma":[0.93221474,0.056490865,0.00773649,0.0013832616,0.0007198993,0.0014547749],"candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.005938036,0.00065643055,0.00791707,0.0003724058,0.00018258161,0.000056636713,0.0015238047,0.0003303895,0.0004556123],"category_scores_gemma":[0.33625516,0.00029609274,0.0016482973,0.00045395715,0.004339438,0.00007249213,0.00018851494,0.002935078,0.000051933243],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018629915,0.00021193249,0.000109769455,0.00038888742,0.00094658276,0.00160833,0.00016936449,7.511266e-9,3.352375e-8,0.000013605799,0.0623589,0.9340063],"study_design_scores_gemma":[0.0017075606,0.0046884986,0.000171143,0.21326685,0.0040832213,0.0046918057,0.00040701745,9.0616703e-7,1.3807183e-7,0.000071335155,0.7706907,0.00022086372],"about_ca_topic_score_codex":0.00016952425,"about_ca_topic_score_gemma":0.000007014335,"teacher_disagreement_score":0.93378544,"about_ca_system_score_codex":0.0001651866,"about_ca_system_score_gemma":0.0021207216,"threshold_uncertainty_score":0.9999491},"labels":[],"label_agreement":null},{"id":"W2591739012","doi":"10.1080/14739879.2006.11864048","title":"The development of reasoning skills and expertise in primary care","year":2006,"lang":"en","type":"article","venue":"Education for Primary Care","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Smiths Detection (Canada)","funders":"","keywords":"Context (archaeology); Cognition; Psychology; Medical education; Cognitive skill; Professional development; Principal (computer security); Computer science; Medicine; Pedagogy","score_opus":0.008773081349516454,"score_gpt":0.30052210277245234,"score_spread":0.2917490214229359,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2591739012","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97022307,0.015227083,0.00008521078,0.0002885743,0.0003254897,0.0005673735,0.000009187117,0.00001950569,0.013254483],"genre_scores_gemma":[0.98528904,0.000087438624,0.012350554,0.00030174069,0.00016367219,0.00018728635,0.00079807604,0.000017571414,0.0008046241],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99910116,0.000019710336,0.00037352633,0.00019387396,0.00015586973,0.00015587812],"domain_scores_gemma":[0.99824023,0.001173788,0.000097158656,0.00019106228,0.00023257032,0.00006521444],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00014635512,0.00009737099,0.00019851045,0.000054932974,0.000109697205,0.000014784705,0.000056069322,0.000067241745,0.000004111186],"category_scores_gemma":[0.0015712245,0.000073765885,0.00005061632,0.00009599456,0.000057576035,0.000032761644,0.000028406466,0.00007870445,0.0000017433371],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008438384,0.00032237763,0.19219325,0.0004410527,0.000011451002,7.437883e-7,0.008481453,0.0000030114165,0.00021787488,0.00022466242,0.002344144,0.7956756],"study_design_scores_gemma":[0.0010828989,0.00007717323,0.9830368,0.0019424368,0.00003093661,0.0000050283093,0.004048241,0.0000022877086,0.0015335758,0.00006989901,0.008061753,0.000108924665],"about_ca_topic_score_codex":0.000050644638,"about_ca_topic_score_gemma":0.00006354695,"teacher_disagreement_score":0.7955667,"about_ca_system_score_codex":0.0004271819,"about_ca_system_score_gemma":0.0015165412,"threshold_uncertainty_score":0.30080855},"labels":[],"label_agreement":null},{"id":"W2592733038","doi":"10.1007/s13341-017-0729-1","title":"Comment les médecins urgentistes raisonnent-ils ? Synthèse des principaux résultats d’une recherche qualitative multicentrique et multidisciplinaire sur la prise de décision en médecine d’urgence","year":2017,"lang":"fr","type":"article","venue":"Annales françaises de médecine d’urgence","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Université de Montréal; Université de Sherbrooke","funders":"","keywords":"Humanities; Philosophy","score_opus":0.15520410340939741,"score_gpt":0.4608920885890881,"score_spread":0.30568798517969065,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2592733038","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8451969,0.043123696,0.030765232,0.07509036,0.001938458,0.0020179464,0.00063706696,0.0003027223,0.0009276207],"genre_scores_gemma":[0.9290663,0.0395261,0.024304599,0.0015841842,0.0009841414,0.00032990405,0.0002059405,0.00026101415,0.0037377896],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.98444986,0.0056329006,0.0028588078,0.0022362252,0.0015809241,0.0032413076],"domain_scores_gemma":[0.8515358,0.13845886,0.0019187887,0.003241273,0.0016154824,0.0032298095],"candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity","insufficient_payload"],"consensus_categories":["metaepi_narrow","sts","research_integrity"],"category_scores_codex":[0.016628137,0.0016598037,0.002452182,0.00040801844,0.0021779921,0.000550433,0.0020847537,0.001642812,0.0011588829],"category_scores_gemma":[0.3520102,0.0015580731,0.0011309449,0.00090716884,0.0032250215,0.0011346138,0.0013423626,0.002916371,0.0005815989],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0038451536,0.01713512,0.6479603,0.0022606126,0.0012621733,0.0039208787,0.085009,0.001568396,0.0058243973,0.014462419,0.019459508,0.19729203],"study_design_scores_gemma":[0.015790317,0.0017657158,0.7884128,0.04690259,0.0017578274,0.0005663284,0.016559653,0.012698915,0.006525813,0.010715625,0.09560412,0.002700284],"about_ca_topic_score_codex":0.020547163,"about_ca_topic_score_gemma":0.0044577457,"teacher_disagreement_score":0.33538204,"about_ca_system_score_codex":0.0012180837,"about_ca_system_score_gemma":0.002309708,"threshold_uncertainty_score":0.9997542},"labels":[],"label_agreement":null},{"id":"W2594094256","doi":"10.1108/ijem-04-2016-0068","title":"Questioning clerkship: applying Popper’s evolutionary analysis of learning to medical student training","year":2017,"lang":"en","type":"article","venue":"International Journal of Educational Management","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Ottawa; Western University","funders":"","keywords":"Originality; Positivism; Value (mathematics); Process (computing); Psychology; Sociology; Pedagogy; Mathematics education; Epistemology; Medical education; Computer science; Medicine; Creativity; Social psychology","score_opus":0.041425693331280974,"score_gpt":0.4476590554873327,"score_spread":0.4062333621560517,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2594094256","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9289177,0.00027325994,0.002996487,0.052241754,0.0023775718,0.0002668586,0.000008047761,0.000011004031,0.012907316],"genre_scores_gemma":[0.9909994,0.00015227334,0.0063243913,0.00070382154,0.00071631867,0.000021783682,0.00003773578,0.000011028135,0.0010332688],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99690354,0.000052421245,0.0008463899,0.00019841448,0.0018395953,0.0001596183],"domain_scores_gemma":[0.9956849,0.0020030292,0.00078993273,0.00022879937,0.00089204195,0.0004012485],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011881513,0.00011726134,0.00039555022,0.00080054795,0.00015557467,0.00008419484,0.00064480084,0.00005978531,0.0012346598],"category_scores_gemma":[0.02805604,0.000107098844,0.00032976267,0.00019407566,0.00008876708,0.0001389118,0.00020589547,0.00028667058,0.000018371486],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00033426512,0.0012512775,0.86230457,0.00002153665,0.008099233,0.00015944582,0.0012774244,0.009607277,0.000036238172,0.029096888,0.0021743304,0.08563752],"study_design_scores_gemma":[0.0009387445,0.00016063829,0.9899752,0.0019315946,0.0010499184,0.00007457549,0.0014369874,0.001083033,0.00000767624,0.000682476,0.0025592833,0.000099841],"about_ca_topic_score_codex":0.0000508306,"about_ca_topic_score_gemma":0.000008534833,"teacher_disagreement_score":0.12767066,"about_ca_system_score_codex":0.0002389949,"about_ca_system_score_gemma":0.00040009615,"threshold_uncertainty_score":0.9996784},"labels":[],"label_agreement":null},{"id":"W2599801844","doi":"10.1097/acm.0000000000001587","title":"Clinical Reasoning and Threshold Concepts","year":2017,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of British Columbia","funders":"","keywords":"Context (archaeology); Action (physics); Affordance; Value (mathematics); Psychology; Cognitive psychology; Computer science","score_opus":0.09691729339596887,"score_gpt":0.46204457496984075,"score_spread":0.36512728157387186,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2599801844","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.016242474,0.014391809,0.000034699748,0.9387143,0.0028753278,0.0006454635,0.000017544526,0.00017695449,0.026901389],"genre_scores_gemma":[0.020840349,0.01758291,0.00016716824,0.8688005,0.051373698,0.000034556266,0.0005201009,0.0001179513,0.040562775],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99548656,0.00015315218,0.0015131162,0.0011598422,0.0009587213,0.00072862924],"domain_scores_gemma":[0.97946316,0.017465772,0.00089465524,0.0013411792,0.00017933881,0.0006559091],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0025271976,0.00057276705,0.0022688266,0.00015588911,0.0001859468,0.000018690878,0.00049555243,0.0040755365,0.000553462],"category_scores_gemma":[0.1519539,0.00040428768,0.00022164301,0.00008360817,0.0022801666,0.00006350591,0.00024028745,0.012280344,0.00016485305],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000059702696,0.000006867752,0.0456483,0.00012725077,0.00015809383,0.0022596046,0.0001693195,2.1921855e-8,0.0000035960923,0.0001094234,0.9418937,0.009564118],"study_design_scores_gemma":[0.004501114,0.00080659805,0.013319708,0.028515631,0.0013479231,0.0010666922,0.00006323068,0.000053769567,0.000003420254,0.0011037875,0.9488071,0.0004110165],"about_ca_topic_score_codex":0.000097445845,"about_ca_topic_score_gemma":7.9392487e-7,"teacher_disagreement_score":0.1494267,"about_ca_system_score_codex":0.000053533804,"about_ca_system_score_gemma":0.00032975854,"threshold_uncertainty_score":0.9998409},"labels":[],"label_agreement":null},{"id":"W2604133518","doi":"10.3233/978-1-61499-742-9-167","title":"Expert Medical Decision-Making: How the Data-Frame Theory Can Explain Physician Sense-Making","year":2017,"lang":"en","type":"article","venue":"Studies in health technology and informatics","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Frame (networking); Medical decision making; Computer science; Clinical decision making; Task (project management); Decision theory; Naturalism; Making-of; Medical information; Management science; Psychology; Artificial intelligence; Knowledge management; Medicine; Epistemology; Mathematics; Family medicine","score_opus":0.09960240327436683,"score_gpt":0.468209270198994,"score_spread":0.3686068669246272,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2604133518","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.48687354,0.03209396,0.005175481,0.46536213,0.0028525775,0.001850441,0.00006927338,0.00046528212,0.00525734],"genre_scores_gemma":[0.9661316,0.010957607,0.0049454332,0.017747995,0.00014418074,0.00003367568,0.000006515136,0.000011106387,0.000021841808],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.99826556,0.00006240471,0.0006904974,0.00019505691,0.0003450406,0.00044144335],"domain_scores_gemma":[0.985669,0.012083086,0.0004835282,0.0015757659,0.00009040514,0.00009818566],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0026700434,0.00017798996,0.00060803996,0.00016714397,0.00095452927,0.000037784983,0.0005522443,0.00028697756,0.000005782977],"category_scores_gemma":[0.18428656,0.0001073448,0.000033511722,0.00017325973,0.0019121201,0.00012853096,0.0012141197,0.0008580658,0.00000528022],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010462905,0.00009528174,0.010360932,0.00023726257,0.00013424376,0.00008932131,0.008451345,0.0000013043594,1.411209e-7,0.031560864,0.013741684,0.935223],"study_design_scores_gemma":[0.009071784,0.002629193,0.043264735,0.0936732,0.00021559623,0.0016176682,0.37261558,0.020619553,0.00002907986,0.27815098,0.17678298,0.0013296325],"about_ca_topic_score_codex":0.000012750464,"about_ca_topic_score_gemma":0.00011099962,"teacher_disagreement_score":0.9338934,"about_ca_system_score_codex":0.00006742627,"about_ca_system_score_gemma":0.00024393556,"threshold_uncertainty_score":0.8225845},"labels":[],"label_agreement":null},{"id":"W2604684877","doi":"","title":"Supervision of clinical reasoning","year":2010,"lang":"en","type":"article","venue":"Canadian Family Physician","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Witness; Context (archaeology); Action (physics); Computer science; Clinical Practice; Medicine; Data science; Medical education; Nursing","score_opus":0.03724041657779519,"score_gpt":0.34725069697051075,"score_spread":0.31001028039271555,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2604684877","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.83386326,0.00009749716,0.0000069771136,0.0005531653,0.00079550286,0.0001393302,0.000025543692,0.000026822941,0.16449194],"genre_scores_gemma":[0.9652944,0.000025950929,0.0012028153,0.03241141,0.0007542166,0.0000040832183,0.000048517228,0.000025126314,0.00023347064],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9986579,0.000033366447,0.000492649,0.0002815925,0.00020177041,0.0003327671],"domain_scores_gemma":[0.99649435,0.0017125484,0.000096017095,0.00051261,0.00016858127,0.0010158834],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00046020892,0.00012400953,0.00047851674,0.00010129503,0.000058197053,0.000011053079,0.00013016068,0.00020302241,0.000024227442],"category_scores_gemma":[0.007749997,0.000112165086,0.00023907435,0.0002209481,0.00018179308,0.000039850795,0.000018893616,0.0005597362,0.00015767664],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000115621224,0.00046877773,0.17256959,0.00003627313,0.00012467163,0.0002091542,0.00008744643,0.000002959003,0.009438144,0.008673038,0.06028122,0.7479931],"study_design_scores_gemma":[0.0014718836,0.00034153802,0.90947235,0.0008454911,0.000111955684,0.000003657906,0.00016545007,0.00035826646,0.00050194305,0.0004346195,0.086086035,0.00020681658],"about_ca_topic_score_codex":0.012476169,"about_ca_topic_score_gemma":0.0057197227,"teacher_disagreement_score":0.7477863,"about_ca_system_score_codex":0.000022668673,"about_ca_system_score_gemma":0.0010095979,"threshold_uncertainty_score":0.99409986},"labels":[],"label_agreement":null},{"id":"W2607394511","doi":"10.1007/s11423-017-9521-6","title":"Success, failure and emotions: examining the relationship between performance feedback and emotions in diagnostic reasoning","year":2017,"lang":"en","type":"article","venue":"Educational Technology Research and Development","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":65,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Western Hospital; Alberta Advanced Education; University of Alberta; University Health Network; McGill University","funders":"Social Sciences and Humanities Research Council of Canada; Fonds de Recherche du Québec-Société et Culture; Canada Research Chairs","keywords":"Psychology; Cluster (spacecraft); Variety (cybernetics); Outcome (game theory); Task (project management); Negative emotion; Psychological intervention; Cognitive psychology; Developmental psychology; Computer science; Artificial intelligence","score_opus":0.13213841421505226,"score_gpt":0.418773756142733,"score_spread":0.2866353419276807,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2607394511","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93049186,0.00071953627,0.000016499518,0.06804034,0.00004032473,0.0003017897,0.0000024725816,0.000020570817,0.0003666269],"genre_scores_gemma":[0.9940901,0.0006096607,0.0043066465,0.000037944472,0.00009093465,0.0001481206,0.00003355697,0.000008843846,0.0006742053],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9988196,0.000049552098,0.00025279957,0.00032283503,0.00024882852,0.00030638627],"domain_scores_gemma":[0.98029184,0.018966597,0.000078378354,0.00031561713,0.00019452324,0.00015302394],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.0013307168,0.000102803715,0.00017295919,0.00031363763,0.0014582842,0.0001145333,0.00014993781,0.00016250393,0.000019459048],"category_scores_gemma":[0.12660164,0.00007888105,0.0000065581785,0.00025563675,0.0010177196,0.00011580513,0.0002555959,0.000688068,0.000019418192],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000006977715,0.00006054966,0.9789905,0.000025559506,0.00001377219,0.000003314871,0.00023324894,1.996125e-7,0.000002132273,0.0049468526,0.00053034746,0.0151865715],"study_design_scores_gemma":[0.000539114,0.00009859669,0.98971343,0.0013893738,0.000007897226,0.0000652597,0.0006523841,0.000009955042,0.000014145619,0.0066426266,0.0007869159,0.000080288904],"about_ca_topic_score_codex":0.000036461286,"about_ca_topic_score_gemma":0.00007969113,"teacher_disagreement_score":0.12527092,"about_ca_system_score_codex":0.0000659981,"about_ca_system_score_gemma":0.0006076941,"threshold_uncertainty_score":0.9998417},"labels":[],"label_agreement":null},{"id":"W2607539611","doi":"10.1111/acem.13209","title":"A Novel Approach to Study Medical Decision Making in the Clinical Setting: The “Own‐point‐of‐view” Perspective","year":2017,"lang":"en","type":"article","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":23,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Université de Sherbrooke","funders":"","keywords":"Perspective (graphical); Subject (documents); Medicine; Context (archaeology); Point (geometry); Field (mathematics); Cognition; Qualitative research; Medical education; Applied psychology; Psychology; Computer science; Psychiatry; Artificial intelligence","score_opus":0.15337048251571367,"score_gpt":0.5193815282424449,"score_spread":0.3660110457267312,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2607539611","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9128994,0.002888626,0.0034246838,0.05928507,0.0020936085,0.0024201507,0.000004669586,0.000039103405,0.016944734],"genre_scores_gemma":[0.9917669,0.0019783261,0.0007002257,0.0032363273,0.002063714,0.00011552569,0.000003870485,0.000030269865,0.00010480633],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9938843,0.00044858782,0.0024248466,0.0007559809,0.0020002883,0.0004859851],"domain_scores_gemma":[0.9777355,0.018880341,0.0007953525,0.0018573623,0.00033216234,0.00039928645],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.018660022,0.0003004803,0.000955552,0.00013365089,0.00033110465,0.0000030026813,0.0017980045,0.0004136449,0.00046740036],"category_scores_gemma":[0.4409146,0.0001405166,0.00023122643,0.0004749921,0.0006236186,0.00006475006,0.00043585838,0.0026071458,0.000027892787],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006436149,0.0028207756,0.8656023,0.000033236443,0.00023522894,0.00006429423,0.025510866,0.0000059969157,0.00003899187,0.002757695,0.039809372,0.06247759],"study_design_scores_gemma":[0.004257823,0.0011668524,0.9658871,0.0033963928,0.0004310926,0.00011584812,0.020627337,0.00042394525,0.0000032694495,0.0023082474,0.0012023402,0.00017970157],"about_ca_topic_score_codex":0.00047644717,"about_ca_topic_score_gemma":0.00006215424,"teacher_disagreement_score":0.42225456,"about_ca_system_score_codex":0.0000559106,"about_ca_system_score_gemma":0.00022512082,"threshold_uncertainty_score":0.9996939},"labels":[],"label_agreement":null},{"id":"W2608756344","doi":"10.1016/j.annemergmed.2017.03.018","title":"Psychological Skills to Improve Emergency Care Providers’ Performance Under Stress","year":2017,"lang":"en","type":"article","venue":"Annals of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":104,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Manitoba","funders":"","keywords":"Medicine; Set (abstract data type); Focus (optics); Resuscitation; Cognition; Communication skills; Critically ill; Medical education; Medical emergency; Nursing; Psychiatry; Intensive care medicine; Emergency medicine","score_opus":0.13382823029359572,"score_gpt":0.4713578927715114,"score_spread":0.3375296624779157,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2608756344","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.946625,0.0019633584,0.00008456606,0.021298997,0.0045973086,0.00068984885,0.00003776425,0.000057072073,0.024646115],"genre_scores_gemma":[0.9842145,0.008469767,0.0001519305,0.0009761718,0.0012773334,0.00005308119,0.000048899452,0.00003495954,0.0047733993],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99693847,0.000042299533,0.0011231278,0.0006458255,0.00067643676,0.0005738657],"domain_scores_gemma":[0.9959681,0.00039675506,0.00049658987,0.0015305401,0.00092308427,0.0006849049],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0005068032,0.00032319367,0.00078584364,0.000131093,0.00024828658,0.0000050940357,0.0005212668,0.00020980425,0.0047063245],"category_scores_gemma":[0.03210198,0.00022728188,0.00027489846,0.00019177273,0.00022228638,0.00012824024,0.00014983003,0.000319824,0.00012318007],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00077404885,0.0019565583,0.608993,0.00068283826,0.0003225145,0.00004941234,0.0028071704,0.000033889755,0.0033470015,0.00056238903,0.2867265,0.09374466],"study_design_scores_gemma":[0.0017785315,0.005578844,0.9801281,0.0027168614,0.00019662015,0.0000055644414,0.0008506117,0.000026105668,0.0035375473,0.0004895523,0.004354497,0.0003371693],"about_ca_topic_score_codex":0.00032069953,"about_ca_topic_score_gemma":0.000044756343,"teacher_disagreement_score":0.3711351,"about_ca_system_score_codex":0.000014798418,"about_ca_system_score_gemma":0.000060439732,"threshold_uncertainty_score":0.9962035},"labels":[],"label_agreement":null},{"id":"W2616587551","doi":"10.1017/cem.2017.330","title":"P128: The novel application of eye-tracking for the cognitive task analysis of expert physician decision-making while leading real-world traumatic resuscitations","year":2017,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Debriefing; Eye tracking; Cognition; Task (project management); Protocol analysis; Recall; Protocol (science); Medicine; Think aloud protocol; Variety (cybernetics); Psychology; Applied psychology; Computer science; Medical education; Human–computer interaction; Cognitive psychology; Artificial intelligence; Psychiatry; Cognitive science; Usability","score_opus":0.1465108396222113,"score_gpt":0.4724967543027614,"score_spread":0.3259859146805501,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2616587551","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7901694,0.0045812693,0.17050004,0.02105648,0.0044061528,0.001670237,0.00020916504,0.000010165531,0.007397098],"genre_scores_gemma":[0.99813384,0.00034627967,0.00044597866,0.00024258833,0.0007037441,0.000019808336,0.000015195275,0.000016964746,0.000075616954],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9979898,0.000037819835,0.0011890654,0.00015841692,0.00039325058,0.00023166211],"domain_scores_gemma":[0.9842759,0.011984454,0.0016559317,0.0005232171,0.0012850006,0.00027549628],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0015576271,0.00013513588,0.0006753524,0.00045730558,0.00048414854,0.000010155011,0.00040288782,0.000052580737,0.0001465952],"category_scores_gemma":[0.056081448,0.00007699125,0.00037738544,0.0007215605,0.000372914,0.00008929782,0.000011387084,0.00022325819,8.6620815e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010474938,0.0006825977,0.3229467,0.0003100083,0.0126046315,0.000038316997,0.056782614,0.00368984,0.009362489,0.0077630146,0.038358092,0.5464142],"study_design_scores_gemma":[0.0019598044,0.00048068716,0.9550258,0.009544302,0.008356625,0.0000032332478,0.010634998,0.010639098,0.00012673151,0.0011867948,0.0018806253,0.0001613377],"about_ca_topic_score_codex":0.007849064,"about_ca_topic_score_gemma":0.04985859,"teacher_disagreement_score":0.63207906,"about_ca_system_score_codex":0.0000706686,"about_ca_system_score_gemma":0.0004230311,"threshold_uncertainty_score":0.9987578},"labels":[],"label_agreement":null},{"id":"W2619928717","doi":"10.2147/amep.s128321","title":"Investigating a self-scoring interview simulation for learning and assessment in the medical consultation","year":2017,"lang":"en","type":"article","venue":"Advances in Medical Education and Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Generalizability theory; Reliability (semiconductor); Medical education; Computer science; Variance (accounting); Objective structured clinical examination; Psychology; Applied psychology; Medicine","score_opus":0.06238171820043661,"score_gpt":0.5473635698218368,"score_spread":0.4849818516214002,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2619928717","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.56866163,0.016682532,0.014646217,0.3740969,0.0014224357,0.00203404,0.0000011631732,0.00006838617,0.02238668],"genre_scores_gemma":[0.96609086,0.010055988,0.016391786,0.007110949,0.00018725726,0.000118764554,0.000013689674,0.000005907392,0.000024805571],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9986258,0.00034477332,0.00032861024,0.00018960221,0.00040066708,0.00011051436],"domain_scores_gemma":[0.9156787,0.08327941,0.00039408664,0.0001806804,0.00017947619,0.0002876777],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0052013956,0.00006976983,0.00015507922,0.000039563143,0.00017053513,0.000069270245,0.00006634948,0.000093148694,0.000033997036],"category_scores_gemma":[0.64998764,0.000050433948,0.000014863671,0.000065733,0.00014112827,0.00038453026,0.000033014043,0.00055844174,7.1934824e-7],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000043083633,0.00032767397,0.048982028,0.00013443032,0.0000066792763,0.000009324544,0.0022853708,0.000015395493,7.142934e-7,0.0026003523,0.000027417525,0.94556755],"study_design_scores_gemma":[0.012760701,0.0007673504,0.368929,0.016591696,0.00025553492,0.000678185,0.024424002,0.25105262,0.0000035037965,0.009879049,0.31430325,0.00035513123],"about_ca_topic_score_codex":0.00006335496,"about_ca_topic_score_gemma":0.00006175582,"teacher_disagreement_score":0.9452124,"about_ca_system_score_codex":0.000028559187,"about_ca_system_score_gemma":0.0006574613,"threshold_uncertainty_score":0.35296068},"labels":[],"label_agreement":null},{"id":"W2621630726","doi":"","title":"Diagnosis : Interpreting the Shadows","year":2017,"lang":"en","type":"book","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":29,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Computer science; Geography","score_opus":0.035883805407639274,"score_gpt":0.3499391546846364,"score_spread":0.3140553492769971,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2621630726","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000117850934,0.0012028653,0.00002375431,0.008448703,0.0006277952,0.0004123404,0.000014465503,0.00009995682,0.9890523],"genre_scores_gemma":[0.004792861,0.0010694248,0.00014697065,0.0046372176,0.0011112012,0.000085775115,0.00007191561,0.000050214974,0.9880344],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9987299,0.000023198296,0.0003488798,0.0003518893,0.00031068875,0.00023546637],"domain_scores_gemma":[0.9758833,0.022432653,0.00024888318,0.0011513841,0.00009922748,0.0001845584],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00044004872,0.00024206477,0.00054262066,0.000038619288,0.00016211461,0.000083494306,0.00035376553,0.00036700835,0.0037048706],"category_scores_gemma":[0.08703965,0.0001295646,0.00038572075,0.000013586813,0.00023680434,0.000018091401,0.00020800337,0.00076570385,0.0011431191],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000021522346,0.000043433836,0.0026542668,0.000026129184,0.00010627087,0.00008581273,0.00006639535,1.12196275e-7,1.2482244e-7,0.0013697838,0.90659994,0.089026205],"study_design_scores_gemma":[0.00043738855,0.00014710515,0.0022750238,0.013317852,0.00039345468,0.000045210803,0.00001698063,0.000031027095,0.000009605848,0.0031430891,0.9799846,0.00019869194],"about_ca_topic_score_codex":0.000058157173,"about_ca_topic_score_gemma":0.000020775788,"teacher_disagreement_score":0.08882751,"about_ca_system_score_codex":0.000088677756,"about_ca_system_score_gemma":0.00045883784,"threshold_uncertainty_score":0.9996346},"labels":[],"label_agreement":null},{"id":"W2622241629","doi":"10.1080/0142159x.2017.1331034","title":"Diagnosis and management of clinical reasoning difficulties: Part II. Clinical reasoning difficulties: Management and remediation strategies","year":2017,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":52,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Université de Montréal","funders":"","keywords":"Management science; Engineering ethics; Psychology; Computer science; Engineering","score_opus":0.070722513415465,"score_gpt":0.41007939403807026,"score_spread":0.3393568806226053,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2622241629","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9839854,0.0015494947,0.00013639434,0.0033604512,0.00086346274,0.000728726,0.000005267537,0.00011552958,0.009255243],"genre_scores_gemma":[0.94601154,0.04253931,0.004916539,0.00032799976,0.0010961521,0.00016107067,0.000050639,0.00004846086,0.0048482595],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99520165,0.00030607753,0.0018654782,0.0009304526,0.0011872225,0.000509151],"domain_scores_gemma":[0.9942781,0.0028486622,0.0007774523,0.0010045011,0.00011794423,0.0009733724],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.004752915,0.00036422184,0.0012401713,0.00007285374,0.00044585406,0.00012833787,0.000347034,0.00056929304,0.00042631722],"category_scores_gemma":[0.03822638,0.00027505192,0.00029419892,0.000086583575,0.0014077349,0.00017686829,0.00078046514,0.00088343705,0.000015242992],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002422699,0.0013654624,0.7733245,0.00035323587,0.0006223633,0.0003259185,0.00035298488,0.000001042564,1.9216021e-7,0.0023491853,0.01394539,0.20711747],"study_design_scores_gemma":[0.006879496,0.0006422797,0.96769905,0.0096020065,0.0007978135,0.000024762725,0.0022802453,0.0012802368,0.000003118071,0.00018185475,0.010296292,0.00031281845],"about_ca_topic_score_codex":0.000112641705,"about_ca_topic_score_gemma":0.000028476665,"teacher_disagreement_score":0.20680466,"about_ca_system_score_codex":0.000035078534,"about_ca_system_score_gemma":0.000093618524,"threshold_uncertainty_score":0.99997014},"labels":[],"label_agreement":null},{"id":"W2622613082","doi":"","title":"Reflections on the Neuroscience of Decision Making","year":2012,"lang":"en","type":"article","venue":"Dialnet (Universidad de la Rioja)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Defence Research and Development Canada","funders":"","keywords":"Cognitive neuroscience; Cognitive science; Neuroscience; Prefrontal cortex; Neurolaw; Cognition; Psychology; Computational neuroscience; Field (mathematics); Neuroeconomics; Neuroethics; Systems neuroscience; Computer science; Social neuroscience; Social cognition","score_opus":0.04961302997467656,"score_gpt":0.39211686930722894,"score_spread":0.3425038393325524,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2622613082","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.91559696,0.00005479179,0.0017019312,0.0021936651,0.0003101687,0.000111388654,0.0000043173986,0.00003486552,0.07999189],"genre_scores_gemma":[0.99667716,0.000048645412,0.00070886326,0.0021616328,0.0001114172,0.0000017302627,0.000001121793,0.00000799356,0.00028142196],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"observational","domain_scores_codex":[0.9992516,0.00009086344,0.000101632926,0.0001222062,0.00022476079,0.00020895466],"domain_scores_gemma":[0.9875975,0.011915069,0.000067601046,0.0002598415,0.000036939662,0.000123005],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0005388378,0.00006815896,0.00012856271,0.0000918671,0.00012133203,0.000012382505,0.00011306671,0.000076790886,0.00013769254],"category_scores_gemma":[0.017770564,0.000048953618,0.000087671106,0.00037460524,0.0001848908,0.0000581745,0.000044029886,0.00020191098,0.00003679941],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0015389489,0.0028415853,0.25493744,0.00003810993,0.00008709629,0.00030128323,0.00739648,0.0005044987,0.013392885,0.42327264,0.09389804,0.20179099],"study_design_scores_gemma":[0.0008753593,0.00037125262,0.96278274,0.00088565,0.00011319777,0.00009753521,0.00059190625,0.00038081873,0.0006730193,0.0022998857,0.030803308,0.00012530539],"about_ca_topic_score_codex":0.000031032196,"about_ca_topic_score_gemma":0.000004679018,"teacher_disagreement_score":0.70784533,"about_ca_system_score_codex":0.000047314545,"about_ca_system_score_gemma":0.00007062484,"threshold_uncertainty_score":0.9905032},"labels":[],"label_agreement":null},{"id":"W2623877750","doi":"","title":"What counts and what is being counted – The social organization of knowledge on the front lines of emergency medical services.","year":2013,"lang":"en","type":"article","venue":"Research Portal (Queen's University Belfast)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Front (military); Business; Front line; Public relations; Political science; Geography","score_opus":0.023109579418651206,"score_gpt":0.3350648172202939,"score_spread":0.3119552378016427,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2623877750","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97381794,0.00019646602,0.000009543798,0.022024527,0.00020708815,0.00033750746,0.000008181504,0.00001459978,0.0033841687],"genre_scores_gemma":[0.98847306,0.0075148195,0.000006164125,0.000160453,0.00012507378,7.878521e-7,0.000028402575,0.000010151567,0.0036810858],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99857056,0.00016984987,0.00018072722,0.00016023402,0.0007360624,0.0001825594],"domain_scores_gemma":[0.9963215,0.0022032892,0.0000849679,0.0001971196,0.0010549439,0.00013817126],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.000727015,0.00007859387,0.0001854157,0.00007846114,0.00023447462,0.000027861533,0.00018787844,0.00012981401,0.004621771],"category_scores_gemma":[0.0028376244,0.00004973483,0.000050960127,0.00037003795,0.00028590852,0.00023160451,0.00015478533,0.00035237698,0.000095435906],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00053862925,0.0034811571,0.3156663,0.0013360326,0.001370844,0.00028959377,0.04598111,0.0000032783475,0.000505633,0.048039157,0.50397056,0.07881772],"study_design_scores_gemma":[0.004978887,0.0012794515,0.7426843,0.012357617,0.0004903094,0.000017008288,0.1802021,0.0017946986,0.0026417298,0.00143023,0.051537298,0.0005863621],"about_ca_topic_score_codex":0.0006458663,"about_ca_topic_score_gemma":0.000025835277,"teacher_disagreement_score":0.45243326,"about_ca_system_score_codex":0.000030450878,"about_ca_system_score_gemma":0.00022410323,"threshold_uncertainty_score":0.9962881},"labels":[],"label_agreement":null},{"id":"W2625152421","doi":"10.1188/17.onf.e152-e167","title":"Antecedents and Outcomes of Uncertainty in Older Adults With Cancer: A Scoping Review of the Literature","year":2017,"lang":"en","type":"review","venue":"Oncology nursing forum","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":36,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Canadian Partnership Against Cancer; University of Toronto","funders":"","keywords":"Medicine; Anticipation (artificial intelligence); Wright; Cancer; Etiology; Event (particle physics); Psychiatry; Internal medicine; History; Art history","score_opus":0.04372948571687779,"score_gpt":0.46652283492755414,"score_spread":0.42279334921067635,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2625152421","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00037600935,0.99425346,4.143676e-7,0.0023777152,0.0004480688,0.0019366499,0.000044026565,0.000006293551,0.0005573385],"genre_scores_gemma":[0.00038080866,0.9988279,0.00006312542,0.00039624042,0.00003554695,0.00009161599,0.000032816668,0.0000222183,0.00014973166],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.998284,0.00016494805,0.0007739898,0.00032059135,0.00019822893,0.00025828197],"domain_scores_gemma":[0.9953002,0.0028008502,0.0011315121,0.0005521997,0.00013428,0.000080986785],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0003999551,0.0002559212,0.0031620201,0.00008912491,0.00005208786,0.0000066110415,0.00020568346,0.000390689,0.000014239661],"category_scores_gemma":[0.008770346,0.00012903762,0.0002967415,0.00019811402,0.000454298,0.000023746261,0.000057406294,0.0005233178,4.3059927e-7],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000053866395,0.00022730815,0.0055778124,0.11960877,0.0000792318,0.000021047419,0.00016798855,8.78937e-8,1.2164624e-8,0.000010133359,0.0004988012,0.8737549],"study_design_scores_gemma":[0.0014309956,0.0003107844,0.0037735396,0.9784721,0.0012327064,0.00012433066,0.00003311107,0.0000010298147,2.219923e-7,0.000009408928,0.014503106,0.000108674205],"about_ca_topic_score_codex":0.000087069704,"about_ca_topic_score_gemma":0.00018416061,"teacher_disagreement_score":0.87364626,"about_ca_system_score_codex":0.00021195081,"about_ca_system_score_gemma":0.0014015494,"threshold_uncertainty_score":0.9995792},"labels":[],"label_agreement":null},{"id":"W2685756418","doi":"10.11575/prism/34808","title":"More questions than answers; using an observational study to count reference activity","year":2012,"lang":"en","type":"book-chapter","venue":"PRISM (University of Calgary)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Observational study; Count data; Statistics; Mathematics","score_opus":0.15172550538399962,"score_gpt":0.3420441721899169,"score_spread":0.19031866680591727,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2685756418","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9328702,0.00015005564,0.018032523,0.00077621586,0.00019728935,0.0010394558,0.000021213416,0.000100211786,0.046812814],"genre_scores_gemma":[0.7819913,0.00023973276,0.042992067,0.00046283152,0.00024278216,0.0000013619845,0.00040953906,0.00009530548,0.1735651],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9986437,0.000038079666,0.0001748184,0.0003943213,0.00054305745,0.0002060053],"domain_scores_gemma":[0.9977267,0.0008083767,0.00021308524,0.00051087316,0.00028469163,0.0004562758],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00027889697,0.00022731475,0.0005551552,0.00016408334,0.00014878246,0.000007921125,0.00019153993,0.00026987694,0.00048881886],"category_scores_gemma":[0.0008332833,0.00026174315,0.00013779022,0.000053429714,0.00015478233,0.00014996738,0.0001549903,0.00047371222,0.00004980754],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0049072783,0.017891891,0.3192614,0.00043923006,0.002820138,0.0016089771,0.020520307,0.00003407113,0.0009456854,0.06046533,0.00733205,0.56377363],"study_design_scores_gemma":[0.0017124556,0.0014120275,0.97241557,0.0019503746,0.0016539191,0.000019790981,0.00025950483,0.002020813,0.000006972275,0.0006964171,0.017280178,0.0005719992],"about_ca_topic_score_codex":0.0033429186,"about_ca_topic_score_gemma":0.000120039156,"teacher_disagreement_score":0.65315413,"about_ca_system_score_codex":0.00019870234,"about_ca_system_score_gemma":0.00036854224,"threshold_uncertainty_score":0.9999835},"labels":[],"label_agreement":null},{"id":"W2732175898","doi":"10.12740/app/70659","title":"From foundations to frontiers: the status and perception of formulation teaching in Canadian English-speaking psychiatry residency programs","year":2017,"lang":"en","type":"article","venue":"Archives of Psychiatry and Psychotherapy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Curriculum; Competence (human resources); Perception; Medical education; Psychology; Standardization; Residency training; Context (archaeology); Medicine; Pedagogy; Political science; Social psychology; Continuing education","score_opus":0.019163682096853346,"score_gpt":0.3428483601187969,"score_spread":0.32368467802194356,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2732175898","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9819412,0.0011702217,0.00086514757,0.011318702,0.00085774687,0.00047239254,0.000025433963,0.000010673268,0.0033384569],"genre_scores_gemma":[0.94915646,0.0016279031,0.04842743,0.00031355012,0.00037921293,0.00001488219,0.000034300178,0.0000147584815,0.000031488144],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9989406,0.000068733985,0.00035271115,0.0002580074,0.00011697007,0.00026298023],"domain_scores_gemma":[0.9985957,0.0004949696,0.00019002253,0.00044244094,0.000020689231,0.00025619028],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00026394968,0.00011917911,0.00023718612,0.00015673332,0.0002699577,0.000053167045,0.00012460025,0.0000657339,0.00001740794],"category_scores_gemma":[0.0005517694,0.00009102824,0.00006807781,0.000051710384,0.00014830388,0.00011221065,0.0000073123047,0.00020999888,4.2968222e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00031819323,0.00013267515,0.835154,0.000019850899,0.000042311734,1.7788868e-7,0.013668605,0.0000018959344,0.000046383448,0.0076160384,0.00012237547,0.1428775],"study_design_scores_gemma":[0.0014836875,0.0003609145,0.9078876,0.0014635621,0.00003574067,0.0000011132155,0.0017960116,0.00015451635,0.0000015253356,0.085536286,0.0011901632,0.00008891639],"about_ca_topic_score_codex":0.09057763,"about_ca_topic_score_gemma":0.29287267,"teacher_disagreement_score":0.20229502,"about_ca_system_score_codex":0.000009065126,"about_ca_system_score_gemma":0.00026675683,"threshold_uncertainty_score":0.9154783},"labels":[],"label_agreement":null},{"id":"W2734948659","doi":"10.6084/m9.figshare.5123776.v1","title":"Supplementary Material for: Alternate-Form Reliability of the Montreal Cognitive Assessment Screening Test in a Clinical Setting","year":2012,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Reliability (semiconductor); Test (biology); Montreal Cognitive Assessment; Cognition; Psychology; Reliability engineering; Computer science; Cognitive impairment; Medicine; Psychiatry; Engineering","score_opus":0.05897627331373613,"score_gpt":0.4052762191826984,"score_spread":0.3462999458689623,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2734948659","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6651513,0.0000146204275,0.0000036103752,0.00037540283,0.00017005121,0.00080585305,0.33286315,0.000013712592,0.0006022932],"genre_scores_gemma":[0.9329237,0.0000022418385,0.00086695905,0.00041719756,0.0004690989,0.00017858359,0.06510061,0.000014911119,0.000026751053],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9985241,0.000076553304,0.00065226585,0.00020518113,0.00022646195,0.00031541393],"domain_scores_gemma":[0.9718337,0.027244957,0.0002853352,0.0002884116,0.00016672477,0.00018088591],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00071969564,0.00011576281,0.00031701842,0.000023404036,0.00004237229,0.000010281181,0.00010510709,0.00010240851,0.020758567],"category_scores_gemma":[0.11854328,0.00007881515,0.00019381166,0.000077289,0.000023131824,0.000054875807,0.0001621507,0.00024203685,0.000010008058],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00017660928,0.00057455216,0.9613233,0.00008946242,0.00002271227,0.00000374498,0.000110211295,0.0000013209639,0.0000074763066,0.000001866833,0.024714468,0.012974271],"study_design_scores_gemma":[0.0026681956,0.0002842593,0.9852814,0.009609296,0.000075507785,0.000007286061,0.00017032507,0.00035955198,0.0006025696,0.000036095542,0.0008192378,0.00008629159],"about_ca_topic_score_codex":0.00027678706,"about_ca_topic_score_gemma":0.00009665661,"teacher_disagreement_score":0.26777235,"about_ca_system_score_codex":0.000048729926,"about_ca_system_score_gemma":0.00011385109,"threshold_uncertainty_score":0.9801366},"labels":[],"label_agreement":null},{"id":"W2736907930","doi":"10.1016/j.jsurg.2017.07.012","title":"Defining and Measuring Decision-Making for the Management of Trauma Patients","year":2017,"lang":"en","type":"article","venue":"Journal of surgical education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University Health Centre","funders":"McGill University Health Centre","keywords":"Interpersonal communication; Cognition; Psychology; Medicine; Qualitative research; Social skills; Clinical psychology; Applied psychology; Medical education; Social psychology; Psychiatry","score_opus":0.039327957037094984,"score_gpt":0.38984688162448894,"score_spread":0.350518924587394,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2736907930","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99565774,0.0010576982,0.00041358863,0.0010718065,0.00056862057,0.00012941477,8.329965e-7,0.0000013275902,0.0010989672],"genre_scores_gemma":[0.9928841,0.00051075825,0.006390696,0.000039936727,0.00013566509,0.0000033934987,5.908897e-7,0.0000046890295,0.000030146133],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9992996,0.000009263465,0.0003384473,0.0000611716,0.00022360691,0.00006791499],"domain_scores_gemma":[0.9940849,0.004975665,0.00045275685,0.00014416892,0.0002729344,0.00006956253],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0006511188,0.000044658875,0.00016449303,0.0000338211,0.00012626215,0.000030122535,0.00008264176,0.000029797482,0.000013007065],"category_scores_gemma":[0.007138494,0.00002616603,0.00009231577,0.000021759259,0.00004402612,0.00005689605,0.000022771244,0.00008467438,6.0640974e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00033407263,0.00028612657,0.075826555,0.000037325943,0.000063798856,0.0000038003682,0.00009402023,0.0000071285917,5.801751e-7,0.00073010084,0.00029568555,0.9223208],"study_design_scores_gemma":[0.0019478349,0.00022760771,0.98704404,0.0061560776,0.00024635045,0.00004621368,0.00029378932,0.000076089775,0.000025313337,0.0015041505,0.002396789,0.000035733756],"about_ca_topic_score_codex":0.0000035454254,"about_ca_topic_score_gemma":3.2542766e-7,"teacher_disagreement_score":0.9222851,"about_ca_system_score_codex":0.000020975514,"about_ca_system_score_gemma":0.000060665345,"threshold_uncertainty_score":0.85459596},"labels":[],"label_agreement":null},{"id":"W2736987381","doi":"10.1177/070674370204700620","title":"Dr Fine Replies","year":2002,"lang":"en","type":"article","venue":"The Canadian Journal of Psychiatry","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Psychology; Medicine","score_opus":0.030778650536660755,"score_gpt":0.2947982900398004,"score_spread":0.26401963950313967,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2736987381","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4934728,0.01247374,0.000019733288,0.45789656,0.0039573433,0.000118922464,0.000009242857,0.000012359078,0.0320393],"genre_scores_gemma":[0.988454,0.000071128095,0.001455425,0.0043878327,0.0014415467,6.1251154e-7,8.9042925e-7,0.00001342227,0.0041751624],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9992146,0.000027971068,0.00032818943,0.000067756315,0.00016289887,0.00019861314],"domain_scores_gemma":[0.99821806,0.0006607012,0.00015143062,0.0002536838,0.00010955955,0.0006065652],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00037674542,0.00007278759,0.00019043172,0.00009109619,0.0001372341,0.000026668145,0.0001426296,0.000059942828,0.0016451261],"category_scores_gemma":[0.0063297707,0.000043830703,0.00014944682,0.00013658918,0.00012386992,0.000023691939,0.0000039626248,0.00036350123,0.00012041462],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000030274507,0.000060597402,0.25103343,0.000008440619,0.00009374944,0.00015624386,0.00028339578,0.000012398429,0.0000024566623,0.0069167977,0.7330603,0.008341917],"study_design_scores_gemma":[0.006011414,0.0024192806,0.68296075,0.0049609733,0.0010179573,0.009241166,0.0007562443,0.00020710746,0.000057754092,0.05459059,0.2373081,0.00046866792],"about_ca_topic_score_codex":0.0025548455,"about_ca_topic_score_gemma":0.09155759,"teacher_disagreement_score":0.49575222,"about_ca_system_score_codex":0.000047020076,"about_ca_system_score_gemma":0.00083054026,"threshold_uncertainty_score":0.9992675},"labels":[],"label_agreement":null},{"id":"W2738295307","doi":"10.1017/cem.2017.357","title":"Teaching diagnostic reasoning: using simulation and mixed practice to build competence","year":2017,"lang":"en","type":"review","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":30,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Medicine; Medical diagnosis; Competence (human resources); Medical education; Acute abdominal pain; Vital signs; Emergency department; Variety (cybernetics); Clinical clerkship; Patient care; Simulated patient; Medical emergency; Abdominal pain; Nursing; Psychology; Curriculum; Computer science; Pedagogy; Surgery; Artificial intelligence","score_opus":0.2885986407626282,"score_gpt":0.5161361898804296,"score_spread":0.22753754911780144,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2738295307","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000415677,0.98813987,0.0005124417,0.0019940883,0.0068217753,0.0005908049,0.000018221734,0.000008667904,0.0014984803],"genre_scores_gemma":[0.0024094833,0.9907442,0.0019907919,0.00028164667,0.004000288,0.000005337977,0.000025440657,0.000077358294,0.00046546385],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99630845,0.00035018293,0.0019306308,0.00039382413,0.00051294244,0.0005039801],"domain_scores_gemma":[0.9747837,0.018171437,0.0022649555,0.0006568916,0.0008551808,0.003267792],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0027157431,0.00045488696,0.002519093,0.0007015834,0.0003910859,0.00002658543,0.0003603075,0.00031746752,0.00033093654],"category_scores_gemma":[0.5590153,0.00033866076,0.00034727564,0.00025883378,0.00016975966,0.00017780333,0.000039811315,0.0013664843,0.000014994553],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003749812,0.000063896696,0.0011337486,0.0027328627,0.000521918,0.0019946827,0.00063043024,0.00017419351,7.039352e-7,0.00016507608,0.011131645,0.98141336],"study_design_scores_gemma":[0.0004717366,0.000660917,0.00030439088,0.15455513,0.004661059,0.0008205164,0.00012110812,0.0001446756,4.7203486e-8,0.0000321968,0.83797413,0.00025409317],"about_ca_topic_score_codex":0.011055787,"about_ca_topic_score_gemma":0.0016127608,"teacher_disagreement_score":0.98115927,"about_ca_system_score_codex":0.00032111938,"about_ca_system_score_gemma":0.0030956601,"threshold_uncertainty_score":0.99990654},"labels":[],"label_agreement":null},{"id":"W2738698971","doi":"10.1007/s10459-017-9787-9","title":"Erratum to: Hemispheric activation differences in novice and expert clinicians during clinical decision making","year":2017,"lang":"en","type":"erratum","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria; University of Calgary","funders":"","keywords":"Clinical decision making; Medical education; Medical physics; Psychology; Computer science; Medicine; Family medicine","score_opus":0.052019184816592796,"score_gpt":0.49413491007692983,"score_spread":0.44211572526033704,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2738698971","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.85675466,0.05230073,0.0003986712,0.01299355,0.06155993,0.0022092313,0.0000070881038,0.00008412965,0.013691984],"genre_scores_gemma":[0.8878873,0.08940324,0.010062952,0.0047555864,0.00309791,0.00021762731,0.00004653262,0.00003714445,0.0044916924],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9957396,0.00018897193,0.0016497021,0.0011973735,0.0006378802,0.000586473],"domain_scores_gemma":[0.99125063,0.0066927783,0.0009736907,0.0005709251,0.00013892521,0.00037301745],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0026000983,0.0003074398,0.0009854456,0.00038897042,0.00038734815,0.0001219508,0.00041570133,0.00042707252,0.000022057182],"category_scores_gemma":[0.059916418,0.00026497708,0.000079969555,0.00068734423,0.00034999484,0.00046317966,0.00012442587,0.0011332944,0.000009282207],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020651367,0.0005871376,0.41534412,0.00040989718,0.000003694784,0.0000045094935,0.00073608116,0.00004172725,9.743314e-7,0.000040199666,0.018961167,0.56366396],"study_design_scores_gemma":[0.00081079535,0.00086176314,0.8974068,0.06977479,0.000010245521,0.0000150969345,0.0014136506,0.0013849114,0.0000011165931,0.0019781783,0.025941871,0.0004007688],"about_ca_topic_score_codex":0.0010350326,"about_ca_topic_score_gemma":0.0018004595,"teacher_disagreement_score":0.56326324,"about_ca_system_score_codex":0.0003482779,"about_ca_system_score_gemma":0.0036060996,"threshold_uncertainty_score":0.9999803},"labels":[],"label_agreement":null},{"id":"W2738834467","doi":"10.1177/0844562116676560","title":"Promoting Clinical Reasoning Among Nursing Students: Why Aren’t Clinical Decision Support Systems a Popular Option?","year":2016,"lang":"en","type":"editorial","venue":"Canadian Journal of Nursing Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Psychology; Clinical decision support system; Management science; Decision support system; Nursing; Medical education; Medicine; Computer science; Engineering; Artificial intelligence","score_opus":0.10097420868947567,"score_gpt":0.5296590584439895,"score_spread":0.4286848497545139,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2738834467","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.12152741,0.0045227795,0.00064418855,0.0023047237,0.8682816,0.0013236948,0.00012985729,0.000046986574,0.0012187383],"genre_scores_gemma":[0.1392902,0.0015746701,0.0012647387,0.00008141068,0.8554527,0.000019623436,0.00017367997,0.00029583168,0.0018471269],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9821485,0.002927032,0.0053197197,0.0012635081,0.0061363797,0.0022048126],"domain_scores_gemma":[0.88702196,0.08459035,0.0033312538,0.0021422894,0.011123192,0.011790971],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["metaresearch","research_integrity"],"category_scores_codex":[0.03976863,0.00063485274,0.0028556539,0.0019297212,0.0007768926,0.0009390968,0.0015983223,0.0031105748,0.00013665439],"category_scores_gemma":[0.4749306,0.0005033291,0.0012673243,0.0008207406,0.002421441,0.0003313935,0.000084289975,0.010153567,0.000117254516],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.0006511336,0.0007947402,0.07017602,0.0000613252,0.00022563887,0.003677409,0.0005465891,0.0000023652126,8.0185487e-7,0.000023885714,0.7841612,0.13967887],"study_design_scores_gemma":[0.010498309,0.009293812,0.04802285,0.62543076,0.0014233458,0.0013043447,0.001592188,0.00020658277,0.000001888206,0.0014841174,0.29978207,0.0009597508],"about_ca_topic_score_codex":0.0013877423,"about_ca_topic_score_gemma":0.00045030564,"teacher_disagreement_score":0.6253694,"about_ca_system_score_codex":0.004960173,"about_ca_system_score_gemma":0.01942539,"threshold_uncertainty_score":0.99974185},"labels":[],"label_agreement":null},{"id":"W2739066951","doi":"10.5430/jnep.v7n12p28","title":"Addressing the development of both knowledge and clinical reasoning in nursing through the perspective of script concordance: An integrative literature review","year":2017,"lang":"en","type":"article","venue":"Journal of Nursing Education and Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":23,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"PsycINFO; Concordance; Perspective (graphical); Cognition; Psychology; Situated; Scripting language; MEDLINE; Scopus; Situational ethics; Knowledge management; Nursing; Computer science; Medicine; Social psychology; Artificial intelligence; Psychiatry","score_opus":0.1864704666557684,"score_gpt":0.5757031419841134,"score_spread":0.38923267532834505,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2739066951","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.36826053,0.50262684,0.0006604879,0.08031008,0.0024138428,0.00063313585,0.0000023744979,0.0000054201246,0.045087267],"genre_scores_gemma":[0.94145864,0.036790352,0.020480536,0.0008112914,0.00036251522,0.0000020688456,0.0000015767566,0.000008855152,0.00008418742],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9983079,0.00060275185,0.00069677667,0.00012796883,0.0001785284,0.00008606826],"domain_scores_gemma":[0.9881038,0.008842121,0.0017144844,0.00024960405,0.0009996981,0.000090311885],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0033460662,0.00009341136,0.00038631444,0.000041125884,0.00030951816,0.0000642561,0.00011738132,0.00006635416,0.0000036896083],"category_scores_gemma":[0.098310396,0.00004924208,0.000060334696,0.00011363338,0.0005079683,0.00046775996,0.000015899157,0.00064832345,1.0493469e-7],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00086582324,0.0031785266,0.0058465526,0.00018463637,0.00010493496,0.0000056322137,0.20213501,2.7197376e-7,0.00006920234,0.004240117,0.0041300394,0.77923924],"study_design_scores_gemma":[0.0016196518,0.0008474622,0.30063885,0.5539694,0.0010164026,0.0023241115,0.1314546,0.00006387197,0.00011590143,0.0029551883,0.004859913,0.00013466962],"about_ca_topic_score_codex":0.000020479049,"about_ca_topic_score_gemma":0.000007980673,"teacher_disagreement_score":0.7791046,"about_ca_system_score_codex":0.00010641451,"about_ca_system_score_gemma":0.0011390924,"threshold_uncertainty_score":0.90928495},"labels":[],"label_agreement":null},{"id":"W2739316696","doi":"10.1177/070674370404901112","title":"Book Review: Quick Cognitive Screening for Clinicians","year":2004,"lang":"en","type":"article","venue":"The Canadian Journal of Psychiatry","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Cognition; Psychology; Medicine; Psychiatry","score_opus":0.033547094559935806,"score_gpt":0.3565252292431342,"score_spread":0.32297813468319836,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2739316696","genre_codex":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0058883945,0.5203565,0.0024282904,0.4665532,0.0023811983,0.0007375426,0.000057980735,0.0000110309775,0.001585863],"genre_scores_gemma":[0.11757223,0.017686723,0.021403246,0.83400327,0.0065481733,0.000025354699,0.000035906385,0.000101751684,0.002623326],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9987932,0.00003974809,0.0005939774,0.0001066067,0.00017963826,0.00028681892],"domain_scores_gemma":[0.9965811,0.0018156861,0.0003494923,0.00017046709,0.00035576988,0.0007275036],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0009820604,0.000109622524,0.00036919437,0.00008263085,0.00020273353,0.000019722802,0.00015103197,0.0000818068,0.00026064718],"category_scores_gemma":[0.015937285,0.0000719891,0.00036153337,0.00013101129,0.00014868256,0.000042624382,0.0000044046965,0.00042970493,0.000026906759],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021568486,0.0000788128,0.016299613,0.00013766107,0.00023859502,0.000075730095,0.00015611999,0.000023677423,3.403384e-7,0.001353167,0.97300106,0.008419546],"study_design_scores_gemma":[0.019126493,0.0054356153,0.0809891,0.17524907,0.00479202,0.003402515,0.0012207469,0.000013240677,0.000038673046,0.026114946,0.6829962,0.0006213957],"about_ca_topic_score_codex":0.002026232,"about_ca_topic_score_gemma":0.033989016,"teacher_disagreement_score":0.50266975,"about_ca_system_score_codex":0.00007026761,"about_ca_system_score_gemma":0.005024781,"threshold_uncertainty_score":0.9923519},"labels":[],"label_agreement":null},{"id":"W2740162871","doi":"10.1080/16549716.2017.1350452","title":"How patients navigate the diagnostic ecosystem in a fragmented health system: a qualitative study from India","year":2017,"lang":"en","type":"article","venue":"Global Health Action","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":53,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; McGill University Health Centre","funders":"Bill and Melinda Gates Foundation","keywords":"Medicine; Focus group; Qualitative research; Health care; Test (biology); Nursing; Family medicine; Business; Economic growth; Marketing; Sociology","score_opus":0.049126624095884384,"score_gpt":0.4539791557388327,"score_spread":0.4048525316429483,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2740162871","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9819077,0.00034153223,0.000085644875,0.012535564,0.0016681219,0.0028055317,0.00045127506,0.00007689531,0.00012769291],"genre_scores_gemma":[0.9981442,0.00010721388,0.00004742492,0.001045864,0.00024482596,0.00019698785,0.0001816177,0.000013542121,0.00001833964],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.996785,0.0010239964,0.0006841171,0.00045690185,0.0005579035,0.00049212156],"domain_scores_gemma":[0.993143,0.0045665624,0.0009810249,0.0007146897,0.00013369032,0.0004610136],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0015678002,0.0002046985,0.0006561148,0.00003950919,0.00059622014,0.00013169851,0.0001895043,0.0001000504,0.000004954148],"category_scores_gemma":[0.021528233,0.00014338877,0.00008210219,0.00016404032,0.000047690482,0.00013103873,0.000069962756,0.0003555865,0.000054029457],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00058254955,0.0022747947,0.8983187,0.00038348083,0.000115393304,0.000049042326,0.013421152,0.000002641712,1.4984889e-7,0.0001387739,0.0038650266,0.08084829],"study_design_scores_gemma":[0.0050915233,0.0018058564,0.95463926,0.0050387834,0.000038712504,0.00000565457,0.03280381,0.00016321591,5.8036727e-7,0.00012445691,0.00018735124,0.000100789366],"about_ca_topic_score_codex":0.032327622,"about_ca_topic_score_gemma":0.00854727,"teacher_disagreement_score":0.0807475,"about_ca_system_score_codex":0.0020322008,"about_ca_system_score_gemma":0.0005750117,"threshold_uncertainty_score":0.9867138},"labels":[],"label_agreement":null},{"id":"W2740848311","doi":"","title":"Observational learning of a mixed model is subject to contextual interference effects","year":2014,"lang":"en","type":"article","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"","keywords":"Observational study; Observational learning; Set (abstract data type); Task (project management); Computer science; Psychology; Action (physics); Cognitive psychology; Artificial intelligence; Mathematics education; Medicine; Experiential learning; Engineering","score_opus":0.058206272390923346,"score_gpt":0.3458972803209503,"score_spread":0.28769100793002694,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2740848311","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.87672675,0.000015602378,0.11711391,0.0011026345,0.00006517357,0.00013659085,0.0000014970814,0.00003998695,0.0047978433],"genre_scores_gemma":[0.98841894,0.0000032366831,0.0052953153,0.0027427438,0.00004853393,0.000011551232,0.00001056218,0.000008757314,0.0034603747],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9992222,0.000041620155,0.00023457146,0.00018076038,0.00018644454,0.0001344056],"domain_scores_gemma":[0.98935467,0.010125453,0.000050439503,0.0001508754,0.00015600014,0.0001625362],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00030537232,0.000083466955,0.00028094737,0.000042375952,0.000019697522,0.0000061491974,0.000060046077,0.000058764308,0.000121857054],"category_scores_gemma":[0.06543047,0.00006428547,0.00007551386,0.000098327415,0.000034183395,0.000019640256,0.00004254017,0.00013279766,0.00008088784],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013991896,0.0009571072,0.68505126,0.00025946714,0.00020181396,0.000011073254,0.0020914709,0.0035242497,0.022616945,0.050297897,0.03720633,0.19638316],"study_design_scores_gemma":[0.0050194478,0.0041149007,0.5266678,0.004210865,0.000160019,0.00001188262,0.0001021071,0.41687834,0.039052505,0.0027173965,0.00072815316,0.0003366084],"about_ca_topic_score_codex":0.000054594682,"about_ca_topic_score_gemma":0.0000073716683,"teacher_disagreement_score":0.4133541,"about_ca_system_score_codex":0.000013740136,"about_ca_system_score_gemma":0.00006735549,"threshold_uncertainty_score":0.9424418},"labels":[],"label_agreement":null},{"id":"W2741148953","doi":"10.1097/acm.0000000000001794","title":"More on the Causes of Errors in Clinical Reasoning","year":2017,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Dalhousie University","funders":"","keywords":"Cognitive bias; Cognition; Confirmation bias; Psychology; Normative; Rationality; Cognitive psychology; Perspective (graphical); Odds; Irrational number; Scrutiny; Debiasing; Social psychology; Positive economics; Epistemology; Medicine; Law; Psychiatry; Political science; Computer science; Economics","score_opus":0.12261473758224085,"score_gpt":0.4516439406480528,"score_spread":0.32902920306581196,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2741148953","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.13910343,0.0015417681,0.0000028728957,0.8510808,0.0013194839,0.0005861746,0.000014943807,0.00003949491,0.0063110734],"genre_scores_gemma":[0.12003681,0.0044825617,0.00003344674,0.85045666,0.014961219,0.00004993825,0.000212653,0.00009023848,0.009676468],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.995347,0.00033440176,0.0019268082,0.00070681015,0.0011362865,0.00054868095],"domain_scores_gemma":[0.949768,0.0471038,0.0012118215,0.0015695652,0.0001380502,0.00020873563],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.00392735,0.00043931746,0.0019397439,0.00023956766,0.00007534147,0.0000041553394,0.00077852816,0.0025237433,0.0003956771],"category_scores_gemma":[0.3532703,0.00023696212,0.00027222716,0.00018172755,0.0017265407,0.000026222724,0.00013268695,0.011927077,0.00007751332],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011994187,0.000021986267,0.042883407,0.00009801934,0.00010380337,0.0010983023,0.00044252368,6.1150865e-7,0.0000047727603,0.00013877467,0.9533969,0.0016909457],"study_design_scores_gemma":[0.006014228,0.0019395758,0.12469978,0.11955914,0.0012799939,0.00029292933,0.00064878253,0.000088699955,0.00006160107,0.0011163148,0.7437574,0.0005415455],"about_ca_topic_score_codex":0.00042424657,"about_ca_topic_score_gemma":0.0000050723547,"teacher_disagreement_score":0.34934294,"about_ca_system_score_codex":0.00007231671,"about_ca_system_score_gemma":0.00037355607,"threshold_uncertainty_score":0.9987712},"labels":[],"label_agreement":null},{"id":"W2741786738","doi":"10.1111/medu.13396","title":"Eyeballing: the use of visual appearance to diagnose ‘sick’","year":2017,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":31,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Joseph’s Healthcare Hamilton; McMaster University","funders":"McMaster University","keywords":"Medicine; Optometry; Psychology","score_opus":0.054458923338343006,"score_gpt":0.43615718386874386,"score_spread":0.38169826053040085,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2741786738","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.92469037,0.00019560443,0.0004603965,0.07269849,0.0010600787,0.00027279195,0.0000011833862,0.000019370706,0.0006017391],"genre_scores_gemma":[0.98627937,0.00015377665,0.00046854932,0.011387146,0.0007025927,0.000043454304,0.000008904317,0.000009871658,0.00094635313],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9988576,0.00003745354,0.0002670163,0.0001769509,0.00051105325,0.00014993173],"domain_scores_gemma":[0.9959755,0.0028278988,0.00013489756,0.00058768044,0.00015790362,0.00031610063],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0004652869,0.00007624462,0.00019431088,0.000027651657,0.000150494,0.00003870301,0.00020932055,0.000093726005,0.00027656913],"category_scores_gemma":[0.30706382,0.000048068054,0.00006991112,0.000065139604,0.00021536562,0.0000512142,0.000060952043,0.00019410819,0.000101839265],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015548436,0.0024215688,0.17779315,0.000045414687,0.0000365147,0.000011768388,0.00076875393,0.000007693567,0.00006514852,0.00092313014,0.12747803,0.6902933],"study_design_scores_gemma":[0.0009419672,0.0004057514,0.8490659,0.0048805736,0.00010069786,0.00001993385,0.00011048699,0.0011684225,0.0005397932,0.0001999513,0.14243579,0.00013071232],"about_ca_topic_score_codex":0.00064801326,"about_ca_topic_score_gemma":0.00003781102,"teacher_disagreement_score":0.6901626,"about_ca_system_score_codex":0.000021763712,"about_ca_system_score_gemma":0.00079104584,"threshold_uncertainty_score":0.6987731},"labels":[],"label_agreement":null},{"id":"W2741817809","doi":"10.1097/acm.0000000000001809","title":"In Reply to Croskerry and to Patel and Bergl","year":2017,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"","keywords":"Debiasing; Hindsight bias; Cognitive bias; Heuristics; Cognition; Psychology; Commit; Confirmation bias; Cognitive psychology; Subject (documents); Social psychology; Computer science; Psychiatry","score_opus":0.05339765191851489,"score_gpt":0.40298218551198717,"score_spread":0.3495845335934723,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2741817809","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08991871,0.0028712095,0.000015368843,0.9037601,0.00076678663,0.0007033919,0.00001446838,0.00003762148,0.0019123431],"genre_scores_gemma":[0.011220022,0.0027476116,0.00014684224,0.95309585,0.00867129,0.000066747874,0.00007326138,0.000060556074,0.023917817],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9971865,0.000057634235,0.0007581416,0.0009011816,0.0005507362,0.00054584094],"domain_scores_gemma":[0.9917918,0.0064623407,0.00019788883,0.0007591483,0.00007882766,0.0007100083],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0011360503,0.0003824363,0.0012631783,0.00036099032,0.00006664756,0.000013181123,0.00023999157,0.001272417,0.00012853829],"category_scores_gemma":[0.103839666,0.0002851329,0.000039631275,0.00014452654,0.0003245296,0.000044662575,0.0002494067,0.0038289006,0.00009461373],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000092041395,0.000004532953,0.013427752,0.00013102865,0.000026913356,0.0016968232,0.0010044093,6.8461986e-8,0.000056731053,0.000011671151,0.97758746,0.0059605693],"study_design_scores_gemma":[0.002134314,0.000689105,0.038021345,0.015573643,0.00020416225,0.00042942492,0.00006146813,0.0000036011963,0.000015622809,0.00068993407,0.9418882,0.00028917682],"about_ca_topic_score_codex":0.0010393461,"about_ca_topic_score_gemma":0.00000677091,"teacher_disagreement_score":0.102703616,"about_ca_system_score_codex":0.000063734726,"about_ca_system_score_gemma":0.00011083496,"threshold_uncertainty_score":0.99996006},"labels":[],"label_agreement":null},{"id":"W2747780714","doi":"10.1177/0898010117725428","title":"Intuition in Clinical Decision Making: Differences Among Practicing Nurses","year":2017,"lang":"en","type":"article","venue":"Journal of Holistic Nursing","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":32,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Trinity College","funders":"","keywords":"Intuition; Clinical decision making; Nursing; Psychology; Clinical judgment; Patient care; Medicine; Family medicine","score_opus":0.1203274997809839,"score_gpt":0.5075878673708714,"score_spread":0.38726036758988747,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2747780714","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98738426,0.0001746118,0.0054175253,0.00085179653,0.0031737909,0.00009758695,8.4849415e-7,0.000008305773,0.0028912714],"genre_scores_gemma":[0.99446356,0.00013207331,0.0042778216,0.000103281796,0.00096987316,4.3724197e-7,6.4691585e-7,0.000013261748,0.00003904582],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99790126,0.00010165445,0.0011581433,0.00017282298,0.00044560173,0.00022052515],"domain_scores_gemma":[0.98502934,0.012759019,0.0014199063,0.00032985333,0.00024514672,0.00021674817],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0016060743,0.00012084042,0.00061062886,0.00017905154,0.00015999017,0.00015108273,0.00019327913,0.00014802859,0.00006399892],"category_scores_gemma":[0.21840374,0.000089796275,0.00019630384,0.000051461077,0.00036658678,0.00026968023,0.000024323366,0.00072818843,0.000006929977],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006231698,0.00056200195,0.68705565,0.000007983764,0.000019425772,0.00045313078,0.00021662252,0.000008209732,0.000013567566,0.0001718712,0.0005159148,0.31035244],"study_design_scores_gemma":[0.0018714094,0.00068758125,0.95262164,0.032183208,0.00021658975,0.00052134175,0.0003107762,0.0006573888,0.000015780892,0.010791736,0.000035321013,0.000087200155],"about_ca_topic_score_codex":0.000018202363,"about_ca_topic_score_gemma":0.000011660954,"teacher_disagreement_score":0.31026524,"about_ca_system_score_codex":0.00012706239,"about_ca_system_score_gemma":0.00017726148,"threshold_uncertainty_score":0.78818},"labels":[],"label_agreement":null},{"id":"W2749690026","doi":"10.1080/13814788.2017.1358709","title":"Reliability and validity of the script concordance test for postgraduate students of general practice","year":2017,"lang":"en","type":"article","venue":"European Journal of General Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":32,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Concordance; Medicine; Reliability (semiconductor); Test (biology); General practice; Medical education; Validity; Family medicine; Medical physics; Clinical psychology; Psychometrics; Internal medicine","score_opus":0.06868847330248289,"score_gpt":0.40386472420065433,"score_spread":0.33517625089817144,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2749690026","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97873867,0.00040113603,0.00039501322,0.014135602,0.00095335866,0.00034837174,0.000031061187,0.000005024899,0.0049917744],"genre_scores_gemma":[0.97300977,0.00071089057,0.022942757,0.0015755346,0.0007225845,9.966311e-7,0.0000014516377,0.00002747149,0.0010085693],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99684507,0.0010632369,0.0010145665,0.00022143699,0.0006652087,0.00019046199],"domain_scores_gemma":[0.97700596,0.016062632,0.0037917385,0.0008066517,0.0021003399,0.00023267504],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.008004397,0.00015723427,0.00050167454,0.000032494623,0.000215817,0.00008161079,0.0005341785,0.000037568167,0.000007041942],"category_scores_gemma":[0.5134595,0.000101866965,0.00022943839,0.000060347185,0.00038505284,0.0005125144,0.00024076695,0.00048192442,0.0000025369052],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.018389385,0.01133563,0.73670506,0.0005943031,0.0016003248,0.0013667379,0.0019505323,0.00035735266,0.0914873,0.00095007423,0.088735946,0.04652735],"study_design_scores_gemma":[0.0071325726,0.004439015,0.93526113,0.0013019288,0.0016861882,0.0013298076,0.00020735877,0.00019801724,0.006714258,0.00014458777,0.041405827,0.00017930784],"about_ca_topic_score_codex":0.0001067986,"about_ca_topic_score_gemma":0.0000019966403,"teacher_disagreement_score":0.50545514,"about_ca_system_score_codex":0.000034009165,"about_ca_system_score_gemma":0.00021532687,"threshold_uncertainty_score":0.49063882},"labels":[],"label_agreement":null},{"id":"W2749959514","doi":"10.1186/s12913-017-2539-y","title":"A scoping review of the potential for chart stimulated recall as a clinical research method","year":2017,"lang":"en","type":"review","venue":"BMC Health Services Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":37,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"University of Cambridge; Health Service Executive; National Institute for Health and Care Research; Health Research Board","keywords":"Medicine; Medical education; Certification; Competence (human resources); Family medicine; Systematic review; MEDLINE; Psychology; Social psychology","score_opus":0.6097883879467679,"score_gpt":0.7288125776889304,"score_spread":0.11902418974216245,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2749959514","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000010536834,0.9751823,0.000072387826,0.0018260295,0.00061888294,0.021533348,0.0001176503,0.000038211037,0.0006006442],"genre_scores_gemma":[0.0000029218954,0.9894367,0.0041361675,0.0014087815,0.0013891499,0.0014953938,0.00026396217,0.00013276887,0.0017341616],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.97419643,0.014595064,0.0043299026,0.0015247983,0.003466433,0.0018873585],"domain_scores_gemma":[0.75559485,0.22396822,0.003333989,0.0074180346,0.0071150586,0.002569842],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.09631581,0.00047236908,0.0060037947,0.0005189564,0.0010850017,0.00011363067,0.0022568542,0.0010572325,0.00023254755],"category_scores_gemma":[0.14990129,0.00028641545,0.0020770913,0.0013228356,0.0006357064,0.0000611014,0.0014125378,0.004473528,0.00028423217],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014230618,0.00025501393,0.000039672872,0.5450831,0.00008984122,0.000009246738,0.000026144764,1.3180697e-7,2.218153e-8,0.000027381971,0.0010665595,0.45326054],"study_design_scores_gemma":[0.0008822984,0.0010037895,0.00012518447,0.6572113,0.00027884636,0.00004712398,0.000030279669,0.00013787334,1.6791239e-7,0.00009270126,0.3400584,0.00013202621],"about_ca_topic_score_codex":0.0011544925,"about_ca_topic_score_gemma":0.00011238438,"teacher_disagreement_score":0.45312852,"about_ca_system_score_codex":0.00034202228,"about_ca_system_score_gemma":0.027628135,"threshold_uncertainty_score":0.9999588},"labels":[],"label_agreement":null},{"id":"W2751599728","doi":"10.1177/0840470417716949","title":"Cognitive bias in health leaders","year":2017,"lang":"en","type":"article","venue":"Healthcare Management Forum","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Cognitive bias; Cognition; Confirmation bias; Psychology; Cognitive psychology; Public relations; Social psychology; Computer science; Management science; Political science; Economics","score_opus":0.16095774854723116,"score_gpt":0.45095727371789407,"score_spread":0.2899995251706629,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2751599728","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.22969155,0.0018100545,0.00032893877,0.7169055,0.0010046668,0.0026664229,0.000028757264,0.00018296689,0.04738116],"genre_scores_gemma":[0.9741928,0.0013236182,0.00038295842,0.02065125,0.00007669866,0.00006081041,0.000047014262,0.000021898573,0.0032429548],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99816114,0.000069948124,0.0004537346,0.000382128,0.00030419146,0.0006288834],"domain_scores_gemma":[0.997962,0.00085341826,0.00022975453,0.0005573014,0.000053888125,0.00034363547],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0007353457,0.00014546137,0.0003843014,0.00014601968,0.00030403057,0.000043837477,0.00014206172,0.00007660159,0.00006075312],"category_scores_gemma":[0.00607127,0.00013284564,0.00009233619,0.00008926887,0.00014665861,0.000059168076,0.00014266527,0.00027715438,0.00017271003],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021012653,0.00028269025,0.6092358,0.00060123374,0.00005099885,0.00034194355,0.0002734874,0.0000017196256,8.0015575e-8,0.007923093,0.011661236,0.36941758],"study_design_scores_gemma":[0.0046075205,0.0007536942,0.9587309,0.010451815,0.000030553743,0.000013237722,0.0061645727,0.00013450698,0.000005063643,0.0012747332,0.017649774,0.0001836703],"about_ca_topic_score_codex":0.002986749,"about_ca_topic_score_gemma":0.0013745663,"teacher_disagreement_score":0.74450123,"about_ca_system_score_codex":0.00013577021,"about_ca_system_score_gemma":0.00011453782,"threshold_uncertainty_score":0.72683156},"labels":[],"label_agreement":null},{"id":"W2751975655","doi":"10.1016/j.survophthal.2017.08.007","title":"Clinical decision-making: heuristics and cognitive biases for the ophthalmologist","year":2017,"lang":"en","type":"review","venue":"Survey of Ophthalmology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":37,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Dalhousie University","funders":"","keywords":"Heuristics; Compromise; Cognition; Health care; Vulnerability (computing); Clinical decision making; Cognitive bias; Process (computing); Human error; Psychology; Computer science; Risk analysis (engineering); Medicine; Cognitive psychology; Intensive care medicine; Psychiatry","score_opus":0.47449890672105394,"score_gpt":0.5812232543067353,"score_spread":0.10672434758568133,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2751975655","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0025524904,0.99130476,0.00017955521,0.000033328033,0.0015591957,0.0017923743,0.0015570156,0.000015587957,0.0010057209],"genre_scores_gemma":[0.05633737,0.94125247,0.0009204135,0.000032021566,0.00034142827,0.00013605334,0.00051230023,0.00007233047,0.00039563494],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9955159,0.0010858965,0.0018719436,0.000842201,0.00025194391,0.00043214762],"domain_scores_gemma":[0.44364637,0.5532038,0.001406149,0.0009180844,0.0006327218,0.00019282752],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0066112685,0.00049395056,0.0044154124,0.00012889106,0.0002349464,0.000035399775,0.00052249135,0.0010528122,0.00014752048],"category_scores_gemma":[0.7971795,0.00029450044,0.0009117465,0.00014581333,0.0023044143,0.000019273084,0.00035933295,0.000768421,0.000024994317],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00058876834,0.0003950189,0.043540154,0.0013768424,0.00077643443,0.0012207101,0.00000410843,1.0319324e-7,9.376235e-10,0.000021467198,0.0010110656,0.9510653],"study_design_scores_gemma":[0.0028094663,0.005478851,0.6779773,0.1879763,0.009720788,0.044078097,0.000020872962,0.000036178095,5.9361017e-8,0.00084591244,0.07028545,0.00077070086],"about_ca_topic_score_codex":0.00020930366,"about_ca_topic_score_gemma":0.000005424453,"teacher_disagreement_score":0.9502946,"about_ca_system_score_codex":0.00001984231,"about_ca_system_score_gemma":0.0009222032,"threshold_uncertainty_score":0.9999507},"labels":[],"label_agreement":null},{"id":"W2752029368","doi":"10.1051/pmed/2017010","title":"Vers une perspective sociocognitiviste de l’étude du raisonnement clinique pour mieux capter l’expertise. Un éclairage particulier à partir de l’étude du raisonnement en médecine d’urgence","year":2016,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Humanities; Philosophy","score_opus":0.03307408580536836,"score_gpt":0.3393827520165277,"score_spread":0.3063086662111593,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2752029368","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.67789215,0.007262561,0.022273393,0.28662974,0.0016798163,0.0013212867,0.00017278724,0.00027039964,0.0024978516],"genre_scores_gemma":[0.9619161,0.016378907,0.0026605295,0.009298928,0.0041235406,0.00064859516,0.00004084084,0.00018232949,0.0047502313],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99103004,0.0012325918,0.0018694388,0.0017501714,0.0012771611,0.0028406258],"domain_scores_gemma":[0.9746744,0.02024606,0.00056125823,0.0013705787,0.00065652916,0.002491142],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0035101776,0.0010682265,0.001543381,0.00018523385,0.00058988,0.00013438809,0.0008396606,0.0008829838,0.00819411],"category_scores_gemma":[0.07775941,0.0008412601,0.0009394305,0.00051723566,0.001980524,0.00022435913,0.00064837193,0.0010575476,0.002197023],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0028308968,0.022964064,0.51177526,0.00040546648,0.0035652195,0.007138181,0.069991685,0.00017734009,0.017193414,0.010617763,0.11290358,0.24043714],"study_design_scores_gemma":[0.05072484,0.0050042486,0.6718899,0.019824792,0.004986652,0.0005788575,0.022744086,0.0036742329,0.04072005,0.0114792315,0.16451807,0.0038550254],"about_ca_topic_score_codex":0.001755646,"about_ca_topic_score_gemma":0.00031000224,"teacher_disagreement_score":0.2840239,"about_ca_system_score_codex":0.0028841617,"about_ca_system_score_gemma":0.0019189301,"threshold_uncertainty_score":0.99940383},"labels":[],"label_agreement":null},{"id":"W2752121763","doi":"10.3138/jvme.1115-184r4","title":"Using the Virtual World of Second Life in Veterinary Medicine: Student and Faculty Perceptions","year":2017,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Facilitator; Likert scale; Medical education; Perception; Curriculum; Psychology; Clinical Practice; Faculty development; Medicine; Professional development; Pedagogy; Nursing","score_opus":0.1907355652815787,"score_gpt":0.5131976684139734,"score_spread":0.3224621031323947,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2752121763","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.987459,0.000508579,0.000032750155,0.010335616,0.0010009011,0.0001190521,0.0000021891904,0.0000026688613,0.0005392203],"genre_scores_gemma":[0.9974172,0.00031676813,0.00033310577,0.00092276605,0.0008043628,0.0000035766334,0.0000037775103,0.000008949105,0.00018946952],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982705,0.00012523316,0.00075290113,0.00012232529,0.0005997604,0.00012924973],"domain_scores_gemma":[0.9969512,0.0015577102,0.000593682,0.00028450842,0.00018897007,0.0004239056],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001353344,0.000106065134,0.00041484117,0.0001572936,0.00012224236,0.00001923655,0.0002386933,0.000088464854,0.0008125785],"category_scores_gemma":[0.032662798,0.0000646658,0.000082484825,0.0000820499,0.000504431,0.00012300814,0.00008396252,0.00046974487,0.0000014211283],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002687577,0.0061301724,0.7274791,0.0003769055,0.0003789596,0.0008713587,0.014054806,0.000024297953,0.0131778745,0.00044933523,0.021339271,0.21303032],"study_design_scores_gemma":[0.0019188086,0.0027806635,0.98266166,0.00454237,0.00010505237,0.0018516994,0.0032575685,0.00012994002,0.000006185106,0.00006589071,0.0026155761,0.0000646007],"about_ca_topic_score_codex":0.00009229923,"about_ca_topic_score_gemma":0.000018815257,"teacher_disagreement_score":0.25518253,"about_ca_system_score_codex":0.000066370856,"about_ca_system_score_gemma":0.0009660525,"threshold_uncertainty_score":0.9754855},"labels":[],"label_agreement":null},{"id":"W2752659514","doi":"10.1051/pmed/2017011","title":"Troisième Conférence internationale de Montréal sur le raisonnement clinique : une pluralité de perspectives pour des enjeux partagés","year":2016,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McGill University; Concordia University; Université de Sherbrooke; Université de Montréal","funders":"","keywords":"Humanities; Philosophy","score_opus":0.08225692620621279,"score_gpt":0.35879236301439077,"score_spread":0.276535436808178,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2752659514","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.688342,0.076181985,0.033640306,0.19547597,0.00091270165,0.0006531188,0.0002068867,0.00022136574,0.0043657],"genre_scores_gemma":[0.84244424,0.0281448,0.007142485,0.0020098847,0.0017553908,0.00017206473,0.000020893158,0.00006979003,0.11824045],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99622804,0.00043646395,0.0008598366,0.00073589984,0.0005567431,0.0011830038],"domain_scores_gemma":[0.9756443,0.02203906,0.0002607501,0.0005304203,0.0004909306,0.0010345458],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0016226615,0.00039647377,0.0006795995,0.00009688109,0.00019933716,0.000064761436,0.0003981098,0.0003716419,0.002440724],"category_scores_gemma":[0.084318995,0.00031067865,0.00040318083,0.00021201979,0.0012780899,0.00016386904,0.00018144856,0.0004650911,0.0006209256],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007241949,0.0057247533,0.2587143,0.0001845083,0.00063965295,0.0008702617,0.008790034,0.0000814548,0.005476045,0.03343859,0.05593303,0.6294232],"study_design_scores_gemma":[0.02517221,0.0024585933,0.4787789,0.027137622,0.0008864349,0.00050050777,0.01627482,0.003959955,0.014973561,0.043166,0.38501242,0.0016790008],"about_ca_topic_score_codex":0.011403707,"about_ca_topic_score_gemma":0.0014899631,"teacher_disagreement_score":0.6277442,"about_ca_system_score_codex":0.0007607538,"about_ca_system_score_gemma":0.0024016646,"threshold_uncertainty_score":0.99993455},"labels":[],"label_agreement":null},{"id":"W2752862788","doi":"","title":"84. Pathological familiarity and false recognition: A case study","year":2001,"lang":"en","type":"article","venue":"Brain and Cognition","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Psychology; Pathological; Cognitive psychology; Developmental psychology","score_opus":0.0723001318875849,"score_gpt":0.3560221237145963,"score_spread":0.2837219918270114,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2752862788","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99567115,0.00021306664,0.00009738192,0.001429756,0.000044623208,0.00039128918,0.000013064977,0.00006600569,0.0020736754],"genre_scores_gemma":[0.9965287,0.0003031278,0.00018189986,0.002621447,0.00012459664,0.00003855615,0.000048046288,0.0000064740043,0.00014715495],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99923265,0.00009096224,0.00017734399,0.00026472466,0.00010735767,0.00012696764],"domain_scores_gemma":[0.9971764,0.002444965,0.000032416636,0.00008591041,0.00007812567,0.00018214856],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00035889813,0.00009563404,0.00018620836,0.000037728285,0.00010378248,0.000027154287,0.000010644074,0.00009279638,0.0000977867],"category_scores_gemma":[0.009360172,0.0000771144,0.000031185657,0.00010196544,0.00009431074,0.00003811699,0.000026473253,0.00015084216,0.00002529805],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00042789103,0.001995025,0.21626718,0.000023609628,0.000040886312,0.031090094,0.00068000704,2.645271e-8,0.00012582114,0.000031363976,0.00070701836,0.7486111],"study_design_scores_gemma":[0.01304486,0.0046124137,0.8837741,0.00084382895,0.000604481,0.076133884,0.012244532,0.00026697625,0.000034778917,0.0072471765,0.00078680733,0.0004061813],"about_ca_topic_score_codex":0.0000820027,"about_ca_topic_score_gemma":0.000053030428,"teacher_disagreement_score":0.7482049,"about_ca_system_score_codex":0.0000057277275,"about_ca_system_score_gemma":0.000014413784,"threshold_uncertainty_score":0.9989844},"labels":[],"label_agreement":null},{"id":"W2753445517","doi":"10.22374/cjgim.v12i2.166","title":"To Err is Human: A Case-Based Review of Cognitive Bias and its Role in Clinical Decision Making","year":2017,"lang":"en","type":"review","venue":"Canadian Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Juravinski Hospital; McMaster University","funders":"","keywords":"Medicine; Cognition; Clinical decision making; Cognitive bias; Intensive care medicine; Cognitive psychology; Psychiatry; Psychology","score_opus":0.264717265330888,"score_gpt":0.5263437014958388,"score_spread":0.26162643616495074,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2753445517","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0063543683,0.9900658,0.00003172957,0.0011459459,0.00084765983,0.00065417896,0.00008312809,0.0000023223292,0.00081487006],"genre_scores_gemma":[0.012586636,0.9795965,0.0002793314,0.005562835,0.0016064164,0.000009179858,0.000017280388,0.000048307887,0.00029348827],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9948582,0.00034308436,0.0035580555,0.00039054957,0.00047378195,0.00037634544],"domain_scores_gemma":[0.98416114,0.010143784,0.00240581,0.00039204283,0.0009002193,0.0019970085],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0038294813,0.0004076569,0.0044797924,0.0009834535,0.00006769827,0.000022257664,0.00036907813,0.00034857015,0.0005791416],"category_scores_gemma":[0.15748008,0.0002740665,0.0007161912,0.0002479693,0.00027037584,0.000043387718,0.00004975297,0.0013212209,0.000013592313],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009615126,0.00004985705,0.0015357562,0.0075804926,0.00022698432,0.022089114,0.00007745154,1.9039352e-7,1.5302382e-7,0.000011108488,0.0132310465,0.95510167],"study_design_scores_gemma":[0.0016630918,0.0020294506,0.00064862875,0.7768594,0.0019190564,0.01057535,0.00002541122,0.000012579199,6.4214373e-7,0.000027287006,0.20606588,0.00017320969],"about_ca_topic_score_codex":0.0042420677,"about_ca_topic_score_gemma":0.0057547884,"teacher_disagreement_score":0.95492846,"about_ca_system_score_codex":0.00022238783,"about_ca_system_score_gemma":0.002798677,"threshold_uncertainty_score":0.99997115},"labels":[],"label_agreement":null},{"id":"W2754163561","doi":"10.1016/j.jclinepi.2017.09.007","title":"Decision making about healthcare-related tests and diagnostic test strategies. Paper 3: a systematic review shows limitations in most tools designed to assess quality and develop recommendations","year":2017,"lang":"en","type":"review","venue":"Journal of Clinical Epidemiology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; Health Sciences Centre; University of British Columbia; Impact; McMaster University","funders":"Agency for Healthcare Research and Quality","keywords":"Operationalization; Terminology; Systematic review; Computer science; Quality (philosophy); Test (biology); Health care; Confusion; Critical appraisal; Resource (disambiguation); Management science; Data science; MEDLINE; Process management; Medicine; Psychology; Alternative medicine; Engineering","score_opus":0.7257361177039493,"score_gpt":0.6401018406470225,"score_spread":0.08563427705692683,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2754163561","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00013032209,0.9814024,0.0011213747,0.013639689,0.00074267946,0.002745101,0.000057415924,0.000016930191,0.0001440747],"genre_scores_gemma":[0.0010196896,0.9738806,0.016100807,0.008461672,0.00022514892,0.00017101002,0.00005294708,0.00004653893,0.000041574036],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.96840006,0.011536148,0.018406408,0.0007572129,0.00035145442,0.0005487083],"domain_scores_gemma":[0.055290073,0.9300543,0.011602054,0.0009099213,0.0010798336,0.0010637837],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.065556265,0.00055939803,0.016479507,0.0003395069,0.00016378665,0.00008925401,0.00043197648,0.0012041082,0.00003442971],"category_scores_gemma":[0.9878362,0.00036967997,0.0008426717,0.00048298773,0.00030256665,0.00024624105,0.00020255285,0.0025699218,0.000024019912],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004195133,0.00032342423,0.01130621,0.1410634,0.00027065162,0.00015352631,0.000022103655,0.0000015483719,2.052555e-8,0.00070928776,0.0015508549,0.844557],"study_design_scores_gemma":[0.0009545157,0.00096266106,0.088439845,0.8871738,0.002699325,0.0009239122,0.000031093103,0.0000101193045,1.6680759e-9,0.0026889367,0.015809165,0.00030663834],"about_ca_topic_score_codex":0.000034231874,"about_ca_topic_score_gemma":0.0002475685,"teacher_disagreement_score":0.92227995,"about_ca_system_score_codex":0.00017006842,"about_ca_system_score_gemma":0.0028755108,"threshold_uncertainty_score":0.9998755},"labels":[],"label_agreement":null},{"id":"W2754197083","doi":"10.1093/ageing/afx144.194","title":"193Rates of Refusal of Cognitive Screening Tests in Older Medical Inpatients","year":2017,"lang":"en","type":"article","venue":"Age and Ageing","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Medicine; Cognition; Cognitive impairment; Psychiatry; Family medicine; Gerontology","score_opus":0.03288262370633362,"score_gpt":0.3701295457405744,"score_spread":0.3372469220342408,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2754197083","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9972271,0.00016778002,0.000043864824,0.00017351007,0.000036037836,0.000086591856,0.0000034245077,0.0000050416165,0.002256682],"genre_scores_gemma":[0.9994243,0.00009403494,0.00021325453,0.0001414803,0.000043488984,0.0000024773565,0.000010087363,0.0000056786566,0.00006524353],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9992822,0.000020345082,0.00025453244,0.00012574409,0.00020654981,0.00011064278],"domain_scores_gemma":[0.9960713,0.0035115008,0.00014132725,0.00012444714,0.000051018724,0.0001004169],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0002639378,0.00006043268,0.00026423568,0.000043374424,0.00004069931,0.000008003386,0.00005531286,0.000077005694,0.000044298286],"category_scores_gemma":[0.060470887,0.00004756495,0.000032824995,0.00003611188,0.00018181339,0.000042866606,0.000078222336,0.00013627396,0.0000011231793],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000090155816,0.00022751479,0.8190454,0.00005074676,0.00002667728,0.003301585,0.0006129554,2.150867e-7,0.00028569202,0.000035781337,0.000036287536,0.17628694],"study_design_scores_gemma":[0.002325859,0.00018921823,0.98367554,0.012612793,0.000035633948,0.000006976251,0.00013337194,0.000098516306,0.0007805108,0.00008884659,0.000007066001,0.000045638735],"about_ca_topic_score_codex":0.00047905298,"about_ca_topic_score_gemma":0.000069270296,"teacher_disagreement_score":0.17624131,"about_ca_system_score_codex":0.000002520583,"about_ca_system_score_gemma":0.000031672582,"threshold_uncertainty_score":0.9474432},"labels":[],"label_agreement":null},{"id":"W2754922575","doi":"10.1016/j.jclinepi.2017.09.003","title":"Decision making about healthcare-related tests and diagnostic test strategies. Paper 2: a review of methodological and practical challenges","year":2017,"lang":"en","type":"review","venue":"Journal of Clinical Epidemiology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":47,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; McMaster University; Victoria Hospital; Impact","funders":"","keywords":"Health care; Test (biology); MEDLINE; Computer science; Management science; Diagnostic test; Clinical study design; Research design; Data science; Clinical trial; Medicine; Pathology; Engineering","score_opus":0.7493340506059935,"score_gpt":0.6877203522985208,"score_spread":0.06161369830747265,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2754922575","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000086446096,0.97999215,0.00019735645,0.017644295,0.0009690397,0.00074331905,0.000017726175,0.000014115651,0.0003355377],"genre_scores_gemma":[0.00035470704,0.9678396,0.027300846,0.003521465,0.0008990824,0.000016589456,0.000008904387,0.0000455213,0.000013292832],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.971346,0.012860612,0.013890654,0.0008682092,0.0004609125,0.00057356735],"domain_scores_gemma":[0.0748137,0.90707564,0.0154538695,0.00087682955,0.0007157782,0.0010641946],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","research_integrity"],"consensus_categories":["metaresearch","research_integrity"],"category_scores_codex":[0.08215181,0.0006092048,0.020716202,0.00018282072,0.00008669477,0.00001862852,0.0003053229,0.002411277,0.00010428269],"category_scores_gemma":[0.9875231,0.00034773143,0.0021125819,0.00012382741,0.0014519747,0.00014048353,0.00031910403,0.0046141525,0.000011740266],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014861833,0.00045055765,0.0041987165,0.03470726,0.0004197893,0.00084502314,0.0000065533472,8.645178e-8,1.035454e-8,0.0019156141,0.0014895011,0.9558183],"study_design_scores_gemma":[0.0011219946,0.0040338677,0.0543979,0.64446604,0.0040654326,0.009467812,0.000009854522,0.000005430207,2.8763647e-9,0.010921806,0.27127486,0.00023501896],"about_ca_topic_score_codex":0.000010301964,"about_ca_topic_score_gemma":0.000005163399,"teacher_disagreement_score":0.9555833,"about_ca_system_score_codex":0.000045296358,"about_ca_system_score_gemma":0.0018725827,"threshold_uncertainty_score":0.9998975},"labels":[],"label_agreement":null},{"id":"W2758973470","doi":"10.1111/jep.12831","title":"Reasoning, evidence, and clinical decision‐making: The great debate moves forward","year":2017,"lang":"en","type":"editorial","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Rehabilitation Institute; University of Toronto; University Health Network; Dalhousie University","funders":"","keywords":"Context (archaeology); Rationality; Epistemology; Value (mathematics); Psychology; Engineering ethics; Clinical Practice; Evidence-based medicine; Health care; MEDLINE; Medicine; Political science; Law; Computer science; Nursing; History","score_opus":0.1266264399672913,"score_gpt":0.5604650979883113,"score_spread":0.43383865802102006,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2758973470","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.017490728,0.020155197,0.000596796,0.0156231895,0.9397397,0.0014833679,0.000016673763,0.00002391105,0.004870457],"genre_scores_gemma":[0.032643322,0.08907231,0.017137272,0.0023598555,0.8577465,0.00003060995,0.000021293361,0.00010381553,0.0008850661],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9814251,0.0048406767,0.007205047,0.0008612084,0.005227159,0.00044079078],"domain_scores_gemma":[0.0647583,0.9104295,0.01339697,0.0018481556,0.0088296635,0.0007374633],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["metaresearch","research_integrity"],"category_scores_codex":[0.12636912,0.00044582356,0.0022345076,0.0002520194,0.0002719033,0.00041580093,0.0007266461,0.0021062528,0.00024120962],"category_scores_gemma":[0.98857313,0.00027455046,0.001118079,0.00023030084,0.00078163657,0.00091337605,0.00032984672,0.007826221,0.000060769646],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.005080617,0.00073183124,0.024026027,0.00003057029,0.0004714146,0.00041968597,0.000048221733,0.000018634186,3.218683e-8,0.00001896225,0.79938793,0.16976607],"study_design_scores_gemma":[0.008905801,0.0040882137,0.1428232,0.03806237,0.008544505,0.00056218705,0.0001235991,0.0023215371,1.3328372e-7,0.004182991,0.79002726,0.00035821335],"about_ca_topic_score_codex":0.000050938783,"about_ca_topic_score_gemma":0.000053199383,"teacher_disagreement_score":0.9166668,"about_ca_system_score_codex":0.00026443187,"about_ca_system_score_gemma":0.005275153,"threshold_uncertainty_score":0.9999707},"labels":[],"label_agreement":null},{"id":"W2761814284","doi":"10.1093/pch/8.suppl_b.49b","title":"Evaluation Of An Assessment-Driven Algorithm To Reduce The Lenght Of Hospitalization In Hospitalized Asthma Patient","year":2003,"lang":"en","type":"article","venue":"Paediatrics & Child Health","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hospital for Sick Children","funders":"","keywords":"Asthma; Medicine; Pediatrics; Intensive care medicine; Algorithm; Computer science; Internal medicine","score_opus":0.02355474948784214,"score_gpt":0.38430751960901466,"score_spread":0.3607527701211725,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2761814284","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9905047,0.0012828673,0.0019511376,0.0029125097,0.00038321718,0.0017485988,0.000034501496,0.000017778073,0.0011647133],"genre_scores_gemma":[0.9867937,0.0006321092,0.011635114,0.00066478824,0.00010514467,0.00005084437,0.00009501354,0.000019406614,0.0000038930466],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99668896,0.00070033077,0.0008664374,0.00029428586,0.0011864117,0.00026355422],"domain_scores_gemma":[0.9972215,0.0010738256,0.00045208586,0.00041022082,0.00060284964,0.00023950996],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0041246815,0.00013009107,0.00041618964,0.00018681634,0.00006284808,0.0000074584086,0.00009157993,0.00009201093,0.000037019894],"category_scores_gemma":[0.018158088,0.00009944484,0.00008495077,0.0008253489,0.00003706194,0.00005237284,0.000023733026,0.00020859584,0.0000028094764],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007892034,0.0038726472,0.29930383,0.00013302156,0.00006935543,0.000004403953,0.004872344,0.0044923676,0.000015082416,0.0026108513,0.001288941,0.68325824],"study_design_scores_gemma":[0.0036507545,0.0027702635,0.9808724,0.00053911307,0.00014375971,0.000006010898,0.0004457109,0.010275096,0.00014946176,0.00073037617,0.0002788771,0.00013817116],"about_ca_topic_score_codex":0.0003088927,"about_ca_topic_score_gemma":0.000030579038,"teacher_disagreement_score":0.6831201,"about_ca_system_score_codex":0.00025866367,"about_ca_system_score_gemma":0.0011146419,"threshold_uncertainty_score":0.99011236},"labels":[],"label_agreement":null},{"id":"W2763889301","doi":"10.1093/pch/9.suppl_a.50a","title":"102 Self-Directed Learning Versus a Structured Cme Course to Assess Physicians' Knowledge of Sedation","year":2004,"lang":"en","type":"article","venue":"Paediatrics & Child Health","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Sedation; Medicine; Test (biology); Multiple choice; Curriculum; Focus group; Family medicine; Medical education; Psychology; Anesthesia; Significant difference; Pedagogy; Internal medicine","score_opus":0.0265728899991191,"score_gpt":0.36233762062782593,"score_spread":0.3357647306287068,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2763889301","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98054904,0.002868669,0.0010257256,0.008932087,0.0017859461,0.001163731,0.000049360853,0.00092057924,0.0027048907],"genre_scores_gemma":[0.98899806,0.0007980173,0.007843613,0.0009610852,0.0011139776,0.000013913289,0.00017929246,0.000040336574,0.0000517369],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99810714,0.0001284163,0.00056506705,0.00038519318,0.0003646979,0.00044947633],"domain_scores_gemma":[0.99571824,0.0028546948,0.00033929557,0.0003035519,0.0003181491,0.00046606807],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00059130805,0.00019941844,0.0005820484,0.00017091255,0.00016387746,0.000014412239,0.000109499924,0.00014627299,0.000024188026],"category_scores_gemma":[0.012383505,0.00019151147,0.00014080679,0.0010142558,0.000030250176,0.00004408886,0.000047092853,0.0005537701,0.00006628759],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004665547,0.015455024,0.40115184,0.0026972943,0.0012274989,0.00008853836,0.042816598,0.012344161,0.00023403399,0.007889873,0.035530433,0.47589916],"study_design_scores_gemma":[0.017657354,0.0043447497,0.96555364,0.0022334585,0.0006217209,0.000018824408,0.0008041271,0.0011310152,0.0002686966,0.00025198905,0.0065746317,0.0005398097],"about_ca_topic_score_codex":0.00023221786,"about_ca_topic_score_gemma":0.0001079519,"teacher_disagreement_score":0.5644018,"about_ca_system_score_codex":0.00030066265,"about_ca_system_score_gemma":0.0011164151,"threshold_uncertainty_score":0.9959356},"labels":[],"label_agreement":null},{"id":"W2765208643","doi":"10.1016/j.cjca.2017.10.018","title":"Avoiding Clinical Errors With Bedside Echocardiography: A Randomized Clinical Study","year":2017,"lang":"en","type":"article","venue":"Canadian Journal of Cardiology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"McMaster University; McMaster University Medical Centre; St. Michael's Hospital; Mount Sinai Hospital; University Health Network; University of Toronto; Ted Rogers Centre for Heart Research","funders":"","keywords":"Medicine; Odds ratio; Confidence interval; Diagnostic accuracy; Clinical history; Diagnostic test; Medical history; Concordance; Internal medicine; Radiology; Emergency medicine","score_opus":0.06385024251420539,"score_gpt":0.4007800564671011,"score_spread":0.3369298139528957,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2765208643","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98262566,0.00055387773,0.0019519591,0.0021756766,0.00279248,0.0006436429,0.000008573666,0.000012630323,0.009235493],"genre_scores_gemma":[0.9947549,0.00040802042,0.00059617485,0.00054404896,0.0035677545,0.000009133052,0.000002996629,0.00003137236,0.00008556215],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9941657,0.002838287,0.0018351271,0.00036616466,0.00028680897,0.0005079212],"domain_scores_gemma":[0.9737527,0.021477558,0.0011388878,0.000881028,0.0006333988,0.0021164631],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0198532,0.0002360597,0.0033272319,0.00031944632,0.0004939512,0.000077495824,0.00046628522,0.0003422176,0.000020670728],"category_scores_gemma":[0.24062572,0.00015656039,0.0017097472,0.00010635534,0.0018269315,0.00008599194,0.000035199962,0.0019404856,0.000018038889],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.013686743,0.00003448152,0.96145034,0.0000031730024,0.0025698706,0.0052245385,0.000078121986,0.000068235844,2.2310387e-7,0.000054601358,0.0052096327,0.011620052],"study_design_scores_gemma":[0.13758421,0.00297512,0.8469005,0.0006204846,0.0024995029,0.0017119964,0.00053083454,0.000017829558,9.53058e-7,0.00027302996,0.006695205,0.0001902828],"about_ca_topic_score_codex":0.0014376384,"about_ca_topic_score_gemma":0.0012381487,"teacher_disagreement_score":0.22077252,"about_ca_system_score_codex":0.000055841283,"about_ca_system_score_gemma":0.002809788,"threshold_uncertainty_score":0.8430552},"labels":[],"label_agreement":null},{"id":"W2765374369","doi":"10.2196/mededu.8100","title":"A Clinical Reasoning Tool for Virtual Patients: Design-Based Research Study","year":2017,"lang":"en","type":"article","venue":"JMIR Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":60,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"European Commission","keywords":"Computer science; Artificial intelligence","score_opus":0.15694378289530433,"score_gpt":0.54538772508619,"score_spread":0.38844394219088574,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2765374369","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9852098,0.00002707545,0.0019870545,0.007240879,0.0016685388,0.0032113264,0.000003837489,0.000060203438,0.0005912898],"genre_scores_gemma":[0.9904886,0.000012987114,0.0033655085,0.0018743165,0.0016455393,0.0014269018,0.00008424854,0.000030793162,0.0010711114],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99646074,0.00043349754,0.0007320419,0.0005603588,0.001399618,0.00041373243],"domain_scores_gemma":[0.9700786,0.02613401,0.00029357822,0.0012664588,0.0012638606,0.0009634956],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.007316101,0.00014650828,0.00041744387,0.00011199329,0.0005504081,0.000110240566,0.000393077,0.00031849608,0.00039446514],"category_scores_gemma":[0.60347855,0.00011682497,0.00016770078,0.00011388806,0.00040660988,0.00008404221,0.00008796113,0.0007295969,0.00009059461],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008587133,0.012922835,0.57544357,0.00001751915,0.000033294033,0.000006618767,0.00022284125,4.566194e-7,7.472151e-7,0.00017744927,0.06898285,0.34133312],"study_design_scores_gemma":[0.008963059,0.00619632,0.97430634,0.0021931273,0.00009068598,0.0000021190965,0.00047222277,0.0016757048,0.000009618084,0.00025389376,0.005676585,0.0001603],"about_ca_topic_score_codex":0.00008229676,"about_ca_topic_score_gemma":0.000008425436,"teacher_disagreement_score":0.59616244,"about_ca_system_score_codex":0.00011734499,"about_ca_system_score_gemma":0.004835688,"threshold_uncertainty_score":0.8578303},"labels":[],"label_agreement":null},{"id":"W2765463029","doi":"10.1503/jpn.170194","title":"The value of a skeptical approach to neurosciences in psychiatric training and practice","year":2017,"lang":"en","type":"editorial","venue":"Journal of Psychiatry and Neuroscience","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Douglas Mental Health University Institute","funders":"McGill University","keywords":"Skepticism; Value (mathematics); Psychology; Psychiatry; Medicine; Computer science; Philosophy; Epistemology; Machine learning","score_opus":0.04164125239608706,"score_gpt":0.3876738325724239,"score_spread":0.3460325801763368,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2765463029","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.043891802,0.0038036513,0.00026188747,0.014458217,0.9346717,0.0003746619,0.000014743575,0.000007093105,0.0025162343],"genre_scores_gemma":[0.29961222,0.031939905,0.037442796,0.0041331127,0.62568635,0.000020437712,0.000002207642,0.00009443711,0.0010685151],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9969962,0.00022991093,0.0008684092,0.00044118884,0.0011416266,0.00032265525],"domain_scores_gemma":[0.9811888,0.016711868,0.0011197743,0.00033987462,0.00021995415,0.00041976932],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0040781787,0.00020187149,0.00063284417,0.00022868246,0.00030914327,0.0001854163,0.0005306561,0.00023027479,3.2301674e-7],"category_scores_gemma":[0.23813282,0.00012256154,0.00012677407,0.0003802545,0.0008293348,0.00020870026,0.000114754694,0.001269831,2.5104163e-7],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.005663888,0.005447076,0.027706202,0.000720217,0.00006044992,0.00030146496,0.0019292108,0.00018569554,0.0009093066,0.0076149176,0.9194736,0.029987967],"study_design_scores_gemma":[0.007982733,0.022252472,0.21037106,0.0086908415,0.0012880644,0.0060069542,0.0024303626,0.0014007811,0.000015989446,0.0038773078,0.73470026,0.0009832002],"about_ca_topic_score_codex":0.000014943512,"about_ca_topic_score_gemma":0.000005625498,"teacher_disagreement_score":0.30898535,"about_ca_system_score_codex":0.000012015524,"about_ca_system_score_gemma":0.001553355,"threshold_uncertainty_score":0.76828474},"labels":[],"label_agreement":null},{"id":"W2765472514","doi":"10.3138/jvme.1016-153r","title":"Clinical Reasoning by Veterinary Students in the First-Opinion Setting: Is It Encouraged? Is It Practiced?","year":2017,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Preceptor; Promotion (chess); Medical education; Scale (ratio); Psychology; Critical thinking; Medicine; Pedagogy","score_opus":0.14152396698406988,"score_gpt":0.5301250776628448,"score_spread":0.3886011106787749,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2765472514","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.80124366,0.0005758396,0.00003233749,0.19395404,0.002899201,0.00022860504,0.000003327352,0.0000062832887,0.0010567028],"genre_scores_gemma":[0.9560298,0.0038303104,0.0008321545,0.036674425,0.0021483034,0.000018690085,0.000013471826,0.000023537856,0.00042933665],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99552083,0.00045040637,0.0015579531,0.00034206585,0.001795755,0.00033297107],"domain_scores_gemma":[0.98854977,0.007863219,0.0017717037,0.0008393636,0.00032958796,0.00064636033],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0075671226,0.00022339124,0.00058233977,0.00010997493,0.00032035698,0.00017752305,0.0009880265,0.0003556758,0.00087759946],"category_scores_gemma":[0.14002097,0.00015475604,0.0003038167,0.000109380955,0.0002505406,0.00039847533,0.00019372108,0.0016223458,0.000057698595],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011697194,0.004845531,0.18796459,0.00009678176,0.00012486022,0.00074048346,0.004144906,2.2471292e-7,0.000038587867,0.000008043047,0.7314748,0.069391444],"study_design_scores_gemma":[0.003730982,0.0066237324,0.57182676,0.010944294,0.00017756732,0.0061027184,0.0032448908,0.00011907836,0.000010225756,0.00005608248,0.39695218,0.00021146539],"about_ca_topic_score_codex":0.000114729846,"about_ca_topic_score_gemma":0.0000032239705,"teacher_disagreement_score":0.3838622,"about_ca_system_score_codex":0.00013083025,"about_ca_system_score_gemma":0.001441313,"threshold_uncertainty_score":0.9609099},"labels":[],"label_agreement":null},{"id":"W2765760314","doi":"10.1016/j.ijpsycho.2017.10.010","title":"The application of reward learning in the real world: Changes in the reward positivity amplitude reflect learning in a medical education context","year":2017,"lang":"en","type":"article","venue":"International Journal of Psychophysiology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary; University of Victoria","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Context (archaeology); Psychology; Cognitive psychology; Geography","score_opus":0.028232134295214668,"score_gpt":0.44284586767958584,"score_spread":0.4146137333843712,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2765760314","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.92101336,0.00018590699,0.00006794056,0.07629193,0.00052975345,0.00017546177,6.589599e-7,0.0000033428691,0.0017316275],"genre_scores_gemma":[0.9961751,0.0011901507,0.000053781478,0.0018189619,0.00063835096,0.00003147064,0.000009173204,0.0000067049027,0.00007628974],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9978747,0.0006673391,0.00058506656,0.00015393028,0.000561873,0.00015710096],"domain_scores_gemma":[0.99121624,0.0074677924,0.0007653852,0.00025278734,0.00025216755,0.000045650926],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0032748415,0.00009250369,0.00028200814,0.0001980387,0.00009195881,0.000038367656,0.0007771997,0.000097369964,0.000012650348],"category_scores_gemma":[0.025119603,0.00004926218,0.00009788708,0.0001392471,0.00022608216,0.000054921842,0.000047544596,0.0010689183,0.0000039895335],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0029501817,0.0011927912,0.50563776,0.000012394478,0.000101697384,0.00010273048,0.003731822,0.000117456766,0.0034707526,0.0042792954,0.00088044856,0.4775227],"study_design_scores_gemma":[0.0015377998,0.0004252619,0.98834974,0.0009241702,0.000016182288,0.000102532555,0.0010116382,0.00015324232,0.000034712608,0.0038807131,0.0035144475,0.000049539463],"about_ca_topic_score_codex":0.0013458398,"about_ca_topic_score_gemma":0.005380368,"teacher_disagreement_score":0.48271203,"about_ca_system_score_codex":0.000087217595,"about_ca_system_score_gemma":0.00031230281,"threshold_uncertainty_score":0.98309225},"labels":[],"label_agreement":null},{"id":"W2766908455","doi":"10.1016/j.jval.2017.08.046","title":"Cost-Effectiveness of Diagnosis: from Tests to Long-Term Outcomes and Costs","year":2017,"lang":"en","type":"article","venue":"Value in Health","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Watchful waiting; Context (archaeology); Cost effectiveness; Medicine; Decision tree; Cost–benefit analysis; Incremental cost-effectiveness ratio; Cost database; Prostate cancer; Computer science; Cost-effectiveness analysis; Machine learning; Risk analysis (engineering); Cancer","score_opus":0.14750244572191867,"score_gpt":0.4600771578112557,"score_spread":0.312574712089337,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2766908455","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9939149,0.00089238567,0.00003409461,0.003527185,0.00029839782,0.0011615762,0.000048378362,0.000018581688,0.000104468825],"genre_scores_gemma":[0.9979613,0.00064931554,0.0004155564,0.0007402883,0.000058664264,0.00013087486,0.000014162545,0.000014865985,0.00001498094],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9986944,0.00018648637,0.0003842558,0.0002978356,0.00019297306,0.00024404185],"domain_scores_gemma":[0.97807014,0.020779958,0.00017645248,0.0005438822,0.000053882122,0.00037568514],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010376868,0.000119221855,0.0006139211,0.00007156592,0.000082958395,0.000021645326,0.00011807223,0.0000815414,0.000015331356],"category_scores_gemma":[0.055165313,0.000097670025,0.000051101382,0.000054729502,0.00008565515,0.00003821207,0.0000878211,0.00015469838,0.000015785745],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013764924,0.00030361212,0.9472646,0.00014711243,0.00001867969,0.000039687326,0.00011568794,0.00001632373,0.000010162624,0.00008424057,0.00006309134,0.05179916],"study_design_scores_gemma":[0.0024477788,0.00048718767,0.9834509,0.013053768,0.000021043235,0.0000036480385,0.000009941573,0.000037045207,0.0002200692,0.0001690798,0.000019626927,0.00007993363],"about_ca_topic_score_codex":0.019055499,"about_ca_topic_score_gemma":0.00033386867,"teacher_disagreement_score":0.054127626,"about_ca_system_score_codex":0.00013215553,"about_ca_system_score_gemma":0.00018330781,"threshold_uncertainty_score":0.9874767},"labels":[],"label_agreement":null},{"id":"W2766931373","doi":"10.1007/s10459-017-9799-5","title":"Generalization and the qualitative–quantitative debate","year":2017,"lang":"en","type":"editorial","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Generalization; Qualitative research; Psychology; Mathematics education; Epistemology; Computer science; Sociology; Social science; Philosophy","score_opus":0.0516131661214722,"score_gpt":0.5374798819452475,"score_spread":0.48586671582377533,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2766931373","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0022702855,0.1109019,0.0004614549,0.01592717,0.8648981,0.0012894355,0.000026671343,0.000028531227,0.0041964487],"genre_scores_gemma":[0.017389346,0.4270777,0.022228349,0.0034841686,0.5237603,0.0007092702,0.00079134665,0.000053647844,0.0045058494],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.997569,0.0004986535,0.000584158,0.00048417342,0.00059874606,0.0002652714],"domain_scores_gemma":[0.95239586,0.045272022,0.0012760698,0.00044178974,0.00041598186,0.00019828968],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005160922,0.00016095974,0.00048786862,0.00014222678,0.0005020511,0.00009230865,0.00021728479,0.00016104747,0.000008511257],"category_scores_gemma":[0.14625818,0.00010088973,0.00004356738,0.00024496356,0.0015564706,0.00029563895,0.000038111706,0.00038934732,0.0000069416265],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005815418,0.00031396424,0.0024640823,0.0009446677,0.000016553673,0.0000016490819,0.01945995,0.00013283505,2.5642927e-7,0.0455751,0.8278628,0.10264663],"study_design_scores_gemma":[0.002142554,0.00092347816,0.0014094175,0.009764855,0.000040496518,0.000004919214,0.004378155,0.0008196408,0.0000015348645,0.036481127,0.94379056,0.0002432407],"about_ca_topic_score_codex":0.0019247305,"about_ca_topic_score_gemma":0.00060970325,"teacher_disagreement_score":0.3411378,"about_ca_system_score_codex":0.00013150566,"about_ca_system_score_gemma":0.0040852022,"threshold_uncertainty_score":0.86093324},"labels":[],"label_agreement":null},{"id":"W2769043608","doi":"10.1108/lhs-02-2017-0010","title":"Leadership in crisis situations: merging the interdisciplinary silos","year":2017,"lang":"en","type":"article","venue":"Leadership in health services","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":28,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McGill University; Montreal Children's Hospital; McGill University Health Centre; Université de Montréal; Centre Hospitalier Universitaire Sainte-Justine","funders":"","keywords":"CLARITY; Distributed leadership; Context (archaeology); Originality; Negotiation; Public relations; Psychology; Value (mathematics); Qualitative research; Shared leadership; Social psychology; Political science; Transactional leadership; Sociology; Computer science","score_opus":0.2859703605735468,"score_gpt":0.4315939895336905,"score_spread":0.1456236289601437,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2769043608","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.21156925,0.0055243974,0.000025531072,0.7794542,0.00054963795,0.00090342626,0.0000053592325,0.00008619749,0.0018820107],"genre_scores_gemma":[0.9649601,0.0002451869,0.00024276873,0.03394063,0.0002813094,0.00007781097,0.00001678289,0.00002592186,0.00020951591],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9973897,0.00023090736,0.0007401206,0.00044228972,0.00043295746,0.0007640544],"domain_scores_gemma":[0.99333435,0.005032628,0.00037024575,0.0009549756,0.00007250166,0.00023529568],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0025788085,0.00020220467,0.00047229393,0.0001864045,0.00044891913,0.00010208451,0.0005129434,0.00016715797,0.0000800794],"category_scores_gemma":[0.0040060375,0.00015001054,0.00010320398,0.00022673613,0.00016104263,0.00026426825,0.00018182075,0.00074660993,0.00010856557],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008854387,0.0012989694,0.8215395,0.014196787,0.00008631219,0.00029537902,0.110880636,0.0003875909,0.000022477998,0.0019529698,0.0021030118,0.046350908],"study_design_scores_gemma":[0.002453317,0.00031942618,0.74071383,0.04116586,0.00003460751,0.000024034221,0.20747586,0.0033081572,0.000031078966,0.003994867,0.00022715834,0.0002517934],"about_ca_topic_score_codex":0.004409089,"about_ca_topic_score_gemma":0.01229513,"teacher_disagreement_score":0.75339085,"about_ca_system_score_codex":0.00020500201,"about_ca_system_score_gemma":0.00024089885,"threshold_uncertainty_score":0.6860971},"labels":[],"label_agreement":null},{"id":"W2769315238","doi":"10.1097/acm.0000000000002081","title":"Managing Multiplicity: Conceptualizing Physician Cognition in Multipatient Environments","year":2017,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":32,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia; McMaster University","funders":"","keywords":"Think aloud protocol; Cognition; Protocol analysis; Construct (python library); Conceptual framework; Psychology; Prioritization; Qualitative research; Protocol (science); Medical education; Applied psychology; Medicine; Computer science; Human–computer interaction; Management science; Alternative medicine","score_opus":0.06770001074902376,"score_gpt":0.3855000675257742,"score_spread":0.31780005677675044,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2769315238","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9767597,0.00064149115,0.00064699637,0.008767879,0.00046934755,0.0005812191,0.000004233375,0.00004899149,0.012080193],"genre_scores_gemma":[0.991372,0.0012513231,0.00010011313,0.0056734285,0.0006511515,0.00003907312,0.00006193149,0.000024814271,0.0008261538],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99821573,0.000047170794,0.0005377889,0.00040238633,0.00043245297,0.00036444765],"domain_scores_gemma":[0.9969701,0.0019312333,0.00031898954,0.00050097425,0.000020570345,0.0002581244],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.000541698,0.00019066082,0.0004967389,0.00010495479,0.00016660127,0.0000073890537,0.00020455074,0.00020115564,0.000085547166],"category_scores_gemma":[0.020136898,0.00015629691,0.000051530653,0.00005474277,0.0004752976,0.0001104941,0.00011223781,0.00074380694,0.00011046697],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005918134,0.0003081602,0.56725246,0.00012790036,0.0001664014,0.000487008,0.010298828,0.000044615328,0.025560047,0.00081006566,0.007456729,0.38689595],"study_design_scores_gemma":[0.028751908,0.0009273651,0.90695184,0.026560688,0.0005537278,0.000044711705,0.0049668294,0.0076241195,0.0055819564,0.0047802427,0.012620101,0.0006364859],"about_ca_topic_score_codex":0.0002553499,"about_ca_topic_score_gemma":0.000010655886,"teacher_disagreement_score":0.38625947,"about_ca_system_score_codex":0.00008657752,"about_ca_system_score_gemma":0.000023008288,"threshold_uncertainty_score":0.9881169},"labels":[],"label_agreement":null},{"id":"W2769893261","doi":"10.1111/jep.12847","title":"Gating the holes in the <scp>Swiss</scp> cheese (part I): <scp>E</scp>xpanding professor <scp>Reason's</scp> model for patient safety","year":2017,"lang":"en","type":"article","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":46,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Saskatchewan Health Authority; Western University; Dalhousie University; Saskatchewan Health; South Bruce Grey Health Centre; University of Saskatchewan","funders":"","keywords":"Cognition; Covert; Harm; Psychology; Health care; Rational analysis; Patient safety; Cognitive bias; Cognitive psychology; Social psychology; Psychiatry","score_opus":0.15587891449811095,"score_gpt":0.4915328957958423,"score_spread":0.3356539812977314,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2769893261","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9603301,0.0012772535,0.0020626087,0.021803929,0.0020598827,0.0026331109,0.000025807663,0.000022789876,0.009784503],"genre_scores_gemma":[0.9850808,0.0012818789,0.0060320236,0.004767572,0.0019292178,0.00018285832,0.000023503284,0.000059796494,0.00064238795],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9886857,0.0031084097,0.0041852933,0.0006129442,0.0026540754,0.0007535658],"domain_scores_gemma":[0.40130374,0.58173305,0.010622225,0.0019440121,0.003736615,0.0006603361],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.06604089,0.00039250142,0.0011336039,0.00018954117,0.0007352037,0.00037135667,0.00082794664,0.000524971,0.000010756465],"category_scores_gemma":[0.93236315,0.00023982044,0.00067465554,0.0003396542,0.00045411015,0.0010170522,0.0002168056,0.0030468267,0.000033408498],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001521955,0.01762681,0.38908136,0.0004697795,0.0017218844,0.0006305121,0.038157538,0.033578802,0.0002871045,0.0059241056,0.3538828,0.15711735],"study_design_scores_gemma":[0.024759587,0.003524457,0.44987398,0.008867104,0.0038065277,0.00064667786,0.048216354,0.38694274,0.0001179167,0.018485881,0.054614924,0.00014382314],"about_ca_topic_score_codex":0.000045845045,"about_ca_topic_score_gemma":0.00007171868,"teacher_disagreement_score":0.8663222,"about_ca_system_score_codex":0.00034296623,"about_ca_system_score_gemma":0.0013774054,"threshold_uncertainty_score":0.9992532},"labels":[],"label_agreement":null},{"id":"W2771318787","doi":"10.1111/jep.12852","title":"Clinical judgement in the era of big data and predictive analytics","year":2017,"lang":"en","type":"article","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":61,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Public Health Ontario; University of Toronto","funders":"","keywords":"Big data; Judgement; Predictive analytics; Clinical judgement; Data science; Analytics; Medicine; Psychology; Computer science; Family medicine; Data mining; Political science","score_opus":0.3703019327459908,"score_gpt":0.5908871838340184,"score_spread":0.22058525108802757,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2771318787","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9251921,0.0010854339,0.0010192416,0.064959146,0.0021994445,0.00068186846,0.000019873727,0.0000023667394,0.0048404834],"genre_scores_gemma":[0.9892396,0.0041776486,0.003316911,0.0020840052,0.0011501524,0.000002694584,0.000009189183,0.000006968342,0.000012861541],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9928466,0.0020811812,0.0033414252,0.00028226443,0.0013103932,0.0001381018],"domain_scores_gemma":[0.75084513,0.23741812,0.0067978664,0.0023333984,0.0023043463,0.00030113783],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.083766125,0.0000969697,0.00065921666,0.00007888994,0.00005793699,0.000055855668,0.00048525076,0.00018960693,0.000034022163],"category_scores_gemma":[0.8271296,0.00006162751,0.00014427389,0.0001251378,0.00037418955,0.0004093046,0.00021602715,0.0017145138,0.0000040247064],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019135708,0.002658682,0.736755,0.000013836698,0.00018123807,0.00012289349,0.00018576825,0.00007954712,6.1792355e-7,0.00006254488,0.005005487,0.2530208],"study_design_scores_gemma":[0.007205009,0.0019457804,0.9626297,0.0012489302,0.0013521493,0.00010498977,0.00070570083,0.019330723,6.2538004e-7,0.001368476,0.004050919,0.000056994308],"about_ca_topic_score_codex":0.000045595374,"about_ca_topic_score_gemma":0.000045571644,"teacher_disagreement_score":0.7433635,"about_ca_system_score_codex":0.000039904968,"about_ca_system_score_gemma":0.00074048695,"threshold_uncertainty_score":0.9434556},"labels":[],"label_agreement":null},{"id":"W2773652717","doi":"10.5430/jnep.v8n4p70","title":"An exploratory mixed methods study of urban and rural registered nurses’ experience of clinical reasoning","year":2017,"lang":"en","type":"article","venue":"Journal of Nursing Education and Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"University of Lethbridge","funders":"","keywords":"Context (archaeology); Exploratory research; Psychology; Qualitative research; Descriptive statistics; Nursing; Applied psychology; Medical education; Medicine","score_opus":0.26143525360923225,"score_gpt":0.6129977857660135,"score_spread":0.35156253215678124,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2773652717","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9937101,0.0011653893,0.00019134066,0.0016183951,0.0018841673,0.00012905517,4.401136e-7,0.000003303129,0.0012977857],"genre_scores_gemma":[0.9640623,0.0003318568,0.035043743,0.00012742676,0.00034932134,0.0000017322349,7.8021463e-7,0.000008360899,0.00007452576],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9979744,0.0007467151,0.0008272995,0.00013260814,0.0002354114,0.00008354087],"domain_scores_gemma":[0.98868245,0.007637039,0.0022868607,0.00041658772,0.00067428465,0.0003027709],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.003552402,0.000081673265,0.0004815898,0.00007750469,0.00015416552,0.000045085166,0.00009023482,0.000078414756,0.0000059778677],"category_scores_gemma":[0.14527713,0.00006627784,0.0000620551,0.000050796054,0.00027635356,0.00052482553,0.00001726619,0.00028901227,1.2184245e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.003569026,0.01512231,0.26260015,0.000041904612,0.00014794638,0.000018105848,0.076172255,9.4150477e-7,0.0009003654,0.00012457346,0.0017573406,0.6395451],"study_design_scores_gemma":[0.0021885333,0.0037826938,0.7751772,0.0040539736,0.0007335569,0.0009118651,0.21204795,0.000071792514,0.0002120342,0.000072347815,0.00067597977,0.00007206553],"about_ca_topic_score_codex":0.000036047655,"about_ca_topic_score_gemma":8.553958e-7,"teacher_disagreement_score":0.639473,"about_ca_system_score_codex":0.000020854053,"about_ca_system_score_gemma":0.00046861317,"threshold_uncertainty_score":0.86192256},"labels":[],"label_agreement":null},{"id":"W2779375299","doi":"","title":"Richtlijn 'Diagnostiek acute buikpijn bij volwassenen'","year":2014,"lang":"nl","type":"article","venue":"Data Archiving and Networked Services (DANS)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute of Infection and Immunity","funders":"","keywords":"Medicine; Guideline; Emergency department; Abdominal pain; Septic shock; Acute abdominal pain; Intensive care medicine; Shock (circulatory); Sepsis; Emergency medicine; Surgery; Radiology; Pathology; Nursing","score_opus":0.01794104506023565,"score_gpt":0.2983735319656723,"score_spread":0.28043248690543665,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2779375299","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9540711,0.010128499,0.007256598,0.0036167316,0.0040418613,0.0010462521,0.0027376572,0.0005792802,0.016522024],"genre_scores_gemma":[0.9595259,0.017521592,0.0028997601,0.0044261566,0.0052389847,0.00003419726,0.007020895,0.0001952749,0.0031372393],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9933721,0.0007480549,0.0013022848,0.0021114103,0.00084171275,0.0016244765],"domain_scores_gemma":[0.95423424,0.040208057,0.00055903743,0.003654231,0.00012308321,0.0012213468],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0024729026,0.000950655,0.0015605292,0.00021560567,0.0006651367,0.0003975947,0.0021312519,0.00042042043,0.00024508973],"category_scores_gemma":[0.0076789693,0.00081152807,0.00021606589,0.0006584005,0.0005017034,0.00041033444,0.002414128,0.0012289849,0.0012074766],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017507733,0.002146632,0.6316657,0.0022322002,0.0033468415,0.0010037062,0.0053519406,0.0009149141,0.0005602444,0.0017952679,0.108858675,0.24037309],"study_design_scores_gemma":[0.0044481205,0.001423284,0.41723397,0.025864586,0.004622105,0.0002858778,0.0010125017,0.45389137,0.000041488373,0.0013409207,0.08816577,0.0016700326],"about_ca_topic_score_codex":0.002256466,"about_ca_topic_score_gemma":0.0010057047,"teacher_disagreement_score":0.45297644,"about_ca_system_score_codex":0.000049585866,"about_ca_system_score_gemma":0.00017797967,"threshold_uncertainty_score":0.9995702},"labels":[],"label_agreement":null},{"id":"W2782103841","doi":"","title":"The Effects of Disease Category on Diagnostic Problem Solving in Mammography","year":2008,"lang":"en","type":"article","venue":"eScholarship (California Digital Library)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"Mammography; Task (project management); Disease; Domain (mathematical analysis); Think aloud protocol; Medical imaging; Medicine; Cognition; Computer science; Psychology; Medical physics; Radiology; Pathology; Human–computer interaction; Breast cancer; Mathematics; Psychiatry","score_opus":0.010694539955290669,"score_gpt":0.2354618077504156,"score_spread":0.22476726779512496,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2782103841","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9892378,0.0020503579,0.000009738774,0.0011141485,0.00013913334,0.00071292877,0.00017224083,0.00014215278,0.006421512],"genre_scores_gemma":[0.99760526,0.00072888244,0.000107957814,0.000704934,0.00010269175,0.00006892335,0.00017766785,0.00006034429,0.0004433178],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9979763,0.00009760963,0.00058287825,0.00039272825,0.00047553045,0.00047495958],"domain_scores_gemma":[0.9678107,0.030955857,0.0001406566,0.0004951605,0.000036602258,0.000561016],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00018695398,0.0002637153,0.0003937665,0.00018275653,0.00014267038,0.00013093479,0.0002535989,0.00011441674,0.000033970366],"category_scores_gemma":[0.05936715,0.00017786647,0.00029453507,0.0005779677,0.00031884413,0.00046202404,0.000117239026,0.0005562094,0.00031392855],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006125539,0.0009246843,0.986153,0.00018850922,0.00004025518,0.0006766125,0.0000286048,0.0000061763626,0.000015518639,0.0008245475,0.0030945363,0.007435013],"study_design_scores_gemma":[0.0023114139,0.0007670285,0.9747962,0.0044852532,0.00006428183,0.000019640678,0.000015431304,0.000018821096,0.0009866523,0.008653693,0.007585036,0.0002965914],"about_ca_topic_score_codex":0.0000059100275,"about_ca_topic_score_gemma":8.765769e-7,"teacher_disagreement_score":0.059180196,"about_ca_system_score_codex":0.000023830742,"about_ca_system_score_gemma":0.00022778318,"threshold_uncertainty_score":0.9485562},"labels":[],"label_agreement":null},{"id":"W2785652257","doi":"10.1164/rccm.201710-2076le","title":"No Room for Error in Medicine—A Case of Déjà Vu","year":2018,"lang":"en","type":"letter","venue":"American Journal of Respiratory and Critical Care Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Déjà vu; Medicine; MEDLINE; Psychiatry","score_opus":0.04710349060111842,"score_gpt":0.40539370218101584,"score_spread":0.35829021157989743,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2785652257","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.19753116,0.027292317,0.0006097638,0.76622653,0.003336407,0.0010202953,0.000172348,0.00003448942,0.0037767082],"genre_scores_gemma":[0.39466333,0.00027307306,0.0012023122,0.5794859,0.024025472,0.000024031373,0.00004371311,0.00010363994,0.0001785121],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99604875,0.0003277084,0.0019010556,0.0004469726,0.0007627471,0.00051274756],"domain_scores_gemma":[0.97195435,0.023872506,0.00076849107,0.00043096842,0.002388674,0.00058499764],"candidate_categories":["metaresearch","metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.0015618027,0.00042751615,0.0027390663,0.00056897115,0.00006165642,0.000006913931,0.00020429897,0.00044938375,0.00033421448],"category_scores_gemma":[0.11369809,0.0002826406,0.0002891463,0.00036892376,0.004273109,0.0000531576,0.00004487999,0.001937227,0.000003447027],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017726091,0.00015425812,0.0052153394,0.0018779263,0.00020299706,0.03616754,0.0012644653,2.7241305e-7,0.00009036959,0.000105351915,0.90009516,0.053053703],"study_design_scores_gemma":[0.01009879,0.09188645,0.0015702492,0.034567196,0.0030611912,0.011083887,0.008064248,0.000014255117,0.000057212485,0.00043800066,0.83861315,0.0005453409],"about_ca_topic_score_codex":0.00020619632,"about_ca_topic_score_gemma":0.00001920484,"teacher_disagreement_score":0.19713217,"about_ca_system_score_codex":0.00009898651,"about_ca_system_score_gemma":0.00046998492,"threshold_uncertainty_score":0.99996257},"labels":[],"label_agreement":null},{"id":"W2787272078","doi":"10.5430/jha.v7n1p43","title":"Improved clinical training of internal medicine students in hospital wards","year":2018,"lang":"en","type":"article","venue":"Journal of Hospital Administration","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Quality assurance; Quality (philosophy); Scale (ratio); Medicine; Medical education; Rating scale; Family medicine; Point (geometry); Quality management; Psychology; External quality assessment; Operations management; Mathematics","score_opus":0.033405592728752924,"score_gpt":0.41495287750186144,"score_spread":0.3815472847731085,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2787272078","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9935899,0.00009888881,0.0003537222,0.0027051626,0.0021923794,0.00018890196,0.0000030951885,0.000007762067,0.00086021074],"genre_scores_gemma":[0.99632734,0.00008146757,0.0011740497,0.00020794786,0.002097705,0.000001803169,0.000005045579,0.000014162288,0.000090491165],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99696815,0.00007245307,0.0019285679,0.00018712014,0.0006485974,0.00019511809],"domain_scores_gemma":[0.99629074,0.0015959976,0.0010499562,0.00016946522,0.0006135064,0.00028030635],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0018521723,0.00014699857,0.0007152258,0.00014519776,0.000020843445,0.000015858843,0.0002040326,0.00015759768,0.00010722497],"category_scores_gemma":[0.03138448,0.00010745997,0.00023158587,0.00017526576,0.00038769792,0.00013712882,0.000036402358,0.00046544507,0.00000523422],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0015793551,0.0049500708,0.9710851,0.000025177407,0.00019686134,0.0003451413,0.0027963545,3.331811e-7,0.00030686017,0.000095659605,0.0019579506,0.016661191],"study_design_scores_gemma":[0.007038578,0.15001887,0.83614844,0.003021352,0.00018531938,0.00013839557,0.0016405109,0.00012163272,0.0010350319,0.00036202406,0.000154664,0.00013514896],"about_ca_topic_score_codex":0.000023231649,"about_ca_topic_score_gemma":0.000014981374,"teacher_disagreement_score":0.14506881,"about_ca_system_score_codex":0.000057854762,"about_ca_system_score_gemma":0.00042504384,"threshold_uncertainty_score":0.9767746},"labels":[],"label_agreement":null},{"id":"W2787485483","doi":"","title":"Utvikling og bruk av kvalitetsindikatorer for spesialisthelsetjenesten","year":2008,"lang":"no","type":"article","venue":"Duo Research Archive (University of Oslo)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Geography","score_opus":0.08969067402087152,"score_gpt":0.3531705222659434,"score_spread":0.26347984824507187,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2787485483","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9387813,0.0024995252,0.013340402,0.009268153,0.0006504979,0.0028515388,0.0017062273,0.00012283272,0.030779488],"genre_scores_gemma":[0.94995576,0.004679574,0.016845366,0.00023543267,0.0010286934,0.0000043396176,0.0006411804,0.00009770525,0.026511945],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99472654,0.00082112924,0.00048153737,0.0009548018,0.0015865136,0.0014294984],"domain_scores_gemma":[0.96477306,0.031666324,0.0002744541,0.000991277,0.0010720241,0.0012228797],"candidate_categories":["metaresearch","metaepi_narrow","sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0029852027,0.00034142067,0.0010092696,0.00078064465,0.0012720857,0.000030697593,0.00088192994,0.00031764864,0.0011208344],"category_scores_gemma":[0.029880412,0.00038737064,0.00071114185,0.00074388186,0.0028839002,0.00017730144,0.0005722053,0.0011358067,0.0006052816],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.023714403,0.008659243,0.297096,0.0027674101,0.0031906373,0.012467461,0.055994254,0.00011529167,0.008301495,0.03661619,0.4303652,0.120712385],"study_design_scores_gemma":[0.029770982,0.008760831,0.65510327,0.012986487,0.0010823153,0.0004650227,0.015983844,0.007788141,0.00085900596,0.023552658,0.24188143,0.0017660276],"about_ca_topic_score_codex":0.002843742,"about_ca_topic_score_gemma":0.00024136661,"teacher_disagreement_score":0.35800722,"about_ca_system_score_codex":0.00019394599,"about_ca_system_score_gemma":0.001599511,"threshold_uncertainty_score":0.99985784},"labels":[],"label_agreement":null},{"id":"W2787751426","doi":"10.3138/jvme.1115-187r","title":"Quantifying Novice and Expert Differences in Visual Diagnostic Reasoning in Veterinary Pathology Using Eye-Tracking Technology","year":2018,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Eye tracking; Protocol analysis; Think aloud protocol; Eye movement; Protocol (science); Cognition; Psychology; Pathology; Medical physics; Artificial intelligence; Medicine; Computer science; Human–computer interaction; Cognitive science","score_opus":0.115784588992179,"score_gpt":0.4808644110946066,"score_spread":0.3650798221024276,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2787751426","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9935439,0.0025454548,0.00020226542,0.0025304134,0.0010180215,0.00011697195,2.520176e-7,0.000013106988,0.00002961693],"genre_scores_gemma":[0.99391955,0.0010325197,0.003680246,0.0006538546,0.0006839081,0.000008155082,0.000002039959,0.00001634656,0.0000034033558],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99806577,0.0001973081,0.0008324088,0.0002519801,0.0003615135,0.00029104753],"domain_scores_gemma":[0.99509305,0.0039258744,0.00034721158,0.00014097891,0.0002007959,0.0002921154],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0011663046,0.00015477993,0.00052955764,0.00061489263,0.00005622336,0.000019561565,0.00012852933,0.0003078129,0.000086558175],"category_scores_gemma":[0.08627734,0.00013055712,0.000051499723,0.00042875242,0.00035114022,0.00015030953,0.00009375511,0.000649038,0.000002865659],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005946424,0.0014063925,0.8253288,0.000070895105,0.000019275338,0.0026917425,0.0018138433,7.714292e-7,0.009668326,0.000041544332,0.00003239638,0.15833133],"study_design_scores_gemma":[0.0019353019,0.0059436182,0.9577686,0.016011834,0.000037815415,0.0133867245,0.0023840123,0.001890288,0.00011970675,0.00012769538,0.0002226799,0.00017168035],"about_ca_topic_score_codex":0.00009597495,"about_ca_topic_score_gemma":0.00001563999,"teacher_disagreement_score":0.15815964,"about_ca_system_score_codex":0.00013125666,"about_ca_system_score_gemma":0.0008413973,"threshold_uncertainty_score":0.9214193},"labels":[],"label_agreement":null},{"id":"W2788437953","doi":"10.2196/mededu.9137","title":"Increasing Reasoning Awareness: Video Analysis of Students’ Two-Party Virtual Patient Interactions","year":2018,"lang":"en","type":"article","venue":"JMIR Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":20,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Virtual patient; Medical education; Psychology; Computer science; Medicine","score_opus":0.019102227677183704,"score_gpt":0.4274921596446074,"score_spread":0.4083899319674237,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2788437953","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99539244,0.000058749763,0.0004143363,0.0012434005,0.00095615385,0.00023618506,0.000005149737,0.00005086617,0.0016427326],"genre_scores_gemma":[0.99631715,0.00002312788,0.0007364999,0.0017796049,0.00065121724,0.00008533127,0.00019771534,0.0000141132405,0.00019522122],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9976829,0.00016222804,0.00063698366,0.00032694868,0.0009851638,0.00020577382],"domain_scores_gemma":[0.994655,0.003598138,0.00027067002,0.00041295364,0.00049293024,0.0005703021],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00068630616,0.00013263927,0.00044829218,0.0003427645,0.00010012858,0.000023184484,0.00014306385,0.00011304783,0.002232626],"category_scores_gemma":[0.06692447,0.00011028424,0.0001949597,0.0010242406,0.00023560364,0.000080377096,0.000079756515,0.00026546055,0.000057425383],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00029444246,0.0033829652,0.877865,0.000013773853,0.00061025005,0.000004952523,0.0024169877,0.0000053383246,0.00016209962,0.0004291252,0.0051122066,0.10970284],"study_design_scores_gemma":[0.00138879,0.0007771808,0.9799232,0.0036002127,0.0019188073,0.00005163285,0.0026166323,0.0039834497,0.0004218254,0.000067690824,0.0050573437,0.00019324763],"about_ca_topic_score_codex":0.001078887,"about_ca_topic_score_gemma":0.00016566378,"teacher_disagreement_score":0.10950959,"about_ca_system_score_codex":0.00012984271,"about_ca_system_score_gemma":0.0012735504,"threshold_uncertainty_score":0.99867946},"labels":[],"label_agreement":null},{"id":"W2789314200","doi":"10.1111/jep.12907","title":"The 3 faces of clinical reasoning: Epistemological explorations of disparate error reduction strategies","year":2018,"lang":"en","type":"article","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":37,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Heuristic; Perspective (graphical); Categorization; Class (philosophy); Logical reasoning; Process (computing); Cognition; Cognitive bias; Computer science; Population; Artificial intelligence; Cognitive psychology; Psychology; Medicine","score_opus":0.269337700662447,"score_gpt":0.5784932144895651,"score_spread":0.30915551382711814,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2789314200","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9763414,0.00077327265,0.0012667202,0.013159051,0.0025387777,0.00043795622,0.0000028675702,0.000007658887,0.005472242],"genre_scores_gemma":[0.98837805,0.0014416854,0.008612569,0.00017438606,0.0012757549,0.000008539151,0.000004667303,0.000010196396,0.00009414322],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9893544,0.0034100842,0.0055339793,0.00025297623,0.0012643792,0.00018419113],"domain_scores_gemma":[0.8158142,0.16491613,0.009490481,0.00074526196,0.008676246,0.00035767537],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.0464365,0.00012999085,0.00088498864,0.000099500176,0.00009608872,0.000036729856,0.00018511826,0.0002993668,0.00012805169],"category_scores_gemma":[0.68182623,0.00007908179,0.00042158004,0.00042006618,0.001112594,0.00053469907,0.00004714757,0.0010052355,0.000017436667],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.059928183,0.025567394,0.45536086,0.00018670008,0.0026510432,0.00020340223,0.0031838743,0.0042533395,0.0007674065,0.036422867,0.04541887,0.36605605],"study_design_scores_gemma":[0.009997795,0.0161811,0.919551,0.0029669218,0.0022625136,0.00060714077,0.012402069,0.010922243,0.00031364133,0.017700642,0.0068622935,0.00023264906],"about_ca_topic_score_codex":0.000029762012,"about_ca_topic_score_gemma":0.000021547714,"teacher_disagreement_score":0.6353897,"about_ca_system_score_codex":0.000045730267,"about_ca_system_score_gemma":0.0016597866,"threshold_uncertainty_score":0.9818943},"labels":[],"label_agreement":null},{"id":"W2789349133","doi":"10.3389/fvets.2018.00043","title":"Morbidity and Mortality Conferences: A Mini Review and Illustrated Application in Veterinary Medicine","year":2018,"lang":"en","type":"review","venue":"Frontiers in Veterinary Science","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":28,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Blame; Multidisciplinary approach; Patient care; Medicine; Patient safety; Medical education; Psychology; Health care; Nursing; Political science; Sociology; Social science; Psychiatry","score_opus":0.1571470539890557,"score_gpt":0.4415223493757952,"score_spread":0.2843752953867395,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2789349133","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.026573222,0.9704264,0.00009106316,0.00016464507,0.0004840858,0.0014886039,0.000025384195,0.000028898547,0.0007177309],"genre_scores_gemma":[0.005814152,0.9921018,0.001412287,0.0002757242,0.00009889332,0.00014845529,0.00006892375,0.00001666577,0.000063129875],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99681926,0.00023879884,0.00100622,0.0011101912,0.00040346105,0.00042209527],"domain_scores_gemma":[0.99753076,0.0009915128,0.00036280305,0.0006595844,0.000105465835,0.00034989178],"candidate_categories":["metaresearch","metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.002757027,0.00039502984,0.0022418792,0.00048447246,0.00007688697,0.000025790998,0.00035584898,0.00025316814,0.000024068238],"category_scores_gemma":[0.009010566,0.0002972421,0.000076232885,0.0015521781,0.003044627,0.00014913222,0.00024757316,0.00047297394,0.000005517617],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006941577,0.00026687965,0.011969454,0.044199355,0.000054440643,0.00059130706,0.00016870815,5.9267375e-8,0.000007871602,0.00004813458,0.00349535,0.93912905],"study_design_scores_gemma":[0.0012032206,0.0046765907,0.079017825,0.35038117,0.0010732703,0.0016108814,0.00011896187,0.00031428417,1.2043948e-7,0.000272518,0.5606135,0.00071764307],"about_ca_topic_score_codex":0.00031917528,"about_ca_topic_score_gemma":0.000007417138,"teacher_disagreement_score":0.93841136,"about_ca_system_score_codex":0.0001632242,"about_ca_system_score_gemma":0.0008620013,"threshold_uncertainty_score":0.99994797},"labels":[],"label_agreement":null},{"id":"W2790519506","doi":"10.1051/pmed/2017024","title":"Une rencontre épistémologique","year":2017,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Humanities; Qualitative research; Philosophy; Sociology; Social science","score_opus":0.0670799353941154,"score_gpt":0.39939250994366554,"score_spread":0.33231257454955015,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2790519506","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.10638489,0.11136352,0.0021004207,0.32861707,0.018505577,0.0009617327,0.00021364754,0.00039789203,0.43145525],"genre_scores_gemma":[0.5012768,0.01245026,0.0022559366,0.0037684608,0.005379878,0.000044938028,0.000049710485,0.000068225534,0.47470582],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9967422,0.00023448009,0.0006974959,0.00073874084,0.00049493863,0.0010921556],"domain_scores_gemma":[0.9838481,0.012338644,0.00040733704,0.0021442606,0.0001800108,0.0010816734],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0015491045,0.00037976648,0.0008671634,0.00005259443,0.0006618877,0.00022973679,0.00076956657,0.00082069705,0.0035481928],"category_scores_gemma":[0.23263049,0.00033912327,0.00039746016,0.00010591796,0.0019926047,0.00015069387,0.00037380945,0.0010376029,0.004677128],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018256089,0.0013676404,0.08744367,0.0001399611,0.00019425391,0.0027416302,0.00026388318,0.000007151664,0.000089508794,0.0052325404,0.19194472,0.7103925],"study_design_scores_gemma":[0.005855469,0.0012242491,0.36697862,0.011040142,0.00077048404,0.0002646021,0.00027779804,0.00076093466,0.00073646853,0.019557457,0.5917845,0.0007493017],"about_ca_topic_score_codex":0.0039933124,"about_ca_topic_score_gemma":0.0002079367,"teacher_disagreement_score":0.7096432,"about_ca_system_score_codex":0.0001659394,"about_ca_system_score_gemma":0.0007452961,"threshold_uncertainty_score":0.99990606},"labels":[],"label_agreement":null},{"id":"W2791813984","doi":"10.1111/medu.13519","title":"A critical narrative review of transfer of basic science knowledge in health professions education","year":2018,"lang":"en","type":"review","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":63,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; McMaster University; University of Toronto; University Health Network","funders":"","keywords":"Knowledge transfer; Narrative; Narrative review; Medical education; Health professions; Engineering ethics; Medicine; Psychology; Pedagogy; Knowledge management; Computer science; Health care; Political science; Engineering; Philosophy; Psychotherapist","score_opus":0.0729375704928514,"score_gpt":0.5494712821799597,"score_spread":0.47653371168710823,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2791813984","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00006425535,0.9862077,0.000034898756,0.008228457,0.0015634254,0.0014203001,0.000008998718,0.000010755451,0.0024611682],"genre_scores_gemma":[0.00043620862,0.99608266,0.00035937552,0.0019295213,0.0003613868,0.00033970713,0.00017821707,0.000019179592,0.00029371184],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9959981,0.000536831,0.0018778275,0.00046949857,0.0008358935,0.00028188658],"domain_scores_gemma":[0.9879287,0.008434741,0.00046712448,0.00075158954,0.0015925394,0.00082529313],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0041020475,0.00022031892,0.0019331027,0.00041407804,0.000059658305,0.0000040919463,0.00027973484,0.00029013897,0.0009440233],"category_scores_gemma":[0.26317874,0.00015558721,0.000272097,0.0017027066,0.0012308104,0.000058081943,0.00004130699,0.00062655506,0.000036174017],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000005102769,0.0030039398,0.000015394362,0.17248906,0.000009301692,2.9631434e-7,0.0011017215,6.46715e-10,7.007178e-8,0.0012430607,0.024610609,0.7975214],"study_design_scores_gemma":[0.00012557802,0.00023018122,0.000098727614,0.6646025,0.00018029318,0.000022929848,0.00024319795,0.000002066326,9.964574e-7,0.00006186274,0.33435017,0.00008148678],"about_ca_topic_score_codex":0.0000892744,"about_ca_topic_score_gemma":0.000010553519,"teacher_disagreement_score":0.79743993,"about_ca_system_score_codex":0.0003390545,"about_ca_system_score_gemma":0.111154675,"threshold_uncertainty_score":0.99996924},"labels":[],"label_agreement":null},{"id":"W2791886998","doi":"10.1097/acm.0000000000002142","title":"Drawing Boundaries: The Difficulty in Defining Clinical Reasoning","year":2018,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":169,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Interdisciplinary Research in Rehabilitation; McGill University","funders":"","keywords":"Conceptualization; Analogy; Construct (python library); Psychology; Deductive reasoning; Management science; Epistemology; Cognitive science; Computer science; Artificial intelligence","score_opus":0.05913878509975403,"score_gpt":0.4251610115444137,"score_spread":0.3660222264446597,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2791886998","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9580367,0.0022461493,0.00035906004,0.025551483,0.001592756,0.00033537467,0.0000017063322,0.00012109371,0.011755668],"genre_scores_gemma":[0.9813929,0.0007737903,0.0003007056,0.012631563,0.00367798,0.000020257246,0.000025254652,0.00003106862,0.0011464643],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9968324,0.00021229761,0.0012941727,0.0005223412,0.00057421584,0.0005646303],"domain_scores_gemma":[0.9811885,0.017466057,0.0002595145,0.00055229955,0.00015235528,0.00038130846],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0044648647,0.00023093213,0.0007386528,0.00010685708,0.00030048034,0.000019606083,0.00029380355,0.0004329788,0.0003711082],"category_scores_gemma":[0.21338145,0.00012959953,0.000105168925,0.0005674259,0.0025826707,0.000060851846,0.00012710904,0.002080713,0.00026004924],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00042405975,0.000057651923,0.8606267,0.000023834169,0.00008235537,0.00012855079,0.00871186,0.0000012377903,0.00017935646,0.0099331355,0.0866601,0.03317113],"study_design_scores_gemma":[0.009695434,0.0019396293,0.8196572,0.01680552,0.00043857394,0.00040689867,0.005461036,0.0016231866,0.00012039074,0.0027702763,0.14067695,0.00040486778],"about_ca_topic_score_codex":0.00030465165,"about_ca_topic_score_gemma":0.000037197067,"teacher_disagreement_score":0.20891659,"about_ca_system_score_codex":0.000076559074,"about_ca_system_score_gemma":0.00034885434,"threshold_uncertainty_score":0.9515957},"labels":[],"label_agreement":null},{"id":"W2792472626","doi":"10.1093/jcag/gwy009.150","title":"A150 IS FOCAL ACTIVE COLITIS OF GREATER SIGNIFICANCE IN PEDIATRIC PATIENTS? A RETROSPECTIVE REVIEW OF 68 CASES WITH CLINICAL CORRELATION.","year":2018,"lang":"en","type":"review","venue":"Journal of the Canadian Association of Gastroenterology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Children's Hospital of Western Ontario; Western University","funders":"","keywords":"Correlation; Medicine; Retrospective cohort study; Clinical significance; Internal medicine; Mathematics","score_opus":0.029109165634616307,"score_gpt":0.33496441335097166,"score_spread":0.30585524771635536,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2792472626","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.17145872,0.81057423,0.00003800592,0.0056848726,0.0033801617,0.004560669,0.0036600498,0.000008063652,0.0006352383],"genre_scores_gemma":[0.06002214,0.9383833,0.00010887316,0.00094244926,0.00033144775,0.000010028102,0.000033822107,0.000026646632,0.00014130968],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9954248,0.0010102945,0.002493178,0.00021638928,0.0005607329,0.0002945872],"domain_scores_gemma":[0.9857877,0.0037154641,0.008018437,0.00027597114,0.0019337161,0.00026871916],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0018601697,0.0002173288,0.0029218746,0.00034964035,0.000026527718,0.0000035507753,0.0002462677,0.00038369582,0.00010771513],"category_scores_gemma":[0.038020257,0.00014182455,0.00084294815,0.00039386295,0.000212202,0.000043016535,0.0000309125,0.0008119971,0.0000031060272],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00042520007,0.00020902336,0.97713506,0.0026947842,0.00046997002,0.000024395136,0.00004630486,4.906925e-7,2.9188219e-9,0.0000035446649,0.015661579,0.0033296684],"study_design_scores_gemma":[0.002317301,0.0054306653,0.9012345,0.07985278,0.0034165438,0.00009576663,0.0000071658683,0.0000010011113,2.6503915e-7,0.000013149691,0.0075539276,0.00007694244],"about_ca_topic_score_codex":0.0018594742,"about_ca_topic_score_gemma":0.024251796,"teacher_disagreement_score":0.12780906,"about_ca_system_score_codex":0.001526967,"about_ca_system_score_gemma":0.0023409838,"threshold_uncertainty_score":0.99355304},"labels":[],"label_agreement":null},{"id":"W2792822433","doi":"10.1111/jep.12905","title":"Developing the experts we need: Fostering adaptive expertise through education","year":2018,"lang":"en","type":"article","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":191,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; University of Toronto","funders":"","keywords":"Excellence; Framing (construction); Curriculum; Context (archaeology); Set (abstract data type); Health care; Engineering ethics; Knowledge management; Medical education; Computer science; Psychology; Medicine; Pedagogy; Political science","score_opus":0.317056414422352,"score_gpt":0.5676272083374496,"score_spread":0.2505707939150976,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2792822433","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.67660576,0.007611504,0.024566082,0.25115126,0.010093918,0.0013378162,8.779113e-7,0.000028365233,0.028604388],"genre_scores_gemma":[0.9412316,0.0016739335,0.040731527,0.013115399,0.0031201215,0.0000222876,0.0000013893218,0.000016228596,0.00008750476],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99570006,0.0010996249,0.0018582224,0.0002062582,0.0009631064,0.00017271179],"domain_scores_gemma":[0.9270841,0.066457085,0.00200604,0.0004287447,0.0038278862,0.00019617265],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.011630636,0.00011960531,0.0003739125,0.00008173387,0.000095096526,0.00004594184,0.00013184581,0.00015765772,0.00019331832],"category_scores_gemma":[0.45921478,0.00007693974,0.00017190703,0.00031260328,0.00017822148,0.0005944936,0.00007678179,0.00058037654,0.000043332064],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0035470063,0.001560377,0.0058705737,0.000013883299,0.00020862195,0.000051045692,0.006536532,0.000087315464,0.000056718367,0.0014564064,0.03154832,0.9490632],"study_design_scores_gemma":[0.03373553,0.019226903,0.26938942,0.038007632,0.004796353,0.006974251,0.11438205,0.02809216,0.0013868152,0.10786813,0.3746821,0.0014586471],"about_ca_topic_score_codex":0.000047902624,"about_ca_topic_score_gemma":0.000012390628,"teacher_disagreement_score":0.94760454,"about_ca_system_score_codex":0.00037411868,"about_ca_system_score_gemma":0.002456049,"threshold_uncertainty_score":0.5453405},"labels":[],"label_agreement":null},{"id":"W2793137171","doi":"10.22454/fammed.2018.221401","title":"Education to Improve Dementia Care:","year":2018,"lang":"en","type":"article","venue":"Family Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Hamilton Health Sciences; Health Sciences Centre; Centre for Family Medicine; Western University","funders":"Canadian Institutes of Health Research","keywords":"Dementia; Cognition; Cognitive impairment; Psychology; Clinical Practice; Medicine; MEDLINE; Clinical psychology; Psychiatry; Family medicine; Disease","score_opus":0.02869562680057204,"score_gpt":0.37933953716221447,"score_spread":0.35064391036164244,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2793137171","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8856407,0.0035595605,0.00075967074,0.009684635,0.004214354,0.0005810159,0.000004934558,0.00011951068,0.09543562],"genre_scores_gemma":[0.96889013,0.0000849826,0.0019726201,0.022984795,0.0042285044,0.00004053307,0.000061672145,0.000020196667,0.0017165391],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9988571,0.000018555247,0.00029007587,0.00029898167,0.00030645335,0.00022886158],"domain_scores_gemma":[0.99798065,0.0007371505,0.000050639504,0.0004001016,0.0003978692,0.00043356756],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00021540486,0.00012296441,0.00027011023,0.000099284545,0.000057238387,0.0000052521364,0.000082149294,0.00007480972,0.0003383311],"category_scores_gemma":[0.01621883,0.000089649344,0.000043053657,0.00025606793,0.00015020992,0.000019832078,0.000036748967,0.00011548522,0.0007671504],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002449056,0.00028104236,0.035146784,0.000051315314,0.00008831481,0.000014415954,0.0019109711,1.7743308e-7,0.014107713,0.00061787106,0.37233377,0.5752027],"study_design_scores_gemma":[0.0030547536,0.005196869,0.6591777,0.0030523543,0.0005390778,0.000012868554,0.001971563,0.000017247818,0.0022535734,0.00025432347,0.32426196,0.00020766685],"about_ca_topic_score_codex":0.0004992795,"about_ca_topic_score_gemma":0.000016720833,"teacher_disagreement_score":0.62403095,"about_ca_system_score_codex":0.000056623947,"about_ca_system_score_gemma":0.00024911948,"threshold_uncertainty_score":0.992068},"labels":[],"label_agreement":null},{"id":"W2793613924","doi":"10.1007/s11606-018-4343-8","title":"General Internists in Pursuit of Diagnostic Excellence in Primary Care: a #ProudtobeGIM Thread That Unites Us All","year":2018,"lang":"en","type":"letter","venue":"Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Mount Sinai Hospital; University of Toronto","funders":"VA National Center for Patient Safety; Center for Innovations in Quality, Effectiveness and Safety; Agency for Healthcare Research and Quality; Cancer Research UK; Gordon and Betty Moore Foundation; Health Services Research and Development; U.S. Department of Veterans Affairs","keywords":"Medicine; Excellence; Primary care; Family medicine; Law","score_opus":0.043495713408561155,"score_gpt":0.34328646684368264,"score_spread":0.29979075343512146,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2793613924","genre_codex":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.78733236,0.014981529,0.00012276559,0.18829016,0.0047736415,0.00058247003,0.000038473525,0.000017896735,0.0038606916],"genre_scores_gemma":[0.3570446,0.01774802,0.0032238273,0.5630586,0.04527318,0.000051807718,0.0005721332,0.0002935334,0.012734301],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9933608,0.00035707193,0.0028992062,0.0005923906,0.0020319019,0.0007586043],"domain_scores_gemma":[0.98519474,0.010892376,0.00207579,0.0006274442,0.00081299664,0.00039664694],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0015390213,0.00070689805,0.0027904015,0.0013360087,0.00001936339,0.000033740685,0.0011056146,0.0009051821,0.00040259992],"category_scores_gemma":[0.0252893,0.0004975593,0.00051323965,0.0003864559,0.00077693874,0.00012727777,0.0002789066,0.0041604615,0.00002032346],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013336402,0.00043576927,0.24603032,0.0009361773,0.00040010418,0.022939337,0.0012136571,0.00002154134,0.00084837805,0.000007695707,0.7190576,0.006775788],"study_design_scores_gemma":[0.025921177,0.020704316,0.339142,0.23997496,0.0021688181,0.009608407,0.0004032367,0.00019955785,0.0025915315,0.0011294427,0.35668647,0.0014700943],"about_ca_topic_score_codex":0.003687362,"about_ca_topic_score_gemma":0.00026495857,"teacher_disagreement_score":0.43028778,"about_ca_system_score_codex":0.0008443375,"about_ca_system_score_gemma":0.0005990545,"threshold_uncertainty_score":0.99974763},"labels":[],"label_agreement":null},{"id":"W2793736931","doi":"10.1080/09593985.2017.1419521","title":"Development and initial evaluation of an instrument to assess physiotherapists’ clinical reasoning focused on clients’ behavior change","year":2018,"lang":"en","type":"article","venue":"Physiotherapy Theory and Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Delphi method; Relevance (law); Content validity; Process (computing); Behavior change; Delphi; Behaviour change; Psychology; Physical therapy; Medicine; Applied psychology; Computer science; Psychometrics; Clinical psychology; Nursing; Psychological intervention; Artificial intelligence; Social psychology","score_opus":0.23576674371976497,"score_gpt":0.5235143761880771,"score_spread":0.2877476324683121,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2793736931","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9967807,0.00013652144,0.00017101326,0.00033616775,0.00025213396,0.0010112725,0.0000051924026,0.000029588056,0.0012774343],"genre_scores_gemma":[0.9835022,0.00016580756,0.010413789,0.0049358937,0.0006879874,0.000217669,0.000015674636,0.000031014682,0.000029954881],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9970421,0.001350046,0.00043623775,0.00043832505,0.0005411196,0.00019216667],"domain_scores_gemma":[0.9904883,0.008309084,0.0002579472,0.00032811047,0.00034504329,0.00027149584],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00653399,0.00018244918,0.00036057603,0.000090681526,0.0001378916,0.000035802346,0.00007003976,0.00010923689,0.00013457266],"category_scores_gemma":[0.012880518,0.00014939002,0.000048563164,0.0001656166,0.00018838161,0.0002350487,0.000034544806,0.00018213253,0.000011640783],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.011183315,0.002350652,0.001892876,0.000026200894,0.00014325096,0.0000051611205,0.0038471208,4.4459603e-7,0.0029905976,0.0036710496,0.000018956805,0.9738704],"study_design_scores_gemma":[0.024891613,0.023320358,0.88398194,0.0052397037,0.0011439244,0.000103902115,0.0030287579,0.0007617366,0.0382975,0.0032694803,0.014916015,0.001045039],"about_ca_topic_score_codex":0.000024785959,"about_ca_topic_score_gemma":0.0000055945397,"teacher_disagreement_score":0.97282535,"about_ca_system_score_codex":0.000039639588,"about_ca_system_score_gemma":0.00014498558,"threshold_uncertainty_score":0.9954344},"labels":[],"label_agreement":null},{"id":"W2793744238","doi":"10.1136/bmjopen-2017-019241","title":"Does exposure to simulated patient cases improve accuracy of clinicians’ predictive value estimates of diagnostic test results? A within-subjects experiment at St Michael’s Hospital, Toronto, Canada","year":2018,"lang":"en","type":"article","venue":"BMJ Open","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Toronto; St. Michael's Hospital; Toronto Metropolitan University","funders":"University of Toronto","keywords":"Medicine; Gold standard (test); Test (biology); Predictive value; Pre- and post-test probability; Predictive value of tests; Diagnostic test; Positive predicative value; Statistics; Pediatrics; Internal medicine; Mathematics","score_opus":0.028648575816303545,"score_gpt":0.38036961109992484,"score_spread":0.3517210352836213,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2793744238","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98648745,0.00022766764,0.000007802462,0.0010781593,0.0012107972,0.008155161,0.0019406528,0.00003136869,0.00086094823],"genre_scores_gemma":[0.9968847,0.00003098286,0.0013552505,0.00071925356,0.00022323764,0.0003914786,0.0001490316,0.00004180884,0.00020426491],"study_design_codex":"observational","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.9966952,0.0000964093,0.0015223383,0.00064215466,0.0006530615,0.00039084506],"domain_scores_gemma":[0.9269369,0.06939416,0.0010830637,0.0011479694,0.0008955313,0.0005423627],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00059082895,0.00031975098,0.00093208556,0.00003329725,0.00009934831,0.00002892806,0.00037554788,0.00014631725,0.00016440294],"category_scores_gemma":[0.3679802,0.00020625214,0.000107954715,0.00016136802,0.00021789479,0.00012521376,0.00066415017,0.00012261583,0.000009393504],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.04892589,0.016210003,0.63119763,0.001109214,0.0026458881,0.0028627003,0.07216194,0.003273933,0.014556551,0.00012618834,0.17730772,0.029622313],"study_design_scores_gemma":[0.021681422,0.11143905,0.2650614,0.03320147,0.0013003522,0.00006956474,0.019556703,0.003002894,0.54101825,0.0001503142,0.001869455,0.0016491147],"about_ca_topic_score_codex":0.53636575,"about_ca_topic_score_gemma":0.32734755,"teacher_disagreement_score":0.5264617,"about_ca_system_score_codex":0.00042718402,"about_ca_system_score_gemma":0.0015460647,"threshold_uncertainty_score":0.8410718},"labels":[],"label_agreement":null},{"id":"W2800142590","doi":"10.1017/cem.2018.337","title":"P139: How available is availability bias? Examining factors that influence diagnostic error","year":2018,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hamilton Health Sciences","funders":"","keywords":"Medical diagnosis; Cognitive bias; Test (biology); Medicine; Cognition; Priming (agriculture); Cognitive test; Clinical psychology; Psychiatry; Pathology","score_opus":0.23193414757240152,"score_gpt":0.3795447185811967,"score_spread":0.14761057100879518,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2800142590","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9769149,0.0029071833,0.000052934145,0.007925047,0.006975238,0.00024614946,0.00002678377,0.000020594845,0.004931171],"genre_scores_gemma":[0.9910993,0.0006520657,0.00015018957,0.001248054,0.0016706618,0.0000038288326,0.0000129266145,0.000041779003,0.005121221],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9969125,0.00012515248,0.0011089268,0.00040810963,0.0006967351,0.0007486074],"domain_scores_gemma":[0.9876356,0.006295962,0.0007507151,0.0006906738,0.0014668426,0.0031601852],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0015765203,0.00034923665,0.0009036784,0.0004319686,0.0002363459,0.00001685041,0.00036825275,0.00021901485,0.03565128],"category_scores_gemma":[0.23981021,0.00025042737,0.00020947457,0.0006702681,0.0007365871,0.00023028119,0.00002397455,0.0006353767,0.00025368124],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000029639517,0.000038726816,0.77562726,0.000035313402,0.00008946771,0.0001730586,0.0014565968,8.569782e-7,0.00014286912,0.000010243342,0.22169891,0.00069705275],"study_design_scores_gemma":[0.0017278662,0.0033504253,0.8908606,0.0041903593,0.0005516715,0.00010351026,0.0016575946,0.00003226692,0.0006725793,0.00019984641,0.09631974,0.00033353094],"about_ca_topic_score_codex":0.012104218,"about_ca_topic_score_gemma":0.008923545,"teacher_disagreement_score":0.23823369,"about_ca_system_score_codex":0.0002274707,"about_ca_system_score_gemma":0.0018406644,"threshold_uncertainty_score":0.9999948},"labels":[],"label_agreement":null},{"id":"W2800234261","doi":"10.1111/jep.12931","title":"Clinical judgement: Multidisciplinary perspectives","year":2018,"lang":"en","type":"editorial","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University Health Network; University of Toronto","funders":"","keywords":"Judgement; Clinical judgement; Intuition; Epistemology; Psychology; Medicine; Philosophy; Family medicine","score_opus":0.10001924984893909,"score_gpt":0.5645408340565281,"score_spread":0.46452158420758904,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2800234261","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0106915105,0.0048164255,0.00020980228,0.009107766,0.96491826,0.0010009543,0.000026063024,0.000025090003,0.009204127],"genre_scores_gemma":[0.007935407,0.013414892,0.014418,0.00062935485,0.9626056,0.000019881043,0.000091396905,0.000083744075,0.00080175477],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9794899,0.0054484783,0.008599009,0.0009295728,0.005106707,0.00042633677],"domain_scores_gemma":[0.1841654,0.76925105,0.015522124,0.0018268496,0.02785006,0.001384497],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch","research_integrity"],"category_scores_codex":[0.101032645,0.00042432736,0.0023187092,0.00041599834,0.000090923335,0.00008773038,0.0004007148,0.0023784589,0.0010569834],"category_scores_gemma":[0.948758,0.0003395079,0.0014549885,0.00045991808,0.00064342,0.0005141088,0.00023347125,0.0073051364,0.0003166864],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.005657257,0.006597705,0.006464545,0.000044960183,0.0007177449,0.0004745543,0.0003031816,0.000015877611,2.6479842e-7,0.000017455877,0.95404303,0.025663411],"study_design_scores_gemma":[0.018559633,0.010276609,0.038319662,0.0073655653,0.0063967304,0.00016680591,0.001839436,0.0010163735,5.3643777e-7,0.001182639,0.9144559,0.000420136],"about_ca_topic_score_codex":0.00002540801,"about_ca_topic_score_gemma":0.00001471074,"teacher_disagreement_score":0.84772533,"about_ca_system_score_codex":0.0005784545,"about_ca_system_score_gemma":0.006202353,"threshold_uncertainty_score":0.9999057},"labels":[],"label_agreement":null},{"id":"W2801623497","doi":"10.1017/cem.2018.215","title":"P017: When the rules hit the road: how emergency physicians make decisions in the era of the Clinical Decision Rules","year":2018,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hamilton Health Sciences","funders":"","keywords":"Think aloud protocol; Cognition; Medicine; Emergency department; Medical education; Applied psychology; Psychology; Psychiatry; Computer science","score_opus":0.10641912680135486,"score_gpt":0.40595936419991047,"score_spread":0.2995402373985556,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2801623497","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8218495,0.005357328,0.00021742716,0.15696563,0.011878181,0.0004931662,0.000035768368,0.000004847624,0.0031981042],"genre_scores_gemma":[0.9852169,0.0053733345,0.00019353209,0.002793439,0.005557485,0.000009083131,0.0000064318265,0.00003229466,0.00081749493],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9953465,0.0006626638,0.0020482736,0.00027535498,0.001088542,0.00057870196],"domain_scores_gemma":[0.9861763,0.009554121,0.0011445752,0.0014342082,0.0010407829,0.00065000425],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.005780182,0.00027740572,0.0007493333,0.00014760687,0.00049355463,0.000013354615,0.0015789341,0.00017987276,0.002242962],"category_scores_gemma":[0.1155665,0.000102406746,0.00062811514,0.00074125023,0.0011776404,0.00006357868,0.000062366176,0.0014607734,0.000033841912],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015531214,0.00021004799,0.14898247,0.000008387501,0.00020963443,0.00006934548,0.0069607017,0.000007646114,0.000037463426,0.0009897143,0.6340861,0.2082832],"study_design_scores_gemma":[0.0014982283,0.0010868717,0.86695945,0.0032708992,0.0005664807,0.00008739082,0.004168515,0.00008636122,0.000010203324,0.011904642,0.110218756,0.00014220856],"about_ca_topic_score_codex":0.005636882,"about_ca_topic_score_gemma":0.06583172,"teacher_disagreement_score":0.717977,"about_ca_system_score_codex":0.00007644612,"about_ca_system_score_gemma":0.0011284043,"threshold_uncertainty_score":0.99866915},"labels":[],"label_agreement":null},{"id":"W2802381110","doi":"10.1007/s40037-018-0425-x","title":"Lies, damned lies, and statistics","year":2018,"lang":"en","type":"article","venue":"Perspectives on Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Statistics; Data science; Computer science; Medical education; Psychology; Medicine; Mathematics","score_opus":0.01499620737032751,"score_gpt":0.38465735838833165,"score_spread":0.3696611510180041,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2802381110","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.89287573,0.0034446071,0.006028668,0.0446931,0.0021491477,0.0006283777,0.000034440363,0.0002157804,0.04993013],"genre_scores_gemma":[0.98100257,0.0012250406,0.008629585,0.004974743,0.0018752198,0.000024088087,0.00006350313,0.000018755873,0.0021865217],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9988326,0.000044435983,0.00019167438,0.00031468514,0.000449137,0.0001674829],"domain_scores_gemma":[0.9962685,0.0026218481,0.000055243712,0.0002270467,0.00027928536,0.0005480819],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00029128458,0.0001064753,0.00019242211,0.00007289725,0.0000874997,0.000020807262,0.000056338493,0.0001197627,0.001727943],"category_scores_gemma":[0.11849223,0.00008542604,0.000026584163,0.00013206645,0.00051954837,0.000026282763,0.000023735356,0.00024729528,0.0001662025],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007493022,0.0046828007,0.022589615,0.000058808557,0.000115805684,0.00002383067,0.017945832,8.5762906e-8,0.000039025435,0.13843988,0.2319057,0.5834493],"study_design_scores_gemma":[0.005374925,0.0067725237,0.8328723,0.005012836,0.00039642976,0.00026863933,0.04348963,0.001742221,0.00029056353,0.017015327,0.08606164,0.0007029973],"about_ca_topic_score_codex":0.00014887273,"about_ca_topic_score_gemma":0.000021444992,"teacher_disagreement_score":0.81028265,"about_ca_system_score_codex":0.00008858665,"about_ca_system_score_gemma":0.0010640994,"threshold_uncertainty_score":0.9991846},"labels":[],"label_agreement":null},{"id":"W2802488858","doi":"10.1017/cem.2018.313","title":"P115: Bounceback reports-improving patient care","year":2018,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Medicine; Emergency department; Complaint; Medical emergency; Family medicine; Emergency medicine; Nursing","score_opus":0.06469549533490818,"score_gpt":0.3646307087611592,"score_spread":0.29993521342625107,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2802488858","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9154675,0.012133956,0.00016114379,0.0049314457,0.028695716,0.00027051067,0.000008844494,0.000017406277,0.038313486],"genre_scores_gemma":[0.9927982,0.00018502017,0.00031300925,0.00077717117,0.0049871765,0.0000019242748,0.000012973983,0.000026395908,0.0008981372],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99760145,0.000037062826,0.0012735024,0.00021373655,0.00045471845,0.0004195141],"domain_scores_gemma":[0.99503297,0.00030944077,0.00058852264,0.00037212984,0.0017655464,0.0019313848],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00053306436,0.0001722728,0.00049728394,0.00025587803,0.00014919577,0.0000050084823,0.000113126625,0.00011204251,0.007972145],"category_scores_gemma":[0.047682893,0.00012480236,0.00017559329,0.00034485033,0.00027076158,0.00006239557,0.000010947003,0.00035057502,0.000036725804],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00016096645,0.00008180798,0.24069527,0.00012753932,0.00022112146,0.005742082,0.010504928,0.000005427442,0.0008191989,0.0001370839,0.64982146,0.0916831],"study_design_scores_gemma":[0.0033966627,0.020746531,0.08364261,0.009381907,0.0012209532,0.0025296619,0.009525599,0.00004481946,0.0008428869,0.0008660329,0.8672056,0.0005967102],"about_ca_topic_score_codex":0.008197565,"about_ca_topic_score_gemma":0.009671071,"teacher_disagreement_score":0.21738414,"about_ca_system_score_codex":0.0002214405,"about_ca_system_score_gemma":0.0021949692,"threshold_uncertainty_score":0.99840695},"labels":[],"label_agreement":null},{"id":"W2802788593","doi":"10.1017/cem.2018.78","title":"LO16: Showing your work: experiences with mind maps and faculty teaching","year":2018,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Cognition; Concept map; Critical thinking; Task (project management); Cognitive map; Psychology; Cognitive psychology; Medicine; Mathematics education; Psychiatry","score_opus":0.15324870265944923,"score_gpt":0.4071301056592043,"score_spread":0.2538814029997551,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2802788593","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98328096,0.0021716512,0.00019009602,0.0066721854,0.0021853095,0.0000981779,0.0000034637362,0.0000047142526,0.0053934506],"genre_scores_gemma":[0.9946817,0.000057359204,0.00095064484,0.0003884746,0.0022383723,0.0000020750847,0.000006535184,0.000017016031,0.0016578113],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99850285,0.000041381532,0.0005992763,0.00018940472,0.00029600103,0.00037109453],"domain_scores_gemma":[0.99734634,0.0003596011,0.00023089122,0.00016124072,0.0003340982,0.0015678505],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0007335078,0.00016594984,0.00045814778,0.00021844581,0.0001994665,0.0000083422,0.00012441028,0.000080685524,0.0033009762],"category_scores_gemma":[0.014610308,0.00010255607,0.000060926675,0.0002505364,0.00040066466,0.00009036619,0.000008931525,0.00040699277,0.000009359795],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003152913,0.00008395354,0.6820321,0.00004003577,0.00036808322,0.0010925394,0.090876035,0.000002401689,0.00015917823,0.00013080718,0.19516516,0.029734418],"study_design_scores_gemma":[0.012622531,0.019783989,0.45531452,0.021843437,0.0034345104,0.002662812,0.22682327,0.00006492924,0.00040807016,0.00048321942,0.2552242,0.0013345184],"about_ca_topic_score_codex":0.0024695203,"about_ca_topic_score_gemma":0.0027436756,"teacher_disagreement_score":0.22671758,"about_ca_system_score_codex":0.000054160762,"about_ca_system_score_gemma":0.00054890045,"threshold_uncertainty_score":0.99761015},"labels":[],"label_agreement":null},{"id":"W2802789733","doi":"10.15694/mep.2018.0000091.1","title":"The use of 'Precision Teaching' in enhancing medical students’ dermatological diagnostic skills","year":2018,"lang":"en","type":"article","venue":"MedEdPublish","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Queen's University; Queen's University Belfast","keywords":"Fluency; Medicine; Intervention (counseling); Medical education; Test (biology); Medical physics; Psychology; Dermatology; Mathematics education; Nursing","score_opus":0.03254768886495296,"score_gpt":0.3676839079772386,"score_spread":0.33513621911228564,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2802789733","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9932877,0.0001593046,0.00053813163,0.0038652166,0.00074101123,0.0003660418,0.0000036347474,0.00006052231,0.0009784602],"genre_scores_gemma":[0.99563515,0.0002413639,0.0008064322,0.0024879458,0.00035105884,0.00003200611,0.000019070867,0.000018914407,0.0004080508],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99623704,0.00033811762,0.00091852125,0.0003658669,0.0016838742,0.00045657594],"domain_scores_gemma":[0.8869729,0.11153871,0.00020209365,0.0006570943,0.00022662144,0.00040257935],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0032469712,0.00017196535,0.0005101964,0.00012281377,0.00014993366,0.00018064794,0.000561457,0.00028956085,0.00030594217],"category_scores_gemma":[0.6529489,0.00010095835,0.00013401145,0.00031946777,0.00050272467,0.00017132751,0.00037683474,0.0006866163,0.000058808753],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012050014,0.001404562,0.86970776,0.000013086343,0.000038622788,0.00019185965,0.00040716334,0.0000011911535,0.000025425752,0.0007276455,0.04649042,0.08087174],"study_design_scores_gemma":[0.0027967377,0.000583602,0.952021,0.003883938,0.00007017168,0.00015170354,0.00015680159,0.00027236578,0.0007874804,0.0005064108,0.038580418,0.00018938268],"about_ca_topic_score_codex":0.00016780545,"about_ca_topic_score_gemma":0.00056669954,"teacher_disagreement_score":0.64970195,"about_ca_system_score_codex":0.000056577224,"about_ca_system_score_gemma":0.00018658889,"threshold_uncertainty_score":0.4116962},"labels":[],"label_agreement":null},{"id":"W2803654236","doi":"10.1097/acm.0000000000002292","title":"Shifting and Sharing: Academic Physicians’ Strategies for Navigating Underperformance and Failure","year":2018,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":25,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; The Wilson Centre; University of Ottawa; Medical Council of Canada","funders":"","keywords":"Psychology; Academic medicine; MEDLINE; Medical education; Medicine; Family medicine; Political science","score_opus":0.0537097805176973,"score_gpt":0.4002192514744242,"score_spread":0.3465094709567269,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2803654236","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9821554,0.0032835456,0.0012428978,0.011895995,0.00020356254,0.00042069342,0.000003482832,0.00014015895,0.00065428484],"genre_scores_gemma":[0.98944116,0.0014359427,0.00093026983,0.004955833,0.002867815,0.000031272586,0.000018518524,0.000032007574,0.0002871587],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99828887,0.00001683886,0.00053062855,0.0004883629,0.00025682064,0.00041850278],"domain_scores_gemma":[0.9954915,0.0037570784,0.00018790271,0.00019253654,0.00011214413,0.00025880462],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0008852211,0.00021809917,0.00049119367,0.000052115236,0.00019550574,0.000010787912,0.00011609765,0.00035492048,0.000028417135],"category_scores_gemma":[0.009470365,0.00016697506,0.00003693125,0.00018383695,0.00077517104,0.00017231221,0.00009758624,0.001231493,0.000006808805],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007506584,0.000042947737,0.62592787,0.0015810198,0.00038861015,0.00003081036,0.038080767,0.0000055613473,0.025799608,0.045986596,0.03518391,0.22622167],"study_design_scores_gemma":[0.044074494,0.013971933,0.31386125,0.19057903,0.0038659258,0.0014567818,0.12428754,0.07598024,0.007878982,0.17737713,0.043534875,0.003131814],"about_ca_topic_score_codex":0.000032796295,"about_ca_topic_score_gemma":0.000005512908,"teacher_disagreement_score":0.31206658,"about_ca_system_score_codex":0.000024777823,"about_ca_system_score_gemma":0.00006874082,"threshold_uncertainty_score":0.9988733},"labels":[],"label_agreement":null},{"id":"W2804613946","doi":"10.1016/j.jclinepi.2018.04.024","title":"Qualitative research is a fundamental scientific process","year":2018,"lang":"en","type":"article","venue":"Journal of Clinical Epidemiology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":46,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Michael's Hospital; University of Toronto","funders":"University of Toronto","keywords":"Qualitative research; Framing (construction); Epistemology; Engineering ethics; Process (computing); Management science; Sociology; Computer science; Social science; Engineering; Philosophy","score_opus":0.7230013453676123,"score_gpt":0.7349159798712153,"score_spread":0.011914634503602972,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2804613946","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.954559,0.00043745674,0.0004486114,0.03931487,0.0023378392,0.00016014819,0.0000069526836,0.0000114246795,0.002723682],"genre_scores_gemma":[0.9819022,0.00015794822,0.0051564453,0.0077369027,0.0035050479,0.000003470319,0.0000036580486,0.0000150289425,0.0015193201],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9862063,0.007242086,0.0047296826,0.00046642576,0.0007390014,0.00061652873],"domain_scores_gemma":[0.25246495,0.73463744,0.0036823591,0.0008439964,0.0064432295,0.0019280085],"candidate_categories":["metaresearch","sts","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.13043542,0.00013633908,0.002188294,0.0002544143,0.0001930643,0.000019658008,0.000308585,0.00039851578,0.0009670939],"category_scores_gemma":[0.8467732,0.00008830959,0.00071521744,0.000504886,0.0044365553,0.000083344145,0.00009060091,0.0019011538,0.00052518776],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004800435,0.0024249768,0.61033505,0.000066731816,0.00046969118,0.0001901092,0.007963952,0.0000014746619,0.000054243596,0.0037191145,0.33190244,0.03807177],"study_design_scores_gemma":[0.011809628,0.035061695,0.5624988,0.005454883,0.0003890969,0.0011151931,0.010029035,0.00073963054,0.0003654684,0.28300044,0.08914131,0.00039483458],"about_ca_topic_score_codex":0.000017817265,"about_ca_topic_score_gemma":0.0000052098453,"teacher_disagreement_score":0.7337413,"about_ca_system_score_codex":0.000055616623,"about_ca_system_score_gemma":0.0012228832,"threshold_uncertainty_score":0.9999462},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"qualitative","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2805513289","doi":"10.1097/acm.0000000000002303","title":"How Different Theories of Clinical Reasoning Influence Teaching and Assessment","year":2018,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":44,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Psychology; MEDLINE; Medical education; Mathematics education; Medicine; Chemistry","score_opus":0.05453971887788701,"score_gpt":0.457705741486721,"score_spread":0.403166022608834,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2805513289","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97702634,0.00087813206,0.0013047697,0.0139373075,0.00041185843,0.00021480981,0.0000020595917,0.000053157666,0.0061715795],"genre_scores_gemma":[0.99190956,0.0013677857,0.0015712737,0.002229888,0.0017756053,0.000008296834,0.000009697846,0.000017251616,0.0011106174],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9977264,0.0002599923,0.0008598881,0.00038655353,0.0004973897,0.0002697972],"domain_scores_gemma":[0.98328525,0.015407318,0.00037754272,0.0003557215,0.00017180425,0.00040236174],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.003269396,0.00018841091,0.00088536425,0.00007748329,0.000101487916,0.0000067383035,0.00013586851,0.00028694666,0.000055218992],"category_scores_gemma":[0.17194131,0.00011510221,0.000063507534,0.00009849213,0.0014388274,0.00007273511,0.0001172902,0.0017550141,0.0000028126267],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00019316391,0.000049099148,0.8629294,0.000067914414,0.00010024004,0.000015831038,0.001553776,1.1246829e-7,0.0012174454,0.04188585,0.0025858479,0.089401305],"study_design_scores_gemma":[0.003981488,0.0027833532,0.9674103,0.009722612,0.0004119904,0.00009486992,0.0018581892,0.00038018683,0.00039781939,0.0073233503,0.005461454,0.00017435417],"about_ca_topic_score_codex":0.00007234812,"about_ca_topic_score_gemma":0.0000019829488,"teacher_disagreement_score":0.1686719,"about_ca_system_score_codex":0.00002076118,"about_ca_system_score_gemma":0.000103881954,"threshold_uncertainty_score":0.8350338},"labels":[],"label_agreement":null},{"id":"W2806800554","doi":"10.1016/j.jns.2018.06.003","title":"Miming neurological syndromes improves medical student's long-term retention and delayed recall of neurology","year":2018,"lang":"en","type":"article","venue":"Journal of the Neurological Sciences","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":33,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Semiology; Neurology; Recall; Psychology; Class (philosophy); Neurological examination; Test (biology); Medicine; Pediatrics; Psychiatry; Cognitive psychology; Artificial intelligence","score_opus":0.04963455002730425,"score_gpt":0.37211239113282224,"score_spread":0.32247784110551797,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2806800554","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9748111,0.00023810839,0.000037980815,0.023579642,0.0010319673,0.00012958948,7.9558265e-7,0.000012844803,0.00015798886],"genre_scores_gemma":[0.99322045,0.0003050516,0.00019252147,0.0059461948,0.00030636508,0.000001589226,1.623055e-7,0.0000048393836,0.000022844597],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9971241,0.00043521204,0.00078593136,0.000324891,0.00102153,0.00030833916],"domain_scores_gemma":[0.9925465,0.0061578364,0.0006000624,0.0001828057,0.00018397899,0.00032879977],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0022254682,0.00014098601,0.0005139484,0.00008351438,0.00017531619,0.000032331678,0.0005596034,0.00023771718,0.00016722454],"category_scores_gemma":[0.04228683,0.00006482558,0.00023565322,0.00030440703,0.002563716,0.00007913276,0.00034249184,0.00068874843,0.0000027051904],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00091986376,0.00042981948,0.9803224,0.000009074439,0.00003395306,0.0004206986,0.000043168027,0.0000035826067,0.0014585836,0.00010578917,0.00024132343,0.01601174],"study_design_scores_gemma":[0.00082274765,0.020414956,0.97315454,0.00012853426,0.000120163255,0.003947599,0.0000065693075,0.00051512674,0.00011137336,0.0006634947,0.000055765562,0.00005914185],"about_ca_topic_score_codex":0.0000044007143,"about_ca_topic_score_gemma":0.00000414857,"teacher_disagreement_score":0.040061362,"about_ca_system_score_codex":0.0000036991394,"about_ca_system_score_gemma":0.00010263497,"threshold_uncertainty_score":0.9657804},"labels":[],"label_agreement":null},{"id":"W2806932928","doi":"10.1212/wnl.88.16_supplement.p1.048","title":"Bad News or Life Altering Diagnoses? Neurology Residents’ Attitudes towards Disclosing Diagnoses (P1.048)","year":2017,"lang":"en","type":"article","venue":"Neurology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta; Alberta Health Services","funders":"","keywords":"Medical diagnosis; Neurology; Medicine; Family medicine; Psychiatry; Psychology; Pathology","score_opus":0.06814587694670039,"score_gpt":0.38686077821166476,"score_spread":0.3187149012649644,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2806932928","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9447183,0.0004197648,0.0000709456,0.048566386,0.0024948418,0.00044629743,0.000016469785,0.00021457688,0.0030524288],"genre_scores_gemma":[0.96532464,0.0021688114,0.00019361741,0.030116988,0.0016756796,0.000098232995,0.000021421862,0.000088569774,0.0003120426],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9966058,0.00027529136,0.0007519595,0.0010515386,0.00039411092,0.0009212944],"domain_scores_gemma":[0.98562956,0.011459177,0.00044045591,0.0016595504,0.00010483096,0.00070643594],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00037143152,0.00042968683,0.0010349451,0.00015841785,0.00053739554,0.00017219501,0.0006760782,0.00047617077,0.0005916945],"category_scores_gemma":[0.12670842,0.00033850118,0.00026069194,0.00010003049,0.0006710293,0.0001962267,0.0006776463,0.0009667346,0.0002562227],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0026751456,0.0005488473,0.9619238,0.000048949558,0.00009576247,0.00438687,0.00007543666,0.000024528908,0.0001791302,0.00027606092,0.0070596123,0.022705883],"study_design_scores_gemma":[0.0029888968,0.003980645,0.9801274,0.000261208,0.00018817087,0.0005229044,0.000004007833,0.00009377833,0.00017419622,0.00026490225,0.011112825,0.00028108727],"about_ca_topic_score_codex":0.0032218026,"about_ca_topic_score_gemma":0.0014802178,"teacher_disagreement_score":0.12633699,"about_ca_system_score_codex":0.000018015384,"about_ca_system_score_gemma":0.00031357346,"threshold_uncertainty_score":0.9999067},"labels":[],"label_agreement":null},{"id":"W2807627726","doi":"10.1007/s10459-018-9838-x","title":"Optimizing self-regulation of performance: is mental effort a cue?","year":2018,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":22,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Jewish General Hospital; Western University","funders":"","keywords":"Certainty; Judgement; Task (project management); Psychology; Cognitive psychology; Rating scale; Path (computing); Scale (ratio); Applied psychology; Social psychology; Computer science; Developmental psychology","score_opus":0.020299733399530024,"score_gpt":0.41908537193386813,"score_spread":0.3987856385343381,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2807627726","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9881702,0.002553218,0.00011579819,0.0021255375,0.0011534463,0.0003563933,0.0000010377199,0.000024172825,0.005500198],"genre_scores_gemma":[0.9705185,0.002641082,0.025434667,0.0010822063,0.00020555749,0.000019188803,0.000008474855,0.000003830153,0.00008646831],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9988308,0.000019874065,0.0004224691,0.00022998427,0.00028931993,0.00020758554],"domain_scores_gemma":[0.9990202,0.00042607108,0.00021477524,0.0001320934,0.0001035119,0.00010334451],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00081875705,0.000067711604,0.00016939982,0.00014034781,0.00014194958,0.000007438285,0.00007794977,0.00003315021,0.000086622975],"category_scores_gemma":[0.0008705648,0.000056654037,0.000024272733,0.0005611836,0.00027018943,0.00029507335,0.0000151905015,0.00006593597,0.000013464547],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005738593,0.0005218055,0.7288516,0.00025765857,0.0000032194719,6.072447e-8,0.005502674,0.0001304295,0.000044686374,0.0009180929,0.00054861896,0.26316375],"study_design_scores_gemma":[0.0014385113,0.004240459,0.92179203,0.008388234,0.000022877837,0.00003564765,0.002446797,0.02828012,0.0047842474,0.0016702758,0.026643714,0.00025705883],"about_ca_topic_score_codex":0.00012630214,"about_ca_topic_score_gemma":0.000023893515,"teacher_disagreement_score":0.2629067,"about_ca_system_score_codex":0.00012553053,"about_ca_system_score_gemma":0.0009585467,"threshold_uncertainty_score":0.23102845},"labels":[],"label_agreement":null},{"id":"W2810200535","doi":"10.1080/0142159x.2018.1484084","title":"Examining the role of self-regulation and emotion in clinical reasoning: Implications for developing expertise","year":2018,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":26,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Reciprocal; Subcategory; Psychology; Context (archaeology); Clinical Practice; Cognitive psychology; Social psychology; Developmental psychology; Medicine","score_opus":0.07115469651765942,"score_gpt":0.39962767977959984,"score_spread":0.3284729832619404,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2810200535","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9882825,0.00028919976,0.0044138143,0.005211223,0.00009219721,0.00031837198,6.363435e-7,0.000039185143,0.0013528491],"genre_scores_gemma":[0.9947602,0.00007883231,0.0041095186,0.00037070186,0.00049601414,0.000049400667,0.0000120455725,0.00001230242,0.00011097934],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99876314,0.000111672976,0.00054186257,0.00022972893,0.00019234845,0.00016125171],"domain_scores_gemma":[0.9962764,0.0031221316,0.00011648766,0.00022420833,0.00011152463,0.00014925336],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0020025305,0.00007831826,0.00024417762,0.00003882683,0.00006005074,0.000006932351,0.00007377591,0.00019750939,0.00007965912],"category_scores_gemma":[0.0606168,0.000052120995,0.000049744274,0.0001336082,0.0002546479,0.00002670959,0.000043890162,0.00019026244,0.000004115712],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000046345678,0.00017227394,0.763561,0.000007863925,0.000022096903,4.1047755e-7,0.0007948645,7.307991e-8,0.00004308115,0.0020370383,0.0004175566,0.23289739],"study_design_scores_gemma":[0.0013658808,0.00026724808,0.9834336,0.00083094597,0.0000499674,0.000015800048,0.0002788979,0.009794806,0.0000898365,0.0017487973,0.0020583575,0.00006587821],"about_ca_topic_score_codex":0.000033378812,"about_ca_topic_score_gemma":0.000015282056,"teacher_disagreement_score":0.2328315,"about_ca_system_score_codex":0.00003635684,"about_ca_system_score_gemma":0.00017091124,"threshold_uncertainty_score":0.947296},"labels":[],"label_agreement":null},{"id":"W2810255362","doi":"10.1542/peds.2018-1238","title":"Just Do It: Incorporating Bedside Teaching Into Every Patient Encounter","year":2018,"lang":"en","type":"review","venue":"PEDIATRICS","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Medicine; Medical education; Patient care; Medical emergency; Medical physics; Nursing","score_opus":0.056520010541392246,"score_gpt":0.39505613851025045,"score_spread":0.33853612796885824,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2810255362","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0002875036,0.99056184,0.00014849701,0.00013119674,0.001891483,0.0007399388,0.000042649724,0.00012114874,0.0060757687],"genre_scores_gemma":[0.00023015283,0.98658556,0.004353043,0.0007765288,0.0069221696,0.000079445526,0.00033370176,0.00011668237,0.00060269365],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99623376,0.00022860922,0.0015606458,0.0007685809,0.0007446795,0.0004637261],"domain_scores_gemma":[0.9830398,0.014395463,0.0010541908,0.0008055904,0.00024096946,0.00046398494],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012511828,0.00057722925,0.0019693302,0.0003490049,0.00019408822,0.000097767785,0.00026577164,0.00069735036,0.000115955816],"category_scores_gemma":[0.08564802,0.00043585635,0.00076042005,0.00044692258,0.00010495948,0.00007569048,0.00032458067,0.001541743,0.0008728236],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000009540344,0.00021731592,0.0009838907,0.0064729545,0.00006399442,0.00015069905,0.00033001835,3.3702386e-7,1.1705667e-8,0.00004128483,0.063456595,0.9282734],"study_design_scores_gemma":[0.0003199727,0.00042363332,0.000016320713,0.033031184,0.0032459204,0.00010900221,0.00009433031,0.000009059345,8.797072e-8,0.00013447691,0.96214074,0.00047529466],"about_ca_topic_score_codex":0.00010110044,"about_ca_topic_score_gemma":0.000009090845,"teacher_disagreement_score":0.9277981,"about_ca_system_score_codex":0.00034075018,"about_ca_system_score_gemma":0.0010604981,"threshold_uncertainty_score":0.9999051},"labels":[],"label_agreement":null},{"id":"W2845768007","doi":"10.1016/j.japh.2018.06.016","title":"Perceived preparedness for full-scope pharmacist services among recent Doctor of Pharmacy graduates from Ontario schools of pharmacy","year":2018,"lang":"en","type":"article","venue":"Journal of the American Pharmacists Association","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Pharmacy; Scope of practice; Pharmacist; Pharmacy practice; Preparedness; Feeling; Graduation (instrument); Family medicine; Nursing; Scope (computer science); Clinical pharmacy; Medical education; Health care; Psychology","score_opus":0.0397087546811891,"score_gpt":0.38450837794267045,"score_spread":0.34479962326148134,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2845768007","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99590737,0.00022538965,0.000050350816,0.0012781371,0.0015388405,0.00060182496,0.00020541663,0.00001937297,0.0001733127],"genre_scores_gemma":[0.9954252,0.0011603057,0.0010115571,0.00081025856,0.0011211459,0.000016170694,0.00003759562,0.000036761056,0.0003809722],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9966388,0.00027649355,0.0014797223,0.00027372694,0.0009804097,0.00035086568],"domain_scores_gemma":[0.9889086,0.0028637182,0.005390654,0.00028280774,0.0022548796,0.00029935673],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0011403142,0.00026936742,0.0010539268,0.00013376736,0.00012082749,0.00003922383,0.000494034,0.000042081076,0.00037631058],"category_scores_gemma":[0.0052293446,0.0001946734,0.00058939867,0.000450139,0.00025294966,0.0002682429,0.00009051904,0.0004908505,0.000009626756],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00693881,0.0009190031,0.8457446,0.000082312145,0.00123022,0.000008164585,0.0028168343,0.000027030821,0.11925887,9.615056e-7,0.010099551,0.012873615],"study_design_scores_gemma":[0.03383433,0.0017145786,0.649505,0.0021124852,0.005816443,0.00008206365,0.0006378592,0.0033058152,0.28329796,0.00024238166,0.018893257,0.0005578263],"about_ca_topic_score_codex":0.004590282,"about_ca_topic_score_gemma":0.00037805692,"teacher_disagreement_score":0.19623962,"about_ca_system_score_codex":0.0009623259,"about_ca_system_score_gemma":0.00076683634,"threshold_uncertainty_score":0.7938551},"labels":[],"label_agreement":null},{"id":"W2864065416","doi":"","title":"What Counts as 'What Works': Expertise, Mechanisms and Values in Evidence-Based Medicine","year":2018,"lang":"en","type":"dissertation","venue":"Durham e-Theses (Durham University)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Queen's University; University of Toronto; McMaster University","keywords":"Engineering ethics; Epistemology; Evidence-based medicine; Movement (music); Alternative medicine; Psychology; Randomized controlled trial; Medicine; Philosophy; Engineering; Pathology","score_opus":0.0433940663525598,"score_gpt":0.3289481851396284,"score_spread":0.2855541187870686,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2864065416","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.898071,0.05078447,0.00016477666,0.004816395,0.007367097,0.0016829504,0.000011786324,0.00035660274,0.03674491],"genre_scores_gemma":[0.61177826,0.25017163,0.0009936779,0.0038507404,0.001247759,0.000020627658,0.0010580604,0.00020481994,0.13067442],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99774957,0.00012751756,0.00038758785,0.00073062617,0.00061201525,0.00039266297],"domain_scores_gemma":[0.99442464,0.004162868,0.000252067,0.0005004972,0.00029744027,0.00036247124],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0003822348,0.00042336577,0.0008229417,0.00074368616,0.00011960972,0.00017381569,0.00025661415,0.00062807667,0.0019417086],"category_scores_gemma":[0.009829048,0.0003861591,0.00015867036,0.0006181331,0.00028709293,0.00091413344,0.000052620053,0.0004614063,0.00024650348],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.02230481,0.00298326,0.006394943,0.0028062165,0.0014906373,0.0136458175,0.024459172,0.00003032611,0.0018261024,0.012547492,0.046059135,0.8654521],"study_design_scores_gemma":[0.017185964,0.0055508474,0.018984763,0.7570183,0.0029245822,0.00015160322,0.11800776,0.0007992191,0.004494786,0.007520137,0.06450652,0.0028555028],"about_ca_topic_score_codex":0.00027239454,"about_ca_topic_score_gemma":0.0001817702,"teacher_disagreement_score":0.8625966,"about_ca_system_score_codex":0.00025932648,"about_ca_system_score_gemma":0.0004397612,"threshold_uncertainty_score":0.99985904},"labels":[],"label_agreement":null},{"id":"W2879709479","doi":"10.1051/pmed/2018002","title":"La rédaction de questions à choix multiple et de questions à réponse ouverte et courte pour les examens écrits dans les programmes de formation en santé : une étude docimologique des consignes de rédaction","year":2017,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval; Université de Sherbrooke","funders":"","keywords":"Humanities; Philosophy; Redaction; Theology","score_opus":0.06771558771740381,"score_gpt":0.4106548740306845,"score_spread":0.3429392863132807,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2879709479","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.63872313,0.004734667,0.32287616,0.030195296,0.00045312993,0.0007476589,0.00019400405,0.0004994945,0.0015764677],"genre_scores_gemma":[0.9321074,0.026972981,0.035744056,0.0010697566,0.0006828262,0.00040183228,0.00014297063,0.0000876006,0.0027905602],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9938396,0.0031034367,0.00080052455,0.0005812346,0.0003971381,0.0012780615],"domain_scores_gemma":[0.97118217,0.026297314,0.00054053345,0.00074293517,0.0003611171,0.00087594317],"candidate_categories":["metaresearch","metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.0050350823,0.0005135369,0.0006379142,0.00018046252,0.0013340021,0.00034550365,0.00040133146,0.0011352196,0.00010800839],"category_scores_gemma":[0.14292291,0.0005152084,0.00031990948,0.00018980428,0.0014176087,0.00061957835,0.00016378441,0.0015544121,0.000064048996],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00056163693,0.0035162054,0.5218253,0.0006147801,0.0003634013,0.000628223,0.027336549,0.0028278448,0.013021163,0.0045941295,0.010428421,0.41428235],"study_design_scores_gemma":[0.0034694078,0.0007350085,0.9217348,0.008542085,0.0010850141,0.0018916298,0.011649479,0.023209153,0.0046112244,0.0076680835,0.0147420075,0.0006620711],"about_ca_topic_score_codex":0.03132971,"about_ca_topic_score_gemma":0.025383174,"teacher_disagreement_score":0.4136203,"about_ca_system_score_codex":0.0022120695,"about_ca_system_score_gemma":0.0028000514,"threshold_uncertainty_score":0.99996614},"labels":[],"label_agreement":null},{"id":"W2881778650","doi":"10.3934/publichealth.2018.3.235","title":"Criterion scores, construct validity and reliability of a web-based instrument to assess physiotherapists’ clinical reasoning focused on behaviour change: ‘Reasoning 4 Change’","year":2018,"lang":"en","type":"article","venue":"AIMS Public Health","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Reliability (semiconductor); Construct validity; Construct (python library); Psychology; Context (archaeology); Applied psychology; Test (biology); Clinical psychology; Internal consistency; Convergent validity; Physical therapy; Psychometrics; Medicine; Computer science","score_opus":0.2400223441486042,"score_gpt":0.4518813126106048,"score_spread":0.2118589684620006,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2881778650","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9730818,0.000108172244,0.00033049926,0.023927493,0.00073396723,0.0014418241,0.00007742947,0.00014804302,0.00015074284],"genre_scores_gemma":[0.974275,0.00018428754,0.009452894,0.014790453,0.0010028196,0.00016567808,0.0000673351,0.00004916259,0.000012401513],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99573725,0.0007840053,0.0011213018,0.0008945639,0.0006894344,0.00077341666],"domain_scores_gemma":[0.9926463,0.0040123956,0.0005008856,0.0008964095,0.0004994923,0.0014445208],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.005110593,0.00032458917,0.0010709874,0.00022413315,0.00021978778,0.00007007108,0.00018239954,0.00026388682,0.00012031173],"category_scores_gemma":[0.028318934,0.00027719472,0.00018765024,0.00051288784,0.0005230563,0.00015032804,0.000117973635,0.00051421113,0.000017910532],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00041829475,0.0011148475,0.8459528,0.00020873817,0.00002178219,0.000008691567,0.0005867107,1.0373909e-7,0.000022651595,0.00027129127,0.001003912,0.15039013],"study_design_scores_gemma":[0.0044641537,0.008047876,0.97878814,0.0052449782,0.00003639083,0.00001640105,0.00008726089,0.0009728852,0.00014990888,0.000045603156,0.0018914506,0.00025497543],"about_ca_topic_score_codex":0.0021098386,"about_ca_topic_score_gemma":0.00012736542,"teacher_disagreement_score":0.15013516,"about_ca_system_score_codex":0.0002688405,"about_ca_system_score_gemma":0.0013140189,"threshold_uncertainty_score":0.99996805},"labels":[],"label_agreement":null},{"id":"W2884353789","doi":"10.1017/cem.2018.419","title":"Emergency medicine definitions, CAEP, and the journey to excellence","year":2018,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; St. Michael's Hospital","funders":"","keywords":"Excellence; Medicine; Content (measure theory); Action (physics); World Wide Web; Computer science; Law","score_opus":0.17455831729560806,"score_gpt":0.39971946787157436,"score_spread":0.2251611505759663,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2884353789","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.42426142,0.043621715,0.001453489,0.3872528,0.041988872,0.00093889766,0.000026661455,0.000031685442,0.10042443],"genre_scores_gemma":[0.96158,0.019477142,0.0002548372,0.0031272115,0.010688358,0.0000099793715,0.000008098737,0.000040374438,0.0048139924],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99691176,0.0001558122,0.001617593,0.0002569627,0.00054857205,0.0005092734],"domain_scores_gemma":[0.9932458,0.0016176763,0.00048283229,0.0004396332,0.0013769038,0.0028371508],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0025906987,0.00025153652,0.00077241065,0.00044212714,0.00035291698,0.000003898822,0.00030947238,0.00011196208,0.024902277],"category_scores_gemma":[0.090956606,0.00014004574,0.00015178072,0.00079339487,0.0010564836,0.000067515386,0.000024291588,0.00050656346,0.00008490446],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00028887167,0.000033064898,0.027818264,0.0000326824,0.00014937155,0.00013493578,0.0043719746,0.0000013344542,0.0001389496,0.0038735948,0.9605642,0.0025927774],"study_design_scores_gemma":[0.016453387,0.011128908,0.15108578,0.0127362665,0.0025427672,0.0014603519,0.0071313772,0.00010805234,0.000080156686,0.02446473,0.7720798,0.0007283859],"about_ca_topic_score_codex":0.010748936,"about_ca_topic_score_gemma":0.011553094,"teacher_disagreement_score":0.5373186,"about_ca_system_score_codex":0.000078464545,"about_ca_system_score_gemma":0.0007531275,"threshold_uncertainty_score":0.9958386},"labels":[],"label_agreement":null},{"id":"W2884774198","doi":"10.1080/0142159x.2018.1481281","title":"Twelve tips for developing key-feature questions (KFQ) for effective assessment of clinical reasoning","year":2018,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":32,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Waterloo; University of Toronto","funders":"","keywords":"Key (lock); Feature (linguistics); Psychology; Computer science; Medical education; Medicine; Philosophy; Linguistics","score_opus":0.05833185039754752,"score_gpt":0.49036635430181824,"score_spread":0.43203450390427073,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2884774198","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.49025664,0.00043679893,0.46621937,0.02390434,0.0026222982,0.0036661495,0.0000412273,0.0002456495,0.012607533],"genre_scores_gemma":[0.85997194,0.00006012734,0.13013734,0.0017776664,0.0041982895,0.00054669194,0.00012671942,0.00005811886,0.0031231414],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99759215,0.00019482725,0.0007776859,0.0004796559,0.00056793896,0.0003877128],"domain_scores_gemma":[0.9809565,0.017252056,0.00027339763,0.00038059035,0.0006009619,0.0005365146],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004332645,0.00019320051,0.0008310226,0.000059366423,0.00011819574,0.000011538972,0.00015789334,0.0005754054,0.00028265023],"category_scores_gemma":[0.2667919,0.00014349542,0.0004043938,0.0001632486,0.00045762624,0.00003258488,0.00006874829,0.0006198884,0.00001323287],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011290879,0.0019003928,0.47145918,0.00032667085,0.00083136576,0.00002807808,0.000507535,7.92311e-7,0.00003300881,0.018850364,0.15749471,0.3474388],"study_design_scores_gemma":[0.019744596,0.007379255,0.6637546,0.01556628,0.0013058329,0.000103311824,0.0002774826,0.040708415,0.00037156043,0.0039797192,0.2461174,0.00069154636],"about_ca_topic_score_codex":0.000027858247,"about_ca_topic_score_gemma":0.000020214346,"teacher_disagreement_score":0.36971527,"about_ca_system_score_codex":0.000091868285,"about_ca_system_score_gemma":0.00092835416,"threshold_uncertainty_score":0.73938423},"labels":[],"label_agreement":null},{"id":"W2888772709","doi":"","title":"How to Think in Medicine: Reasoning, Decision Making, and Communication in Health Sciences and Professions","year":2018,"lang":"en","type":"book","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Health care; Argumentation theory; Decision aids; Psychology; Engineering ethics; Medical education; Medicine; Alternative medicine; Epistemology; Engineering; Political science","score_opus":0.05501307370034094,"score_gpt":0.4329903343861373,"score_spread":0.3779772606857964,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2888772709","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.09264938,0.03675086,0.0010600286,0.22439627,0.00083456276,0.0059117544,0.000021556594,0.00024006398,0.63813555],"genre_scores_gemma":[0.16742323,0.024734717,0.10853497,0.022264715,0.00067233335,0.00010869784,0.00014440856,0.000108702625,0.6760082],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9980038,0.00015783559,0.0005763435,0.000543383,0.00042614085,0.00029248834],"domain_scores_gemma":[0.98521763,0.0136616565,0.0002432625,0.00049364474,0.00008656611,0.0002972322],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.003148297,0.00021943574,0.000749133,0.0004814036,0.00015239857,0.000045097957,0.00018007815,0.00030906303,0.000082773084],"category_scores_gemma":[0.062295236,0.00014795687,0.000024906723,0.0002733658,0.0006507191,0.000059039,0.0002825196,0.0005768864,0.000006903402],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00028673973,0.00024691503,0.08339357,0.00034151674,0.000021668124,0.000043375272,0.006023247,0.0000012351468,0.0000016024258,0.029799676,0.701354,0.17848642],"study_design_scores_gemma":[0.005518776,0.0047915936,0.33328095,0.48509794,0.00008565846,0.00014977567,0.0029258064,0.0007751037,0.0000016091342,0.06434849,0.10231493,0.0007093449],"about_ca_topic_score_codex":0.00045895207,"about_ca_topic_score_gemma":0.0025110052,"teacher_disagreement_score":0.59903914,"about_ca_system_score_codex":0.00016317294,"about_ca_system_score_gemma":0.0007192405,"threshold_uncertainty_score":0.94560343},"labels":[],"label_agreement":null},{"id":"W2889026671","doi":"10.7759/cureus.3205","title":"MedsOnCall Pager App: A Pilot Project for Practicing Safe Clinical Decision-making","year":2018,"lang":"en","type":"article","venue":"Cureus","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; Ottawa Hospital","funders":"University of Ottawa","keywords":"Pager; Preparedness; Medicine; Medical education; Clinical decision making; Family medicine; Computer science","score_opus":0.11532986777428891,"score_gpt":0.491314050658109,"score_spread":0.3759841828838201,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2889026671","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.83746165,0.001555908,0.09340996,0.0051924167,0.005644561,0.0040590884,0.000043210173,0.00066930946,0.05196388],"genre_scores_gemma":[0.9240892,0.0001514024,0.06754881,0.0036036088,0.0035498207,0.000082542705,0.000014427307,0.000051954794,0.0009082129],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99778837,0.00008493027,0.0008008353,0.0005416025,0.00035028803,0.00043396052],"domain_scores_gemma":[0.940352,0.05812354,0.00026814465,0.0006255273,0.00036902147,0.0002617683],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0019419607,0.00019266046,0.0005453164,0.000057527723,0.0001468762,0.00004265905,0.00016068069,0.00019819198,0.00029744933],"category_scores_gemma":[0.3039463,0.00014486678,0.0002475349,0.00018883686,0.00021002586,0.000084758096,0.0000952275,0.00033393528,0.00023443125],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00652888,0.002833639,0.046550076,0.00007299266,0.00038457455,0.00017103019,0.0005947235,0.0000011405858,0.00016108279,0.0013933135,0.3650037,0.57630485],"study_design_scores_gemma":[0.023359839,0.03850041,0.13751723,0.020353341,0.002532348,0.0006511612,0.0009988537,0.009256394,0.00046291648,0.012893884,0.75192815,0.0015454899],"about_ca_topic_score_codex":0.00004309205,"about_ca_topic_score_gemma":0.000056233337,"teacher_disagreement_score":0.57475936,"about_ca_system_score_codex":0.000069086425,"about_ca_system_score_gemma":0.0003896713,"threshold_uncertainty_score":0.7019169},"labels":[],"label_agreement":null},{"id":"W2889406775","doi":"10.1007/s10459-018-9848-8","title":"Is the mouth the mirror of the mind?","year":2018,"lang":"en","type":"editorial","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Psychology; Medical education; Computer science; Mathematics education; Dentistry; Medicine","score_opus":0.02915197083426731,"score_gpt":0.46108983840610374,"score_spread":0.43193786757183644,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2889406775","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.011375679,0.017367765,0.0000034779466,0.047228083,0.92068905,0.0010621258,0.00003483302,0.0000074404006,0.002231556],"genre_scores_gemma":[0.096824825,0.0307925,0.0025258858,0.014007845,0.84749305,0.00042683422,0.00008539321,0.000045816014,0.007797859],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99739134,0.00021601544,0.00066508335,0.00039466677,0.0010286145,0.00030425098],"domain_scores_gemma":[0.97083485,0.026724255,0.0010615076,0.00088836084,0.0003766625,0.00011435471],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.003135359,0.00014482724,0.00030644753,0.00006328179,0.0003806919,0.000025645053,0.0007444818,0.00016428283,0.00006516834],"category_scores_gemma":[0.058329128,0.000058582315,0.00010304302,0.00081858324,0.0015931862,0.00008106752,0.00006850838,0.00055657915,0.00001806449],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000023601393,0.00016443539,0.0071110576,0.00014556093,0.000003965764,5.421691e-8,0.002434463,0.000003950367,2.9738095e-7,0.00012675759,0.9464319,0.043553967],"study_design_scores_gemma":[0.00018229755,0.0002663658,0.010784248,0.0044700224,0.000026086665,0.0000017793184,0.0015415725,0.000029990295,0.000018626757,0.0019623183,0.9806444,0.000072288865],"about_ca_topic_score_codex":0.0020869342,"about_ca_topic_score_gemma":0.0007073157,"teacher_disagreement_score":0.085449144,"about_ca_system_score_codex":0.00013392183,"about_ca_system_score_gemma":0.007033553,"threshold_uncertainty_score":0.99859565},"labels":[],"label_agreement":null},{"id":"W2889907846","doi":"10.15621/ijphy/2015/v2i6/80761","title":"MANAGEMENT OF A GUILLAIN BARRE SYNDROME PATIENT THROUGH THREE TRACK REASONING: A CASE STUDY","year":2015,"lang":"en","type":"article","venue":"International Journal of Physiotherapy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Reasoning system; Psychology of reasoning; Verbal reasoning; Deductive reasoning; Opportunistic reasoning; Process (computing); Model-based reasoning; Adaptive reasoning; Case-based reasoning; Analytic reasoning; Qualitative reasoning; Causal reasoning; Inductive reasoning; Computer science; Artificial intelligence; Medicine; Cognition; Knowledge representation and reasoning; Psychiatry","score_opus":0.0406988620319853,"score_gpt":0.39977195226175644,"score_spread":0.35907309022977113,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2889907846","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9965972,0.0005317425,0.00059133786,0.0006557609,0.0006728951,0.00026102568,0.0000073779593,0.000009806075,0.00067285734],"genre_scores_gemma":[0.9946852,0.00011170415,0.0046585626,0.00025924103,0.00021033338,0.000003605549,0.0000025261834,0.00001660946,0.000052261646],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9980417,0.00005257734,0.0006825765,0.00014671194,0.00095609436,0.000120336554],"domain_scores_gemma":[0.9978808,0.00049364613,0.0004984891,0.0002028832,0.0007548017,0.00016936606],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00034047576,0.00013238016,0.00038874007,0.00009390249,0.000016262175,0.000020263218,0.00020544628,0.000040693438,0.00012815131],"category_scores_gemma":[0.0007534083,0.00009836876,0.00021631487,0.000119682336,0.000048084807,0.00009761906,0.000032377106,0.0001602742,0.00000720558],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008438563,0.021543728,0.28282297,0.000093834926,0.0124219535,0.31066087,0.019584607,0.0024459667,0.00042395425,0.0013529918,0.008107687,0.3321029],"study_design_scores_gemma":[0.09428175,0.051272225,0.57908386,0.022098111,0.0015159727,0.16594386,0.058609255,0.0020428575,0.0017878315,0.017034749,0.0050951527,0.001234376],"about_ca_topic_score_codex":0.00029533752,"about_ca_topic_score_gemma":0.000016007307,"teacher_disagreement_score":0.3308685,"about_ca_system_score_codex":0.00010366216,"about_ca_system_score_gemma":0.00012332633,"threshold_uncertainty_score":0.40113616},"labels":[],"label_agreement":null},{"id":"W2890213943","doi":"10.1515/dx-2018-0033","title":"Improving diagnosis by improving education: a policy brief on education in healthcare professions","year":2018,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":82,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Health care; Process (computing); Quality (philosophy); Medical education; Medicine; Knowledge base; Patient safety; Psychology; Nursing; Computer science; Political science","score_opus":0.020678051955281877,"score_gpt":0.39254283439134036,"score_spread":0.37186478243605847,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2890213943","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9081085,0.0016906299,0.00002093458,0.084608994,0.0013994954,0.0012761318,0.000050480823,0.00014295724,0.0027018823],"genre_scores_gemma":[0.9673678,0.00076439546,0.00071096944,0.02511605,0.0020390402,0.002566794,0.00017449952,0.000059001326,0.0012014478],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9976231,0.00014010606,0.00061193714,0.00069001404,0.00036968247,0.0005651546],"domain_scores_gemma":[0.99331015,0.004836515,0.00024203827,0.00067204476,0.0003611074,0.0005781598],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.000374456,0.0002815061,0.00043779775,0.00040075256,0.00022935582,0.00005852273,0.0001791716,0.00027566057,0.00021050121],"category_scores_gemma":[0.046277396,0.0002589467,0.00013756962,0.0008376503,0.00014129978,0.00012835977,0.000086969485,0.00047503813,0.0001935032],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000044379478,0.0024439741,0.46332854,0.000071993745,0.000008910163,0.0000025250424,0.0002652292,1.1797224e-7,0.000041424773,0.0009240728,0.068959884,0.46390894],"study_design_scores_gemma":[0.00308763,0.0030575267,0.9200578,0.014539984,0.0001876366,0.000030695057,0.0015229319,0.00011192046,0.00916792,0.0013887164,0.045969173,0.00087809784],"about_ca_topic_score_codex":0.023229383,"about_ca_topic_score_gemma":0.000600099,"teacher_disagreement_score":0.46303084,"about_ca_system_score_codex":0.000537044,"about_ca_system_score_gemma":0.0043709837,"threshold_uncertainty_score":0.9999863},"labels":[],"label_agreement":null},{"id":"W2890605886","doi":"10.1007/s11606-018-4667-4","title":"General Internists in Pursuit of Diagnostic Excellence","year":2018,"lang":"en","type":"letter","venue":"Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Mount Sinai Hospital","funders":"Health Services Research and Development","keywords":"Medicine; Excellence; MEDLINE; Family medicine; Intensive care medicine; Epistemology","score_opus":0.035173202361742116,"score_gpt":0.35944918870884796,"score_spread":0.32427598634710586,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2890605886","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.34472308,0.0066181556,0.0005170764,0.6315061,0.009778483,0.00043053145,0.000031880263,0.00001971002,0.006375022],"genre_scores_gemma":[0.15087992,0.008371888,0.0033823717,0.6593397,0.118271634,0.000020154806,0.00014124779,0.00023613752,0.059356928],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9938387,0.00024091618,0.0030386345,0.0004377839,0.0018770151,0.0005669283],"domain_scores_gemma":[0.9836838,0.012066533,0.002251793,0.0005524358,0.0010462928,0.00039912327],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0016348042,0.000536707,0.0024022853,0.00089544064,0.000020219612,0.000022384822,0.00093874,0.00077304925,0.0014839899],"category_scores_gemma":[0.07360683,0.0003640277,0.0005734174,0.00029178313,0.00076465384,0.000078979814,0.0001875456,0.00349477,0.000043570195],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005706116,0.0002326181,0.016232615,0.00023821306,0.00031752145,0.012315134,0.00022076571,0.0000062210747,0.0009109279,0.000008385995,0.964111,0.0048360084],"study_design_scores_gemma":[0.013015183,0.019020047,0.039309118,0.12650262,0.0018081341,0.0115954,0.000081387276,0.00035717877,0.0027492307,0.0011808154,0.78357494,0.0008059164],"about_ca_topic_score_codex":0.0011098956,"about_ca_topic_score_gemma":0.000030355312,"teacher_disagreement_score":0.19384316,"about_ca_system_score_codex":0.0002836782,"about_ca_system_score_gemma":0.00039030143,"threshold_uncertainty_score":0.99988115},"labels":[],"label_agreement":null},{"id":"W2891885683","doi":"10.1016/j.hpe.2018.09.001","title":"Guidelines for Creating Written Clinical Reasoning Exams: Insight from a Delphi Study","year":2018,"lang":"en","type":"article","venue":"Health Professions Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université de Sherbrooke","funders":"","keywords":"Likert scale; Delphi method; Guideline; Medical education; Scale (ratio); Delphi; Psychology; Snowball sampling; Relevance (law); Medicine; Computer science; Artificial intelligence; Pathology","score_opus":0.28227288140519496,"score_gpt":0.5888860901688526,"score_spread":0.30661320876365766,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2891885683","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9559922,0.00091331784,0.0022743302,0.03258677,0.0037400604,0.0031566322,0.000016574098,0.00015811139,0.0011620127],"genre_scores_gemma":[0.9050304,0.00017858649,0.06522082,0.018677428,0.0073097725,0.0007770522,0.0003294301,0.000049212984,0.0024272767],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9967709,0.0004547958,0.001531569,0.00057371164,0.0002815124,0.00038751398],"domain_scores_gemma":[0.9838374,0.012438157,0.0006273427,0.00073459366,0.0017026857,0.00065980724],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0027945812,0.00017890413,0.0005615578,0.00010832591,0.0006932094,0.000023209537,0.00011741804,0.00019225117,0.00014790498],"category_scores_gemma":[0.15636936,0.00013794744,0.00011655739,0.000281156,0.000092220886,0.00007094681,0.00006757242,0.0003688218,0.00008616518],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003289413,0.004307866,0.5070366,0.000059269445,0.00006461317,0.0000011354512,0.009992363,2.3807698e-7,0.000011347619,0.00024114338,0.17364639,0.30431005],"study_design_scores_gemma":[0.0043885866,0.004107812,0.901889,0.013531398,0.00025100054,0.00000819973,0.042845245,0.0010208419,0.000022438924,0.0008776877,0.030773705,0.00028404468],"about_ca_topic_score_codex":0.0017899842,"about_ca_topic_score_gemma":0.0002478868,"teacher_disagreement_score":0.3948524,"about_ca_system_score_codex":0.00013723379,"about_ca_system_score_gemma":0.00392441,"threshold_uncertainty_score":0.8507369},"labels":[],"label_agreement":null},{"id":"W2893892086","doi":"10.1186/s41235-018-0131-6","title":"The easy-to-hard training advantage with real-world medical images","year":2018,"lang":"en","type":"article","venue":"Cognitive Research Principles and Implications","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria","funders":"Natural Sciences and Engineering Research Council of Canada; National Science Foundation","keywords":"Categorization; Computer science; Schedule; Usability; Construct (python library); Perception; Artificial intelligence; Machine learning; Medical physics; Medicine; Psychology; Human–computer interaction","score_opus":0.2005530890453507,"score_gpt":0.5004577524330057,"score_spread":0.299904663387655,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2893892086","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7004485,0.0004588148,0.0030401708,0.06646241,0.000060524,0.0020139166,0.00013002787,0.00014247638,0.22724314],"genre_scores_gemma":[0.99025434,0.0015155211,0.0008076454,0.00079488993,0.00043432353,0.00037378308,0.000035167315,0.000025063175,0.005759279],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9979741,0.00016076324,0.00026042195,0.00041284045,0.0006413456,0.00055055827],"domain_scores_gemma":[0.97228867,0.024820345,0.000051427432,0.00041819506,0.0014073236,0.0010140552],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0018523008,0.00011623538,0.00020359716,0.00013781627,0.00092317245,0.000095886855,0.00017425514,0.000057765825,0.0001552846],"category_scores_gemma":[0.064036995,0.00006791364,0.00004131421,0.00061896164,0.0015308851,0.000045444292,0.00022253396,0.0004727922,0.00015082156],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009382676,0.00044444646,0.1883704,0.000037240443,0.00022847636,0.00006370613,0.0012269766,1.4972035e-7,0.00083706493,0.046466343,0.0063881557,0.75499874],"study_design_scores_gemma":[0.000988242,0.0009885023,0.94544417,0.0014660545,0.00004327777,0.00007256063,0.001598805,0.00004737639,0.00045580513,0.0007346956,0.048045788,0.0001147259],"about_ca_topic_score_codex":0.000085292326,"about_ca_topic_score_gemma":0.00047291326,"teacher_disagreement_score":0.75707376,"about_ca_system_score_codex":0.000040927913,"about_ca_system_score_gemma":0.00065706245,"threshold_uncertainty_score":0.943847},"labels":[],"label_agreement":null},{"id":"W2895214953","doi":"10.1503/cmaj.109-5663","title":"Do doctors experiencing burnout make more errors?","year":2018,"lang":"en","type":"article","venue":"Canadian Medical Association Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":31,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Toronto Public Health","funders":"","keywords":"Burnout; Medicine; Family medicine; Clinical psychology","score_opus":0.015390028798740194,"score_gpt":0.32401566758618355,"score_spread":0.30862563878744337,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2895214953","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94487524,0.0002714098,0.00007862265,0.038617186,0.0021604756,0.000103360246,0.000009839701,0.000038936167,0.01384492],"genre_scores_gemma":[0.96596014,0.00012218415,0.00027943234,0.02542563,0.0066156187,0.000004921369,0.000009570123,0.000021527592,0.0015609508],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.99693525,0.00009375978,0.0005575524,0.00021235489,0.0015771971,0.0006238759],"domain_scores_gemma":[0.9930987,0.002148929,0.00020908022,0.00016947385,0.00065258355,0.0037212619],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0018146662,0.00013828672,0.00031559457,0.00016929416,0.00032009042,0.00009420434,0.0001647115,0.0004649081,0.011319655],"category_scores_gemma":[0.13087365,0.000113946626,0.00016408408,0.00028289048,0.00013450776,0.000055426462,0.000018174049,0.0009641369,0.00047895187],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000044086584,0.00011655625,0.48745018,0.0000072943712,0.00021608692,0.0022717852,0.0049725845,0.0000010700011,0.000023300994,0.000324676,0.3055265,0.19904588],"study_design_scores_gemma":[0.003351501,0.00050205085,0.24503006,0.0018250007,0.00012877097,0.0022214924,0.004421873,0.0003313326,0.00008884753,0.0002767908,0.7414585,0.0003638092],"about_ca_topic_score_codex":0.0020806736,"about_ca_topic_score_gemma":0.0074001024,"teacher_disagreement_score":0.43593198,"about_ca_system_score_codex":0.0009938581,"about_ca_system_score_gemma":0.0034586028,"threshold_uncertainty_score":0.98958415},"labels":[],"label_agreement":null},{"id":"W2895951022","doi":"10.15694/mep.2018.0000236.1","title":"Learning-by-Concordance for Family Physicians: Revealing its Value for Continuing Professional Development in Dermatology","year":2018,"lang":"en","type":"article","venue":"MedEdPublish","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Concordance; Continuing professional development; Primary care; Continuing medical education; Medical education; Service (business); Professional development; Medicine; Task (project management); Family medicine; Value (mathematics); Psychology; Continuing education; Computer science; Engineering; Business","score_opus":0.02401553511128097,"score_gpt":0.34283092488728617,"score_spread":0.3188153897760052,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2895951022","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98944426,0.00057977525,0.003664025,0.002753986,0.0015340738,0.0012057427,0.000015346423,0.000094030765,0.0007087844],"genre_scores_gemma":[0.9805522,0.000022302162,0.009263767,0.0032090768,0.00060008845,0.00061387056,0.00023636765,0.000041804262,0.0054605417],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99816066,0.0000657709,0.0005802779,0.00043739486,0.00023087037,0.0005250241],"domain_scores_gemma":[0.9929127,0.0061095557,0.00020299549,0.00015193394,0.0005074408,0.000115349234],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0007967025,0.00017365787,0.0005152479,0.000095911935,0.00013200079,0.000043471286,0.0001448066,0.0001844992,0.000015890762],"category_scores_gemma":[0.05106018,0.00015771593,0.000096804186,0.00020930918,0.00009336943,0.00013437204,0.00005176268,0.00026579548,0.00001726419],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019647616,0.0016923018,0.08907103,0.00084587384,0.00020550116,0.00003878618,0.0073505747,0.000023405692,0.0138896145,0.0038212594,0.5108564,0.3702405],"study_design_scores_gemma":[0.021198478,0.0016943186,0.064968385,0.01943587,0.00017332028,0.00005525008,0.002113579,0.0067626718,0.070226826,0.0012585281,0.8110826,0.0010301488],"about_ca_topic_score_codex":0.000013857354,"about_ca_topic_score_gemma":0.000023124205,"teacher_disagreement_score":0.36921036,"about_ca_system_score_codex":0.00008353112,"about_ca_system_score_gemma":0.00027457494,"threshold_uncertainty_score":0.95693314},"labels":[],"label_agreement":null},{"id":"W2896824827","doi":"10.17483/2368-6669.1156","title":"Cognitive Companionship: A Potential Pedagogical Approach to Developing Clinical Reasoning in Nursing?","year":2018,"lang":"en","type":"article","venue":"Quality Advancement in Nursing Education - Avancées en formation infirmière","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"Social Sciences and Humanities Research Council of Canada; Fonds de Recherche du Québec-Société et Culture; Université de Montréal","keywords":"Humanities; Psychology; Philosophy","score_opus":0.14648829945396522,"score_gpt":0.528075832229984,"score_spread":0.38158753277601876,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2896824827","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8845539,0.00016971379,0.08924197,0.0059298505,0.001828869,0.0013737276,0.00001148454,0.00011325727,0.016777204],"genre_scores_gemma":[0.95468163,0.00007352538,0.041871127,0.0018408854,0.0006869993,0.00015813863,0.000444936,0.000025352962,0.00021742788],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99599606,0.0005884308,0.0017675414,0.0005819236,0.00055826973,0.0005077602],"domain_scores_gemma":[0.996171,0.0022926866,0.0003999312,0.00033357608,0.0004816842,0.00032115413],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.002176671,0.00027373398,0.0006043663,0.0004452417,0.00018678616,0.00006866859,0.0001515499,0.00023383343,0.000064580774],"category_scores_gemma":[0.01528274,0.00027981141,0.00011786579,0.0008026621,0.00029898394,0.0004333007,0.000033744534,0.0005493171,0.00012171374],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0025189256,0.015129756,0.13666227,0.00026016525,0.000042792297,0.00000436502,0.04071912,0.00043322775,0.00005790808,0.09304938,0.0033635832,0.7077585],"study_design_scores_gemma":[0.005799575,0.0009382277,0.89552313,0.038811497,0.00011300406,0.00010411186,0.040517323,0.0045499094,0.00026277773,0.0075595626,0.005008967,0.000811893],"about_ca_topic_score_codex":0.000045663455,"about_ca_topic_score_gemma":0.000020422283,"teacher_disagreement_score":0.7588609,"about_ca_system_score_codex":0.0009646954,"about_ca_system_score_gemma":0.00059205753,"threshold_uncertainty_score":0.9999654},"labels":[],"label_agreement":null},{"id":"W2897228760","doi":"10.1073/pnas.1806905115","title":"Deep neural network improves fracture detection by clinicians","year":2018,"lang":"en","type":"article","venue":"Proceedings of the National Academy of Sciences","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":617,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Radiography; Orthopedic surgery; Medicine; Emergency department; Diagnostic accuracy; Wrist; MEDLINE; Emergency medicine; Radiology; Surgery; Psychiatry","score_opus":0.0307677605064417,"score_gpt":0.3539758753027643,"score_spread":0.3232081147963226,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2897228760","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9855651,0.00023521853,0.000015483824,0.009484828,0.000089315625,0.00017689259,0.000003074871,0.000020621972,0.004409431],"genre_scores_gemma":[0.9950281,0.000022500259,0.00069750164,0.0034385826,0.0006841358,0.0000042020424,1.6441705e-7,0.000003957104,0.00012084868],"study_design_codex":"bench_or_experimental","study_design_gemma":"observational","domain_scores_codex":[0.9985377,0.0000048117568,0.00033400417,0.00020951913,0.0007534005,0.00016052053],"domain_scores_gemma":[0.9978225,0.001392803,0.00037519698,0.000008159074,0.00034152498,0.000059765094],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010618581,0.00007644066,0.00017127213,0.000045427314,0.00017947268,0.000013693351,0.00027631788,0.00012770567,0.000021444546],"category_scores_gemma":[0.016056068,0.000045838162,0.0000954932,0.0005229022,0.0010307932,0.00013328737,0.00005665917,0.00022397887,0.0000029759783],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007456601,0.00078936404,0.24993677,0.00016664193,0.00018668953,5.0751233e-8,0.000813443,0.00048166834,0.54630727,0.0068516824,0.07785095,0.11586978],"study_design_scores_gemma":[0.0010139818,0.0012979833,0.63132423,0.00072792347,0.00013260634,0.000035241083,0.00023387327,0.034680862,0.28221157,0.04550383,0.0026327178,0.00020518055],"about_ca_topic_score_codex":0.000010802374,"about_ca_topic_score_gemma":2.1330055e-7,"teacher_disagreement_score":0.38138744,"about_ca_system_score_codex":0.00001731298,"about_ca_system_score_gemma":0.000019363923,"threshold_uncertainty_score":0.9922321},"labels":[],"label_agreement":null},{"id":"W2897321108","doi":"10.1051/pmed/2018011","title":"Diagnostic et prise en charge des difficultés de raisonnement clinique. Guide AMEE n° 117 (version courte)","year":2017,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Université de Montréal","funders":"","keywords":"Humanities; Gynecology; Psychology; Philosophy; Medicine","score_opus":0.062261693888650604,"score_gpt":0.41226383842584274,"score_spread":0.35000214453719214,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2897321108","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8632434,0.03431289,0.003908708,0.06830408,0.006338505,0.0016283965,0.0003948689,0.00038243094,0.021486737],"genre_scores_gemma":[0.7146106,0.10643168,0.0076004327,0.007982577,0.0060890494,0.00025635015,0.00028964246,0.00022189527,0.15651779],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99438864,0.00065649033,0.001258272,0.0010328314,0.0009572679,0.0017064702],"domain_scores_gemma":[0.9482554,0.047017332,0.0006562057,0.0019149523,0.00032251817,0.0018335653],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0039335187,0.0006413408,0.0011236845,0.00010432458,0.00097156625,0.0003816794,0.000995037,0.0008275669,0.0051101656],"category_scores_gemma":[0.39416796,0.00058110536,0.0005884322,0.000114658105,0.0016861691,0.00026454165,0.0007486209,0.0012501254,0.003708512],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00028594842,0.004760867,0.37145188,0.00044348356,0.00032987449,0.002010167,0.0026032974,0.00003019074,0.00040422785,0.0011206272,0.39296404,0.2235954],"study_design_scores_gemma":[0.008242475,0.0014461196,0.45800182,0.019632328,0.0012772542,0.00010666204,0.00095090334,0.0015625465,0.0016093579,0.0011960993,0.505062,0.00091242266],"about_ca_topic_score_codex":0.001375948,"about_ca_topic_score_gemma":0.00010140627,"teacher_disagreement_score":0.39023444,"about_ca_system_score_codex":0.0006380995,"about_ca_system_score_gemma":0.0010862151,"threshold_uncertainty_score":0.999664},"labels":[],"label_agreement":null},{"id":"W2897681013","doi":"10.1051/pmed/2018012","title":"Diagnostic et prise en charge des difficultés de raisonnement clinique. Guide AMEE n <sup>o</sup>  117 (version courte)","year":2017,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Université de Montréal","funders":"","keywords":"Humanities; Gynecology; Psychology; Philosophy; Medicine","score_opus":0.05666630648463463,"score_gpt":0.39597806121473944,"score_spread":0.33931175473010483,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2897681013","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.89512336,0.026212515,0.003269424,0.055237774,0.0029522046,0.001660754,0.0004346782,0.00039157522,0.014717727],"genre_scores_gemma":[0.74975675,0.091771066,0.007755006,0.007960026,0.006477903,0.00030843457,0.0003827039,0.00026415967,0.13532394],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99339736,0.00078679586,0.0014790603,0.0012240543,0.0011264898,0.001986255],"domain_scores_gemma":[0.944588,0.050084677,0.00067842746,0.002175048,0.00036942298,0.002104411],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0043198843,0.0007735099,0.0013234816,0.00013172055,0.001108743,0.00043981418,0.0011658558,0.0009794823,0.0056874487],"category_scores_gemma":[0.38158402,0.0007037513,0.000695413,0.00014559853,0.0018648535,0.00031685416,0.00088170014,0.0014815578,0.0042029084],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00036601225,0.005566325,0.35866138,0.00054291665,0.00043709594,0.0022743002,0.005081945,0.00025523562,0.00017365003,0.0010169593,0.42072687,0.2048973],"study_design_scores_gemma":[0.01253207,0.0021245491,0.32800463,0.026416695,0.0019618275,0.00017192081,0.0023872827,0.010524408,0.0012363138,0.001479126,0.61175025,0.0014109376],"about_ca_topic_score_codex":0.0014007904,"about_ca_topic_score_gemma":0.000058298603,"teacher_disagreement_score":0.37726414,"about_ca_system_score_codex":0.00072580506,"about_ca_system_score_gemma":0.0012019732,"threshold_uncertainty_score":0.99954134},"labels":[],"label_agreement":null},{"id":"W2899003206","doi":"10.1016/j.aju.2018.10.052","title":"[5] Is there a correlation between cognitive and technical skills during urology objective structured clinical examinations (OSCEs)?","year":2018,"lang":"en","type":"article","venue":"Arab Journal of Urology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University Health Centre","funders":"","keywords":"Medicine; Cognition; Correlation; Medical physics; Urology; Medical education; Psychiatry","score_opus":0.022984440794139666,"score_gpt":0.36093306394147256,"score_spread":0.33794862314733287,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2899003206","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99337834,0.000505786,0.0023082998,0.0022978464,0.00039972283,0.00018924297,0.000015365573,0.000019890014,0.0008855305],"genre_scores_gemma":[0.9967469,0.00029033097,0.00037629082,0.0013881406,0.0011128251,0.0000027698443,0.0000071939107,0.000017831468,0.000057735942],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99802065,0.00035726657,0.0008721263,0.00028540616,0.00020617688,0.0002583801],"domain_scores_gemma":[0.9830816,0.015309759,0.000576455,0.00015207122,0.0006081031,0.00027200274],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00087887124,0.00015267708,0.00063313404,0.00018322215,0.00011377295,0.000010468241,0.00008097584,0.00045648435,0.00019686933],"category_scores_gemma":[0.03660683,0.00011956233,0.0001627787,0.00016204422,0.0008576607,0.00007169132,0.00006458713,0.00093409064,0.000029978155],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008954463,0.0002663948,0.9857267,0.000004917431,0.00031849265,0.00019396626,0.00043936583,0.0000016209665,0.0001582656,0.00010755124,0.00032033995,0.011566942],"study_design_scores_gemma":[0.004446933,0.005489411,0.9835418,0.00012452924,0.0008132403,0.0026296708,0.000082860766,0.000055886816,0.00014336129,0.002440809,0.0001339624,0.000097573866],"about_ca_topic_score_codex":0.0000062134473,"about_ca_topic_score_gemma":0.000011876526,"teacher_disagreement_score":0.03572796,"about_ca_system_score_codex":0.000027309623,"about_ca_system_score_gemma":0.00015823697,"threshold_uncertainty_score":0.97150826},"labels":[],"label_agreement":null},{"id":"W2899004899","doi":"10.1007/s10459-018-9859-5","title":"Comfort with uncertainty: reframing our conceptions of how clinicians navigate complex clinical situations","year":2018,"lang":"en","type":"review","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":144,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Cognitive reframing; Psychology; Engineering ethics; MEDLINE; Medical education; Medicine; Social psychology; Engineering; Political science","score_opus":0.17242005164032068,"score_gpt":0.5695529757285183,"score_spread":0.39713292408819756,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2899004899","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00036233003,0.990603,0.00039239024,0.004003984,0.0018370374,0.0017573476,0.00005596661,0.00004235406,0.0009455498],"genre_scores_gemma":[0.00274114,0.9795327,0.015245522,0.00086807815,0.0007274927,0.00017761244,0.0004245922,0.000023598923,0.0002592635],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99555653,0.00047251544,0.0021277391,0.0008086581,0.00055627647,0.0004782837],"domain_scores_gemma":[0.98725545,0.009089557,0.002301182,0.0005909952,0.0003786224,0.00038416887],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00359159,0.0003070752,0.0020219136,0.00030765435,0.00028098538,0.000034455206,0.00035441646,0.00023614307,0.000023099428],"category_scores_gemma":[0.014239879,0.00021911247,0.00027320578,0.0015435796,0.0012178132,0.0002226333,0.000046930778,0.00062910316,0.00001895687],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000014642937,0.0005624564,0.0064716404,0.002926852,0.000018220175,9.805805e-7,0.00016267296,0.00003401212,7.074877e-9,0.0008198816,0.0013617604,0.98762685],"study_design_scores_gemma":[0.00044811313,0.0019888768,0.0013746384,0.11584218,0.00021864341,0.000043960972,0.0021095546,0.00013501795,3.4263994e-8,0.00048178976,0.8771111,0.0002461296],"about_ca_topic_score_codex":0.00029891642,"about_ca_topic_score_gemma":0.0005916928,"teacher_disagreement_score":0.98738074,"about_ca_system_score_codex":0.00027130466,"about_ca_system_score_gemma":0.009876312,"threshold_uncertainty_score":0.9957368},"labels":[],"label_agreement":null},{"id":"W2899041345","doi":"10.4212/cjhp.v71i5.2844","title":"Is It Necessary for Pharmacists to Evaluate Other Health Professionals’ Satisfaction with Pharmacist Services?","year":2018,"lang":"en","type":"article","venue":"The Canadian Journal of Hospital Pharmacy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Pharmacist; Health professionals; Family medicine; Patient satisfaction; Nursing; Clinical pharmacy; Health services; Medicine; Pharmacy; Health care; Political science; Environmental health","score_opus":0.06785641093805622,"score_gpt":0.43871653529341076,"score_spread":0.37086012435535454,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2899041345","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.86182886,0.0005302425,0.00017273585,0.13404223,0.0017096319,0.0010634467,0.00012346375,0.00001237507,0.0005169979],"genre_scores_gemma":[0.91989785,0.00004478777,0.00060703827,0.07712492,0.0015865043,0.000024121699,0.000007675605,0.000042899264,0.00066420855],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99816215,0.000094873496,0.00055086333,0.00021525743,0.00047835338,0.0004984904],"domain_scores_gemma":[0.99614537,0.00071889756,0.00044469695,0.00021086524,0.00088389684,0.0015962686],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011781692,0.00020787815,0.00038366197,0.00015765948,0.00045289184,0.00007042662,0.00024728946,0.000034986904,0.0011629654],"category_scores_gemma":[0.0006584547,0.00012624684,0.00015251624,0.00027032467,0.00015560136,0.00016237711,0.000015907603,0.0003698639,0.000141694],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0031534089,0.00032367778,0.18215331,0.00016305945,0.00074567046,0.00017027797,0.010170574,0.000032989912,0.0001859808,0.00003731993,0.74721724,0.055646487],"study_design_scores_gemma":[0.06953163,0.026171168,0.17519034,0.016117347,0.0029981094,0.0033495463,0.003392421,0.0050087613,0.016311316,0.001696492,0.6782157,0.0020171686],"about_ca_topic_score_codex":0.006731726,"about_ca_topic_score_gemma":0.006957968,"teacher_disagreement_score":0.069001555,"about_ca_system_score_codex":0.00027340063,"about_ca_system_score_gemma":0.0028960994,"threshold_uncertainty_score":0.9998825},"labels":[],"label_agreement":null},{"id":"W2899080483","doi":"10.1097/acm.0000000000002512","title":"A Philosophical Approach to Addressing Uncertainty in Medical Education","year":2018,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":69,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Public Health Ontario","funders":"","keywords":"MEDLINE; Medical education; Psychology; Epistemology; Engineering ethics; Medicine; Philosophy; Political science; Law; Engineering","score_opus":0.09629249498382905,"score_gpt":0.4426868394559438,"score_spread":0.3463943444721147,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2899080483","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.72849625,0.0010137757,0.0009936875,0.17171063,0.0010129737,0.0007180399,0.0000014773211,0.00012789955,0.09592527],"genre_scores_gemma":[0.94422567,0.00017327513,0.0004713015,0.04860414,0.005757269,0.000062302504,0.00005176804,0.000022883554,0.0006313933],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99731874,0.000098693556,0.0007128145,0.00051120605,0.0009498645,0.000408699],"domain_scores_gemma":[0.99578226,0.002487568,0.00008783634,0.00035489257,0.00016960256,0.0011178314],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0016631412,0.0001880104,0.00055274524,0.00026968226,0.000051321276,0.0000034551624,0.00023746412,0.0005224001,0.0007482265],"category_scores_gemma":[0.14647235,0.00013393503,0.00004322338,0.00071961415,0.00053367513,0.000036463658,0.00008456138,0.0012048682,0.00017000356],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0023482456,0.0017170778,0.13356173,0.0003259525,0.0001313631,0.0001926686,0.016432555,0.000021940568,0.0011938901,0.067002766,0.3954797,0.3815921],"study_design_scores_gemma":[0.045870986,0.010166455,0.48429617,0.16225109,0.0011852154,0.004635525,0.009303743,0.036656294,0.00056760584,0.114008695,0.12831303,0.002745179],"about_ca_topic_score_codex":0.00022375313,"about_ca_topic_score_gemma":0.0000072945913,"teacher_disagreement_score":0.3788469,"about_ca_system_score_codex":0.00012740561,"about_ca_system_score_gemma":0.0009334402,"threshold_uncertainty_score":0.8607173},"labels":[],"label_agreement":null},{"id":"W2900010494","doi":"10.14196/mjiri.32.75","title":"Development and psychometrics of script concordance test (SCT) in midwifery","year":2018,"lang":"en","type":"article","venue":"Medical Journal of the Islamic Republic of Iran","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"Iran University of Medical Sciences","keywords":"Concordance; Test (biology); Psychometrics; Psychology; Medicine; Obstetrics; Clinical psychology; Internal medicine; Biology","score_opus":0.0331837527264752,"score_gpt":0.3296400457796719,"score_spread":0.29645629305319665,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2900010494","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9908664,0.0008161471,0.00017349438,0.0055649634,0.0006935508,0.000100192905,0.0000015268018,0.000003800349,0.0017799305],"genre_scores_gemma":[0.995843,0.0003550746,0.002463563,0.00088959234,0.00024795328,7.972746e-7,5.438312e-7,0.000009704442,0.00018980331],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9971104,0.00007951994,0.0013505545,0.00014128821,0.0011206061,0.00019759382],"domain_scores_gemma":[0.9922995,0.0058382647,0.0008423934,0.00033421628,0.00045345168,0.00023215654],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0028274136,0.00011155519,0.0006164814,0.00030388002,0.000026801446,0.000008017466,0.00034435006,0.00014270195,0.00010996251],"category_scores_gemma":[0.15183568,0.00006489066,0.00012637774,0.00080311106,0.0007006335,0.000052401912,0.00009158938,0.00047111514,0.0000015195772],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00045483056,0.00182257,0.6547129,0.00017626592,0.00016645576,0.00012930219,0.0015810647,6.6037677e-7,0.0015264252,0.00046787475,0.009662547,0.3292991],"study_design_scores_gemma":[0.0050142305,0.0011288832,0.97446436,0.007961775,0.00007396301,0.0007248919,0.0002869324,0.00011382379,0.004604245,0.00083297613,0.0046864115,0.000107494496],"about_ca_topic_score_codex":0.000031929285,"about_ca_topic_score_gemma":0.000041062594,"teacher_disagreement_score":0.3291916,"about_ca_system_score_codex":0.000052769265,"about_ca_system_score_gemma":0.00072992826,"threshold_uncertainty_score":0.8553088},"labels":[],"label_agreement":null},{"id":"W2900691345","doi":"10.1007/s10459-018-9856-8","title":"Clinical practice, deliberate practice, and “big data”","year":2018,"lang":"en","type":"editorial","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Clinical Practice; Medical education; MEDLINE; Big data; Medicine; Data science; Psychology; Computer science; Family medicine; Data mining; Chemistry","score_opus":0.06980051183933747,"score_gpt":0.545383094992556,"score_spread":0.47558258315321855,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2900691345","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0002338153,0.058592677,0.00011071542,0.015100871,0.9149922,0.00075386843,0.000064117885,0.000042041895,0.010109655],"genre_scores_gemma":[0.00015966584,0.25437564,0.058937743,0.007257464,0.67767173,0.000053216438,0.0007931612,0.000024668487,0.00072671054],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99464005,0.00059584016,0.0016287143,0.0014307329,0.0011937354,0.00051091186],"domain_scores_gemma":[0.70169204,0.28624347,0.0047742883,0.0028894627,0.002831732,0.0015690111],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.014026465,0.00025728953,0.00072393625,0.0002208107,0.00026498796,0.00012221209,0.0004839266,0.00047404002,0.000021709631],"category_scores_gemma":[0.6004569,0.00021538821,0.00004143319,0.0007173055,0.0010319654,0.0013550412,0.00022103197,0.0012001157,0.00004653063],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011102246,0.0005051156,0.0015787655,0.00014961226,0.000006118662,0.0000033559131,0.00007546062,0.0000011453642,8.807244e-9,0.000060786864,0.59341794,0.40409064],"study_design_scores_gemma":[0.000594414,0.0014553776,0.0009962976,0.004890144,0.00007831682,0.000035976478,0.0005490712,0.00009720243,1.0860146e-7,0.0005682032,0.9905428,0.00019206558],"about_ca_topic_score_codex":0.0009966259,"about_ca_topic_score_gemma":0.00030045572,"teacher_disagreement_score":0.58643043,"about_ca_system_score_codex":0.00016917863,"about_ca_system_score_gemma":0.015734516,"threshold_uncertainty_score":0.98984534},"labels":[],"label_agreement":null},{"id":"W2901128264","doi":"10.1080/0142159x.2018.1512746","title":"Key features to assess clinical decisions","year":2018,"lang":"en","type":"letter","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Key (lock); Computer science; Feature (linguistics); Clinical decision making; Medical education; Data science; Psychology; Medical physics; Medicine; Family medicine; Computer security","score_opus":0.15053325892628397,"score_gpt":0.4672472517837163,"score_spread":0.3167139928574323,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2901128264","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.010042802,0.000347197,0.0005674753,0.9438784,0.0040104627,0.00057766936,0.000019908668,0.0002719712,0.040284123],"genre_scores_gemma":[0.0008675137,0.00021539947,0.0020218012,0.8455573,0.05041252,0.00006759432,0.00045099558,0.00013188936,0.100274995],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9930724,0.00056013605,0.0013521044,0.0012123315,0.0029243666,0.0008786578],"domain_scores_gemma":[0.96439064,0.030635737,0.00024805297,0.0018306725,0.0003854026,0.0025095246],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity","insufficient_payload"],"category_scores_codex":[0.0033040936,0.0005277668,0.0016475986,0.00018780815,0.000096452066,0.00006442786,0.0007447849,0.004703431,0.022002308],"category_scores_gemma":[0.5358192,0.00036422798,0.00081929157,0.0003203734,0.00060402666,0.000022997627,0.00040718802,0.008516607,0.0061512454],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007985493,0.00041346866,0.011774742,0.000016915534,0.00019193206,0.0033017902,0.000046303,2.0216218e-8,1.3248363e-7,0.000009244755,0.94878453,0.03538104],"study_design_scores_gemma":[0.0014723057,0.00076992164,0.01605048,0.00557664,0.00041593218,0.00025646112,0.000011250614,0.000026027017,0.0000013118347,0.00021422228,0.974807,0.0003984029],"about_ca_topic_score_codex":0.00006925984,"about_ca_topic_score_gemma":0.000017098178,"teacher_disagreement_score":0.5325151,"about_ca_system_score_codex":0.000105533814,"about_ca_system_score_gemma":0.0010632537,"threshold_uncertainty_score":0.99988097},"labels":[],"label_agreement":null},{"id":"W2902592658","doi":"10.5847/wjem.j.1920-8642.2019.01.002","title":"The use of outcome feedback by emergency medicine physicians: Results of a physician survey","year":2018,"lang":"en","type":"article","venue":"World Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University; University of Toronto; McMaster University","funders":"Dalhousie University","keywords":"Medicine; Observational study; Emergency department; Family medicine; Outcome (game theory); MEDLINE; Nursing; Internal medicine","score_opus":0.15011200159547777,"score_gpt":0.4130960909491343,"score_spread":0.2629840893536565,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2902592658","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.87451375,0.027538111,0.00080254884,0.04843112,0.030444916,0.0011445878,0.000247303,0.000055470173,0.016822176],"genre_scores_gemma":[0.96332264,0.0115928305,0.00018450357,0.0005413207,0.004624175,0.0000037599073,0.00005692614,0.0000628495,0.019610997],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99211943,0.00040188746,0.0051332396,0.000332252,0.00151319,0.00050001295],"domain_scores_gemma":[0.98382586,0.007887358,0.0038537295,0.0008712532,0.0031587898,0.00040300356],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0041179974,0.0003809346,0.0017422207,0.00032536683,0.00013874569,0.0000020275659,0.000507013,0.00010469349,0.0013600951],"category_scores_gemma":[0.09558323,0.00021291073,0.00043527296,0.0018658527,0.0010953306,0.00013874171,0.000070263,0.0005991056,0.00001567471],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0025862148,0.00064037764,0.120154075,0.00008864092,0.00050689426,0.000009039462,0.0007296363,0.0000049671858,0.004560453,0.00010476374,0.8604158,0.010199119],"study_design_scores_gemma":[0.012772943,0.011363208,0.7024284,0.013108071,0.0018292123,0.000021525977,0.00076249393,0.00014749165,0.0017651699,0.00075577525,0.2545812,0.00046448284],"about_ca_topic_score_codex":0.0015394967,"about_ca_topic_score_gemma":0.0010048249,"teacher_disagreement_score":0.6058346,"about_ca_system_score_codex":0.000048757225,"about_ca_system_score_gemma":0.00011287803,"threshold_uncertainty_score":0.9995528},"labels":[],"label_agreement":null},{"id":"W2902630798","doi":"10.1111/medu.13637","title":"The perils of teaching medical triads","year":2018,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute for Clinical Evaluative Sciences; Sunnybrook Health Science Centre; University of Toronto","funders":"","keywords":"Medical education; MEDLINE; Medicine; Psychology; Family medicine; Political science","score_opus":0.015892351065476443,"score_gpt":0.41818406873466535,"score_spread":0.4022917176691889,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2902630798","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8477156,0.00070889207,0.0006137488,0.1095697,0.0030270647,0.00023368577,7.798571e-7,0.000053766624,0.03807673],"genre_scores_gemma":[0.98803115,0.00017104953,0.0003530477,0.007409991,0.0023656304,0.00001988948,0.000014478429,0.000009669962,0.0016250924],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9978443,0.00012445771,0.0004235016,0.00015620662,0.001272815,0.00017874876],"domain_scores_gemma":[0.9882923,0.010239302,0.00013528658,0.00043405755,0.00029318113,0.0006058716],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0028764177,0.00007210902,0.00019957579,0.00003365569,0.00013607666,0.000010108595,0.00017260404,0.00020464172,0.0022405784],"category_scores_gemma":[0.46727636,0.000041084215,0.000080063466,0.00011721131,0.00056076644,0.000018500812,0.00003860786,0.000436114,0.00009621428],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011371008,0.00078745204,0.011322528,0.000014556926,0.000026599222,0.0000032693554,0.00045953688,9.571816e-9,0.00001731615,0.0056297756,0.15300919,0.8286161],"study_design_scores_gemma":[0.0046729594,0.0019943125,0.1692251,0.0075386846,0.00025450933,0.00035003704,0.0016041367,0.0024591815,0.0007886493,0.0034401428,0.80738664,0.00028564048],"about_ca_topic_score_codex":0.00016028504,"about_ca_topic_score_gemma":0.000023472461,"teacher_disagreement_score":0.8283304,"about_ca_system_score_codex":0.000031541706,"about_ca_system_score_gemma":0.0032385634,"threshold_uncertainty_score":0.99867153},"labels":[],"label_agreement":null},{"id":"W2909110100","doi":"10.1177/0008417418820367","title":"Book Review: Clinical reasoning in occupational therapy","year":2018,"lang":"en","type":"article","venue":"Canadian Journal of Occupational Therapy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Université de Sherbrooke","funders":"","keywords":"Occupational therapy; Psychology; Psychotherapist; Psychiatry","score_opus":0.17799766363676084,"score_gpt":0.4774489135856892,"score_spread":0.29945124994892836,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2909110100","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.60037935,0.3838635,0.00038328927,0.010839241,0.0018774397,0.0006959298,0.000053723146,0.000018798717,0.0018887012],"genre_scores_gemma":[0.53935367,0.1608433,0.0048631406,0.2863846,0.0065238154,0.000025143092,0.00014891091,0.00008392967,0.0017734594],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9975714,0.00019600475,0.0012200208,0.00022416204,0.0004631441,0.00032526822],"domain_scores_gemma":[0.994728,0.0025989385,0.00047714933,0.00025911478,0.0010258273,0.0009109541],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0018531603,0.00017623029,0.0005725848,0.0003126392,0.00010145506,0.000027677272,0.00020223002,0.00015731543,0.0051508164],"category_scores_gemma":[0.014374281,0.0001399635,0.00031512784,0.0003656149,0.0003020624,0.00014784251,0.000005552341,0.0004623839,0.00008406059],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011961301,0.00013632051,0.58020097,0.000006913254,0.000083523344,0.00013263912,0.00006584567,0.0000025555623,0.0000067022984,0.00009882541,0.38475543,0.033314176],"study_design_scores_gemma":[0.0028053117,0.0009373565,0.65583557,0.0028721776,0.000015036904,0.0001835584,0.000006754032,0.000021181977,0.000032188764,0.00007364162,0.3371088,0.000108444816],"about_ca_topic_score_codex":0.00060033577,"about_ca_topic_score_gemma":0.0006570002,"teacher_disagreement_score":0.27554536,"about_ca_system_score_codex":0.00016501242,"about_ca_system_score_gemma":0.005442129,"threshold_uncertainty_score":0.9957586},"labels":[],"label_agreement":null},{"id":"W2909429215","doi":"10.1093/neuros/nyy457","title":"Reasons to Believe: Biostatistics and Methodology for the Neurosurgeon","year":2018,"lang":"en","type":"article","venue":"Neurosurgery","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Biostatistics; Neurosurgery; MEDLINE; Medical physics; Surgery; Pathology; Epidemiology; Law","score_opus":0.13496586522917106,"score_gpt":0.4090202353865766,"score_spread":0.2740543701574055,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2909429215","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8976721,0.0004185843,0.0402951,0.05343602,0.005466732,0.0013402859,0.00018965566,0.00020138481,0.0009801367],"genre_scores_gemma":[0.9311959,0.00037593595,0.020344427,0.044227924,0.0014080638,0.00011491934,0.000020987092,0.000069707065,0.0022421384],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9987702,0.00011969462,0.00028974906,0.00035446178,0.00014655208,0.0003193486],"domain_scores_gemma":[0.9190185,0.08010436,0.00006806455,0.00037599585,0.00016026349,0.00027281107],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0009100281,0.00013379386,0.00032517707,0.000067565124,0.00017133294,0.000028286036,0.00006760631,0.00008090311,0.000051347248],"category_scores_gemma":[0.1823501,0.0000903021,0.00009218658,0.00020926421,0.00025638615,0.000015140892,0.00006685516,0.00014262315,0.00003971479],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016391608,0.0002791858,0.14696972,0.00005705355,0.000080680424,0.00013954884,0.00025444987,0.000006017692,0.0065227044,0.004206184,0.6295922,0.21025309],"study_design_scores_gemma":[0.0010112977,0.0016956412,0.6914798,0.00027839871,0.0003067929,0.00035761698,0.00003454514,0.000602604,0.0023378427,0.0008803682,0.30077696,0.00023814544],"about_ca_topic_score_codex":0.000035399116,"about_ca_topic_score_gemma":0.000009173172,"teacher_disagreement_score":0.54451007,"about_ca_system_score_codex":0.0000071079908,"about_ca_system_score_gemma":0.00007121943,"threshold_uncertainty_score":0.82453734},"labels":[],"label_agreement":null},{"id":"W2911290735","doi":"10.1097/acm.0000000000002618","title":"Clinical Reasoning Assessment Methods: A Scoping Review and Practical Guidance","year":2019,"lang":"en","type":"review","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":338,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Concordance; Computer science; Competence (human resources); Construct (python library); Construct validity; Medical physics; Management science; Medical education; Data science; Psychology; Medicine; Psychometrics; Clinical psychology; Engineering","score_opus":0.35190693664255196,"score_gpt":0.666586725573005,"score_spread":0.31467978893045306,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2911290735","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[6.489845e-7,0.9819067,0.0025848725,0.0062882374,0.0010768783,0.004268547,0.0000064882643,0.000121296274,0.0037463407],"genre_scores_gemma":[2.544088e-7,0.95338196,0.03071996,0.011493901,0.0021197624,0.00025542665,0.00014527491,0.00011517124,0.0017683029],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9903767,0.0021860865,0.004424804,0.0015196309,0.0008056851,0.0006870939],"domain_scores_gemma":[0.9073637,0.088313565,0.0019181739,0.0011846508,0.00021032795,0.0010095949],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.017107204,0.0008032337,0.009829086,0.0001882502,0.00007740402,0.000011573187,0.00026295372,0.0016020457,0.0005720332],"category_scores_gemma":[0.3706289,0.0005150483,0.00059403497,0.0005742666,0.0006124651,0.0000881968,0.0003691345,0.0060642073,0.00017735126],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000009485734,0.000027075443,0.00023624305,0.16814281,0.00031208043,0.00012148404,0.000017930166,5.925831e-9,6.4520314e-8,0.0004350301,0.03187638,0.7988214],"study_design_scores_gemma":[0.0007213966,0.00034582082,0.00003729386,0.5584401,0.013080412,0.0012820318,0.000011839628,0.000008301459,2.101414e-8,0.00003662615,0.42578325,0.00025294043],"about_ca_topic_score_codex":0.000013721397,"about_ca_topic_score_gemma":3.177665e-7,"teacher_disagreement_score":0.7985685,"about_ca_system_score_codex":0.00016488416,"about_ca_system_score_gemma":0.0027339319,"threshold_uncertainty_score":0.9997301},"labels":[],"label_agreement":null},{"id":"W2912003064","doi":"10.1177/1062860618820687","title":"An IDEA: Safety Training to Improve Critical Thinking by Individuals and Teams","year":2019,"lang":"en","type":"article","venue":"American Journal of Medical Quality","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hospital for Sick Children; University of Toronto","funders":"","keywords":"Cognition; Harm; Context (archaeology); Cognitive bias; Correctness; Medicine; Applied psychology; Scale (ratio); Patient safety; Cognitive psychology; Psychology; Health care; Social psychology; Computer science; Psychiatry","score_opus":0.02315600201409811,"score_gpt":0.4198769921141,"score_spread":0.3967209901000019,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2912003064","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9560201,0.00019807446,0.004262447,0.03856774,0.00036748784,0.00013490199,0.000019198844,0.000023602137,0.00040642553],"genre_scores_gemma":[0.97885644,0.00010833335,0.0050740116,0.015447375,0.00045614204,0.0000017847266,0.0000063978478,0.000019764388,0.00002975797],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9955186,0.0005056197,0.0011961816,0.00032322036,0.0020512484,0.00040512072],"domain_scores_gemma":[0.9766024,0.02002238,0.00042379028,0.00032650476,0.00022502293,0.0023999233],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.008953873,0.00017553994,0.0010698473,0.0000783209,0.00007176169,0.00004664241,0.00030102124,0.00014939718,0.00039743926],"category_scores_gemma":[0.11391017,0.00012756386,0.00015312698,0.00021480951,0.0005647011,0.00010955004,0.0000828182,0.0009830227,0.0000175398],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00071656494,0.00078696833,0.17432253,0.000046930047,0.00019196126,0.00013601551,0.0057192952,0.0000016036604,0.00091114914,0.005071109,0.0021230592,0.8099728],"study_design_scores_gemma":[0.016611712,0.03722335,0.83924526,0.01155329,0.00084848073,0.00246369,0.060008895,0.0003668279,0.0006324733,0.009471694,0.020146525,0.0014277888],"about_ca_topic_score_codex":0.00019121116,"about_ca_topic_score_gemma":0.0000031051716,"teacher_disagreement_score":0.80854505,"about_ca_system_score_codex":0.0000614696,"about_ca_system_score_gemma":0.0005658617,"threshold_uncertainty_score":0.89355373},"labels":[],"label_agreement":null},{"id":"W2912476504","doi":"10.1111/medu.13801","title":"Accuracy of self‐monitoring: does experience, ability or case difficulty matter?","year":2019,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":30,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Seniority; Aptitude; Context (archaeology); Self-monitoring; Psychology; Psychological intervention; Confidence interval; Social psychology; Applied psychology; Medicine; Developmental psychology; Psychiatry","score_opus":0.01824775107992284,"score_gpt":0.3958603817714509,"score_spread":0.37761263069152806,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2912476504","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99324375,0.000073743235,0.000011703557,0.0034585572,0.0019234559,0.00030596842,0.0000022773143,0.000045634817,0.0009348813],"genre_scores_gemma":[0.9969398,0.000058998347,0.00062050694,0.000873744,0.00043293185,0.00005009327,0.000016195414,0.000009431602,0.0009983034],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9986254,0.00005099823,0.00041920797,0.00026563965,0.00047269688,0.00016605157],"domain_scores_gemma":[0.99393874,0.0050199153,0.00012957233,0.0004274375,0.00016461426,0.00031973145],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00031313545,0.000101643804,0.00027615923,0.000039199735,0.000029508512,0.000009437282,0.000092529335,0.00014472193,0.005653044],"category_scores_gemma":[0.06638946,0.00005675446,0.00007380129,0.00016785898,0.0000991546,0.000054631717,0.000035899895,0.00019611618,0.0001958397],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001043477,0.0023455622,0.94783103,0.00015559023,0.000017522165,0.000057738132,0.0022151335,1.5379803e-7,0.00017670126,0.00002528428,0.004676376,0.042394575],"study_design_scores_gemma":[0.001427362,0.00029145,0.98225206,0.0020669003,0.00008186716,0.00080417894,0.0036813754,0.00009261143,0.0019419028,0.000053758107,0.0071506975,0.00015581213],"about_ca_topic_score_codex":0.00031842824,"about_ca_topic_score_gemma":0.0000053602644,"teacher_disagreement_score":0.06607632,"about_ca_system_score_codex":0.00006476766,"about_ca_system_score_gemma":0.0011225098,"threshold_uncertainty_score":0.99525595},"labels":[],"label_agreement":null},{"id":"W2912548934","doi":"10.1016/j.bjpt.2019.01.010","title":"Higher order thinking about differential diagnosis","year":2019,"lang":"en","type":"review","venue":"Brazilian Journal of Physical Therapy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":57,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval","funders":"","keywords":"CLARITY; Medical diagnosis; Differential diagnosis; Set (abstract data type); Health care; Process (computing); Medicine; Computer science; Intensive care medicine; Psychology; Pathology","score_opus":0.06574197211650293,"score_gpt":0.40220126432081277,"score_spread":0.3364592922043098,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2912548934","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0012842222,0.9952191,0.000033447624,0.0003392973,0.0017078905,0.0005786676,0.000021116733,0.00003093118,0.0007852939],"genre_scores_gemma":[0.0006795012,0.9914645,0.00018864311,0.000462185,0.0059983656,0.000024112187,0.000032226006,0.00013091622,0.00101956],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99668705,0.0002751718,0.00126132,0.0004012089,0.0009412616,0.0004339852],"domain_scores_gemma":[0.9875088,0.009869554,0.0012553218,0.00055600953,0.00034894227,0.00046141105],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00033078564,0.0006194985,0.004295724,0.00020670453,0.00006808935,0.00009400983,0.0005143511,0.0003999468,0.0011680905],"category_scores_gemma":[0.0029458266,0.0003629528,0.0029587732,0.00042587056,0.00015453126,0.00009204378,0.00006276256,0.0015747988,0.00031133476],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012605826,0.0012632157,0.000105653504,0.0011827208,0.001122726,0.00013805187,0.00007901274,7.468138e-7,0.0000021160665,0.00033416046,0.0042306553,0.9914149],"study_design_scores_gemma":[0.0024639948,0.0015863664,0.00063017994,0.07743863,0.0020555484,0.000118518634,0.000003934152,0.0000027825326,0.000007976977,0.0006197598,0.9147049,0.00036736397],"about_ca_topic_score_codex":0.0000034344193,"about_ca_topic_score_gemma":1.2567384e-7,"teacher_disagreement_score":0.9910475,"about_ca_system_score_codex":0.00010587832,"about_ca_system_score_gemma":0.0006037193,"threshold_uncertainty_score":0.9998822},"labels":[],"label_agreement":null},{"id":"W2913333113","doi":"10.1007/s11606-019-04839-5","title":"Practicing Medicine with Colleagues: Pitfalls from Social Psychology Science","year":2019,"lang":"en","type":"article","venue":"Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"Canadian Institutes of Health Research; BrightFocus Foundation","keywords":"Attribution; Viewpoints; Interpersonal communication; Medicine; Perspective (graphical); Situational ethics; Social psychology; Perception; Psychology","score_opus":0.040509104935208366,"score_gpt":0.4249271351773229,"score_spread":0.3844180302421145,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2913333113","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9331495,0.0008485161,0.0008808489,0.03833091,0.00273709,0.00019195935,0.000003074362,0.000017783477,0.023840303],"genre_scores_gemma":[0.9761541,0.00032361367,0.0025444005,0.009959396,0.0071108392,0.0000016275004,0.000006145047,0.0000313668,0.003868515],"study_design_codex":"bench_or_experimental","study_design_gemma":"observational","domain_scores_codex":[0.9964463,0.00009486311,0.0011980633,0.00037928726,0.0014765027,0.0004050014],"domain_scores_gemma":[0.9936991,0.0032166496,0.0011399843,0.00030824513,0.0010920819,0.000543959],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0018396521,0.0002560886,0.0011881005,0.00045522887,0.00008862371,0.000017583112,0.0004132591,0.00012810159,0.0018045362],"category_scores_gemma":[0.016203908,0.00013998392,0.0001250155,0.0005472328,0.0012007493,0.00019829498,0.000059885388,0.0009064344,0.000064903375],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.016024228,0.0011485172,0.304803,0.00007268883,0.0013494178,0.003982745,0.0049650874,0.00004110844,0.36066577,0.0012218944,0.27363902,0.032086514],"study_design_scores_gemma":[0.050127998,0.044386275,0.82082903,0.027394952,0.0020384884,0.010448944,0.002539675,0.00064928917,0.005736826,0.0018217397,0.033361282,0.000665511],"about_ca_topic_score_codex":0.00081089116,"about_ca_topic_score_gemma":0.000016031969,"teacher_disagreement_score":0.516026,"about_ca_system_score_codex":0.00017048784,"about_ca_system_score_gemma":0.00044181026,"threshold_uncertainty_score":0.99910796},"labels":[],"label_agreement":null},{"id":"W2914287857","doi":"10.1007/s10459-019-09875-8","title":"Debiasing versus knowledge retrieval checklists to reduce diagnostic error in ECG interpretation","year":2019,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":44,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Impact; McMaster University","funders":"Royal College of Physicians and Surgeons of Canada","keywords":"Debiasing; Checklist; Cognitive bias; Cognition; Heuristics; Interpretation (philosophy); MEDLINE; Psychology; Patient safety; Overconfidence effect; Computer science; Medicine; Cognitive psychology; Social psychology; Psychiatry; Health care","score_opus":0.036371792652310324,"score_gpt":0.4666605353569062,"score_spread":0.4302887427045959,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2914287857","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9780711,0.0065760575,0.00010026827,0.0044498746,0.0046875016,0.0008755837,0.0000010732333,0.000025976877,0.005212596],"genre_scores_gemma":[0.9932533,0.00097215164,0.00432919,0.0010541788,0.00014201125,0.000042855896,0.000013974679,0.000008566266,0.00018379741],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9981682,0.00011562229,0.00057714706,0.0004892935,0.00027353858,0.0003762277],"domain_scores_gemma":[0.9876705,0.011533151,0.00018525847,0.0002362736,0.00010787294,0.000266954],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001488841,0.000115811985,0.0002765027,0.0003460019,0.00006603335,0.000025795207,0.0001492933,0.000058192534,0.00005149294],"category_scores_gemma":[0.06295212,0.00010802263,0.00003169614,0.0014564552,0.000103697304,0.0003349121,0.00003118929,0.00020220394,0.00013549857],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00087135646,0.0008365952,0.45354685,0.00032106735,0.0000021955746,0.0000019450142,0.005303283,0.002461568,0.0000684511,0.0007098766,0.00030842065,0.5355684],"study_design_scores_gemma":[0.0025525978,0.0026401866,0.9583621,0.016582314,0.00001117977,0.000014574066,0.004240953,0.009599394,0.0003120118,0.0010347894,0.0043237945,0.0003260753],"about_ca_topic_score_codex":0.00035574226,"about_ca_topic_score_gemma":0.00053630216,"teacher_disagreement_score":0.5352423,"about_ca_system_score_codex":0.0005530731,"about_ca_system_score_gemma":0.0018878891,"threshold_uncertainty_score":0.94494104},"labels":[],"label_agreement":null},{"id":"W2916375756","doi":"10.1515/dx-2014-0009","title":"Diagnostic conversations: Clinical Decision Making in surgery – Part 2","year":2014,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Royal Australasian College of Physicians; Royal College of Physicians and Surgeons of Canada; Royal Australasian College of Surgeons; U.S. Consumer Product Safety Commission","keywords":"Clinical decision making; Psychology; Computer science; Medicine; Intensive care medicine","score_opus":0.056840258419547374,"score_gpt":0.38960377286245956,"score_spread":0.3327635144429122,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2916375756","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.992599,0.0005969059,0.00075949664,0.00281947,0.0014503481,0.0003558717,0.00000830102,0.00011141937,0.0012991655],"genre_scores_gemma":[0.98930514,0.0027149345,0.0012019326,0.0058002183,0.0006574116,0.0002179455,0.00004165503,0.000029794306,0.000030950858],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9974715,0.00023865054,0.0010600898,0.0004928127,0.00036169274,0.00037522276],"domain_scores_gemma":[0.5532988,0.44539067,0.00021141181,0.0006260212,0.00013093045,0.00034217886],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002337087,0.00018188162,0.00070650707,0.00020504993,0.000056020643,0.000033374145,0.00010137854,0.00023286454,0.0006727107],"category_scores_gemma":[0.5454742,0.00016250205,0.0003267566,0.0003989666,0.0001310037,0.000070563765,0.0000629784,0.0003364926,0.00061723276],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000053015272,0.00052130275,0.7504821,0.000011909318,0.00001623372,0.00007080912,0.000025299933,0.000012783075,9.6738454e-8,0.00016891319,0.05046238,0.19817515],"study_design_scores_gemma":[0.0013042777,0.00016502081,0.9639368,0.0061667357,0.000101148034,0.000012687456,0.000027266362,0.00044289022,0.000017363538,0.001418062,0.026221516,0.00018620925],"about_ca_topic_score_codex":0.000059249727,"about_ca_topic_score_gemma":0.00006261233,"teacher_disagreement_score":0.5431371,"about_ca_system_score_codex":0.00006079682,"about_ca_system_score_gemma":0.000119034295,"threshold_uncertainty_score":0.79334843},"labels":[],"label_agreement":null},{"id":"W2916706452","doi":"","title":"1st year Psychology Student Calibration on Exams: Insights into confidence calibration, question complexity effects, and (in)consistencies with the Dunning-Kruger effect","year":2018,"lang":"en","type":"article","venue":"2018 Conference of the Canadian Society for the Study of Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University; Mount Saint Vincent University","funders":"","keywords":"Calibration; Psychology; Social psychology; Computer science; Mathematics; Statistics","score_opus":0.040971988485144197,"score_gpt":0.35536562213811007,"score_spread":0.31439363365296585,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2916706452","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9897639,0.00016858042,0.0002005195,0.007273636,0.00037167317,0.001984675,0.0000035730106,0.0000063417833,0.00022708344],"genre_scores_gemma":[0.9983482,0.000027604801,0.0001315793,0.0010994723,0.000118572716,0.0001462731,0.000009379891,0.000008927604,0.00010999972],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99898237,0.00018836686,0.0002482935,0.00021972535,0.00022514185,0.00013608021],"domain_scores_gemma":[0.99622965,0.0027661102,0.00021190809,0.00042140795,0.000283287,0.000087660985],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0006413955,0.00012214103,0.00024075329,0.000037136226,0.0003526748,0.000033494664,0.00019249956,0.00007863058,0.000011461779],"category_scores_gemma":[0.0021314884,0.00006138556,0.00008540316,0.00018276162,0.0009130364,0.00005266206,0.000022591412,0.00014787128,8.7578417e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014159435,0.0027658343,0.7468537,0.00071506796,0.0010713224,0.0000010267868,0.15493311,0.00007262666,0.0014710131,0.03955861,0.035901833,0.015239912],"study_design_scores_gemma":[0.0015088778,0.003156088,0.9797472,0.0012805591,0.00033267986,0.0000026787054,0.011216258,0.00063595915,0.0005795493,0.0012345772,0.00019934397,0.000106253276],"about_ca_topic_score_codex":0.031859823,"about_ca_topic_score_gemma":0.18768778,"teacher_disagreement_score":0.23289348,"about_ca_system_score_codex":0.00010135532,"about_ca_system_score_gemma":0.00075448886,"threshold_uncertainty_score":0.9745871},"labels":[],"label_agreement":null},{"id":"W2917561174","doi":"10.1111/medu.13814","title":"Experts’ responses in script concordance tests: a response process validity investigation","year":2019,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":26,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"College of Family Physicians of Canada","funders":"","keywords":"Concordance; Context (archaeology); Consistency (knowledge bases); Test (biology); Psychology; Process (computing); Action (physics); Qualitative property; Applied psychology; Medicine; Social psychology; Computer science; Artificial intelligence; Machine learning","score_opus":0.041013671760452015,"score_gpt":0.402577520766441,"score_spread":0.361563849005989,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2917561174","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97181255,0.00025157465,0.000018042703,0.025803143,0.00087139994,0.00045013972,0.0000015250104,0.000064124775,0.0007275089],"genre_scores_gemma":[0.9830665,0.00004384769,0.0002974561,0.013814319,0.00022512741,0.00013041675,0.00004375787,0.000015778864,0.0023627721],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99799865,0.00034538392,0.00043523172,0.0003493803,0.0006534702,0.00021786224],"domain_scores_gemma":[0.9896086,0.009254192,0.00012278561,0.00039301018,0.0002180543,0.0004033731],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0019052263,0.00012181718,0.00027532823,0.00014668432,0.00002895422,0.000017061595,0.000111476445,0.00019248735,0.0007605194],"category_scores_gemma":[0.33630812,0.00010334247,0.000046422978,0.00041996175,0.00015080976,0.00009455466,0.00002021583,0.0002917759,0.0001775846],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0025259766,0.0009839227,0.9593017,0.00008079535,0.000006111387,0.000022553264,0.0019309906,0.0000013340948,0.0010952278,0.000113391354,0.02143259,0.012505371],"study_design_scores_gemma":[0.0021902898,0.0005951999,0.980946,0.005360806,0.000020859383,0.00007512316,0.0012670661,0.00050687976,0.0013284091,0.00098994,0.0065427506,0.00017669002],"about_ca_topic_score_codex":0.00016234221,"about_ca_topic_score_gemma":0.000024527282,"teacher_disagreement_score":0.3344029,"about_ca_system_score_codex":0.000171969,"about_ca_system_score_gemma":0.0050143604,"threshold_uncertainty_score":0.88952607},"labels":[],"label_agreement":null},{"id":"W2920942474","doi":"10.1093/ptj/pzz044","title":"Predictors of Clinical Reasoning Using the Reasoning 4 Change Instrument With Physical Therapist Students","year":2019,"lang":"en","type":"article","venue":"Physical Therapy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":20,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Psychology; Metacognition; Curriculum; Biopsychosocial model; Competence (human resources); Cognition; Multilevel model; Clinical psychology; Developmental psychology; Social psychology; Psychotherapist; Pedagogy","score_opus":0.08881890418088285,"score_gpt":0.4224207290277642,"score_spread":0.33360182484688133,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2920942474","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9973799,0.0001665853,0.000040652936,0.00040085532,0.00033226493,0.001048957,0.000008377141,0.00007042974,0.00055199664],"genre_scores_gemma":[0.996595,0.00023977611,0.0003620496,0.0007144708,0.0018786999,0.00004920622,0.000013079901,0.000060877872,0.000086799664],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.997243,0.00024374617,0.00051232084,0.00052882615,0.0010599382,0.00041219278],"domain_scores_gemma":[0.99459666,0.0038797033,0.00036461573,0.00076005794,0.00015105646,0.00024791263],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0006953838,0.0003133683,0.0010297284,0.000039842504,0.00009333733,0.0000377665,0.00033539764,0.00009891783,0.000038935894],"category_scores_gemma":[0.0011105394,0.00016582911,0.0004365887,0.0003428801,0.0003500899,0.00010480059,0.000095784075,0.0005876781,0.000029812047],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017642125,0.003076682,0.9637215,0.000020588584,0.00068092067,0.000010970166,0.0019401351,0.000028768194,0.00103039,0.000848315,0.000036625315,0.026840897],"study_design_scores_gemma":[0.009147048,0.00502449,0.9717012,0.0028490257,0.00027976392,0.000017351627,0.0004254226,0.0070049376,0.002269265,0.00024369283,0.00067969965,0.00035808762],"about_ca_topic_score_codex":0.0001610112,"about_ca_topic_score_gemma":0.0000011336649,"teacher_disagreement_score":0.02648281,"about_ca_system_score_codex":0.000044532557,"about_ca_system_score_gemma":0.00012817829,"threshold_uncertainty_score":0.6762315},"labels":[],"label_agreement":null},{"id":"W2921092794","doi":"10.1515/dx-2018-0069","title":"Experienced physician descriptions of intuition in clinical reasoning: a typology","year":2019,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":45,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Impact; McMaster University","funders":"","keywords":"Intuition; Typology; Medical diagnosis; Psychology; Phenomenon; Medicine; Social psychology; Epistemology; Cognitive psychology; Pathology; Cognitive science; Sociology","score_opus":0.04170661413683197,"score_gpt":0.3857682627652473,"score_spread":0.3440616486284153,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2921092794","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9916834,0.00014055712,0.000035270124,0.00070819556,0.0003949114,0.00030908664,0.000006218464,0.000030322259,0.0066920184],"genre_scores_gemma":[0.99742496,0.0003458775,0.0006529723,0.0012242588,0.00010348409,0.000120602774,0.00002754295,0.000012407639,0.00008787237],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99873227,0.00009138978,0.0005477712,0.00027892273,0.00014834965,0.00020130123],"domain_scores_gemma":[0.9943518,0.0050056614,0.00012901107,0.00031494224,0.00008060502,0.000117954034],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00035844388,0.00009548574,0.00045720045,0.000101703925,0.000014601674,0.0000048425904,0.000075481665,0.000153718,0.00039303338],"category_scores_gemma":[0.018253893,0.00008627019,0.00016968182,0.00029052034,0.00014229004,0.000047373534,0.00003619353,0.00021753667,0.00018625834],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009577496,0.0010307893,0.9766839,0.000012767966,0.000019244126,0.000012705375,0.00031936253,0.000004684887,0.00015400171,0.0035555935,0.0014488634,0.016662313],"study_design_scores_gemma":[0.0020060407,0.0008783241,0.9919188,0.0014802534,0.00004855048,0.0000047881117,0.0007061884,0.0001569655,0.0012496875,0.00060878333,0.0008355897,0.000106039224],"about_ca_topic_score_codex":0.00015969893,"about_ca_topic_score_gemma":0.000024791536,"teacher_disagreement_score":0.017895449,"about_ca_system_score_codex":0.000036289755,"about_ca_system_score_gemma":0.00008323344,"threshold_uncertainty_score":0.99001575},"labels":[],"label_agreement":null},{"id":"W2921154254","doi":"10.1097/acm.0000000000002046","title":"“Rising to the Level of Your Incompetence”: What Physicians’ Self-Assessment of Their Performance Reveals About the Imposter Syndrome in Medicine","year":2017,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":307,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; The Wilson Centre; University of Ottawa; Medical Council of Canada","funders":"","keywords":"Affect (linguistics); Credibility; Psychology; Feeling; Anxiety; Self-assessment; Clinical psychology; Medical education; Social psychology; Medicine; Psychiatry","score_opus":0.11681573995274157,"score_gpt":0.4186364759335918,"score_spread":0.30182073598085024,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2921154254","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9136818,0.0023582864,0.00010690291,0.08106868,0.0007135295,0.0007941235,0.000007767672,0.000018523668,0.001250402],"genre_scores_gemma":[0.98545736,0.007379832,0.00016453948,0.005873918,0.00070139405,0.000041741307,0.000009513888,0.000026846992,0.00034483097],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9968148,0.00015519024,0.0013287507,0.00038767533,0.0008776391,0.0004359621],"domain_scores_gemma":[0.9910614,0.0061968896,0.0009287565,0.0013988621,0.00022960843,0.00018450042],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004184252,0.0002870019,0.0011994772,0.00015064963,0.00020116715,0.000010932589,0.0010257208,0.00020815678,0.000072190516],"category_scores_gemma":[0.0101934755,0.00012957468,0.000093059636,0.00032408515,0.0008123545,0.00020663234,0.0002607188,0.0011411299,0.0000079963675],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00031030798,0.00019214103,0.84399974,0.00054907607,0.00034068283,0.00006592112,0.03649272,0.000052196825,0.010576188,0.0011826152,0.008229868,0.09800851],"study_design_scores_gemma":[0.0026099884,0.0011063014,0.95239246,0.035806376,0.00021774172,0.00012167173,0.005481333,0.000550243,0.0005048732,0.00037777086,0.0006984993,0.00013272287],"about_ca_topic_score_codex":0.00034972926,"about_ca_topic_score_gemma":0.000016504158,"teacher_disagreement_score":0.10839271,"about_ca_system_score_codex":0.00008717866,"about_ca_system_score_gemma":0.00017505343,"threshold_uncertainty_score":0.9981441},"labels":[],"label_agreement":null},{"id":"W2924990228","doi":"10.21815/jde.019.075","title":"The Culture of Certainty in Dentistry and Its Impact on Dental Education and Practice","year":2019,"lang":"en","type":"article","venue":"Journal of Dental Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Certainty; Dental care; Dental education; Dental practice; Dentistry; Medicine; Psychology; Medical education; Philosophy; Epistemology","score_opus":0.012157776221609127,"score_gpt":0.41553745643016415,"score_spread":0.403379680208555,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2924990228","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99418604,0.0028494685,4.5550152e-7,0.0006315554,0.0010437038,0.00015011584,0.0000015343517,0.0000011201583,0.0011360022],"genre_scores_gemma":[0.9979273,0.0007507664,0.000103594655,0.00023419423,0.00017735659,0.0000016720579,0.000005451312,0.000005019522,0.00079465786],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99929905,0.000056553472,0.0002923709,0.00007568975,0.00020579013,0.00007056629],"domain_scores_gemma":[0.9973073,0.001987121,0.0003490027,0.00007252834,0.00017367608,0.00011037825],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00042200487,0.00006495046,0.00014751742,0.00006403086,0.000024768358,0.00002774906,0.00004002482,0.000057027937,0.000027450253],"category_scores_gemma":[0.010774581,0.000039016795,0.000050675335,0.000077982964,0.000025604166,0.0001421011,0.000011938488,0.00021015531,0.000009656127],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0020039137,0.003773101,0.92361075,0.000090879745,0.00011192782,0.000014748706,0.0027310369,0.0000099763765,0.004760849,0.00053146627,0.0064064777,0.055954862],"study_design_scores_gemma":[0.0010604945,0.00076520175,0.97431123,0.0015805474,0.00010514447,0.0038220864,0.01711277,0.0000234955,0.0007045919,0.00014862906,0.000314503,0.00005132958],"about_ca_topic_score_codex":0.00007918757,"about_ca_topic_score_gemma":0.0000066109405,"teacher_disagreement_score":0.055903535,"about_ca_system_score_codex":0.00009463955,"about_ca_system_score_gemma":0.00048523786,"threshold_uncertainty_score":0.99755806},"labels":[],"label_agreement":null},{"id":"W2926946235","doi":"","title":"Reassurance is not for the faint of heart","year":2019,"lang":"en","type":"article","venue":"Canadian Family Physician","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Computer science; World Wide Web; Data science; Medicine","score_opus":0.0273401175826326,"score_gpt":0.29551974187209035,"score_spread":0.26817962428945774,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2926946235","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.90593576,0.0008658477,0.000043708842,0.027971251,0.0008077762,0.0014307597,0.00038741837,0.000029583158,0.06252791],"genre_scores_gemma":[0.83002067,0.0000142418985,0.00013142181,0.1686708,0.00017847797,0.00002005347,0.000013690415,0.000016241569,0.00093440065],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9992014,0.000011393318,0.00017902184,0.000189869,0.00014114834,0.00027712443],"domain_scores_gemma":[0.9964503,0.002666843,0.000050807,0.00044409634,0.0001230729,0.000264905],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00015467034,0.00009446777,0.00027727752,0.0000392324,0.000048647013,0.000008896884,0.0001068131,0.00005919038,0.000013048252],"category_scores_gemma":[0.0009207823,0.00006852585,0.00017476443,0.00013211371,0.000073240626,0.00001959935,0.0000075667904,0.00011013044,0.00016522381],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003541411,0.00017024636,0.018570615,0.00017080667,0.00034612903,0.000017271868,0.00043922613,0.00006901297,0.015641086,0.0075793327,0.798365,0.15827715],"study_design_scores_gemma":[0.0018654119,0.0004347899,0.4854704,0.0010843656,0.00015450701,0.0000017116798,0.000518229,0.0009255709,0.0032433658,0.0004878848,0.50559294,0.00022080398],"about_ca_topic_score_codex":0.012160571,"about_ca_topic_score_gemma":0.0010309931,"teacher_disagreement_score":0.46689978,"about_ca_system_score_codex":0.000062491235,"about_ca_system_score_gemma":0.00044976853,"threshold_uncertainty_score":0.99441755},"labels":[],"label_agreement":null},{"id":"W2937181177","doi":"10.1093/ajcp/142.suppl1.073","title":"Resident Case Load Assessment between Academic General and Subspecialty Sign-out Sites at the University of Toronto Anatomical Pathology Program","year":2014,"lang":"en","type":"article","venue":"American Journal of Clinical Pathology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Toronto","funders":"","keywords":"Subspecialty; Sign (mathematics); Medicine; Anatomical pathology; Pathology; Medical education","score_opus":0.04460860764136058,"score_gpt":0.4289113755324037,"score_spread":0.3843027678910431,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2937181177","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9944061,0.0005433851,0.00065036217,0.0034299225,0.0002581329,0.00020840867,0.0000132711775,0.000015522748,0.00047487568],"genre_scores_gemma":[0.991621,0.0016470844,0.0048710597,0.00074602966,0.0009584051,9.4796974e-7,0.0000061803794,0.000015286816,0.00013398122],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99580926,0.0016997282,0.0014498796,0.0003550678,0.00035896339,0.00032710467],"domain_scores_gemma":[0.97718835,0.02010561,0.0014589569,0.0003689409,0.00034883825,0.0005293176],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0048338226,0.00017799063,0.0014810154,0.000033423083,0.00010314413,0.000005667625,0.00022086843,0.00028637407,0.00018470209],"category_scores_gemma":[0.021124188,0.00012157801,0.00038004885,0.00007889767,0.0026919825,0.00004827293,0.00021692981,0.0010765619,0.000008463546],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006692796,0.00021932865,0.76069444,0.0000064404635,0.00008601124,0.003411482,0.00031458633,0.0000023209268,0.0001688741,0.00009802199,0.0013950573,0.23293413],"study_design_scores_gemma":[0.0028777244,0.008557957,0.9775585,0.00012078196,0.00057097187,0.005571659,0.0005758672,0.00007809479,0.000027395874,0.00025887298,0.003675041,0.00012715264],"about_ca_topic_score_codex":0.00024969658,"about_ca_topic_score_gemma":0.0006488145,"teacher_disagreement_score":0.23280697,"about_ca_system_score_codex":0.00014898428,"about_ca_system_score_gemma":0.00038901804,"threshold_uncertainty_score":0.9918722},"labels":[],"label_agreement":null},{"id":"W2937315522","doi":"10.1515/dx-2018-0073","title":"Understanding diagnosis through ACTion: evaluation of a point-of-care checklist for junior emergency medical residents","year":2019,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Impact; St. Michael's Hospital; University of Toronto; McMaster University; North York General Hospital","funders":"","keywords":"Medical diagnosis; Checklist; Logistic regression; Psychological intervention; Rating scale; Medicine; Psychology; Scale (ratio); Clinical psychology; Psychiatry; Developmental psychology; Pathology","score_opus":0.21533579399065822,"score_gpt":0.44777129675403665,"score_spread":0.23243550276337843,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2937315522","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9851967,0.002316013,0.00084917265,0.0037273162,0.0015902224,0.0018124839,0.000083058956,0.00004685499,0.0043781726],"genre_scores_gemma":[0.9940681,0.0039088973,0.0006279699,0.00021148636,0.00023155044,0.0006881449,0.00012582297,0.000035143592,0.000102891965],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99692076,0.00012318083,0.00080707384,0.00039983095,0.0014770793,0.00027209322],"domain_scores_gemma":[0.9899057,0.008344398,0.00036241888,0.00046227346,0.0007318136,0.00019337497],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012272233,0.00018519629,0.00060445996,0.00011461469,0.000060807262,0.000007768995,0.00015963975,0.00027061097,0.0049503306],"category_scores_gemma":[0.061105214,0.00016660285,0.00038655422,0.00035169156,0.00009382957,0.0001263109,0.000058999238,0.00017036739,0.000037137866],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002634595,0.0007284498,0.96116894,0.0006273815,0.00021134401,0.000003493768,0.0009766949,0.000020723552,0.00003161539,0.001518033,0.023045573,0.011404313],"study_design_scores_gemma":[0.020759802,0.0040129586,0.8958582,0.020586433,0.004351696,0.000018101624,0.012903065,0.0014238319,0.024974205,0.01106203,0.0032572958,0.0007923779],"about_ca_topic_score_codex":0.00025241906,"about_ca_topic_score_gemma":0.000116653464,"teacher_disagreement_score":0.06531072,"about_ca_system_score_codex":0.00026623366,"about_ca_system_score_gemma":0.0003429587,"threshold_uncertainty_score":0.9959593},"labels":[],"label_agreement":null},{"id":"W2939069044","doi":"10.1007/s10459-019-09886-5","title":"The disconnect between knowing and doing in health professions education and practice","year":2019,"lang":"en","type":"review","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":48,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; University of British Columbia; Royal College of Physicians and Surgeons of Canada; Royal Roads University; Royal Society of Canada","funders":"","keywords":"Action (physics); Normative; Engineering ethics; Process (computing); Psychology; Health care; Knowledge management; Cognition; Computer science; Epistemology; Political science","score_opus":0.09732570189779916,"score_gpt":0.5663693783170898,"score_spread":0.4690436764192906,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2939069044","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00020259402,0.97843206,0.000020738435,0.016769016,0.0015642579,0.0019528725,0.0000025677446,0.000010906815,0.0010450133],"genre_scores_gemma":[0.000941449,0.9937814,0.003522464,0.00095840043,0.00029737805,0.00023223655,0.00003718887,0.000012408343,0.00021707518],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99681073,0.0008379481,0.0011080954,0.00056901743,0.0002687448,0.00040543888],"domain_scores_gemma":[0.95078415,0.047519546,0.0010760301,0.0002871609,0.00006421138,0.0002688969],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0063097966,0.00019013742,0.00086784904,0.00027160687,0.0004459426,0.000069512316,0.00014114616,0.00009835079,0.0000017255398],"category_scores_gemma":[0.036534052,0.00011949555,0.000041058003,0.0008889162,0.00026389625,0.00044702485,0.000059823575,0.0005397698,0.0000042486563],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000032363141,0.000116851654,0.0062015997,0.006505659,0.000002101323,9.233808e-8,0.00057379395,0.0000011963691,2.0594808e-9,0.0011985536,0.00007611528,0.9853208],"study_design_scores_gemma":[0.000119662676,0.00023618185,0.0035674942,0.1494768,0.000033581855,0.000037498266,0.0040286663,0.000008911942,9.753695e-9,0.0005753276,0.8418046,0.00011124829],"about_ca_topic_score_codex":0.001049577,"about_ca_topic_score_gemma":0.00029466508,"teacher_disagreement_score":0.9852095,"about_ca_system_score_codex":0.00036992558,"about_ca_system_score_gemma":0.013286757,"threshold_uncertainty_score":0.992307},"labels":[],"label_agreement":null},{"id":"W2940064032","doi":"10.1515/dx-2018-0084","title":"A simulation-based approach to training in heuristic clinical decision-making","year":2019,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":20,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Alberta Health Services; Rockyview General Hospital; University of Calgary","funders":"","keywords":"Debriefing; Cognitive bias; Odds; Session (web analytics); Heuristic; Psychology; Experiential learning; Computer science; Cognition; Applied psychology; Cognitive psychology; Social psychology; Artificial intelligence; Machine learning; Mathematics education","score_opus":0.06825140882401998,"score_gpt":0.4176728520204016,"score_spread":0.34942144319638163,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2940064032","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98405504,0.000081751736,0.009478464,0.00044769046,0.00042330715,0.0007607731,0.000008277497,0.00007929249,0.004665394],"genre_scores_gemma":[0.9800247,0.000009952895,0.0149705475,0.004583647,0.00019329348,0.00013458285,0.000018789424,0.000031897955,0.000032605658],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9979101,0.00008922706,0.0007463486,0.00054781575,0.00036755853,0.00033896163],"domain_scores_gemma":[0.86037093,0.13874654,0.000087356995,0.00042575208,0.000086863096,0.00028256487],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0009995056,0.00016528156,0.00060117146,0.00020663819,0.000030445926,0.000028730788,0.00012743764,0.00017792011,0.00024137498],"category_scores_gemma":[0.2225575,0.00014670138,0.00021469514,0.0005108023,0.000036323265,0.000032056163,0.000042258664,0.00031888395,0.00049726194],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018566784,0.000660482,0.7594005,0.000016369975,0.000013978888,0.000028931157,0.0002299075,0.13440694,3.2907946e-7,0.00008014727,0.00046349782,0.104513235],"study_design_scores_gemma":[0.0032205298,0.00053093786,0.7490185,0.0052159154,0.000076319724,0.0000019598456,0.00020480194,0.23955628,0.0000039983256,0.00037309766,0.0015282677,0.00026943145],"about_ca_topic_score_codex":0.000024866706,"about_ca_topic_score_gemma":0.000008429858,"teacher_disagreement_score":0.22155799,"about_ca_system_score_codex":0.000067294095,"about_ca_system_score_gemma":0.0002176386,"threshold_uncertainty_score":0.7839912},"labels":[],"label_agreement":null},{"id":"W2940064549","doi":"10.17483/2368-6669.1152","title":"Clinical Reasoning on an Assignment: Baccalaureate Nursing Students' Perceptions","year":2019,"lang":"en","type":"article","venue":"Quality Advancement in Nursing Education - Avancées en formation infirmière","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Perception; Psychology; Theme (computing); Patient care; Medical education; Nursing; Medicine; Computer science","score_opus":0.04319354663981405,"score_gpt":0.48959830792127673,"score_spread":0.44640476128146267,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2940064549","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9754116,0.000103910905,0.003183221,0.0032205626,0.0024046616,0.0012036428,0.000012216823,0.00014685822,0.014313355],"genre_scores_gemma":[0.98999256,0.00025066806,0.005607787,0.001441446,0.00041416156,0.000103965176,0.00069985614,0.00003696463,0.0014525692],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9960671,0.0005690157,0.0015167589,0.0005764451,0.00084470696,0.00042595714],"domain_scores_gemma":[0.996219,0.002018507,0.00047696472,0.0007389915,0.00020740215,0.00033915145],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0019165957,0.0002882882,0.00055017974,0.00030407062,0.00016886694,0.00010498764,0.00020752351,0.00022521873,0.00052489917],"category_scores_gemma":[0.0034209907,0.00028719145,0.00016410102,0.00035562486,0.0001279382,0.00081299903,0.000019961133,0.0005744608,0.00048427578],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006919683,0.016567979,0.48534173,0.0001216432,0.000036557627,0.0000014643464,0.00765063,0.0015340637,0.00015194615,0.02070246,0.0016536209,0.46554592],"study_design_scores_gemma":[0.004110524,0.0014692117,0.94892377,0.021368798,0.00011164513,0.000023970335,0.012941352,0.0016098088,0.000117212985,0.0025907806,0.006215723,0.0005172207],"about_ca_topic_score_codex":0.000024267809,"about_ca_topic_score_gemma":0.0000062573463,"teacher_disagreement_score":0.4650287,"about_ca_system_score_codex":0.0011749691,"about_ca_system_score_gemma":0.00026855766,"threshold_uncertainty_score":0.99995804},"labels":[],"label_agreement":null},{"id":"W2942985582","doi":"10.1017/cem.2019.99","title":"LO56: Measuring cognitive load on shift: Application of cognitive load theory during clinical work in the emergency department","year":2019,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Medicine; Cognition; Cognitive load; Distraction; Emergency department; Affect (linguistics); Effects of sleep deprivation on cognitive performance; Observational study; Emergency medicine; Physical therapy; Family medicine; Psychiatry; Psychology; Internal medicine","score_opus":0.08289948953400209,"score_gpt":0.3867230320339284,"score_spread":0.30382354249992627,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2942985582","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9838176,0.002918425,0.000101838086,0.0021383306,0.0029964135,0.00076676026,0.000016705255,0.0000057355114,0.0072382055],"genre_scores_gemma":[0.99701816,0.0011655042,0.000009630654,0.0003769523,0.0010906197,0.00001925031,0.000016793216,0.000026992655,0.00027612635],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9958795,0.00039614577,0.0020323545,0.0003243399,0.00092610694,0.00044161003],"domain_scores_gemma":[0.99137455,0.005574133,0.0009147301,0.0003533923,0.00096470735,0.00081850763],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004651507,0.00024934852,0.0008211325,0.00025907476,0.000074720294,0.0000027412652,0.0002671484,0.00016727466,0.0024849959],"category_scores_gemma":[0.050147906,0.00016552719,0.0002926683,0.0006891711,0.00019939378,0.000059855156,0.000015006231,0.0008489438,0.00008899588],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0025286057,0.00046576298,0.9624963,0.00008715199,0.00034513403,0.0002475912,0.004345841,0.000008932621,0.00002316439,0.00037814432,0.0024675739,0.026605846],"study_design_scores_gemma":[0.0050699366,0.0019609896,0.979327,0.008970817,0.00054349075,0.000028160772,0.002516663,0.000009607603,0.00006123335,0.0008301987,0.00049994106,0.00018192618],"about_ca_topic_score_codex":0.0012458957,"about_ca_topic_score_gemma":0.003585956,"teacher_disagreement_score":0.0454964,"about_ca_system_score_codex":0.00024328614,"about_ca_system_score_gemma":0.001342465,"threshold_uncertainty_score":0.99842685},"labels":[],"label_agreement":null},{"id":"W2943533685","doi":"10.1097/acm.0000000000002768","title":"Management Reasoning: Implications for Health Professions Educators and a Research Agenda","year":2019,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":90,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Hamilton Health Sciences","funders":"","keywords":"Negotiation; Medical diagnosis; Plan (archaeology); Knowledge management; Cognition; Management science; Psychology; Computer science; Medicine; Pathology","score_opus":0.1429927648167414,"score_gpt":0.5329605326184251,"score_spread":0.38996776780168374,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2943533685","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.54153883,0.0051255464,0.0005592042,0.43660793,0.0006354761,0.0039972872,0.000015616253,0.00011261759,0.011407466],"genre_scores_gemma":[0.9628694,0.005060106,0.00130224,0.009052987,0.0006108749,0.0004192412,0.00010662929,0.000035028435,0.020543467],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99806255,0.000100096084,0.0005017697,0.0004743553,0.0003422524,0.00051896105],"domain_scores_gemma":[0.9917084,0.0070090955,0.000119656856,0.00046358327,0.00015171885,0.0005475689],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0029010426,0.00013213008,0.0004261526,0.00023740981,0.00019371971,0.0000040875184,0.00014052913,0.00016727117,0.00013000544],"category_scores_gemma":[0.017301507,0.00009423242,0.000041003983,0.00043301767,0.00022365588,0.00003014433,0.00010414118,0.00079595705,0.000057135316],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001996146,0.00009929254,0.16868085,0.0005532804,0.000113067435,0.0000044750864,0.004703391,3.695766e-7,0.00026240115,0.08279656,0.68313783,0.059448853],"study_design_scores_gemma":[0.00817013,0.0024532264,0.62925804,0.020173104,0.00023343088,0.00013909198,0.012870982,0.000110916306,0.000051212777,0.018094117,0.3081532,0.00029252504],"about_ca_topic_score_codex":0.000059739235,"about_ca_topic_score_gemma":0.0000012036734,"teacher_disagreement_score":0.4605772,"about_ca_system_score_codex":0.00010627508,"about_ca_system_score_gemma":0.00029224582,"threshold_uncertainty_score":0.99097615},"labels":[],"label_agreement":null},{"id":"W2944436640","doi":"10.2196/13151","title":"Evaluation of a Health Information Technology–Enabled Collective Intelligence Platform to Improve Diagnosis in Primary Care and Urgent Care Settings: Protocol for a Pragmatic Randomized Controlled Trial","year":2019,"lang":"en","type":"article","venue":"JMIR Research Protocols","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Agency for Healthcare Research and Quality; National Cancer Institute; Gordon and Betty Moore Foundation","keywords":"Randomized controlled trial; Medicine; Intervention (counseling); Collective intelligence; Health care; Usability; Nursing; Family medicine; Artificial intelligence; Computer science; Pathology","score_opus":0.09874572585468087,"score_gpt":0.525575500077347,"score_spread":0.4268297742226661,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2944436640","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.016146256,0.000019418714,0.00012228286,0.00076722464,0.000013617469,0.9819038,0.000024823388,0.000038142844,0.0009644669],"genre_scores_gemma":[0.014580708,0.0000015472966,0.0015725457,0.00012390655,0.000032626394,0.9835813,0.000050499882,0.000018611594,0.000038234783],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9944289,0.0010925647,0.0018065444,0.00041507167,0.0017008429,0.0005561137],"domain_scores_gemma":[0.980846,0.0140098585,0.0006472998,0.00041736537,0.0038424379,0.00023707066],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01918241,0.00021241608,0.0018320001,0.00095188903,0.00011484324,0.000064845764,0.00016708445,0.0002508617,0.000054209246],"category_scores_gemma":[0.1646894,0.00015228757,0.00020471991,0.00096086395,0.00015375673,0.00019081803,0.00014972298,0.00046804445,0.000009411126],"study_design_candidate":"randomized_trial","study_design_consensus":"randomized_trial","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.83793783,0.00071655907,0.0008306409,0.006295945,0.00008568814,5.9144594e-7,0.0066350014,0.000034636774,0.000029804352,0.00011409732,0.0003748307,0.1469444],"study_design_scores_gemma":[0.94584644,0.015213497,0.00020837615,0.028944556,0.000037997568,0.000001344317,0.004654396,0.0018555789,0.00087007333,0.0014412734,0.0007809108,0.00014553803],"about_ca_topic_score_codex":0.00004995969,"about_ca_topic_score_gemma":0.000016163358,"teacher_disagreement_score":0.14679886,"about_ca_system_score_codex":0.001554733,"about_ca_system_score_gemma":0.006366383,"threshold_uncertainty_score":0.9992666},"labels":[],"label_agreement":null},{"id":"W2945422642","doi":"10.1007/s10459-019-09895-4","title":"Productive failure as an instructional approach to promote future learning","year":2019,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":67,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University Health Network; The Wilson Centre; University of Toronto","funders":"","keywords":"Pharmacy; Instructional design; Intervention (counseling); Knowledge acquisition; Mathematics education; Computer science; Psychology; Knowledge management; Medical education; Medicine; Nursing","score_opus":0.014666452347729335,"score_gpt":0.3966998131881986,"score_spread":0.38203336084046924,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2945422642","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9789914,0.0010768883,0.00007061849,0.010760292,0.0016673594,0.00096595846,7.174608e-7,0.00003881045,0.0064279945],"genre_scores_gemma":[0.92740184,0.00032643627,0.06869432,0.0018036218,0.0009704191,0.00009727109,0.000039315215,0.000008405441,0.0006583642],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99843895,0.000077228164,0.00029055562,0.0005561169,0.00036529024,0.00027185187],"domain_scores_gemma":[0.99904436,0.00027441865,0.00013721944,0.00017258308,0.0001194026,0.00025201877],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00092378183,0.000095439864,0.00019300239,0.00016885811,0.00014758781,0.000024261602,0.00011558603,0.000049295228,0.000042627315],"category_scores_gemma":[0.004196772,0.000078006706,0.000021476138,0.00086905627,0.00009618234,0.00046391296,0.000017917419,0.00027253848,0.000059072197],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000096658434,0.0010020337,0.27986377,0.00022302741,0.0000029183318,2.841775e-7,0.0040370966,0.006301548,0.000039334933,0.017674807,0.0002594477,0.69049907],"study_design_scores_gemma":[0.0016291535,0.008046842,0.5522907,0.0043998356,0.000015259558,0.00024095584,0.04013504,0.0031891381,0.00020637049,0.015299572,0.37389183,0.0006553484],"about_ca_topic_score_codex":0.00015797919,"about_ca_topic_score_gemma":0.0000291597,"teacher_disagreement_score":0.6898437,"about_ca_system_score_codex":0.00014612346,"about_ca_system_score_gemma":0.0016840376,"threshold_uncertainty_score":0.5024231},"labels":[],"label_agreement":null},{"id":"W2945812467","doi":"","title":"SPIRALS: a new cognitive approach to clinical reasoning in the emergency department.","year":2018,"lang":"en","type":"article","venue":"The Journal of Macrodynamic Analysis (Memorial University of Newfoundland)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Memorial University of Newfoundland","funders":"","keywords":"Emergency department; Cognition; Computer science; Medical emergency; Medicine; Psychology; Psychiatry","score_opus":0.033481159605731975,"score_gpt":0.3354868305381905,"score_spread":0.30200567093245856,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2945812467","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97400606,0.000113421185,0.019619614,0.0014230636,0.0004798561,0.00024919203,0.0000094635425,0.0000066730277,0.0040926305],"genre_scores_gemma":[0.9958567,0.00040501627,0.0013219229,0.00026013207,0.0014571559,9.020852e-8,0.0000156437,0.000008055823,0.000675278],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9975815,0.00059433444,0.00077447155,0.00020680622,0.000590861,0.0002520672],"domain_scores_gemma":[0.9943867,0.0037268721,0.0007536501,0.00037522658,0.00047652578,0.00028103913],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003605709,0.0001570677,0.00081355515,0.00030450913,0.00014820124,0.000013749859,0.00054665876,0.00011421436,0.00023365345],"category_scores_gemma":[0.008092152,0.00009694735,0.0006801326,0.0014372278,0.00027027237,0.00007624833,0.000099434576,0.00045494348,0.000026228014],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.016362589,0.0022641139,0.93395627,0.00002030003,0.0070117917,0.00023817453,0.009089097,0.001075834,0.00004612317,0.0004370918,0.014523144,0.014975489],"study_design_scores_gemma":[0.0073844055,0.0025502897,0.96141356,0.00045891394,0.014451059,0.00014778624,0.0070243315,0.004348524,0.000002187908,0.00047493345,0.0015253493,0.00021866128],"about_ca_topic_score_codex":0.0026224952,"about_ca_topic_score_gemma":0.003687519,"teacher_disagreement_score":0.027457308,"about_ca_system_score_codex":0.0000980217,"about_ca_system_score_gemma":0.00029854113,"threshold_uncertainty_score":0.96876466},"labels":[],"label_agreement":null},{"id":"W2946222857","doi":"10.1053/j.semtcvs.2019.05.011","title":"Medical Error and “Psycho-Physiology”","year":2019,"lang":"en","type":"letter","venue":"Seminars in Thoracic and Cardiovascular Surgery","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Manitoba; Winnipeg Regional Health Authority","funders":"","keywords":"Workload; Medicine; Cognition; Root cause analysis; Heart rate variability; Harm; Medical emergency; Heart rate; Emergency medicine; Psychiatry; Internal medicine; Psychology; Blood pressure","score_opus":0.025751142252221304,"score_gpt":0.3207664065328046,"score_spread":0.2950152642805833,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2946222857","genre_codex":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.41871578,0.32356992,0.00006152852,0.24897398,0.003622318,0.001102774,0.000057760917,0.00017038049,0.0037255774],"genre_scores_gemma":[0.08693413,0.17624249,0.00040296928,0.7202457,0.008973076,0.00019948145,0.0013162044,0.0003125712,0.005373363],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99615794,0.00045729158,0.0006404642,0.0010748893,0.0010597673,0.00060963084],"domain_scores_gemma":[0.98440754,0.014150665,0.00011106722,0.0010178109,0.00006925032,0.00024368113],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.00300389,0.00046332847,0.0024067557,0.0003061356,0.000048746,0.000026666168,0.00012024665,0.002121785,0.00016659667],"category_scores_gemma":[0.017704314,0.00038770091,0.0011981818,0.00024585292,0.00043975687,0.000039844163,0.00015577966,0.0025122494,0.00005895823],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011727321,0.000057279758,0.024330648,0.0007595353,0.0010075229,0.0058073215,0.00005566367,9.839449e-7,0.0000017095635,0.0000015015918,0.7615296,0.20633094],"study_design_scores_gemma":[0.0009986529,0.00012300418,0.040348463,0.022711867,0.0014950519,0.0030531299,0.000120378514,0.00006442323,0.000012728066,0.00020405231,0.93002045,0.00084782543],"about_ca_topic_score_codex":0.00005886799,"about_ca_topic_score_gemma":0.0000012399183,"teacher_disagreement_score":0.47127175,"about_ca_system_score_codex":0.000035871235,"about_ca_system_score_gemma":0.00032957838,"threshold_uncertainty_score":0.9998575},"labels":[],"label_agreement":null},{"id":"W2946543631","doi":"10.1080/10401334.2019.1610658","title":"Pimping in Residency: The Emotional Roller-Coaster of a Pedagogical Method – A Qualitative Study Using Interviews and Rich Picture Drawings","year":2019,"lang":"en","type":"article","venue":"Teaching and Learning in Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"London Health Sciences Centre; Western University","funders":"Western University of Health Sciences","keywords":"Psychology; Qualitative research; Anxiety; Value (mathematics); Roller coaster; Social psychology; Medical education; Medicine","score_opus":0.12174119049274099,"score_gpt":0.4942782947269374,"score_spread":0.3725371042341964,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2946543631","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99254996,0.0010128844,0.0009131194,0.004530476,0.000070251044,0.00043969223,3.2345184e-7,0.000018220893,0.00046509906],"genre_scores_gemma":[0.99600047,0.000028776874,0.0027346138,0.00053496147,0.00011763728,0.000006873951,0.000003592537,0.000014636453,0.00055845134],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.9955279,0.0029183396,0.0006330795,0.00036048656,0.00035067246,0.0002094702],"domain_scores_gemma":[0.98107845,0.018429628,0.0001965567,0.00016790297,0.00004398864,0.00008348012],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.012086712,0.00017012475,0.0007442811,0.00022049726,0.00010549764,0.000012058867,0.00007289678,0.00009571249,0.0000691005],"category_scores_gemma":[0.038868856,0.00009788084,0.000039895356,0.0002072252,0.00013486136,0.000046249283,0.00009717618,0.0020294022,0.0000013116363],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00022675651,0.00019568054,0.3835817,0.00008858853,0.000038717593,0.000027193677,0.60339504,0.00021420082,0.00029447532,0.00013572408,0.000025074101,0.011776837],"study_design_scores_gemma":[0.005823536,0.002039849,0.5436499,0.010107405,0.0001290498,0.00009189579,0.4293471,0.008058459,0.0000024878302,0.00049712876,0.00010852936,0.00014462082],"about_ca_topic_score_codex":0.001646837,"about_ca_topic_score_gemma":0.00007289966,"teacher_disagreement_score":0.17404795,"about_ca_system_score_codex":0.000036821468,"about_ca_system_score_gemma":0.00006361358,"threshold_uncertainty_score":0.96922714},"labels":[],"label_agreement":null},{"id":"W2946797420","doi":"10.1111/medu.13903","title":"Aligning direct observation and assessment","year":2019,"lang":"en","type":"letter","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Clinical Practice; Psychology; Medical education; Medicine; Nursing","score_opus":0.033054270191953464,"score_gpt":0.397296711810977,"score_spread":0.3642424416190235,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2946797420","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.009155139,0.0006455464,0.00006184171,0.9689608,0.002352045,0.00039210153,0.0000041170224,0.000066232475,0.018362151],"genre_scores_gemma":[0.0030163408,0.00051192066,0.0017506059,0.97368443,0.006026534,0.00007560014,0.00205137,0.00003746244,0.012845739],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9980042,0.00009387875,0.0003780413,0.00041473279,0.0008908666,0.00021831122],"domain_scores_gemma":[0.9933597,0.0058492175,0.00016318457,0.0003376576,0.00014684099,0.0001434145],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00056202605,0.00018359763,0.00044694438,0.000090932794,0.000039950555,0.000033931265,0.000079306745,0.0008753843,0.0006808197],"category_scores_gemma":[0.05346424,0.00015020501,0.00008196193,0.000113357084,0.000075551616,0.000036149137,0.00003795464,0.0014965134,0.00009002977],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000041708636,0.00012357446,0.015296308,0.0001554789,0.000031393956,0.000023941946,0.000035013145,4.6356e-8,0.0000014553144,0.000020283447,0.91648173,0.06782659],"study_design_scores_gemma":[0.000472107,0.0001305782,0.051586594,0.0050874455,0.00020099823,0.000047782938,0.00001758079,0.0002392289,0.0000038867825,0.000092549366,0.941967,0.00015423473],"about_ca_topic_score_codex":0.00017212477,"about_ca_topic_score_gemma":0.0000018855658,"teacher_disagreement_score":0.06767236,"about_ca_system_score_codex":0.0001322976,"about_ca_system_score_gemma":0.0034951635,"threshold_uncertainty_score":0.95450884},"labels":[],"label_agreement":null},{"id":"W2947307719","doi":"10.1186/s12909-019-1558-2","title":"Just-in-time faculty development: a mobile application helps clinical teachers verify and describe clinical reasoning difficulties","year":2019,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université de Montréal; Université de Sherbrooke","funders":"Université de Sherbrooke","keywords":"Process (computing); Medical education; Psychology; Identification (biology); Mobile device; Value (mathematics); Mathematics education; Computer science; Medicine; World Wide Web","score_opus":0.04424399847219175,"score_gpt":0.41711357572547775,"score_spread":0.372869577253286,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2947307719","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9956916,0.00044572222,0.00033120124,0.0006529889,0.00093261315,0.000975274,0.0000024223855,0.000101919715,0.0008662883],"genre_scores_gemma":[0.9801576,0.00042194157,0.01213551,0.0020939691,0.0007473996,0.0003169066,0.000524137,0.00003858855,0.003563926],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.996434,0.0003055498,0.0013918502,0.00074034784,0.0007677962,0.00036047687],"domain_scores_gemma":[0.99027616,0.007848043,0.0002846777,0.0004618274,0.0001368673,0.0009924135],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0034558908,0.00022757219,0.0006925022,0.00009441109,0.00006593659,0.00003736505,0.00019226238,0.0006285647,0.00044159257],"category_scores_gemma":[0.07242872,0.00018832498,0.00015216872,0.00023950073,0.00025707934,0.00008343384,0.00010456074,0.0008056166,0.00065546506],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015677881,0.0017592679,0.8209481,0.00006454074,0.000026756305,0.0000024912686,0.0006115429,0.0000020203897,0.000008723945,0.000077405006,0.0050694803,0.17127289],"study_design_scores_gemma":[0.0033885771,0.00031292363,0.9726867,0.0022892768,0.00008443842,0.00003760779,0.0011446786,0.0017193871,0.000017075745,0.000028745753,0.018044919,0.00024566846],"about_ca_topic_score_codex":0.00009225931,"about_ca_topic_score_gemma":0.000047013724,"teacher_disagreement_score":0.17102721,"about_ca_system_score_codex":0.0001290795,"about_ca_system_score_gemma":0.0033912205,"threshold_uncertainty_score":0.93538463},"labels":[],"label_agreement":null},{"id":"W2948281555","doi":"10.1016/j.jcrc.2019.06.002","title":"Technology-mediated macrocognition: Investigating how physicians, nurses, and respiratory therapists make critical decisions","year":2019,"lang":"en","type":"article","venue":"Journal of Critical Care","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"North York General Hospital; SickKids Foundation; Hospital for Sick Children; University of Toronto","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Sensemaking; Anticipation (artificial intelligence); Critical thinking; Process (computing); Medicine; Intensive care; Coding (social sciences); Decision support system; Clinical decision making; Decision-making; Knowledge management; Psychology; Computer science; Operations management; Artificial intelligence","score_opus":0.033238151883537945,"score_gpt":0.3763557653505219,"score_spread":0.34311761346698394,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2948281555","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95873684,0.0066128396,0.00010406003,0.032810763,0.00057861634,0.00016368357,0.000032706026,0.00004442598,0.00091606297],"genre_scores_gemma":[0.9917027,0.00003967223,0.0032479262,0.004487686,0.0004534336,0.000004669814,0.0000070314263,0.000025404264,0.000031483483],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99833745,0.000108039036,0.00047563954,0.0002339472,0.0005330614,0.00031186242],"domain_scores_gemma":[0.9813164,0.0165997,0.00010242852,0.00022166198,0.0013207077,0.00043913416],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0003428571,0.00016518593,0.0005182529,0.0001820283,0.00010202475,0.00006306629,0.00010866111,0.00027216793,0.00010478848],"category_scores_gemma":[0.17298596,0.00013147054,0.00015473888,0.00027361294,0.0007304567,0.00010230345,0.000045851262,0.000997955,0.000030815852],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010104907,0.002234637,0.3922406,0.0019522836,0.00026439727,0.0048889457,0.0031241141,0.0000030501917,0.037913963,0.03163666,0.008730989,0.51599985],"study_design_scores_gemma":[0.037251744,0.029780619,0.5333222,0.08828711,0.0046159592,0.009873747,0.11051875,0.00046537168,0.026527539,0.11463958,0.04153172,0.0031856298],"about_ca_topic_score_codex":7.5067004e-7,"about_ca_topic_score_gemma":0.0000011324577,"teacher_disagreement_score":0.5128142,"about_ca_system_score_codex":0.000052416508,"about_ca_system_score_gemma":0.00022199488,"threshold_uncertainty_score":0.8339803},"labels":[],"label_agreement":null},{"id":"W2949993659","doi":"10.1186/s12909-019-1638-3","title":"Does providing the correct diagnosis as feedback after self-explanation improve medical students diagnostic performance?","year":2019,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":22,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université de Sherbrooke","funders":"Université de Sherbrooke","keywords":"Medical diagnosis; Diagnostic accuracy; Medicine; Control (management); Medical physics; Computer science; Artificial intelligence; Internal medicine; Radiology","score_opus":0.007607194685092862,"score_gpt":0.3264679375921091,"score_spread":0.31886074290701627,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2949993659","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9813072,0.0003263884,0.000046410198,0.00976548,0.0060058124,0.0013531058,0.0000029553314,0.00012766138,0.0010649945],"genre_scores_gemma":[0.9845042,0.0017059797,0.00017880969,0.008841058,0.0016089143,0.0011680843,0.000099240584,0.00003552045,0.0018581436],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9951622,0.00020723426,0.0006554359,0.00053388084,0.0030275434,0.00041371054],"domain_scores_gemma":[0.9683058,0.02958265,0.00022112313,0.0006276816,0.00026267843,0.001000036],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0020748435,0.00024847905,0.00039602353,0.00009608388,0.00011212508,0.00009260687,0.00048383023,0.0003813728,0.006864549],"category_scores_gemma":[0.20623276,0.00012605885,0.00015258756,0.0002852597,0.00013228526,0.0001372464,0.00016289062,0.0006359501,0.0016308395],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014904914,0.001748034,0.9227417,0.00021165886,0.00005274766,0.000010440338,0.00065859023,0.0000011430533,0.000001585253,0.000049482842,0.0080691,0.06630647],"study_design_scores_gemma":[0.0031921058,0.0008327003,0.97353315,0.00806121,0.00037842637,0.00009992292,0.0011021513,0.0023400157,0.00029053108,0.0001464056,0.009665096,0.00035826705],"about_ca_topic_score_codex":0.00024072242,"about_ca_topic_score_gemma":0.00008244839,"teacher_disagreement_score":0.20415792,"about_ca_system_score_codex":0.00020857272,"about_ca_system_score_gemma":0.004517111,"threshold_uncertainty_score":0.9991465},"labels":[],"label_agreement":null},{"id":"W2951839764","doi":"10.1136/bmjqs-2018-009147","title":"Safety-I, Safety-II and burnout: how complexity science can help clinician wellness","year":2019,"lang":"en","type":"article","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":63,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Burnout; Medicine; Patient safety; Workplace safety; Medical education; Nursing; Applied psychology; Occupational safety and health; Health care; Clinical psychology; Psychology; Pathology","score_opus":0.06828426058351167,"score_gpt":0.40663189149668033,"score_spread":0.33834763091316866,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2951839764","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.88368416,0.00029748408,0.00052050815,0.084416784,0.0015379359,0.0020265535,0.00037092547,0.00025464827,0.026890982],"genre_scores_gemma":[0.98913574,0.00042653488,0.00093472406,0.004648403,0.00055724196,0.000015048854,0.00018631769,0.000041802254,0.0040542074],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9942726,0.0004530136,0.0018028414,0.0012287371,0.0012932473,0.00094956165],"domain_scores_gemma":[0.9849296,0.01154506,0.00054141495,0.0014995917,0.00043447214,0.0010498657],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.010159934,0.00043532875,0.0013712303,0.0001483585,0.00083485484,0.00009733658,0.00045044365,0.0002873095,0.0005348327],"category_scores_gemma":[0.044154845,0.00036512656,0.00029709106,0.0007681687,0.0017980947,0.00019362917,0.00056486775,0.00073027285,0.00019362835],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008636277,0.0019886598,0.6803813,0.0006883307,0.0002965766,0.00014171444,0.003839821,0.00004183188,0.0011244267,0.11276124,0.0051066806,0.18499313],"study_design_scores_gemma":[0.007524235,0.00087930856,0.943211,0.0021906935,0.00017052621,0.00010173788,0.002304573,0.00042729595,0.0003434331,0.0033082585,0.038735252,0.000803667],"about_ca_topic_score_codex":0.0012766804,"about_ca_topic_score_gemma":0.00039135088,"teacher_disagreement_score":0.2628297,"about_ca_system_score_codex":0.00030319943,"about_ca_system_score_gemma":0.0013234845,"threshold_uncertainty_score":0.9998801},"labels":[],"label_agreement":null},{"id":"W2953967017","doi":"10.1111/medu.13919","title":"The information distortion bias: implications for medical decisions","year":2019,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Interior Health","funders":"","keywords":"Medical diagnosis; Distortion (music); Interpretation (philosophy); Psychology; Certainty; Action (physics); Social psychology; Computer science; Medicine; Epistemology; Pathology","score_opus":0.025993065093966157,"score_gpt":0.4020740150221425,"score_spread":0.37608094992817637,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2953967017","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5643253,0.0005859069,0.028894763,0.39160475,0.0047097607,0.0020620709,0.000020019637,0.00016908975,0.007628401],"genre_scores_gemma":[0.97950935,0.00060649525,0.0006701807,0.016044836,0.0006113324,0.00050282595,0.00085767213,0.0000134179,0.0011839127],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9983554,0.000038985014,0.00054651295,0.00015418943,0.00070362416,0.00020132626],"domain_scores_gemma":[0.9752035,0.023062283,0.00017951967,0.00053953286,0.00045581517,0.0005593463],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0013333634,0.000087146596,0.00016064013,0.000057684596,0.00018795213,0.00003918804,0.00017630911,0.00021444197,0.0005925179],"category_scores_gemma":[0.4127867,0.000054262404,0.000116437586,0.00020636378,0.00008443731,0.00011350989,0.000030764226,0.00022767983,0.00042429264],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000062755316,0.00033132886,0.021800661,0.000011246258,0.000015283733,9.589028e-8,0.00009690678,7.647603e-7,0.0000022280988,0.010727512,0.15340076,0.8135505],"study_design_scores_gemma":[0.0016226677,0.000251036,0.36050564,0.0015419879,0.00007828622,0.000058260655,0.00025798444,0.0020382742,0.000015324149,0.0048623085,0.6286392,0.00012902143],"about_ca_topic_score_codex":0.000039598017,"about_ca_topic_score_gemma":0.00002967123,"teacher_disagreement_score":0.8134214,"about_ca_system_score_codex":0.00010134097,"about_ca_system_score_gemma":0.002994112,"threshold_uncertainty_score":0.64876556},"labels":[],"label_agreement":null},{"id":"W2955744164","doi":"10.1515/dx-2018-0107","title":"Competencies for improving diagnosis: an interprofessional framework for education and training in health care","year":2019,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":97,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Teamwork; Patient safety; Medicine; Health care; Quality (philosophy); Set (abstract data type); Medical education; Focus group; Nursing; Psychology; Computer science","score_opus":0.033928516371815785,"score_gpt":0.3959838954297819,"score_spread":0.36205537905796614,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2955744164","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98727196,0.0036990205,0.0008031732,0.005350457,0.00089312217,0.0018201671,0.00007061583,0.000045163364,0.000046297573],"genre_scores_gemma":[0.9520912,0.00023913331,0.039676294,0.004264923,0.00031026494,0.0031502498,0.00020423236,0.000030621668,0.000033081695],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.998709,0.000045053028,0.0003956851,0.00040337685,0.00012980013,0.0003170698],"domain_scores_gemma":[0.9835862,0.015717546,0.00013699225,0.00020987482,0.00012014747,0.00022923923],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00039658585,0.00014788716,0.0004269516,0.00011548887,0.000087535846,0.00003051057,0.00007429018,0.00014729839,0.00004139313],"category_scores_gemma":[0.017502265,0.00013181717,0.0000958186,0.00010726936,0.000039344122,0.000090063724,0.000030644904,0.00019057929,0.0000034722423],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013992023,0.00038048136,0.6071324,0.00049874134,0.000011859849,4.910082e-7,0.008353565,0.000003536918,0.000005822239,0.0037921397,0.00042686472,0.37925422],"study_design_scores_gemma":[0.006454099,0.006634106,0.8233534,0.036070168,0.00017214227,0.0000141733835,0.10269509,0.0015544408,0.000493195,0.013246393,0.008585131,0.000727677],"about_ca_topic_score_codex":0.0003333241,"about_ca_topic_score_gemma":0.00018261025,"teacher_disagreement_score":0.37852654,"about_ca_system_score_codex":0.00011467613,"about_ca_system_score_gemma":0.0007871458,"threshold_uncertainty_score":0.99077374},"labels":[],"label_agreement":null},{"id":"W2960173995","doi":"10.1051/pmed/2019018","title":"S’instrumenter pour évaluer","year":2018,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec en Outaouais; Université de Montréal","funders":"","keywords":"Humanities; Valuation (finance); Philosophy; Economics","score_opus":0.0638670924343104,"score_gpt":0.38388807829630794,"score_spread":0.32002098586199756,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2960173995","genre_codex":"empirical","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.42553246,0.053464834,0.0029258817,0.11764173,0.026680157,0.0008181062,0.000153859,0.00044717902,0.3723358],"genre_scores_gemma":[0.30336338,0.0032485325,0.004508274,0.013749643,0.017972898,0.000029050332,0.00004584382,0.00009881532,0.65698355],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99666584,0.00017966742,0.0006877938,0.0006787518,0.00062664086,0.0011612872],"domain_scores_gemma":[0.9915588,0.0062183165,0.00013460666,0.00078786927,0.00023845644,0.001061967],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0008643726,0.00035458425,0.0006037306,0.00009045089,0.00017318956,0.000069618734,0.0002855964,0.00047399383,0.03449414],"category_scores_gemma":[0.03899544,0.00031689854,0.00033746427,0.00029281995,0.0014422657,0.00009291061,0.00022722513,0.0005871526,0.035872463],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014616782,0.0012428189,0.034956872,0.000056395547,0.00016389738,0.00025503343,0.0003080562,7.5199887e-7,0.00016338142,0.0011545941,0.4345256,0.5270264],"study_design_scores_gemma":[0.0044391397,0.0014565883,0.037051678,0.006545461,0.00058187795,0.00018197966,0.00037053984,0.0013910473,0.0027692337,0.0022562223,0.94246674,0.0004895164],"about_ca_topic_score_codex":0.00014903008,"about_ca_topic_score_gemma":0.00001134872,"teacher_disagreement_score":0.5265369,"about_ca_system_score_codex":0.00016700792,"about_ca_system_score_gemma":0.00056773314,"threshold_uncertainty_score":0.9999283},"labels":[],"label_agreement":null},{"id":"W2961947282","doi":"10.1080/0142159x.2019.1635686","title":"The terminology of clinical reasoning in health professions education: Implications and considerations","year":2019,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":86,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Centre for Interdisciplinary Research in Rehabilitation; McGill University Health Centre","funders":"","keywords":"Terminology; Verbal reasoning; Categorization; Context (archaeology); Psychology; Analytic reasoning; Coding (social sciences); Deductive reasoning; Medical education; Computer science; Medicine; Artificial intelligence; Linguistics; Cognition; Sociology","score_opus":0.07095953125832484,"score_gpt":0.4746362165772631,"score_spread":0.4036766853189383,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2961947282","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.85671335,0.0010989977,0.000031441035,0.13362928,0.00036097487,0.00044809244,0.0000010689076,0.000020703545,0.0076960707],"genre_scores_gemma":[0.99281996,0.0005396195,0.00055838167,0.0023157832,0.00014620533,0.000051232142,0.000010609977,0.000007878204,0.0035503276],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.998278,0.0003090259,0.0008234336,0.00021390014,0.00019413074,0.00018147063],"domain_scores_gemma":[0.989704,0.009303752,0.00018868707,0.0003822285,0.000077729965,0.000343593],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0021169828,0.000068124304,0.00035867922,0.00003805127,0.000084377825,0.000006153951,0.00006343127,0.00017485346,0.0007058688],"category_scores_gemma":[0.094593,0.000042247782,0.00005636218,0.00010537682,0.00032295048,0.000017309241,0.00005490864,0.0005721611,0.000025557949],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000015493923,0.00047397023,0.90609556,0.000016249596,0.000013382918,0.000001507249,0.00018954845,1.0075174e-7,0.000001270244,0.018188331,0.007922141,0.06708242],"study_design_scores_gemma":[0.0011697062,0.00020238986,0.98239535,0.0014282321,0.000014998536,0.00007907419,0.00051827775,0.0003889918,0.000001176706,0.0031749003,0.010575369,0.00005153919],"about_ca_topic_score_codex":0.00013419842,"about_ca_topic_score_gemma":0.00007651246,"teacher_disagreement_score":0.13610658,"about_ca_system_score_codex":0.000020723306,"about_ca_system_score_gemma":0.0020357897,"threshold_uncertainty_score":0.9130336},"labels":[],"label_agreement":null},{"id":"W2963198629","doi":"10.1007/s10459-019-09903-7","title":"It’s the destination: diagnostic accuracy and reasoning","year":2019,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Competence (human resources); Reliability (semiconductor); Diagnostic test; Test (biology); Computer science; Retrospective cohort study; Medical diagnosis; Psychology; Medicine; Pediatrics; Social psychology; Pathology","score_opus":0.024076915041461578,"score_gpt":0.4310750063265015,"score_spread":0.4069980912850399,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2963198629","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.90469545,0.020194432,0.00010565775,0.06232451,0.001543943,0.0009290885,7.6900335e-7,0.000028653945,0.010177489],"genre_scores_gemma":[0.98003095,0.008755977,0.0031673182,0.0076440037,0.00012687243,0.000055281347,0.0000050137255,0.00000389278,0.00021068419],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99891776,0.00007192186,0.00029707077,0.00026715983,0.00023343034,0.00021267969],"domain_scores_gemma":[0.9735028,0.025905346,0.00019200695,0.0001949043,0.00007215619,0.00013279315],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0013658466,0.00006942445,0.00013941998,0.00006991947,0.00016899608,0.00003268774,0.00009745868,0.000025609714,0.000047537276],"category_scores_gemma":[0.06677007,0.000045753168,0.0000152461835,0.0004720201,0.00020208735,0.00027928074,0.00002282387,0.00013802198,0.000028724762],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000007454815,0.00009131717,0.77054,0.00007206189,6.584198e-7,4.4545243e-7,0.00064466713,0.00017110248,0.0000010425823,0.0055825124,0.0004995546,0.22238918],"study_design_scores_gemma":[0.00047069913,0.00053400954,0.9533044,0.0048750336,0.000007673092,0.00009043226,0.002080795,0.001800381,0.000011842783,0.004834593,0.03187545,0.0001147064],"about_ca_topic_score_codex":0.00019189189,"about_ca_topic_score_gemma":0.000063375024,"teacher_disagreement_score":0.22227447,"about_ca_system_score_codex":0.000068777925,"about_ca_system_score_gemma":0.0009207842,"threshold_uncertainty_score":0.94109094},"labels":[],"label_agreement":null},{"id":"W2964216090","doi":"10.2196/13386","title":"The Impact of Medical Students’ Individual Teaching Format Choice on the Learning Outcome Related to Clinical Reasoning","year":2019,"lang":"en","type":"article","venue":"JMIR Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Formative assessment; Key (lock); Feature (linguistics); Outcome (game theory); Computer science; Medical education; Mathematics education; Psychology; Artificial intelligence; Medicine; Mathematics; Linguistics","score_opus":0.03757859690385167,"score_gpt":0.5054586306448167,"score_spread":0.467880033740965,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2964216090","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9611377,0.00009064499,0.000013578078,0.03366306,0.001261037,0.0008636782,0.0000013000636,0.00006349663,0.0029055022],"genre_scores_gemma":[0.9928324,0.000081157414,0.000038435377,0.0049156174,0.00057755696,0.00008664331,0.00004261468,0.000028678527,0.001396857],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99348366,0.0007147186,0.0012995164,0.00035773756,0.0037296982,0.0004146524],"domain_scores_gemma":[0.9589703,0.03870897,0.00040755162,0.0006567598,0.00015479,0.0011016158],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.013077921,0.00020561201,0.0005111784,0.000086674336,0.00024531884,0.00008432907,0.0009404853,0.00047219265,0.0011785337],"category_scores_gemma":[0.39710462,0.00009941316,0.00036650113,0.00029681763,0.00019792389,0.000058041445,0.00028525697,0.002916087,0.00043499575],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013273467,0.0011671572,0.8935295,0.0000122072715,0.00013901651,0.0000029629523,0.0010967721,0.000008557431,7.9446005e-7,0.0014823695,0.020974424,0.08145355],"study_design_scores_gemma":[0.0016425925,0.0013309906,0.98411876,0.0031206594,0.000066173445,0.00003866056,0.0013451852,0.0011044741,0.0000016925735,0.00008935564,0.007017516,0.00012393938],"about_ca_topic_score_codex":0.00028115732,"about_ca_topic_score_gemma":0.000007076735,"teacher_disagreement_score":0.3840267,"about_ca_system_score_codex":0.00013238224,"about_ca_system_score_gemma":0.0017957295,"threshold_uncertainty_score":0.9997345},"labels":[],"label_agreement":null},{"id":"W2965081066","doi":"10.1097/acm.0000000000002907","title":"Reconsidering Basic: Integrating Social and Behavioral Sciences to Support Learning","year":2019,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":29,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Women's College Hospital; The Wilson Centre; Centre for Addiction and Mental Health; University of Toronto; Toronto Dementia Research Alliance; University Health Network","funders":"","keywords":"Curriculum; Scope (computer science); Health care; Psychology; Basic science; Behavioural sciences; Cognition; Medical education; Medicine; Computer science; Cognitive psychology; Psychotherapist; Pedagogy; Neuroscience","score_opus":0.18181794554515876,"score_gpt":0.44561440066779345,"score_spread":0.2637964551226347,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2965081066","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9833358,0.00016063607,0.00008181676,0.011168831,0.00032823076,0.00025649735,7.1495265e-7,0.00007739036,0.004590069],"genre_scores_gemma":[0.9922084,0.00009387702,0.00054289796,0.004015808,0.00062933913,0.000009311331,0.000009013565,0.000016009455,0.0024753336],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99845123,0.000045869012,0.00042679984,0.00038571493,0.00036167153,0.00032872846],"domain_scores_gemma":[0.996571,0.0028874571,0.00010432974,0.000092194634,0.00005487505,0.000290119],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0012543715,0.00014414638,0.00044382756,0.0001366014,0.00013722405,0.000008652516,0.00008923902,0.00017458564,0.0007824742],"category_scores_gemma":[0.018717255,0.000105076666,0.00003431654,0.0002585197,0.0002657707,0.0000623993,0.00008386439,0.0009041366,0.00011496263],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007262928,0.000015402997,0.8592375,0.000041607425,0.000017825414,0.000038539038,0.0110051,0.000006248339,0.007046615,0.0006354936,0.011061331,0.110821724],"study_design_scores_gemma":[0.028846068,0.04284635,0.5974307,0.04303076,0.0018707779,0.003302523,0.12678692,0.003974001,0.0067737442,0.0035705913,0.13868676,0.002880796],"about_ca_topic_score_codex":0.00007850048,"about_ca_topic_score_gemma":0.00000511327,"teacher_disagreement_score":0.2618068,"about_ca_system_score_codex":0.000033641507,"about_ca_system_score_gemma":0.000108979286,"threshold_uncertainty_score":0.9895485},"labels":[],"label_agreement":null},{"id":"W2965808027","doi":"10.1002/aet2.10379","title":"Showing Your Thinking: Using Mind Maps to Understand the Gaps Between Experienced Emergency Physicians and Their Students","year":2019,"lang":"en","type":"article","venue":"AEM Education and Training","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":20,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; University of Manitoba; McMaster University; University of British Columbia","funders":"","keywords":"Mind map; Interpretation (philosophy); Concept map; Psychology; Interview; Epistemology; Mathematics education; Cognitive psychology; Computer science","score_opus":0.11110486817981069,"score_gpt":0.4105669897331808,"score_spread":0.2994621215533701,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2965808027","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9957319,0.0003671027,0.00016240777,0.0017256706,0.00046706357,0.00029866182,0.0000030448778,0.000014719271,0.00122941],"genre_scores_gemma":[0.9979498,0.000040119343,0.00050967815,0.0008452243,0.00026424488,0.000006014174,0.00001208319,0.000014232021,0.00035859563],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99913466,0.000033386445,0.00020502908,0.00025564907,0.00018358299,0.00018770494],"domain_scores_gemma":[0.9987787,0.0008037379,0.000063917825,0.00016664529,0.000045357912,0.00014162078],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.000328106,0.00011858924,0.00020777271,0.000050286897,0.00017970761,0.000060044942,0.00008050225,0.000047839374,0.000052805004],"category_scores_gemma":[0.00095970504,0.000082199476,0.00004064624,0.00015307292,0.00004344985,0.000063278356,0.000043038774,0.00012680917,0.0000067742744],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000020124851,0.00014352697,0.50397617,0.000016563316,0.0000789759,8.838954e-7,0.2693233,0.0000053955437,0.00067296776,0.00030814268,0.00019457516,0.2252594],"study_design_scores_gemma":[0.00077466544,0.00016783622,0.65334374,0.0011653103,0.00010558517,0.000018656285,0.34227067,0.00008507176,0.00023700124,0.00085180404,0.00078610214,0.00019357173],"about_ca_topic_score_codex":0.000038057628,"about_ca_topic_score_gemma":0.0000047587932,"teacher_disagreement_score":0.22506583,"about_ca_system_score_codex":0.000027036715,"about_ca_system_score_gemma":0.00010332664,"threshold_uncertainty_score":0.33519974},"labels":[],"label_agreement":null},{"id":"W2968386989","doi":"10.4300/jgme-d-19-00464.1","title":"Generalizability Theory Made Simple(r): An Introductory Primer to G-Studies","year":2019,"lang":"en","type":"editorial","venue":"Journal of Graduate Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":24,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; St. Joseph’s Healthcare Hamilton","funders":"","keywords":"Generalizability theory; Competence (human resources); Objective structured clinical examination; Medical education; Psychology; Validity; External validity; Educational measurement; Computer science; Applied psychology; Psychometrics; Social psychology; Curriculum; Medicine; Clinical psychology; Pedagogy","score_opus":0.050091625625308445,"score_gpt":0.44163069179937203,"score_spread":0.3915390661740636,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2968386989","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.14338024,0.0036183337,0.00014438783,0.016445493,0.8358127,0.00041881183,0.000015976831,0.000026546275,0.00013753225],"genre_scores_gemma":[0.009331276,0.00328202,0.0011636018,0.0042328676,0.9792611,0.0000211125,0.00034958558,0.0000779667,0.0022804528],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9934781,0.0008629822,0.0016372534,0.0005185206,0.0031273796,0.00037573825],"domain_scores_gemma":[0.9462569,0.041140895,0.0014487136,0.0015438359,0.0064719203,0.0031377464],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0088157,0.00037746356,0.0014851689,0.00033139688,0.000066765904,0.00003579218,0.00043503282,0.00082753255,0.00047799124],"category_scores_gemma":[0.65392256,0.00026453857,0.00040160702,0.00026623358,0.00027454385,0.00011799982,0.00010065118,0.0023120185,0.00010603452],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007869533,0.0014078913,0.0004734034,0.00021285056,0.00032539043,0.000023683753,0.00061462214,0.0000038111025,0.000028543349,0.00006495195,0.956923,0.039134897],"study_design_scores_gemma":[0.0018753668,0.0026698243,0.0047294046,0.005628934,0.0013175503,0.00013215769,0.00076011854,0.000013022422,0.00007984308,0.004566731,0.97781384,0.00041322116],"about_ca_topic_score_codex":0.00006566897,"about_ca_topic_score_gemma":0.0000066202665,"teacher_disagreement_score":0.64510685,"about_ca_system_score_codex":0.0005940812,"about_ca_system_score_gemma":0.01356235,"threshold_uncertainty_score":0.9999897},"labels":[],"label_agreement":null},{"id":"W2969250926","doi":"10.1080/0142159x.2019.1652260","title":"Are rating scales really better than checklists for measuring increasing levels of expertise?","year":2019,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":20,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"University of Ottawa","keywords":"Checklist; Rating scale; Psychology; Test (biology); Measure (data warehouse); Objective structured clinical examination; Medical education; Rating system; Applied psychology; Medicine; Clinical psychology; Computer science; Data mining; Cognitive psychology; Developmental psychology","score_opus":0.06439051680890505,"score_gpt":0.35095250320841537,"score_spread":0.2865619863995103,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2969250926","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9915117,0.00032541918,0.00072786544,0.0032440408,0.00022160816,0.0003904917,0.000005524073,0.00007025879,0.003503081],"genre_scores_gemma":[0.99509895,0.000008571979,0.0019528947,0.0010724594,0.0005974965,0.000030566593,0.000012091248,0.00003988164,0.0011870612],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99778515,0.00012601713,0.0005744087,0.00035021198,0.0008146662,0.00034952728],"domain_scores_gemma":[0.9955441,0.0033194243,0.00024609276,0.00034913063,0.00013756992,0.0004036847],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001920825,0.00017216783,0.0006126277,0.00005631998,0.00004319624,0.000013111409,0.00014893194,0.0002833462,0.0010573092],"category_scores_gemma":[0.11287552,0.00013093068,0.0002209994,0.00009507133,0.00013432723,0.000043082164,0.000072375995,0.0003655052,0.000036506666],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00023332215,0.00033200628,0.94457376,0.00018075439,0.00008984844,0.00003383043,0.00041168762,9.0807504e-7,0.0020092514,0.00003563159,0.0013712108,0.050727785],"study_design_scores_gemma":[0.004992517,0.00027100847,0.9769023,0.010037327,0.00011811043,0.00006948859,0.00037751446,0.0032869931,0.0026560107,0.00021189489,0.0007918962,0.00028495636],"about_ca_topic_score_codex":0.00010812789,"about_ca_topic_score_gemma":0.000020932406,"teacher_disagreement_score":0.110954694,"about_ca_system_score_codex":0.00006168526,"about_ca_system_score_gemma":0.000123569,"threshold_uncertainty_score":0.9998559},"labels":[],"label_agreement":null},{"id":"W2970167379","doi":"10.1111/medu.13872","title":"Critical thinking, biases and dual processing: The enduring myth of generalisable skills","year":2019,"lang":"en","type":"review","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":89,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Psychology; Mythology; Context (archaeology); Cognitive skill; Popularity; Critical thinking; Cognition; Medical education; Pedagogy; Social psychology; Medicine","score_opus":0.07028802884795039,"score_gpt":0.45516530068263755,"score_spread":0.3848772718346872,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2970167379","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00042469703,0.9914914,0.00001839441,0.0055680526,0.0009268096,0.0004941789,0.000007774105,0.000030574054,0.0010381135],"genre_scores_gemma":[0.0003552542,0.9928796,0.0004136044,0.0036803996,0.0012001022,0.00008264204,0.0001762062,0.00003751067,0.001174681],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9976265,0.00016801486,0.0007509547,0.00038665364,0.0008144467,0.000253462],"domain_scores_gemma":[0.9825815,0.016066626,0.0003054478,0.00047356586,0.00022783734,0.00034501986],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0009834409,0.00024544264,0.0012004712,0.000096230724,0.00009416193,0.000044344375,0.00016827165,0.00044697872,0.0003807658],"category_scores_gemma":[0.21537866,0.00013760917,0.0002187488,0.00022650919,0.00041519268,0.000046451605,0.00009591066,0.0006579577,0.000036906702],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000037457655,0.0006899638,0.000075830256,0.00919859,0.000041036765,0.0000062085687,0.0002701424,5.9182792e-8,5.2740106e-8,0.00063165603,0.022718921,0.9663638],"study_design_scores_gemma":[0.00025231822,0.00012102291,0.00013192625,0.1641682,0.0015336788,0.00036566693,0.00007087292,0.000033171505,0.0000017407757,0.00021119378,0.8329477,0.00016256626],"about_ca_topic_score_codex":0.000073858704,"about_ca_topic_score_gemma":0.0000018815891,"teacher_disagreement_score":0.96620125,"about_ca_system_score_codex":0.000052745796,"about_ca_system_score_gemma":0.005580244,"threshold_uncertainty_score":0.9899114},"labels":[],"label_agreement":null},{"id":"W2970977809","doi":"10.11157/fohpe.v20i2.321","title":"Script concordance test examinations: Student perception and study approaches","year":2019,"lang":"en","type":"article","venue":"Focus on Health Professional Education A Multi-Professional Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Concordance; Construct (python library); Multiple choice; Test (biology); Psychology; Construct validity; Medical education; Perception; Psychometrics; Medicine; Clinical psychology; Computer science; Significant difference; Internal medicine","score_opus":0.10711796087989668,"score_gpt":0.4778008864963227,"score_spread":0.370682925616426,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2970977809","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9532461,0.00087770395,0.0000707717,0.03623723,0.006281496,0.0028251426,0.000018352954,0.000058122314,0.0003850423],"genre_scores_gemma":[0.9755314,0.0002540784,0.0035677264,0.0062666168,0.00082140934,0.00030318098,0.00006999466,0.00004723664,0.013138343],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99562275,0.0007804911,0.0011661947,0.00066128507,0.0012600031,0.0005092907],"domain_scores_gemma":[0.9926098,0.004773694,0.0007460132,0.00038482383,0.0005555583,0.00093010196],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0023287975,0.00034786848,0.000638218,0.00035690985,0.0008423687,0.000067419256,0.00020114667,0.00020546193,0.00063101965],"category_scores_gemma":[0.006969909,0.00025080665,0.00012648599,0.0002993415,0.00010924352,0.00022706752,0.00012341407,0.0016629242,0.00030604764],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004672826,0.019645847,0.83167654,0.00015112715,0.00006325598,0.000011278298,0.006572718,0.0000052767214,0.00015236477,0.00045891997,0.016289419,0.12450596],"study_design_scores_gemma":[0.0054665124,0.0024208762,0.9382606,0.0061674286,0.000045282002,0.00020870366,0.04571664,0.00038096367,0.000007289968,0.0002905142,0.00079473585,0.00024041347],"about_ca_topic_score_codex":0.00007988515,"about_ca_topic_score_gemma":0.000022244973,"teacher_disagreement_score":0.124265544,"about_ca_system_score_codex":0.00056091265,"about_ca_system_score_gemma":0.0045513595,"threshold_uncertainty_score":0.9999944},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2971249398","doi":"10.1016/j.annemergmed.2019.06.023","title":"A Scoping Review of Physicians’ Clinical Reasoning in Emergency Departments","year":2019,"lang":"en","type":"review","venue":"Annals of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":35,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Centre for Interdisciplinary Research in Rehabilitation; St. Michael's Hospital; McGill University Health Centre; Montreal Children's Hospital","funders":"","keywords":"Medicine; Malpractice; Cognition; Specialty; Workload; Context (archaeology); Causal reasoning; Family medicine; Psychiatry","score_opus":0.4080202960413548,"score_gpt":0.5860777957770291,"score_spread":0.1780574997356743,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2971249398","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00009169526,0.98655236,0.000008471365,0.00074961275,0.0037881634,0.0031035284,0.000039456747,0.00002519947,0.005641543],"genre_scores_gemma":[0.000016074462,0.9969365,0.000088106455,0.0004688082,0.0009466692,0.00012323896,0.0004100072,0.00009011657,0.0009204748],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.988471,0.0006034782,0.008421337,0.00089085655,0.0009986546,0.00061471394],"domain_scores_gemma":[0.98908335,0.0045536957,0.003928357,0.0013927987,0.0007492567,0.0002925642],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0043941974,0.0006748334,0.009119728,0.00039854043,0.000023084578,6.818337e-7,0.0004492992,0.00049342733,0.0039589233],"category_scores_gemma":[0.104118876,0.0005001605,0.002244392,0.0015740364,0.00017544065,0.00006403163,0.00014876518,0.00084391714,0.00009549737],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000040702107,0.0007804431,0.0038199294,0.51549405,0.00052252517,0.000026107065,0.00003255161,1.5054906e-7,2.5242386e-7,0.00005379828,0.061226774,0.41800272],"study_design_scores_gemma":[0.000664723,0.0010930783,0.0006410984,0.86679876,0.0027184733,0.0000070534525,0.000010573724,0.0000026066366,0.0000010679298,0.000039283652,0.12772182,0.00030145954],"about_ca_topic_score_codex":0.000110954934,"about_ca_topic_score_gemma":0.00001834337,"teacher_disagreement_score":0.41770127,"about_ca_system_score_codex":0.00002463612,"about_ca_system_score_gemma":0.00074011664,"threshold_uncertainty_score":0.999745},"labels":[],"label_agreement":null},{"id":"W2971458922","doi":"10.2196/14044","title":"Differential Diagnosis Assessment in Ambulatory Care With an Automated Medical History–Taking Device: Pilot Randomized Controlled Trial","year":2019,"lang":"en","type":"article","venue":"JMIR Medical Informatics","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":24,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Université de Genève; Hôpitaux Universitaires de Genève","keywords":"Medical diagnosis; Anamnesis; Medicine; Randomized controlled trial; Differential diagnosis; Ambulatory; Outpatient clinic; Gold standard (test); Physical therapy; Emergency medicine; Pediatrics; Surgery; Internal medicine; Radiology; Pathology","score_opus":0.022089495260936075,"score_gpt":0.3709031590523351,"score_spread":0.34881366379139905,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2971458922","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9864421,0.00017717131,0.00017496353,0.00047474614,0.0012312803,0.0065220804,0.000004101331,0.00044715672,0.0045264303],"genre_scores_gemma":[0.9928775,0.0001613593,0.0011713421,0.0030759927,0.0003433089,0.0019883618,0.00024548417,0.000043559314,0.00009310563],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.99271697,0.00044267843,0.0024919417,0.00030233437,0.0034528845,0.00059320265],"domain_scores_gemma":[0.9827045,0.01429808,0.00083400484,0.0005811778,0.00024509197,0.0013371236],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0022244425,0.00042896866,0.003083302,0.00023735911,0.000051294555,0.00004404402,0.00034666958,0.0006074683,0.0065474766],"category_scores_gemma":[0.020253053,0.00026207088,0.00046023348,0.00021733406,0.0004284522,0.00019954197,0.00011651641,0.0014139598,0.000086842716],"study_design_candidate":"randomized_trial","study_design_consensus":"randomized_trial","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.9621726,0.0035549304,0.018340679,0.00067004765,0.000589735,0.00023867183,0.0050512925,0.000015942223,0.0000014288686,0.0007575749,0.0019390302,0.006668041],"study_design_scores_gemma":[0.86794657,0.0029683602,0.0033664934,0.005350817,0.00036349372,0.000035326146,0.0010469558,0.118152805,0.0000025568438,0.0000100569105,0.0004507922,0.000305774],"about_ca_topic_score_codex":0.000118123884,"about_ca_topic_score_gemma":0.00005980349,"teacher_disagreement_score":0.11813687,"about_ca_system_score_codex":0.0006066091,"about_ca_system_score_gemma":0.0030374115,"threshold_uncertainty_score":0.99998313},"labels":[],"label_agreement":null},{"id":"W2973076393","doi":"10.3917/cips.121.0063","title":"L’ignorance et l’opinion, un couple heureux","year":2019,"lang":"fr","type":"article","venue":"Les cahiers internationaux de psychologie sociale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; Université de Montréal","funders":"","keywords":"Humanities; Ignorance; Philosophy; Epistemology","score_opus":0.029186396303886728,"score_gpt":0.3825385486476824,"score_spread":0.35335215234379563,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2973076393","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7175314,0.020545442,0.009163296,0.08188861,0.03370171,0.0008232911,0.00022880624,0.00029771848,0.13581975],"genre_scores_gemma":[0.89156437,0.010804695,0.0036761598,0.017547362,0.0024033925,0.000054045555,0.00020135441,0.00007986482,0.07366874],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.9971389,0.00027492258,0.0006244011,0.00069133565,0.00052580726,0.00074462936],"domain_scores_gemma":[0.9914425,0.0070608696,0.00031308897,0.00054308143,0.0003411405,0.00029929812],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0012671944,0.00032897602,0.00052410684,0.00011456244,0.00014764798,0.00008139482,0.0004959466,0.0008848192,0.0037290056],"category_scores_gemma":[0.024822893,0.0003521314,0.00035104374,0.00032963528,0.0007200377,0.00013497763,0.00007779407,0.0012603363,0.0026274514],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00038416582,0.0017779863,0.29924184,0.00014595367,0.00067042524,0.00015381999,0.0026542395,0.00028443322,0.00037304987,0.2332196,0.20578434,0.25531015],"study_design_scores_gemma":[0.010683832,0.0011273763,0.4355212,0.0056081684,0.0002443624,0.00017815808,0.0031829546,0.0013005852,0.00043433215,0.06411127,0.47655314,0.0010546419],"about_ca_topic_score_codex":0.00036052536,"about_ca_topic_score_gemma":0.000027275126,"teacher_disagreement_score":0.2707688,"about_ca_system_score_codex":0.00071316946,"about_ca_system_score_gemma":0.00042788233,"threshold_uncertainty_score":0.99989307},"labels":[],"label_agreement":null},{"id":"W2973721819","doi":"10.1007/s11096-019-00906-3","title":"Community pharmacists’ clinical reasoning: a protocol analysis","year":2019,"lang":"en","type":"article","venue":"International Journal of Clinical Pharmacy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Medicine; Community pharmacy; Protocol (science); Pharmacy; Clinical pharmacy; Family medicine; Alternative medicine; Pathology","score_opus":0.166141387282063,"score_gpt":0.5838873433814872,"score_spread":0.41774595609942416,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2973721819","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97506106,0.000031483236,0.0005192217,0.0038099345,0.0038433666,0.006458063,0.000018751116,0.000047172245,0.010210929],"genre_scores_gemma":[0.98722106,0.0001645912,0.002226355,0.0056526586,0.0029856118,0.00031854477,0.00002732235,0.000029471756,0.0013743908],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99239284,0.0012825717,0.0043633874,0.00030309058,0.0013853104,0.00027279905],"domain_scores_gemma":[0.97281337,0.021261169,0.0023839648,0.0005195757,0.0021380638,0.00088388287],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.010043962,0.00023570424,0.0015443349,0.00035564846,0.00005232508,0.00007494716,0.00092823984,0.00016383528,0.0034002543],"category_scores_gemma":[0.071019575,0.00017919231,0.0025898325,0.00044468348,0.00024017185,0.00022929716,0.00022284662,0.0029679425,0.0003717963],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.007116475,0.0040899687,0.8995499,0.00001630254,0.005690037,0.0004487688,0.00004593259,0.00005270101,0.00004876639,0.00007057666,0.013612531,0.06925801],"study_design_scores_gemma":[0.048687875,0.0022122185,0.732451,0.0011909378,0.0035917242,0.0007518572,0.0000644884,0.004801037,0.00040390706,0.0006013419,0.20487097,0.00037264067],"about_ca_topic_score_codex":0.00003598495,"about_ca_topic_score_gemma":0.0000015151596,"teacher_disagreement_score":0.19125843,"about_ca_system_score_codex":0.00007519083,"about_ca_system_score_gemma":0.00072395883,"threshold_uncertainty_score":0.99933225},"labels":[],"label_agreement":null},{"id":"W2975356529","doi":"10.1051/pmed/2019015","title":"Se réapproprier l’incertitude du marais : réintroduire la réflexivité dans l’enseignement du raisonnement clinique","year":2018,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Humanities; Philosophy","score_opus":0.03631660426821781,"score_gpt":0.3474952210744522,"score_spread":0.3111786168062344,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2975356529","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5033312,0.030566776,0.03279746,0.21780585,0.019307839,0.0037109505,0.0004725269,0.0011051723,0.19090223],"genre_scores_gemma":[0.88868797,0.012419368,0.006723057,0.013399523,0.024799848,0.00034779147,0.00020002501,0.00024126733,0.053181164],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99277633,0.0007506237,0.0017413532,0.001551373,0.0013866529,0.001793651],"domain_scores_gemma":[0.985205,0.010600524,0.00043526012,0.0015245803,0.0005664116,0.0016682267],"candidate_categories":["metaresearch","metaepi_narrow","sts","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0033224137,0.0008351696,0.0013001164,0.00015654185,0.00053232396,0.00017349752,0.0006816627,0.00084799505,0.014860106],"category_scores_gemma":[0.038628384,0.00075141,0.00064395904,0.00052818516,0.0031685694,0.00015529743,0.00068964047,0.001158567,0.004394573],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009428058,0.007504479,0.051507004,0.00032254463,0.0007589397,0.0014839447,0.0070555527,0.000026594536,0.0007099005,0.005646618,0.47382858,0.45021302],"study_design_scores_gemma":[0.006608019,0.0031047142,0.08810788,0.0047660465,0.0010786522,0.0002679822,0.0010903775,0.0024757197,0.0023920136,0.0016665746,0.88744205,0.0009999534],"about_ca_topic_score_codex":0.0008040441,"about_ca_topic_score_gemma":0.00029419825,"teacher_disagreement_score":0.44921306,"about_ca_system_score_codex":0.0006652975,"about_ca_system_score_gemma":0.0014027742,"threshold_uncertainty_score":0.99954426},"labels":[],"label_agreement":null},{"id":"W2976538850","doi":"10.1007/s10459-019-09916-2","title":"A view from the sender side of feedback: anticipated receptivity to clinical feedback when changing prior physicians’ clinical decisions—a mixed methods study","year":2019,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Communication source; Receptivity; Context (archaeology); Medicine; Credibility; Family medicine; Exploratory research; Psychology; Medical education; Social psychology; Internal medicine","score_opus":0.11598715787885458,"score_gpt":0.5528642506234847,"score_spread":0.43687709274463005,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2976538850","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97899103,0.0039076847,0.0016292492,0.007523289,0.005480564,0.0021487512,0.000006386043,0.000028215987,0.00028480266],"genre_scores_gemma":[0.9081525,0.0034471587,0.0813807,0.0062947604,0.0005154349,0.00007321586,0.000012669886,0.000014678238,0.00010886287],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9937225,0.0022784416,0.0019914508,0.0008844823,0.00060839247,0.0005147108],"domain_scores_gemma":[0.93398285,0.0638762,0.0008109695,0.0007459125,0.0002664397,0.00031760643],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.017662132,0.00019117263,0.0009926487,0.00016641467,0.00018754648,0.000030290164,0.00042305014,0.00010334058,0.00007236955],"category_scores_gemma":[0.07444201,0.00012893084,0.00015775746,0.0015988722,0.0003524558,0.00023783192,0.00015807601,0.00046584537,0.00007584061],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000084968764,0.0012678204,0.4447854,0.000017362136,0.000008664385,2.552812e-7,0.0015672086,0.00004638541,0.0000034318884,0.000035672103,0.0002798024,0.55190307],"study_design_scores_gemma":[0.0011529952,0.0019927996,0.9713576,0.0047166054,0.000043686883,0.000001254076,0.013545011,0.00060080737,0.000025526257,0.0015229055,0.0048910007,0.00014984071],"about_ca_topic_score_codex":0.0009687808,"about_ca_topic_score_gemma":0.00064325926,"teacher_disagreement_score":0.5517532,"about_ca_system_score_codex":0.000094011855,"about_ca_system_score_gemma":0.0015583951,"threshold_uncertainty_score":0.9333544},"labels":[],"label_agreement":null},{"id":"W2977053186","doi":"10.2196/16047","title":"Improving Team-Based Decision Making Using Data Analytics and Informatics: Protocol for a Collaborative Decision Support Design","year":2019,"lang":"en","type":"article","venue":"JMIR Research Protocols","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":32,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Western University of Health Sciences; Chapman University","keywords":"Protocol (science); Health informatics; Analytics; Informatics; Health care; Decision support system; Computer science; Knowledge management; Medicine; Data science; Artificial intelligence; Nursing; Public health; Engineering","score_opus":0.43725466349680525,"score_gpt":0.6149080005643458,"score_spread":0.17765333706754055,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2977053186","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0010401671,7.6781015e-7,0.14472504,0.000053723732,0.0000073839515,0.8538047,0.000059065893,0.000050114217,0.0002590093],"genre_scores_gemma":[0.00087825506,1.2132168e-7,0.28685156,0.000115390176,0.00008078659,0.7119439,0.000033314675,0.000038996182,0.000057655365],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99596786,0.00023967178,0.0009989658,0.0006249202,0.0014041094,0.0007644702],"domain_scores_gemma":[0.9653226,0.0311549,0.00032854525,0.0014704561,0.0013473755,0.00037612737],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.007047712,0.00026247208,0.0005887651,0.00043558306,0.00027348756,0.00036630026,0.0005260294,0.0002663609,0.0001581089],"category_scores_gemma":[0.06376204,0.00019884355,0.000075299446,0.0008543141,0.00019831337,0.00041969505,0.00065495184,0.0006105851,0.000048442183],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0616993,0.0033618093,0.054612998,0.006930416,0.00015647103,0.0001040709,0.00091163826,0.0016194122,0.0025001457,0.00025232582,0.06897846,0.79887295],"study_design_scores_gemma":[0.027141748,0.010641911,0.0017739987,0.057660937,0.00003121667,0.000034838602,0.0005766524,0.8537159,0.0012772797,0.0012550983,0.045417786,0.00047266038],"about_ca_topic_score_codex":0.000008748093,"about_ca_topic_score_gemma":0.0000043247906,"teacher_disagreement_score":0.85209644,"about_ca_system_score_codex":0.00021792411,"about_ca_system_score_gemma":0.0027155378,"threshold_uncertainty_score":0.9441243},"labels":[],"label_agreement":null},{"id":"W2977593000","doi":"10.1016/j.nepr.2019.102632","title":"Theoretical foundations of educational strategies used in e-learning environments for developing clinical reasoning in nursing students: A scoping review","year":2019,"lang":"en","type":"review","venue":"Nurse Education in Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":38,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre Hospitalier de l’Université de Montréal; Montreal Heart Institute; McGill University Health Centre; Université de Montréal; Université du Québec à Montréal","funders":"Fonds de Recherche du Québec - Santé; Fonds de Recherche du Québec-Société et Culture; Université de Montréal","keywords":"Perspective (graphical); Cognition; Psychology; Nurse education; Interpersonal communication; Nursing; Medicine; Computer science; Artificial intelligence; Social psychology","score_opus":0.09101381756252269,"score_gpt":0.5839391524621433,"score_spread":0.49292533489962065,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2977593000","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00041800184,0.9860715,0.00024342058,0.0026199135,0.001304137,0.0051314333,0.0000041759695,0.00000975076,0.004197634],"genre_scores_gemma":[0.0007918817,0.9868548,0.01005175,0.00030675204,0.00026379098,0.00096178136,0.000487811,0.000058987065,0.00022245584],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9943753,0.001304104,0.002756211,0.00070624775,0.00050565443,0.00035249503],"domain_scores_gemma":[0.93000853,0.067723714,0.0015410557,0.00044817268,0.00015901944,0.00011950133],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.004311999,0.00033702617,0.0020394777,0.0004691263,0.00004600065,0.00006436807,0.00025941886,0.00039738405,0.00017370803],"category_scores_gemma":[0.19677052,0.00033145127,0.00028759523,0.00071095285,0.00022253812,0.00037909925,0.000034324778,0.0012211923,0.000047387053],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001472892,0.008200332,0.026758594,0.047864474,0.00014193336,0.000004251556,0.00081830594,0.000018566865,1.2494934e-7,0.12470419,0.00026667753,0.7910753],"study_design_scores_gemma":[0.0011580276,0.00014726871,0.0032114224,0.9732015,0.0012845775,0.0000439488,0.0024211828,0.000021523538,1.09748065e-7,0.00090446835,0.017305125,0.00030084266],"about_ca_topic_score_codex":0.000030161978,"about_ca_topic_score_gemma":0.000013353366,"teacher_disagreement_score":0.925337,"about_ca_system_score_codex":0.0008896669,"about_ca_system_score_gemma":0.013687651,"threshold_uncertainty_score":0.99991375},"labels":[],"label_agreement":null},{"id":"W2980463893","doi":"","title":"Audit to investigate junior doctors' knowledge of how to administer and score the Montreal Cognitive Assessment (MoCA).","year":2018,"lang":"en","type":"article","venue":"PubMed","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Montreal Cognitive Assessment; Medicine; Session (web analytics); Test (biology); Audit; Cognition; Medical education; Family medicine; Cognitive impairment; Psychiatry","score_opus":0.05267143783826358,"score_gpt":0.3411474492699295,"score_spread":0.28847601143166596,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2980463893","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9671758,0.00009353228,0.00007172442,0.018394204,0.00033151137,0.0013611851,0.000034921657,0.00002915645,0.012507957],"genre_scores_gemma":[0.9908945,0.000012407803,0.0002306172,0.0035497607,0.00050394767,0.00069392595,0.000008923676,0.00001670938,0.0040892255],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9988986,0.0000733117,0.00022579964,0.00029757805,0.0001831472,0.00032153295],"domain_scores_gemma":[0.99464965,0.004008233,0.00007531902,0.00026866604,0.00027900175,0.00071911694],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00055211224,0.00013863105,0.0003114089,0.00007108779,0.000060446975,0.000034692468,0.00009455688,0.00006718237,0.000028654096],"category_scores_gemma":[0.02629886,0.00008702923,0.000056414472,0.00023445259,0.00032785806,0.00002844136,0.0001425494,0.00013010447,0.000034673714],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00071119965,0.0006304531,0.20094046,0.00007113174,0.00025883483,0.00004285296,0.0033015066,2.2725837e-7,0.00006318004,0.0004038136,0.07285238,0.720724],"study_design_scores_gemma":[0.0012332272,0.00061216427,0.99033046,0.00048630586,0.00018629835,0.0000137942125,0.00022399321,0.000010777108,0.00096902903,0.00010713033,0.0057210466,0.000105802974],"about_ca_topic_score_codex":0.00009808288,"about_ca_topic_score_gemma":0.0002778484,"teacher_disagreement_score":0.78938997,"about_ca_system_score_codex":0.000035148518,"about_ca_system_score_gemma":0.00012911642,"threshold_uncertainty_score":0.981903},"labels":[],"label_agreement":null},{"id":"W2980600521","doi":"10.1051/pmed/2019019","title":"Concevoir une formation par concordance pour développer le raisonnement professionnel : quelles étapes faut-il parcourir ?","year":2018,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Université de Sherbrooke; Université de Montréal","funders":"","keywords":"Humanities; Philosophy","score_opus":0.07166401914607201,"score_gpt":0.3706747281902101,"score_spread":0.29901070904413807,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2980600521","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6695107,0.04572566,0.019216308,0.23698209,0.009417263,0.0018748522,0.00018528398,0.0004890455,0.016598748],"genre_scores_gemma":[0.9260054,0.0066494467,0.0047679655,0.005778127,0.0044510346,0.00018665618,0.00014591783,0.00009153671,0.051923893],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99504423,0.0004271079,0.0013935125,0.00083592086,0.00090465276,0.0013945469],"domain_scores_gemma":[0.9891062,0.008030892,0.0005028814,0.0008030965,0.0005884659,0.0009684589],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0013994863,0.0005696735,0.0009729157,0.000107110405,0.00068587455,0.000089587214,0.00043367178,0.00056517235,0.0026333844],"category_scores_gemma":[0.026650852,0.0004926339,0.00031438243,0.00051316863,0.0018324198,0.00031309677,0.00029002276,0.0007671159,0.0037325898],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006481567,0.0036634214,0.030221641,0.00061212527,0.00030314765,0.00024878158,0.009578188,0.000027711454,0.0025819554,0.003971863,0.5848954,0.36324763],"study_design_scores_gemma":[0.014338926,0.002384533,0.055576887,0.02399965,0.00065161445,0.00013115582,0.022600973,0.009895615,0.026715258,0.007212432,0.8349887,0.0015042234],"about_ca_topic_score_codex":0.000629549,"about_ca_topic_score_gemma":0.00020926847,"teacher_disagreement_score":0.3617434,"about_ca_system_score_codex":0.00026681257,"about_ca_system_score_gemma":0.0016714776,"threshold_uncertainty_score":0.9997525},"labels":[],"label_agreement":null},{"id":"W2980714529","doi":"10.30476/jamp.2019.83101.1083","title":"Combination of different clinical reasoning tests in a national exam.","year":2019,"lang":"en","type":"article","venue":"PubMed","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Olympiad; Cronbach's alpha; Test (biology); Shahid; Reliability (semiconductor); Concordance; Medical education; Descriptive statistics; Medicine; Psychology; Mathematics education; Statistics; Mathematics; Internal medicine","score_opus":0.051523777251230814,"score_gpt":0.3506013692919797,"score_spread":0.2990775920407489,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2980714529","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9862326,0.000052256823,0.0000061569394,0.0006596931,0.0002809167,0.000572421,0.0000023654313,0.000020628642,0.012173003],"genre_scores_gemma":[0.99855083,0.00002897622,0.00009544662,0.00027953513,0.00007497081,0.00015576722,0.000024820642,0.000008457451,0.00078119204],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9987387,0.00006491482,0.00047494192,0.00019445874,0.0003518489,0.00017511728],"domain_scores_gemma":[0.99198353,0.0074978354,0.0001251743,0.00014611412,0.00012203646,0.00012528525],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010169962,0.000070917355,0.0003158332,0.00009009568,0.000006776077,0.0000053815816,0.000052704392,0.000096209755,0.000049918217],"category_scores_gemma":[0.054894697,0.00005749679,0.000088089975,0.00014661103,0.00004207504,0.00002879236,0.000030066216,0.00018379022,0.000025174815],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009646461,0.00057042576,0.94928485,0.0000146772,0.000015586596,0.0000035782382,0.00001732536,0.0000018042176,0.000006932817,0.0017622581,0.00032252254,0.047903575],"study_design_scores_gemma":[0.0033865594,0.00011167643,0.9947751,0.00034154323,0.000017475353,0.0000044129615,0.000010310804,0.0002956622,0.00007392002,0.00083307683,0.00009529918,0.00005497594],"about_ca_topic_score_codex":0.000020333458,"about_ca_topic_score_gemma":0.000008377494,"teacher_disagreement_score":0.0538777,"about_ca_system_score_codex":0.00006131058,"about_ca_system_score_gemma":0.000056895897,"threshold_uncertainty_score":0.95306635},"labels":[],"label_agreement":null},{"id":"W2981241231","doi":"10.4236/health.2019.1110103","title":"Understanding and Promoting Clinical Reasoning in Chronic and Multimorbid Conditions: A Call for GPs and Healthcare Professionals","year":2019,"lang":"en","type":"article","venue":"Health","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Montfort Hospital; Université de Montréal","funders":"","keywords":"Health professionals; Action (physics); Health care; Multimorbidity; Medicine; Nursing; Psychology; Chronic disease; Family medicine","score_opus":0.16831872256946703,"score_gpt":0.49028889101684375,"score_spread":0.3219701684473767,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2981241231","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97065437,0.0045052306,0.00014131436,0.022618188,0.00014427147,0.0018230983,0.000024539959,0.000034841778,0.00005417387],"genre_scores_gemma":[0.9927577,0.0017834905,0.0019512677,0.0031575407,0.00012743454,0.00005454268,0.000046631718,0.000019314664,0.000102125654],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9983924,0.00014995609,0.0005626385,0.00041513122,0.00012577185,0.00035410974],"domain_scores_gemma":[0.9915303,0.007720656,0.00016990965,0.00013125388,0.000041776842,0.00040610076],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0018584435,0.00011629781,0.0005038382,0.00007291282,0.0001362781,0.000016992799,0.000020731013,0.00014108885,0.000009827745],"category_scores_gemma":[0.008751933,0.00009820372,0.000032675867,0.00008174064,0.00009526246,0.00004771495,0.000043263026,0.00034839823,0.0000020614395],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021438464,0.00019374158,0.97762185,0.0021100629,0.000024045918,0.000013000493,0.0012648812,0.0000012252167,0.000019321844,0.004502351,0.0011160076,0.012919128],"study_design_scores_gemma":[0.01182175,0.0034224591,0.93437755,0.02899856,0.0000349649,0.00013467541,0.002009714,0.016329862,0.000006443925,0.002116187,0.0005079878,0.0002398396],"about_ca_topic_score_codex":0.00024710066,"about_ca_topic_score_gemma":0.00026603142,"teacher_disagreement_score":0.04324429,"about_ca_system_score_codex":0.00024333505,"about_ca_system_score_gemma":0.0007240682,"threshold_uncertainty_score":0.9995978},"labels":[],"label_agreement":null},{"id":"W2981877078","doi":"10.1097/acm.0000000000002797","title":"In Reply to Eichbaum","year":2019,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Impact; McMaster University","funders":"","keywords":"Debiasing; Hindsight bias; Psychology; Cognitive bias; Perspective (graphical); Confirmation bias; Experiential learning; Metacognition; Malpractice; Cognition; Cognitive psychology; Overconfidence effect; Credibility; Social psychology; Epistemology; Computer science; Psychiatry; Pedagogy; Political science","score_opus":0.04798426041355285,"score_gpt":0.3901999257639128,"score_spread":0.34221566535035997,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2981877078","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00449486,0.0014149297,0.000026506468,0.97578424,0.002747384,0.0010446411,0.000011461041,0.00008725147,0.01438873],"genre_scores_gemma":[0.001360868,0.0007916278,0.0000699827,0.86208534,0.020469032,0.000059551137,0.0003807323,0.000087738765,0.11469514],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9959256,0.00011137148,0.0012160884,0.00095556414,0.0010175002,0.00077385677],"domain_scores_gemma":[0.9825046,0.015780378,0.00024649533,0.001031438,0.00011302846,0.0003240263],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity","insufficient_payload"],"category_scores_codex":[0.00127863,0.00045708966,0.0016456556,0.0005773068,0.000018292523,0.0000035711826,0.00039588555,0.0020207257,0.0016440937],"category_scores_gemma":[0.11108399,0.00033679916,0.00012508943,0.0005749164,0.00016047074,0.00003695026,0.00012955896,0.008356788,0.002840229],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000782574,0.000007578985,0.0029222001,0.00015203854,0.000037998343,0.0016895203,0.00044269185,5.8062506e-7,0.000048186113,0.000022597673,0.9936775,0.0009208334],"study_design_scores_gemma":[0.0026025479,0.00072754204,0.0015334834,0.017883163,0.00023206115,0.00024514837,0.000052306277,0.000008999387,0.000022207898,0.00036248897,0.9760242,0.0003058293],"about_ca_topic_score_codex":0.0005414513,"about_ca_topic_score_gemma":0.0000017511306,"teacher_disagreement_score":0.11369891,"about_ca_system_score_codex":0.00020901716,"about_ca_system_score_gemma":0.0002696026,"threshold_uncertainty_score":0.9999084},"labels":[],"label_agreement":null},{"id":"W2981940088","doi":"10.1007/s12186-019-09234-z","title":"Investigating the Role of Cognitive Feedback in Practice-Oriented Learning for Clinical Diagnostics","year":2019,"lang":"en","type":"article","venue":"Vocations and Learning","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria","funders":"","keywords":"Cognition; Task (project management); Set (abstract data type); Competence (human resources); Computer science; Process (computing); Psychology; Key (lock); Cognitive psychology; Social psychology","score_opus":0.03341280975573986,"score_gpt":0.38909105261910587,"score_spread":0.355678242863366,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2981940088","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9932999,0.0014751669,0.00033837103,0.0011208201,0.00007918311,0.0005026062,0.0000017515016,0.00002128873,0.0031609088],"genre_scores_gemma":[0.99607813,0.0013476227,0.0016914112,0.0003182443,0.00009829182,0.00003843019,0.000050784394,0.000014339161,0.00036273137],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9988325,0.0002548503,0.0004299289,0.00020360107,0.00012766066,0.0001514082],"domain_scores_gemma":[0.8944576,0.1047324,0.00028357905,0.00011264508,0.00031921561,0.0000945613],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0015831995,0.00007988551,0.0002401104,0.000053695105,0.0001266522,0.000016059606,0.00003479426,0.00008874865,0.000017857585],"category_scores_gemma":[0.38013026,0.00006021295,0.00006116921,0.00022804606,0.00013519956,0.000053959324,0.000035703575,0.00069858355,0.00001621556],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000049580747,0.00015589411,0.9448233,0.000015816864,0.0000385308,5.9792205e-7,0.0018244674,0.00015275874,0.000024898196,0.0011701076,0.000040015995,0.051704008],"study_design_scores_gemma":[0.0030120588,0.0011266201,0.9200865,0.0029885862,0.0002851886,0.0000141260825,0.03413394,0.016967677,0.00005896859,0.00018160747,0.021013374,0.00013135443],"about_ca_topic_score_codex":0.000090614674,"about_ca_topic_score_gemma":0.0000059506992,"teacher_disagreement_score":0.37854707,"about_ca_system_score_codex":0.000010254058,"about_ca_system_score_gemma":0.00012216934,"threshold_uncertainty_score":0.62509114},"labels":[],"label_agreement":null},{"id":"W2982372929","doi":"10.1002/aet2.10404","title":"Emergency Medicine Education Research Since the 2012 Consensus Conference: How Far Have We Come and What’s Next?","year":2019,"lang":"en","type":"article","venue":"AEM Education and Training","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Certification; Medical education; Task force; Political science; State (computer science); Consensus conference; Medicine; Engineering ethics; Public relations; Engineering; Computer science; Public administration","score_opus":0.1866956348707654,"score_gpt":0.4558083382260652,"score_spread":0.2691127033552998,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2982372929","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7790031,0.022663971,0.000008682723,0.18823119,0.0021033653,0.00049244304,0.0000010981142,0.000023565939,0.0074726134],"genre_scores_gemma":[0.9692278,0.01606207,0.00014457086,0.0010684432,0.00048504543,0.000040151677,0.00003812763,0.000013780528,0.0129200425],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99872935,0.00015527087,0.00026039197,0.00030674215,0.0003007925,0.0002474286],"domain_scores_gemma":[0.99470407,0.0041799876,0.00010395784,0.00029305567,0.00045306305,0.00026584862],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0011516449,0.00012414683,0.00024138136,0.00011373373,0.0001876221,0.00012768348,0.000072387345,0.00009346528,0.0005998301],"category_scores_gemma":[0.0077523906,0.00008442073,0.000028201372,0.00018817691,0.00034767244,0.0002377033,0.000030320616,0.00036339837,0.000033936634],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000054077365,0.00032423274,0.028451102,0.00009092634,0.00004372078,0.0000010644887,0.020298349,2.1827796e-7,0.00035642472,0.020075593,0.039523695,0.89078057],"study_design_scores_gemma":[0.0016624791,0.0006702109,0.39501768,0.004811264,0.00015572848,0.00019349465,0.36936814,0.00039487626,0.000031299973,0.0060172365,0.22140163,0.00027596025],"about_ca_topic_score_codex":0.00015989097,"about_ca_topic_score_gemma":0.000030210698,"teacher_disagreement_score":0.89050466,"about_ca_system_score_codex":0.000026276299,"about_ca_system_score_gemma":0.0009018884,"threshold_uncertainty_score":0.92808956},"labels":[],"label_agreement":null},{"id":"W2982844391","doi":"10.1007/s10459-019-09939-9","title":"The compatibility principle: on philosophies in the assessment of clinical competence","year":2019,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":39,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Women's College Hospital; Sunnybrook Health Science Centre; The Wilson Centre; University Health Network; University of Toronto; Health Sciences Centre","funders":"","keywords":"Epistemology; Competence (human resources); Philosophical methodology; Commit; Philosophy of science; Engineering ethics; Philosophical analysis; Psychology; Computer science; Sociology; Social psychology; Philosophy","score_opus":0.080111345698781,"score_gpt":0.5569840745406762,"score_spread":0.47687272884189524,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2982844391","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9676006,0.0014498078,0.000009894323,0.02166501,0.001230682,0.0007513183,0.0000010637639,0.000005279442,0.007286347],"genre_scores_gemma":[0.992511,0.003169359,0.0015885648,0.0025793337,0.00009180363,0.000034462442,0.0000032769656,0.0000018482056,0.00002035038],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.997779,0.00039767486,0.00088331336,0.00028071154,0.00046551035,0.00019379593],"domain_scores_gemma":[0.9729691,0.026104493,0.00039444191,0.00039515668,0.000072009054,0.00006477745],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.007856621,0.00006569357,0.00026669193,0.000055833836,0.000108534914,0.000012260758,0.00027254404,0.000028696892,0.000010597164],"category_scores_gemma":[0.013544132,0.00003326833,0.000039466267,0.00054307847,0.0005431163,0.00009275243,0.000021427546,0.00028302707,0.000005850931],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000021172027,0.0005424851,0.924954,0.000047109228,5.3392426e-7,1.0009428e-7,0.00016463929,0.00036637578,2.843925e-7,0.034572236,0.000028019578,0.03930306],"study_design_scores_gemma":[0.00030199706,0.0009530921,0.9846107,0.001181034,0.0000014203479,0.0000014986138,0.0010181271,0.0016689196,0.000001909967,0.0074552544,0.0027726009,0.00003349507],"about_ca_topic_score_codex":0.00019824192,"about_ca_topic_score_gemma":0.00026383725,"teacher_disagreement_score":0.05965667,"about_ca_system_score_codex":0.00007371573,"about_ca_system_score_gemma":0.0017441469,"threshold_uncertainty_score":0.9947652},"labels":[],"label_agreement":null},{"id":"W2983799067","doi":"10.1016/j.jpurol.2019.10.022","title":"Further medical experience will be required to validate these results: How experience -based medicine shapes the validity of medical evidence","year":2019,"lang":"en","type":"article","venue":"Journal of Pediatric Urology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"SickKids Foundation; Hospital for Sick Children; University of Toronto","funders":"Novo Nordisk Fonden","keywords":"Medicine; Evidence-based medicine; Critical appraisal; Alternative medicine; MEDLINE; Medical education; Pathology","score_opus":0.09773452569652415,"score_gpt":0.3995022030945357,"score_spread":0.3017676773980116,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2983799067","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7666059,0.0013388882,0.0010924643,0.22953846,0.0010475848,0.00026152103,0.0000062797244,0.000018399662,0.00009050209],"genre_scores_gemma":[0.9823686,0.0024393762,0.00030543114,0.012002936,0.0027521774,0.00001416385,0.0000024300793,0.000024344456,0.00009053333],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9929172,0.00071323966,0.0016530097,0.0004664531,0.0037273655,0.000522731],"domain_scores_gemma":[0.9541929,0.04200554,0.0011971431,0.00071270566,0.0007311258,0.0011606106],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0070757526,0.00028222776,0.0010722153,0.00027735817,0.00005140184,0.000017402876,0.0012826204,0.00048576793,0.002050891],"category_scores_gemma":[0.33761472,0.00015306266,0.00029160368,0.0007330354,0.0007670803,0.00018337516,0.00021753227,0.0009891309,0.000027676793],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.023128903,0.004603487,0.7623223,0.00031860816,0.00079291296,0.0062809587,0.060895648,0.0003377976,0.003018762,0.00071736565,0.11860162,0.018981617],"study_design_scores_gemma":[0.09050194,0.1012695,0.6017333,0.009924765,0.0077654473,0.010594338,0.019710064,0.015881531,0.0035883023,0.0015944691,0.13433258,0.003103715],"about_ca_topic_score_codex":0.00007122928,"about_ca_topic_score_gemma":0.0000102066315,"teacher_disagreement_score":0.33053896,"about_ca_system_score_codex":0.000053893276,"about_ca_system_score_gemma":0.0012172763,"threshold_uncertainty_score":0.9988614},"labels":[],"label_agreement":null},{"id":"W2984557443","doi":"10.1097/nne.0000000000000755","title":"Usefulness of Abductive Reasoning in Nursing Education","year":2019,"lang":"en","type":"article","venue":"Nurse Educator","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Thompson Rivers University","funders":"","keywords":"Abductive reasoning; Psychology; Deductive reasoning; Medical education; Medicine; Computer science; Artificial intelligence","score_opus":0.016388076526622226,"score_gpt":0.36934687639202396,"score_spread":0.35295879986540174,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2984557443","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9736419,0.00033037414,0.000008744089,0.0009810345,0.0025447397,0.00033427638,0.0000020117668,0.000021294982,0.022135643],"genre_scores_gemma":[0.99645215,0.000009330338,0.0013535226,0.00018598275,0.0003254962,0.000025492704,0.000018103778,0.00002013319,0.0016097734],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99904263,0.000038234033,0.0002855401,0.00025533943,0.00018292399,0.00019534411],"domain_scores_gemma":[0.99808943,0.0012081654,0.00011066548,0.0003241705,0.00012053189,0.00014700522],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00020523358,0.00010463841,0.00031626463,0.00015236984,0.000014041052,0.0000066483904,0.00006529549,0.00009759384,0.0003962349],"category_scores_gemma":[0.007097871,0.00009721398,0.00007274692,0.00029795605,0.00008101063,0.00005765454,0.000005278485,0.00020040321,0.00014485096],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00024244867,0.004484813,0.8930339,0.00006500589,0.000021896863,0.0000028510162,0.004610356,0.0000063695975,0.00081401336,0.011505782,0.0052974666,0.079915054],"study_design_scores_gemma":[0.0013861785,0.00026604178,0.9654607,0.014021347,0.0000840534,0.000047321264,0.012322471,0.000044571065,0.004222317,0.0015206981,0.00044766723,0.00017662044],"about_ca_topic_score_codex":0.00011939868,"about_ca_topic_score_gemma":0.000002118058,"teacher_disagreement_score":0.07973843,"about_ca_system_score_codex":0.00015514018,"about_ca_system_score_gemma":0.0011107604,"threshold_uncertainty_score":0.8497327},"labels":[],"label_agreement":null},{"id":"W2984758147","doi":"10.1002/aet2.10408","title":"Emergency Medicine Oral Case Presentations: Evaluation of a Novel Curriculum","year":2019,"lang":"en","type":"article","venue":"AEM Education and Training","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Research Manitoba; Manitoba Health; Children's Hospital Research Institute of Manitoba; University of Manitoba","funders":"","keywords":"Curriculum; Medicine; Randomized controlled trial; Test (biology); Intervention (counseling); Presentation (obstetrics); Medical education; Focus group; Physical therapy; Nursing; Psychology; Internal medicine; Pedagogy; Surgery","score_opus":0.14851521891832206,"score_gpt":0.4779076363776144,"score_spread":0.3293924174592924,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2984758147","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98220617,0.0007289936,0.00009478407,0.001187969,0.0008325388,0.0003665154,0.0000019340932,0.000015098551,0.014566014],"genre_scores_gemma":[0.9978607,0.000072863586,0.00093994715,0.0001245885,0.00015271021,0.000033583598,0.000056646568,0.000007404123,0.00075151795],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9991261,0.00003943892,0.00031779113,0.00016223919,0.00026456753,0.00008988086],"domain_scores_gemma":[0.99858195,0.0006381205,0.00012450285,0.00013949067,0.00040612655,0.00010980492],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0007613661,0.00006878987,0.00017710081,0.00008318722,0.000032610147,0.000002772196,0.000020558342,0.00004154806,0.0013685882],"category_scores_gemma":[0.011123548,0.000055900924,0.0000335921,0.00018939287,0.00004135343,0.000053204953,0.000007762878,0.00007238208,0.000009787265],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003630812,0.0021749632,0.47536585,0.00014970267,0.00013165994,0.00000576274,0.016790193,0.000051725725,0.00066528533,0.0072614267,0.009691664,0.48767546],"study_design_scores_gemma":[0.012816683,0.0010335285,0.86220694,0.0030174355,0.001781186,0.0020902874,0.07520296,0.03681177,0.0000974191,0.0020081291,0.0025651455,0.0003685039],"about_ca_topic_score_codex":0.0003877863,"about_ca_topic_score_gemma":0.000017465278,"teacher_disagreement_score":0.48730695,"about_ca_system_score_codex":0.000013795631,"about_ca_system_score_gemma":0.00030890928,"threshold_uncertainty_score":0.9995443},"labels":[],"label_agreement":null},{"id":"W2985262800","doi":"10.1016/j.tcr.2019.100255","title":"“Morbidity and Mortality”: A new section in Trauma Case Reports to help us learn from our mistakes","year":2019,"lang":"en","type":"editorial","venue":"Trauma Case Reports","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Calgary","funders":"","keywords":"Section (typography); Medicine; Forensic engineering; Intensive care medicine; Engineering; Business; Advertising","score_opus":0.04814602006902247,"score_gpt":0.36351113801600604,"score_spread":0.3153651179469836,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2985262800","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.68046474,0.00064643024,0.000013804461,0.000164662,0.31663576,0.0010019498,0.00013820318,0.0001370088,0.0007974615],"genre_scores_gemma":[0.65765786,0.000063846564,0.00024358377,0.00010147218,0.3371348,0.000055703265,0.0009679403,0.00012600169,0.0036487672],"study_design_codex":"case_report","study_design_gemma":"case_report","domain_scores_codex":[0.993303,0.00018592911,0.0023976087,0.0021072158,0.0013027054,0.0007035358],"domain_scores_gemma":[0.99021256,0.0050708,0.0012062311,0.0018127777,0.00034705744,0.0013505773],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0018096523,0.0007823564,0.0020352742,0.00032988755,0.00012167752,0.00015344478,0.00004495605,0.0018502452,0.00015210421],"category_scores_gemma":[0.065847084,0.00075212563,0.0004305889,0.00036857638,0.000063503605,0.000116145195,0.000070354465,0.0019999256,0.000024151343],"study_design_candidate":"case_report","study_design_consensus":"case_report","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014158078,0.00029756423,0.010727464,0.00012398807,0.00016742891,0.5887058,0.00036455703,0.0000200602,0.000008540331,1.597021e-7,0.39275858,0.006684319],"study_design_scores_gemma":[0.0025580262,0.0009170393,0.025201024,0.004776723,0.0015641393,0.7059198,0.00082749076,0.000025192137,0.00012561453,0.00010927585,0.25693756,0.0010381276],"about_ca_topic_score_codex":0.16190259,"about_ca_topic_score_gemma":0.030206919,"teacher_disagreement_score":0.13582103,"about_ca_system_score_codex":0.0004446979,"about_ca_system_score_gemma":0.0013359897,"threshold_uncertainty_score":0.999493},"labels":[],"label_agreement":null},{"id":"W2987182583","doi":"10.1182/blood-2019-123284","title":"Exploring the Basis and Acceptability of Practice Variation Among Thrombosis Medicine Specialists: A Mixed-Methods Study","year":2019,"lang":"en","type":"article","venue":"Blood","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Ottawa Hospital; University of Calgary","funders":"","keywords":"Medicine; Thrombosis; Subspecialty; Venous thrombosis; Pulmonary embolism; Family medicine; Deep vein; Intensive care medicine; Surgery","score_opus":0.07118699647964441,"score_gpt":0.40732098251271126,"score_spread":0.3361339860330669,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2987182583","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9929043,0.00012773402,0.00015111368,0.0016726971,0.00042143377,0.00071225344,0.0000024759256,0.00002890801,0.00397909],"genre_scores_gemma":[0.9946687,0.00018805944,0.0045713265,0.00016588732,0.0003109802,0.000045412482,0.0000029844955,0.00001303886,0.000033573066],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982581,0.00050870306,0.000446991,0.00029992146,0.00034716766,0.00013909694],"domain_scores_gemma":[0.9716314,0.027312638,0.00022298393,0.00054480205,0.00018803425,0.00010011609],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0034116362,0.000117661104,0.00044424622,0.00004644827,0.000050894265,0.000012936209,0.00008802996,0.000047408626,0.0003720673],"category_scores_gemma":[0.110042654,0.00007206925,0.000059443064,0.00027283205,0.00014199966,0.00016136034,0.00008568892,0.00020018124,0.0000068812706],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000849768,0.0024241353,0.9523748,0.00004258914,0.00030748363,0.0000044769117,0.0072079846,0.000001115398,0.00035032525,0.00019098431,0.000072283736,0.03693882],"study_design_scores_gemma":[0.0032326025,0.000945145,0.9847175,0.00021907156,0.0013100373,0.000007821228,0.008563883,0.000015146366,0.000658388,0.00009456053,0.0001675118,0.000068365436],"about_ca_topic_score_codex":0.0011409404,"about_ca_topic_score_gemma":0.00003393959,"teacher_disagreement_score":0.10663102,"about_ca_system_score_codex":0.000017594539,"about_ca_system_score_gemma":0.000041905318,"threshold_uncertainty_score":0.89745384},"labels":[],"label_agreement":null},{"id":"W2989702233","doi":"10.1007/s10459-019-09932-2","title":"Scoping reviews in health professions education: challenges, considerations and lessons learned about epistemology and methodology","year":2019,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":128,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Université de Montréal; Centre for Interdisciplinary Research in Rehabilitation; McGill University Health Centre","funders":"","keywords":"Epistemology; Engineering ethics; Meaning (existential); Scholarship; Foundation (evidence); Perspective (graphical); Rationality; Sociology; Systematic review; Management science; Psychology; MEDLINE; Computer science; Political science; Philosophy","score_opus":0.3039440632008207,"score_gpt":0.5800463904718719,"score_spread":0.2761023272710512,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2989702233","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.09665046,0.6065205,0.00010454585,0.2906604,0.0018704152,0.0023467953,0.0000013938923,0.000019848772,0.0018256331],"genre_scores_gemma":[0.4183323,0.52701616,0.04343668,0.010414164,0.0001225563,0.0003321186,0.000013752362,0.000007973753,0.00032429874],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99718297,0.0010122849,0.0008226661,0.00053545245,0.00011835931,0.00032826528],"domain_scores_gemma":[0.99025196,0.00882446,0.00041085677,0.00021806563,0.00006130487,0.00023333653],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004962843,0.00011149413,0.0005593604,0.00026243483,0.00020877292,0.000014822105,0.000055088232,0.00007638158,0.000037597933],"category_scores_gemma":[0.020115126,0.00009448234,0.000017571887,0.00036330434,0.00030938495,0.00021025533,0.00003246282,0.00026344802,0.000007974092],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000016520995,0.00040476143,0.16832697,0.0027593377,0.000001426059,3.9569065e-7,0.004019069,0.00002960049,0.000005240381,0.10515804,0.00022341954,0.71905524],"study_design_scores_gemma":[0.0017838535,0.0019428387,0.69456035,0.11206319,0.0000128938855,0.00031039864,0.01778791,0.00033800985,0.000008239496,0.13413607,0.03663616,0.00042009618],"about_ca_topic_score_codex":0.00089256314,"about_ca_topic_score_gemma":0.0022187987,"teacher_disagreement_score":0.71863514,"about_ca_system_score_codex":0.00013007213,"about_ca_system_score_gemma":0.0052878037,"threshold_uncertainty_score":0.98813885},"labels":[],"label_agreement":null},{"id":"W2990732773","doi":"10.1016/j.annemergmed.2019.09.002","title":"Becoming Less Wrong (and More Rational) in Clinical Decisionmaking","year":2019,"lang":"en","type":"letter","venue":"Annals of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Medicine; Emergency physician; Core competency; Emergency department; Core (optical fiber); MEDLINE; Medical emergency; Psychiatry; Management","score_opus":0.33222665799568124,"score_gpt":0.5152843948473019,"score_spread":0.1830577368516207,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2990732773","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.16053681,0.0072559575,0.000070449096,0.8244701,0.003959206,0.0005609841,0.00002382003,0.000024257095,0.0030984227],"genre_scores_gemma":[0.08750326,0.034503017,0.00046410135,0.84481424,0.018967971,0.000045513698,0.0012437305,0.0001703983,0.012287787],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.994931,0.00018155044,0.0027909663,0.00071992027,0.0009170114,0.00045958423],"domain_scores_gemma":[0.97944146,0.01843731,0.00078827934,0.0006955861,0.0004603074,0.00017707628],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0025151018,0.00038046396,0.0018133805,0.00039432995,0.000035965597,0.0000027826027,0.0002225341,0.00093273283,0.0014807642],"category_scores_gemma":[0.09331072,0.0002908629,0.00037851464,0.00033576673,0.00026470472,0.000056198543,0.000118502736,0.0019526168,0.000031515963],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000105361614,0.00012156098,0.115490206,0.00026211186,0.00012687738,0.00020067628,0.00019299309,0.000004109993,0.00000521155,0.000042410906,0.87372386,0.009724641],"study_design_scores_gemma":[0.0052990164,0.0017478577,0.4212234,0.035684977,0.0006115371,0.000051205345,0.00044693923,0.000536628,0.00002199509,0.0018597319,0.53181636,0.0007003782],"about_ca_topic_score_codex":0.00030720074,"about_ca_topic_score_gemma":0.000028906165,"teacher_disagreement_score":0.3419075,"about_ca_system_score_codex":0.000015995061,"about_ca_system_score_gemma":0.00020762139,"threshold_uncertainty_score":0.99995434},"labels":[],"label_agreement":null},{"id":"W2991213201","doi":"10.1503/cmaj.190506","title":"Deconstructing the diagnostic reasoning of human versus artificial intelligence","year":2019,"lang":"en","type":"article","venue":"Canadian Medical Association Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":33,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Computer science; Artificial intelligence; Artificial intelligence, situated approach; Key (lock); Human intelligence; Artificial psychology; Model-based reasoning; Applications of artificial intelligence; Marketing and artificial intelligence; Data science; Cognitive science; Artificial Intelligence System; Intelligent decision support system; Knowledge representation and reasoning; Psychology; Computer security","score_opus":0.024520645660322634,"score_gpt":0.32537056434059547,"score_spread":0.30084991868027283,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2991213201","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9804941,0.00014828437,0.00006807574,0.011158476,0.0017045593,0.00013915356,0.0000046972373,0.000011313208,0.0062713837],"genre_scores_gemma":[0.99702764,0.00005782196,0.00015516528,0.0015151273,0.0011151336,0.000002399639,0.0000055439687,0.000011226628,0.00010994606],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99773204,0.00012701491,0.0006260646,0.00013736838,0.0010079518,0.00036954667],"domain_scores_gemma":[0.9790316,0.019065488,0.0003687565,0.00016668793,0.0003030894,0.0010643298],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0023324457,0.00009392559,0.00027871638,0.000094958436,0.00020424943,0.000045018973,0.00017771988,0.00025171865,0.0036177938],"category_scores_gemma":[0.21435264,0.00006882831,0.00014610616,0.00023175387,0.00008662648,0.00004343781,0.00001778684,0.0010000955,0.00020371504],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006250634,0.00007692799,0.72101283,0.000012499202,0.0002544823,0.00029059924,0.0005197933,0.000021396001,0.00003689453,0.009951167,0.006184333,0.26157656],"study_design_scores_gemma":[0.00889466,0.0036599059,0.8892088,0.015026835,0.001264524,0.0037449424,0.020605035,0.0055506597,0.0017216472,0.01371214,0.035256203,0.0013546477],"about_ca_topic_score_codex":0.0017176132,"about_ca_topic_score_gemma":0.004860428,"teacher_disagreement_score":0.26022193,"about_ca_system_score_codex":0.00038358525,"about_ca_system_score_gemma":0.0019782227,"threshold_uncertainty_score":0.99729306},"labels":[],"label_agreement":null},{"id":"W2991923043","doi":"10.1136/bmjstel-2019-000515","title":"Risk orientation predicts hypoxic time during difficult airway simulation: a mixed-methods pilot study","year":2019,"lang":"en","type":"article","venue":"BMJ Simulation & Technology Enhanced Learning","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Alberta","funders":"","keywords":"Debriefing; Thematic analysis; Airway; Medicine; Personality; Intervention (counseling); Airway management; Psychology; Clinical psychology; Nursing; Qualitative research; Social psychology; Anesthesia; Medical education","score_opus":0.015579374097022711,"score_gpt":0.3782558540813377,"score_spread":0.362676479984315,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2991923043","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95422345,0.00005748482,0.0411052,0.00014175988,0.00035701934,0.0021416938,0.000002405041,0.0012078473,0.0007631402],"genre_scores_gemma":[0.9910865,0.000010269853,0.005850861,0.00003723671,0.00019172857,0.00012366465,0.000064978514,0.00006435673,0.0025704186],"study_design_codex":"simulation_or_modeling","study_design_gemma":"observational","domain_scores_codex":[0.9969064,0.00046104603,0.00088270864,0.0008213717,0.00047847588,0.0004500235],"domain_scores_gemma":[0.9886616,0.009270454,0.0006386918,0.00064889895,0.0006349825,0.0001453778],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0010044553,0.00032231494,0.0006559969,0.00049437926,0.00030313662,0.00003374944,0.00014808548,0.00032651,0.00045313913],"category_scores_gemma":[0.07974521,0.00030693616,0.00009790152,0.0012490705,0.0001031737,0.00015860327,0.00013128686,0.0009870582,0.0008695107],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000573417,0.00067458453,0.19320984,0.000022638833,0.00010462873,0.000012895126,0.00037523964,0.7342247,0.012164359,0.000008550826,0.0000066893026,0.058622487],"study_design_scores_gemma":[0.007307611,0.0052600754,0.5190662,0.00061190425,0.00027081335,0.000011221706,0.0009918873,0.4592409,0.0066028144,0.0001542945,0.00009775788,0.0003844937],"about_ca_topic_score_codex":0.0000098106,"about_ca_topic_score_gemma":0.000002092879,"teacher_disagreement_score":0.32585636,"about_ca_system_score_codex":0.00015856109,"about_ca_system_score_gemma":0.00008906124,"threshold_uncertainty_score":0.99993825},"labels":[],"label_agreement":null},{"id":"W2992787682","doi":"10.36834/cmej.36631","title":"Using script theory to cultivate illness script formation and clinical reasoning in health professions education","year":2015,"lang":"en","type":"article","venue":"Canadian Medical Education Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":140,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Université de Montréal; McGill University","funders":"","keywords":"Scripting language; Sketch; Interpretation (philosophy); Computer science; Foundation (evidence); Sociology of health and illness; Psychology; Health care","score_opus":0.06919590731136298,"score_gpt":0.44638243542674527,"score_spread":0.3771865281153823,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2992787682","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8376304,0.0023040944,0.00306652,0.14648153,0.0076753106,0.000854463,0.000005605988,0.000029271238,0.0019528054],"genre_scores_gemma":[0.88801825,0.00037612522,0.010623894,0.09792756,0.0019872063,0.000033133438,0.000056824723,0.000034016848,0.00094295945],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9971925,0.00059661927,0.000976589,0.0002458988,0.00054771337,0.000440668],"domain_scores_gemma":[0.9807067,0.00079535187,0.0002777343,0.00022615993,0.0006090093,0.017385067],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0058138235,0.00013909886,0.0003509331,0.000488142,0.00025765007,0.000070866816,0.00012263033,0.00023242894,0.00053791486],"category_scores_gemma":[0.30624712,0.000118305725,0.000058777183,0.00044056796,0.00011666435,0.00021234342,0.00002502083,0.0009815416,0.000027328731],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012426128,0.0007097563,0.06341468,0.000050908126,0.000022411527,0.000032929827,0.0063631674,0.000008330277,0.0000027646722,0.0034840656,0.15092403,0.7748627],"study_design_scores_gemma":[0.008700658,0.0014127422,0.37117714,0.062043544,0.00021861213,0.008419654,0.1121049,0.006830806,0.000016900458,0.012670514,0.4152781,0.0011264252],"about_ca_topic_score_codex":0.010759235,"about_ca_topic_score_gemma":0.006908502,"teacher_disagreement_score":0.7737363,"about_ca_system_score_codex":0.0010777212,"about_ca_system_score_gemma":0.09784206,"threshold_uncertainty_score":0.9958282},"labels":[],"label_agreement":null},{"id":"W2993100816","doi":"10.1177/0840470419869032","title":"The role of artificial intelligence in learning and professional development for healthcare professionals","year":2019,"lang":"en","type":"review","venue":"Healthcare Management Forum","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":91,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; McMaster University Medical Centre; Island Health","funders":"","keywords":"Health professionals; Health care; Professional development; Medical education; Psychology; Knowledge management; Engineering ethics; Medicine; Political science; Computer science; Engineering","score_opus":0.07794808491829462,"score_gpt":0.461033838228527,"score_spread":0.38308575331023237,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2993100816","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00003252623,0.9872735,0.00006123207,0.0058319685,0.0007901672,0.0057923654,0.000017658458,0.0000268509,0.00017369483],"genre_scores_gemma":[0.0005954412,0.9931995,0.0013179521,0.0002920315,0.00008463814,0.001253011,0.00026190784,0.00005066055,0.0029448017],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9965514,0.00026616806,0.0015468197,0.0005522001,0.00045936924,0.0006240706],"domain_scores_gemma":[0.99243444,0.006319356,0.0006087926,0.0002991226,0.0001504237,0.00018787709],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0015593217,0.00032158569,0.0013751148,0.00022589613,0.0002782758,0.00001836453,0.00017634683,0.00031671728,0.000009930525],"category_scores_gemma":[0.0023467017,0.0002032276,0.00021835198,0.00029645374,0.000073816496,0.000020990485,0.00031405452,0.00070800685,0.00001913276],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009516081,0.00015595576,0.00074552954,0.03119274,0.00007429031,0.0000048357024,0.000119726064,2.4579487e-7,1.040421e-8,0.019471461,0.00019784366,0.9479422],"study_design_scores_gemma":[0.0001660273,0.00026069957,0.00015208567,0.112133145,0.00013188229,0.0000056456547,0.0027987522,0.000059420272,0.0000011175197,0.0016858664,0.8824293,0.00017607203],"about_ca_topic_score_codex":0.00003504902,"about_ca_topic_score_gemma":0.00012203574,"teacher_disagreement_score":0.9477661,"about_ca_system_score_codex":0.00018427194,"about_ca_system_score_gemma":0.00091174775,"threshold_uncertainty_score":0.8287381},"labels":[],"label_agreement":null},{"id":"W2993418133","doi":"10.1097/acm.0000000000003098","title":"Making Decisions in the Era of the Clinical Decision Rule: How Emergency Physicians Use Clinical Decision Rules","year":2019,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":41,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Manitoba; University of Ottawa; Manitoba Health; Medical Council of Canada; McMaster University","funders":"","keywords":"Test (biology); Decision aids; Cognition; Clinical decision making; Psychology; Think aloud protocol; MEDLINE; Emergency department; Medical education; Medicine; Medical emergency; Family medicine; Computer science; Psychiatry; Alternative medicine; Pathology","score_opus":0.1298368193813043,"score_gpt":0.4701392533778751,"score_spread":0.3403024339965708,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2993418133","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9741627,0.0015509558,0.0014209546,0.017299958,0.003959206,0.0010034471,0.000013576476,0.00003681212,0.0005524095],"genre_scores_gemma":[0.97424245,0.01486342,0.0015071877,0.0072097275,0.0015848118,0.000017217342,0.000020410456,0.00004211615,0.0005126325],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9931782,0.0007834134,0.0029614556,0.0007900324,0.0017095457,0.00057735073],"domain_scores_gemma":[0.8755185,0.12135409,0.0008725309,0.001727497,0.00027299908,0.00025435397],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.009091039,0.00034334822,0.0013893839,0.00013242426,0.0001095421,0.000011060357,0.0011066031,0.0007895485,0.00033813165],"category_scores_gemma":[0.3481492,0.00016512843,0.00056743075,0.0009348834,0.00063177396,0.00013976946,0.0003394745,0.003966578,0.0001640097],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00046168905,0.00018782169,0.6698743,0.000008732899,0.000041699073,0.000022239397,0.00055250584,0.000009280884,0.00005357717,0.00054055196,0.0511443,0.2771033],"study_design_scores_gemma":[0.004880461,0.0006908,0.9571576,0.015596804,0.0003165596,0.000052965133,0.0010318813,0.0005201218,0.000010431406,0.0076925256,0.011869234,0.00018061223],"about_ca_topic_score_codex":0.000053974763,"about_ca_topic_score_gemma":0.000021796033,"teacher_disagreement_score":0.33905816,"about_ca_system_score_codex":0.000048345348,"about_ca_system_score_gemma":0.00021788508,"threshold_uncertainty_score":0.9983313},"labels":[],"label_agreement":null},{"id":"W2994400412","doi":"","title":"Power & type II errors in recently published ophthalmology research","year":2013,"lang":"en","type":"article","venue":"Investigative Ophthalmology & Visual Science","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Optometry; Type (biology); Ophthalmology; Medicine; Geology","score_opus":0.1095661599293697,"score_gpt":0.450274757805428,"score_spread":0.34070859787605834,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2994400412","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9444173,0.0001260967,1.13965484e-7,0.007977054,0.0007246969,0.0007913082,0.0000020879736,0.00004791632,0.045913447],"genre_scores_gemma":[0.99570626,0.000007931723,0.001135739,0.00073077885,0.000086761625,0.0001216774,0.000013875449,0.000026648378,0.0021703476],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9949297,0.0006017257,0.0006890579,0.0012698141,0.0010607505,0.0014489676],"domain_scores_gemma":[0.9897542,0.006265248,0.0001845549,0.00072159874,0.0018445428,0.0012298604],"candidate_categories":["metaresearch","sts","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0044482043,0.0002872056,0.00062906736,0.000824063,0.00048190507,0.00009653059,0.0007286134,0.00037269515,0.0069349064],"category_scores_gemma":[0.16314808,0.00023467223,0.00007945186,0.004705522,0.01017105,0.00057126366,0.0007445387,0.0013542679,0.0011342675],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018964449,0.0015350214,0.78127795,0.000014840647,0.000031219894,0.0022998503,0.0017828852,0.000009087892,0.20660873,0.0008227019,0.0049043368,0.0005237355],"study_design_scores_gemma":[0.0010598091,0.004911149,0.95330584,0.00046044434,0.000010143316,0.0020757357,0.0006399035,0.00017949576,0.018071301,0.018863369,0.00013522404,0.00028760757],"about_ca_topic_score_codex":0.0010465456,"about_ca_topic_score_gemma":0.0000028819284,"teacher_disagreement_score":0.18853743,"about_ca_system_score_codex":0.0003330932,"about_ca_system_score_gemma":0.0017669844,"threshold_uncertainty_score":0.99964345},"labels":[],"label_agreement":null},{"id":"W2995367892","doi":"10.1136/bmj.l6763","title":"Can academic satire exist in the age of “fake news?” Tracking the citation record of a “holiday review” paper","year":2019,"lang":"en","type":"article","venue":"BMJ","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McGill University Health Centre; Montreal Children's Hospital","funders":"","keywords":"Citation; Tracking (education); Fake news; Association (psychology); Library science; History; Psychology; Media studies; Computer science; Sociology; Pedagogy","score_opus":0.07317359500264478,"score_gpt":0.4016918184888096,"score_spread":0.3285182234861648,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2995367892","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9466628,0.002300802,0.000002217227,0.036259532,0.00014135428,0.00096532,0.0000060368384,0.000006909817,0.013655048],"genre_scores_gemma":[0.9878707,0.0026472234,0.00009715056,0.008679134,0.00011848796,0.000027881804,0.000019440788,0.000007939406,0.00053205015],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99882495,0.00016581468,0.0005271697,0.00011819724,0.0002555699,0.00010827112],"domain_scores_gemma":[0.99251646,0.006822807,0.00022711973,0.00035549383,0.000052107604,0.00002602365],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00146634,0.00006924311,0.00029362415,0.00002612449,0.000012252382,0.000003013254,0.00012450977,0.00007364662,0.00009366926],"category_scores_gemma":[0.033568084,0.000035695757,0.0000945842,0.00021527712,0.000057426467,0.000017581619,0.000014738808,0.00029175702,0.000016954],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015106317,0.00017099804,0.69537073,0.0008663523,0.000044243698,0.000024571473,0.0050082603,0.000007689756,0.002257615,0.0021056845,0.04695473,0.24703804],"study_design_scores_gemma":[0.0009625911,0.00020493484,0.9666397,0.01765601,0.000118653836,0.000015849568,0.0008305447,0.000014980852,0.00009012759,0.00095556694,0.012435839,0.000075191434],"about_ca_topic_score_codex":0.00057654537,"about_ca_topic_score_gemma":0.00023319144,"teacher_disagreement_score":0.27126896,"about_ca_system_score_codex":0.000014756174,"about_ca_system_score_gemma":0.00006510875,"threshold_uncertainty_score":0.9745726},"labels":[],"label_agreement":null},{"id":"W2995889517","doi":"10.1111/medu.14024","title":"Myths in medicine: How did we get here?","year":2019,"lang":"en","type":"letter","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Children's Hospital of Eastern Ontario; Centre for Addiction and Mental Health; University of Ottawa","funders":"","keywords":"Mythology; Medical education; Alternative medicine; Psychology; Engineering ethics; Medicine; History; Engineering; Classics","score_opus":0.021924770775028884,"score_gpt":0.36565835229978333,"score_spread":0.3437335815247544,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2995889517","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0019841522,0.0031045729,0.000036907066,0.9766078,0.005817565,0.00059148774,0.000006075827,0.00007239205,0.011779077],"genre_scores_gemma":[0.004268699,0.002078325,0.00016398191,0.9370384,0.018010462,0.00012489257,0.0018103742,0.00007472548,0.036430135],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99618906,0.00020368758,0.0006642142,0.0007116902,0.0017118835,0.00051945774],"domain_scores_gemma":[0.98705184,0.011339194,0.00026571404,0.00083281286,0.00018599229,0.00032446257],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity","insufficient_payload"],"category_scores_codex":[0.0009045207,0.00037255464,0.0010745892,0.00031369668,0.000027873324,0.000023481334,0.00031026616,0.0019236914,0.004718615],"category_scores_gemma":[0.18328524,0.00027729318,0.00017233858,0.0003232952,0.00030026134,0.00004438229,0.00006020463,0.0036452806,0.00079538603],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000020883717,0.00030711686,0.0047654537,0.000325891,0.000030070289,0.0002223094,0.00020960181,6.942786e-8,8.017627e-7,0.00001849828,0.9456643,0.048435025],"study_design_scores_gemma":[0.001556081,0.0002656099,0.0044113332,0.018474014,0.00016189543,0.00014201015,0.00020551607,0.00005845724,0.0000019783301,0.00025891614,0.9742024,0.00026178095],"about_ca_topic_score_codex":0.0003115635,"about_ca_topic_score_gemma":0.000016570688,"teacher_disagreement_score":0.1823807,"about_ca_system_score_codex":0.0002733359,"about_ca_system_score_gemma":0.00424444,"threshold_uncertainty_score":0.9999826},"labels":[],"label_agreement":null},{"id":"W2997373298","doi":"10.26481/dis.20190620as","title":"Through the eyes of the physician","year":2019,"lang":"en","type":"dissertation","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Cognition; Cognitive load; Psychology; Pupil; Medicine; Medical education; Psychiatry","score_opus":0.02068656990782289,"score_gpt":0.35818056979232166,"score_spread":0.33749399988449874,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2997373298","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.38264453,0.0006559417,0.0000072500543,0.0034819809,0.0017116442,0.0005973127,0.000010831767,0.000025488818,0.610865],"genre_scores_gemma":[0.8573337,0.00012870549,0.00004624318,0.003474781,0.0002185771,0.000010746503,0.0001436367,0.000020440213,0.13862316],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9992285,0.000025707976,0.0002256557,0.0001416063,0.00027844103,0.000100097466],"domain_scores_gemma":[0.9949431,0.004290787,0.00017284034,0.0004777333,0.00010016439,0.000015415151],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00007972673,0.000112283684,0.0002984722,0.000008029678,0.00003548907,0.0000061153332,0.00015934011,0.00014325918,0.00016807108],"category_scores_gemma":[0.007081331,0.000043909044,0.00027052363,0.00010615158,0.000052482214,0.000010428361,0.000015923433,0.00026809008,0.00011096238],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012633482,0.0025464385,0.028243152,0.0015599692,0.0019187453,0.000013200141,0.016464489,0.00004487228,0.0028504916,0.11432829,0.6267066,0.20406042],"study_design_scores_gemma":[0.0038802228,0.0008848049,0.81143486,0.028749466,0.0034589358,0.0000091683805,0.019989008,0.00012311923,0.045391027,0.011223281,0.074139625,0.0007164681],"about_ca_topic_score_codex":0.00028483485,"about_ca_topic_score_gemma":0.000054218417,"teacher_disagreement_score":0.78319174,"about_ca_system_score_codex":0.000009330772,"about_ca_system_score_gemma":0.00016199794,"threshold_uncertainty_score":0.84775263},"labels":[],"label_agreement":null},{"id":"W2997456271","doi":"10.1016/j.amjsurg.2019.12.027","title":"Mechanizing medicine – Tomorrows history started yesterday","year":2019,"lang":"en","type":"article","venue":"The American Journal of Surgery","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Yesterday; Craft; Subject (documents); History of medicine; History; Medical history; Medical practice; State (computer science); Medicine; Engineering ethics; Engineering; Computer science; Surgery; Intensive care medicine; Classics; Library science","score_opus":0.053614132741942724,"score_gpt":0.30510107711451756,"score_spread":0.25148694437257485,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2997456271","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98795295,0.0013181311,0.00011552726,0.007824981,0.001520027,0.000071489325,6.866035e-7,0.00001905337,0.001177184],"genre_scores_gemma":[0.99255717,0.00040763355,0.00018227138,0.0053343633,0.00061047956,9.4391464e-7,0.000001887389,0.000023123612,0.0008821387],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9983291,0.00016091227,0.0006703999,0.000119804004,0.00049008324,0.00022971713],"domain_scores_gemma":[0.98207134,0.01618755,0.00093050214,0.00033535925,0.0002347303,0.00024051811],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002074713,0.00012732603,0.0008937995,0.00018526166,0.000023345361,0.000004399952,0.00013445463,0.000025191957,0.0008482234],"category_scores_gemma":[0.011602284,0.00007024461,0.00026030195,0.00024643604,0.0003120806,0.00005388474,0.000025635594,0.00037342042,0.00008491666],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004405208,0.00061511016,0.4509318,0.00007623396,0.0008794008,0.0014342711,0.0029431598,0.000036980156,0.01796985,0.00033033724,0.23720379,0.2831739],"study_design_scores_gemma":[0.005285882,0.0059923176,0.85698026,0.015696848,0.001624653,0.007081338,0.0074289697,0.00028809792,0.0010669521,0.0010747939,0.09667032,0.0008095337],"about_ca_topic_score_codex":0.000083608975,"about_ca_topic_score_gemma":5.071173e-7,"teacher_disagreement_score":0.4060485,"about_ca_system_score_codex":0.00015275294,"about_ca_system_score_gemma":0.00034010422,"threshold_uncertainty_score":0.9967234},"labels":[],"label_agreement":null},{"id":"W2998363599","doi":"10.1097/acm.0000000000003142","title":"Philosophy of Science Series: Harnessing the Multidisciplinary Edge Effect by Exploring Paradigms, Ontologies, Epistemologies, Axiologies, and Methodologies","year":2019,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":119,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Multidisciplinary approach; Ontology; Epistemology; Engineering ethics; Generative grammar; GRASP; Rhetoric; Axiology; Philosophy of science; Sociology; Computer science; Data science; Social science; Artificial intelligence; Philosophy; Engineering","score_opus":0.14644577611081144,"score_gpt":0.4060008068098453,"score_spread":0.25955503069903385,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2998363599","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95670325,0.022639591,0.00016204978,0.015483708,0.001158659,0.0008379165,0.000008785691,0.0001886263,0.0028174235],"genre_scores_gemma":[0.98771137,0.009835206,0.00096714415,0.0004856784,0.00027052985,0.00008417175,0.000024311512,0.00002359215,0.00059798115],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9968197,0.000374112,0.0008471307,0.0007470167,0.00059939816,0.00061266875],"domain_scores_gemma":[0.9672283,0.031239478,0.0004148093,0.00077314547,0.00012198667,0.00022227825],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.005604354,0.00035525902,0.0012335825,0.00017451595,0.00024317089,0.000009355484,0.0006226662,0.00032897553,0.00003830806],"category_scores_gemma":[0.1219954,0.0001861295,0.00007892635,0.00063690904,0.006622601,0.00027824577,0.0005486668,0.0011199665,0.0000121325465],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0029000768,0.00008091495,0.7338038,0.0015503527,0.00032443574,0.00011910183,0.0097817015,0.000029111949,0.11565693,0.014384526,0.013472985,0.10789604],"study_design_scores_gemma":[0.020675818,0.023772022,0.71364474,0.039509222,0.0018865907,0.0023804347,0.043817848,0.00063210446,0.089668825,0.04646122,0.0154398875,0.0021113164],"about_ca_topic_score_codex":0.00013350896,"about_ca_topic_score_gemma":9.02575e-7,"teacher_disagreement_score":0.11639105,"about_ca_system_score_codex":0.00009792776,"about_ca_system_score_gemma":0.00021667947,"threshold_uncertainty_score":0.9960808},"labels":[],"label_agreement":null},{"id":"W2998759873","doi":"10.1111/medu.14056","title":"The cognitive process of test takers when using the script concordance test rating scale","year":2020,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Rating scale; Concordance; Psychology; Test (biology); Cognition; Scale (ratio); Likert scale; Applied psychology; Competence (human resources); Context (archaeology); Clinical psychology; Medical education; Social psychology; Medicine; Developmental psychology; Psychiatry","score_opus":0.029263393459681163,"score_gpt":0.3777618976822826,"score_spread":0.3484985042226014,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2998759873","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.86889315,0.0010257529,0.00094126596,0.12616894,0.0004930174,0.00053111673,0.000008079237,0.00003752045,0.0019011489],"genre_scores_gemma":[0.9836419,0.00007080031,0.00028112505,0.015175954,0.0006413939,0.000033602868,0.000017249551,0.000011894656,0.00012604483],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9986773,0.000059296835,0.0003655426,0.00018884047,0.000544522,0.00016446372],"domain_scores_gemma":[0.9592083,0.038853377,0.00038837723,0.00029129532,0.000761292,0.0004973734],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0006188909,0.000088897745,0.00018700497,0.000010581944,0.00016561022,0.000021623293,0.00014163414,0.00007133676,0.00013210882],"category_scores_gemma":[0.4577569,0.00004852589,0.000051547715,0.00022749502,0.00047038775,0.000034234225,0.000027381077,0.00032026856,0.00001299428],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00024469552,0.0021134906,0.5963092,0.00032301678,0.00007516717,0.00001347579,0.016527768,0.0000069995326,0.00095216045,0.000111763904,0.05301999,0.3303023],"study_design_scores_gemma":[0.016949173,0.0060034874,0.47722903,0.06353476,0.0030534843,0.0006085384,0.20587263,0.15967593,0.028757354,0.00454345,0.032158777,0.0016133977],"about_ca_topic_score_codex":0.000105390325,"about_ca_topic_score_gemma":0.000014798499,"teacher_disagreement_score":0.457138,"about_ca_system_score_codex":0.000025934612,"about_ca_system_score_gemma":0.0021449868,"threshold_uncertainty_score":0.5468106},"labels":[],"label_agreement":null},{"id":"W3004890774","doi":"10.1007/s40596-020-01186-0","title":"Curriculum Mapping: an Innovative Approach to Mapping the Didactic Lecture Series at the University of Toronto Postgraduate Psychiatry","year":2020,"lang":"en","type":"article","venue":"Academic Psychiatry","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Centre for Addiction and Mental Health; University of Toronto","funders":"","keywords":"Curriculum; Medical education; Series (stratigraphy); Psychology; Medicine; Pedagogy","score_opus":0.03689876626027671,"score_gpt":0.2902811978241722,"score_spread":0.2533824315638955,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3004890774","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.82181156,0.005252807,0.0009852501,0.16457948,0.0009539122,0.00077919726,0.000027047756,0.00011359198,0.0054971627],"genre_scores_gemma":[0.97418547,0.000507193,0.005098848,0.01879563,0.0009900997,0.000007245442,0.000058040154,0.00003487088,0.000322584],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9981426,0.00015365754,0.00043057598,0.00051959086,0.00038684547,0.00036670332],"domain_scores_gemma":[0.9982272,0.00046415717,0.00027702958,0.0005390057,0.00016624026,0.00032638386],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00038684503,0.0002593056,0.00041768808,0.000028419714,0.0003259972,0.000011632207,0.0005477598,0.00027996505,0.00010277436],"category_scores_gemma":[0.0028761553,0.00016146564,0.00015589545,0.00079364574,0.00025010327,0.00014732787,0.0002319607,0.0010388077,0.000047109366],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00081524206,0.00072927814,0.75102466,0.0002799157,0.00055326044,0.000005137103,0.03645643,0.000097288306,0.0007550436,0.0102651715,0.1928374,0.0061811986],"study_design_scores_gemma":[0.0061942204,0.002200534,0.70737857,0.0031059822,0.0008104307,0.0003674551,0.12787597,0.0016599975,0.00012143466,0.004044855,0.14496925,0.0012712763],"about_ca_topic_score_codex":0.0005720634,"about_ca_topic_score_gemma":0.00014907878,"teacher_disagreement_score":0.15237394,"about_ca_system_score_codex":0.00009932927,"about_ca_system_score_gemma":0.00036114312,"threshold_uncertainty_score":0.6584377},"labels":[],"label_agreement":null},{"id":"W3005271500","doi":"10.1097/adm.0000000000000632","title":"Guidance for Writing Case Reports in Addiction Medicine","year":2020,"lang":"en","type":"editorial","venue":"Journal of Addiction Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Scholarship; Medicine; Observational study; MEDLINE; Clinical trial; Medical education; Systematic review; Value (mathematics); Addiction; Psychiatry; Pathology","score_opus":0.02716902306448952,"score_gpt":0.37481189472824206,"score_spread":0.3476428716637525,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3005271500","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0018058957,0.018069228,0.0019667053,0.022011792,0.9533017,0.0010274949,0.00010251715,0.000098410914,0.0016162],"genre_scores_gemma":[0.01359861,0.0039318795,0.0013102979,0.0010638776,0.97865593,0.000042537817,0.00063193106,0.00011644153,0.0006484806],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9921834,0.00016891558,0.004556407,0.0006147854,0.0020277498,0.0004487439],"domain_scores_gemma":[0.9667526,0.025381878,0.0040806057,0.00047425422,0.002457683,0.00085301086],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.005178625,0.0005131482,0.0025068556,0.0008209128,0.00011856669,0.000016059996,0.00012717929,0.0010569599,0.0005276397],"category_scores_gemma":[0.37162894,0.00037747028,0.00041647334,0.00069469487,0.00031801843,0.00014531481,0.00003997703,0.002756959,0.0000053997555],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013557336,0.0001594333,0.0004102589,0.00021846317,0.00030001914,0.045668937,0.0002853252,0.000019255913,0.00011870443,0.0000056949248,0.93877316,0.01268504],"study_design_scores_gemma":[0.011633529,0.007978053,0.00064510247,0.05262648,0.0026226344,0.03266985,0.00048718063,0.00021335231,0.000025940579,0.00015507953,0.8906655,0.00027734568],"about_ca_topic_score_codex":0.00024254296,"about_ca_topic_score_gemma":0.000042993976,"teacher_disagreement_score":0.3664503,"about_ca_system_score_codex":0.0005491757,"about_ca_system_score_gemma":0.001016392,"threshold_uncertainty_score":0.99986774},"labels":[],"label_agreement":null},{"id":"W3007499719","doi":"10.1136/bmjqs-2019-010673","title":"Allowing failure so trainees can thrive: the importance of guided autonomy in medical education","year":2020,"lang":"en","type":"letter","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Patient safety; Mistake; Autonomy; Medicine; Commit; Medical education; Quality (philosophy); Variety (cybernetics); Nursing; Best practice; Obligation; Health care; Computer science","score_opus":0.05492717542364938,"score_gpt":0.42390810620824765,"score_spread":0.36898093078459826,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3007499719","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.010522915,0.0004172713,0.0000887281,0.9854236,0.00047150298,0.0010671914,0.00008588332,0.00006296868,0.0018599244],"genre_scores_gemma":[0.05387687,0.0001609803,0.0007707383,0.9382814,0.0043424414,0.0001330434,0.0017433267,0.000053037038,0.000638141],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9944918,0.00058968394,0.0027111399,0.0006014286,0.001209556,0.00039638748],"domain_scores_gemma":[0.98338735,0.0145711955,0.0007109406,0.0008690343,0.00019489788,0.0002665591],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0029875217,0.0003471219,0.0012357623,0.000069227826,0.00007518332,0.000021118782,0.00044338405,0.0008210852,0.00063681224],"category_scores_gemma":[0.12303903,0.00024229125,0.0004519296,0.00032075125,0.00040262175,0.000032201053,0.00011154852,0.002377515,0.000016417138],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007319403,0.0001719298,0.027276475,0.00034636748,0.00012621972,0.0002631209,0.0019329414,0.0000019392367,0.000004680339,0.00042021656,0.9591763,0.010206618],"study_design_scores_gemma":[0.0025639522,0.00013499043,0.09444886,0.008987412,0.00032210143,0.00008956969,0.0014135767,0.000095617455,0.000016496673,0.0016127634,0.8897607,0.00055395986],"about_ca_topic_score_codex":0.0015489034,"about_ca_topic_score_gemma":0.00071169814,"teacher_disagreement_score":0.1200515,"about_ca_system_score_codex":0.0002611049,"about_ca_system_score_gemma":0.011859793,"threshold_uncertainty_score":0.99992406},"labels":[],"label_agreement":null},{"id":"W3007534255","doi":"10.3352/jeehp.2020.17.5","title":"Potential of feedback during objective structured clinical examination to evoke an emotional response in medical students in Canada","year":2020,"lang":"en","type":"article","venue":"Journal of Educational Evaluation for Health Professions","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Medical Council of Canada; University of Ottawa","funders":"","keywords":"Embarrassment; Objective structured clinical examination; Psychology; Emotional reaction; Clinical psychology; Medical education; Medicine; Social psychology; Psychiatry","score_opus":0.08914017321305964,"score_gpt":0.5361213036606929,"score_spread":0.4469811304476332,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3007534255","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.89125,0.000061207385,0.00011810165,0.10668473,0.00088085415,0.000966881,0.00002978975,0.0000014129503,0.0000070224373],"genre_scores_gemma":[0.99440736,0.0000204077,0.0014656429,0.003451501,0.0005087741,0.00004228167,0.00006790846,0.0000108239055,0.000025289757],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99395794,0.0015515581,0.0018819403,0.00021857074,0.0021896358,0.00020033335],"domain_scores_gemma":[0.9883319,0.008517657,0.00082985923,0.00011341623,0.0013701108,0.0008370145],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.009015798,0.00009695058,0.00045917564,0.00024714583,0.00007423351,0.0000061251044,0.00018375256,0.000112463444,0.00049111806],"category_scores_gemma":[0.15037292,0.000083918494,0.00008431124,0.00039360722,0.000031200485,0.000097120916,0.000037670306,0.0004959175,0.0000015770845],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.011794735,0.00248401,0.9628084,0.00016423661,0.0000677751,0.000009768049,0.0037617455,0.003788145,0.00017457064,0.00010638538,0.003119453,0.01172081],"study_design_scores_gemma":[0.0059275664,0.0011886577,0.98554134,0.0020570017,0.000031967422,0.00002059581,0.0023100742,0.0025578937,0.000015003553,0.0002760919,0.000013237658,0.00006055621],"about_ca_topic_score_codex":0.0058851545,"about_ca_topic_score_gemma":0.03597429,"teacher_disagreement_score":0.14135712,"about_ca_system_score_codex":0.0012818692,"about_ca_system_score_gemma":0.04006302,"threshold_uncertainty_score":0.9816167},"labels":[],"label_agreement":null},{"id":"W3009248208","doi":"10.1007/s10459-020-09963-0","title":"A think-aloud study to inform the design of radiograph interpretation practice","year":2020,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"SickKids Foundation; Hospital for Sick Children; University of Toronto","funders":"","keywords":"Interpretation (philosophy); Think aloud protocol; Protocol analysis; Computer science; Psychology; Radiology; Medical education; Medicine; Human–computer interaction; Cognitive science; Usability","score_opus":0.047002687748520705,"score_gpt":0.4666914205843243,"score_spread":0.41968873283580355,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3009248208","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7482671,0.013864905,0.040866397,0.186622,0.0019239209,0.0060413964,0.0000023314933,0.00007007203,0.002341861],"genre_scores_gemma":[0.96047777,0.00088584103,0.023588601,0.014934613,0.00007733015,0.000020812207,0.0000015839668,0.0000031545078,0.000010321506],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.9986012,0.00018541767,0.00048645763,0.0002142949,0.00036025542,0.00015238598],"domain_scores_gemma":[0.9907724,0.008469217,0.0002869261,0.0001522787,0.00014725923,0.00017188712],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0023308583,0.000065783875,0.00017647241,0.00010662488,0.00011284231,0.000017346223,0.00016172855,0.000014857085,0.000009721487],"category_scores_gemma":[0.054037165,0.000042733995,0.000023494511,0.0016790037,0.000121368415,0.00033610367,0.000021255271,0.00012740844,0.000008513221],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00088950474,0.0024891354,0.080189444,0.00020409361,0.000015618034,0.0000014169706,0.36341795,0.02048345,0.000010410413,0.0034237057,0.0016590413,0.52721626],"study_design_scores_gemma":[0.0021232364,0.027354315,0.24459526,0.004438358,0.00010292518,0.00003469921,0.68692565,0.022274796,0.000075937445,0.0050987336,0.0066106184,0.00036548622],"about_ca_topic_score_codex":0.00032148726,"about_ca_topic_score_gemma":0.000025933447,"teacher_disagreement_score":0.52685076,"about_ca_system_score_codex":0.000048653954,"about_ca_system_score_gemma":0.0009936518,"threshold_uncertainty_score":0.9539311},"labels":[],"label_agreement":null},{"id":"W3009341746","doi":"10.1111/acem.13955","title":"Shared Decision Making for Syncope in the Emergency Department: A Randomized Controlled Feasibility Trial","year":2020,"lang":"en","type":"article","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":28,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital","funders":"National Heart, Lung, and Blood Institute","keywords":"Medicine; Emergency department; Randomized controlled trial; Confidence interval; Syncope (phonology); Clinical equipoise; Emergency medicine; Physical therapy; Internal medicine","score_opus":0.10536285440653394,"score_gpt":0.44316656752769534,"score_spread":0.3378037131211614,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3009341746","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8982652,0.018141557,0.0059015146,0.045359768,0.004613807,0.023805894,0.00003869519,0.00014867993,0.0037248847],"genre_scores_gemma":[0.9843046,0.00696363,0.00033892298,0.003102411,0.0029794706,0.0020053252,0.0001060198,0.000038689977,0.00016091816],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.99337244,0.00063103554,0.0037599076,0.0007490686,0.00095531385,0.0005322574],"domain_scores_gemma":[0.9641423,0.03362382,0.0009033153,0.00060783507,0.0002880097,0.00043468713],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.008393858,0.0003915302,0.0027536892,0.00012093278,0.00011791737,0.0000015930311,0.0004636167,0.00036408525,0.0030981705],"category_scores_gemma":[0.44797572,0.00020759896,0.0008598043,0.0007330128,0.00016114584,0.00007565355,0.00006789106,0.0009250632,0.00003968133],"study_design_candidate":"randomized_trial","study_design_consensus":"randomized_trial","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.91777134,0.0002640544,0.014136071,0.00008468071,0.0002047217,0.000017569357,0.0016689358,0.000007834206,0.000071736824,0.0006172271,0.06357514,0.0015806726],"study_design_scores_gemma":[0.9760912,0.0017952641,0.0050316555,0.0017310028,0.0014463684,0.000008230057,0.000392546,0.0034162502,0.000004936969,0.008382371,0.0014597581,0.00024042618],"about_ca_topic_score_codex":0.000028858924,"about_ca_topic_score_gemma":0.000011728818,"teacher_disagreement_score":0.43958187,"about_ca_system_score_codex":0.00006463465,"about_ca_system_score_gemma":0.00014544117,"threshold_uncertainty_score":0.9978131},"labels":[],"label_agreement":null},{"id":"W3012224981","doi":"10.12968/hmed.2019.0406","title":"Teaching and learning clinical reasoning: a teacher's toolbox to meet different learning needs","year":2020,"lang":"en","type":"article","venue":"British Journal of Hospital Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université de Sherbrooke","funders":"","keywords":"Toolbox; Medicine; Coaching; Table (database); Clinical Practice; Medical education; Computer science; Psychology; Nursing","score_opus":0.022824248840775406,"score_gpt":0.32746823033277606,"score_spread":0.30464398149200067,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3012224981","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95577157,0.003675028,0.0007387865,0.036657687,0.0005805415,0.000227875,0.0000011109033,0.00007079786,0.0022765922],"genre_scores_gemma":[0.98934394,0.000731627,0.0018487569,0.0034946383,0.003870717,0.0000033229576,0.00000620088,0.000052216103,0.000648569],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99652267,0.0005822597,0.001404228,0.00034678032,0.0007442552,0.00039980048],"domain_scores_gemma":[0.99186504,0.005431911,0.0005484323,0.00011541439,0.0002721454,0.0017670528],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0028360963,0.00025828098,0.0013001626,0.00013918617,0.0002364888,0.00006853456,0.00017808066,0.00018100183,0.00013678953],"category_scores_gemma":[0.28422678,0.00021928233,0.00025524994,0.00023325472,0.00021699743,0.00011640802,0.00011335278,0.0033042058,0.000008293314],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00072901975,0.0007329495,0.40136847,0.00008599974,0.00037318727,0.0039971326,0.0068820175,0.000043340013,0.0001550968,0.00013779546,0.028306875,0.55718815],"study_design_scores_gemma":[0.02271639,0.09242441,0.76650923,0.044606626,0.001864731,0.009716448,0.024032922,0.0023108646,0.000038837305,0.00020712403,0.034541413,0.0010310144],"about_ca_topic_score_codex":0.00019510025,"about_ca_topic_score_gemma":0.0000040711493,"teacher_disagreement_score":0.5561571,"about_ca_system_score_codex":0.000048942933,"about_ca_system_score_gemma":0.00009059254,"threshold_uncertainty_score":0.9989952},"labels":[],"label_agreement":null},{"id":"W3014969104","doi":"10.1186/s12909-020-02012-9","title":"Mapping clinical reasoning literature across the health professions: a scoping review","year":2020,"lang":"en","type":"review","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":128,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Centre for Interdisciplinary Research in Rehabilitation; McGill University Health Centre","funders":"Fonds de Recherche du Québec - Santé; Santen; Faculty of Medicine, McGill University; McGill University","keywords":"Medical education; MEDLINE; Health professions; Medicine; Psychology; Engineering ethics; Health care; Political science; Engineering","score_opus":0.15963247321797686,"score_gpt":0.5739068030996893,"score_spread":0.4142743298817125,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3014969104","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000021590456,0.9662373,0.0003047528,0.025573708,0.0027483597,0.0048012543,0.000014208127,0.0001292298,0.00018901785],"genre_scores_gemma":[7.8377445e-7,0.9548851,0.0021283587,0.03610618,0.004325985,0.0009568278,0.0010264617,0.00007277959,0.0004974788],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9917106,0.0023661356,0.0031538182,0.000976961,0.0011909734,0.0006014814],"domain_scores_gemma":[0.9336714,0.057488773,0.0029339455,0.0020514554,0.00049682736,0.0033576007],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.007946371,0.00051132357,0.003809514,0.00006693533,0.00033228786,0.000083071565,0.00053406216,0.0008885053,0.00034358157],"category_scores_gemma":[0.54120946,0.0002816568,0.0012180813,0.0013123486,0.0002437835,0.00005394642,0.00026473682,0.0033252072,0.00029025547],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000007157984,0.00027690476,0.00009063369,0.17511623,0.00004859957,0.000016781083,0.0002598916,2.2247355e-9,2.528655e-10,0.00006009267,0.05149927,0.77262443],"study_design_scores_gemma":[0.0001957128,0.0000806972,0.000049683385,0.56262165,0.00024296033,0.00030686925,0.000107645785,0.0000063332773,1.5298891e-9,0.000010013801,0.436243,0.00013542462],"about_ca_topic_score_codex":0.000021472079,"about_ca_topic_score_gemma":0.000015013187,"teacher_disagreement_score":0.772489,"about_ca_system_score_codex":0.00021667997,"about_ca_system_score_gemma":0.05568808,"threshold_uncertainty_score":0.9999636},"labels":[],"label_agreement":null},{"id":"W3015880472","doi":"10.1002/art.41278","title":"Stand Together and Deliver: Challenges and Opportunities for Rheumatology Education During the <scp>COVID</scp>‐19 Pandemic","year":2020,"lang":"en","type":"article","venue":"Arthritis & Rheumatology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":36,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"National Center for Advancing Translational Sciences","keywords":"Accreditation; Pandemic; Medicine; Medical education; Isolation (microbiology); Health care; Population; Graduate medical education; Coronavirus disease 2019 (COVID-19); Family medicine; Psychology; Nursing; Disease; Internal medicine; Political science","score_opus":0.07352747539879376,"score_gpt":0.31431771299267275,"score_spread":0.240790237593879,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3015880472","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8682488,0.12052384,0.00007997607,0.009993681,0.00016904139,0.00040781446,0.00001855899,0.00009256536,0.00046567657],"genre_scores_gemma":[0.7537495,0.24410646,0.00018769871,0.0016756625,0.000026191938,0.000100367404,0.000014527943,0.000022161537,0.000117441115],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9986011,0.00012905597,0.00041121547,0.00041873186,0.000103663486,0.0003362621],"domain_scores_gemma":[0.98929065,0.009868881,0.00014133342,0.00019545347,0.00010211147,0.00040153763],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0003028614,0.00018783493,0.0005770861,0.000053964584,0.00019617975,0.000023272856,0.00008216355,0.00020267499,0.000022635515],"category_scores_gemma":[0.016983712,0.00015093198,0.00006152805,0.000039937288,0.00046882447,0.00006204732,0.00008751494,0.00020482067,0.000007730695],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015893437,0.00015488954,0.63599694,0.0008026878,0.00018460143,0.00005855027,0.01213521,8.977688e-7,0.000060630675,0.026132893,0.006019554,0.3182942],"study_design_scores_gemma":[0.030762946,0.0039794208,0.42945203,0.00906701,0.00056997227,0.043318078,0.08975907,0.00074710185,0.000094079776,0.024091508,0.3673392,0.0008195967],"about_ca_topic_score_codex":0.00008444341,"about_ca_topic_score_gemma":0.0000975691,"teacher_disagreement_score":0.36131963,"about_ca_system_score_codex":0.000018926128,"about_ca_system_score_gemma":0.00027678596,"threshold_uncertainty_score":0.99129665},"labels":[],"label_agreement":null},{"id":"W3016904076","doi":"10.1096/fasebj.2020.34.s1.05702","title":"Neurocognition of Teaching and Learning Clinical Reasoning in Veterinary Pathology Using Eye‐tracking and Electroencephalography","year":2020,"lang":"en","type":"article","venue":"The FASEB Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Calgary","funders":"","keywords":"Eye tracking; Neurocognitive; Electroencephalography; Eye movement; Cognition; Artificial intelligence; Computer science; Psychology; Medicine; Cognitive psychology; Medical education; Neuroscience","score_opus":0.06101769726291117,"score_gpt":0.3819770505118859,"score_spread":0.32095935324897473,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3016904076","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9973921,0.0006563801,0.000746696,0.0009631096,0.000056182005,0.00007858416,5.1092724e-7,0.000012527114,0.00009391973],"genre_scores_gemma":[0.9957419,0.000594197,0.0026359234,0.00078212086,0.00023071292,6.502882e-7,0.0000014994038,0.000011362565,0.0000016313827],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99823195,0.0008381418,0.0004479139,0.00016742521,0.00012488206,0.00018971521],"domain_scores_gemma":[0.995223,0.004288918,0.00022028238,0.000056494515,0.000041505333,0.00016982794],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002203781,0.00008993066,0.00030404102,0.000070321505,0.00018658998,0.000026016527,0.000046889374,0.000069655245,0.0000063220086],"category_scores_gemma":[0.029670691,0.00006520329,0.00007576868,0.00010868774,0.00017499627,0.000065782944,0.000052198335,0.0018953375,3.6560138e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004983097,0.000076900775,0.85421675,0.000023440669,0.000032811357,0.000719494,0.001600286,0.000056887253,0.07272542,0.000028351766,0.000005895243,0.07001546],"study_design_scores_gemma":[0.0029596249,0.0036500376,0.9682003,0.002259781,0.0002359713,0.006030282,0.0016343626,0.014013697,0.0005432811,0.0003055505,0.000029644816,0.00013743425],"about_ca_topic_score_codex":0.000013199949,"about_ca_topic_score_gemma":4.4713002e-7,"teacher_disagreement_score":0.113983594,"about_ca_system_score_codex":0.000007165368,"about_ca_system_score_gemma":0.000039441824,"threshold_uncertainty_score":0.9785028},"labels":[],"label_agreement":null},{"id":"W3017384349","doi":"10.1096/fasebj.2020.34.s1.01825","title":"Cognitive Integration: The Value of Explicitly Communicating the Connections Between Anatomy and Clinical Science","year":2020,"lang":"en","type":"article","venue":"The FASEB Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Humber Polytechnic","funders":"","keywords":"Cognition; Psychology; Disease; Medical education; Medicine; Pathology; Neuroscience","score_opus":0.10256573128159854,"score_gpt":0.42196625691462025,"score_spread":0.3194005256330217,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3017384349","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94101256,0.00036130034,0.0013907753,0.055777945,0.000082432205,0.0001814686,0.0000074476643,0.000010022868,0.0011760425],"genre_scores_gemma":[0.9957967,0.00034674056,0.00021224294,0.003167864,0.00045431577,0.0000028588047,0.0000016997313,0.000005552447,0.000012064422],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9986506,0.00034719746,0.00045499287,0.00010857197,0.00030404597,0.00013458639],"domain_scores_gemma":[0.9575524,0.041253276,0.00032081036,0.0003052084,0.0003361459,0.00023218311],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0032550185,0.0000721914,0.0002102089,0.000019327827,0.00078056887,0.00006283195,0.00035902945,0.000035179437,0.000028969784],"category_scores_gemma":[0.061791442,0.000028880482,0.000107900654,0.0003386509,0.0017212029,0.000050478906,0.00014961265,0.0009773807,0.0000069673983],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004839567,0.0003697292,0.70553535,0.00001490568,0.0006997827,0.00002872057,0.018292392,0.00003725334,0.0031290487,0.015872093,0.002637875,0.2528989],"study_design_scores_gemma":[0.004282782,0.002028041,0.921398,0.0026586247,0.0016890223,0.0007442471,0.042989142,0.010259203,0.010183704,0.0030149021,0.0005441819,0.0002081734],"about_ca_topic_score_codex":0.00004526942,"about_ca_topic_score_gemma":0.000003432582,"teacher_disagreement_score":0.25269073,"about_ca_system_score_codex":0.000012427687,"about_ca_system_score_gemma":0.00024722054,"threshold_uncertainty_score":0.9461115},"labels":[],"label_agreement":null},{"id":"W3017575403","doi":"10.1017/cem.2020.23","title":"Just the Facts: Diagnosing growth plate fractures in the emergency department","year":2020,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Saskatchewan; Hospital for Sick Children; Royal University Hospital","funders":"","keywords":"Ankle; Neurovascular bundle; Emergency department; Medicine; Lateral malleolus; Plantar flexion; Orthodontics; Anatomy","score_opus":0.1174647469861402,"score_gpt":0.389292132019897,"score_spread":0.2718273850337568,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3017575403","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7123975,0.012613696,0.00008721961,0.26465043,0.0052376403,0.00042194038,0.000015352987,0.000008614878,0.0045675663],"genre_scores_gemma":[0.9855636,0.0039845123,0.000025024845,0.0075127715,0.0028014693,0.000007759187,0.000013772909,0.000021171094,0.00006997244],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99743336,0.00016780144,0.0012334498,0.00018962008,0.0005334385,0.00044230028],"domain_scores_gemma":[0.9960148,0.0018808311,0.00040180358,0.00025614517,0.00028253687,0.0011639296],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001126005,0.00021477896,0.00049974,0.00014168404,0.00017128958,0.000007385507,0.00042671507,0.00009535614,0.00561079],"category_scores_gemma":[0.05404503,0.00010385323,0.00020712365,0.0006501961,0.00013139578,0.00006955975,0.000012530767,0.000820988,0.000025373807],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000106877196,0.00007874732,0.3813429,0.000047881294,0.00014483443,0.0012694766,0.009870302,0.000042833493,0.0000370953,0.00034375963,0.60292006,0.003795261],"study_design_scores_gemma":[0.0022457219,0.0017862179,0.8776399,0.0016294584,0.0006204131,0.00014538533,0.004848404,0.000080986334,0.00004221093,0.0009710214,0.109754056,0.00023624567],"about_ca_topic_score_codex":0.004395987,"about_ca_topic_score_gemma":0.007835356,"teacher_disagreement_score":0.496297,"about_ca_system_score_codex":0.00007589161,"about_ca_system_score_gemma":0.00062182813,"threshold_uncertainty_score":0.9952982},"labels":[],"label_agreement":null},{"id":"W3019726490","doi":"10.1503/cmaj.75308","title":"Sample sizes in COVID-19–related research","year":2020,"lang":"en","type":"letter","venue":"Canadian Medical Association Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Coronavirus disease 2019 (COVID-19); 2019-20 coronavirus outbreak; Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); Clinical trial; Coronavirus Infections; Sample size determination; Betacoronavirus; Argument (complex analysis); Sample (material); Coronavirus; Medicine; Computer science; Data science; Family medicine; Disease; Statistics; Virology; Pathology; Internal medicine; Mathematics; Infectious disease (medical specialty); Outbreak","score_opus":0.0614344497346976,"score_gpt":0.3920385698115993,"score_spread":0.3306041200769017,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3019726490","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.001043399,0.0005689008,0.0000197193,0.9931858,0.0012958898,0.00027240376,0.00012071048,0.000035698762,0.0034574838],"genre_scores_gemma":[0.0040291552,0.0009577132,0.00010656138,0.98119104,0.011664513,0.00001660053,0.00045511182,0.00005349572,0.0015257959],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.991716,0.0011718239,0.001112907,0.0004491399,0.004314922,0.0012351507],"domain_scores_gemma":[0.9369447,0.054621764,0.00034962574,0.00024532838,0.000654704,0.007183866],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity","insufficient_payload"],"category_scores_codex":[0.00612521,0.00023730886,0.0008188434,0.0007548397,0.00030172186,0.00017202224,0.0004048204,0.0030595711,0.023491712],"category_scores_gemma":[0.7272058,0.00021689676,0.0002859795,0.0009873384,0.00016993101,0.000051912888,0.00004342018,0.018900733,0.0009822957],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.000014208748,0.000025543803,0.017018594,0.0000317976,0.00012141193,0.040120766,0.0002467968,0.0000010076294,1.0984724e-7,0.000022363143,0.9340303,0.008367087],"study_design_scores_gemma":[0.00230622,0.00015959435,0.002901657,0.0017931657,0.000058524354,0.0008827341,0.00016759064,0.00011879859,2.0560556e-7,0.0020765045,0.98933905,0.0001959497],"about_ca_topic_score_codex":0.027746584,"about_ca_topic_score_gemma":0.019611524,"teacher_disagreement_score":0.7210806,"about_ca_system_score_codex":0.005923936,"about_ca_system_score_gemma":0.03600252,"threshold_uncertainty_score":0.99979556},"labels":[],"label_agreement":null},{"id":"W3020864904","doi":"10.1136/jnnp-2019-bnpa.28","title":"28 Using experimental simulation of dementia to understand functional cognitive disorders","year":2019,"lang":"en","type":"article","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Dementia; Computer science; Cognition; Psychology; Medicine; Neuroscience","score_opus":0.05786041986399412,"score_gpt":0.3767031247273309,"score_spread":0.31884270486333677,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3020864904","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9641268,0.00006516816,0.029327324,0.0001008289,0.00014311635,0.0002652724,0.0000032253413,0.000014386572,0.0059538693],"genre_scores_gemma":[0.9987181,0.0000019134284,0.00053835846,0.0003800729,0.00003064616,0.0000011181049,0.0000194762,0.0000070584324,0.00030328394],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99944663,0.000010203138,0.00015027726,0.00013139469,0.00017547453,0.000086002845],"domain_scores_gemma":[0.997016,0.002756005,0.000033205513,0.00006280601,0.000054250424,0.00007773332],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00006546911,0.000055953795,0.00012311473,0.000040032253,0.000017611694,0.0000040291607,0.000011939719,0.000034052868,0.002073647],"category_scores_gemma":[0.0022537871,0.00004762644,0.00005066045,0.00007844879,0.00002512916,0.000027977572,0.000018052178,0.000034444765,0.00006412861],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0035537833,0.0026689728,0.8657861,0.00006284529,0.0007644373,0.000004954623,0.002088624,0.08558945,0.027479855,0.0023105836,0.0010252128,0.0086652115],"study_design_scores_gemma":[0.028465007,0.005946602,0.627782,0.0044959667,0.0011758013,0.000013429389,0.041980628,0.23688415,0.050880305,0.0011483165,0.0003729247,0.00085490895],"about_ca_topic_score_codex":0.000045543118,"about_ca_topic_score_gemma":0.000004502684,"teacher_disagreement_score":0.2380041,"about_ca_system_score_codex":0.000029157483,"about_ca_system_score_gemma":0.00004213195,"threshold_uncertainty_score":0.9988386},"labels":[],"label_agreement":null},{"id":"W3022875342","doi":"10.1007/s11606-020-05864-5","title":"Pitfalls from Psychology Science that Worsen with Practice","year":2020,"lang":"en","type":"editorial","venue":"Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Institute for Clinical Evaluative Sciences; Sunnybrook Health Science Centre","funders":"Institute of Health Services and Policy Research; Canada Research Chairs; Canadian Institutes of Health Research; Sunnybrook Research Institute","keywords":"Medicine; MEDLINE; Medical education; Family medicine","score_opus":0.036894231820841075,"score_gpt":0.42651304540608687,"score_spread":0.3896188135852458,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3022875342","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.007207238,0.004613217,0.0012229707,0.051753227,0.9198825,0.00025245914,0.000044643522,0.00003466191,0.014989088],"genre_scores_gemma":[0.010444325,0.0033992676,0.008216241,0.010570205,0.96440786,0.0000034233058,0.000063280124,0.00007977993,0.0028156433],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9936242,0.00016472231,0.001414378,0.0006388671,0.0036863533,0.00047148103],"domain_scores_gemma":[0.972772,0.01973479,0.0027027372,0.00058769796,0.0028175174,0.0013852611],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0019175244,0.00047882303,0.0018881205,0.00043464216,0.00006715945,0.000061667284,0.000893419,0.00054646045,0.00048226028],"category_scores_gemma":[0.19495846,0.00027656698,0.00025801623,0.00046355216,0.0011995082,0.00024361261,0.00014323516,0.0037437382,0.000056425695],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0061026253,0.00025426215,0.0012301621,0.000023890701,0.0006287023,0.0062102345,0.0002663074,0.0000024611718,0.0008903783,0.000011051993,0.97967595,0.0047039883],"study_design_scores_gemma":[0.007920831,0.011622371,0.0031311142,0.016031474,0.0021795847,0.0018870835,0.0001407067,0.000024159246,0.00024797043,0.0001618643,0.95636106,0.00029175592],"about_ca_topic_score_codex":0.001056532,"about_ca_topic_score_gemma":0.000020181133,"teacher_disagreement_score":0.19304094,"about_ca_system_score_codex":0.0002655891,"about_ca_system_score_gemma":0.0019011219,"threshold_uncertainty_score":0.99996865},"labels":[],"label_agreement":null},{"id":"W3024174351","doi":"10.1017/cem.2020.237","title":"P030: Assessment of lab results on emergency department patients that leave without seeing a physician","year":2020,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; Sunnybrook Health Science Centre","funders":"","keywords":"Medicine; Emergency department; Medical emergency; Medical record; Emergency medicine; Chart; Family medicine; Nursing; Surgery","score_opus":0.11069794248129879,"score_gpt":0.39421264537442396,"score_spread":0.2835147028931252,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3024174351","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94950885,0.0011291973,0.00018674685,0.020074654,0.0062817386,0.0004887146,0.00011703651,0.000014117288,0.022198953],"genre_scores_gemma":[0.9957932,0.0007297522,0.00018354514,0.0013698727,0.0015438504,0.0000038251032,0.000069832146,0.000034536075,0.00027156185],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9967735,0.00010260401,0.0015711888,0.00029225857,0.0008291543,0.00043132398],"domain_scores_gemma":[0.99568594,0.00036254324,0.0009586259,0.00029954853,0.0006937812,0.0019995493],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0006136741,0.0002659545,0.0008616715,0.00021885887,0.0000875763,0.0000026423083,0.00022039241,0.00010029295,0.0013471305],"category_scores_gemma":[0.014074489,0.00019761246,0.0002691458,0.00045002348,0.00009655775,0.000066474306,0.000016689108,0.0004976453,0.000015360474],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005296584,0.00041736543,0.7373189,0.0001368945,0.00048287582,0.00029263992,0.0021512457,0.00010285726,0.00026666798,0.00019567803,0.24667822,0.011427008],"study_design_scores_gemma":[0.010795013,0.013711703,0.9209926,0.0059307315,0.0009940657,0.0000102211325,0.0009820572,0.00044251318,0.00038813596,0.00022183465,0.045114253,0.00041687442],"about_ca_topic_score_codex":0.0012064079,"about_ca_topic_score_gemma":0.0010874465,"teacher_disagreement_score":0.20156397,"about_ca_system_score_codex":0.00017028277,"about_ca_system_score_gemma":0.00094159547,"threshold_uncertainty_score":0.9995658},"labels":[],"label_agreement":null},{"id":"W3024447389","doi":"10.1017/cem.2020.131","title":"LO77: Performance enhancing psychological skills in clinical simulation","year":2020,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Cognition; Anxiety; Cognitive load; Medicine; Protocol (science); Physical therapy; Baseline (sea); Applied psychology; Physical medicine and rehabilitation; Clinical psychology; Psychology; Psychiatry","score_opus":0.17652984846059674,"score_gpt":0.46518334412574686,"score_spread":0.2886534956651501,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3024447389","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9796422,0.0010426205,0.0005395365,0.011733953,0.0039149644,0.00014580307,0.0000018094895,0.00000830179,0.0029708205],"genre_scores_gemma":[0.9931344,0.00087968854,0.00021033193,0.0026267935,0.0029545783,9.989218e-7,0.0000060433317,0.000015473102,0.00017169552],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9971329,0.000098744706,0.0019050079,0.00022329172,0.000303454,0.00033662643],"domain_scores_gemma":[0.9954951,0.0016985267,0.00037784164,0.00016756519,0.000311361,0.0019496424],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0013486107,0.00014769948,0.0006628887,0.00018429644,0.000042806736,0.0000025526379,0.00015833598,0.00017148082,0.0042786347],"category_scores_gemma":[0.070928164,0.000107786196,0.00017689203,0.0005175454,0.00012858157,0.00007407547,0.000007798139,0.0007560628,0.000054637872],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00022522509,0.00015372306,0.95162517,0.00004147481,0.000045327484,0.0005986033,0.0008894983,0.0011152629,0.00007859156,0.000029653398,0.020419933,0.024777517],"study_design_scores_gemma":[0.003616563,0.0043901424,0.9734175,0.0023641859,0.00013560275,0.000046514386,0.00019034685,0.002803148,0.000020817432,0.00009478147,0.012760078,0.00016031916],"about_ca_topic_score_codex":0.0002947378,"about_ca_topic_score_gemma":0.0008483096,"teacher_disagreement_score":0.06957955,"about_ca_system_score_codex":0.00006835196,"about_ca_system_score_gemma":0.00048650088,"threshold_uncertainty_score":0.99663156},"labels":[],"label_agreement":null},{"id":"W3024893209","doi":"","title":"FM-MAP: A Novel In-Training Examination Predicts Success on Family Medicine Certification Examination.","year":2017,"lang":"en","type":"article","venue":"PubMed","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Toronto","funders":"","keywords":"Certification; Test (biology); Medicine; Medical education; Family medicine; Final examination; Physical examination; Psychology; Internal medicine; Management","score_opus":0.1615150107733623,"score_gpt":0.3512279207919043,"score_spread":0.18971291001854199,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3024893209","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9482419,0.00012655774,0.000655109,0.008352879,0.0009558398,0.0012471615,0.000010928566,0.00010638988,0.040303264],"genre_scores_gemma":[0.9964138,0.000057049714,0.00012156346,0.00065098284,0.00059585925,0.00069081277,0.00011205957,0.000026565138,0.0013312873],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9979534,0.00007098397,0.0005128654,0.00047440754,0.0005888737,0.0003994592],"domain_scores_gemma":[0.99479544,0.0036700165,0.00033603897,0.0007260512,0.00019480065,0.00027762496],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0023096965,0.00018130481,0.0003764985,0.0002527123,0.00015900191,0.00006632101,0.00024924832,0.00017430817,0.000025881069],"category_scores_gemma":[0.08403726,0.00015229071,0.00005203524,0.00014700281,0.0002288887,0.00019343533,0.000046441954,0.00028174988,0.000028017643],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009327555,0.0004620045,0.04629147,0.000041095616,0.000027517717,0.000031593387,0.00077214465,0.000012107697,0.0002645151,0.00096359896,0.00064763165,0.950393],"study_design_scores_gemma":[0.0038200922,0.00014288988,0.99257874,0.0008829063,0.000066144865,0.0000075218363,0.0001846748,0.0007676142,0.00029310616,0.00014705991,0.0009701918,0.0001390314],"about_ca_topic_score_codex":0.00018304716,"about_ca_topic_score_gemma":0.000030107722,"teacher_disagreement_score":0.950254,"about_ca_system_score_codex":0.00016506745,"about_ca_system_score_gemma":0.000053732627,"threshold_uncertainty_score":0.9236783},"labels":[],"label_agreement":null},{"id":"W3025240286","doi":"10.1111/jgs.16559","title":"Clinical Judgment Is Paramount When Performing Cognitive Screening during <scp>COVID</scp> ‐19","year":2020,"lang":"en","type":"letter","venue":"Journal of the American Geriatrics Society","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Dalhousie University; Université de Montréal; Concordia University","funders":"","keywords":"Medicine; Coronavirus disease 2019 (COVID-19); 2019-20 coronavirus outbreak; Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); Cognition; Clinical judgment; Betacoronavirus; Pandemic; MEDLINE; Coronavirus Infections; Intensive care medicine; Virology; Psychiatry; Medical physics; Internal medicine; Infectious disease (medical specialty); Disease","score_opus":0.04627235303054417,"score_gpt":0.3480727927678365,"score_spread":0.3018004397372923,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3025240286","genre_codex":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5139067,0.001683401,0.0015213514,0.479756,0.0019938457,0.0007029892,0.00014411392,0.000072634764,0.00021892185],"genre_scores_gemma":[0.0263019,0.0057014506,0.012042922,0.93365103,0.0210055,0.000011813362,0.000036178997,0.00015660666,0.0010926172],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99384826,0.00046626345,0.0022861594,0.00066516665,0.0019204535,0.0008137041],"domain_scores_gemma":[0.9641773,0.027706927,0.005899572,0.0006281402,0.0007091907,0.0008788838],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0018028239,0.0006030794,0.0023122246,0.000108045,0.000406876,0.00013761077,0.00077165005,0.0005928537,0.000054412703],"category_scores_gemma":[0.062249027,0.0004157262,0.004315267,0.0007308005,0.00080130796,0.00009654223,0.000525658,0.007901648,0.00003196759],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010722069,0.00014030552,0.09688388,0.00016611433,0.0014550661,0.0006807154,0.0016812708,0.0000116892015,0.000012193331,2.16804e-7,0.8953302,0.0035311382],"study_design_scores_gemma":[0.011337775,0.0039714607,0.12418344,0.007835108,0.0128845535,0.0025598097,0.00943124,0.00062825216,0.00010249894,0.0003711942,0.82584405,0.0008505992],"about_ca_topic_score_codex":0.00012527365,"about_ca_topic_score_gemma":5.750285e-7,"teacher_disagreement_score":0.48760483,"about_ca_system_score_codex":0.00057996647,"about_ca_system_score_gemma":0.00154523,"threshold_uncertainty_score":0.9998295},"labels":[],"label_agreement":null},{"id":"W3025260314","doi":"10.3138/jvme.0818-097r1","title":"Use of Inductive, Problem-Based Clinical Reasoning Enhances Diagnostic Accuracy in Final-Year Veterinary Students","year":2020,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Medical Research Council; European Commission; Leverhulme Trust; Wellcome Trust; National Institutes of Health; National Science Foundation","keywords":"Inductive reasoning; Inductive method; Diagnostic test; Medicine; Test (biology); Cognition; Medical education; Mathematics education; Artificial intelligence; Psychology; Veterinary medicine; Computer science; Teaching method","score_opus":0.22277068814463055,"score_gpt":0.48854279336710704,"score_spread":0.2657721052224765,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3025260314","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9922026,0.00035588874,0.00019403615,0.006051578,0.00081363425,0.00030963478,0.0000028096426,0.000012675525,0.000057187117],"genre_scores_gemma":[0.9869669,0.00097006763,0.008824113,0.0023908168,0.000778266,0.000016608657,0.000013653381,0.000021864013,0.000017722305],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9961006,0.0004612041,0.001779356,0.00028390353,0.0011413851,0.0002335561],"domain_scores_gemma":[0.97528934,0.021843191,0.0011539281,0.00025614654,0.00047295247,0.0009844273],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0015830569,0.00018295537,0.00077619817,0.00017322527,0.000028577797,0.000026579568,0.0003104525,0.00024366209,0.00028640963],"category_scores_gemma":[0.3279165,0.00014968243,0.00025115872,0.0003925796,0.00021178903,0.00027035648,0.00010278401,0.00096230267,0.000012316022],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0058554094,0.0093722325,0.83256775,0.00040395098,0.00015901556,0.0013837871,0.0008451651,0.000036807374,0.001362768,0.000029370329,0.004715211,0.1432685],"study_design_scores_gemma":[0.0054539475,0.01988372,0.9478834,0.017854804,0.0002163832,0.0005805943,0.00044323085,0.00035783733,0.00022884014,0.000072144554,0.0067974865,0.00022759434],"about_ca_topic_score_codex":0.00004432377,"about_ca_topic_score_gemma":0.0000011834783,"teacher_disagreement_score":0.32633343,"about_ca_system_score_codex":0.000082534585,"about_ca_system_score_gemma":0.0026613935,"threshold_uncertainty_score":0.67774475},"labels":[],"label_agreement":null},{"id":"W3025342191","doi":"10.1017/cem.2020.170","title":"MP22: Using galvanic skin response to identify resuscitation expertise in a pulmonary embolism simulation exercise","year":2020,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Medicine; Cognitive load; Cognition; Pulmonary embolism; Schema (genetic algorithms); Cardiopulmonary resuscitation; Descriptive statistics; Physical therapy; Medical emergency; Resuscitation; Emergency medicine; Psychiatry; Computer science; Cardiology; Statistics","score_opus":0.14849545693951754,"score_gpt":0.439227725576182,"score_spread":0.29073226863666446,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3025342191","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97200596,0.0030561488,0.0012172178,0.02086972,0.002347205,0.0003153888,0.0000055611795,0.000008165358,0.00017461997],"genre_scores_gemma":[0.9969747,0.0002029964,0.00028010437,0.0012009423,0.0011925153,0.0000032752205,0.000007763147,0.000029244055,0.00010847051],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9975792,0.00016224197,0.0012402727,0.00023813671,0.00043289468,0.0003472125],"domain_scores_gemma":[0.9959205,0.0009900901,0.0002874244,0.00019656455,0.00044412073,0.0021613413],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011821629,0.00016910529,0.00054444594,0.0006165006,0.0000658508,0.000006240441,0.00014688804,0.00011630566,0.0013431563],"category_scores_gemma":[0.07651121,0.00014827208,0.00012994942,0.001006303,0.00005307584,0.00012749457,0.000011262411,0.00032996878,0.000025166168],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.03477714,0.0007445579,0.27856457,0.0006239397,0.0003938555,0.022703169,0.11819196,0.101786084,0.15639648,0.00010902412,0.12754932,0.1581599],"study_design_scores_gemma":[0.0057853614,0.0022153493,0.9197675,0.015134958,0.0005293148,0.00010903025,0.003612285,0.042952776,0.00027548362,0.000368381,0.008742321,0.0005072722],"about_ca_topic_score_codex":0.0034177452,"about_ca_topic_score_gemma":0.0011614024,"teacher_disagreement_score":0.6412029,"about_ca_system_score_codex":0.000261943,"about_ca_system_score_gemma":0.000985751,"threshold_uncertainty_score":0.9995698},"labels":[],"label_agreement":null},{"id":"W3025479060","doi":"10.22374/cjgim.v15i2.442","title":"COVID-19 and Canadian General Internists","year":2020,"lang":"en","type":"article","venue":"Canadian Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Coronavirus disease 2019 (COVID-19); Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); 2019-20 coronavirus outbreak; Family medicine; Virology; Internal medicine; Infectious disease (medical specialty); Outbreak","score_opus":0.0544627132220234,"score_gpt":0.35146421109551607,"score_spread":0.29700149787349267,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3025479060","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.73608077,0.0040656934,0.00061821524,0.24948375,0.0015264215,0.00016793024,0.00004265403,0.000013621781,0.008000943],"genre_scores_gemma":[0.891685,0.00020254824,0.00067003444,0.10188635,0.003598604,0.000001313743,0.000015722919,0.000026946567,0.0019134543],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99818605,0.000068119625,0.00073679845,0.00024348758,0.00029404,0.00047147818],"domain_scores_gemma":[0.98216075,0.0005797092,0.00022554802,0.00015055545,0.0002711029,0.016612327],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00044529382,0.00021752747,0.0006384655,0.00040129758,0.00009647011,0.000036977133,0.0002458243,0.00013318987,0.0023239902],"category_scores_gemma":[0.042437617,0.00016455998,0.00013334263,0.00019223584,0.00039536555,0.00006171731,0.000019133775,0.00062253134,0.000026380621],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004114493,0.000022742179,0.23920625,0.00009976231,0.00037144232,0.019545099,0.003075143,0.00006242839,0.00094436575,0.0012228476,0.71733683,0.017701626],"study_design_scores_gemma":[0.010288663,0.006747881,0.08701342,0.0034933758,0.0006377278,0.016864426,0.0007170835,0.002087114,0.00033297244,0.0008578793,0.87038386,0.0005756083],"about_ca_topic_score_codex":0.57542187,"about_ca_topic_score_gemma":0.39487818,"teacher_disagreement_score":0.18054366,"about_ca_system_score_codex":0.00051306677,"about_ca_system_score_gemma":0.0057817735,"threshold_uncertainty_score":0.9998545},"labels":[],"label_agreement":null},{"id":"W3025977132","doi":"10.1097/acm.0b013e3182816880","title":"In the Real World, Faster Diagnoses Are Not Necessarily More Accurate","year":2013,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University; Hamilton Health Sciences","funders":"","keywords":"Context (archaeology); Medical diagnosis; Cognition; Test (biology); Cognitive psychology; Psychology; Agency (philosophy); Empirical research; Social psychology; Medicine; Epistemology; Psychiatry","score_opus":0.07356878552081053,"score_gpt":0.3905377803136941,"score_spread":0.3169689947928836,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3025977132","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.028437061,0.0017712393,0.00000967411,0.9590605,0.0015925389,0.0012769513,0.000035499987,0.00009320952,0.0077233096],"genre_scores_gemma":[0.034079652,0.0069646453,0.000018756778,0.8959099,0.024881838,0.00034903304,0.00058501674,0.000104846666,0.037106305],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9944766,0.00037187344,0.0015582913,0.00096456695,0.0016302727,0.0009984],"domain_scores_gemma":[0.9572873,0.04029526,0.0007662574,0.0011860783,0.00018144191,0.00028365792],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0014237269,0.00074533356,0.0018249357,0.00051497685,0.0000796434,0.000025976504,0.0009670841,0.0020098982,0.0018710439],"category_scores_gemma":[0.047403876,0.00040634145,0.00020240972,0.00078905094,0.00076039595,0.00013838377,0.00015489578,0.011252331,0.00069603126],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000610095,0.0000248566,0.019509636,0.0001753643,0.000072921495,0.002736915,0.0014571813,8.1635966e-7,0.000007951818,0.000041838597,0.97453827,0.0013732555],"study_design_scores_gemma":[0.003466044,0.00029608724,0.07788803,0.017754605,0.00072455563,0.0001992359,0.0010548767,0.000038756938,0.000015714257,0.0005798518,0.89748454,0.0004977339],"about_ca_topic_score_codex":0.0019502132,"about_ca_topic_score_gemma":0.000050510404,"teacher_disagreement_score":0.07705374,"about_ca_system_score_codex":0.00013633985,"about_ca_system_score_gemma":0.00020548668,"threshold_uncertainty_score":0.9998388},"labels":[],"label_agreement":null},{"id":"W3029550683","doi":"10.1111/tct.13178","title":"Breaking good news: a medical student perspective","year":2020,"lang":"en","type":"letter","venue":"The Clinical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Harm; Perspective (graphical); Medical diagnosis; Curriculum; Context (archaeology); Foundation (evidence); Medical education; Psychology; Medicine; Bridging (networking); Causality (physics); Social psychology; Pedagogy; Computer science; Pathology","score_opus":0.09389692026702968,"score_gpt":0.4638885493567049,"score_spread":0.36999162908967526,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3029550683","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0019584317,0.000882778,0.00007167597,0.9587374,0.001489508,0.00077867595,0.000015280124,0.00034199367,0.0357243],"genre_scores_gemma":[0.011327971,0.00060789695,0.00015422123,0.93598765,0.03665518,0.000063053034,0.00010220764,0.00014584391,0.014956002],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9931199,0.0012110951,0.0016478674,0.0012121536,0.002114825,0.00069414754],"domain_scores_gemma":[0.9409805,0.05549598,0.000608772,0.0016736487,0.00029471968,0.00094641413],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity","insufficient_payload"],"category_scores_codex":[0.0022708906,0.0005608084,0.0018191859,0.000052131414,0.00013140804,0.0000733234,0.0009905268,0.0021532974,0.0036280511],"category_scores_gemma":[0.19957758,0.00032790305,0.0014232874,0.00025811876,0.00093269686,0.000025149517,0.0005787688,0.013699587,0.0020699007],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012509889,0.00065473915,0.0106655015,0.000024851312,0.0008562177,0.0038691137,0.00045725406,9.7797496e-8,9.3050495e-8,0.00014241847,0.9670114,0.016193248],"study_design_scores_gemma":[0.0033922144,0.0011352933,0.0113990735,0.0029670773,0.0018755327,0.0002536111,0.0007414101,0.00007513272,3.9883776e-7,0.001872093,0.9757853,0.00050282077],"about_ca_topic_score_codex":0.00031006607,"about_ca_topic_score_gemma":0.000013100813,"teacher_disagreement_score":0.1973067,"about_ca_system_score_codex":0.00020667564,"about_ca_system_score_gemma":0.0009801605,"threshold_uncertainty_score":0.9999173},"labels":[],"label_agreement":null},{"id":"W3029841721","doi":"10.1097/acm.0000000000003524","title":"From Theory to Practice: The Application of Cognitive Load Theory to the Practice of Medicine","year":2020,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":153,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Parallels; Cognition; Cognitive load; Cognitive architecture; Learning theory; Process (computing); LIDA; Cognitive science; Psychology; Cognitive psychology; Computer science; Psychiatry","score_opus":0.033637478782433484,"score_gpt":0.4181533337642616,"score_spread":0.3845158549818281,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3029841721","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.019024039,0.005944791,0.12557858,0.83512235,0.0004442453,0.0023010895,0.000044109936,0.00006234556,0.011478432],"genre_scores_gemma":[0.81852937,0.0009987716,0.0005683194,0.17637542,0.0030058837,0.000104859944,0.00004821863,0.000031148422,0.00033802446],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9959853,0.0008615439,0.001098765,0.0004925448,0.001287272,0.00027453233],"domain_scores_gemma":[0.8223445,0.17485395,0.0007199242,0.0006150169,0.00096638786,0.0005002418],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0074399337,0.00024038652,0.0007691277,0.000062114996,0.00007298892,0.000002190042,0.0005079356,0.0002236431,0.00040855954],"category_scores_gemma":[0.7115709,0.000114283124,0.00007121793,0.0008812561,0.00087170955,0.00007834396,0.00017335951,0.0011562302,0.00017067294],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.020188343,0.00024160312,0.0011274966,0.00011563238,0.0011200703,0.00004809329,0.20648582,0.00003545853,0.010476627,0.050687205,0.25268167,0.45679197],"study_design_scores_gemma":[0.01085904,0.0072406386,0.005900419,0.012765848,0.008185919,0.00014618182,0.217688,0.00033088727,0.0035452289,0.02168241,0.7112453,0.00041015176],"about_ca_topic_score_codex":0.000599831,"about_ca_topic_score_gemma":0.0000028056652,"teacher_disagreement_score":0.7995053,"about_ca_system_score_codex":0.00004670717,"about_ca_system_score_gemma":0.0002417947,"threshold_uncertainty_score":0.5023309},"labels":[],"label_agreement":null},{"id":"W3031249971","doi":"10.1096/fasebj.2018.32.1_supplement.507.23","title":"Creating Instruction to Optimize Learning of the Anatomical Sciences","year":2018,"lang":"en","type":"article","venue":"The FASEB Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; Toronto Rehabilitation Institute; University of Toronto; University Health Network","funders":"","keywords":"Cognition; Test (biology); Task (project management); Basic science; Psychology; Medicine; Medical education; Cognitive psychology; Neuroscience","score_opus":0.03079937232272738,"score_gpt":0.34076771976446113,"score_spread":0.30996834744173374,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3031249971","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99190414,0.000031849144,0.0005197838,0.002521365,0.00035424775,0.00006208031,3.916587e-7,0.000008472391,0.0045976923],"genre_scores_gemma":[0.99633616,0.000014073038,0.0023844193,0.00044359616,0.00060597295,6.135174e-7,1.248829e-7,0.0000042496913,0.00021081524],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9991893,0.00009872153,0.00021369355,0.00007833962,0.0002830771,0.00013687384],"domain_scores_gemma":[0.9977058,0.0018389398,0.00013120516,0.00010729268,0.000113396294,0.000103352795],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010647149,0.000049395985,0.000120110744,0.00003575127,0.0004721971,0.000029494902,0.00014920355,0.00002967314,0.00015250433],"category_scores_gemma":[0.026541924,0.000022214817,0.00008118289,0.00024793783,0.00041145424,0.0000270501,0.00005530371,0.00032792834,0.000016748269],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00091747835,0.00026280148,0.73191255,0.000011921425,0.00016909858,0.000021948788,0.005475859,0.005293947,0.07397431,0.0033153887,0.0034058318,0.17523885],"study_design_scores_gemma":[0.006807849,0.0052929013,0.644689,0.008584515,0.0005822078,0.005332036,0.008445126,0.028287126,0.2876187,0.002646219,0.0013250926,0.00038921603],"about_ca_topic_score_codex":0.000028817009,"about_ca_topic_score_gemma":0.0000016395808,"teacher_disagreement_score":0.21364439,"about_ca_system_score_codex":0.000023949622,"about_ca_system_score_gemma":0.00012851633,"threshold_uncertainty_score":0.9816579},"labels":[],"label_agreement":null},{"id":"W3033827820","doi":"10.20429/amtp.2014.18","title":"What Can Business Leaders Learn from Medical Misdiagnoses?","year":2014,"lang":"en","type":"article","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Computer science; Business; Political science; Public relations","score_opus":0.0338994187866585,"score_gpt":0.3361544289093267,"score_spread":0.3022550101226682,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3033827820","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8888885,0.0005610464,0.0009552403,0.09521485,0.0017742094,0.00016920868,0.0000035922328,0.0002360795,0.012197289],"genre_scores_gemma":[0.9672664,0.0010521003,0.0005457702,0.024601066,0.0011657877,0.000012872299,0.00012424751,0.000030306297,0.0052014706],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99827063,0.000064132204,0.00031796095,0.00037039968,0.00066272676,0.00031416753],"domain_scores_gemma":[0.9826191,0.015999151,0.00005346889,0.00043967963,0.00015407478,0.0007345362],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00038272096,0.00015831648,0.0004051692,0.00004593875,0.000047093257,0.00005595812,0.0001386812,0.00025997125,0.0058411034],"category_scores_gemma":[0.115866326,0.00011616466,0.00010199846,0.0001917928,0.0002542129,0.00007485264,0.000051216553,0.00031504757,0.0005508261],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00022344293,0.0012232893,0.24454981,0.000051188174,0.00022974532,0.00044043193,0.0009437161,0.000018366014,0.00007465769,0.0015211824,0.17546114,0.575263],"study_design_scores_gemma":[0.01131586,0.00060905184,0.82961637,0.009784914,0.00047815824,0.00013047608,0.0034306827,0.0021874525,0.0010205883,0.0020202836,0.13863724,0.0007689177],"about_ca_topic_score_codex":0.0018775482,"about_ca_topic_score_gemma":0.000505031,"teacher_disagreement_score":0.58506656,"about_ca_system_score_codex":0.000028693103,"about_ca_system_score_gemma":0.00022215357,"threshold_uncertainty_score":0.9950677},"labels":[],"label_agreement":null},{"id":"W3035608132","doi":"10.1017/cem.2020.406","title":"“Once more unto the breach, dear friends” – Reflections on the novel coronavirus disease and emergency medicine in Canada","year":2020,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Canadian Association of Emergency Physicians; University of Manitoba","funders":"","keywords":"Content (measure theory); Action (physics); Coronavirus; Coronavirus disease 2019 (COVID-19); Medicine; Internet privacy; Disease; Computer science; Infectious disease (medical specialty); Internal medicine","score_opus":0.21566859563785803,"score_gpt":0.41755426955158864,"score_spread":0.2018856739137306,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3035608132","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5595428,0.009997567,0.00009348767,0.42002308,0.0052986746,0.0003983808,0.000058529004,0.0000079567235,0.0045795757],"genre_scores_gemma":[0.9886001,0.0021723902,0.000011906854,0.006760356,0.0020809276,0.0000087530325,0.000012626823,0.000026945429,0.00032603738],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99753493,0.00010111356,0.0011326898,0.00026317773,0.00053795835,0.00043011646],"domain_scores_gemma":[0.9948942,0.0016728174,0.00034661268,0.00033290195,0.00030888655,0.0024445902],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00070130493,0.00024849342,0.0005179348,0.00012748195,0.00023813827,0.000003027915,0.000306282,0.000056613026,0.0049316813],"category_scores_gemma":[0.032691985,0.00012548883,0.00009353379,0.00078386854,0.0002879651,0.000054040785,0.000018580102,0.000817257,0.0000039059496],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009043194,0.00015856665,0.4439771,0.00014861868,0.0006327157,0.0015437817,0.019014487,0.0004176717,0.0012193392,0.0038042902,0.52091897,0.007260154],"study_design_scores_gemma":[0.003196754,0.0016963026,0.8373844,0.0034976527,0.00083096925,0.00011146549,0.015819773,0.0007527303,0.000016482214,0.0005381525,0.13581386,0.00034146087],"about_ca_topic_score_codex":0.82050675,"about_ca_topic_score_gemma":0.9291415,"teacher_disagreement_score":0.4290573,"about_ca_system_score_codex":0.00039163834,"about_ca_system_score_gemma":0.004827025,"threshold_uncertainty_score":0.99597794},"labels":[],"label_agreement":null},{"id":"W3041967012","doi":"10.1515/dx-2020-0079","title":"<i>Sapere aude</i> in the diagnostic process","year":2020,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Process (computing); Mathematics; Computer science; Programming language","score_opus":0.03937130185738696,"score_gpt":0.3404461604698064,"score_spread":0.30107485861241945,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3041967012","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8887279,0.0013017829,0.000014198374,0.10423142,0.00009855984,0.0005876725,0.000014310084,0.00009039334,0.004933742],"genre_scores_gemma":[0.95331013,0.0010001609,0.00006409104,0.044729434,0.00035576575,0.00047039764,0.000029531142,0.000021351008,0.000019149502],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9985745,0.00007257698,0.00032842602,0.00034366752,0.00037796167,0.0003028699],"domain_scores_gemma":[0.9690877,0.030220546,0.000061901905,0.00029696547,0.000063989435,0.00026888202],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00025578254,0.00016209378,0.00032991773,0.000032396496,0.000050636037,0.000035539215,0.00023130969,0.00010005335,0.0003451612],"category_scores_gemma":[0.1206444,0.00010565961,0.00012872269,0.0005600277,0.000078034864,0.000049482813,0.000034003675,0.00036652092,0.00038487976],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000050896917,0.00061911886,0.9366869,0.000065035645,0.000019426318,0.0004437506,0.0019563362,0.000030213907,0.0000058424957,0.00017220118,0.046098262,0.013852065],"study_design_scores_gemma":[0.003592105,0.0010241709,0.9524464,0.002207094,0.00028236557,0.000057993664,0.0015241359,0.0002818173,0.0009427816,0.00062171853,0.03663553,0.00038387647],"about_ca_topic_score_codex":0.00006244755,"about_ca_topic_score_gemma":0.000019139463,"teacher_disagreement_score":0.12038861,"about_ca_system_score_codex":0.000019171197,"about_ca_system_score_gemma":0.000084870124,"threshold_uncertainty_score":0.8867628},"labels":[],"label_agreement":null},{"id":"W3044245418","doi":"10.1016/j.jtcvs.2020.06.146","title":"Are maladaptive brain changes the reason for burnout and medical error?","year":2020,"lang":"en","type":"article","venue":"Journal of Thoracic and Cardiovascular Surgery","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Boniface Hospital; University of Manitoba","funders":"Pfizer Canada","keywords":"Burnout; Psychology; Clinical psychology; Cognitive psychology","score_opus":0.05866353813175967,"score_gpt":0.3414684489825735,"score_spread":0.28280491085081383,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3044245418","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.76390827,0.10289151,0.0038382928,0.1285396,0.0004082491,0.00028891742,0.0000121460735,0.000019848201,0.00009314398],"genre_scores_gemma":[0.98357683,0.008079455,0.00022253662,0.006009581,0.0020289095,0.000008242716,0.0000026101654,0.000021724634,0.00005011242],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9983953,0.00020076682,0.00035612885,0.00018736973,0.0006745229,0.00018586314],"domain_scores_gemma":[0.9888681,0.009971231,0.00026582374,0.00016887304,0.00021266489,0.0005132865],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0036271096,0.0001338855,0.0008988849,0.000050121515,0.000081341175,0.000029210545,0.00006751773,0.00014037508,0.000018115123],"category_scores_gemma":[0.06707296,0.00007781824,0.0008477554,0.00011893564,0.00014684242,0.000041123614,0.000046939578,0.00033516958,0.0000012048744],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016691502,0.00009272443,0.057652887,0.00033744014,0.0034269604,0.0017666039,0.0006398358,0.000013686083,0.000014569601,0.000033618777,0.04209017,0.89226234],"study_design_scores_gemma":[0.0061531034,0.0022415442,0.5382121,0.020771055,0.010639814,0.017708715,0.008749374,0.0021583934,0.00068192807,0.00081937393,0.39080757,0.0010570139],"about_ca_topic_score_codex":0.000006341505,"about_ca_topic_score_gemma":0.0000019155511,"teacher_disagreement_score":0.8912054,"about_ca_system_score_codex":0.000010641719,"about_ca_system_score_gemma":0.00012775374,"threshold_uncertainty_score":0.94078547},"labels":[],"label_agreement":null},{"id":"W3046224457","doi":"10.1007/s40037-020-00603-2","title":"A&amp;nbsp;layered analysis of self-explanation and structured reflection to support clinical reasoning in medical students","year":2020,"lang":"en","type":"article","venue":"Perspectives on Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Sherbrooke","funders":"Université de Sherbrooke","keywords":"Reflection (computer programming); Psychological intervention; Context (archaeology); Curriculum; Engineering ethics; Epistemology; Computer science; Psychology; Pedagogy; Engineering","score_opus":0.04058826536004433,"score_gpt":0.4744902999609369,"score_spread":0.43390203460089255,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3046224457","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9763288,0.00012188876,0.00067773997,0.021470578,0.00027762912,0.00035271453,0.0000052229348,0.000058645317,0.0007067632],"genre_scores_gemma":[0.99095744,0.0005649104,0.0018193031,0.006025003,0.00042961186,0.000027895467,0.00012223102,0.000015328771,0.000038286453],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99614817,0.0002580604,0.00085388776,0.00064214866,0.001890956,0.00020676969],"domain_scores_gemma":[0.9952231,0.0026762725,0.00020965547,0.0002559869,0.00027549974,0.0013595088],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0018573322,0.00016177996,0.0006862207,0.00043231106,0.000039779512,0.000022899741,0.00019200945,0.0003306304,0.0010947764],"category_scores_gemma":[0.17575775,0.00014149962,0.00015314443,0.0014380757,0.000111137204,0.000049258648,0.00006965757,0.0005959143,0.000019669893],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011001941,0.0038466894,0.9025976,0.00004536555,0.0008004539,0.00002406472,0.019985456,0.000026288875,0.000033141892,0.0011284013,0.0034665018,0.066945866],"study_design_scores_gemma":[0.0020426894,0.0012386568,0.9874078,0.0005841802,0.0006042771,0.000013033855,0.0050116647,0.0022400585,0.000019478628,0.00006011512,0.0006310649,0.00014695502],"about_ca_topic_score_codex":0.00017802688,"about_ca_topic_score_gemma":0.00019981695,"teacher_disagreement_score":0.17390043,"about_ca_system_score_codex":0.00017558788,"about_ca_system_score_gemma":0.0011732992,"threshold_uncertainty_score":0.9998184},"labels":[],"label_agreement":null},{"id":"W3047050422","doi":"10.1111/medu.14304","title":"Warning bells: How clinicians leverage their discomfort to manage moments of uncertainty","year":2020,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":49,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Grounded theory; Psychology; Leverage (statistics); Cognition; Applied psychology; Medical education; Medicine; Qualitative research; Computer science; Psychiatry","score_opus":0.03239432769441617,"score_gpt":0.3634797715138036,"score_spread":0.3310854438193874,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3047050422","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8521012,0.00012155343,0.0020429892,0.14284998,0.0006235513,0.00034278273,0.000008752077,0.000049080933,0.0018600883],"genre_scores_gemma":[0.96802783,0.00010485232,0.000419303,0.02978717,0.00058691844,0.000027716278,0.00012062597,0.000016316286,0.00090924103],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9985569,0.000056635952,0.0003756506,0.00028539894,0.00052458374,0.00020080271],"domain_scores_gemma":[0.9968452,0.0020964975,0.00011507315,0.00023878866,0.000060519662,0.00064395886],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00043330994,0.000118959106,0.00035132657,0.00004923669,0.000038249003,0.000011007471,0.000151206,0.00011792988,0.00051285006],"category_scores_gemma":[0.07099285,0.0000893233,0.00011240491,0.0002777019,0.00007391032,0.000034044166,0.00006889016,0.00027300502,0.000039359405],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00031650718,0.0023037018,0.101189375,0.00025132962,0.00015176761,0.000030711904,0.009557704,0.00010724832,0.00017072013,0.0003202295,0.18158792,0.7040128],"study_design_scores_gemma":[0.007694308,0.004441156,0.31523526,0.014150725,0.0006681772,0.000046282654,0.013907454,0.017379427,0.0013904218,0.0010234338,0.6230367,0.0010266657],"about_ca_topic_score_codex":0.000118015785,"about_ca_topic_score_gemma":0.000009937942,"teacher_disagreement_score":0.7029861,"about_ca_system_score_codex":0.00004739495,"about_ca_system_score_gemma":0.00055005716,"threshold_uncertainty_score":0.93683255},"labels":[],"label_agreement":null},{"id":"W3047936085","doi":"10.1007/s12028-020-01071-6","title":"Cognitive Biases to Consider in Non-convulsive Status Epilepticus Presenting in a Hematologic Malignancy","year":2020,"lang":"en","type":"article","venue":"Neurocritical Care","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University Health Network; University of Toronto","funders":"","keywords":"Medicine; Status epilepticus; Neurology; Hematologic malignancy; Cognition; Malignancy; Epilepsy; Intensive care medicine; Pediatrics; Anesthesia; Psychiatry; Internal medicine","score_opus":0.06705451192476251,"score_gpt":0.3876192612131569,"score_spread":0.32056474928839435,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3047936085","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98981947,0.00031798787,0.000029758323,0.0059435293,0.00009314704,0.00071614044,0.000059425765,0.000055161425,0.002965354],"genre_scores_gemma":[0.98289484,0.000023742894,0.00022219431,0.016611358,0.000091789625,0.000081721846,0.000037244903,0.000024947029,0.000012148219],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99799424,0.00010797006,0.00047117344,0.0005540461,0.00026736438,0.0006051862],"domain_scores_gemma":[0.9785412,0.020379893,0.000027681332,0.00013787005,0.00023862784,0.000674709],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.000056713,0.00018721528,0.00056534394,0.000095985895,0.000025830408,0.000025446005,0.000063930696,0.0001089396,0.00014592426],"category_scores_gemma":[0.29504323,0.00017174317,0.000094677875,0.0004315704,0.0001099833,0.000040187846,0.00011915162,0.0004712169,0.00016806289],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007508192,0.0005616415,0.9569928,0.0004123076,0.000017512255,0.028538665,0.0056913993,0.00002836931,0.00079136994,0.00041971344,0.001305651,0.0044897464],"study_design_scores_gemma":[0.010612006,0.003529345,0.9373883,0.014909137,0.00021429206,0.00026771353,0.021606723,0.0008142309,0.009416378,0.00039797695,0.0001486252,0.0006952441],"about_ca_topic_score_codex":0.00011596315,"about_ca_topic_score_gemma":0.00006246207,"teacher_disagreement_score":0.29498652,"about_ca_system_score_codex":0.000039737806,"about_ca_system_score_gemma":0.00014001274,"threshold_uncertainty_score":0.71089494},"labels":[],"label_agreement":null},{"id":"W3048342385","doi":"10.1097/acm.0000000000003634","title":"“I Was Worried About the Patient, but I Wasn’t Feeling Worried”: How Physicians Judge Their Comfort in Settings of Uncertainty","year":2020,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":35,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Feeling; Grounded theory; Narrative; Psychology; Coding (social sciences); Qualitative research; Medical education; Applied psychology; Social psychology; Medicine; Sociology","score_opus":0.044038855035464446,"score_gpt":0.31579527680966446,"score_spread":0.2717564217742,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3048342385","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.82813287,0.0032723625,0.00018557666,0.1654619,0.000334108,0.0006865388,0.000028289576,0.00008661839,0.0018117211],"genre_scores_gemma":[0.9406362,0.0014588027,0.00006788334,0.05636908,0.001038193,0.000035765646,0.00009370593,0.000041555875,0.00025878655],"study_design_codex":"not_applicable","study_design_gemma":"qualitative","domain_scores_codex":[0.99707764,0.00015714204,0.0010010453,0.0005190811,0.00069888216,0.000546221],"domain_scores_gemma":[0.98821485,0.010306115,0.00048533548,0.00046263463,0.00016684287,0.0003642021],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010051071,0.00033128212,0.0010454918,0.00008573882,0.00008180921,0.0000061394885,0.00038260766,0.00031549748,0.00009612097],"category_scores_gemma":[0.056013867,0.00019534104,0.00014887717,0.00071986794,0.0006106822,0.00006803903,0.00013110899,0.0017711101,0.000013514147],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0037572652,0.0004230608,0.15600163,0.00078057003,0.0007527179,0.00040677376,0.1845974,0.001531822,0.023338217,0.0033970638,0.3588871,0.26612636],"study_design_scores_gemma":[0.05548829,0.011909265,0.09041764,0.10377539,0.0037636778,0.0003712386,0.3566107,0.0516972,0.020042691,0.005282566,0.29738194,0.0032594034],"about_ca_topic_score_codex":0.00026406557,"about_ca_topic_score_gemma":0.0000073168276,"teacher_disagreement_score":0.26286697,"about_ca_system_score_codex":0.00007624092,"about_ca_system_score_gemma":0.00015722774,"threshold_uncertainty_score":0.95193774},"labels":[],"label_agreement":null},{"id":"W3048499909","doi":"10.15694/mep.2020.000166.1","title":"COVID-19: Unveiling the Role of Uncertainty in Medical Education","year":2020,"lang":"en","type":"article","venue":"MedEdPublish","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Coronavirus disease 2019 (COVID-19); Pandemic; Phrase; Value (mathematics); Everyday life; Health care; Public relations; Health professionals; Psychology; Medical education; Political science; Medicine; Computer science; Pathology; Disease; Law","score_opus":0.026997574674668335,"score_gpt":0.36057498661155135,"score_spread":0.333577411936883,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3048499909","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6956181,0.0013888007,0.00007114516,0.28461412,0.00033607686,0.00038936397,0.0000061796404,0.000075864125,0.017500363],"genre_scores_gemma":[0.9518507,0.00006150064,0.00007640521,0.04753753,0.00032590464,0.000021662458,0.000047674,0.000008669638,0.00006991353],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.9986428,0.000088603614,0.00037811307,0.00021020396,0.0005168193,0.00016345078],"domain_scores_gemma":[0.9919886,0.006439009,0.00011639619,0.0002976609,0.00014376409,0.0010145608],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.000850147,0.00007937964,0.0002492092,0.000050164115,0.000029819092,0.000020822676,0.00019880965,0.00012037173,0.0013888364],"category_scores_gemma":[0.4521922,0.000051991556,0.00007512907,0.00047644487,0.0001327475,0.000049727543,0.00006182862,0.0003274962,0.000015352352],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000525765,0.0017341591,0.46154842,0.00019822642,0.000076871955,0.00006637643,0.0068509826,0.00048429632,0.00015044646,0.016211782,0.15473288,0.3574198],"study_design_scores_gemma":[0.011269257,0.0011604558,0.1320551,0.003519103,0.00031099698,0.00014831219,0.019387413,0.055730112,0.00089138123,0.022714535,0.7521435,0.000669803],"about_ca_topic_score_codex":0.0009367243,"about_ca_topic_score_gemma":0.00012028527,"teacher_disagreement_score":0.5974106,"about_ca_system_score_codex":0.000065536915,"about_ca_system_score_gemma":0.004145071,"threshold_uncertainty_score":0.99952406},"labels":[],"label_agreement":null},{"id":"W3049240114","doi":"10.1515/dx-2019-0091","title":"The effect of prior experience on diagnostic reasoning: exploration of availability bias","year":2020,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"Royal College of Physicians and Surgeons of Canada","keywords":"Medical diagnosis; Medicine; Radiology","score_opus":0.06774312574498485,"score_gpt":0.3509576239459068,"score_spread":0.2832144982009219,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3049240114","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9941142,0.00062493834,0.000043426106,0.0037950529,0.00015693891,0.0006932264,0.000015785085,0.000045500816,0.00051092764],"genre_scores_gemma":[0.99807465,0.0009483144,0.00009595335,0.00036654124,0.00010848688,0.00035731986,0.0000118077605,0.000017170616,0.000019757965],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982425,0.00022287977,0.0005662088,0.00032550687,0.000443682,0.00019920098],"domain_scores_gemma":[0.88794523,0.11092101,0.00027540204,0.00048714108,0.0001370454,0.00023419355],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.000614999,0.00016480028,0.0004937063,0.000029270668,0.0000791583,0.000013568013,0.00016152515,0.00009049479,0.00009371634],"category_scores_gemma":[0.45265916,0.00010156514,0.00021708511,0.00032083504,0.00027418663,0.00007104885,0.000054675293,0.0001781151,0.00005350857],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007756435,0.00026909134,0.95375943,0.00010455964,0.000043244854,0.00001316282,0.0013406073,0.000031451083,0.000146941,0.00022717779,0.00243509,0.040853593],"study_design_scores_gemma":[0.003135555,0.013004116,0.72720546,0.0035572727,0.00033502074,0.0000027486178,0.00048159805,0.0004195404,0.24888878,0.0000756633,0.0026556961,0.0002385792],"about_ca_topic_score_codex":0.000064398555,"about_ca_topic_score_gemma":0.0000039776064,"teacher_disagreement_score":0.45204416,"about_ca_system_score_codex":0.000027094524,"about_ca_system_score_gemma":0.000055156597,"threshold_uncertainty_score":0.5519513},"labels":[],"label_agreement":null},{"id":"W3049707794","doi":"10.7759/cureus.9727","title":"SPIRALS: An Approach to Non-Linear Thinking for Medical Students in the Emergency Department","year":2020,"lang":"en","type":"article","venue":"Cureus","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Memorial University of Newfoundland","funders":"","keywords":"Medicine; Emergency department; Delphi method; Inclusion (mineral); Context (archaeology); Medical education; Delphi; Curriculum; Set (abstract data type); Nursing; Psychology; Pedagogy","score_opus":0.07935842846227878,"score_gpt":0.42143543884958307,"score_spread":0.3420770103873043,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3049707794","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97375035,0.00032340703,0.006778173,0.014812944,0.00033241793,0.0014281913,0.000011101016,0.00005173464,0.0025116727],"genre_scores_gemma":[0.9773484,0.000075790704,0.0041719927,0.0171389,0.00092957314,0.00019855413,0.00007652978,0.000018655202,0.00004164333],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982234,0.00005014727,0.00035508888,0.00030518902,0.0008251688,0.00024095617],"domain_scores_gemma":[0.9982851,0.00094499066,0.000039281862,0.00026359715,0.000044219454,0.00042277778],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010904038,0.00011310114,0.00025190815,0.000016107497,0.000051580984,0.000017741122,0.0004907914,0.00009832244,0.000098891134],"category_scores_gemma":[0.021987611,0.00007114923,0.00008773701,0.00016954163,0.000018814506,0.00002766342,0.00009507292,0.00019561645,0.00004087504],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010970864,0.012226223,0.7847777,0.00025421675,0.0003096368,0.00030550157,0.028453752,0.00028509164,0.000065183165,0.0034051223,0.14035551,0.02846495],"study_design_scores_gemma":[0.017803103,0.010198141,0.8431087,0.0018710971,0.00073813344,0.00007229073,0.0063430653,0.04154598,0.00021098493,0.0025233582,0.07449695,0.0010881801],"about_ca_topic_score_codex":0.000040802734,"about_ca_topic_score_gemma":0.00001825243,"teacher_disagreement_score":0.065858565,"about_ca_system_score_codex":0.000023847646,"about_ca_system_score_gemma":0.000070337854,"threshold_uncertainty_score":0.9862506},"labels":[],"label_agreement":null},{"id":"W3081089509","doi":"10.1186/s12874-020-01095-8","title":"Measuring clinical uncertainty and equipoise by applying the agreement study methodology to patient management decisions","year":2020,"lang":"en","type":"article","venue":"BMC Medical Research Methodology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":20,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; University of Alberta Hospital; Centre Hospitalier de l’Université de Montréal; Health Sciences Centre; Ottawa Hospital","funders":"","keywords":"Reliability (semiconductor); Clinical equipoise; Dilemma; Medicine; Warrant; Clinical trial; Randomized controlled trial; Inter-rater reliability; Variety (cybernetics); Portfolio; Medical physics; Computer science; Psychology; Artificial intelligence; Rating scale; Surgery; Pathology; Mathematics","score_opus":0.7981267890516146,"score_gpt":0.6120079939446796,"score_spread":0.18611879510693496,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3081089509","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6512899,0.001916385,0.2887691,0.048972655,0.000656707,0.0068493774,0.000012319915,0.000099956495,0.001433623],"genre_scores_gemma":[0.6480188,0.0029149526,0.31866595,0.02581617,0.0010551413,0.0030655684,0.000025735306,0.00007770939,0.00035999407],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9542947,0.037458923,0.0017804463,0.0015228016,0.0036589885,0.001284134],"domain_scores_gemma":[0.26849335,0.7245437,0.00021499985,0.0013660344,0.0006146541,0.0047672843],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.08108921,0.00028002873,0.0012739332,0.00019553502,0.0003592251,0.000041324,0.0007096586,0.00035003707,0.00065285724],"category_scores_gemma":[0.8043064,0.00017256415,0.00022445957,0.0008440009,0.00091908034,0.000023141101,0.002257766,0.0020110502,0.00019626177],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002807884,0.0011502078,0.02968855,0.00006692499,0.0003581626,0.0005387837,0.0016743797,0.000042551408,0.00007842569,0.0010086066,0.02122583,0.9413597],"study_design_scores_gemma":[0.04865035,0.07546539,0.29321924,0.006986723,0.0026088424,0.00059629633,0.12304079,0.014483036,0.0005948404,0.01549575,0.41647357,0.0023851783],"about_ca_topic_score_codex":0.00039854305,"about_ca_topic_score_gemma":0.00022561796,"teacher_disagreement_score":0.9389745,"about_ca_system_score_codex":0.000100236124,"about_ca_system_score_gemma":0.0005717848,"threshold_uncertainty_score":0.94621205},"labels":[],"label_agreement":null},{"id":"W3081762009","doi":"10.1002/nop2.608","title":"Comparing feeling of competence regarding humanistic caring in Belgian nurses and nursing students: A comparative cross‐sectional study conducted in a French Belgian teaching hospital","year":2020,"lang":"en","type":"article","venue":"Nursing Open","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Feeling; Competence (human resources); Cross-sectional study; Nursing; Humanism; Psychology; Medicine; Social psychology","score_opus":0.10274201679513024,"score_gpt":0.4425888555934571,"score_spread":0.3398468387983269,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3081762009","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9946403,0.0002324663,0.00008151901,0.0003921489,0.00034240546,0.0015212377,0.0000059334393,0.000044976012,0.0027390018],"genre_scores_gemma":[0.9982149,0.0000035317464,0.0015057866,0.000070786395,0.000101979414,0.000026453323,0.000024653687,0.00002802362,0.000023885143],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9970491,0.00025792044,0.000996167,0.0007364146,0.0005457099,0.00041472437],"domain_scores_gemma":[0.9971385,0.0019487722,0.0002807112,0.0002397909,0.00013458458,0.00025761154],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00086916774,0.00027759903,0.001065093,0.00020661682,0.00026892294,0.00031526864,0.00041484335,0.00010507429,0.000013619834],"category_scores_gemma":[0.0047354936,0.00028779948,0.00006570907,0.0004085592,0.00034939026,0.00029161133,0.00018672789,0.00092860754,0.0000026270552],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00027076746,0.0023942264,0.93781775,0.000034209028,0.000058084715,0.000069585956,0.057453137,0.00057619833,0.0003240676,0.000176899,0.00003387069,0.00079118117],"study_design_scores_gemma":[0.0068905842,0.0011105923,0.9497998,0.014484379,0.00009878794,0.000024526304,0.021434495,0.0057014115,0.00007548456,0.00010997119,0.0000025478234,0.00026743286],"about_ca_topic_score_codex":0.003090345,"about_ca_topic_score_gemma":0.0003318129,"teacher_disagreement_score":0.036018644,"about_ca_system_score_codex":0.0003892413,"about_ca_system_score_gemma":0.00014383286,"threshold_uncertainty_score":0.99995744},"labels":[],"label_agreement":null},{"id":"W3088067868","doi":"10.1007/s11606-020-06228-9","title":"Finding Out Your Doctor Made a Mistake","year":2020,"lang":"en","type":"letter","venue":"Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"University of Toronto","keywords":"Medicine; Mistake; MEDLINE; Family medicine; Law","score_opus":0.10421357675402737,"score_gpt":0.39361460405394155,"score_spread":0.2894010272999142,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3088067868","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.009796259,0.0025905054,0.00063730706,0.97022945,0.007695872,0.00020663936,0.000031224874,0.00003590723,0.008776821],"genre_scores_gemma":[0.0036654922,0.00077809964,0.0016591764,0.79431874,0.103366375,0.000003561778,0.00009244738,0.00010783521,0.096008256],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99509877,0.00014377416,0.0021544045,0.00041894612,0.0016428829,0.00054123596],"domain_scores_gemma":[0.9931311,0.003454651,0.0016824724,0.00037070838,0.0005899714,0.0007710759],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0005593093,0.0005522648,0.0021715013,0.00042691,0.0000483755,0.0000428695,0.00054552814,0.00063767534,0.002041049],"category_scores_gemma":[0.035494365,0.0003565698,0.0008099171,0.00017198786,0.00026436028,0.000064141495,0.000091319074,0.005687132,0.00016431567],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005623729,0.00006800644,0.0013400369,0.00016794376,0.0006844626,0.03301019,0.00034701775,0.0000013444834,0.0016442988,0.000006704591,0.958969,0.0031986213],"study_design_scores_gemma":[0.004666717,0.0042483285,0.0012027081,0.020724652,0.0015577907,0.0061971084,0.000060781345,0.00007712455,0.00037474558,0.00023817235,0.96033186,0.00032000293],"about_ca_topic_score_codex":0.00012985441,"about_ca_topic_score_gemma":0.0000022733732,"teacher_disagreement_score":0.17591071,"about_ca_system_score_codex":0.00027819918,"about_ca_system_score_gemma":0.00040202367,"threshold_uncertainty_score":0.9998886},"labels":[],"label_agreement":null},{"id":"W3088431699","doi":"10.1503/cmaj.201426","title":"Misdiagnosis related to premature diagnostic closure during the COVID-19 pandemic","year":2020,"lang":"en","type":"article","venue":"Canadian Medical Association Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"McMaster University; Hamilton General Hospital","funders":"","keywords":"Chills; Medicine; Coronavirus disease 2019 (COVID-19); Pandemic; Closure (psychology); Exertional dyspnea; Medical history; 2019-20 coronavirus outbreak; Intensive care medicine; Medical emergency; Surgery; Outbreak; Pathology; Infectious disease (medical specialty)","score_opus":0.019015333356984944,"score_gpt":0.30469454271643603,"score_spread":0.28567920935945107,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3088431699","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5374359,0.00048928545,0.000018952116,0.4606993,0.00045485867,0.00028169868,0.00003430279,0.000058569523,0.00052715675],"genre_scores_gemma":[0.77792084,0.0005467661,0.000043981297,0.21934348,0.001579759,0.00002103735,0.000016028762,0.000024720155,0.0005033595],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9966369,0.0002990707,0.00066279486,0.0002819236,0.0015172361,0.00060202315],"domain_scores_gemma":[0.9717732,0.017692862,0.00023196879,0.00017900683,0.00030204494,0.009820933],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0018585658,0.00017612588,0.00039072442,0.00009523175,0.00044073033,0.00009291787,0.00029301428,0.0006216884,0.005184373],"category_scores_gemma":[0.5833496,0.00012414988,0.00022446651,0.00059451914,0.000054335254,0.00004714848,0.000034610857,0.0021579885,0.00056778424],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000033907858,0.000026354805,0.6908827,0.000016505632,0.0001276756,0.0013100455,0.0009481041,0.00004576833,0.0000049528744,0.000030018595,0.30424422,0.002329779],"study_design_scores_gemma":[0.0027972653,0.00020444967,0.60347456,0.00059083186,0.00018378516,0.0021216006,0.00041794806,0.00008068151,0.0000050188155,0.00013019024,0.389782,0.00021166683],"about_ca_topic_score_codex":0.0009841055,"about_ca_topic_score_gemma":0.0045392984,"teacher_disagreement_score":0.581491,"about_ca_system_score_codex":0.0013334734,"about_ca_system_score_gemma":0.0042118276,"threshold_uncertainty_score":0.99572504},"labels":[],"label_agreement":null},{"id":"W3089560638","doi":"10.1016/j.nedt.2020.104607","title":"Learning strategies used by undergraduate nursing students in the context of a digitial educational strategy based on script concordance: A descriptive study","year":2020,"lang":"en","type":"article","venue":"Nurse Education Today","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":31,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Concordance; Context (archaeology); Scripting language; Metacognition; Psychology; Qualitative research; Medical education; Computer science; Medicine; Cognition","score_opus":0.03651313723307676,"score_gpt":0.3861388579056112,"score_spread":0.34962572067253445,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3089560638","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96909034,0.00012222584,0.00007959951,0.025111677,0.00043316738,0.0009610304,0.000009786823,0.000025696485,0.0041664816],"genre_scores_gemma":[0.9972194,0.0000063034545,0.000041360418,0.0018931159,0.0002408375,0.000104828025,0.00013345995,0.000019407453,0.00034128973],"study_design_codex":"observational","study_design_gemma":"qualitative","domain_scores_codex":[0.9979849,0.00035062578,0.00046387265,0.00034821345,0.0006494098,0.00020296965],"domain_scores_gemma":[0.9962445,0.0029953597,0.00020978958,0.00020397714,0.00018619155,0.00016018492],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00032784534,0.00018047717,0.00032864168,0.00007779127,0.00008142415,0.000107127926,0.00019406418,0.00006386608,0.00008953732],"category_scores_gemma":[0.006029255,0.00014097692,0.000077595796,0.00037560338,0.00016180084,0.000110252535,0.0000057744915,0.00042469034,0.000020446063],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017146782,0.031747088,0.7635348,0.00006747953,0.00013145928,0.0000139106205,0.12286055,0.00082927174,0.00094714324,0.0074381866,0.03731616,0.033399276],"study_design_scores_gemma":[0.0050517726,0.0048586205,0.40209755,0.0021238625,0.0001618712,0.0000034576128,0.5833352,0.0008088189,0.00024342778,0.00089445827,0.00019830733,0.00022264046],"about_ca_topic_score_codex":0.00028256443,"about_ca_topic_score_gemma":0.000022378694,"teacher_disagreement_score":0.46047467,"about_ca_system_score_codex":0.00016386659,"about_ca_system_score_gemma":0.0021532897,"threshold_uncertainty_score":0.7218017},"labels":[],"label_agreement":null},{"id":"W3090300286","doi":"10.22454/fammed.2020.271478","title":"The Critical Thinking Skills of Practicing Family Physicians: A Population-Based Cross-Sectional Study","year":2020,"lang":"en","type":"article","venue":"Family Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Alberta","funders":"","keywords":"Family medicine; Cross-sectional study; Medicine; Univariate; Population; Multivariate analysis; Regression analysis; Multivariate statistics; Psychology; Internal medicine; Pathology; Environmental health","score_opus":0.06818993975072282,"score_gpt":0.417784001645849,"score_spread":0.3495940618951262,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3090300286","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9838813,0.0006286259,0.0013252876,0.011081406,0.00083854544,0.000638275,0.000011267153,0.000115844545,0.0014794661],"genre_scores_gemma":[0.98557824,0.000031248903,0.00051318604,0.012341564,0.0013475347,0.000029110663,0.000050368726,0.000033640576,0.00007508607],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9967559,0.00017794716,0.0010078499,0.00044919507,0.0012694229,0.00033970294],"domain_scores_gemma":[0.96537775,0.0332144,0.0002701243,0.0004021709,0.0004650234,0.00027054705],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0012920491,0.00022294067,0.0006205858,0.0000634073,0.0003429516,0.00003483769,0.00020564164,0.000113211456,0.00004698987],"category_scores_gemma":[0.14002441,0.00014284701,0.00015656187,0.00047000896,0.0003775072,0.00007609469,0.000053845306,0.0005434173,0.0000270397],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00047534338,0.00079847104,0.99153113,0.000055637796,0.000106171385,0.00006736979,0.0011728695,0.00016664847,0.00078480406,0.00038839973,0.002238151,0.0022149715],"study_design_scores_gemma":[0.0036715209,0.002175912,0.9879568,0.0011850039,0.0002545976,0.000005765939,0.002213457,0.0016355759,0.00004025053,0.00034709266,0.0003859699,0.00012802093],"about_ca_topic_score_codex":0.00069656636,"about_ca_topic_score_gemma":0.000007614795,"teacher_disagreement_score":0.13873236,"about_ca_system_score_codex":0.000058239973,"about_ca_system_score_gemma":0.00024487142,"threshold_uncertainty_score":0.8672195},"labels":[],"label_agreement":null},{"id":"W3092611646","doi":"10.1051/pmed/2020035","title":"Former en ligne au recrutement de patients partenaires : l’apport des formations par concordance","year":2020,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Humanities; Political science; Art","score_opus":0.06701499413443834,"score_gpt":0.35346482402660645,"score_spread":0.2864498298921681,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3092611646","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.88145256,0.013843375,0.011512329,0.06806444,0.002401769,0.0014070184,0.00029145452,0.0002952717,0.020731762],"genre_scores_gemma":[0.97009856,0.0063066273,0.0041804453,0.010176101,0.0021044747,0.00018974778,0.00022666964,0.00006193632,0.006655421],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.9964715,0.0002272058,0.0010141216,0.0004950744,0.00058972783,0.0012023482],"domain_scores_gemma":[0.9913415,0.0060554766,0.0002531705,0.00040123524,0.00022526739,0.0017233884],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0008531107,0.00034903418,0.0006434491,0.000045168046,0.0002938216,0.000053948927,0.0002846565,0.0003062526,0.0012374341],"category_scores_gemma":[0.054224715,0.00031897545,0.0002877617,0.00032697708,0.00064565655,0.00022916948,0.00018565815,0.0005067867,0.0012974406],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014710444,0.0012871943,0.46627653,0.0003693112,0.00017209782,0.000094002535,0.0075148065,0.00003712087,0.000035474462,0.00023818757,0.13954115,0.384287],"study_design_scores_gemma":[0.005876299,0.002734271,0.4789987,0.006406182,0.0008500159,0.00002375206,0.0020416782,0.0033687945,0.001484354,0.0014289777,0.4961136,0.00067337917],"about_ca_topic_score_codex":0.0003050365,"about_ca_topic_score_gemma":0.0000492075,"teacher_disagreement_score":0.38361365,"about_ca_system_score_codex":0.00045217026,"about_ca_system_score_gemma":0.0011385045,"threshold_uncertainty_score":0.9999262},"labels":[],"label_agreement":null},{"id":"W3094655505","doi":"","title":"What every teacher needs to know about clinical reasoning","year":2005,"lang":"en","type":"article","venue":"Educación Médica","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":114,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"sort; Context (archaeology); Psychology; Core (optical fiber); Medical education; Computer science; Medicine; Information retrieval","score_opus":0.028213286774424168,"score_gpt":0.38426516575570296,"score_spread":0.35605187898127877,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3094655505","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.90008533,0.005584809,0.00010807287,0.066969186,0.003790396,0.00046991956,0.0000035237404,0.00026915772,0.022719577],"genre_scores_gemma":[0.87948024,0.002441262,0.00864776,0.024236284,0.009501019,0.00006339852,0.00003850088,0.00008805527,0.07550349],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99729174,0.00014206488,0.0008521772,0.0005642117,0.0005065939,0.0006432261],"domain_scores_gemma":[0.99283296,0.0046951147,0.00013299387,0.0008183575,0.00014880003,0.0013718038],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0013594831,0.00028337608,0.0006810273,0.00019999889,0.000117674594,0.00012566922,0.00022937339,0.0003383644,0.0028315843],"category_scores_gemma":[0.04707655,0.00024328876,0.00037301783,0.00049880357,0.0001701414,0.00021877231,0.00012339941,0.00080337265,0.004208291],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00022694371,0.0013497613,0.09498855,0.000010463432,0.00008554683,0.000038158276,0.00069340435,0.000009490694,0.000031320775,0.0003377954,0.47390988,0.42831868],"study_design_scores_gemma":[0.0016751965,0.0005743632,0.18387786,0.0044118674,0.00020144692,0.00009439787,0.00095782336,0.000098703196,0.00011588023,0.000048843394,0.8075957,0.00034796342],"about_ca_topic_score_codex":0.000038093418,"about_ca_topic_score_gemma":0.000009235471,"teacher_disagreement_score":0.4279707,"about_ca_system_score_codex":0.00011764258,"about_ca_system_score_gemma":0.0005220732,"threshold_uncertainty_score":0.99807996},"labels":[],"label_agreement":null},{"id":"W3096970678","doi":"10.7759/cureus.11363","title":"Rater Training in Medical Education: A Scoping Review","year":2020,"lang":"en","type":"review","venue":"Cureus","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":22,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Manitoba; St. Boniface Hospital","funders":"","keywords":"Medicine; Medical education; Training (meteorology)","score_opus":0.14891686617388797,"score_gpt":0.5060802185157247,"score_spread":0.3571633523418367,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3096970678","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[2.0466368e-7,0.98817194,0.000004924325,0.0033511643,0.0006587901,0.0013965497,0.0000030496285,0.000048485454,0.00636492],"genre_scores_gemma":[6.715902e-7,0.9898613,0.00020187386,0.008046147,0.0011025576,0.00033581353,0.00018736652,0.000044555043,0.00021972577],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9975807,0.00017109547,0.0011033141,0.00047065257,0.0004157018,0.00025850825],"domain_scores_gemma":[0.9943595,0.004355829,0.0002330855,0.00040673008,0.00005246807,0.0005923938],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0006736816,0.00029283654,0.002818262,0.000054701588,0.000018629597,0.000014287285,0.00020548924,0.00037733893,0.0027224852],"category_scores_gemma":[0.1207756,0.00021321989,0.00044566416,0.0003983396,0.000055069642,0.000023910017,0.000074820324,0.0007985382,0.0004059922],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000015298904,0.000079963494,0.000005721204,0.07806109,0.000029162034,0.00023959977,0.000043027958,1.4034265e-9,7.50988e-10,0.000060500413,0.008102238,0.91337717],"study_design_scores_gemma":[0.00012347607,0.00003219495,0.000001244831,0.5267858,0.0004165989,0.00023359219,0.000006019018,5.579161e-7,5.9517387e-9,0.000010380568,0.47229016,0.000099999364],"about_ca_topic_score_codex":0.000011045323,"about_ca_topic_score_gemma":0.000014475718,"teacher_disagreement_score":0.91327715,"about_ca_system_score_codex":0.00011758028,"about_ca_system_score_gemma":0.012196717,"threshold_uncertainty_score":0.99818915},"labels":[],"label_agreement":null},{"id":"W3098363273","doi":"10.1111/epi.16739","title":"Precision Medicine: Academic dreaming or clinical reality?","year":2020,"lang":"en","type":"review","venue":"Epilepsia","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":26,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hotchkiss Brain Institute; University of Calgary","funders":"","keywords":"Clinical neurology; Psychology; Psychoanalysis; Medicine; Cognitive science; Neuroscience","score_opus":0.2545166491254092,"score_gpt":0.5307896273623853,"score_spread":0.2762729782369761,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3098363273","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000005833643,0.9866728,0.00019773128,0.003337882,0.001863215,0.0014357931,0.000046364457,0.00028350524,0.0061568297],"genre_scores_gemma":[0.000012528016,0.9858299,0.00063897704,0.0017463714,0.0071929917,0.00009033756,0.0005343885,0.00014576052,0.003808752],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99284846,0.00076723093,0.0036202858,0.0013287545,0.0008693315,0.0005659355],"domain_scores_gemma":[0.92398113,0.072478674,0.0010558739,0.0011069671,0.0001267035,0.0012506378],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity","insufficient_payload"],"category_scores_codex":[0.0029473742,0.0007294802,0.006223587,0.00018314744,0.00008616357,0.00001806088,0.0005435495,0.0018917504,0.0013609484],"category_scores_gemma":[0.31652316,0.00042520568,0.001295963,0.0007838777,0.00039863348,0.000050119812,0.00029309423,0.003971638,0.0018696528],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012573485,0.00008779108,0.00015876182,0.0027663347,0.000188458,0.0005756792,0.00003306632,2.461303e-8,2.1378924e-8,0.00013697911,0.15407126,0.8418559],"study_design_scores_gemma":[0.0010522994,0.0009554823,0.00037861237,0.19122466,0.0033454013,0.0004001723,0.00001350594,0.000015823825,5.28413e-8,0.00010440551,0.80218387,0.00032571287],"about_ca_topic_score_codex":0.00003218392,"about_ca_topic_score_gemma":0.000001685319,"teacher_disagreement_score":0.8415302,"about_ca_system_score_codex":0.00016343627,"about_ca_system_score_gemma":0.0014126707,"threshold_uncertainty_score":0.99982},"labels":[],"label_agreement":null},{"id":"W3098609939","doi":"10.1111/medu.14416","title":"How do cognitive processes influence script concordance test responses?","year":2020,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Psychology; Test (biology); Cognition; Deliberation; Concordance; Comprehension; Social psychology; Cognitive psychology; Scripting language; Process (computing); Applied psychology; Developmental psychology; Computer science; Medicine","score_opus":0.02433325861366146,"score_gpt":0.36717964701015715,"score_spread":0.34284638839649567,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3098609939","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7842422,0.0016556378,0.0003755592,0.21170847,0.00028043383,0.00032229564,0.000010779191,0.000102863705,0.0013017617],"genre_scores_gemma":[0.94020647,0.00030032004,0.00028235113,0.057459183,0.000655129,0.00007551885,0.00004835736,0.000013843278,0.00095884275],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99859273,0.000050990624,0.00023937343,0.0003298007,0.00059669185,0.00019041859],"domain_scores_gemma":[0.9742787,0.022520676,0.00022051983,0.00029027683,0.0012686771,0.0014211325],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00024699373,0.000118421456,0.0002383269,0.0000324927,0.00005225928,0.000052807853,0.000105833264,0.00012678209,0.0002964442],"category_scores_gemma":[0.7960297,0.00009619408,0.000039918876,0.00043029405,0.0003427027,0.000097370495,0.00003015738,0.00033139583,0.00014151903],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009897208,0.0026405677,0.57792467,0.0006213924,0.000052033443,0.00013859798,0.002806238,3.9367242e-7,0.00031743268,0.00032426696,0.1494887,0.26469597],"study_design_scores_gemma":[0.0074807545,0.004068039,0.76445204,0.028350735,0.00054273376,0.0003611558,0.008236304,0.00041412187,0.00499655,0.00071791786,0.17952543,0.0008541978],"about_ca_topic_score_codex":0.00003647776,"about_ca_topic_score_gemma":0.0000049633168,"teacher_disagreement_score":0.7957827,"about_ca_system_score_codex":0.000032399952,"about_ca_system_score_gemma":0.0048405956,"threshold_uncertainty_score":0.8587009},"labels":[],"label_agreement":null},{"id":"W3102342516","doi":"10.31661/gmj.v9i0.2048","title":"Artificial Intelligence, Important Assistant of Scientists and Physicians","year":2020,"lang":"en","type":"article","venue":"Galen Medical Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Guelph","funders":"","keywords":"Medicine; Medical education","score_opus":0.04329076480531882,"score_gpt":0.3506717717858603,"score_spread":0.3073810069805415,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3102342516","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95710266,0.0011166224,0.009286616,0.03037008,0.00051148894,0.00012107253,0.000012391109,0.000027297823,0.0014517712],"genre_scores_gemma":[0.99444425,0.00070322014,0.0012915563,0.0026573422,0.00085959025,9.013703e-7,0.0000033817057,0.000009702978,0.000030081546],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99779147,0.000045173383,0.00074943405,0.00018707963,0.0010017005,0.00022514415],"domain_scores_gemma":[0.99727124,0.0013074806,0.00021139592,0.00009672726,0.00012535711,0.0009877975],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0008464301,0.00009582476,0.00037691236,0.000046488276,0.0000642565,0.000023906203,0.00011500284,0.00009964592,0.0010698927],"category_scores_gemma":[0.040449467,0.00006935497,0.00012493935,0.00022013589,0.0003939049,0.000031603045,0.000051778792,0.000540835,0.000023695437],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00063614524,0.0008699307,0.08159925,0.00010600973,0.00018293143,0.0031936602,0.0010250806,0.000002501792,0.0022574791,0.0024087594,0.015315345,0.8924029],"study_design_scores_gemma":[0.0073347054,0.014505622,0.76774484,0.034374878,0.0026792667,0.016580733,0.0058964007,0.03980851,0.031116758,0.042889956,0.035057038,0.0020113077],"about_ca_topic_score_codex":0.000004514491,"about_ca_topic_score_gemma":0.0000023663179,"teacher_disagreement_score":0.8903916,"about_ca_system_score_codex":0.000014394046,"about_ca_system_score_gemma":0.00035823573,"threshold_uncertainty_score":0.99984324},"labels":[],"label_agreement":null},{"id":"W3102773901","doi":"10.1007/s10459-020-10013-y","title":"Exploring why we learn from productive failure: insights from the cognitive and learning sciences","year":2020,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":39,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; University of Toronto; University Health Network","funders":"","keywords":"Cognition; Psychology; Mathematics education; Medicine; Psychiatry","score_opus":0.10315907975330381,"score_gpt":0.4032614221234406,"score_spread":0.30010234237013683,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3102773901","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8766563,0.032818023,0.00016832857,0.0889039,0.00056966976,0.0004185452,0.000003089999,0.00003415754,0.0004280031],"genre_scores_gemma":[0.97524154,0.014158796,0.0032457293,0.006606074,0.0006410788,0.00007442676,0.0000147718265,0.000005580482,0.000012014903],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9981902,0.00024724452,0.00032545737,0.0006294675,0.00036988064,0.00023770455],"domain_scores_gemma":[0.98791033,0.011497707,0.00021375435,0.00008558759,0.00008293165,0.0002096892],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00062968145,0.00011238257,0.00022433537,0.00006029707,0.00060306385,0.000064721484,0.0001411787,0.000026815065,0.000020800919],"category_scores_gemma":[0.032503556,0.0000745259,0.00002143233,0.0009211087,0.0008306757,0.0007312997,0.000048799797,0.00038328156,0.000012838154],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010008063,0.00019839901,0.3868753,0.000059503433,0.000009126497,0.0000018145836,0.048675288,0.0008049299,0.000048517482,0.0013720941,0.0003426305,0.56151235],"study_design_scores_gemma":[0.0019066153,0.0037814523,0.57322836,0.0155458525,0.000070542636,0.000008091624,0.30609247,0.0050397357,0.0007580999,0.023047237,0.0699145,0.0006070298],"about_ca_topic_score_codex":0.0023534363,"about_ca_topic_score_gemma":0.0003338673,"teacher_disagreement_score":0.5609053,"about_ca_system_score_codex":0.000044217897,"about_ca_system_score_gemma":0.0008736941,"threshold_uncertainty_score":0.9756461},"labels":[],"label_agreement":null},{"id":"W3103286534","doi":"10.1080/0142159x.2020.1842343","title":"Consensus statement on the content of clinical reasoning curricula in undergraduate medical education","year":2020,"lang":"en","type":"review","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":142,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Queen's University; Swansea University; Queen's University Belfast; University of Glasgow; Newcastle University","keywords":"Curriculum; Statement (logic); Medical education; Apprenticeship; Psychology; Medicine; Pedagogy; Epistemology","score_opus":0.2072499453889439,"score_gpt":0.5000426008882051,"score_spread":0.29279265549926115,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3103286534","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0012402795,0.9395982,0.000023068917,0.051006198,0.001188682,0.0014527818,0.000008398029,0.000057450936,0.0054249186],"genre_scores_gemma":[0.0010061811,0.9930727,0.000103353894,0.00322951,0.000994183,0.00018262949,0.00014403582,0.00006247302,0.0012049422],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9907616,0.0019682792,0.003577711,0.0007449613,0.0025237198,0.00042376568],"domain_scores_gemma":[0.96658635,0.028805802,0.0013642366,0.0009896506,0.00019736035,0.002056603],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0071713175,0.00046255183,0.003418892,0.00011372885,0.00003682473,0.000012943571,0.00046092446,0.001051436,0.002643816],"category_scores_gemma":[0.48034173,0.00024458376,0.0009779023,0.00044805498,0.00068276934,0.000008646232,0.00018763989,0.003596575,0.00023070972],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000058353766,0.0022020054,0.002826552,0.001096606,0.0002465165,0.00038716223,0.000036899106,4.329234e-8,2.8513045e-9,0.0021790504,0.053700496,0.9372663],"study_design_scores_gemma":[0.0014936525,0.00042529427,0.0003220977,0.17244634,0.00071959087,0.00011441735,0.000099286204,0.00031720145,5.964802e-8,0.00011067335,0.8237311,0.00022027134],"about_ca_topic_score_codex":0.00020205912,"about_ca_topic_score_gemma":0.00002047285,"teacher_disagreement_score":0.93704605,"about_ca_system_score_codex":0.00019451432,"about_ca_system_score_gemma":0.0085453065,"threshold_uncertainty_score":0.99870217},"labels":[],"label_agreement":null},{"id":"W3104321308","doi":"10.1136/bmjqs-2020-011405","title":"Incidence, nature and causes of avoidable significant harm in primary care in England: retrospective case note review","year":2020,"lang":"en","type":"article","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":69,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute of Population and Public Health","funders":"Patient Safety Translational Research Centre; Department of Health and Social Care; NIHR Greater Manchester Patient Safety Translational Research Centre; National Institute for Health and Care Research","keywords":"Harm; Medicine; Incidence (geometry); Retrospective cohort study; Primary care; Patient safety; Family medicine; Emergency medicine; Health care; Pediatrics; Surgery","score_opus":0.03585353319219315,"score_gpt":0.39628756367659923,"score_spread":0.3604340304844061,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3104321308","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96821594,0.021278752,0.000019244319,0.007558181,0.00007270726,0.0014075567,0.00018947174,0.00003156388,0.0012265593],"genre_scores_gemma":[0.98425967,0.009474601,0.00036883197,0.0056506875,0.00009944768,0.000028872888,0.000083227664,0.000012655942,0.00002199409],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9975324,0.00036015818,0.0011425272,0.0004204495,0.00033534536,0.0002090868],"domain_scores_gemma":[0.99290735,0.006157174,0.00023320902,0.00029336114,0.00018784383,0.00022106172],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0019469565,0.00016108422,0.0009492067,0.00004561324,0.000028340393,0.0000058266705,0.00006896236,0.00020725213,0.000042836884],"category_scores_gemma":[0.0674059,0.00013246646,0.00009248409,0.00043507427,0.00010641965,0.000052355714,0.000089635134,0.00077494414,0.000004661743],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0022209578,0.0003336181,0.9598498,0.008637168,0.00004113151,0.0053404635,0.008228093,0.000008997396,0.00034680538,0.00030477927,0.001074307,0.013613912],"study_design_scores_gemma":[0.004389718,0.0004779198,0.97060883,0.020929217,0.00011641506,0.00015130332,0.0010734894,0.000025170932,0.00031385038,0.00010415784,0.0015625162,0.00024740488],"about_ca_topic_score_codex":0.0015860581,"about_ca_topic_score_gemma":0.0009784341,"teacher_disagreement_score":0.065458946,"about_ca_system_score_codex":0.00021431307,"about_ca_system_score_gemma":0.00049259426,"threshold_uncertainty_score":0.9404497},"labels":[],"label_agreement":null},{"id":"W3107172992","doi":"10.22541/au.160562668.84453336/v1","title":"A Cognitive Autopsy Approach towards Understanding Diagnostic Failure","year":2020,"lang":"en","type":"preprint","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; Nova Scotia Health Authority","funders":"","keywords":"Context (archaeology); Cognition; Medical diagnosis; Cognitive bias; Cognitive psychology; Narrative; Psychology; Process (computing); Feature (linguistics); Identification (biology); Variety (cybernetics); Root cause; Confirmation bias; Medicine; Social psychology; Applied psychology; Computer science; Artificial intelligence; Operations management; Pathology; Psychiatry; Engineering; Linguistics; History","score_opus":0.12876505524569887,"score_gpt":0.3653253802732935,"score_spread":0.2365603250275946,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3107172992","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.03682189,0.0012999144,0.51145476,0.040042344,0.0016299391,0.004443683,0.00037940207,0.0018293769,0.4020987],"genre_scores_gemma":[0.9836346,0.00023828275,0.009768425,0.003304684,0.00081368175,0.00017363504,0.0010023923,0.000079411264,0.0009848909],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99713403,0.00009492397,0.00059031975,0.0010878936,0.0006381759,0.000454634],"domain_scores_gemma":[0.9795226,0.018906428,0.00020195107,0.00046547593,0.00016706288,0.0007365256],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00032210152,0.00049813284,0.001134757,0.00012059057,0.00007748156,0.00011032543,0.00019671409,0.00072514464,0.0006930908],"category_scores_gemma":[0.15442955,0.00040387188,0.00050033734,0.00024022076,0.0002026027,0.000027084627,0.0007015704,0.0017962222,0.00022158668],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004164834,0.013288553,0.12827243,0.012324501,0.013981956,0.011821248,0.018149056,0.00076085975,0.00011014857,0.17127296,0.5192169,0.10663657],"study_design_scores_gemma":[0.05643042,0.0077123167,0.27184847,0.1793558,0.031517103,0.002066731,0.056372143,0.07108923,0.0013118088,0.3031177,0.0072341473,0.011944128],"about_ca_topic_score_codex":0.00017859046,"about_ca_topic_score_gemma":0.000008808137,"teacher_disagreement_score":0.9468127,"about_ca_system_score_codex":0.00030371634,"about_ca_system_score_gemma":0.0009090369,"threshold_uncertainty_score":0.99984133},"labels":[],"label_agreement":null},{"id":"W3108433231","doi":"10.1051/pmed/2020041","title":"L’apprentissage du raisonnement clinique infirmier dans le cadre d’un dispositif éducatif numérique basé sur la concordance de scripts","year":2020,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Humanities; Art","score_opus":0.03820034848394254,"score_gpt":0.319301001343273,"score_spread":0.28110065285933045,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3108433231","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7370187,0.028272977,0.038552452,0.16436893,0.0040707965,0.0014150181,0.00042379196,0.00045230894,0.025425058],"genre_scores_gemma":[0.9492723,0.007306156,0.0022507578,0.012259222,0.003804631,0.00020081812,0.00020517317,0.0001624359,0.024538556],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.994105,0.0011011193,0.0014239663,0.0012090873,0.0007373553,0.0014235191],"domain_scores_gemma":[0.97991043,0.016188022,0.00038300414,0.00085810455,0.00024200663,0.0024184072],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0018576225,0.00067245075,0.0012357471,0.0000627571,0.0003344405,0.00016815467,0.000664052,0.0007601863,0.0034107019],"category_scores_gemma":[0.060512822,0.00068937225,0.0005629978,0.00049552484,0.0011540931,0.000217288,0.0004298724,0.0015927815,0.0006620953],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006711521,0.006620258,0.38615242,0.0009874941,0.0006683477,0.00461279,0.022025207,0.00014893615,0.0040675947,0.004281867,0.34102356,0.22874036],"study_design_scores_gemma":[0.010092613,0.0015155913,0.13770087,0.00910676,0.0006264049,0.00040970778,0.008865539,0.006451517,0.015200614,0.00086761033,0.8076969,0.001465905],"about_ca_topic_score_codex":0.0013197702,"about_ca_topic_score_gemma":0.0001270048,"teacher_disagreement_score":0.4666733,"about_ca_system_score_codex":0.00030828067,"about_ca_system_score_gemma":0.0023613744,"threshold_uncertainty_score":0.99955577},"labels":[],"label_agreement":null},{"id":"W3109788745","doi":"10.1007/s10459-020-10018-7","title":"Where we’ve come from, where we might go","year":2020,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Summative assessment; Formative assessment; Operationalization; Curriculum; Psychology; Distance education; Mathematics education; Medical education; Pedagogy; Medicine; Epistemology","score_opus":0.03572322375038325,"score_gpt":0.423735829557509,"score_spread":0.38801260580712577,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3109788745","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.14320175,0.40458807,0.00028537717,0.44149688,0.0032023862,0.00078962324,0.000018597095,0.000107501,0.0063098213],"genre_scores_gemma":[0.7208812,0.2431144,0.016069388,0.017835949,0.0012946192,0.00008493331,0.000042420623,0.000018484285,0.00065856555],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.9981246,0.000092042166,0.0005398172,0.0005074281,0.0003905467,0.00034555446],"domain_scores_gemma":[0.99552596,0.003492635,0.00023721586,0.00020168556,0.00007023413,0.00047227775],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004700933,0.00012653004,0.000326393,0.00007067353,0.0001476179,0.000030987045,0.0002076491,0.00006527011,0.00035703022],"category_scores_gemma":[0.0059597925,0.000105070896,0.000046249083,0.0007767016,0.00024604978,0.00029979055,0.000031174273,0.0002478729,0.00018006391],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009189081,0.00059493416,0.53743386,0.0009173736,0.0000057600437,0.0000051907937,0.0053840564,0.0002558657,0.000021493544,0.002159676,0.03702415,0.41610578],"study_design_scores_gemma":[0.001520536,0.0018283112,0.07924015,0.020928621,0.000025744866,0.000013551636,0.00830187,0.0022997705,0.00008294221,0.009861314,0.8754974,0.00039978643],"about_ca_topic_score_codex":0.0005444762,"about_ca_topic_score_gemma":0.0006417326,"teacher_disagreement_score":0.83847326,"about_ca_system_score_codex":0.00012341875,"about_ca_system_score_gemma":0.0015183215,"threshold_uncertainty_score":0.7134859},"labels":[],"label_agreement":null},{"id":"W3110129881","doi":"10.1007/s11739-020-02580-0","title":"Diagnostic reasoning in internal medicine: a practical reappraisal","year":2020,"lang":"en","type":"review","venue":"Internal and Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":29,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University Hospital Foundation","funders":"Università degli Studi di Pavia; Fondazione IRCCS Policlinico San Matteo","keywords":"Medicine; MEDLINE; Intensive care medicine","score_opus":0.09853712744191005,"score_gpt":0.47470206740769244,"score_spread":0.3761649399657824,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3110129881","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00010238186,0.9604537,0.00021212065,0.007278343,0.0037182628,0.0007458395,0.000015510634,0.00009518919,0.027378656],"genre_scores_gemma":[0.00061029685,0.990504,0.00017705989,0.0003816185,0.0055365628,0.0001206898,0.00013593391,0.00010812412,0.002425695],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9941918,0.00031506442,0.0028366302,0.0011408195,0.0008572372,0.0006584176],"domain_scores_gemma":[0.97828513,0.019072663,0.0007778622,0.00049838424,0.0001598615,0.001206105],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011384747,0.00087618845,0.004308071,0.00045813504,0.00005429848,0.0000092842965,0.00031242936,0.0005255452,0.004558583],"category_scores_gemma":[0.24624667,0.0005519535,0.0004461471,0.00070561044,0.00052964373,0.00008612191,0.00026968555,0.0024074381,0.0001992436],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00031332741,0.00030977695,0.005533933,0.0070194104,0.00057334336,0.0064065834,0.0006965108,8.449622e-8,0.0000010618256,0.0013163844,0.10312505,0.87470454],"study_design_scores_gemma":[0.0016506811,0.0016213129,0.0011823314,0.23871014,0.0021093935,0.0027577442,0.00012682192,0.000058954305,1.5405354e-7,0.00015469671,0.7512631,0.00036467175],"about_ca_topic_score_codex":0.002072851,"about_ca_topic_score_gemma":0.00007840563,"teacher_disagreement_score":0.8743399,"about_ca_system_score_codex":0.00010225081,"about_ca_system_score_gemma":0.00034945196,"threshold_uncertainty_score":0.999894},"labels":[],"label_agreement":null},{"id":"W3110986760","doi":"10.1051/pmed/2020055","title":"Le déploiement de la formation à distance au sein des facultés de médecine dans le contexte de la crise sanitaire liée à la COVID-19 : et après ?","year":2020,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Coronavirus disease 2019 (COVID-19); Political science; Humanities; 2019-20 coronavirus outbreak; Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); Medicine; Outbreak; Philosophy; Virology; Infectious disease (medical specialty)","score_opus":0.11939697675753275,"score_gpt":0.4197854189295969,"score_spread":0.3003884421720642,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3110986760","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.33036506,0.012213558,0.55625933,0.08282446,0.00020227669,0.00061804766,0.00031399165,0.00035172328,0.016851556],"genre_scores_gemma":[0.9698448,0.0038551013,0.0054923575,0.0155189,0.00060347765,0.0001186874,0.00014780188,0.00008111898,0.004337727],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.99509364,0.0020078197,0.000783692,0.00055200147,0.0005143352,0.0010485072],"domain_scores_gemma":[0.97212106,0.024996333,0.00023252565,0.0004005493,0.00010344155,0.0021460967],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.003664022,0.00040594483,0.00066942576,0.000054348773,0.0004960305,0.00017345432,0.000367278,0.00052311644,0.00020747338],"category_scores_gemma":[0.116826005,0.0004066552,0.00028767547,0.00034193785,0.0018376218,0.00025733357,0.00019701921,0.0009529906,0.00011815368],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0022940182,0.009592321,0.30148914,0.0039041773,0.0006816363,0.004581198,0.13538316,0.0062022023,0.008145572,0.10392156,0.23594235,0.18786266],"study_design_scores_gemma":[0.020103145,0.0013057621,0.1288218,0.006593436,0.0008710406,0.0005271145,0.029230403,0.018728103,0.004266055,0.017716477,0.77069545,0.0011411967],"about_ca_topic_score_codex":0.0038586701,"about_ca_topic_score_gemma":0.0004077007,"teacher_disagreement_score":0.63947976,"about_ca_system_score_codex":0.0011419987,"about_ca_system_score_gemma":0.0055652964,"threshold_uncertainty_score":0.99983853},"labels":[],"label_agreement":null},{"id":"W3112383600","doi":"10.46747/cfp.6612904","title":"Virtual versus in-person primary care visits","year":2020,"lang":"en","type":"article","venue":"Canadian Family Physician","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"College of Family Physicians of Canada; Alberta Medical Association; University of Alberta","funders":"","keywords":"Primary care; Diagnostic accuracy; Medicine; Virtual patient; Computer science; Medical physics; Family medicine; Medical education; Radiology","score_opus":0.036183160045199364,"score_gpt":0.27621946871252584,"score_spread":0.24003630866732648,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3112383600","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6390213,0.0010819408,0.000005106621,0.0035885684,0.00046704084,0.00029120577,0.00008509506,0.000063835956,0.3553959],"genre_scores_gemma":[0.84146786,0.000016497554,0.000051667943,0.15766543,0.0005127416,0.000008009968,0.00019392018,0.00002698937,0.000056903016],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9988662,0.000024033434,0.0001628876,0.00034060996,0.00019424135,0.00041203754],"domain_scores_gemma":[0.9979413,0.0006411837,0.00003144293,0.0002099668,0.000060393875,0.0011157176],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.000036328493,0.00015805711,0.0003309362,0.00008592938,0.00005128344,0.00001900341,0.000111426096,0.000103084385,0.0000068317627],"category_scores_gemma":[0.0009970937,0.00016555651,0.00010797086,0.00035525885,0.00004572934,0.000046745517,0.000013762114,0.0002953922,0.00038447414],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012018319,0.00018042902,0.010968115,0.00016120805,0.00016874835,0.0023755073,0.004564277,0.00013399113,0.0023333956,0.0018954969,0.082718074,0.8932989],"study_design_scores_gemma":[0.017770551,0.0039548106,0.7755584,0.0031105229,0.0003294182,0.0000022130275,0.021075862,0.0005116422,0.0004699175,0.000028206287,0.1759961,0.0011923441],"about_ca_topic_score_codex":0.0067040017,"about_ca_topic_score_gemma":0.004474265,"teacher_disagreement_score":0.8921066,"about_ca_system_score_codex":0.00035442002,"about_ca_system_score_gemma":0.0011142386,"threshold_uncertainty_score":0.9999104},"labels":[],"label_agreement":null},{"id":"W3116879138","doi":"10.1051/pmed/2020054","title":"Pandémie COVID-19 : le défi de la responsabilité sociale des facultés de médecine","year":2020,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Francophone University Association","funders":"","keywords":"Humanities; Political science; Coronavirus disease 2019 (COVID-19); 2019-20 coronavirus outbreak; Philosophy; Medicine; Virology","score_opus":0.11538673531580239,"score_gpt":0.43154916137325944,"score_spread":0.31616242605745704,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3116879138","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5941989,0.043287054,0.023026079,0.32308927,0.00074865785,0.0007720983,0.00036521925,0.00063501776,0.013877704],"genre_scores_gemma":[0.82610476,0.011101056,0.014014087,0.066217914,0.005791301,0.00014696512,0.00013734678,0.00020605761,0.07628052],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9946632,0.0015661544,0.0008820736,0.0008457257,0.0006127745,0.0014300928],"domain_scores_gemma":[0.9338128,0.06093127,0.00019140911,0.00055169064,0.00015823543,0.0043545817],"candidate_categories":["metaresearch","metaepi_narrow","sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0035997666,0.00044639315,0.0009702603,0.0000711663,0.00043018218,0.000120175224,0.00045165975,0.0008533858,0.003570775],"category_scores_gemma":[0.48331925,0.00043839368,0.00049551914,0.0006329285,0.0028571642,0.000103810184,0.00027230044,0.0011939808,0.00077662047],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0024992698,0.0045773177,0.35730153,0.0012130281,0.00044039244,0.0026804314,0.011744737,0.00040659157,0.0035818932,0.008815786,0.23796609,0.36877292],"study_design_scores_gemma":[0.011790188,0.001401339,0.08198156,0.0019461445,0.00076543214,0.00024711076,0.0038153122,0.0019921758,0.0013201991,0.024333829,0.8696389,0.0007677997],"about_ca_topic_score_codex":0.0014723954,"about_ca_topic_score_gemma":0.000045390574,"teacher_disagreement_score":0.6316728,"about_ca_system_score_codex":0.0008871655,"about_ca_system_score_gemma":0.007829685,"threshold_uncertainty_score":0.9998565},"labels":[],"label_agreement":null},{"id":"W3116946031","doi":"10.1093/jacamr/dlaa107","title":"Cognitive bias: how understanding its impact on antibiotic prescribing decisions can help advance antimicrobial stewardship","year":2020,"lang":"en","type":"review","venue":"JAC-Antimicrobial Resistance","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":34,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute for Clinical Evaluative Sciences; Sunnybrook Hospital; Toronto East General Hospital; University of Toronto; Public Health Ontario; Hotel Dieu Shaver Health and Rehabilitation Centre","funders":"","keywords":"Cognitive bias; Cognition; Nudge theory; Confirmation bias; Psychological intervention; Antimicrobial stewardship; Psychology; Perspective (graphical); Stewardship (theology); Cognitive psychology; Risk analysis (engineering); Social psychology; Medicine; Computer science; Political science; Artificial intelligence; Psychiatry","score_opus":0.152116100299532,"score_gpt":0.39645321615786766,"score_spread":0.24433711585833567,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3116946031","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0020385329,0.97921973,0.0007912256,0.0031103431,0.0019002478,0.004943355,0.0054310276,0.00051183335,0.0020537071],"genre_scores_gemma":[0.050639752,0.9420279,0.0010271475,0.0013184341,0.0012137135,0.0000310639,0.0010437915,0.00040870492,0.0022894933],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99222976,0.0007233172,0.0018109167,0.0024759818,0.0010845537,0.0016754803],"domain_scores_gemma":[0.9664128,0.029036807,0.0016231134,0.0012182266,0.00045029735,0.0012587496],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.0007975943,0.0018338012,0.0059377784,0.0006232192,0.00053400826,0.0005016349,0.00075889856,0.0011101322,0.000085912354],"category_scores_gemma":[0.09656916,0.0014652818,0.0027213695,0.0018388284,0.0004466722,0.00019623645,0.0003059331,0.0025606386,0.00036964324],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008965291,0.0040361034,0.00062501285,0.048242707,0.0067908424,0.0115556875,0.0007979953,0.000062871666,0.0008956003,0.0013471098,0.3124763,0.6042045],"study_design_scores_gemma":[0.0041056033,0.0010030682,0.00019077465,0.683831,0.006175974,0.00044006473,0.00012982768,0.000017629998,0.000602595,0.00003260276,0.30117583,0.0022950561],"about_ca_topic_score_codex":0.00003401857,"about_ca_topic_score_gemma":0.00009753741,"teacher_disagreement_score":0.6355883,"about_ca_system_score_codex":0.0019064919,"about_ca_system_score_gemma":0.0028712065,"threshold_uncertainty_score":0.9997405},"labels":[],"label_agreement":null},{"id":"W3120884580","doi":"10.22454/primer.2021.453158","title":"Ambiguity Tolerance and Prospective Specialty Choice Among Third-Year Medical Students","year":2021,"lang":"en","type":"article","venue":"PRiMER","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Alberta","funders":"Social Sciences and Humanities Research Council of Canada; University of Alberta","keywords":"Specialty; Ambiguity; Medical education; Scale (ratio); Psychology; Ambiguity tolerance; Medicine; Family medicine; Computer science","score_opus":0.016812345403815737,"score_gpt":0.35364189537034557,"score_spread":0.3368295499665298,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3120884580","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9817038,0.00058908464,0.000037106707,0.0015985816,0.00031839838,0.00020603255,0.0000034109032,0.000055471144,0.0154881105],"genre_scores_gemma":[0.99389124,0.00025197794,0.0004035432,0.0018055006,0.0007360651,0.000015212752,0.000009677428,0.000015133354,0.0028716372],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99825555,0.000050703195,0.00024014134,0.00037902745,0.0008580803,0.00021649468],"domain_scores_gemma":[0.9972214,0.0019104321,0.000051240833,0.0002667788,0.00015885447,0.00039130865],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00041073907,0.00011543699,0.00032740308,0.00001816921,0.000059746584,0.00003889057,0.000106320294,0.0001387794,0.0007921896],"category_scores_gemma":[0.04047914,0.00009591889,0.000068615416,0.00014197532,0.00017458723,0.00003999073,0.00018181253,0.0003576001,0.000104150495],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000045483608,0.00042104334,0.99204856,0.000012742784,0.000058738126,0.00015923589,0.00011688005,1.09705866e-7,0.000021814401,0.0001784326,0.0033357814,0.0036011478],"study_design_scores_gemma":[0.0022288535,0.000079246376,0.9924674,0.000614472,0.00008159502,0.000027540638,0.000036334117,0.000018225348,0.0013398743,0.00021526734,0.0027893714,0.00010182715],"about_ca_topic_score_codex":0.00019118388,"about_ca_topic_score_gemma":0.00012896468,"teacher_disagreement_score":0.0400684,"about_ca_system_score_codex":0.000038333517,"about_ca_system_score_gemma":0.00019432134,"threshold_uncertainty_score":0.9676033},"labels":[],"label_agreement":null},{"id":"W3123412951","doi":"","title":"Use of Formal Methods in Medical Decision Making","year":2009,"lang":"en","type":"article","venue":"SSRN Electronic Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"MEDLINE; Computer science; Medical literature; Formal methods; Medicine; Political science","score_opus":0.035840721754714894,"score_gpt":0.43771220823875845,"score_spread":0.4018714864840436,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3123412951","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8055042,0.0019495337,0.18984374,0.0018683383,0.00015926649,0.0001065917,4.4741503e-7,0.000017270326,0.0005505805],"genre_scores_gemma":[0.96539193,0.0032074824,0.030336447,0.00078947714,0.00017188993,7.4999855e-7,0.000001317575,0.000009722763,0.00009098608],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9972135,0.000121451114,0.0006366602,0.00014463051,0.0005345407,0.0013492294],"domain_scores_gemma":[0.9927706,0.0066312104,0.0001615278,0.00016632078,0.00008692567,0.0001834203],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0051784655,0.00010666752,0.00037475835,0.00017870142,0.000038969905,0.000013682099,0.00012461381,0.0001680371,0.00010879408],"category_scores_gemma":[0.058437403,0.0000794948,0.00016594477,0.0002553799,0.000041042174,0.00012852722,0.000026197982,0.0020018353,0.0000054574134],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00048889907,0.00025630035,0.011092709,0.000001982315,0.00003056981,0.00006478092,0.000040553798,0.00001975621,0.00012453106,0.013884707,0.00007557677,0.97391963],"study_design_scores_gemma":[0.01009653,0.0068626762,0.37080568,0.010071182,0.00027660961,0.011667902,0.00054495916,0.0056173196,0.00070801,0.5813016,0.0015846663,0.0004628456],"about_ca_topic_score_codex":0.000018329209,"about_ca_topic_score_gemma":0.00008104074,"teacher_disagreement_score":0.9734568,"about_ca_system_score_codex":0.00030862849,"about_ca_system_score_gemma":0.0020262478,"threshold_uncertainty_score":0.94949377},"labels":[],"label_agreement":null},{"id":"W3123716856","doi":"10.1111/eje.12649","title":"Script concordance tests: A call for action in dental education","year":2021,"lang":"en","type":"article","venue":"European Journal Of Dental Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Université de Montréal","funders":"","keywords":"Concordance; Test (biology); Dental education; Medical education; Psychology; Clinical judgment; Action (physics); Medicine; Medical physics","score_opus":0.0335290089963491,"score_gpt":0.3785386474968958,"score_spread":0.3450096385005467,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3123716856","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98967576,0.0016714084,0.0003238371,0.0012142533,0.003477895,0.00016491361,0.0000027157,0.000007760259,0.003461437],"genre_scores_gemma":[0.9919374,0.00022491445,0.0037228782,0.0013065814,0.0012254333,0.0000056287067,0.00005866752,0.00002186219,0.0014966526],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9988758,0.00014216041,0.0005249623,0.00014543554,0.0001917972,0.000119874814],"domain_scores_gemma":[0.9984697,0.00056403695,0.00030445546,0.00012239488,0.00036061055,0.0001788152],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00049196486,0.00008863264,0.00019196294,0.00011821426,0.00003579419,0.00003759974,0.000068335205,0.000025347374,0.000053504562],"category_scores_gemma":[0.012617188,0.000084535226,0.00012161689,0.00018784989,0.000030987816,0.00014253767,0.000014755108,0.0002141653,0.0000353943],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00053736754,0.0053617214,0.2942975,0.00010422293,0.00005522958,0.00015134925,0.0008195498,0.000011382903,0.016314747,0.00024033191,0.039026268,0.64308035],"study_design_scores_gemma":[0.003023293,0.0008028066,0.9558823,0.0042357324,0.00015299556,0.003224689,0.009946871,0.000031770847,0.0046587656,0.00015753524,0.017724028,0.00015920128],"about_ca_topic_score_codex":0.00002362751,"about_ca_topic_score_gemma":0.000042154727,"teacher_disagreement_score":0.66158485,"about_ca_system_score_codex":0.00025700845,"about_ca_system_score_gemma":0.0014599465,"threshold_uncertainty_score":0.99569994},"labels":[],"label_agreement":null},{"id":"W3123772788","doi":"10.1097/acm.0000000000003841","title":"A Communication-Based Assessment Station as an Alternative to the OSCE","year":2021,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Addiction and Mental Health; University of Toronto; University Health Network","funders":"","keywords":"Objective structured clinical examination; Cronbach's alpha; Psychology; Medical education; Test (biology); Educational measurement; Medicine; Family medicine; Clinical psychology; Psychometrics; Curriculum","score_opus":0.060086623764751365,"score_gpt":0.4618448324584287,"score_spread":0.40175820869367734,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3123772788","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6861389,0.0010646924,0.009885139,0.2925164,0.0003797882,0.00058872293,0.000011682391,0.00008411355,0.009330536],"genre_scores_gemma":[0.94337904,0.00054533756,0.00157062,0.051917903,0.00050455716,0.00006894688,0.0003487669,0.000016630316,0.0016482146],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99827784,0.00026660546,0.00041531085,0.000283582,0.00056424644,0.00019239179],"domain_scores_gemma":[0.9886224,0.009873302,0.00012135659,0.0006992854,0.0003286294,0.0003549787],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0011895729,0.000121964666,0.00028232866,0.0000545053,0.000106206986,0.000007390397,0.00023209225,0.0001094804,0.00071535545],"category_scores_gemma":[0.045349438,0.00007591064,0.000035614146,0.00033294965,0.00013431841,0.000039065297,0.000054729462,0.0007299786,0.000109452136],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001304428,0.0014231112,0.16387244,0.00018998298,0.00071358046,0.00070011703,0.047071356,0.002473402,0.011462686,0.112138405,0.436333,0.2223175],"study_design_scores_gemma":[0.02703035,0.0075360565,0.5722478,0.024323363,0.0019951363,0.0004748682,0.030911352,0.042986926,0.012329385,0.037376754,0.24162887,0.0011591521],"about_ca_topic_score_codex":0.00047720908,"about_ca_topic_score_gemma":0.000038351795,"teacher_disagreement_score":0.40837535,"about_ca_system_score_codex":0.000092863644,"about_ca_system_score_gemma":0.000466413,"threshold_uncertainty_score":0.96269196},"labels":[],"label_agreement":null},{"id":"W3125039613","doi":"10.3138/jvme-2020-0032","title":"Tolerance of Ambiguity in Veterinary Students at the Beginning and End of a Second-year Clinical Pathology Course","year":2021,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Ambiguity; Ambiguity tolerance; Scale (ratio); Clinical pathology; Medicine; Psychology; Course (navigation); Pathology; Medical education; Veterinary medicine; Computer science","score_opus":0.0791960509547792,"score_gpt":0.47612229470818046,"score_spread":0.39692624375340124,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3125039613","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9922602,0.002872316,0.000009915082,0.0036752797,0.00095709274,0.00008259791,0.0000033609304,0.0000019403947,0.00013727679],"genre_scores_gemma":[0.9963836,0.001743898,0.0005744438,0.00088058534,0.00028139312,0.0000032440119,0.0000057020497,0.000007765692,0.00011935028],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99768025,0.00039435626,0.0010831866,0.00015446181,0.00056016404,0.00012758338],"domain_scores_gemma":[0.99497217,0.0037471408,0.0005394499,0.00021910312,0.0002719064,0.00025021756],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0024193567,0.00008718439,0.00053969014,0.00005439294,0.00002322154,0.0000037568059,0.00013878624,0.00015636385,0.00058813894],"category_scores_gemma":[0.026367348,0.00006228836,0.00013003805,0.0001343535,0.00041352364,0.00004475364,0.00014475762,0.00049006316,0.0000015527977],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019239198,0.0052262424,0.87957126,0.00022583362,0.0001095986,0.001878903,0.0010796752,0.0000023964317,0.0029998794,0.000027854427,0.0029438653,0.10401057],"study_design_scores_gemma":[0.0024629838,0.0032459,0.9829495,0.0028405671,0.000083047606,0.0058434014,0.00056895864,0.000025291114,0.00009800715,0.00007029707,0.0017617206,0.00005033348],"about_ca_topic_score_codex":0.000012050102,"about_ca_topic_score_gemma":0.000008630452,"teacher_disagreement_score":0.10396023,"about_ca_system_score_codex":0.000041077303,"about_ca_system_score_gemma":0.001419489,"threshold_uncertainty_score":0.981834},"labels":[],"label_agreement":null},{"id":"W3126255905","doi":"10.1007/s10459-021-10032-3","title":"Are raters influenced by prior information about a learner? A review of assimilation and contrast effects in assessment","year":2021,"lang":"en","type":"review","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Medical Council of Canada; Ottawa Hospital; University of Ottawa","funders":"","keywords":"Contrast (vision); Psychology; Context (archaeology); Cognitive psychology; Conceptualization; Cognition; Context effect; Applied psychology; Social psychology; Computer science; Linguistics","score_opus":0.02609944033284872,"score_gpt":0.48743528107114276,"score_spread":0.46133584073829403,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3126255905","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00027223147,0.9962537,0.00009485321,0.0008041367,0.00043082205,0.0019163956,0.000008719735,0.000008283138,0.00021081785],"genre_scores_gemma":[0.00035570408,0.99521357,0.0017388724,0.002104902,0.00002927501,0.00034850804,0.00019325966,0.0000057351413,0.000010154943],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9972662,0.0003635237,0.0014354875,0.0003260849,0.00037547626,0.00023327308],"domain_scores_gemma":[0.9932219,0.0044834004,0.0018350543,0.00017743981,0.00014191416,0.0001403122],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002166431,0.00018477449,0.001392137,0.0002770918,0.000058222537,0.000027097478,0.00009494991,0.00010947561,0.000008528319],"category_scores_gemma":[0.021421844,0.0001450408,0.000071457696,0.0011052695,0.00016879228,0.0004981287,0.000023671588,0.00030577905,0.000001500341],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000020699201,0.00009871591,0.006600959,0.08630048,0.0000025173608,4.1190026e-7,0.00006561332,0.000004968055,2.6230373e-8,0.000096019605,0.00019790765,0.90663034],"study_design_scores_gemma":[0.000323409,0.00018519312,0.017185705,0.7605878,0.00005190735,0.0000105082,0.00012476667,0.00004985535,3.1376703e-7,0.000036770954,0.22132364,0.000120150864],"about_ca_topic_score_codex":0.00012456255,"about_ca_topic_score_gemma":0.00003099565,"teacher_disagreement_score":0.9065102,"about_ca_system_score_codex":0.0003393902,"about_ca_system_score_gemma":0.0037825804,"threshold_uncertainty_score":0.9868211},"labels":[],"label_agreement":null},{"id":"W3127034479","doi":"10.1111/medu.14468","title":"Qualitative analysis of medical student reflections on the implicit association test","year":2021,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"National Institute on Minority Health and Health Disparities; Josiah Macy Jr. Foundation; Robert Wood Johnson Foundation","keywords":"Implicit-association test; Psychology; Narrative; Grounded theory; Identification (biology); Social psychology; Test (biology); Session (web analytics); Association (psychology); Implicit bias; Qualitative research; Implicit attitude; Medical education; Applied psychology; Clinical psychology; Psychotherapist; Medicine","score_opus":0.06111876509693844,"score_gpt":0.5642901515508294,"score_spread":0.5031713864538909,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3127034479","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.71272683,0.000158117,0.00036220322,0.2774564,0.0005710575,0.00015156856,0.000011202904,0.000028629056,0.00853399],"genre_scores_gemma":[0.9792676,0.0002465661,0.000091506925,0.018165287,0.00027757237,0.0000585806,0.00017518036,0.0000074287477,0.0017102559],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9966236,0.00033751325,0.000482492,0.00022494385,0.002183846,0.00014761687],"domain_scores_gemma":[0.94655025,0.05169911,0.00026435978,0.000423236,0.00069242716,0.0003706361],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0027480784,0.000081940736,0.00033936958,0.000117845506,0.000074316085,0.00001440798,0.000120599565,0.0002272547,0.0047236597],"category_scores_gemma":[0.5140002,0.00005381612,0.00020130484,0.0013319162,0.00007583092,0.0000145415515,0.000035975347,0.0004148998,0.000048641858],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000108571534,0.0325416,0.42639107,0.000074025906,0.0064549022,0.000049064583,0.04806578,0.000017234974,0.00034907294,0.06401102,0.32253152,0.09940612],"study_design_scores_gemma":[0.0013123082,0.00053449656,0.948158,0.0018757582,0.003778931,0.000019270261,0.030335383,0.001112662,0.0005449086,0.0012024828,0.01093583,0.0001899142],"about_ca_topic_score_codex":0.00022925147,"about_ca_topic_score_gemma":0.0001940442,"teacher_disagreement_score":0.521767,"about_ca_system_score_codex":0.00022704907,"about_ca_system_score_gemma":0.0026580626,"threshold_uncertainty_score":0.99618614},"labels":[],"label_agreement":null},{"id":"W3127256662","doi":"10.1111/medu.14459","title":"Skeptical self‐regulation: Resident experiences of uncertainty about uncertainty","year":2021,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":28,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Centre for Advancing Health Outcomes; University of British Columbia","funders":"","keywords":"Skepticism; Set (abstract data type); Psychology; Grounded theory; Rehearsing; Affordance; Medical education; Cognitive psychology; Social psychology; Computer science; Epistemology; Qualitative research; Medicine; Sociology","score_opus":0.013847685424555578,"score_gpt":0.36190629241857153,"score_spread":0.348058606994016,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3127256662","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97891915,0.0008845781,0.0001994845,0.01371649,0.0011512787,0.00015250075,0.0000012541054,0.000057415848,0.0049178763],"genre_scores_gemma":[0.9929919,0.00018462467,0.0011034034,0.003549675,0.0006187268,0.00006615738,0.0001257989,0.000009587922,0.0013501258],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9978997,0.00010925333,0.0005385321,0.00031114917,0.00094047614,0.00020084898],"domain_scores_gemma":[0.99565464,0.0029759402,0.00010987024,0.00036328967,0.000487959,0.00040832013],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00050596433,0.00011066811,0.00032168604,0.000047494224,0.00005464578,0.000014539221,0.00009554146,0.00019515009,0.0071680336],"category_scores_gemma":[0.09032956,0.00008914097,0.00011931483,0.00032362173,0.00023405452,0.000036890215,0.00004166225,0.00021312581,0.000046844954],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007216026,0.015744016,0.08128897,0.0007188358,0.00040781073,0.00024714952,0.083996765,0.00030492424,0.0005456602,0.037473343,0.37992185,0.39862907],"study_design_scores_gemma":[0.007847624,0.0014751757,0.7522879,0.016142942,0.00087223377,0.0008379656,0.08577544,0.01843607,0.0065440284,0.007558797,0.101172395,0.0010493752],"about_ca_topic_score_codex":0.00013535455,"about_ca_topic_score_gemma":0.000021712165,"teacher_disagreement_score":0.670999,"about_ca_system_score_codex":0.00009916943,"about_ca_system_score_gemma":0.004670638,"threshold_uncertainty_score":0.99373955},"labels":[],"label_agreement":null},{"id":"W3127579986","doi":"10.1186/s12909-021-02522-0","title":"Teaching clinical reasoning to undergraduate medical students by illness script method: a randomized controlled trial","year":2021,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":55,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"Shiraz University; Shiraz University of Medical Sciences; Thomas Jefferson University; Foundation for Advancement of International Medical Education and Research","keywords":"Randomized controlled trial; Concordance; Intervention (counseling); Medicine; Psychological intervention; Test (biology); Clinical trial; Family medicine; Psychiatry; Internal medicine","score_opus":0.0289634510315746,"score_gpt":0.47181737137631213,"score_spread":0.44285392034473753,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3127579986","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6673222,0.0021843265,0.17240286,0.1251101,0.017690573,0.0074181478,0.0000065743757,0.00033451742,0.0075306497],"genre_scores_gemma":[0.65916413,0.0022560554,0.11667719,0.15173152,0.015858365,0.005527616,0.0010997267,0.00024984914,0.047435556],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9849453,0.0063016135,0.0026991395,0.0009705222,0.0045270845,0.00055635686],"domain_scores_gemma":[0.7169476,0.26786378,0.0014100078,0.0020940804,0.0014441549,0.010240391],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.03174421,0.0003543243,0.0031300003,0.0001354451,0.000205975,0.00012742697,0.00050484494,0.0007054483,0.0020024816],"category_scores_gemma":[0.8697659,0.00025651933,0.0010237453,0.00035446856,0.00025094143,0.00006492961,0.00025801893,0.0014609362,0.00016637954],"study_design_candidate":"randomized_trial","study_design_consensus":"randomized_trial","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.69588625,0.010084324,0.0048493077,0.00007737788,0.00064816716,0.00010123062,0.0004167748,0.000001122182,0.000006889206,0.0039662677,0.13134244,0.15261984],"study_design_scores_gemma":[0.97839904,0.000387268,0.00041840962,0.0048521627,0.0010459207,0.00014543,0.00053265906,0.0021854911,0.00001594111,0.0008436493,0.010839966,0.00033408013],"about_ca_topic_score_codex":0.00039535086,"about_ca_topic_score_gemma":0.00008256677,"teacher_disagreement_score":0.8380217,"about_ca_system_score_codex":0.00014581774,"about_ca_system_score_gemma":0.013324918,"threshold_uncertainty_score":0.9999887},"labels":[],"label_agreement":null},{"id":"W3127624731","doi":"10.1038/s41405-021-00067-4","title":"The interrelationship between confidence and correctness in a multiple-choice assessment: pointing out misconceptions and assuring valuable questions","year":2021,"lang":"en","type":"article","venue":"BDJ Open","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Saskatchewan","funders":"","keywords":"Correctness; Low Confidence; Confidence interval; Psychology; Point (geometry); Mathematics education; Sample (material); Computer science; Social psychology; Statistics; Mathematics; Algorithm","score_opus":0.13854468535408662,"score_gpt":0.44543906412743406,"score_spread":0.30689437877334746,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3127624731","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9836617,0.0007164554,0.00121694,0.0051597636,0.0001848921,0.0005424468,0.00001013743,0.00003315019,0.008474524],"genre_scores_gemma":[0.99573183,0.00012346209,0.0016289565,0.00015938682,0.000077174394,0.00006172545,0.000024528053,0.000009879075,0.002183037],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99893886,0.00018289429,0.00031176233,0.00026701612,0.00012349576,0.00017598917],"domain_scores_gemma":[0.95838946,0.041028675,0.00008556006,0.00024591413,0.000114414775,0.00013597625],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0009295067,0.00008832817,0.00023643933,0.00003105547,0.00030200722,0.00025589732,0.00010144943,0.000063845095,0.00002904285],"category_scores_gemma":[0.080942094,0.00007064802,0.000027097316,0.00013307999,0.00012536501,0.00013213792,0.00026145563,0.00032817887,0.0000073719025],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000008922491,0.00004204238,0.9928618,0.000009357085,0.000021286944,0.000013185228,0.00032088818,0.0000118061525,0.00004672252,0.0008663357,0.00021818456,0.0055794716],"study_design_scores_gemma":[0.0013737035,0.00004429536,0.99125403,0.0021425993,0.00007817856,0.000029408451,0.001050698,0.0022517475,0.000039513547,0.00074298703,0.00090550346,0.00008734298],"about_ca_topic_score_codex":0.0012307911,"about_ca_topic_score_gemma":0.0025006584,"teacher_disagreement_score":0.08001259,"about_ca_system_score_codex":0.000037604605,"about_ca_system_score_gemma":0.00022024049,"threshold_uncertainty_score":0.92679954},"labels":[],"label_agreement":null},{"id":"W3127977746","doi":"10.1044/2020_ajslp-20-00073","title":"Confidence and Training of Speech-Language Pathologists in Cognitive-Communication Disorders: Time to Rethink Graduate Education Models?","year":2021,"lang":"en","type":"article","venue":"American Journal of Speech-Language Pathology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":38,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Cognition; Curriculum; Psychology; Consistency (knowledge bases); Scope (computer science); Medical education; Medicine; Computer science; Pedagogy","score_opus":0.03743500064446241,"score_gpt":0.36467436331004066,"score_spread":0.32723936266557824,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3127977746","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98998475,0.0032772606,0.0014781163,0.0033649304,0.000082094564,0.00020359913,0.000016653486,0.000016158636,0.0015764245],"genre_scores_gemma":[0.96694946,0.0011401726,0.030205952,0.0013115453,0.000112441674,0.000007384877,0.00003734014,0.000024233905,0.00021145573],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.99778813,0.00047169955,0.0008402848,0.00031531986,0.00030233324,0.00028223154],"domain_scores_gemma":[0.99473387,0.0035179467,0.0006914283,0.0003823198,0.0004311684,0.0002432591],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0013613916,0.00019085172,0.00087053725,0.00030671206,0.000033428958,0.000017170169,0.00017236944,0.0001072382,0.00014485656],"category_scores_gemma":[0.024257021,0.00017612553,0.00012638919,0.00051957177,0.00059701264,0.00011518747,0.00009913596,0.0006152381,0.0000148611725],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00037101284,0.00040834688,0.003600087,0.000025477644,0.00004404979,0.0024098644,0.031941984,0.000022723536,0.017606081,0.00009918351,0.00006430129,0.9434069],"study_design_scores_gemma":[0.012663457,0.012815486,0.39003253,0.023528166,0.0013854693,0.051311523,0.4546161,0.0014470525,0.033379365,0.016996205,0.00010911869,0.0017155323],"about_ca_topic_score_codex":0.0002643194,"about_ca_topic_score_gemma":0.000073380994,"teacher_disagreement_score":0.94169134,"about_ca_system_score_codex":0.00005119564,"about_ca_system_score_gemma":0.00051513215,"threshold_uncertainty_score":0.98396206},"labels":[],"label_agreement":null},{"id":"W3128689444","doi":"10.1111/jep.13544","title":"Enhancing clinical judgement in virtual care for complex chronic disease","year":2021,"lang":"en","type":"article","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Judgement; Deductive reasoning; Health care; Psychology; Computer science; Artificial intelligence; Epistemology","score_opus":0.17615140402813137,"score_gpt":0.5621877584359806,"score_spread":0.3860363544078492,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3128689444","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9297182,0.008118461,0.009330105,0.04390513,0.005433732,0.0017407303,0.000018760893,0.000017597307,0.0017173078],"genre_scores_gemma":[0.9790305,0.0016877794,0.011692452,0.0050771916,0.002325384,0.000030231,0.000054974265,0.000022522678,0.00007900362],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99180967,0.0020318618,0.004319176,0.0004003588,0.0011680734,0.00027087462],"domain_scores_gemma":[0.8342422,0.15840213,0.0019868652,0.00049018423,0.0042298697,0.00064875185],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.02525103,0.00014923392,0.00088822294,0.00013489633,0.000046073743,0.0000391889,0.00011078509,0.00021568501,0.00044776726],"category_scores_gemma":[0.7502899,0.00013247154,0.0005666675,0.00032121263,0.00010014674,0.0002665767,0.000067775065,0.0011609823,0.000027338587],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.015416151,0.014381199,0.37704214,0.0003418024,0.0005855416,0.0023794987,0.0007995088,0.0065734973,0.00015254473,0.0009272452,0.014536888,0.566864],"study_design_scores_gemma":[0.034317147,0.005220395,0.87826145,0.0051022945,0.0021912558,0.00014498463,0.0032171204,0.018271841,0.000041119325,0.0012013499,0.051750846,0.00028021797],"about_ca_topic_score_codex":0.000009285343,"about_ca_topic_score_gemma":0.00014156286,"teacher_disagreement_score":0.7250389,"about_ca_system_score_codex":0.00051709695,"about_ca_system_score_gemma":0.0045768274,"threshold_uncertainty_score":0.8751554},"labels":[],"label_agreement":null},{"id":"W3129635634","doi":"10.18162/ritpu-2021-v18n1-16","title":"Travaux pratiques d’anatomie et d’histologie à distance","year":2021,"lang":"fr","type":"article","venue":"Revue internationale des technologies en pédagogie universitaire","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Humanities; Geography; Art","score_opus":0.033349975356133175,"score_gpt":0.3273671338578902,"score_spread":0.294017158501757,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3129635634","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5455817,0.13980034,0.015632313,0.21343783,0.0042182924,0.0006542687,0.0005699669,0.0026857473,0.07741953],"genre_scores_gemma":[0.5535098,0.067952946,0.011035346,0.00039257456,0.00024544602,0.000017696471,0.00025703124,0.00005500233,0.3665342],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99732584,0.00020636799,0.00057429203,0.00084317254,0.00039244516,0.00065787876],"domain_scores_gemma":[0.9876527,0.010351603,0.00027690735,0.0008080796,0.0007500073,0.00016066521],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00038702466,0.00041383412,0.0006641142,0.00026623099,0.00018691283,0.00008022087,0.0007533125,0.0008326312,0.0004781904],"category_scores_gemma":[0.074496366,0.00046599976,0.00038356037,0.0006472597,0.001266525,0.0004229133,0.0006121995,0.0010481892,0.00026307555],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011781259,0.000986039,0.036413804,0.00043937188,0.00067099166,0.012195249,0.0012212985,0.00018591539,0.0003117589,0.27554786,0.019115226,0.65279466],"study_design_scores_gemma":[0.002765904,0.0008884612,0.017848462,0.02127928,0.00075854856,0.0017982826,0.05077204,0.0016227185,0.010224264,0.019272452,0.87167716,0.0010924451],"about_ca_topic_score_codex":0.00011285644,"about_ca_topic_score_gemma":0.0005882163,"teacher_disagreement_score":0.8525619,"about_ca_system_score_codex":0.0020227374,"about_ca_system_score_gemma":0.0012711749,"threshold_uncertainty_score":0.99977916},"labels":[],"label_agreement":null},{"id":"W3132306444","doi":"10.1016/j.ijosm.2021.02.001","title":"The development and application of a remediation process in an osteopathic curriculum","year":2021,"lang":"en","type":"article","venue":"International journal of osteopathic medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Sherbrooke","funders":"","keywords":"Remedial education; Curriculum; Medicine; Medical education; Psychological intervention; Process (computing); Intervention (counseling); Osteopathic medicine in the United States; Mathematics education; Pedagogy; Nursing; Alternative medicine; Psychology","score_opus":0.018687297509815164,"score_gpt":0.3654858127832671,"score_spread":0.3467985152734519,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3132306444","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98931706,0.0017426821,0.000832928,0.0072357454,0.0005380155,0.00012032772,9.416156e-7,0.0000047887993,0.0002075392],"genre_scores_gemma":[0.99815255,0.00038961484,0.0006277491,0.00030261028,0.00046724215,0.000008239884,0.000019950816,0.000007949202,0.00002408128],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9977067,0.00010280475,0.0010517276,0.00014509336,0.0008853318,0.00010835133],"domain_scores_gemma":[0.99525666,0.0019523213,0.0006120886,0.00013679666,0.001892922,0.00014922052],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0021110645,0.00009340023,0.00031257514,0.00014808937,0.00003349547,0.000012374096,0.0001607216,0.00007602746,0.000013468353],"category_scores_gemma":[0.027323784,0.000058359128,0.000038421953,0.00020608667,0.00013118719,0.000089946116,0.00003122764,0.00031745294,0.0000014803188],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005234812,0.0006760486,0.63997835,0.000095800446,0.00010111267,0.00045466973,0.0072397795,0.000052108477,0.00455346,0.0011251498,0.00006845985,0.34513158],"study_design_scores_gemma":[0.012493527,0.0022498805,0.93459404,0.013662229,0.00023999505,0.003282879,0.012129815,0.0038021782,0.011417388,0.0039784005,0.0019087695,0.0002408843],"about_ca_topic_score_codex":0.000012652965,"about_ca_topic_score_gemma":0.000082096456,"teacher_disagreement_score":0.34489068,"about_ca_system_score_codex":0.00007190692,"about_ca_system_score_gemma":0.0005524042,"threshold_uncertainty_score":0.9808695},"labels":[],"label_agreement":null},{"id":"W3134009116","doi":"10.1002/aet2.10592","title":"Image interpretation: Learning analytics–informed education opportunities","year":2021,"lang":"en","type":"article","venue":"AEM Education and Training","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; SickKids Foundation; Hospital for Sick Children; University of Toronto","funders":"Hospital for Sick Children","keywords":"Pneumonia; Medicine; Chest radiograph; Radiography; Confidence interval; Interpretation (philosophy); Radiology; Internal medicine","score_opus":0.07781740510588091,"score_gpt":0.39086293487865786,"score_spread":0.31304552977277694,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3134009116","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8069105,0.0013125436,0.00061271514,0.009505471,0.00096023694,0.00013951283,0.0000018267767,0.000098444005,0.18045874],"genre_scores_gemma":[0.97853863,0.0004813017,0.006448541,0.002899579,0.00026398472,0.00002292892,0.0002979054,0.000012544528,0.011034571],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9992833,0.000035560497,0.00025069635,0.00017780665,0.00012276343,0.00012988808],"domain_scores_gemma":[0.99789673,0.0013866606,0.00008539027,0.00012352553,0.00029719554,0.00021052583],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00017212579,0.00009043337,0.00017325416,0.000086641194,0.00010055184,0.00007860714,0.000024414863,0.000054659384,0.00036008662],"category_scores_gemma":[0.013441458,0.000089885085,0.000055110493,0.00013926001,0.000063287065,0.00012125207,0.000015313297,0.00017084679,0.000022231548],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000017458837,0.00023011895,0.006442153,0.000050376057,0.0000349174,0.0000066467096,0.0059186537,0.0000021154376,0.00006451261,0.0029012403,0.0026325015,0.9816993],"study_design_scores_gemma":[0.0031549004,0.0007751138,0.43319604,0.008631416,0.0009316885,0.0018013648,0.3844902,0.007021914,0.0007551062,0.004254297,0.15408848,0.0008994888],"about_ca_topic_score_codex":0.000017095646,"about_ca_topic_score_gemma":0.000006454821,"teacher_disagreement_score":0.9807998,"about_ca_system_score_codex":0.000037816135,"about_ca_system_score_gemma":0.003227039,"threshold_uncertainty_score":0.99486876},"labels":[],"label_agreement":null},{"id":"W3134270660","doi":"10.1515/dx-2020-0127","title":"Real-world virtual patient simulation to improve diagnostic performance through deliberate practice: a prospective quasi-experimental study","year":2021,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":20,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital","funders":"National Institute on Deafness and Other Communication Disorders; Center for Innovative Medicine, Johns Hopkins University","keywords":"Computer science; Psychology","score_opus":0.025065914944485443,"score_gpt":0.37698990767724905,"score_spread":0.3519239927327636,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3134270660","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99202937,0.00025834798,0.00009999731,0.0018114476,0.0005971147,0.0021896516,0.000032869593,0.00014919614,0.0028319848],"genre_scores_gemma":[0.99260014,0.00042331737,0.001359068,0.0029669423,0.00033746407,0.001744252,0.00006293446,0.00006191986,0.00044394317],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.997068,0.00019787751,0.0006915398,0.00087289076,0.0007071779,0.00046253004],"domain_scores_gemma":[0.97362566,0.024554055,0.0001971737,0.0006084045,0.00065292895,0.0003617805],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00028622252,0.00033516434,0.00058710144,0.00009318723,0.00020373803,0.00009705649,0.00009567429,0.00010152862,0.00036626813],"category_scores_gemma":[0.07217759,0.00030807863,0.00015583256,0.0007856433,0.000055944198,0.00032763893,0.00018948474,0.0003444817,0.00026606553],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013559395,0.057785876,0.8731992,0.000034999168,0.0004930607,0.0012191489,0.016401127,0.007355699,0.00016944406,0.0004010323,0.0026184479,0.03896598],"study_design_scores_gemma":[0.014569326,0.07366014,0.7747601,0.0037156823,0.0018084634,0.00006541577,0.025260236,0.007602632,0.08390136,0.00009396924,0.0128262695,0.0017364067],"about_ca_topic_score_codex":0.00045266442,"about_ca_topic_score_gemma":0.00008422361,"teacher_disagreement_score":0.09843914,"about_ca_system_score_codex":0.0003935968,"about_ca_system_score_gemma":0.0002575943,"threshold_uncertainty_score":0.9999371},"labels":[],"label_agreement":null},{"id":"W3134800294","doi":"10.5430/jct.v10n1p36","title":"Exploration of an Effective Method for the Step-by-Step Presentation of Case Information to Guide Grade 4 Medical Students to Develop Clinical Reasoning Skills","year":2021,"lang":"en","type":"article","venue":"Journal of Curriculum and Teaching","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Jinan University","keywords":"Presentation (obstetrics); Competence (human resources); Medical education; Psychology; Computer science; Medicine; Surgery","score_opus":0.02524730887196545,"score_gpt":0.45553534621618613,"score_spread":0.43028803734422066,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3134800294","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.56794536,0.000086108084,0.4302585,0.0010948369,0.00021984962,0.0003565852,0.000006045151,0.0000036332294,0.000029061028],"genre_scores_gemma":[0.79212964,0.00008183522,0.20663285,0.000783077,0.0002955491,0.000021914366,0.00002203943,0.0000098648625,0.000023215616],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9970929,0.0005595513,0.0013006735,0.00012579981,0.0007924672,0.00012860075],"domain_scores_gemma":[0.9862129,0.011664647,0.00065203814,0.00013877993,0.0009759837,0.00035564194],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0077189924,0.00010344326,0.0004722583,0.00009632403,0.00009653504,0.000044233104,0.00012259437,0.00010493639,0.000005711499],"category_scores_gemma":[0.16220458,0.00006676225,0.00011944943,0.00017868518,0.000026779868,0.00040337257,0.00008169454,0.00036230104,5.3333343e-7],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00035774463,0.001551424,0.07793294,0.000116835574,0.00040157884,0.00017459164,0.005041715,0.00025347236,0.00019241881,0.00049660204,0.012229513,0.90125114],"study_design_scores_gemma":[0.04422414,0.022550462,0.67057025,0.03174171,0.005732094,0.015682684,0.08949542,0.06536576,0.012816287,0.000551623,0.04005125,0.0012182815],"about_ca_topic_score_codex":0.00019719497,"about_ca_topic_score_gemma":0.000039014427,"teacher_disagreement_score":0.9000329,"about_ca_system_score_codex":0.000038783095,"about_ca_system_score_gemma":0.00019328634,"threshold_uncertainty_score":0.8448525},"labels":[],"label_agreement":null},{"id":"W3134994971","doi":"10.1080/09593985.2021.1895387","title":"Augmented behavioral medicine competencies in physical therapy students’ clinical reasoning with a targeted curriculum: a final-semester cohort-comparison study","year":2021,"lang":"en","type":"article","venue":"Physiotherapy Theory and Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Curriculum; Cohort; Medical education; Medicine; Physical therapy education; Physical therapy; Psychology; Physical therapist; Pedagogy; Internal medicine","score_opus":0.06256641617932189,"score_gpt":0.4830603475245219,"score_spread":0.4204939313452,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3134994971","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9944025,0.0024840366,0.0001852775,0.0013472999,0.00015694101,0.0009791735,0.0000033391343,0.00007385848,0.00036756453],"genre_scores_gemma":[0.99433273,0.001589589,0.0008341945,0.0023627328,0.00037469537,0.00010657261,0.00004078915,0.000047013426,0.00031168564],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99454844,0.002844392,0.00072169176,0.00076432386,0.0007518214,0.00036930613],"domain_scores_gemma":[0.9776907,0.02082851,0.0003546368,0.0005515384,0.0003135087,0.00026109585],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0034918003,0.000367578,0.0011828003,0.00008012277,0.00014026614,0.00007349806,0.00015587766,0.000105516934,0.00016663231],"category_scores_gemma":[0.010065188,0.00024325267,0.00012516913,0.00053417264,0.00035532878,0.0002396451,0.000068137044,0.0007732388,0.0000074444733],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.012140729,0.027099695,0.93896765,0.000015870912,0.00066556386,0.0005905752,0.006863276,0.0000064097308,0.00084866065,0.0007290732,0.000195338,0.01187715],"study_design_scores_gemma":[0.021276996,0.010165704,0.94424266,0.0015090768,0.0007108367,0.00018340505,0.018977445,0.00017408424,0.00014361595,0.0005255658,0.0016903644,0.0004002345],"about_ca_topic_score_codex":0.0005068317,"about_ca_topic_score_gemma":0.00007319032,"teacher_disagreement_score":0.017984116,"about_ca_system_score_codex":0.00004160865,"about_ca_system_score_gemma":0.00016428748,"threshold_uncertainty_score":0.99827343},"labels":[],"label_agreement":null},{"id":"W3135633933","doi":"10.1016/j.ophtha.2021.03.008","title":"Patient Harm Due to Diagnostic Error of Neuro-Ophthalmologic Conditions","year":2021,"lang":"en","type":"article","venue":"Ophthalmology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":52,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"National Institute of Neurological Disorders and Stroke; National Eye Institute; Washington University in St. Louis; National Institutes of Health; School of Medicine, Emory University; Emory University; Research to Prevent Blindness","keywords":"Medicine; Referral; Neuro-ophthalmology; Pediatrics; Emergency medicine; Ophthalmology; Family medicine; Glaucoma","score_opus":0.044943183053384156,"score_gpt":0.3682393070267502,"score_spread":0.32329612397336605,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3135633933","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9819975,0.0002620435,0.000028702092,0.0040435246,0.00073464325,0.0003333297,0.00008942819,0.0000357837,0.012475071],"genre_scores_gemma":[0.9957619,0.000011696912,0.0013118548,0.0018221964,0.000094014926,0.00008745399,0.00014349249,0.000024997562,0.0007424497],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982581,0.00016453379,0.0005231137,0.00047164317,0.0002160226,0.00036656292],"domain_scores_gemma":[0.97926456,0.019380355,0.00012467333,0.0005473872,0.0003091027,0.00037394144],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00013209067,0.00018098557,0.00063298485,0.00010239073,0.000056667428,0.0000067102314,0.00009464611,0.00018379492,0.003971685],"category_scores_gemma":[0.10943639,0.0001649761,0.00019263853,0.00037658133,0.00016441463,0.00002059297,0.00014716247,0.00024341293,0.00041409297],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006568029,0.008528825,0.6140975,0.00022057298,0.00042235537,0.31710047,0.0008378242,0.0005195772,0.017330592,0.0046918383,0.030306963,0.0052866857],"study_design_scores_gemma":[0.0014802958,0.0039679063,0.8816223,0.0006580877,0.00022820562,0.09979428,0.00021790675,0.00003085173,0.008456971,0.0023250952,0.0009435592,0.0002744952],"about_ca_topic_score_codex":0.00005542188,"about_ca_topic_score_gemma":7.6297124e-7,"teacher_disagreement_score":0.26752484,"about_ca_system_score_codex":0.000031826003,"about_ca_system_score_gemma":0.00022106602,"threshold_uncertainty_score":0.9969388},"labels":[],"label_agreement":null},{"id":"W3138676254","doi":"10.1038/s41746-021-00423-6","title":"The potential of artificial intelligence to improve patient safety: a scoping review","year":2021,"lang":"en","type":"review","venue":"npj Digital Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":288,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Canadian Institutes of Health Research","keywords":"Context (archaeology); Patient safety; Harm; Leverage (statistics); Health care; Computer science; Artificial intelligence; MEDLINE; Adverse effect; Medicine; Risk analysis (engineering); Intensive care medicine; Machine learning; Data science; Psychology","score_opus":0.06301976923425544,"score_gpt":0.4160997882411918,"score_spread":0.35308001900693636,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3138676254","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000002694599,0.99086636,0.00079341215,0.0021674442,0.0012316284,0.00290023,0.00007163677,0.0000336575,0.0019329364],"genre_scores_gemma":[0.000046931327,0.9973849,0.00010266372,0.00097018573,0.00082786253,0.00014234883,0.00028079568,0.000052726533,0.00019157573],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9953353,0.000104759194,0.002630238,0.00062064803,0.00088580186,0.00042328003],"domain_scores_gemma":[0.9859941,0.011391816,0.0007243085,0.00096202386,0.00042880338,0.00049895304],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00072823477,0.000452082,0.0035818429,0.000104240135,0.00008167254,0.00003848979,0.00028517775,0.00021105078,0.00018197233],"category_scores_gemma":[0.13148585,0.00024367338,0.000723413,0.00074355234,0.00032613717,0.000039635634,0.00025927622,0.00053593784,0.00011788901],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000029993535,0.00011212324,2.759915e-7,0.028112013,0.00012698995,0.00011950139,0.000015902293,2.1007072e-7,3.073374e-7,0.00028988943,0.0011750151,0.9700178],"study_design_scores_gemma":[0.00006487929,0.00096486777,5.0719103e-7,0.735522,0.0012982191,0.00012553357,0.00005077211,0.0000018697907,0.00000264709,0.000078590325,0.2617266,0.00016354288],"about_ca_topic_score_codex":0.000012012706,"about_ca_topic_score_gemma":0.0000028674094,"teacher_disagreement_score":0.96985424,"about_ca_system_score_codex":0.000098456556,"about_ca_system_score_gemma":0.0009740477,"threshold_uncertainty_score":0.99367124},"labels":[],"label_agreement":null},{"id":"W3138881536","doi":"10.5430/jnep.v11n6p73","title":"Script concordance testing to understand the hypothesis processes of undergraduate nursing students - Multiple case study","year":2021,"lang":"en","type":"article","venue":"Journal of Nursing Education and Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Concordance; Think aloud protocol; Psychology; Phenomenon; Nursing practice; Clinical Practice; Representation (politics); Nurse education; Nursing; Medicine; Epistemology; Computer science","score_opus":0.20563365206049586,"score_gpt":0.4890247646838655,"score_spread":0.2833911126233697,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3138881536","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9634248,0.0017755564,0.0005495668,0.032111835,0.00070972746,0.00021515648,9.040967e-7,0.0000054529105,0.0012070056],"genre_scores_gemma":[0.988679,0.00013444803,0.00975975,0.0010106334,0.00020118838,0.0000015639255,3.5327577e-7,0.000010962243,0.00020210522],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.9986583,0.0002786514,0.00044440868,0.0001314605,0.00038030636,0.00010685993],"domain_scores_gemma":[0.9632575,0.03414508,0.0006036302,0.00016921277,0.0016295587,0.00019505607],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010022661,0.0000859864,0.00024032228,0.00006857416,0.00021316581,0.000075928314,0.000062067964,0.000028960998,0.0000035792218],"category_scores_gemma":[0.22728002,0.00006189919,0.000033553424,0.00044892688,0.00008130215,0.00017403711,0.000012669848,0.00022082716,6.735456e-7],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0020507004,0.04810777,0.14852147,0.00021299167,0.0005354251,0.0019456082,0.12573677,0.00010292549,0.0013402912,0.00009295283,0.011991288,0.6593618],"study_design_scores_gemma":[0.0031129296,0.00214203,0.08294744,0.0194389,0.002287976,0.10071629,0.78633505,0.000051621213,0.0007023392,0.0013837614,0.00069623336,0.00018539935],"about_ca_topic_score_codex":0.000059084028,"about_ca_topic_score_gemma":0.000010128631,"teacher_disagreement_score":0.6605983,"about_ca_system_score_codex":0.00012258897,"about_ca_system_score_gemma":0.0012765271,"threshold_uncertainty_score":0.7792289},"labels":[],"label_agreement":null},{"id":"W3146393752","doi":"10.1096/fasebj.2019.33.1_supplement.604.2","title":"Making Connections Explicit: The Effect of Self‐explanation on cognitive Integration of Basic and Clinical Sciences in Novices","year":2019,"lang":"en","type":"article","venue":"The FASEB Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; The Wilson Centre; Humber Polytechnic; University Health Network","funders":"","keywords":"Cognition; Task (project management); Basic science; Psychology; Cognitive psychology; Medicine; Neuroscience","score_opus":0.055196079070173314,"score_gpt":0.40572512136153865,"score_spread":0.35052904229136533,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3146393752","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99757046,0.00009280402,0.00007704764,0.0007376257,0.00017831684,0.0002221886,0.000002672327,0.000003989571,0.001114911],"genre_scores_gemma":[0.99954104,0.00012530712,0.000032052718,0.00020209217,0.00008207505,0.00000230112,0.0000012174849,0.0000028479135,0.0000110475585],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9989984,0.00029804904,0.00033284447,0.000087508844,0.00020762773,0.00007556594],"domain_scores_gemma":[0.970082,0.029505473,0.0002477062,0.000070079885,0.0000701232,0.000024659133],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0029721272,0.000055169898,0.00019608118,0.00007747654,0.00007669141,0.00001462364,0.000061107516,0.00003948813,0.000030632324],"category_scores_gemma":[0.01342851,0.000024473386,0.000065275555,0.00016779242,0.00014263281,0.000041109895,0.00001331844,0.00030370784,0.0000049327905],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008383729,0.00025283033,0.961977,0.00002410383,0.000083036786,0.0000049355835,0.0011849721,0.00006603591,0.0021256446,0.0010189469,0.000050415656,0.03237373],"study_design_scores_gemma":[0.004469791,0.009275927,0.94811827,0.007181749,0.00030879088,0.00017160515,0.003528884,0.0052256887,0.02131348,0.00032469316,0.000007430258,0.00007367458],"about_ca_topic_score_codex":0.000027664391,"about_ca_topic_score_gemma":0.000024100716,"teacher_disagreement_score":0.032300055,"about_ca_system_score_codex":0.000013160676,"about_ca_system_score_gemma":0.000052666313,"threshold_uncertainty_score":0.9948818},"labels":[],"label_agreement":null},{"id":"W3146496327","doi":"10.1111/medu.14529","title":"The medical pause: Importance, processes and training","year":2021,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":34,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"Nederlandse Organisatie voor Wetenschappelijk Onderzoek","keywords":"Acknowledgement; Cognition; Relaxation (psychology); Interrupt; Reflection (computer programming); Training (meteorology); Computer science; Curriculum; Function (biology); Cognitive training; Control (management); Psychology; Medical education; Cognitive psychology; Applied psychology; Medicine; Pedagogy; Social psychology; Artificial intelligence","score_opus":0.021679172766640702,"score_gpt":0.3774527011484503,"score_spread":0.3557735283818096,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3146496327","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.73752093,0.012531622,0.00016914688,0.23737368,0.0011589694,0.00016245092,0.0000011505749,0.000075269774,0.011006766],"genre_scores_gemma":[0.9666435,0.004710828,0.00035638447,0.024505321,0.0011106358,0.000061842395,0.00007156277,0.000015014232,0.0025249005],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9984212,0.000046056168,0.00029714787,0.00021895807,0.00082628,0.0001903641],"domain_scores_gemma":[0.98948276,0.009167416,0.00007198931,0.00025908282,0.0003048885,0.0007138831],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0008086014,0.00007736807,0.00017283967,0.000015791587,0.0001238015,0.000030569252,0.00007280085,0.00015693286,0.0010961987],"category_scores_gemma":[0.44640028,0.000047702066,0.000029485449,0.00023327838,0.00022522568,0.000025442927,0.00003419227,0.000314142,0.000018159055],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003103789,0.00057118945,0.04733256,0.000093588984,0.000037646994,0.00016505051,0.0006659361,2.0208438e-8,0.0000052902437,0.0017677569,0.07894315,0.8703868],"study_design_scores_gemma":[0.0035799842,0.00034191823,0.2719717,0.009139091,0.00027790974,0.0042330585,0.00642711,0.00035577238,0.00022955688,0.0064598927,0.6965866,0.00039737238],"about_ca_topic_score_codex":0.000017021142,"about_ca_topic_score_gemma":0.00013290688,"teacher_disagreement_score":0.8699894,"about_ca_system_score_codex":0.000020039812,"about_ca_system_score_gemma":0.009625676,"threshold_uncertainty_score":0.99981695},"labels":[],"label_agreement":null},{"id":"W3151445967","doi":"10.2196/25046","title":"Barriers to the Use of Clinical Decision Support for the Evaluation of Pulmonary Embolism: Qualitative Interview Study","year":2021,"lang":"en","type":"article","venue":"JMIR Human Factors","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"U.S. National Library of Medicine; National Heart, Lung, and Blood Institute; National Institutes of Health","keywords":"Qualitative research; Medicine; Pulmonary embolism; Clinical decision support system; Decision support system; Psychology; Applied psychology; Medical emergency; Surgery; Artificial intelligence; Computer science","score_opus":0.4294143737453224,"score_gpt":0.5712241916642262,"score_spread":0.14180981791890385,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3151445967","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9961305,0.000100244564,0.00038082848,0.00079353567,0.0003710904,0.002087743,0.00006212747,0.000010199374,0.00006368376],"genre_scores_gemma":[0.9987198,0.000019121784,0.00015954801,0.00049390405,0.00010487698,0.00016522738,0.000049429498,0.000015804346,0.00027226727],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9965756,0.0010199602,0.0011055992,0.00029370753,0.0008591801,0.00014598425],"domain_scores_gemma":[0.9488754,0.048783395,0.00031838982,0.00068175455,0.0011210209,0.00022001732],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.006639004,0.0001265075,0.00056521455,0.00005124454,0.00010245202,0.000017133802,0.00017728917,0.00008405447,0.00064880954],"category_scores_gemma":[0.17730355,0.00006473911,0.0003696372,0.00023243227,0.00012828487,0.00004397525,0.00011228545,0.00018854272,0.000006625646],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009966296,0.0038157797,0.52516973,0.000099545,0.0009614954,0.000016554415,0.0788639,0.00018843925,0.0001216734,0.00036737742,0.028007211,0.3613917],"study_design_scores_gemma":[0.0018146321,0.0013924486,0.9582708,0.0006224077,0.00075020525,0.0000017526427,0.03220804,0.00037420916,0.00013722919,0.00024736085,0.0040939166,0.00008700955],"about_ca_topic_score_codex":0.000055627585,"about_ca_topic_score_gemma":0.00008907786,"teacher_disagreement_score":0.43310106,"about_ca_system_score_codex":0.000029425953,"about_ca_system_score_gemma":0.0003874711,"threshold_uncertainty_score":0.8296264},"labels":[],"label_agreement":null},{"id":"W3152785930","doi":"10.1136/bmjebm-2020-111633","title":"Does my patient have SARS-CoV-2 infection? A reminder of clinical probability formulas","year":2021,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Jewish General Hospital","funders":"","keywords":"Bayes' theorem; Conditional probability; Law of total probability; Table (database); Probability theory; Contingency table; Medicine; Mathematics; Computer science; Posterior probability; Calculus (dental); Bayesian probability; Artificial intelligence; Statistics; Data mining","score_opus":0.1913117687147357,"score_gpt":0.47436889575860586,"score_spread":0.28305712704387015,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3152785930","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96181273,0.0008523674,0.0013362578,0.031716183,0.001643763,0.0013000893,0.000009497173,0.00009346643,0.0012356656],"genre_scores_gemma":[0.98647124,0.0004932723,0.003546288,0.007888218,0.0011114527,0.00011533307,0.000048551312,0.000029451767,0.0002962125],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9948089,0.0004699541,0.0023966187,0.0007862552,0.0011334099,0.00040483952],"domain_scores_gemma":[0.96293515,0.033021394,0.0007500449,0.001319625,0.00154906,0.00042475035],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.003713264,0.000300676,0.001436571,0.000120350065,0.00006687285,0.000010085887,0.00011258422,0.0003408685,0.0005234384],"category_scores_gemma":[0.412501,0.00016628749,0.00047966142,0.00047170944,0.00074828265,0.00011051204,0.00008970777,0.0006133913,0.000031619085],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017880611,0.0019641037,0.9112291,0.0006804256,0.0001677899,0.00043372545,0.0002710364,0.000019674813,0.002270046,0.00017410812,0.020851284,0.060150664],"study_design_scores_gemma":[0.013310791,0.008046188,0.7696353,0.07243218,0.0015000924,0.0001774972,0.0003107197,0.001096119,0.11464987,0.0041978667,0.014065153,0.00057822425],"about_ca_topic_score_codex":0.0006257737,"about_ca_topic_score_gemma":0.00023072063,"teacher_disagreement_score":0.40878776,"about_ca_system_score_codex":0.00011718012,"about_ca_system_score_gemma":0.0018659276,"threshold_uncertainty_score":0.6781007},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W3155491546","doi":"10.2196/24642","title":"Monitoring Diagnostic Safety Risks in Emergency Departments: Protocol for a Machine Learning Study","year":2021,"lang":"en","type":"article","venue":"JMIR Research Protocols","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Agency for Healthcare Research and Quality; Health Services Research and Development; U.S. Department of Veterans Affairs","keywords":"Emergency department; Patient safety; Computer science; Health care; Medical record; Data collection; Clinical decision support system; Machine learning; Medicine; MEDLINE; Protocol (science); Medical emergency; Decision support system; Artificial intelligence; Data science; Nursing","score_opus":0.43891745360220696,"score_gpt":0.6441882560261594,"score_spread":0.2052708024239524,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3155491546","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.018491114,0.000008821996,0.000047801343,0.00042708684,0.000018860172,0.9800252,0.00001200824,0.00009382095,0.0008752918],"genre_scores_gemma":[0.024287108,0.000005516148,0.0005367905,0.00001482259,0.00038040153,0.97318107,0.000029324085,0.00006462436,0.0015003174],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9950568,0.0010211681,0.0009925422,0.0008037669,0.0011450357,0.0009806593],"domain_scores_gemma":[0.9828851,0.015092512,0.00013272506,0.00063989346,0.00081182795,0.0004379053],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0035727306,0.00025715964,0.0005726269,0.0002347968,0.00031592618,0.000082829785,0.00023500738,0.00016728112,0.00055625755],"category_scores_gemma":[0.15511893,0.0002203401,0.00017503604,0.0010160918,0.00007200149,0.000102850565,0.0003214983,0.0013981827,0.00007672953],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0023604273,0.009067226,0.9659442,0.0005492047,0.0000522499,0.0006004262,0.00031192088,0.00003014771,0.00016241429,0.000022775363,0.0007933551,0.020105675],"study_design_scores_gemma":[0.029033108,0.009630228,0.8760032,0.0151842255,0.000015974794,0.000023399221,0.0006919928,0.00032765043,0.0018530327,0.00054492214,0.06634076,0.00035151668],"about_ca_topic_score_codex":0.0001683477,"about_ca_topic_score_gemma":0.000115048315,"teacher_disagreement_score":0.1515462,"about_ca_system_score_codex":0.00027165253,"about_ca_system_score_gemma":0.0005787039,"threshold_uncertainty_score":0.8985208},"labels":[],"label_agreement":null},{"id":"W3157388129","doi":"10.1016/j.jss.2021.03.057","title":"The Impact of Surgeon Experience on Script Concordance Test Scoring","year":2021,"lang":"en","type":"article","venue":"Journal of Surgical Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"University of Ottawa","keywords":"Concordance; Medicine; Test (biology); Specialty; Subgroup analysis; Family medicine; Meta-analysis; Internal medicine","score_opus":0.14822989427597108,"score_gpt":0.5095313188142345,"score_spread":0.3613014245382634,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3157388129","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9902497,0.0020540713,0.0000047377903,0.0024657734,0.00017975761,0.000067445646,0.0000027030849,0.0000033316467,0.004972458],"genre_scores_gemma":[0.99388075,0.004643085,0.000047121204,0.000024595827,0.00033170884,0.0000023071436,6.3622593e-7,0.000009504435,0.0010603194],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9971098,0.0002579504,0.00063856854,0.00014890224,0.0014426168,0.00040221214],"domain_scores_gemma":[0.90474665,0.09197557,0.00022101616,0.00037374208,0.0021357592,0.0005472858],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.003758039,0.00008618001,0.00041665556,0.000091951544,0.00012887373,0.000049321992,0.00019220791,0.00006916448,0.0002587737],"category_scores_gemma":[0.18118487,0.000044506432,0.00036667718,0.00058494764,0.00040512485,0.000050922426,0.000080072714,0.0010091802,0.00001726693],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0050359834,0.0027550717,0.8305382,0.00004520161,0.00021327486,0.015100199,0.00045310665,0.000097930504,0.012277414,0.0012762304,0.021611044,0.110596344],"study_design_scores_gemma":[0.00818561,0.01147542,0.8854755,0.013151445,0.000048660222,0.004153841,0.0024668383,0.0004079991,0.05293246,0.0015099924,0.019940753,0.00025145756],"about_ca_topic_score_codex":0.00008474499,"about_ca_topic_score_gemma":0.0000029476914,"teacher_disagreement_score":0.17742684,"about_ca_system_score_codex":0.0001369166,"about_ca_system_score_gemma":0.0010573228,"threshold_uncertainty_score":0.8257123},"labels":[],"label_agreement":null},{"id":"W3159235874","doi":"10.1097/nne.0000000000001028","title":"Script Concordance Approach in Nursing Education","year":2021,"lang":"en","type":"article","venue":"Nurse Educator","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Interdisciplinary Research in Rehabilitation","funders":"","keywords":"Concordance; Clinical Practice; Psychology; Computer science; Nurse education; Medical education; Medicine; Nursing","score_opus":0.023537772484996916,"score_gpt":0.38284334397395375,"score_spread":0.3593055714889568,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3159235874","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8764648,0.0026982515,0.000054227075,0.005364425,0.0031577426,0.00018615469,0.0000021368942,0.000039721148,0.11203258],"genre_scores_gemma":[0.98449177,0.0000487789,0.006379171,0.0021946856,0.00085805834,0.000052030082,0.00007969047,0.000020617443,0.0058752103],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99898905,0.00004816715,0.00024582364,0.00032397677,0.00016567686,0.00022731577],"domain_scores_gemma":[0.99854696,0.00070283195,0.0000480881,0.00033767073,0.00012541021,0.00023902275],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0001392502,0.000104883504,0.0002500231,0.000067523666,0.000029904902,0.000020840758,0.000053018888,0.00009032171,0.00023773748],"category_scores_gemma":[0.008945416,0.000102490536,0.00006888434,0.00034800955,0.00008990065,0.00004016037,0.0000054374395,0.00024167489,0.00010624691],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013279764,0.016673757,0.39644802,0.000084458334,0.000026653961,0.000054319124,0.003521713,0.000003951112,0.0009469942,0.01226257,0.19898534,0.37085944],"study_design_scores_gemma":[0.0025394058,0.00016246001,0.94643074,0.009615656,0.00015648939,0.00054068223,0.019518726,0.00016223193,0.002945646,0.0024678467,0.0150755,0.000384634],"about_ca_topic_score_codex":0.00003443203,"about_ca_topic_score_gemma":0.0000023302061,"teacher_disagreement_score":0.5499827,"about_ca_system_score_codex":0.00017654512,"about_ca_system_score_gemma":0.0024321724,"threshold_uncertainty_score":0.99940264},"labels":[],"label_agreement":null},{"id":"W3159623101","doi":"10.1136/bmjqs-2020-012464","title":"Bridging the feedback gap: a sociotechnical approach to informing clinicians of patients’ subsequent clinical course and outcomes","year":2021,"lang":"en","type":"editorial","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":30,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"Agency for Healthcare Research and Quality","keywords":"Medicine; Overconfidence effect; Emergency department; Audit; Outcome (game theory); Medical diagnosis; MEDLINE; Patient safety; Medical emergency; Hospital medicine; Health care; Nursing; Family medicine","score_opus":0.07642136729722349,"score_gpt":0.4567869062840934,"score_spread":0.3803655389868699,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3159623101","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.1979025,0.0012737755,0.0035466754,0.028830051,0.7513631,0.008325885,0.0030737787,0.00039979877,0.0052844156],"genre_scores_gemma":[0.14180407,0.0028014586,0.0076975017,0.008507658,0.8296993,0.00032796088,0.0067461664,0.00029554605,0.0021203875],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9897182,0.0009922526,0.0055041504,0.0010375626,0.002109475,0.0006383879],"domain_scores_gemma":[0.8777289,0.11576871,0.0020475532,0.0021335515,0.0014253629,0.0008959306],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.014346019,0.0005692999,0.0029208227,0.00006658224,0.0001921249,0.000061095845,0.0004605517,0.0018825296,0.000028653443],"category_scores_gemma":[0.3582356,0.00039200956,0.0011901315,0.00030128486,0.00073870306,0.00006239543,0.0006681969,0.0028560078,0.000022307715],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013706,0.0023097051,0.09694525,0.0005210368,0.0006058267,0.000024892724,0.00049682125,0.000008061372,3.683194e-7,0.00024200855,0.83843756,0.059037898],"study_design_scores_gemma":[0.0105479,0.0009954431,0.6917344,0.0050019803,0.0020846338,0.000009159718,0.0014028929,0.000118506396,0.0000039702168,0.00012958549,0.28691873,0.001052819],"about_ca_topic_score_codex":0.0004228142,"about_ca_topic_score_gemma":0.0000557168,"teacher_disagreement_score":0.59478915,"about_ca_system_score_codex":0.00019109494,"about_ca_system_score_gemma":0.002557699,"threshold_uncertainty_score":0.9998532},"labels":[],"label_agreement":null},{"id":"W3160785994","doi":"10.1007/s43678-021-00137-4","title":"CJEM Debate Series: what’s in a name? It is simply an emergency room, and we are ERPs!","year":2021,"lang":"en","type":"editorial","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Horizon Health Network; Nova Scotia Health Authority; Dalhousie University; University of British Columbia; Island Health","funders":"","keywords":"Medicine; Series (stratigraphy); Emergency department; Medical emergency; Computer security; Psychiatry; Computer science","score_opus":0.06295552953757176,"score_gpt":0.3789861255566007,"score_spread":0.31603059601902894,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3160785994","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.007546821,0.09708702,0.000009528893,0.04755275,0.8463806,0.00031597575,0.00015825297,0.000012357929,0.0009367287],"genre_scores_gemma":[0.0021033643,0.2971748,0.00008853561,0.0014463154,0.6873433,0.000015580443,0.0003010008,0.00016410486,0.011362974],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99380296,0.00023584545,0.0029597876,0.0007491445,0.0012944731,0.0009577822],"domain_scores_gemma":[0.9893117,0.0020489458,0.0014337647,0.00078088645,0.002151739,0.00427299],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0016517318,0.00070967805,0.0023073326,0.0008733287,0.00015063363,0.000036836922,0.0004696052,0.0011498783,0.020463657],"category_scores_gemma":[0.095487714,0.00059554086,0.000379966,0.0008434995,0.00017606627,0.0005134157,0.00004322549,0.0025435993,0.00001589994],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015434533,0.000098072225,0.006654326,0.0003164435,0.0002669568,0.004477691,0.0026594936,0.0000029995826,0.0000060400307,0.000007098481,0.980578,0.0047785337],"study_design_scores_gemma":[0.0026101777,0.0021280805,0.0019009664,0.02739803,0.0008470737,0.00017171526,0.0059560556,0.00001547249,0.0000031806385,0.00051010074,0.95794713,0.00051201566],"about_ca_topic_score_codex":0.013885506,"about_ca_topic_score_gemma":0.16116919,"teacher_disagreement_score":0.2000878,"about_ca_system_score_codex":0.0003897027,"about_ca_system_score_gemma":0.0062453263,"threshold_uncertainty_score":0.9997576},"labels":[],"label_agreement":null},{"id":"W3161461325","doi":"10.1177/14778785211009105","title":"Equipoise and ethics in educational research","year":2021,"lang":"en","type":"article","venue":"Theory and Research in Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Clinical equipoise; Research ethics; Psychology; Educational research; Engineering ethics; Sociology; Pedagogy; Medicine; Clinical trial","score_opus":0.2724044869545877,"score_gpt":0.5867358006625898,"score_spread":0.3143313137080021,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3161461325","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9396982,0.0061987406,0.000001618497,0.041326616,0.00013655426,0.0001949642,0.0000011432335,0.0000030565418,0.0124391],"genre_scores_gemma":[0.98972726,0.0041689295,0.00021362602,0.00034546776,0.00024640613,0.00007076488,0.000030695588,0.000007418708,0.0051894244],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99652284,0.0022859653,0.00018321071,0.00027783695,0.00043897473,0.00029117105],"domain_scores_gemma":[0.917861,0.08102819,0.00001528834,0.00025033907,0.0005962468,0.00024893045],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.017340863,0.000049573195,0.000119531316,0.00036986708,0.00012816438,0.00004007385,0.00004651226,0.00014916626,0.00024840655],"category_scores_gemma":[0.15793014,0.00004368715,0.000012114692,0.00072864676,0.00049510965,0.00005037092,0.00009296661,0.0014968745,0.000026289637],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00064502016,0.0028218206,0.14303836,0.000297458,0.000011243497,0.000041407733,0.009343374,0.0000013847074,0.0014767036,0.7108716,0.0028921918,0.12855943],"study_design_scores_gemma":[0.0006638614,0.0001299176,0.47720662,0.002607825,0.0000039360525,0.000060654544,0.008665703,0.000017160837,0.0004119281,0.50801545,0.0021544297,0.00006250794],"about_ca_topic_score_codex":0.00020309768,"about_ca_topic_score_gemma":0.0001168357,"teacher_disagreement_score":0.33416826,"about_ca_system_score_codex":0.00009979718,"about_ca_system_score_gemma":0.0032848518,"threshold_uncertainty_score":0.849163},"labels":[],"label_agreement":null},{"id":"W3161895874","doi":"10.1080/0142159x.2021.1925642","title":"On the validity of summative entrustment decisions","year":2021,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":34,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta; University of Toronto; University of Ottawa; Medical Council of Canada","funders":"","keywords":"Summative assessment; Argument (complex analysis); Psychology; Process (computing); Health care; External validity; Nursing; Medical education; Formative assessment; Medicine; Social psychology; Computer science; Pedagogy; Political science","score_opus":0.09308446346101831,"score_gpt":0.3871435513230113,"score_spread":0.294059087861993,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3161895874","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9437332,0.00014298908,0.00043370528,0.02480692,0.00021180187,0.00012267366,0.00000333659,0.000021348562,0.030524023],"genre_scores_gemma":[0.9926665,0.00013198836,0.00013342782,0.0026491033,0.00014956194,0.000013277576,0.000015901796,0.000007925175,0.0042323107],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99837357,0.00021324599,0.0002713043,0.00016908691,0.00082981616,0.00014295112],"domain_scores_gemma":[0.9765694,0.022516912,0.000058559428,0.00041985314,0.000108836604,0.00032643572],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0009390453,0.0000773659,0.00024882398,0.000014760368,0.000035791618,0.0000042879137,0.00008719679,0.00011446687,0.021528352],"category_scores_gemma":[0.34654903,0.00004017069,0.00014458311,0.00012838947,0.00016033427,0.0000059722915,0.00006429872,0.00038356858,0.00017485359],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00029466118,0.009801793,0.18947901,0.000030715008,0.0005250868,0.0014280807,0.0016998965,0.000004344891,0.00010187877,0.0712819,0.5900898,0.13526283],"study_design_scores_gemma":[0.019889092,0.004175238,0.66427124,0.02812418,0.0015833508,0.00042596122,0.004248822,0.0073603354,0.019158747,0.09488693,0.1548505,0.001025638],"about_ca_topic_score_codex":0.00001113107,"about_ca_topic_score_gemma":0.000003482153,"teacher_disagreement_score":0.47479218,"about_ca_system_score_codex":0.000030323523,"about_ca_system_score_gemma":0.00026195665,"threshold_uncertainty_score":0.9793661},"labels":[],"label_agreement":null},{"id":"W3162490550","doi":"10.1007/s10459-021-10054-x","title":"New ways of seeing: supplementing existing competency framework development guidelines with systems thinking","year":2021,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":38,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; Hamilton Health Sciences; University Health Network; University of Toronto; McMaster University; Fanshawe College","funders":"","keywords":"Situated; Context (archaeology); Accountability; Systems thinking; Mirroring; Accreditation; Conceptual framework; Engineering ethics; Knowledge management; Adaptation (eye); Curriculum; Computer science; Quality (philosophy); Management science; Psychology; Medicine; Sociology; Pedagogy; Medical education; Epistemology; Engineering","score_opus":0.10127107802946216,"score_gpt":0.4586654335046492,"score_spread":0.35739435547518705,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3162490550","genre_codex":"empirical","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.47369146,0.15015301,0.23987503,0.112596996,0.0073654545,0.0020933282,0.0000044453473,0.00020739224,0.0140128825],"genre_scores_gemma":[0.3036112,0.00079470046,0.6912614,0.003884824,0.0003070937,0.000020005778,0.000019511406,0.00000694122,0.00009430888],"study_design_codex":"observational","study_design_gemma":"systematic_review","domain_scores_codex":[0.99787885,0.00005399015,0.00089719833,0.00034418827,0.0005153505,0.0003104455],"domain_scores_gemma":[0.99531376,0.0035191136,0.00045774857,0.00018348411,0.000375941,0.00014995715],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0015532046,0.00009887249,0.0003031148,0.00010241061,0.00021358152,0.000032204684,0.00011701568,0.00004154689,0.000027790978],"category_scores_gemma":[0.015207995,0.0000782118,0.000020747373,0.000816199,0.000076056494,0.00016197354,0.00003282847,0.00017755528,0.0000018097412],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000014862394,0.00018627492,0.7701287,0.00081603124,0.0000075123517,0.000007163935,0.003239679,0.0029975611,0.000011599632,0.044965126,0.00068681716,0.17693865],"study_design_scores_gemma":[0.003925442,0.002732452,0.22349086,0.3316048,0.00011111549,0.00062935427,0.07912038,0.009000421,0.003158144,0.07194783,0.27268538,0.0015938054],"about_ca_topic_score_codex":0.00062588695,"about_ca_topic_score_gemma":0.0001449333,"teacher_disagreement_score":0.54663783,"about_ca_system_score_codex":0.00011587562,"about_ca_system_score_gemma":0.007016049,"threshold_uncertainty_score":0.99861324},"labels":[],"label_agreement":null},{"id":"W3163230369","doi":"10.7759/cureus.14959","title":"What Are the Personality Types Among Emergency Medicine Physicians?","year":2021,"lang":"en","type":"article","venue":"Cureus","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Calgary","funders":"","keywords":"Personality; Medicine; Interpersonal communication; Emergency department; Context (archaeology); Psychological intervention; Big Five personality traits; Family medicine; Nursing; Psychology; Social psychology","score_opus":0.050466139294528485,"score_gpt":0.36464019661512687,"score_spread":0.3141740573205984,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3163230369","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9265535,0.022176448,0.000024395982,0.03370518,0.0030573858,0.00013543756,0.0000053320227,0.000059771122,0.014282548],"genre_scores_gemma":[0.97913843,0.0058127297,0.00002839114,0.0020948383,0.0015952454,0.000012601642,0.000042007778,0.00001588095,0.011259884],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99897003,0.00006486217,0.00024193544,0.0002466648,0.00028921422,0.00018726947],"domain_scores_gemma":[0.99743634,0.0016535699,0.00010042631,0.00045384414,0.00023839508,0.000117400814],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00024684813,0.00011197913,0.0002742092,0.000008324866,0.000087663495,0.000013301914,0.00009162426,0.00007267297,0.0035728149],"category_scores_gemma":[0.021315325,0.00006732591,0.00013279078,0.00020245955,0.00017172404,0.00008048709,0.000043457185,0.00021027445,0.00009381248],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000061842424,0.0006757265,0.7193798,0.000075550626,0.00035088955,0.00025816023,0.0028338481,0.0000039193683,0.00042482794,0.0015048296,0.25546643,0.018964164],"study_design_scores_gemma":[0.0006592327,0.00011473844,0.96721536,0.0023466398,0.000330099,0.000019642008,0.0068349326,0.000069940776,0.00033522598,0.0012614399,0.020687938,0.00012482113],"about_ca_topic_score_codex":0.00016091253,"about_ca_topic_score_gemma":0.00021671518,"teacher_disagreement_score":0.24783555,"about_ca_system_score_codex":0.00002648862,"about_ca_system_score_gemma":0.00004961822,"threshold_uncertainty_score":0.99733806},"labels":[],"label_agreement":null},{"id":"W3163310228","doi":"10.3171/2020.12.peds20912","title":"Pediatric neurosurgeons’ philosophical approaches to making intraoperative decisions when encountering an uncertainty or a complication while operating on children","year":2021,"lang":"en","type":"article","venue":"Journal of Neurosurgery Pediatrics","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Simon Fraser University; York University","funders":"","keywords":"Harm; Medicine; Grounded theory; Theme (computing); Narrative; Coding (social sciences); Medical education; Qualitative research; Nursing; Psychology; Social psychology","score_opus":0.21522764559049423,"score_gpt":0.3544706078066131,"score_spread":0.13924296221611887,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3163310228","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98979324,0.00017118914,0.00089971675,0.007812891,0.0008207097,0.00028870683,0.000045430188,0.000044291934,0.00012379455],"genre_scores_gemma":[0.98779994,0.00046259357,0.0033795442,0.004518176,0.0036993446,0.000010797654,0.000037744652,0.000057906982,0.000033921013],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9962549,0.00036378877,0.0014517157,0.0005730667,0.0009296369,0.00042687423],"domain_scores_gemma":[0.9846526,0.012949069,0.0005649294,0.0005064227,0.00061853277,0.0007084415],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0010032946,0.00032061673,0.00081718847,0.000602659,0.0002432254,0.00022066777,0.00024174951,0.00017179541,0.00007856579],"category_scores_gemma":[0.07047859,0.00024622632,0.00032455777,0.0012739298,0.00004216853,0.00023195076,0.00013932442,0.0010008033,0.000017365739],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0021905925,0.0065857293,0.9027383,0.000072828465,0.00012639973,0.0050361454,0.0014578747,0.037283305,0.0012375682,0.0004520001,0.0123779625,0.030441334],"study_design_scores_gemma":[0.00354228,0.005009565,0.97429955,0.0012380215,0.0009901384,0.008382873,0.00026136168,0.003837491,0.00045493257,0.00053483224,0.0005248135,0.0009241145],"about_ca_topic_score_codex":0.0000035471464,"about_ca_topic_score_gemma":0.0000033240901,"teacher_disagreement_score":0.071561314,"about_ca_system_score_codex":0.00010821597,"about_ca_system_score_gemma":0.00080839463,"threshold_uncertainty_score":0.999999},"labels":[],"label_agreement":null},{"id":"W3164565344","doi":"10.1503/cjs.014919","title":"Assessing the acceptability of script concordance testing: a nationwide study in otolaryngology","year":2021,"lang":"en","type":"article","venue":"Canadian Journal of Surgery","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"Université de Montréal; McGill University","funders":"","keywords":"Medicine; Concordance; Test (biology); Otorhinolaryngology; Medical education; Family medicine; Medical physics; Surgery; Internal medicine","score_opus":0.11686897260641212,"score_gpt":0.3753410711170639,"score_spread":0.2584720985106518,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3164565344","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9953689,0.00091188966,0.000019777252,0.0021393406,0.0005399035,0.00008364321,0.000002427686,0.0000019366694,0.0009321874],"genre_scores_gemma":[0.9989453,0.000009632841,0.0002748787,0.0006244904,0.0001133917,0.0000018382832,0.000001342769,0.000006824371,0.000022309],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9985225,0.00026322878,0.0007411578,0.00010898784,0.00017403465,0.00019005033],"domain_scores_gemma":[0.97187483,0.026342906,0.0003653729,0.00022519704,0.0008282551,0.00036342596],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0022487794,0.00006823174,0.00046188894,0.0001691488,0.000048905436,0.000030356465,0.00007728712,0.000050295846,0.00030325045],"category_scores_gemma":[0.23542449,0.000050490013,0.00011718073,0.0005400939,0.00014546179,0.00009370371,0.000009859297,0.00034678678,8.8265864e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000097265765,0.00012472988,0.98264515,0.000010877842,0.000026618029,0.0017189793,0.00021696721,0.0000127564845,0.000043866152,0.000008315301,0.00047319444,0.014708795],"study_design_scores_gemma":[0.00041201786,0.000060683615,0.9963789,0.0004243836,0.000044760905,0.00046823267,0.0016492524,0.000023687675,0.00013830193,0.000118810705,0.00023923477,0.000041715157],"about_ca_topic_score_codex":0.004550091,"about_ca_topic_score_gemma":0.013824816,"teacher_disagreement_score":0.23317571,"about_ca_system_score_codex":0.00009540459,"about_ca_system_score_gemma":0.0069714347,"threshold_uncertainty_score":0.9986581},"labels":[],"label_agreement":null},{"id":"W3164655521","doi":"10.1016/j.nepr.2021.103085","title":"Automated essay scoring (AES) of constructed responses in nursing examinations: An evaluation","year":2021,"lang":"en","type":"article","venue":"Nurse Education in Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Reliability (semiconductor); Cohort; Medicine; Psychology; Pathology","score_opus":0.04273014382337106,"score_gpt":0.48851028552074693,"score_spread":0.4457801416973759,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3164655521","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9780867,0.00054989004,0.000035242974,0.003299181,0.0009408845,0.00034763425,0.0000025566787,0.000059795737,0.016678117],"genre_scores_gemma":[0.98911697,0.00006745665,0.0100724865,0.00025313062,0.000083649364,0.00005643612,0.00012576333,0.000012433692,0.00021165528],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9977366,0.0009407537,0.0005265694,0.00027285703,0.00038869138,0.00013456293],"domain_scores_gemma":[0.9845674,0.01346139,0.00028451261,0.0003501354,0.0012405511,0.000096025185],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0019111991,0.00008597508,0.00020459662,0.00029462963,0.000031130505,0.000024221337,0.000045317633,0.000108521126,0.00020534571],"category_scores_gemma":[0.27725863,0.0000965048,0.000023822042,0.0008327169,0.00008450974,0.00040590524,0.000005174288,0.00023143517,0.000009932231],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010268972,0.013477521,0.388285,0.000068292604,0.00004385881,0.000052077252,0.010747769,0.00030768444,0.006083698,0.0114488,0.0010673391,0.56739104],"study_design_scores_gemma":[0.0024719185,0.00014065923,0.93603116,0.0063631735,0.0002763317,0.000244693,0.039138045,0.010435727,0.0037419128,0.00090396014,0.00012319774,0.00012922195],"about_ca_topic_score_codex":0.00018893702,"about_ca_topic_score_gemma":0.000031831678,"teacher_disagreement_score":0.5672618,"about_ca_system_score_codex":0.00039477312,"about_ca_system_score_gemma":0.00560349,"threshold_uncertainty_score":0.994035},"labels":[],"label_agreement":null},{"id":"W3170147743","doi":"","title":"Defining clinical reasoning: preliminary findings from a BEME scoping study","year":2016,"lang":"en","type":"article","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Computer science; Artificial intelligence","score_opus":0.05061490021095451,"score_gpt":0.4013260951371773,"score_spread":0.3507111949262228,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3170147743","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99054843,0.0003286175,0.0008170199,0.0019602396,0.0004580592,0.00062737527,0.00000843434,0.00026621122,0.0049856002],"genre_scores_gemma":[0.9907854,0.000099963385,0.005488197,0.0011074693,0.00034678844,0.000040720544,0.000013728649,0.000037847836,0.002079873],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9974248,0.00014299218,0.000914543,0.00070467085,0.0004093115,0.00040372246],"domain_scores_gemma":[0.96378374,0.03485706,0.00013225398,0.0006412127,0.00010776167,0.00047794427],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0011925194,0.00021561356,0.0006731151,0.00007160667,0.00009508858,0.000030495314,0.00015466899,0.0001931587,0.0012330298],"category_scores_gemma":[0.10927036,0.00012897151,0.0002296498,0.00015167965,0.00012893614,0.000082462466,0.00018544125,0.00027334015,0.00097442186],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00055650703,0.0011009758,0.96313983,0.0000060895295,0.00012057379,0.00018603218,0.00015707094,1.6423112e-7,0.000033439657,0.00031113907,0.0038417813,0.030546386],"study_design_scores_gemma":[0.007249663,0.0038429634,0.97230566,0.0149945095,0.00032086956,0.00002282335,0.0003913217,0.000050525523,0.0001784836,0.00023224507,0.00017074367,0.00024018489],"about_ca_topic_score_codex":0.00023487698,"about_ca_topic_score_gemma":0.000027718996,"teacher_disagreement_score":0.10807784,"about_ca_system_score_codex":0.00004593074,"about_ca_system_score_gemma":0.00021894865,"threshold_uncertainty_score":0.9998034},"labels":[],"label_agreement":null},{"id":"W3174568830","doi":"10.4103/amhs.amhs_106_21","title":"The Unexplored Value of “Normal”","year":2021,"lang":"en","type":"article","venue":"Archives of Medicine and Health Sciences","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Medicine; Categorization; Value (mathematics); Perception; Health care; Medical education; Cognitive psychology; Applied psychology; Epistemology; Psychology; Law","score_opus":0.09164391425160429,"score_gpt":0.4306734100594784,"score_spread":0.3390294958078741,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3174568830","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5511553,0.046614107,0.00019969181,0.3596044,0.00035169112,0.00024363483,0.00000405256,0.000015908137,0.04181119],"genre_scores_gemma":[0.969255,0.025066596,0.0021348754,0.0031945861,0.00011203447,0.0000024417561,0.0000031557058,0.000002499386,0.00022883953],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9989629,0.000059811737,0.00038018666,0.00014167315,0.0002834002,0.00017202964],"domain_scores_gemma":[0.9902173,0.009274802,0.00015018687,0.00014366914,0.000039164162,0.00017483963],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010385789,0.000051047817,0.0002905833,0.000045791374,0.0001636292,0.0000026923422,0.00008001746,0.0000136377375,0.000014584442],"category_scores_gemma":[0.00942829,0.000025872087,0.00003710279,0.00020388032,0.0020707338,0.000016895578,0.000041880205,0.0000820092,3.8709018e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00016594367,0.0002465892,0.17157902,0.00042246756,0.0000547189,0.000020242607,0.007995402,0.000010978948,0.0008914864,0.15608728,0.006957009,0.65556884],"study_design_scores_gemma":[0.0034220715,0.0050566862,0.9268986,0.009154217,0.00007204222,0.000109511435,0.0134603605,0.0009910258,0.0015126772,0.030189225,0.0090400595,0.00009352705],"about_ca_topic_score_codex":0.0003768762,"about_ca_topic_score_gemma":0.000040591858,"teacher_disagreement_score":0.7553196,"about_ca_system_score_codex":0.0000019296676,"about_ca_system_score_gemma":0.0004882553,"threshold_uncertainty_score":0.99891573},"labels":[],"label_agreement":null},{"id":"W3175440889","doi":"10.1080/0142159x.2021.1925101","title":"Diagnosing conflict: Conflicting data, interpersonal conflict, and conflicts of interest in clinical competency committees","year":2021,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital; University of Ottawa; Royal College of Physicians and Surgeons of Canada; University of Alberta; University of Toronto; McMaster University","funders":"","keywords":"Conflict of interest; Interpersonal communication; Political science; Psychology; Conflict resolution; Public relations; Social psychology; Law","score_opus":0.2235986980788205,"score_gpt":0.4456094355921404,"score_spread":0.2220107375133199,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3175440889","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9818858,0.0043275743,0.00013611988,0.006467446,0.00049796974,0.00024382092,0.000026949803,0.00006099031,0.0063533136],"genre_scores_gemma":[0.99259496,0.0018593331,0.00062871387,0.0034081012,0.00047897067,0.000011158555,0.00025083974,0.00003888579,0.0007290581],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.995995,0.00048392513,0.0017036193,0.00080824917,0.00055247906,0.00045675674],"domain_scores_gemma":[0.9818521,0.015405908,0.00034701312,0.0010733365,0.00025409216,0.0010675137],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0033326158,0.0002813412,0.0012944194,0.00010392818,0.00004964859,0.000036343863,0.00037900446,0.00050460984,0.0023874936],"category_scores_gemma":[0.14504746,0.00024350824,0.00016815896,0.0002438317,0.0010764772,0.00008978829,0.0007173536,0.0016254503,0.000029247798],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015297192,0.0015123253,0.9436575,0.00011317079,0.00017565796,0.0011395186,0.0014089765,1.4878448e-7,0.00006061917,0.0011255605,0.0035922064,0.047061298],"study_design_scores_gemma":[0.017885461,0.001013658,0.76472455,0.018242264,0.0004672955,0.0011753795,0.0030442392,0.018281428,0.00039078135,0.00013882398,0.17386338,0.00077276456],"about_ca_topic_score_codex":0.0004113157,"about_ca_topic_score_gemma":0.0003318479,"teacher_disagreement_score":0.17893302,"about_ca_system_score_codex":0.000034878747,"about_ca_system_score_gemma":0.00046785976,"threshold_uncertainty_score":0.9985245},"labels":[],"label_agreement":null},{"id":"W3176216966","doi":"10.2196/17670","title":"Evidence of Construct Validity of Computer-Based Tests for Clinical Reasoning: Instrument Validation Study","year":2021,"lang":"en","type":"article","venue":"JMIR Serious Games","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Construct (python library); Construct validity; Test (biology); Medical education; Psychology; Clinical psychology; Psychometrics; Medicine; Computer science","score_opus":0.15023080325858193,"score_gpt":0.44914620711111164,"score_spread":0.2989154038525297,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3176216966","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99675584,0.00023318203,0.00080529,0.00045543312,0.00047231064,0.0011634979,0.000020826248,0.000042404645,0.000051208783],"genre_scores_gemma":[0.9846818,0.0001026496,0.014740335,0.0001480567,0.00016374841,0.000071625334,0.000035092497,0.000016227568,0.000040482533],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9977971,0.00025249194,0.0009895581,0.00040120335,0.0003735783,0.000186082],"domain_scores_gemma":[0.9824905,0.015759418,0.00044734174,0.0005424755,0.0005962354,0.00016406005],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010202131,0.00015067613,0.00077468937,0.000052260646,0.00003209119,0.000014876804,0.00010611623,0.00012049296,0.00004718002],"category_scores_gemma":[0.03458493,0.0001328426,0.0002433672,0.00021150374,0.00018573734,0.00004202115,0.000067227265,0.00015143142,0.0000028935697],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005192235,0.002771492,0.90645903,0.00022363159,0.00010983453,0.000033895933,0.00013072716,0.000026428452,0.00048008809,0.00005385285,0.0010117954,0.08818002],"study_design_scores_gemma":[0.005388261,0.008027309,0.96438944,0.0056043463,0.00033436736,0.000028321978,0.00019659555,0.0009884983,0.014607716,0.0000814857,0.00020212853,0.00015155759],"about_ca_topic_score_codex":0.000025359277,"about_ca_topic_score_gemma":0.0000101364385,"teacher_disagreement_score":0.08802846,"about_ca_system_score_codex":0.00003214562,"about_ca_system_score_gemma":0.00062752404,"threshold_uncertainty_score":0.97354716},"labels":[],"label_agreement":null},{"id":"W3177136976","doi":"10.3389/fneur.2021.675520","title":"Regret and Therapeutic Decisions in Multiple Sclerosis Care: Literature Review and Research Protocol","year":2021,"lang":"en","type":"article","venue":"Frontiers in Neurology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; St. Michael's Hospital","funders":"F. Hoffmann-La Roche","keywords":"Regret; Health care; Psychology; Protocol (science); Applied psychology; Medicine; Alternative medicine; Computer science","score_opus":0.08755084886117268,"score_gpt":0.3953555347171995,"score_spread":0.3078046858560268,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3177136976","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.47917578,0.40994295,0.00006731143,0.053986635,0.0005909554,0.055176288,0.000030257075,0.000052076834,0.000977734],"genre_scores_gemma":[0.6293073,0.27287897,0.009178646,0.04157761,0.00022046214,0.046142973,0.00007385256,0.000088592984,0.000531544],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99834126,0.00050149317,0.0002813918,0.00043925227,0.00015598928,0.00028061503],"domain_scores_gemma":[0.9948227,0.0045723077,0.000029706265,0.00030990405,0.00013671201,0.0001286948],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0005656067,0.000101750964,0.00043643307,0.00016526553,0.00004435679,0.000018492874,0.00006125798,0.00019888794,0.000012464281],"category_scores_gemma":[0.033979826,0.00008301333,0.000032586126,0.0005969823,0.00020359743,0.000020567795,0.000107399006,0.00083766197,0.0000016920654],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005277022,0.00019610804,0.8530933,0.00043855168,0.000014237579,0.001061125,0.00049774407,9.353823e-7,0.00006892163,0.00004588891,0.017503193,0.12655224],"study_design_scores_gemma":[0.006747871,0.0010669716,0.9199202,0.014281584,0.00006211824,0.00034138112,0.0001251895,0.00020640821,0.00007450191,0.0016200605,0.05539044,0.00016329347],"about_ca_topic_score_codex":0.000013486564,"about_ca_topic_score_gemma":0.000049108447,"teacher_disagreement_score":0.15013155,"about_ca_system_score_codex":0.000018349132,"about_ca_system_score_gemma":0.00010373975,"threshold_uncertainty_score":0.9741574},"labels":[],"label_agreement":null},{"id":"W3177758865","doi":"10.1080/14739879.2021.1931965","title":"Clinical reasoning – a Scotoma in the medical gaze?","year":2021,"lang":"en","type":"article","venue":"Education for Primary Care","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Gaze; Blind spot; Optometry; Psychology; Medicine; Computer science; Medical education; Artificial intelligence","score_opus":0.027060615172655695,"score_gpt":0.4057333241079892,"score_spread":0.3786727089353335,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3177758865","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.919491,0.0062415064,0.00015270707,0.021417553,0.0021800988,0.00068983715,0.000010579597,0.000050660132,0.049766093],"genre_scores_gemma":[0.961304,0.00039035888,0.005412243,0.028166607,0.0018301194,0.00023835698,0.0011815843,0.000024271962,0.0014524353],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9985708,0.00014684562,0.00042874878,0.000290424,0.00038337434,0.00017981537],"domain_scores_gemma":[0.99108595,0.007933603,0.00006745876,0.00039879343,0.00031654222,0.0001976581],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0009042987,0.000088642395,0.00025715216,0.00003742597,0.000058345282,0.000025676098,0.0001255379,0.00016586737,0.00021794063],"category_scores_gemma":[0.065651715,0.00006459899,0.00018562019,0.0002296536,0.000067683184,0.000029997238,0.000034111086,0.00033264369,0.000034156674],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013575208,0.0015071325,0.3732737,0.00021436559,0.00003235153,0.00010761641,0.0016023084,5.9736936e-7,0.0000066800067,0.0028912935,0.067370154,0.55285805],"study_design_scores_gemma":[0.0022767338,0.0002541563,0.89202136,0.0034732337,0.000112447495,0.00029230522,0.005828985,0.000022484011,0.000041548523,0.00045961063,0.095077425,0.00013970834],"about_ca_topic_score_codex":0.000027780383,"about_ca_topic_score_gemma":0.000036822545,"teacher_disagreement_score":0.55271834,"about_ca_system_score_codex":0.00010455438,"about_ca_system_score_gemma":0.004455542,"threshold_uncertainty_score":0.9422187},"labels":[],"label_agreement":null},{"id":"W3180605944","doi":"10.1007/978-3-030-80421-3_22","title":"Expert, Novice, and Intermediate Performance: Exploring the Relationship Between Clinical Reasoning Behaviors and Diagnostic Performance","year":2021,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Computer science; Artificial intelligence","score_opus":0.08044834264296553,"score_gpt":0.35157271966701936,"score_spread":0.27112437702405384,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3180605944","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9845141,0.0017168303,0.0114266025,0.00063419,0.00102408,0.0003516665,0.0000034814882,0.000049856153,0.00027918027],"genre_scores_gemma":[0.9887562,0.0031212703,0.0059552575,0.000755281,0.0011862658,0.000025968282,0.000019595482,0.000035319546,0.00014485982],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99728763,0.00005102524,0.0007361956,0.00094266847,0.0005583173,0.00042413632],"domain_scores_gemma":[0.95567465,0.04308699,0.000207907,0.00059671275,0.00013622441,0.00029749956],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0015408315,0.00036147228,0.00066656986,0.00022268714,0.00033707506,0.00015981107,0.0003167889,0.00029331475,0.000013727679],"category_scores_gemma":[0.026329141,0.00025864475,0.00009706554,0.00024328471,0.0014565277,0.00020603533,0.0005718167,0.0016380528,0.000009256596],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000115910625,0.000015191507,0.61233264,0.000036812417,0.000008358359,0.00004049518,0.0005098041,0.00006417059,5.684503e-7,0.000056841578,0.000005450118,0.38691807],"study_design_scores_gemma":[0.00064581755,0.0004623033,0.98040617,0.010821771,0.00011055374,0.00013268662,0.0000035087332,0.0065381527,0.000036783684,0.00035188318,0.00016325629,0.00032712152],"about_ca_topic_score_codex":0.000017394026,"about_ca_topic_score_gemma":0.000009512914,"teacher_disagreement_score":0.38659093,"about_ca_system_score_codex":0.000085211585,"about_ca_system_score_gemma":0.0003202707,"threshold_uncertainty_score":0.9999866},"labels":[],"label_agreement":null},{"id":"W3184773661","doi":"10.1177/2327857921101058","title":"Can Cognitive Assessment Games Save Us From Cognitive Decline?","year":2021,"lang":"en","type":"article","venue":"Proceedings of the International Symposium on Human Factors and Ergonomics in Health Care","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Schwartz/Reisman Emergency Medicine Institute; North York General Hospital; University of Toronto","funders":"","keywords":"Cognition; Cognitive Assessment System; Cognitive decline; Psychology; Health care; Cognitive neuropsychology; Applied psychology; Medicine; Dementia; Neuropsychology; Cognitive impairment; Psychiatry; Economics; Economic growth","score_opus":0.032477844277442616,"score_gpt":0.3635917760258931,"score_spread":0.3311139317484505,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3184773661","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9913363,0.00021088343,8.1486263e-7,0.0046774275,0.00041995227,0.00026756278,0.0003897416,0.000012071549,0.0026852635],"genre_scores_gemma":[0.9962691,0.00043230757,0.00008912849,0.0026957195,0.00017560039,0.000018581028,0.00021765004,0.000017100876,0.00008482584],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9986595,0.00001269709,0.000503093,0.00036061837,0.00028596507,0.00017812273],"domain_scores_gemma":[0.99686736,0.0020690418,0.00030944613,0.00006159564,0.0005434031,0.00014914444],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00016212623,0.00016054079,0.00036148223,0.000057618112,0.00009607754,0.00005016251,0.00012732971,0.00008805145,0.0000456361],"category_scores_gemma":[0.002934023,0.00012318639,0.00011364519,0.00006782692,0.00009498381,0.000037941027,0.00018296641,0.0003267364,8.523268e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00026936646,0.00019410768,0.99477357,0.000069358146,0.00008942543,0.0000026922248,0.001797253,0.0000035733888,0.00030773546,0.0016669731,0.00010476019,0.00072116224],"study_design_scores_gemma":[0.0017749193,0.00029932612,0.98099995,0.004546367,0.00004920763,0.00000449232,0.0053917416,0.00004674213,0.0060578943,0.0005299996,0.0001799381,0.00011941237],"about_ca_topic_score_codex":0.0017568376,"about_ca_topic_score_gemma":0.0006085356,"teacher_disagreement_score":0.013773633,"about_ca_system_score_codex":0.00038315405,"about_ca_system_score_gemma":0.00026948063,"threshold_uncertainty_score":0.50233954},"labels":[],"label_agreement":null},{"id":"W3188246492","doi":"10.1007/s43678-021-00187-8","title":"Opening the digital front door: digital offerings in a pediatric emergency department","year":2021,"lang":"en","type":"letter","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Emergency department; Medicine; Medical emergency; Front (military); Emergency medicine; Nursing; Engineering","score_opus":0.060463793877834966,"score_gpt":0.32971718779865045,"score_spread":0.2692533939208155,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3188246492","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.1274175,0.04004087,0.000046719222,0.7735605,0.030275475,0.00089434476,0.00025692408,0.000020807238,0.027486814],"genre_scores_gemma":[0.7303591,0.018573416,0.00008343764,0.10717652,0.09292209,0.00007637321,0.0020949298,0.00043134292,0.04828279],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9947513,0.00009114521,0.002688694,0.00048612402,0.0010264115,0.00095636235],"domain_scores_gemma":[0.9950013,0.0014143464,0.001016389,0.0006331588,0.0006181436,0.001316647],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.000750365,0.0005604343,0.0014889006,0.000677328,0.00012717607,0.000048918308,0.00062270684,0.00053804775,0.010558375],"category_scores_gemma":[0.037503794,0.00036637823,0.0006354006,0.0007090413,0.00014678738,0.00019089202,0.00006166695,0.0031832221,0.000075088865],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000025578818,0.000039579172,0.16591783,0.000101279424,0.00016588555,0.009612873,0.00038997724,0.0000031188838,5.026796e-7,0.0000029438518,0.8216388,0.0021015855],"study_design_scores_gemma":[0.0017100594,0.0009200007,0.058994554,0.00329326,0.0008897343,0.0007792242,0.00032665167,0.000010523111,8.343127e-7,0.00018147055,0.932429,0.0004646686],"about_ca_topic_score_codex":0.0020088854,"about_ca_topic_score_gemma":0.0024551773,"teacher_disagreement_score":0.66638404,"about_ca_system_score_codex":0.0004136578,"about_ca_system_score_gemma":0.003041739,"threshold_uncertainty_score":0.9998788},"labels":[],"label_agreement":null},{"id":"W3189012766","doi":"10.7759/cureus.17041","title":"A Cognitive Autopsy Approach Towards Explaining Diagnostic Failure","year":2021,"lang":"en","type":"article","venue":"Cureus","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Medicine; Context (archaeology); Cognition; Medical diagnosis; Cognitive bias; Identification (biology); Cognitive psychology; Clinical psychology; Psychiatry; Pathology; Psychology","score_opus":0.04052406594641275,"score_gpt":0.34037735269461666,"score_spread":0.2998532867482039,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3189012766","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8350176,0.0072769863,0.008157956,0.0041704695,0.00085855543,0.0006315635,0.000072401686,0.000442306,0.14337213],"genre_scores_gemma":[0.98979956,0.00024365811,0.005913665,0.0016537134,0.0004991418,0.00007697539,0.00027896915,0.000031852258,0.0015024501],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9985582,0.00006759199,0.0002815197,0.00044158017,0.0003124803,0.00033862912],"domain_scores_gemma":[0.9886028,0.010402885,0.00006401603,0.00031778967,0.00027375194,0.00033875805],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00021709983,0.00017930665,0.00040467898,0.000028459974,0.00007773457,0.000039016646,0.00007187021,0.00017406882,0.0007612809],"category_scores_gemma":[0.1610311,0.00015514746,0.0001709341,0.000250916,0.00009001899,0.000052736978,0.000088932,0.0003361761,0.00017146427],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007263498,0.011837879,0.33599356,0.00065897306,0.0025234497,0.01775533,0.014442517,0.00006454712,0.0013813806,0.017134454,0.12315063,0.47433093],"study_design_scores_gemma":[0.02474768,0.002469567,0.85358757,0.018561525,0.0038127215,0.0055717602,0.025876181,0.0029659125,0.012025212,0.003282999,0.044722486,0.0023763818],"about_ca_topic_score_codex":0.00004028775,"about_ca_topic_score_gemma":0.000011097235,"teacher_disagreement_score":0.51759404,"about_ca_system_score_codex":0.000052663203,"about_ca_system_score_gemma":0.00039268003,"threshold_uncertainty_score":0.8460359},"labels":[],"label_agreement":null},{"id":"W3191474316","doi":"10.3390/dj9080094","title":"Evaluating Classification Consistency of Oral Lesion Images for Use in an Image Classification Teaching Tool","year":2021,"lang":"en","type":"article","venue":"Dentistry Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"Faculty of Medicine and Dentistry, University of Alberta; University of Alberta","keywords":"Kappa; Rubric; Consistency (knowledge bases); Artificial intelligence; Cohen's kappa; Medicine; Inter-rater reliability; Lesion; Pattern recognition (psychology); Computer science; Mathematics; Machine learning; Statistics; Pathology; Mathematics education","score_opus":0.2608180206976614,"score_gpt":0.490187083475301,"score_spread":0.22936906277763958,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3191474316","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98389685,0.00013259659,0.014816746,0.0005124352,0.00030124374,0.00017045217,0.000019483521,0.000013655005,0.00013651606],"genre_scores_gemma":[0.9098781,0.000060678583,0.08943866,0.000077205485,0.00013305077,0.000009868263,0.00014023948,0.00001450193,0.0002477218],"study_design_codex":"bench_or_experimental","study_design_gemma":"observational","domain_scores_codex":[0.99827963,0.00025167927,0.00072458916,0.00023706179,0.00033099888,0.00017605549],"domain_scores_gemma":[0.99604994,0.0025862975,0.00037228997,0.00025768997,0.00061080273,0.00012295369],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0016715138,0.00009716994,0.00027421265,0.00008332839,0.00013145279,0.00013494684,0.00007021064,0.00010391565,0.000060632952],"category_scores_gemma":[0.0506029,0.00009085354,0.00013245903,0.00010552833,0.00006272948,0.00031232391,0.000019903293,0.00040513903,0.000004451712],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00026773906,0.0011872639,0.31876403,0.00007620878,0.000037294467,0.0002400928,0.00012668082,0.000012594434,0.48660988,0.00063295977,0.0013904041,0.19065486],"study_design_scores_gemma":[0.0033150902,0.00028553957,0.97388244,0.0014199108,0.00017143453,0.0011072892,0.00084295287,0.012429875,0.005846197,0.0004906477,0.000095841104,0.00011276349],"about_ca_topic_score_codex":0.000014601375,"about_ca_topic_score_gemma":0.000011839825,"teacher_disagreement_score":0.6551184,"about_ca_system_score_codex":0.000092208335,"about_ca_system_score_gemma":0.0003394104,"threshold_uncertainty_score":0.95739424},"labels":[],"label_agreement":null},{"id":"W3191534089","doi":"10.18192/uojm.v11i1.5827","title":"La pédagogie médicale au cœur d’une pandémie : les cliniques simulées sont-elles immunes?","year":2021,"lang":"fr","type":"article","venue":"University of Ottawa Journal of Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"Ottawa Heart Institute; Royal Ottawa Mental Health Centre; University of Ottawa","funders":"","keywords":"Humanities; Coronavirus disease 2019 (COVID-19); Philosophy; Medicine","score_opus":0.04262161246292451,"score_gpt":0.3342637592080528,"score_spread":0.2916421467451283,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3191534089","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.647446,0.16022968,0.006404346,0.15626746,0.004165605,0.00023576604,0.000051188894,0.000043071614,0.025156908],"genre_scores_gemma":[0.9061199,0.06363052,0.004126876,0.000627731,0.002628669,7.256964e-8,0.00002725595,0.000036810412,0.022802152],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9959736,0.00083022076,0.0012965727,0.00032696273,0.0009769647,0.0005956813],"domain_scores_gemma":[0.9699483,0.025918437,0.0012839121,0.00044994577,0.0014302799,0.00096913014],"candidate_categories":["metaresearch","metaepi_narrow","sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.003559976,0.00033342943,0.0016939179,0.0002735495,0.00020687516,0.000011548752,0.00044986742,0.0005729479,0.0039516506],"category_scores_gemma":[0.051062346,0.0002995783,0.00062267826,0.00047102472,0.0027780854,0.00016445303,0.00016963764,0.001328638,0.000032749507],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014001778,0.0041074124,0.24113685,0.0010784827,0.00227542,0.03219919,0.010721484,0.00027929663,0.0013262837,0.0046849186,0.13685776,0.5639327],"study_design_scores_gemma":[0.023755828,0.0045553083,0.10528042,0.060810715,0.004630251,0.00245319,0.07127084,0.00040521641,0.0013099103,0.0011529516,0.72388554,0.0004898255],"about_ca_topic_score_codex":0.0026105496,"about_ca_topic_score_gemma":0.00046243952,"teacher_disagreement_score":0.5870278,"about_ca_system_score_codex":0.00028020766,"about_ca_system_score_gemma":0.0021660062,"threshold_uncertainty_score":0.99994564},"labels":[],"label_agreement":null},{"id":"W3194402577","doi":"10.1101/2021.08.18.21262205","title":"Communicating risk to patients: developing an “Efficiency Index” for diagnostic and screening tests","year":2021,"lang":"en","type":"preprint","venue":"medRxiv","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Metric (unit); Diagnostic accuracy; Mace; Test (biology); Cognition; Diagnostic test; Computer science; Psychology; Statistics; Medical physics; Medicine; Pediatrics; Psychiatry; Mathematics; Radiology; Operations management","score_opus":0.05033215158074426,"score_gpt":0.3646754166217851,"score_spread":0.3143432650410408,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3194402577","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95509666,0.000794113,0.041510902,0.00080348016,0.00032393567,0.001212079,0.000054654836,0.00011189812,0.00009228312],"genre_scores_gemma":[0.9205922,0.0006719153,0.07680745,0.0010255326,0.00018466447,0.00026462873,0.00037416324,0.00006208625,0.000017371658],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9974732,0.00019467117,0.0007182972,0.00081633206,0.00035869493,0.00043881877],"domain_scores_gemma":[0.9616944,0.035637915,0.00036220264,0.0011748427,0.0005887778,0.0005418213],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0011585234,0.00033913055,0.00075143203,0.00014699476,0.00025271656,0.00013092525,0.0003412142,0.00031833074,0.000012579925],"category_scores_gemma":[0.33810595,0.00031731473,0.00014163481,0.00023099045,0.00009447072,0.000047267557,0.0011244274,0.00086068973,0.0000041663898],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003879111,0.00027513897,0.9345133,0.00014352454,0.000053468113,0.000012983423,0.00081944675,0.00022621555,0.000006491646,0.000020655429,0.000041915875,0.063848116],"study_design_scores_gemma":[0.0013708312,0.00035035773,0.9837339,0.010719651,0.00023253584,0.000006071737,0.00019971564,0.002539552,0.00006443257,0.00016741795,0.00022813448,0.00038743656],"about_ca_topic_score_codex":0.0005970329,"about_ca_topic_score_gemma":0.00023095298,"teacher_disagreement_score":0.33694744,"about_ca_system_score_codex":0.00007186738,"about_ca_system_score_gemma":0.00029194224,"threshold_uncertainty_score":0.9999279},"labels":[],"label_agreement":null},{"id":"W3195559434","doi":"10.1136/bmjqs-2020-012456","title":"Improving diagnostic performance through feedback: the Diagnosis Learning Cycle","year":2021,"lang":"en","type":"review","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":41,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"Health Services Research and Development; Gordon and Betty Moore Foundation","keywords":"Medicine; Learning cycle; Medical education; Process management; Medical physics; Management","score_opus":0.11505961552406756,"score_gpt":0.45283494592622014,"score_spread":0.3377753304021526,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3195559434","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0004991818,0.99174476,0.00022506117,0.0012955621,0.0008839834,0.002185459,0.00007570758,0.0002181691,0.0028720924],"genre_scores_gemma":[0.000825002,0.9933299,0.0004014321,0.0008411416,0.0015901674,0.0008798019,0.00081035844,0.00013732162,0.0011848984],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9921971,0.0016780499,0.0032059997,0.0011120707,0.000959013,0.0008477548],"domain_scores_gemma":[0.8476312,0.14887872,0.0012912946,0.0015641969,0.00029134718,0.00034323323],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.004782392,0.0007965953,0.0037790623,0.00006004918,0.0005854899,0.0001270226,0.00050226506,0.00069874123,0.0006286411],"category_scores_gemma":[0.3002204,0.00051052344,0.0017497252,0.00075273373,0.00032186194,0.00012931155,0.000516736,0.0024449523,0.00047345128],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000028229288,0.0002680332,0.002574166,0.00847584,0.0003193473,0.00010062091,0.0002303262,0.000026357608,3.2680273e-8,0.00028063738,0.0022953518,0.98540103],"study_design_scores_gemma":[0.00077874993,0.00018050228,0.0020380192,0.10915564,0.0026591679,0.00016213738,0.00033486597,0.000044365235,0.0000011427204,0.000022553393,0.8839798,0.00064305216],"about_ca_topic_score_codex":0.00067215005,"about_ca_topic_score_gemma":0.000021354012,"teacher_disagreement_score":0.984758,"about_ca_system_score_codex":0.0003442837,"about_ca_system_score_gemma":0.0019573176,"threshold_uncertainty_score":0.9998565},"labels":[],"label_agreement":null},{"id":"W3197459981","doi":"10.1111/medu.14636","title":"Artificial intelligence meets clinical skills teaching","year":2021,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Faculty of Medicine, University of British Columbia","keywords":"Medical school; Library science; Medicine; Medical education; Computer science","score_opus":0.04116447251485556,"score_gpt":0.4589902540721953,"score_spread":0.41782578155733974,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3197459981","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.81582296,0.0009574537,0.0076355557,0.14896618,0.0064695324,0.00026851063,0.0000024298347,0.00016681473,0.019710572],"genre_scores_gemma":[0.96274346,0.00033882068,0.0071518626,0.025096638,0.0025943671,0.00002142675,0.00013278729,0.00001868304,0.0019019807],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9976826,0.00023490039,0.00078658154,0.0004003393,0.00066096964,0.0002346462],"domain_scores_gemma":[0.9902762,0.008221059,0.00011070989,0.0004506557,0.00024897515,0.0006923724],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0017615359,0.00011301169,0.0003298228,0.000045089884,0.00008262086,0.000027306178,0.0001054185,0.00027377158,0.0027926604],"category_scores_gemma":[0.41947046,0.00009595791,0.00017344594,0.00017852717,0.00015608795,0.000037748225,0.000058083355,0.0007733124,0.00049116934],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000021506834,0.00393169,0.015420723,0.000013704545,0.000024115358,0.00006783445,0.00014073096,4.821052e-7,0.000013768983,0.0069561787,0.028738515,0.94467074],"study_design_scores_gemma":[0.0015932897,0.001554423,0.4085696,0.016426465,0.00096668175,0.0019266917,0.0028014737,0.0050110044,0.0054791123,0.044314705,0.51006746,0.0012890729],"about_ca_topic_score_codex":0.00004085785,"about_ca_topic_score_gemma":0.00001578118,"teacher_disagreement_score":0.94338167,"about_ca_system_score_codex":0.00005147744,"about_ca_system_score_gemma":0.0037106408,"threshold_uncertainty_score":0.99811894},"labels":[],"label_agreement":null},{"id":"W3199066141","doi":"10.1136/fmch-2020-000798","title":"Multimorbidity and clinical reasoning through the eyes of GPs: a qualitative study","year":2021,"lang":"en","type":"article","venue":"Family Medicine and Community Health","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Context (archaeology); Thematic analysis; Qualitative research; Exploratory research; Global Positioning System; Cognition; Psychology; Medicine; Applied psychology; Computer science; Psychiatry; Sociology","score_opus":0.3016915154585408,"score_gpt":0.5551847459188142,"score_spread":0.2534932304602734,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3199066141","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9787424,0.0073882183,0.0001263765,0.011371712,0.00016720191,0.0004135017,0.000012017841,0.000027668855,0.0017508633],"genre_scores_gemma":[0.9822742,0.007866604,0.00094805006,0.008610129,0.00015161627,0.000009532921,0.000037182628,0.0000116800165,0.00009101571],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.99356616,0.0048127775,0.0008536477,0.00022266668,0.00032403562,0.00022069579],"domain_scores_gemma":[0.9459175,0.052435625,0.00029184058,0.0007732058,0.00028158005,0.00030028203],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.006845124,0.00015296206,0.0009092108,0.00002888538,0.00047244324,0.000007644863,0.0001061565,0.000081943355,0.000020405658],"category_scores_gemma":[0.06691369,0.00009238038,0.00006785561,0.0002498899,0.00081567303,0.00003543653,0.00020046545,0.0013687538,0.0000013873816],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00054426375,0.0065675187,0.39497778,0.00043542488,0.0005381836,0.000078578414,0.5312339,5.761636e-7,0.000020774785,0.0015721922,0.013902125,0.050128642],"study_design_scores_gemma":[0.003565217,0.0043709,0.5026582,0.0023197632,0.00015981254,0.000028386647,0.48534653,0.00004328533,0.0000021345274,0.00054610585,0.0008956403,0.00006406442],"about_ca_topic_score_codex":0.030543331,"about_ca_topic_score_gemma":0.001561581,"teacher_disagreement_score":0.10768038,"about_ca_system_score_codex":0.000022992326,"about_ca_system_score_gemma":0.00040722327,"threshold_uncertainty_score":0.9759124},"labels":[],"label_agreement":null},{"id":"W3199949523","doi":"10.1007/s10459-021-10069-4","title":"Language games and scholarly writing","year":2021,"lang":"en","type":"editorial","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Field (mathematics); Work (physics); Computer science; Sociology; Mathematics education; Psychology; Engineering","score_opus":0.016229827037928607,"score_gpt":0.4465205838009304,"score_spread":0.43029075676300177,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3199949523","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.017748926,0.3048963,0.000020301964,0.0033081586,0.67063737,0.00036717608,0.0000150420565,0.00003755064,0.0029691819],"genre_scores_gemma":[0.02052965,0.15078738,0.013204552,0.002678751,0.81060284,0.00012454469,0.00045663817,0.000029997585,0.0015856451],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9978687,0.00010963242,0.0005447009,0.00057353533,0.0005826166,0.00032080762],"domain_scores_gemma":[0.9862316,0.012730417,0.0003340554,0.00022903088,0.00022011297,0.0002547678],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0018963163,0.00014239394,0.00040900588,0.00018722897,0.00015279824,0.000157199,0.000119117605,0.00021204569,0.000027495793],"category_scores_gemma":[0.09934092,0.00012702165,0.000035240522,0.0005144029,0.00023816747,0.0004822932,0.00004347233,0.00070541876,0.0000052640517],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000014938603,0.00027484322,0.034851056,0.000913233,0.0000038511307,0.000010114442,0.0021324588,0.0000062035706,0.0000024799701,0.00027176875,0.3431815,0.6183376],"study_design_scores_gemma":[0.000635659,0.00055360614,0.010579714,0.029171437,0.000021326578,0.000018278495,0.007736563,0.000035199148,0.000008913643,0.0007849051,0.9501865,0.00026789278],"about_ca_topic_score_codex":0.00069930754,"about_ca_topic_score_gemma":0.00020598077,"teacher_disagreement_score":0.61806965,"about_ca_system_score_codex":0.00017228043,"about_ca_system_score_gemma":0.0057154596,"threshold_uncertainty_score":0.9999212},"labels":[],"label_agreement":null},{"id":"W3200263634","doi":"10.1515/dx-2020-0128","title":"Narrowing the mindware gap in medicine","year":2021,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Variety (cybernetics); Psychology; Cognition; Medicine; Intensive care medicine; Psychiatry; Computer science; Artificial intelligence","score_opus":0.06994617405434375,"score_gpt":0.3739702062088033,"score_spread":0.3040240321544595,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3200263634","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.88468164,0.0052347216,0.000014770623,0.09871294,0.00042060364,0.00018243349,0.0000042771762,0.000041952382,0.010706646],"genre_scores_gemma":[0.98837125,0.0010830828,0.00016512722,0.00914546,0.0003824376,0.00010725316,0.000035929046,0.00001598847,0.00069344573],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.998837,0.00008020539,0.00030638365,0.0002636422,0.00027497232,0.00023778656],"domain_scores_gemma":[0.9845454,0.014780788,0.000041961368,0.00039060213,0.00009821873,0.000143045],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00040486475,0.00011033891,0.0003278447,0.00004998379,0.000054000397,0.000011472931,0.00008474975,0.00008644012,0.0015083338],"category_scores_gemma":[0.07380612,0.00006879492,0.00009942748,0.00044711053,0.000107923646,0.000024612265,0.000049778406,0.0002589246,0.00010190654],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000022473649,0.00033100796,0.90246856,0.000021204627,0.0000373147,0.0010791093,0.0009316144,0.0000082231845,0.00006361033,0.0005012509,0.056670208,0.03786543],"study_design_scores_gemma":[0.0031843472,0.00023570139,0.93020785,0.0064760256,0.00020528438,0.00013602484,0.0021150114,0.000085460924,0.0032665995,0.0009981527,0.052894108,0.00019544674],"about_ca_topic_score_codex":0.00022149904,"about_ca_topic_score_gemma":0.00013514399,"teacher_disagreement_score":0.103689626,"about_ca_system_score_codex":0.000041681986,"about_ca_system_score_gemma":0.00010068906,"threshold_uncertainty_score":0.99940443},"labels":[],"label_agreement":null},{"id":"W3200730176","doi":"10.7759/cureus.c48","title":"A Cognitive Autopsy Approach Towards Explaining Diagnostic Failure","year":2021,"lang":"en","type":"article","venue":"Cureus","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Medicine; Context (archaeology); Cognition; Medical diagnosis; Cognitive bias; Identification (biology); Variance (accounting); Cognitive psychology; Applied psychology; Clinical psychology; Psychiatry; Pathology; Psychology","score_opus":0.04052406594641275,"score_gpt":0.34037735269461666,"score_spread":0.2998532867482039,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3200730176","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8350176,0.0072769863,0.008157956,0.0041704695,0.00085855543,0.0006315635,0.000072401686,0.000442306,0.14337213],"genre_scores_gemma":[0.98979956,0.00024365811,0.005913665,0.0016537134,0.0004991418,0.00007697539,0.00027896915,0.000031852258,0.0015024501],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9985582,0.00006759199,0.0002815197,0.00044158017,0.0003124803,0.00033862912],"domain_scores_gemma":[0.9886028,0.010402885,0.00006401603,0.00031778967,0.00027375194,0.00033875805],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00021709983,0.00017930665,0.00040467898,0.000028459974,0.00007773457,0.000039016646,0.00007187021,0.00017406882,0.0007612809],"category_scores_gemma":[0.1610311,0.00015514746,0.0001709341,0.000250916,0.00009001899,0.000052736978,0.000088932,0.0003361761,0.00017146427],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007263498,0.011837879,0.33599356,0.00065897306,0.0025234497,0.01775533,0.014442517,0.00006454712,0.0013813806,0.017134454,0.12315063,0.47433093],"study_design_scores_gemma":[0.02474768,0.002469567,0.85358757,0.018561525,0.0038127215,0.0055717602,0.025876181,0.0029659125,0.012025212,0.003282999,0.044722486,0.0023763818],"about_ca_topic_score_codex":0.00004028775,"about_ca_topic_score_gemma":0.000011097235,"teacher_disagreement_score":0.51759404,"about_ca_system_score_codex":0.000052663203,"about_ca_system_score_gemma":0.00039268003,"threshold_uncertainty_score":0.8460359},"labels":[],"label_agreement":null},{"id":"W3201457795","doi":"10.1002/aet2.10693","title":"Cognitive load and processes during chest radiograph interpretation in the emergency department across the spectrum of expertise","year":2021,"lang":"en","type":"article","venue":"AEM Education and Training","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Kingston Health Sciences Centre; University of Ottawa; Royal College of Physicians and Surgeons of Canada; Queen's University","funders":"","keywords":"Emergency department; Cognition; Medicine; Confidence interval; Interpretation (philosophy); Psychology; Medical education; Emergency medicine; Psychiatry; Internal medicine; Computer science","score_opus":0.029448287294307474,"score_gpt":0.3686106776756755,"score_spread":0.33916239038136803,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3201457795","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9913419,0.0051661073,0.000034326564,0.002665594,0.000101381265,0.00015082968,0.0000027471146,0.0000055754404,0.00053152005],"genre_scores_gemma":[0.99797714,0.0014792974,0.00006289875,0.00029806275,0.000056857472,0.000053002284,0.000016006947,0.000004194244,0.0000525545],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99946016,0.00004026438,0.00017727827,0.00012894272,0.0001000835,0.00009328304],"domain_scores_gemma":[0.9986013,0.0011486603,0.000056918216,0.00006879968,0.00008598059,0.00003835139],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00022282777,0.000056570163,0.00010341955,0.00001899201,0.000069007496,0.000016578046,0.000022251357,0.000022686083,0.000020316582],"category_scores_gemma":[0.0062140683,0.000035445013,0.000023669601,0.00021185694,0.00006885532,0.000034294895,0.0000102233635,0.000077303615,4.7384995e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014045653,0.00069553027,0.39243513,0.00019921619,0.00006971784,0.000012325584,0.2772228,0.0000012730821,0.00017037135,0.00022541733,0.00007007389,0.32875767],"study_design_scores_gemma":[0.00061322824,0.000060026636,0.9307503,0.00088375853,0.000048803493,0.00008954834,0.066362314,0.000025694577,0.0007859709,0.00029880382,0.000036748606,0.000044812597],"about_ca_topic_score_codex":0.000036815003,"about_ca_topic_score_gemma":0.00011338527,"teacher_disagreement_score":0.5383152,"about_ca_system_score_codex":0.000010045018,"about_ca_system_score_gemma":0.00021168136,"threshold_uncertainty_score":0.7439269},"labels":[],"label_agreement":null},{"id":"W3201523268","doi":"10.46747/cfp.6709652","title":"La politique sur la gestion des tests en Ontario","year":2021,"lang":"fr","type":"article","venue":"Canadian Family Physician","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"Toronto and Region Conservation Authority; Sinai Health System","funders":"","keywords":"Humanities; Political science; Medicine; Gynecology; Philosophy","score_opus":0.02850602464081066,"score_gpt":0.2841788601480446,"score_spread":0.25567283550723396,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3201523268","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.43660948,0.011565993,0.000018151475,0.008145104,0.0015260099,0.00018340406,0.00021173227,0.000048898735,0.54169124],"genre_scores_gemma":[0.901597,0.0009993105,0.00035270565,0.053452395,0.0019321697,0.000022019356,0.00028378388,0.00008386134,0.041276757],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9973654,0.000532271,0.00035546953,0.00050388277,0.00023413992,0.0010088298],"domain_scores_gemma":[0.98415226,0.013132135,0.00006818254,0.00049150985,0.00030123713,0.0018546522],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00046851308,0.00031523805,0.0004992668,0.00011378823,0.00014617873,0.000113820606,0.00014613738,0.0005865444,0.00012759487],"category_scores_gemma":[0.012937556,0.00037176613,0.00022756822,0.0003715697,0.00042564113,0.00013327826,0.000037845148,0.0010047905,0.00068700424],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000026874695,0.00074788113,0.040315937,0.00019889486,0.0003105004,0.010745576,0.002742879,0.00008716579,0.00060617394,0.042973176,0.12787694,0.773368],"study_design_scores_gemma":[0.00069791265,0.000101409714,0.5930795,0.0066385847,0.00022189516,0.00004908772,0.0005526215,0.0000373423,0.00014194137,0.0027131974,0.39548334,0.00028318117],"about_ca_topic_score_codex":0.87367535,"about_ca_topic_score_gemma":0.9372031,"teacher_disagreement_score":0.7730848,"about_ca_system_score_codex":0.0030740495,"about_ca_system_score_gemma":0.01387872,"threshold_uncertainty_score":0.9998734},"labels":[],"label_agreement":null},{"id":"W3202293436","doi":"10.1136/bmjqs-2021-013493","title":"Should electronic differential diagnosis support be used early or late in the diagnostic process? A multicentre experimental study of Isabel","year":2021,"lang":"en","type":"article","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":61,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Impact; McMaster University","funders":"Physicians' Services Incorporated Foundation","keywords":"Medicine; Medical diagnosis; Differential diagnosis; Medical history; Emergency medicine; Family medicine; Pediatrics; Internal medicine; Pathology","score_opus":0.11492802133079599,"score_gpt":0.46121284727268935,"score_spread":0.34628482594189336,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3202293436","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9934369,0.00020730197,0.000015952151,0.003846281,0.00015118012,0.00208987,0.000053667674,0.000047451806,0.00015140738],"genre_scores_gemma":[0.99766403,0.00017253218,0.000019396213,0.0010856449,0.0001425153,0.0005170121,0.00013502303,0.00003287824,0.00023098095],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99526805,0.00085934455,0.0016065625,0.0006037592,0.0010537736,0.0006085422],"domain_scores_gemma":[0.9677601,0.03087451,0.00028834338,0.0007157011,0.00017122076,0.00019011271],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0015577178,0.00030303348,0.0008983132,0.000056668367,0.00009786873,0.000036503287,0.00025474635,0.00016874,0.0011129462],"category_scores_gemma":[0.053377084,0.00019946911,0.00023433511,0.00049484713,0.00012226788,0.000079849364,0.00011757568,0.0005612233,0.00001983191],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0025165963,0.029540066,0.93336904,0.00013817132,0.00026984522,0.0006211763,0.030791774,0.000018118795,0.0003045476,0.0001608864,0.0005632205,0.001706546],"study_design_scores_gemma":[0.01556456,0.0032679879,0.94121313,0.00085140363,0.0004656371,0.00002581562,0.034493204,0.00006601058,0.0034938366,0.000062918625,0.00015149175,0.0003440193],"about_ca_topic_score_codex":0.0009953863,"about_ca_topic_score_gemma":0.0018581285,"teacher_disagreement_score":0.05181937,"about_ca_system_score_codex":0.0001308316,"about_ca_system_score_gemma":0.00074132124,"threshold_uncertainty_score":0.99980015},"labels":[],"label_agreement":null},{"id":"W3205567714","doi":"10.2196/jmir.2232","title":"Correction: Development and Validation of Filters for the Retrieval of Studies of Clinical Examination From Medline","year":2012,"lang":"en","type":"article","venue":"Journal of Medical Internet Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"MEDLINE; Information retrieval; Computer science; Medicine; Biology","score_opus":0.33050163383571723,"score_gpt":0.5593820156705984,"score_spread":0.22888038183488119,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3205567714","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99118954,0.0032333017,0.001829272,0.0023700714,0.0011681198,0.00015686211,0.0000022959123,0.0000013500847,0.000049203216],"genre_scores_gemma":[0.9959856,0.0017922632,0.0012103803,0.00003995425,0.0006957121,0.0000023647146,0.000005023314,0.000005403762,0.00026327453],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99612004,0.00043903454,0.0014349957,0.00008629892,0.0017790803,0.00014055075],"domain_scores_gemma":[0.8633277,0.13288201,0.0008907544,0.00017692747,0.0023603481,0.00036228972],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.025202718,0.000058181984,0.0005185739,0.000117872696,0.000021864902,0.000003912172,0.00014746859,0.00014793882,0.00014042384],"category_scores_gemma":[0.34712,0.00003203528,0.000117709715,0.00014443269,0.00055376056,0.000046136487,0.00009951637,0.0006245312,7.633902e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0061056926,0.00258327,0.28730658,0.00049569074,0.0020882648,0.000018141143,0.009795396,0.0000041606727,0.00067486963,0.00020407152,0.059495512,0.6312284],"study_design_scores_gemma":[0.008368877,0.0057030474,0.7995381,0.016259016,0.0004994121,0.00010956166,0.00881859,0.003449737,0.15186217,0.00023706227,0.0050444957,0.000109926936],"about_ca_topic_score_codex":0.000040839037,"about_ca_topic_score_gemma":0.000004537338,"teacher_disagreement_score":0.6311184,"about_ca_system_score_codex":0.000037861955,"about_ca_system_score_gemma":0.00033725254,"threshold_uncertainty_score":0.8734809},"labels":[],"label_agreement":null},{"id":"W3206827487","doi":"10.1503/cmaj.210794","title":"Physician, heal thyself","year":2021,"lang":"en","type":"article","venue":"Canadian Medical Association Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Alberta Health Services; University of Alberta","funders":"","keywords":"Medicine; Alternative medicine; Medical emergency; Family medicine; Pathology","score_opus":0.011965069798046698,"score_gpt":0.2989281451746716,"score_spread":0.2869630753766249,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3206827487","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.49552557,0.0014291604,0.0001656452,0.4057611,0.0034289374,0.00013472936,0.000036784004,0.0000667377,0.09345135],"genre_scores_gemma":[0.77236176,0.00064910046,0.00054477656,0.21136367,0.0081241885,0.000004774351,0.00006210972,0.000031464195,0.006858166],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9976902,0.00011696575,0.0003676607,0.00015217681,0.0012296122,0.0004433694],"domain_scores_gemma":[0.9942471,0.0020604813,0.000115685885,0.000117636664,0.0006422848,0.002816808],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00096398283,0.00008954586,0.00027631674,0.0000630975,0.00016714571,0.00006191481,0.00007697186,0.00032090867,0.005418457],"category_scores_gemma":[0.07056669,0.000080402184,0.0001756208,0.00025568178,0.000027663036,0.000043280597,0.000011433481,0.0010272248,0.0005813908],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000012782299,0.00017562206,0.112897635,0.000006723893,0.00021059555,0.00636433,0.00014018686,0.0000018544141,0.000022491706,0.000794715,0.71313477,0.1662383],"study_design_scores_gemma":[0.0026563578,0.00012090867,0.193046,0.00085863366,0.00010966248,0.0022193918,0.00037242402,0.00016240356,0.000101604994,0.0007116734,0.7994591,0.00018186911],"about_ca_topic_score_codex":0.0006266745,"about_ca_topic_score_gemma":0.004596973,"teacher_disagreement_score":0.27683616,"about_ca_system_score_codex":0.00066241313,"about_ca_system_score_gemma":0.007921636,"threshold_uncertainty_score":0.99770254},"labels":[],"label_agreement":null},{"id":"W3207657503","doi":"10.2196/32507","title":"Assessing the Performance of a New Artificial Intelligence–Driven Diagnostic Support Tool Using Medical Board Exam Simulations: Clinical Vignette Study","year":2021,"lang":"en","type":"article","venue":"JMIR Medical Informatics","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Medical diagnosis; Medicine; Diagnostic test; Clinical decision support system; Decision support system; Diagnostic accuracy; Vignette; Medical physics; Emergency medicine; Artificial intelligence; Psychology; Computer science; Radiology","score_opus":0.1227083511359059,"score_gpt":0.4719615238526976,"score_spread":0.3492531727167917,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3207657503","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9887909,0.000056916233,0.0075558824,0.0017084024,0.00071621314,0.00080556487,0.0000056681683,0.000069837915,0.0002905749],"genre_scores_gemma":[0.990483,0.00018041531,0.004562397,0.0035875326,0.0010129963,0.000021951493,0.0000596263,0.000025932266,0.00006614153],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9926336,0.00027573112,0.0034298198,0.0002673734,0.0029039893,0.0004894944],"domain_scores_gemma":[0.9628512,0.034220446,0.0006111301,0.0008324475,0.0004545656,0.0010302037],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0028569854,0.0002753833,0.00092531135,0.000096195705,0.00017922252,0.00011010868,0.0004172299,0.00049461203,0.0026775636],"category_scores_gemma":[0.16565193,0.00018658632,0.00028860293,0.00067407085,0.00056618545,0.0003106402,0.0003574494,0.0013441273,0.00010289722],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014735035,0.004695752,0.47233182,0.00028423226,0.00040646963,0.0006772087,0.0062703122,0.0010696778,0.0000081413655,0.00023545633,0.0033980475,0.5104755],"study_design_scores_gemma":[0.003478832,0.0027300639,0.14605145,0.0074174735,0.0010248077,0.0005574516,0.01341993,0.82227415,0.0002576377,0.00033519787,0.0018379905,0.0006150174],"about_ca_topic_score_codex":0.00003248804,"about_ca_topic_score_gemma":0.000021883563,"teacher_disagreement_score":0.8212045,"about_ca_system_score_codex":0.000065563014,"about_ca_system_score_gemma":0.004238674,"threshold_uncertainty_score":0.99823415},"labels":[],"label_agreement":null},{"id":"W3209299781","doi":"10.1097/acm.0000000000004136","title":"O–RI–M: Reporting to Include Data Interpretation","year":2021,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Interpretation (philosophy); Internship; Interpreter; Computer science; Data collection; Medical education; Psychology; Medicine; Sociology; Programming language","score_opus":0.19290785817707373,"score_gpt":0.4655973067503758,"score_spread":0.2726894485733021,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3209299781","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.003539845,0.003302945,0.002345625,0.9809386,0.002375359,0.0006389848,0.000044972974,0.00017962344,0.0066340566],"genre_scores_gemma":[0.010237503,0.0010422097,0.00076973176,0.934145,0.025609372,0.000037690937,0.012142026,0.00010589627,0.015910614],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9934465,0.00016074501,0.0030623483,0.0014054227,0.0013141037,0.0006108552],"domain_scores_gemma":[0.98117745,0.013590614,0.0018939318,0.002497362,0.0003441666,0.00049649226],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0030730222,0.00045637693,0.0017182812,0.00029300514,0.00006580291,0.00001274552,0.0007745785,0.0018177305,0.00097230356],"category_scores_gemma":[0.5326818,0.0003527382,0.00011537081,0.0005293272,0.00017871187,0.000074528,0.0008990686,0.0069836206,0.00024416068],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004492789,0.00000749433,0.0016810987,0.00020797028,0.00013147047,0.0034735883,0.00078622985,8.189692e-7,0.00010273354,0.000005775514,0.97574234,0.01781555],"study_design_scores_gemma":[0.0010961469,0.0003335555,0.00090804446,0.038399123,0.0011327494,0.0011186094,0.00028560116,0.0003981114,0.000045872013,0.00024811106,0.9556646,0.00036950785],"about_ca_topic_score_codex":0.00021863518,"about_ca_topic_score_gemma":0.0000041908306,"teacher_disagreement_score":0.52960885,"about_ca_system_score_codex":0.00016039325,"about_ca_system_score_gemma":0.0005629109,"threshold_uncertainty_score":0.99994093},"labels":[],"label_agreement":null},{"id":"W3209528475","doi":"10.1177/1753495x211045614","title":"Canadian general internal medicine residents’ perception of a pedagogical tool of online cases in obstetric medicine","year":2021,"lang":"en","type":"article","venue":"Obstetric Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université de Sherbrooke","funders":"","keywords":"Likert scale; Medicine; Scale (ratio); Perception; Exploratory research; Family medicine; Medical education; Psychology","score_opus":0.09572577664910718,"score_gpt":0.39508491792112277,"score_spread":0.29935914127201557,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3209528475","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9599235,0.014284669,0.000077225166,0.016237212,0.0015949536,0.00045024164,0.000049579547,0.000036717098,0.0073458757],"genre_scores_gemma":[0.98221546,0.004550482,0.0008190297,0.0023356243,0.0014888715,0.000012990396,0.00057094736,0.00004714821,0.00795945],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99422383,0.00029183194,0.002475037,0.0007522923,0.0015827704,0.0006742384],"domain_scores_gemma":[0.97740334,0.019098923,0.0005632641,0.000888255,0.0010550965,0.0009911194],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011402755,0.00044289057,0.00214613,0.0020222473,0.00005245584,0.0000040197733,0.00031407503,0.00036813493,0.00665861],"category_scores_gemma":[0.29223272,0.00031177906,0.0001874394,0.003866755,0.00079067633,0.000070949696,0.000101463134,0.00076309167,0.00001903208],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00077967794,0.0016012953,0.8127755,0.0008709117,0.0003187034,0.013240291,0.005778665,0.000033246753,0.010046389,0.00092605245,0.04240052,0.11122877],"study_design_scores_gemma":[0.0150138745,0.0042681154,0.9484026,0.016610937,0.00070912373,0.0015295206,0.008711043,0.00045330846,0.00023053306,0.00026550933,0.003463059,0.00034239885],"about_ca_topic_score_codex":0.21759433,"about_ca_topic_score_gemma":0.022889275,"teacher_disagreement_score":0.29109246,"about_ca_system_score_codex":0.00060598494,"about_ca_system_score_gemma":0.0011747117,"threshold_uncertainty_score":0.9999334},"labels":[],"label_agreement":null},{"id":"W3210858561","doi":"10.1016/j.ecns.2021.08.025","title":"Exploring Cognitive Biases and Clinical Reasoning During Simulation With BScN Students","year":2021,"lang":"en","type":"article","venue":"Clinical Simulation in Nursing","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Debriefing; Cognitive bias; Psychology; Cognition; Clinical judgment; Curriculum; Identification (biology); Session (web analytics); Medical education; Confirmation bias; Clinical psychology; Social psychology; Medicine; Pedagogy; Computer science; Psychiatry","score_opus":0.22076105285945188,"score_gpt":0.5144079332455468,"score_spread":0.2936468803860949,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3210858561","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9964815,0.00022782748,0.0016475553,0.00027607262,0.00046262666,0.00037351454,0.0000028789595,0.00009432197,0.00043374448],"genre_scores_gemma":[0.9974523,0.00031234842,0.0011900011,0.00026829058,0.0005783689,0.00001236627,0.0000569936,0.000039536,0.000089780835],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9965765,0.00040651235,0.0014108224,0.0007598955,0.00050186046,0.00034440847],"domain_scores_gemma":[0.9131141,0.08520461,0.00035651508,0.00034667013,0.00047603736,0.00050208805],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0012635298,0.00021406532,0.000708536,0.00010898816,0.00014707657,0.000083887055,0.000056478075,0.0002205864,0.000041735137],"category_scores_gemma":[0.13122329,0.00019560834,0.00016441771,0.00044345888,0.0002937252,0.00026741956,0.00006301134,0.00066328805,0.0000125586],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011607864,0.0011871529,0.8874246,0.000016143333,0.00006244365,0.00021981617,0.0002170302,0.036462937,0.0000026289201,0.000022813529,0.0000014391245,0.07322224],"study_design_scores_gemma":[0.008029242,0.00045404784,0.88031507,0.02240287,0.0002651309,0.00002129391,0.00049839896,0.087688774,0.000032364867,0.00007925095,0.000009983691,0.00020355938],"about_ca_topic_score_codex":0.000007229418,"about_ca_topic_score_gemma":0.000007742719,"teacher_disagreement_score":0.12995976,"about_ca_system_score_codex":0.0000934016,"about_ca_system_score_gemma":0.00012621902,"threshold_uncertainty_score":0.87609476},"labels":[],"label_agreement":null},{"id":"W3211073716","doi":"10.1007/s10639-021-10772-0","title":"Diagnosing virtual patients in a technology-rich learning environment: a sequential Mining of Students’ efficiency and behavioral patterns","year":2021,"lang":"en","type":"article","venue":"Education and Information Technologies","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Context (archaeology); Virtual patient; Metacognition; Educational technology; Computer science; Psychological intervention; Medical education; Psychology; Artificial intelligence; Mathematics education; Machine learning; Medicine; Cognition","score_opus":0.012388542842350952,"score_gpt":0.31436499139590024,"score_spread":0.30197644855354927,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3211073716","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9985843,0.00024252581,0.00015624297,0.00057936314,0.0000852494,0.00015431915,0.0000034671452,0.00009210377,0.00010246944],"genre_scores_gemma":[0.9979458,0.001174083,0.0006797994,0.000060525534,0.0000049914665,0.000033688553,0.000068700254,0.000003512485,0.000028889006],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9992326,0.000013885419,0.00035637332,0.000118347845,0.0001663203,0.00011249533],"domain_scores_gemma":[0.9993939,0.00024410967,0.0001472832,0.0001133949,0.00007446257,0.00002689368],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00011491644,0.000081196056,0.00016678177,0.00033658557,0.000052346888,0.00003124537,0.00005050555,0.00014626651,0.000015832517],"category_scores_gemma":[0.005428887,0.000077410295,0.000016340393,0.0002536629,0.00011904536,0.0002699635,0.00013117018,0.00018306261,0.000003130151],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000059540316,0.00036115572,0.7882569,0.000018399905,0.000003731795,6.3364445e-7,0.0009495974,0.000004870822,0.000025018882,0.00018172851,0.000016275557,0.21017575],"study_design_scores_gemma":[0.0014734056,0.0005514828,0.94777775,0.0009791892,0.00003794499,0.000017734596,0.044101693,0.000112312,0.0041818684,0.000079966754,0.00055962824,0.00012699762],"about_ca_topic_score_codex":0.0000146484235,"about_ca_topic_score_gemma":0.0000018101877,"teacher_disagreement_score":0.21004876,"about_ca_system_score_codex":0.00003947698,"about_ca_system_score_gemma":0.00007624266,"threshold_uncertainty_score":0.64992774},"labels":[],"label_agreement":null},{"id":"W3211667853","doi":"10.22541/au.161860163.36232122/v1","title":"Use of Evidence in Acute Stroke Decision-Making: Implications for Evidence-Based Medicine","year":2021,"lang":"en","type":"preprint","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Ottawa Hospital","funders":"","keywords":"Evidence-based medicine; Best evidence; Medical decision making; Best practice; Medical practice; Process (computing); Medical knowledge; Medical literature; MEDLINE; Psychology; Evidence-based practice; Acute stroke; Medicine; Alternative medicine; Medical education; Family medicine; Psychiatry; Computer science; Emergency department; Management; Political science; Economics","score_opus":0.36355981799656373,"score_gpt":0.4928168163810397,"score_spread":0.12925699838447596,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3211667853","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6677702,0.008140119,0.24450743,0.074282914,0.0010665606,0.0036837545,0.00022634288,0.00013773899,0.0001849249],"genre_scores_gemma":[0.8964472,0.0042325105,0.09585451,0.002354425,0.00018434494,0.00035176158,0.000112540285,0.000034484765,0.00042821563],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99682456,0.00009646342,0.0013890463,0.00086661085,0.0005091467,0.00031416453],"domain_scores_gemma":[0.79705,0.19968872,0.00045200431,0.0014810842,0.0010913472,0.00023681966],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0012126341,0.0003105533,0.0012821879,0.00033341907,0.000034322533,0.0000396548,0.0003286411,0.0004555676,0.00037145842],"category_scores_gemma":[0.48454452,0.00021240402,0.00042527055,0.00037939116,0.00021527444,0.000114637194,0.0003408475,0.0005787714,0.0000039041633],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0036320323,0.0016975971,0.64674467,0.0014595869,0.0007050886,0.00016247232,0.00036482362,0.0037827834,0.003416506,0.0012797466,0.028540524,0.30821416],"study_design_scores_gemma":[0.0019388873,0.0009793198,0.6411457,0.34458715,0.0016030105,0.000023790402,0.0000890172,0.005538459,0.00077084306,0.0026049505,0.00033441637,0.00038443765],"about_ca_topic_score_codex":0.00048751562,"about_ca_topic_score_gemma":0.00016848354,"teacher_disagreement_score":0.4833319,"about_ca_system_score_codex":0.00017037998,"about_ca_system_score_gemma":0.0016245934,"threshold_uncertainty_score":0.8661584},"labels":[],"label_agreement":null},{"id":"W3213447869","doi":"10.3138/jvme-2021-0090","title":"The Use of a Modified Script Concordance Test in Clinical Rounds to Foster and Assess Clinical Reasoning Skills","year":2021,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Formative assessment; Concordance; Likert scale; Test (biology); Medical education; Psychology; Scale (ratio); Mathematics education; Medicine","score_opus":0.20023205139965874,"score_gpt":0.48704153369643943,"score_spread":0.2868094822967807,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3213447869","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9799815,0.00092196785,0.00041139685,0.016180668,0.0022338307,0.00015054281,0.0000020997727,0.000005005738,0.00011298797],"genre_scores_gemma":[0.98263776,0.0027818286,0.0065363254,0.006785151,0.00090528943,0.0000077229115,0.0000051626403,0.000015807382,0.00032495105],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9960986,0.0005649984,0.0021983813,0.00025318074,0.0006653987,0.0002194368],"domain_scores_gemma":[0.94984233,0.04691285,0.0008138447,0.00044376592,0.00091603724,0.0010711463],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0043846113,0.00012924732,0.0007001137,0.000081635735,0.00004925274,0.000052827178,0.00015722061,0.00023477698,0.000042540956],"category_scores_gemma":[0.44204476,0.00008800162,0.00021435502,0.00028138518,0.00029445448,0.0001454289,0.00012308317,0.00088752055,0.000002919831],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010853778,0.0044707633,0.5910712,0.00004754609,0.00007621414,0.00059646333,0.00023442178,0.0000046528876,0.00012296462,0.000105363266,0.0165484,0.38563663],"study_design_scores_gemma":[0.0028307822,0.0040613147,0.95691633,0.007515658,0.00010929003,0.002622069,0.0003314083,0.00052850763,0.000024340494,0.00011823512,0.024832085,0.000109958506],"about_ca_topic_score_codex":0.000020073003,"about_ca_topic_score_gemma":0.00001745356,"teacher_disagreement_score":0.43766016,"about_ca_system_score_codex":0.00006345018,"about_ca_system_score_gemma":0.0031830724,"threshold_uncertainty_score":0.5646634},"labels":[],"label_agreement":null},{"id":"W3216977830","doi":"10.1051/pmed/2021022","title":"Développement du raisonnement clinique chez des étudiantes en sciences infirmières par la méthode SNAPPS-A : une étude pilote","year":2021,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Université Laval","funders":"Fonds de Recherche du Québec - Santé","keywords":"Humanities; Philosophy","score_opus":0.09537892462426356,"score_gpt":0.4058657572337909,"score_spread":0.31048683260952736,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3216977830","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.72551036,0.07820198,0.00918196,0.16769937,0.0034077496,0.00087957963,0.00011803498,0.00032138088,0.014679577],"genre_scores_gemma":[0.82471347,0.10311194,0.05657496,0.006022586,0.0032696314,0.00020222804,0.00010801797,0.0001117382,0.0058854446],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99311775,0.0013893733,0.0014865826,0.0012960044,0.0011804187,0.0015298534],"domain_scores_gemma":[0.9586515,0.03874276,0.00032629434,0.0007461546,0.0003835324,0.0011497893],"candidate_categories":["metaresearch","metaepi_narrow","sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0057424637,0.0006312832,0.0011792666,0.00015320671,0.0006371502,0.000240672,0.000571941,0.00053152867,0.0035980938],"category_scores_gemma":[0.10251431,0.0005688945,0.00046944094,0.0014298063,0.003356728,0.0001711963,0.00072001916,0.00094180275,0.0006476234],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021714225,0.0067099417,0.6882633,0.0006308246,0.00068994524,0.0031068735,0.0060674385,0.00018222269,0.0016630215,0.012229427,0.02881371,0.25142616],"study_design_scores_gemma":[0.008831523,0.0028068246,0.7369415,0.020431273,0.0014122565,0.0004722585,0.007652876,0.0011247494,0.019488137,0.016899152,0.18232386,0.001615566],"about_ca_topic_score_codex":0.00052296097,"about_ca_topic_score_gemma":0.0004822887,"teacher_disagreement_score":0.2498106,"about_ca_system_score_codex":0.00027787947,"about_ca_system_score_gemma":0.0024328972,"threshold_uncertainty_score":0.9996762},"labels":[],"label_agreement":null},{"id":"W3217137518","doi":"10.1051/pmed/2021020","title":"Le raisonnement clinique : regards rétrospectif et prospectif sur le concept","year":2021,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; University of Ottawa","funders":"","keywords":"Humanities; Philosophy; Psychology","score_opus":0.049533610136674615,"score_gpt":0.3562400395800592,"score_spread":0.30670642944338455,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3217137518","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.10556135,0.22536644,0.008229881,0.41855642,0.016120449,0.002048654,0.00046513134,0.0009351504,0.22271653],"genre_scores_gemma":[0.52494526,0.015902713,0.005513171,0.014574227,0.0046817916,0.00019367,0.00037322694,0.00021391794,0.433602],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99426216,0.0006063528,0.0013879578,0.0014011347,0.0009898434,0.0013525701],"domain_scores_gemma":[0.9852941,0.011116766,0.00033542342,0.0013493927,0.00071577146,0.0011885624],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0017616824,0.0005611187,0.0013405646,0.00008101941,0.00033293306,0.00019380052,0.0003424632,0.0007127035,0.0046525146],"category_scores_gemma":[0.08900402,0.0005935593,0.00071113143,0.000693511,0.0009888682,0.0001834611,0.0004362492,0.0015524665,0.00094341167],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00042638145,0.018773269,0.057118967,0.0005051867,0.0010237731,0.006059466,0.0030326578,0.00016547074,0.0017185442,0.18682706,0.52581555,0.19853368],"study_design_scores_gemma":[0.012239156,0.0011446241,0.044842727,0.006536635,0.00050437835,0.00037365817,0.0021438226,0.0003824723,0.026759198,0.0056663095,0.89843726,0.0009697388],"about_ca_topic_score_codex":0.0021299578,"about_ca_topic_score_gemma":0.0007124826,"teacher_disagreement_score":0.4193839,"about_ca_system_score_codex":0.00031494955,"about_ca_system_score_gemma":0.009330993,"threshold_uncertainty_score":0.9998345},"labels":[],"label_agreement":null},{"id":"W329895787","doi":"10.1016/j.amjsurg.2015.03.010","title":"Thinking like an expert: surgical decision making as a cyclical process of being aware","year":2015,"lang":"en","type":"article","venue":"The American Journal of Surgery","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":46,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary; University of Toronto; McMaster University; Western University","funders":"Canadian Institutes of Health Research","keywords":"Flexibility (engineering); Grounded theory; Process (computing); Computer science; Psychology; Knowledge management; Qualitative research; Sociology; Management","score_opus":0.054108970536333574,"score_gpt":0.4002091959845585,"score_spread":0.34610022544822494,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W329895787","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9967699,0.00044198002,0.0005043243,0.0015954296,0.0003484448,0.000057028126,8.27324e-7,0.000019763938,0.00026230857],"genre_scores_gemma":[0.99684745,0.00020837087,0.00096462603,0.0012864835,0.0006549243,0.0000017458689,0.0000014285395,0.000025898053,0.000009045185],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99730307,0.00023211143,0.0009233997,0.00016432347,0.0011042192,0.0002729027],"domain_scores_gemma":[0.97379124,0.023820598,0.0010436721,0.00032833038,0.00054717436,0.00046896754],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.003258695,0.00015035126,0.0009543339,0.00017381675,0.0000648844,0.000030499323,0.0002600029,0.000052199517,0.00003739808],"category_scores_gemma":[0.02438763,0.000087219516,0.00033851087,0.000448916,0.00037114872,0.00012368243,0.000060482405,0.0004402294,0.000006997375],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008452893,0.00096924225,0.44326413,0.000021096857,0.00021250808,0.0028288658,0.0063115777,0.00022457415,0.000074158976,0.00007533171,0.0023214372,0.5352442],"study_design_scores_gemma":[0.012348854,0.013894793,0.69087064,0.080253474,0.0017690937,0.06450247,0.07604223,0.0030835012,0.0021461688,0.047158733,0.005792858,0.00213717],"about_ca_topic_score_codex":0.000046015804,"about_ca_topic_score_gemma":6.349259e-7,"teacher_disagreement_score":0.53310704,"about_ca_system_score_codex":0.000050158862,"about_ca_system_score_gemma":0.00056078547,"threshold_uncertainty_score":0.9838304},"labels":[],"label_agreement":null},{"id":"W335635667","doi":"10.1176/appi.psychotherapy.2006.60.4.375","title":"Making Sense of Client Data: Clinical Experience and Confirmationism Revisited","year":2006,"lang":"en","type":"article","venue":"American Journal of Psychotherapy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Multivariate analysis of variance; Psychology; Think aloud protocol; Affect (linguistics); Situational ethics; Praxis; Social psychology; Variance (accounting); Clinical psychology; Applied psychology; Computer science","score_opus":0.08571847659953946,"score_gpt":0.46570842684537345,"score_spread":0.379989950245834,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W335635667","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9921493,0.00091750815,0.0045921975,0.001563481,0.00020124968,0.00008632398,0.000008794403,0.000007409172,0.00047372666],"genre_scores_gemma":[0.9820253,0.0033322251,0.012355515,0.0018184542,0.00041848898,5.9164023e-7,0.000005501778,0.000012224054,0.000031650627],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99826866,0.00013101105,0.0010280929,0.00017355199,0.0002761852,0.00012250614],"domain_scores_gemma":[0.99606305,0.0022603297,0.000994043,0.00037682077,0.0001850079,0.00012073283],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008991406,0.00009285904,0.0005565972,0.00007477754,0.000025800284,0.00001571658,0.00010979544,0.000036283633,0.00006252413],"category_scores_gemma":[0.001263661,0.000069716385,0.000103687824,0.00018261845,0.0005094036,0.0000658074,0.000008608061,0.00019378922,0.0000020147063],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0023853392,0.001202664,0.5759624,0.000021036674,0.00024656704,0.00016826285,0.00094448327,0.000010906765,0.0011541479,0.0009577594,0.010760802,0.40618563],"study_design_scores_gemma":[0.007850928,0.005247903,0.93793035,0.0034067917,0.00021832656,0.001664103,0.0009745839,0.00032959995,0.00022099585,0.0013448404,0.040549777,0.0002618025],"about_ca_topic_score_codex":0.000033355944,"about_ca_topic_score_gemma":0.0000013247133,"teacher_disagreement_score":0.40592384,"about_ca_system_score_codex":0.000007619304,"about_ca_system_score_gemma":0.000070891096,"threshold_uncertainty_score":0.28429517},"labels":[],"label_agreement":null},{"id":"W4200393022","doi":"10.3389/fmed.2021.765489","title":"Creating Clinical Reasoning Assessment Tools in Different Languages: Adaptation of the Pediatric Emergency Medicine Script Concordance Test to Japanese","year":2021,"lang":"en","type":"article","venue":"Frontiers in Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre Hospitalier Universitaire Sainte-Justine","funders":"Lastentautien Tutkimussäätiö; Japan Foundation for Pediatric Research","keywords":"Concordance; Generalizability theory; Test (biology); Inter-rater reliability; Context (archaeology); Pediatric emergency medicine; Cronbach's alpha; Documentation; Medical education; Medical diagnosis; Medicine; Psychology; Reliability (semiconductor); Emergency department; Computer science; Clinical psychology; Psychometrics; Nursing; Emergency physician","score_opus":0.051022287661406596,"score_gpt":0.40135440312346454,"score_spread":0.35033211546205795,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4200393022","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9761365,0.0044911797,0.0023024555,0.00775677,0.0035012376,0.0006316547,0.000009055782,0.0000284118,0.0051427134],"genre_scores_gemma":[0.9871878,0.0021129597,0.0064729135,0.00095782714,0.001563772,0.000053682947,0.000058927213,0.000029280167,0.001562846],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99614704,0.00034739816,0.0018326651,0.0005311795,0.00077084824,0.00037086537],"domain_scores_gemma":[0.9915834,0.006753189,0.0004372651,0.0006399793,0.0002857476,0.0003004193],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0019039636,0.00025391465,0.0013217849,0.00022153511,0.000050505816,0.0000040664613,0.00021873083,0.0001599574,0.00033946088],"category_scores_gemma":[0.20022339,0.00015702526,0.00013167385,0.0013048552,0.00020428402,0.000051991585,0.00010673005,0.00068771583,0.0000023647267],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008539029,0.00042071677,0.97093743,0.00010402377,0.000027140146,0.000098401215,0.0024852015,0.00006621635,0.00014744901,0.000046860005,0.01763256,0.007948626],"study_design_scores_gemma":[0.004216755,0.00082911685,0.9665766,0.005661042,0.00024383716,0.000016223668,0.01903908,0.002872136,0.000023028391,0.00008348765,0.00029471415,0.00014397675],"about_ca_topic_score_codex":0.00073095143,"about_ca_topic_score_gemma":0.00024295087,"teacher_disagreement_score":0.19831944,"about_ca_system_score_codex":0.00017946352,"about_ca_system_score_gemma":0.0002625642,"threshold_uncertainty_score":0.8065135},"labels":[],"label_agreement":null},{"id":"W4200611129","doi":"10.21203/rs.3.rs-1117083/v1","title":"The Effectiveness of Using Virtual Patient Educational Tools to Improve Medical Students’ Clinical Reasoning Skills: A Systematic","year":2021,"lang":"en","type":"preprint","venue":"Research Square","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"School for Public Health Research; University College London; National Institute for Health and Care Research","keywords":"CINAHL; PsycINFO; Virtual patient; MEDLINE; Inclusion (mineral); Scopus; Medical education; Critical thinking; Psychology; Educational measurement; Medicine; Psychological intervention; Mathematics education; Nursing; Curriculum; Social psychology; Pedagogy","score_opus":0.07384442904551017,"score_gpt":0.5110396626717794,"score_spread":0.4371952336262692,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4200611129","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9897596,0.0018899016,0.00041308382,0.0008679611,0.0017137711,0.0050262557,0.00005892578,0.00002854829,0.00024192306],"genre_scores_gemma":[0.99664813,0.00048060252,0.0006355908,0.0001318829,0.0007326953,0.0010050714,0.00015272628,0.000061702915,0.00015160978],"study_design_codex":"observational","study_design_gemma":"systematic_review","domain_scores_codex":[0.9809007,0.007995659,0.0020827416,0.00110541,0.0071188514,0.0007966617],"domain_scores_gemma":[0.7061313,0.2865959,0.00047929658,0.0019334977,0.003498505,0.0013615399],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.03610086,0.00034682223,0.0016991909,0.00023677244,0.00029383835,0.00040694544,0.0010392421,0.000743332,0.00012596854],"category_scores_gemma":[0.7190976,0.00023150224,0.00070076494,0.0006090892,0.0005173988,0.000045101297,0.0035558918,0.002996898,0.000042457978],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.006336057,0.024412228,0.76668954,0.1327195,0.0055084927,0.0013560451,0.0035284113,0.0009347733,0.00033428706,0.0057033235,0.0022077176,0.050269626],"study_design_scores_gemma":[0.0023438446,0.0026170474,0.3535949,0.63531095,0.00040756987,0.00008106554,0.0031861705,0.0013855422,0.0002927248,0.00034763408,0.000036795896,0.0003957574],"about_ca_topic_score_codex":0.00032449779,"about_ca_topic_score_gemma":0.000024689165,"teacher_disagreement_score":0.68299675,"about_ca_system_score_codex":0.00063516654,"about_ca_system_score_gemma":0.006971186,"threshold_uncertainty_score":0.9993032},"labels":[],"label_agreement":null},{"id":"W4205412635","doi":"10.1097/pcc.0000000000002886","title":"Clinical Reasoning Behind Antibiotic Use in PICUs: A Qualitative Study*","year":2022,"lang":"en","type":"article","venue":"Pediatric Critical Care Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Ottawa; Montreal Children's Hospital; McGill University Health Centre; Université de Montréal; Université Laval; McGill University","funders":"","keywords":"Etiology; Cognition; Critically ill; Clinical judgment; Antibiotics; Disease; Causal reasoning; MEDLINE","score_opus":0.08765681342788624,"score_gpt":0.48468269622299415,"score_spread":0.3970258827951079,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4205412635","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98997486,0.0027901458,0.000073097195,0.0034503092,0.0013478282,0.0010793195,0.000057235473,0.00012243525,0.0011047703],"genre_scores_gemma":[0.99517626,0.00014254854,0.0006631198,0.0016677584,0.0018667215,0.00011252126,0.00011135871,0.000050446517,0.00020929106],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99393725,0.0015193406,0.0015674125,0.00094457413,0.0013659431,0.0006654464],"domain_scores_gemma":[0.9397778,0.058216378,0.00015221776,0.00067407865,0.00040962023,0.00076993083],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0039871526,0.00031875185,0.0012324385,0.00042087634,0.00022636072,0.000023711036,0.0002255851,0.00014292989,0.00087945245],"category_scores_gemma":[0.2670542,0.00026664347,0.0002215024,0.0009459031,0.0004117557,0.00008680921,0.00032008835,0.0017323463,0.00003854186],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00048997503,0.0024778354,0.95904946,0.000119675045,0.000030924304,0.002525908,0.027360251,0.000005681822,0.0000013469721,0.00041516635,0.0039639655,0.0035597796],"study_design_scores_gemma":[0.011515321,0.012259578,0.67989767,0.00052610255,0.0014550495,0.00015358899,0.2924698,0.0001243003,0.0000010276773,0.00020226435,0.0009477158,0.00044759992],"about_ca_topic_score_codex":0.0010589816,"about_ca_topic_score_gemma":0.00014459832,"teacher_disagreement_score":0.27915183,"about_ca_system_score_codex":0.00025805394,"about_ca_system_score_gemma":0.00035346817,"threshold_uncertainty_score":0.9999786},"labels":[],"label_agreement":null},{"id":"W4205418447","doi":"10.1136/bmj-2021-064389","title":"Diagnostic reasoning in cardiovascular medicine","year":2022,"lang":"en","type":"review","venue":"BMJ","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":25,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medical diagnosis; Process (computing); Cognition; Context (archaeology); Experiential learning; Psychology; Diagnostic test; Heuristic; Experiential knowledge; Computer science; Cognitive psychology; Artificial intelligence; Medicine; Psychiatry; Epistemology; Pathology; Mathematics education","score_opus":0.0930780950467053,"score_gpt":0.4187793180822375,"score_spread":0.3257012230355322,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4205418447","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000004360507,0.9802902,0.000010221138,0.00041786354,0.0008858802,0.001812112,0.000014678639,0.00008632687,0.016478388],"genre_scores_gemma":[0.000009640047,0.9953339,0.00018555866,0.00028617005,0.0017160277,0.00095909485,0.00045802817,0.00009902714,0.0009525214],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99617064,0.00047684778,0.00130589,0.0007282115,0.00087117235,0.00044723775],"domain_scores_gemma":[0.9467629,0.051273286,0.00031071945,0.0012987193,0.000036503516,0.00031786165],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0028240944,0.00044647686,0.004387259,0.00034947987,0.00006024597,0.000009887077,0.0002417279,0.00033581466,0.0022134557],"category_scores_gemma":[0.3408891,0.00033373042,0.0014327479,0.0008405478,0.000115788054,0.00001909609,0.00018990363,0.0012337012,0.00021698645],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000007718854,0.00010224489,0.00028707806,0.0034831525,0.0004897351,0.0031831972,0.00003585752,0.0000045886345,1.3080088e-9,0.00008480512,0.020044051,0.9722776],"study_design_scores_gemma":[0.00089201663,0.00019316157,0.0004248481,0.14561856,0.003688439,0.00082107517,0.000027604043,0.0000039355136,7.546379e-9,0.000023285724,0.8480398,0.00026728582],"about_ca_topic_score_codex":0.0003726636,"about_ca_topic_score_gemma":0.0000071265113,"teacher_disagreement_score":0.97201025,"about_ca_system_score_codex":0.00032856292,"about_ca_system_score_gemma":0.00064819434,"threshold_uncertainty_score":0.9999115},"labels":[],"label_agreement":null},{"id":"W4205794204","doi":"10.1097/acm.0000000000000416","title":"Reflecting Upon Reflection in Diagnostic Reasoning","year":2014,"lang":"en","type":"letter","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"","keywords":"Reflection (computer programming); Process (computing); Contrast (vision); Cognition; Computer science; Psychology; Test (biology); Cognitive psychology; Artificial intelligence; Psychiatry","score_opus":0.0837015185357991,"score_gpt":0.42038679158953446,"score_spread":0.33668527305373536,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4205794204","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.018082805,0.004112736,0.00042131334,0.95208603,0.0023899265,0.0009380843,0.0000031194925,0.0003533869,0.021612579],"genre_scores_gemma":[0.07540653,0.0044997525,0.00026380963,0.8777369,0.033382125,0.00014186358,0.00078336254,0.00019448815,0.0075911195],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99439114,0.0003998443,0.0018716541,0.0011580872,0.0011412866,0.0010379999],"domain_scores_gemma":[0.9361736,0.061825264,0.00073551433,0.0007713021,0.00016680262,0.00032750054],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0029312402,0.0006493048,0.0018298007,0.0007484553,0.00009942393,0.0000105434765,0.0003241092,0.0033430555,0.00042929576],"category_scores_gemma":[0.4388643,0.00052367774,0.00017218932,0.0008511973,0.00022968881,0.000061126935,0.000090707246,0.012933918,0.00025948024],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008242683,0.000012207136,0.010463448,0.00026798045,0.0000574679,0.0014830832,0.0005265688,0.0000057907946,0.00010125855,0.000053452,0.98136705,0.0055792527],"study_design_scores_gemma":[0.0043951506,0.0011651994,0.0030372634,0.07116442,0.0008164924,0.0014658622,0.00014436086,0.00034940123,0.000081974744,0.0017632948,0.9149491,0.00066747604],"about_ca_topic_score_codex":0.00064125686,"about_ca_topic_score_gemma":0.000019647772,"teacher_disagreement_score":0.43593305,"about_ca_system_score_codex":0.0004894694,"about_ca_system_score_gemma":0.00026392864,"threshold_uncertainty_score":0.99972147},"labels":[],"label_agreement":null},{"id":"W4206378432","doi":"10.1001/jama.2021.24988","title":"The Rational Diagnostician and Achieving Diagnostic Excellence","year":2022,"lang":"en","type":"article","venue":"JAMA","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Medicine; Excellence; Intensive care medicine; Medical physics; Epistemology","score_opus":0.015979903880900422,"score_gpt":0.2944704211804765,"score_spread":0.2784905172995761,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4206378432","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.935161,0.0037482772,0.00010398732,0.05560755,0.00057427015,0.00035668776,0.000027168653,0.000061644816,0.004359434],"genre_scores_gemma":[0.99429244,0.0007818843,0.00014105144,0.003187771,0.00037241753,0.00012971264,0.000027389791,0.000011883522,0.0010554709],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99897945,0.00008955497,0.00020138796,0.00019886588,0.00033508526,0.00019564784],"domain_scores_gemma":[0.9136107,0.085880816,0.00006538919,0.0002326028,0.000053013835,0.00015747196],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0005523392,0.00008187705,0.00013191064,0.000024670899,0.0006441584,0.00005770123,0.00009333426,0.000029896455,0.00024691064],"category_scores_gemma":[0.15021326,0.00005820433,0.000049260285,0.0001210875,0.00011595051,0.000032352153,0.00017194948,0.00033038625,0.000029051107],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00025246645,0.0004306363,0.50370157,0.00002073177,0.00006835373,0.00039405323,0.00056133367,0.00013589724,0.00012787718,0.010984403,0.07923731,0.40408534],"study_design_scores_gemma":[0.0017219692,0.00045717976,0.8956288,0.00026645063,0.00008549238,0.00018585802,0.00048021632,0.0006219409,0.000031719093,0.0013092998,0.09907841,0.00013266978],"about_ca_topic_score_codex":0.00004096965,"about_ca_topic_score_gemma":0.000007464375,"teacher_disagreement_score":0.4039527,"about_ca_system_score_codex":0.00003681755,"about_ca_system_score_gemma":0.00009295247,"threshold_uncertainty_score":0.85694486},"labels":[],"label_agreement":null},{"id":"W4206717700","doi":"10.2196/29943","title":"A Novel Diagnostic Decision Support System for Medical Professionals: Prospective Feasibility Study","year":2022,"lang":"en","type":"article","venue":"JMIR Formative Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Medicine; Observational study; Context (archaeology); Medical emergency; CLARITY; Health care; Traceability; Transparency (behavior); Decision support system; Emergency department; Complaint; Clinical decision support system; Usability; Nursing; Computer science; Artificial intelligence","score_opus":0.10995411016397517,"score_gpt":0.5223993728625324,"score_spread":0.4124452626985572,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4206717700","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98303086,0.0000681223,0.001779493,0.0015688054,0.0005962072,0.010928437,0.0002505721,0.00012689993,0.0016506158],"genre_scores_gemma":[0.98799896,0.0000042572583,0.00032234553,0.00014495406,0.00019008206,0.010705301,0.00010416763,0.00003325564,0.00049667177],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99290586,0.00064843433,0.0008388019,0.00063135696,0.0042353813,0.0007401931],"domain_scores_gemma":[0.90318745,0.093907855,0.00016569896,0.00070087606,0.001279087,0.0007590432],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.01236206,0.00019789327,0.000608966,0.0002999547,0.0009962962,0.00004015446,0.00040730706,0.00012498083,0.0010018008],"category_scores_gemma":[0.15297978,0.00014450653,0.00020960328,0.00089335797,0.00023213602,0.00012033093,0.00090957095,0.001513665,0.00011572368],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.019801747,0.11402268,0.59473735,0.0014480187,0.0007194464,0.0010474151,0.04662616,0.00002374575,0.00009660317,0.006412893,0.15634772,0.05871619],"study_design_scores_gemma":[0.024014896,0.02450158,0.87558657,0.002793825,0.00008160321,0.00037407287,0.06701045,0.0025831135,0.00012870209,0.0010898109,0.0014865396,0.0003488349],"about_ca_topic_score_codex":0.00005179089,"about_ca_topic_score_gemma":0.000016884216,"teacher_disagreement_score":0.2808492,"about_ca_system_score_codex":0.001173333,"about_ca_system_score_gemma":0.00168925,"threshold_uncertainty_score":0.9999114},"labels":[],"label_agreement":null},{"id":"W4210249571","doi":"10.1007/s12630-021-02164-2","title":"Hippocrates and prophecies: the unfulfilled promise of prediction rules","year":2022,"lang":"en","type":"letter","venue":"Canadian Journal of Anesthesia/Journal canadien d anesthésie","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"London Health Sciences Centre; Lawson Health Research Institute; Western University","funders":"","keywords":"Computer science","score_opus":0.015445414925559271,"score_gpt":0.23881944768960295,"score_spread":0.2233740327640437,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4210249571","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5198729,0.00603645,0.0000016877609,0.47218895,0.0007124027,0.0005684956,0.00010477452,0.0000109595,0.0005033653],"genre_scores_gemma":[0.7943682,0.0052332594,0.0012484626,0.15344688,0.028360153,0.000111293535,0.00064830406,0.00083376176,0.015749644],"study_design_codex":"not_applicable","study_design_gemma":"case_report","domain_scores_codex":[0.99552166,0.00041975628,0.00170498,0.00034804622,0.001101703,0.00090383465],"domain_scores_gemma":[0.99210435,0.0025774536,0.0019131611,0.0004817365,0.001095056,0.0018282193],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002208047,0.0005230832,0.0014317662,0.00091153017,0.0006538481,0.00020599389,0.0006816922,0.0006470961,0.001058475],"category_scores_gemma":[0.0070215613,0.00035415904,0.00060295075,0.0004271469,0.00079497305,0.0001594172,0.000022699875,0.0047082077,0.000005200479],"study_design_candidate":"case_report","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020506252,0.000033771666,0.0031711345,0.00014041131,0.0003266625,0.4397695,0.00094533013,0.000020364678,0.000003436573,0.000027708344,0.55339384,0.0019627805],"study_design_scores_gemma":[0.0011867809,0.0022675327,0.0029971588,0.0018224632,0.0008561786,0.7264853,0.00062784774,0.000016578784,0.0000076177353,0.00041099245,0.2630773,0.00024423635],"about_ca_topic_score_codex":0.006508857,"about_ca_topic_score_gemma":0.0011715079,"teacher_disagreement_score":0.31874204,"about_ca_system_score_codex":0.0008897477,"about_ca_system_score_gemma":0.0124613345,"threshold_uncertainty_score":0.99989104},"labels":[],"label_agreement":null},{"id":"W4210312837","doi":"10.3389/fcvm.2022.767378","title":"Understanding Clinician Macrocognition to Inform the Design of a Congenital Heart Disease Clinical Decision Support System","year":2022,"lang":"en","type":"article","venue":"Frontiers in Cardiovascular Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"North York General Hospital; Canadian Heart Research Centre; University of Toronto","funders":"Natural Sciences and Engineering Research Council of Canada; Hospital for Sick Children","keywords":"Medicine; Intensive care medicine; Clinical decision support system; Disease; Heart disease; Decision support system; Cardiology; Internal medicine; Computer science; Artificial intelligence","score_opus":0.09359844576600695,"score_gpt":0.3487165850772336,"score_spread":0.2551181393112266,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4210312837","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.047951918,0.012131671,0.9188388,0.0064273607,0.008115114,0.004578061,0.000087741195,0.00014595885,0.001723352],"genre_scores_gemma":[0.9914252,0.00011809096,0.0051859645,0.0024452216,0.00047297572,0.00018237068,0.000074035255,0.000036168673,0.00005995067],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.995588,0.00061502313,0.0013678706,0.00048462235,0.0015716532,0.00037283188],"domain_scores_gemma":[0.99285185,0.0054268595,0.00016747472,0.00088879134,0.00014429365,0.00052071153],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.009858225,0.0002147235,0.0013509826,0.00032644556,0.00021406128,0.000008925504,0.00024151051,0.00010094021,0.0000939965],"category_scores_gemma":[0.023833508,0.00015246459,0.0007949916,0.00090978434,0.00032688043,0.000043076383,0.00018043534,0.00060104404,0.000013221246],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0143866455,0.00088002096,0.5606323,0.00032165088,0.0037716972,0.0023736747,0.0017864933,0.0124687385,0.000009321082,0.0005262029,0.30116788,0.1016754],"study_design_scores_gemma":[0.058928315,0.013901435,0.7138292,0.011536665,0.013401123,0.0013132856,0.06435828,0.022362184,0.000025311358,0.0043127057,0.094567716,0.0014637812],"about_ca_topic_score_codex":0.000049971557,"about_ca_topic_score_gemma":0.0000016050778,"teacher_disagreement_score":0.9434733,"about_ca_system_score_codex":0.00050843006,"about_ca_system_score_gemma":0.00039125266,"threshold_uncertainty_score":0.9843892},"labels":[],"label_agreement":null},{"id":"W4210490521","doi":"10.5206/uwomj.v89i2.14361","title":"The influence of fear in emergency medicine","year":2021,"lang":"en","type":"article","venue":"University of Western Ontario Medical Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Psychology; Medical emergency; Medicine; Business","score_opus":0.018784326760218892,"score_gpt":0.29415516561840543,"score_spread":0.27537083885818653,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4210490521","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9894531,0.0003262005,0.000068323345,0.007997286,0.00021681438,0.00003840066,7.375594e-7,0.000003496428,0.0018956737],"genre_scores_gemma":[0.9920569,0.003016982,0.00012725467,0.0001556748,0.00008632509,2.7845266e-8,0.0000026538069,0.000003934895,0.004550261],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9984482,0.00008205486,0.00045423963,0.00011146957,0.00073877047,0.00016527799],"domain_scores_gemma":[0.9970801,0.0017965691,0.00023099732,0.00018615366,0.00033770315,0.0003684661],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0007973755,0.00007006906,0.00033167927,0.00005567837,0.000070462396,0.0000018964506,0.00020930276,0.000105737396,0.0023829564],"category_scores_gemma":[0.014675835,0.000049804257,0.00010691516,0.00016399825,0.00034179832,0.00004135384,0.000083273415,0.00062083657,0.0000063427424],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021818693,0.0002174773,0.9900808,0.000016104988,0.000064565225,0.0010182274,0.0018749387,0.000017456929,0.00012297882,0.000071027214,0.0013606366,0.0049376],"study_design_scores_gemma":[0.002395033,0.00028637578,0.98345435,0.0026688653,0.00008235074,0.00041381607,0.0010384988,0.000008512592,0.000038274273,0.000242255,0.009324845,0.000046839425],"about_ca_topic_score_codex":0.013250979,"about_ca_topic_score_gemma":0.08656564,"teacher_disagreement_score":0.07331467,"about_ca_system_score_codex":0.00009910052,"about_ca_system_score_gemma":0.0012214852,"threshold_uncertainty_score":0.998529},"labels":[],"label_agreement":null},{"id":"W4210524298","doi":"10.2196/35225","title":"Incidence of Diagnostic Errors Among Unexpectedly Hospitalized Patients Using an Automated Medical History–Taking System With a Differential Diagnosis Generator: Retrospective Observational Study","year":2022,"lang":"en","type":"article","venue":"JMIR Medical Informatics","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":32,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Japan Society for the Promotion of Science","keywords":"Observational study; Medicine; Incidence (geometry); Retrospective cohort study; Pediatrics; Medical record; Emergency medicine; Intensive care medicine; Medical emergency; Surgery; Internal medicine","score_opus":0.04375000772609914,"score_gpt":0.3425514200045552,"score_spread":0.298801412278456,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4210524298","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.996522,0.000043006243,0.00051392766,0.00005655448,0.0006659611,0.0016289477,0.000054462816,0.00041994415,0.000095189855],"genre_scores_gemma":[0.99770725,0.00000994047,0.00062669325,0.0003805279,0.00015500616,0.00074985967,0.00031418086,0.0000476285,0.0000089395835],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9914876,0.0003597552,0.0018535355,0.0003501416,0.0054765544,0.00047236847],"domain_scores_gemma":[0.9920891,0.004360515,0.0013168528,0.00057193387,0.0006139657,0.0010476236],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0009752267,0.00036497336,0.0010486845,0.00024044277,0.00025923655,0.00002546228,0.0004759512,0.00026287822,0.001479042],"category_scores_gemma":[0.042413346,0.00029006097,0.00015585915,0.000634391,0.00044854084,0.00028934286,0.00038582121,0.0009421621,0.00000452089],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020192005,0.0043051406,0.98688227,0.00015145093,0.00021329786,0.00023617114,0.006405628,0.00015593099,0.0000011763765,0.00008240753,0.0009464647,0.0004181609],"study_design_scores_gemma":[0.005654903,0.0030324468,0.783896,0.0020418675,0.00027355575,0.000051997915,0.0038926643,0.20085101,0.0000068932463,0.0000021471055,0.000017352597,0.00027915125],"about_ca_topic_score_codex":0.0006637438,"about_ca_topic_score_gemma":0.000072816714,"teacher_disagreement_score":0.20298624,"about_ca_system_score_codex":0.0015632852,"about_ca_system_score_gemma":0.0015651986,"threshold_uncertainty_score":0.9999552},"labels":[],"label_agreement":null},{"id":"W4213369770","doi":"10.1007/s10459-022-10091-0","title":"The concept of errors in medical education: a scoping review","year":2022,"lang":"en","type":"review","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medical education; MEDLINE; Computer science; Psychology; Engineering ethics; Management science; Medicine; Political science; Engineering","score_opus":0.08572745924398721,"score_gpt":0.5622458593667582,"score_spread":0.47651840012277097,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4213369770","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000004553261,0.98756105,0.000001692469,0.0059550633,0.0025246616,0.0022781962,0.0000018809884,0.000008602636,0.001664273],"genre_scores_gemma":[0.000008512904,0.99476355,0.00040595356,0.0034688597,0.00017315257,0.0010145482,0.000055434302,0.000009897194,0.00010006292],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99584085,0.00058768207,0.0019229576,0.00046592337,0.0008516998,0.00033090665],"domain_scores_gemma":[0.9773551,0.020221619,0.0015150124,0.00050899864,0.00008457661,0.00031464966],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0053270687,0.00017565204,0.0012990786,0.00022120599,0.0001840166,0.000010064154,0.00047222522,0.0000875878,0.00041835775],"category_scores_gemma":[0.078409515,0.00011447204,0.00014247138,0.002093084,0.0005603568,0.00010955006,0.00006898944,0.0005673416,0.0000049516],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000014630748,0.00023252892,0.00022099953,0.038019527,0.0000011565245,4.82379e-7,0.00007155789,0.0000014836817,7.618731e-11,0.0010479862,0.0005995054,0.9598033],"study_design_scores_gemma":[0.0000510749,0.00009990892,0.000023118364,0.51178974,0.000016253354,0.000027361151,0.00018488937,0.0000023685711,7.1890267e-9,0.00014833843,0.48760208,0.000054842236],"about_ca_topic_score_codex":0.00032221153,"about_ca_topic_score_gemma":0.0003154138,"teacher_disagreement_score":0.95974845,"about_ca_system_score_codex":0.00038313816,"about_ca_system_score_gemma":0.054577135,"threshold_uncertainty_score":0.95078254},"labels":[],"label_agreement":null},{"id":"W4214526580","doi":"10.1515/dx-2018-0095","title":"Diagnostic Error in Medicine","year":2018,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria","funders":"","keywords":"Medicine; Computer science","score_opus":0.05515093164152968,"score_gpt":0.38950068431997137,"score_spread":0.3343497526784417,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4214526580","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96060926,0.00087234576,0.000042911593,0.018615989,0.00064764754,0.00039571853,0.0000064198807,0.00011302387,0.018696697],"genre_scores_gemma":[0.9916667,0.0006975976,0.00040122273,0.005406137,0.0011606846,0.00021127469,0.000027593609,0.000028216999,0.0004006264],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9984119,0.000052280986,0.00044992846,0.00038927404,0.00031232813,0.00038430252],"domain_scores_gemma":[0.97295094,0.026065713,0.00007284166,0.00043995946,0.00014157394,0.00032900582],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00042481453,0.0001765142,0.00048956624,0.00019016379,0.000049487622,0.000008762735,0.00011602118,0.00013859941,0.0029578642],"category_scores_gemma":[0.16271284,0.00013707626,0.000086088396,0.0005339419,0.00040713773,0.00004342268,0.0000503644,0.00021800752,0.0010009316],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000055780987,0.00045559206,0.91297865,0.000014871626,0.000021143725,0.00023233352,0.00032872005,8.2647756e-7,0.000018435969,0.00044246638,0.06310954,0.02234163],"study_design_scores_gemma":[0.0027634967,0.0012600487,0.96868867,0.0037288123,0.00011506109,0.000034096112,0.00011685057,0.00007822796,0.0006642637,0.0008736631,0.021500712,0.00017608763],"about_ca_topic_score_codex":0.0005722916,"about_ca_topic_score_gemma":0.00021625469,"teacher_disagreement_score":0.16228803,"about_ca_system_score_codex":0.000066445515,"about_ca_system_score_gemma":0.00006730063,"threshold_uncertainty_score":0.9997769},"labels":[],"label_agreement":null},{"id":"W4220956924","doi":"10.1136/bmj.o799","title":"Advancing diagnostic excellence: the cognitive challenge for medicine","year":2022,"lang":"en","type":"letter","venue":"BMJ","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Excellence; Cognition; Data science; Computer science; Medicine; Psychiatry; Political science","score_opus":0.04708524039034935,"score_gpt":0.3829921139710457,"score_spread":0.3359068735806964,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4220956924","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00017910243,0.0056715677,0.00019275352,0.97493804,0.002936806,0.0032585366,0.00017397745,0.00009983979,0.01254938],"genre_scores_gemma":[0.003655332,0.0028595617,0.00012533876,0.93573254,0.038775545,0.0031399056,0.0017842774,0.00013272771,0.01379476],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9970096,0.00018320828,0.0007057861,0.00064549886,0.0008398769,0.000616081],"domain_scores_gemma":[0.7901991,0.2084437,0.00036703155,0.00060642394,0.00023789976,0.00014582378],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0013615818,0.00037969401,0.0008943808,0.00010449965,0.00026893566,0.000012976609,0.0002698215,0.0004582235,0.002723746],"category_scores_gemma":[0.3255003,0.00023710729,0.00033599968,0.00017217577,0.0003211472,0.000020219788,0.00014303104,0.0022741505,0.00009514405],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013630257,0.00008083527,0.000132201,0.00021903162,0.00016010994,0.0011396739,0.00036423127,0.000001382613,4.3265737e-7,0.000020675167,0.98849106,0.009254058],"study_design_scores_gemma":[0.0025577622,0.0020293286,0.0004216243,0.007746641,0.0012484365,0.00016986724,0.0006528,0.00003029474,0.0000025926518,0.00084677926,0.98400104,0.0002928364],"about_ca_topic_score_codex":0.00008702947,"about_ca_topic_score_gemma":0.0000067584465,"teacher_disagreement_score":0.32413873,"about_ca_system_score_codex":0.00011045714,"about_ca_system_score_gemma":0.00032462494,"threshold_uncertainty_score":0.9981879},"labels":[],"label_agreement":null},{"id":"W4220981499","doi":"10.36834/cmej.73556","title":"Raisonnement clinique et simulation : faciliter la priorisation d’hypothèses grâce aux patients simulés. Données d’une recherche quantitative","year":2022,"lang":"fr","type":"article","venue":"Canadian Medical Education Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Université de Montréal; NOSM University; University of Ottawa","funders":"","keywords":"Prioritization; Key (lock); Test (biology); Clinical Practice; Psychology; Medical education; Computer science; Control (management); Clinical trial; Medicine; Family medicine; Artificial intelligence; Management science; Pathology; Engineering","score_opus":0.14879507019797542,"score_gpt":0.45040924557332673,"score_spread":0.3016141753753513,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4220981499","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.47528848,0.021314798,0.0036874963,0.4591512,0.028842486,0.0024981059,0.00047581803,0.000084997635,0.008656625],"genre_scores_gemma":[0.9157952,0.0027901293,0.0021315834,0.07103168,0.0017731164,0.00014861091,0.0005498823,0.00007055379,0.005709207],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.991132,0.004469302,0.0014647783,0.00049048435,0.0017047293,0.00073868455],"domain_scores_gemma":[0.8933936,0.08463736,0.0013070765,0.0007740086,0.0027330935,0.0171549],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.009260791,0.00033748377,0.0005710286,0.00046973603,0.000785609,0.00017627733,0.0003562439,0.0006059337,0.04048134],"category_scores_gemma":[0.6328638,0.00033933145,0.00026579175,0.0008382234,0.0003587946,0.00028204775,0.00007147677,0.0037492118,0.00018499864],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00050874514,0.0058172783,0.039375562,0.00092227233,0.0005071063,0.00034628104,0.050938256,0.009291994,0.000003293305,0.011543975,0.19828492,0.6824603],"study_design_scores_gemma":[0.005254076,0.0021683276,0.05634447,0.015319667,0.0006285892,0.0002768345,0.023547,0.019944405,0.0000051881075,0.0064731482,0.8692482,0.00079011],"about_ca_topic_score_codex":0.021073258,"about_ca_topic_score_gemma":0.011208939,"teacher_disagreement_score":0.6816702,"about_ca_system_score_codex":0.002963118,"about_ca_system_score_gemma":0.090461485,"threshold_uncertainty_score":0.9999059},"labels":[],"label_agreement":null},{"id":"W4221034124","doi":"10.1016/j.jemermed.2021.12.021","title":"Pediatric Musculoskeletal Radiographs: Anatomy and Fractures Prone to Diagnostic Error Among Emergency Physicians","year":2022,"lang":"en","type":"article","venue":"Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"SickKids Foundation; Hospital for Sick Children; University of Toronto","funders":"","keywords":"Radiography; Medicine; Pelvis; Ankle; Radiographic anatomy; Elbow; Tibia; Orthodontics; Humerus; Condyle; Radiology; Anatomy","score_opus":0.021006703841841213,"score_gpt":0.373744209498817,"score_spread":0.35273750565697576,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4221034124","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97340083,0.012373892,0.00016875942,0.007726364,0.0050611505,0.00053142814,0.000012877719,0.00003076419,0.0006939278],"genre_scores_gemma":[0.9870269,0.008050384,0.00022060212,0.00041409535,0.003651544,0.000046482033,0.000016775031,0.000047947975,0.0005253129],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99624723,0.00017211029,0.0015541004,0.00035906586,0.0012207694,0.0004467161],"domain_scores_gemma":[0.99566895,0.0020684053,0.00080170005,0.00033139825,0.00038130666,0.0007482143],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011995657,0.00030682512,0.0009027647,0.00071806833,0.00023221405,0.0000031723969,0.00025200233,0.00007964723,0.0064523662],"category_scores_gemma":[0.034755953,0.00023203972,0.00040796856,0.0014777855,0.000092693095,0.00010644423,0.0001213974,0.00088476914,0.0000072171238],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00024242955,0.0008771668,0.76355153,0.00012953105,0.00027296436,0.0003985776,0.0013847062,0.00014714497,0.00030119813,0.000041227242,0.22400153,0.008651972],"study_design_scores_gemma":[0.0025363127,0.005806427,0.9774156,0.0004700978,0.0015508969,0.00016496849,0.0012831913,0.000047249276,0.000014987766,0.0004305279,0.009996884,0.00028283536],"about_ca_topic_score_codex":0.00012293355,"about_ca_topic_score_gemma":0.000018615241,"teacher_disagreement_score":0.21400464,"about_ca_system_score_codex":0.0000647213,"about_ca_system_score_gemma":0.000111902504,"threshold_uncertainty_score":0.9944559},"labels":[],"label_agreement":null},{"id":"W4221075296","doi":"10.1007/s40037-022-00701-3","title":"Advancing the assessment of clinical reasoning across the health professions: Definitional and methodologic recommendations","year":2022,"lang":"en","type":"review","venue":"Perspectives on Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":36,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University Health Centre; McGill University","funders":"Fonds de Recherche du Québec - Santé","keywords":"Terminology; Health care; Deductive reasoning; Quality (philosophy); MEDLINE; Management science; Psychology; Medical education; Computer science; Medicine; Artificial intelligence","score_opus":0.26378319855564253,"score_gpt":0.6325051687407531,"score_spread":0.3687219701851106,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4221075296","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00008921754,0.9320016,0.00044525383,0.06336784,0.0013658213,0.0013223235,0.00004864424,0.000031999945,0.0013272822],"genre_scores_gemma":[0.00017745378,0.9859848,0.008442652,0.0034768835,0.0006775147,0.00072028756,0.00034683722,0.000027214779,0.00014633135],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99274784,0.0040804306,0.0014212626,0.0005839979,0.00087129493,0.000295195],"domain_scores_gemma":[0.8108229,0.1850488,0.0021254516,0.001016777,0.00028287427,0.0007031924],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.014358819,0.00024423332,0.001419593,0.00007781307,0.00067810615,0.000021445541,0.00027633796,0.0002428041,0.0013170068],"category_scores_gemma":[0.3339626,0.00012113316,0.00044688405,0.000424728,0.00055803935,0.000029169363,0.00019569373,0.0025517098,0.0000046799178],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000016927694,0.0012457141,0.00022798174,0.00053942995,0.00012445482,9.593662e-7,0.0017733251,2.0517031e-7,2.1622713e-9,0.012188308,0.0060936753,0.97778904],"study_design_scores_gemma":[0.0004774534,0.0008100952,0.0053113913,0.037902463,0.0004858526,0.00014733488,0.045635022,0.00005782886,9.487067e-9,0.0009733254,0.9080412,0.00015803546],"about_ca_topic_score_codex":0.00007584648,"about_ca_topic_score_gemma":0.000011040985,"teacher_disagreement_score":0.977631,"about_ca_system_score_codex":0.00035357501,"about_ca_system_score_gemma":0.010757996,"threshold_uncertainty_score":0.9997494},"labels":[],"label_agreement":null},{"id":"W4223450571","doi":"10.1097/pts.0000000000000999","title":"Defining Diagnostic Error: A Scoping Review to Assess the Impact of the National Academies’ Report Improving Diagnosis in Health Care","year":2022,"lang":"en","type":"review","venue":"Journal of Patient Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":24,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto East General Hospital; University of Toronto","funders":"National Institute on Aging; Agency for Healthcare Research and Quality","keywords":"Health care; MEDLINE; Medicine; Data science; Computer science; Medical physics; Political science","score_opus":0.12509668580769764,"score_gpt":0.49385948039961414,"score_spread":0.3687627945919165,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4223450571","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0003628933,0.9946536,0.000009579222,0.00153215,0.000581624,0.0025959332,0.000102826045,0.000006046386,0.00015534877],"genre_scores_gemma":[0.0013812148,0.9965129,0.00019351779,0.001401358,0.0001923158,0.00021193811,0.00005482689,0.000044175264,0.000007787205],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99195695,0.0011093151,0.0046117,0.00033304334,0.001613891,0.00037510335],"domain_scores_gemma":[0.9232212,0.06699583,0.008227566,0.0005832721,0.0005858265,0.00038631933],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005176038,0.0003736765,0.0029530327,0.00025836896,0.0001859813,0.000014590556,0.00048313403,0.000170813,0.00022107709],"category_scores_gemma":[0.32438502,0.00018587112,0.0018816672,0.0010650923,0.00006203995,0.00006106087,0.00035670056,0.0021048298,0.000003887085],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006781665,0.00039739232,0.021504845,0.0639427,0.00037295185,0.00024325785,0.00083336676,0.00066824234,2.1147725e-8,0.00002712753,0.0070162797,0.904926],"study_design_scores_gemma":[0.00041098127,0.0009719408,0.005505662,0.89999133,0.0007391589,0.0012569127,0.00014435129,0.000001930844,1.8405595e-7,0.0000072490366,0.09078991,0.00018038323],"about_ca_topic_score_codex":0.00016142997,"about_ca_topic_score_gemma":0.000018806644,"teacher_disagreement_score":0.90474564,"about_ca_system_score_codex":0.0022833773,"about_ca_system_score_gemma":0.008556865,"threshold_uncertainty_score":0.9970637},"labels":[],"label_agreement":null},{"id":"W4223547123","doi":"10.4045/tidsskr.22.0068","title":"Introduction of new health technologies – is the patient being heard?","year":2022,"lang":"no","type":"article","venue":"Tidsskrift for Den norske legeforening","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute of Health Economics","funders":"","keywords":"Psychology","score_opus":0.02034970041479691,"score_gpt":0.31239114157219583,"score_spread":0.29204144115739894,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4223547123","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6369978,0.025642956,0.0031509786,0.3221643,0.006010291,0.0042130086,0.00039221332,0.0006039964,0.0008244555],"genre_scores_gemma":[0.9875675,0.00040716544,0.0033936384,0.0021018642,0.0014995114,0.00021670667,0.00023044714,0.000093666895,0.0044895047],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9960097,0.00016725932,0.0013110556,0.00074705045,0.00088575453,0.0008791866],"domain_scores_gemma":[0.9935502,0.004217583,0.0008780061,0.0010060114,0.0001693967,0.0001788278],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0016338388,0.00037758204,0.00092313427,0.00024864968,0.0009886536,0.000043687804,0.00046938166,0.00017642726,0.0005661376],"category_scores_gemma":[0.015208817,0.00029473306,0.00049467204,0.0007957289,0.00029616847,0.00008318517,0.00056268426,0.0010826201,0.000032633605],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00072353677,0.00037252915,0.0104417745,0.00024352956,0.0004403844,0.000011318953,0.013943736,0.0007352385,0.00010218014,0.0051220106,0.34212583,0.6257379],"study_design_scores_gemma":[0.0049529755,0.011898978,0.0041418187,0.0025928302,0.0010157182,0.00022868077,0.032284,0.003111588,0.0058581214,0.006967255,0.92616445,0.0007835628],"about_ca_topic_score_codex":0.001512898,"about_ca_topic_score_gemma":0.000025675228,"teacher_disagreement_score":0.62495434,"about_ca_system_score_codex":0.0003439239,"about_ca_system_score_gemma":0.0007967116,"threshold_uncertainty_score":0.99995047},"labels":[],"label_agreement":null},{"id":"W4224021915","doi":"10.1016/j.oooo.2021.12.020","title":"Learning by concordance – a new tool for developing clinical reasoning in oral pathology and dental education","year":2022,"lang":"en","type":"article","venue":"Oral Surgery Oral Medicine Oral Pathology and Oral Radiology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Concordance; Medicine; Pathology; Medical physics; Psychology; Internal medicine","score_opus":0.041758555245611166,"score_gpt":0.37560151334923464,"score_spread":0.33384295810362347,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4224021915","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9809986,0.0050525074,0.0010097867,0.009416795,0.0025171312,0.0007999613,0.00002655275,0.00011693306,0.00006177676],"genre_scores_gemma":[0.977117,0.0015994528,0.008169905,0.010436562,0.0005443016,0.00023422371,0.0007962564,0.000071186565,0.0010311031],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9937872,0.0019362959,0.001644251,0.0013088087,0.0002679516,0.0010554409],"domain_scores_gemma":[0.98787,0.010626881,0.0005321395,0.00032644722,0.0001317569,0.00051277055],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.005552143,0.0005374071,0.0024979357,0.00039587016,0.0004238891,0.0000197074,0.00018635116,0.0005926213,0.0002299659],"category_scores_gemma":[0.03613318,0.00047544396,0.00021730585,0.00053416914,0.0013252258,0.00015519597,0.00028627724,0.0015769205,0.000009090132],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019572298,0.00023573861,0.760897,0.000043146658,0.000051168663,0.0011954546,0.00023488418,0.000006791284,0.00037108155,0.00066984573,0.02941969,0.20491797],"study_design_scores_gemma":[0.017645812,0.012727392,0.85742766,0.0014019214,0.0007043044,0.026766034,0.002631616,0.00093789917,0.00008056921,0.0052312305,0.07284462,0.001600943],"about_ca_topic_score_codex":0.0006520605,"about_ca_topic_score_gemma":0.0002213111,"teacher_disagreement_score":0.20331703,"about_ca_system_score_codex":0.0001837409,"about_ca_system_score_gemma":0.00090386695,"threshold_uncertainty_score":0.99976975},"labels":[],"label_agreement":null},{"id":"W4224248982","doi":"10.3138/jvme-2021-0138","title":"Effect of Romanowsky-Stained Concentrated Preparations versus Direct Smears on Veterinary Students’ Ability to Identify Bacterial Sepsis in Fluid Cytology Samples from Dogs, Cats, and Horses","year":2022,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Cytology; Medicine; Sepsis; Significant difference; Pathology; Surgery; Internal medicine","score_opus":0.06030636737446409,"score_gpt":0.4584693574047977,"score_spread":0.3981629900303336,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4224248982","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9952232,0.000302234,0.0000050898743,0.0014671364,0.0024794866,0.00041654264,0.000046055142,0.00001129737,0.000048952952],"genre_scores_gemma":[0.99890685,0.00015712797,0.00018803682,0.00025981953,0.00028223317,0.000058376278,0.00012319954,0.000013214709,0.0000111173285],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99683344,0.0010692015,0.0008523683,0.00027921904,0.0007719655,0.00019383019],"domain_scores_gemma":[0.98953646,0.009312409,0.00033008438,0.00024961107,0.00013270175,0.00043874292],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0014477979,0.00015950162,0.0006342626,0.00019873503,0.00007630993,0.000019467587,0.00020484213,0.000112031485,0.00082203414],"category_scores_gemma":[0.03866859,0.00013714119,0.000115929695,0.00025090986,0.00013164495,0.000072420735,0.00015242086,0.00036543517,0.0000029110229],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.23093553,0.019427316,0.36502182,0.0005868261,0.001068741,0.0020290578,0.010689207,0.00030145838,0.22076018,0.000046437643,0.017005282,0.13212813],"study_design_scores_gemma":[0.010463087,0.06488062,0.9159777,0.0016131965,0.00034522833,0.00045873545,0.00082304265,0.00006033668,0.0008975808,0.00003582176,0.004236249,0.00020841057],"about_ca_topic_score_codex":0.00047980924,"about_ca_topic_score_gemma":0.000014037716,"teacher_disagreement_score":0.55095583,"about_ca_system_score_codex":0.00029105492,"about_ca_system_score_gemma":0.00092574715,"threshold_uncertainty_score":0.96942914},"labels":[],"label_agreement":null},{"id":"W4225262802","doi":"10.36834/cmej.73361","title":"Four ways to get a grip on making robust decisions from workplace-based assessments","year":2022,"lang":"en","type":"article","venue":"Canadian Medical Education Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Computer science","score_opus":0.06884077317129825,"score_gpt":0.3855496618531936,"score_spread":0.3167088886818954,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4225262802","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5934117,0.00044541524,0.0135290325,0.36423868,0.012262172,0.00074649154,0.00015795403,0.00006738135,0.01514123],"genre_scores_gemma":[0.7686989,0.00001871407,0.009318402,0.21948323,0.0017054898,0.000076893026,0.000112675065,0.00003255697,0.00055316836],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99723,0.00014132803,0.0004732898,0.00029914943,0.0014092866,0.00044692468],"domain_scores_gemma":[0.98280805,0.004978404,0.00013776682,0.00035979023,0.00021501913,0.011500978],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011205163,0.00014462479,0.0002672057,0.00044073974,0.00056421413,0.00008279517,0.0003067619,0.00012364479,0.06515582],"category_scores_gemma":[0.20305318,0.0001328898,0.0001309613,0.00044935234,0.000040146228,0.000029645142,0.00003474861,0.0014410021,0.00017796409],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000163234,0.0005579276,0.010972504,0.0000022232423,0.000049740032,0.00072193076,0.00019640333,0.0012037116,0.0000017498237,0.00037245182,0.78853667,0.19722146],"study_design_scores_gemma":[0.0049646115,0.001593988,0.13409045,0.0076253703,0.0002523627,0.0014332762,0.0023462411,0.0054121558,0.0000052642463,0.0026136513,0.839071,0.00059166993],"about_ca_topic_score_codex":0.0046500443,"about_ca_topic_score_gemma":0.005245897,"teacher_disagreement_score":0.20193265,"about_ca_system_score_codex":0.0013237884,"about_ca_system_score_gemma":0.05109703,"threshold_uncertainty_score":0.95428234},"labels":[],"label_agreement":null},{"id":"W4225414557","doi":"10.1096/fasebj.2022.36.s1.0i610","title":"Explicit Connections Between Anatomy and Clinical Science are Key to Successful Cognitive Integration","year":2022,"lang":"en","type":"article","venue":"The FASEB Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; University of Toronto; Humber Polytechnic","funders":"","keywords":"Curriculum; Cognition; Psychology; Session (web analytics); Computer science; Neuroscience; Pedagogy","score_opus":0.05225721607692946,"score_gpt":0.39830260067255324,"score_spread":0.3460453845956238,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4225414557","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.988431,0.00007960982,0.001382835,0.008826398,0.000357102,0.0002050533,0.00003097891,0.000019918958,0.00066712504],"genre_scores_gemma":[0.99730027,0.00005384908,0.0001420575,0.0018241208,0.0004897694,0.000013129117,0.000006226217,0.0000078767025,0.00016269054],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99860674,0.00017823851,0.00036151978,0.00020116607,0.00044824794,0.00020408955],"domain_scores_gemma":[0.990606,0.008342855,0.00017832241,0.00014731426,0.0002708758,0.00045464133],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0024938388,0.00008172496,0.00022440906,0.00013118432,0.0010039174,0.00008500033,0.00015381089,0.000028012088,0.00024134922],"category_scores_gemma":[0.03257332,0.000053513537,0.000083887215,0.0004901785,0.00026057273,0.00007372827,0.00015038718,0.00082180405,0.00002044037],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002386104,0.00022408502,0.9473384,0.0000018462536,0.00008001749,0.000086737804,0.0008776528,0.000015361898,0.0004232358,0.00045121546,0.0020934697,0.04816937],"study_design_scores_gemma":[0.0020879204,0.00138309,0.97984034,0.0005669655,0.0003536951,0.0008352445,0.011458467,0.00032267495,0.0016848221,0.0007249587,0.0006043962,0.00013740058],"about_ca_topic_score_codex":0.000053724965,"about_ca_topic_score_gemma":0.00001134407,"teacher_disagreement_score":0.048031967,"about_ca_system_score_codex":0.00007995172,"about_ca_system_score_gemma":0.00027003116,"threshold_uncertainty_score":0.97557575},"labels":[],"label_agreement":null},{"id":"W4230148605","doi":"10.3138/ptc.64.3.325","title":"What Does the Cochrane Collaboration Say about Rheumatoid Arthritis?","year":2012,"lang":"en","type":"article","venue":"Physiotherapy Canada","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Rheumatoid arthritis; Medicine; Cochrane collaboration; Physical therapy; Alternative medicine; Internal medicine; Cochrane Library; Pathology","score_opus":0.0073351209721885975,"score_gpt":0.31320071932114607,"score_spread":0.3058655983489575,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4230148605","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93497866,0.036959115,0.000025330475,0.023549838,0.003273453,0.00038932473,0.000021289074,0.000047909285,0.0007550886],"genre_scores_gemma":[0.954558,0.032737277,0.00005292944,0.010969948,0.00062488136,0.000035218665,0.00003396856,0.000024214842,0.0009635317],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9988427,0.000053881802,0.00022664279,0.00015039633,0.00041899734,0.00030735796],"domain_scores_gemma":[0.99678767,0.0024366342,0.00008393875,0.0003476933,0.000114618124,0.00022944558],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00019750303,0.00013224164,0.0002638044,0.000018655259,0.00012951516,0.000068783826,0.000077591874,0.000047560297,0.0005637517],"category_scores_gemma":[0.0012764958,0.00007394751,0.000054625223,0.00021136363,0.00006646032,0.00017426595,0.000009528733,0.00012718585,0.000020026471],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007822404,0.0012208347,0.050116755,0.00012787223,0.00058429735,0.000054178756,0.0054329103,0.000057963927,0.014598006,0.006854601,0.46656978,0.45360056],"study_design_scores_gemma":[0.0087490445,0.00043925602,0.2402458,0.0105174165,0.00006911071,0.00010360578,0.0071913805,0.00037903475,0.024268053,0.00096946297,0.70610553,0.0009623187],"about_ca_topic_score_codex":0.15096408,"about_ca_topic_score_gemma":0.34354448,"teacher_disagreement_score":0.45263824,"about_ca_system_score_codex":0.0001212885,"about_ca_system_score_gemma":0.00066745933,"threshold_uncertainty_score":0.8546897},"labels":[],"label_agreement":null},{"id":"W4232598396","doi":"10.1093/med/9780199685219.003.0002","title":"Basic concepts","year":2014,"lang":"en","type":"book","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Categorical variable; Reliability (semiconductor); Scale (ratio); Computer science; Key (lock); Management science; Psychology; Data science; Artificial intelligence; Machine learning; Engineering; Geography; Cartography; Computer security","score_opus":0.0262458398663024,"score_gpt":0.3535631039750397,"score_spread":0.3273172641087373,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4232598396","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000068524394,0.00033619252,0.00012767327,0.00081068784,0.00041638585,0.00018556307,0.0000061419328,0.00013128172,0.99791753],"genre_scores_gemma":[0.0008932336,0.00008027597,0.0003320229,0.006607009,0.0013227293,0.000006355415,0.00023229429,0.00003893867,0.99048716],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9990209,0.00001293383,0.00026541026,0.00028339663,0.00024960664,0.00016777232],"domain_scores_gemma":[0.9902658,0.008897559,0.00007982734,0.0004133551,0.00008182089,0.00026163267],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00015830352,0.00017601308,0.00055716373,0.000040545678,0.000016511696,0.000008500808,0.00006527829,0.00039685785,0.006452795],"category_scores_gemma":[0.016489698,0.00012854708,0.00021929607,0.000018967306,0.00010712088,0.0000054343295,0.00003204083,0.00036542586,0.0047169104],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000016604361,0.000031705524,0.00017679132,0.000028876442,0.00004297574,0.00004547731,0.0000061400733,9.0556334e-8,5.225442e-7,0.009705865,0.968799,0.021145938],"study_design_scores_gemma":[0.00096321286,0.00028643908,0.00061158853,0.0032593622,0.00023470457,0.00003473098,0.0000015669165,0.000011804493,0.000009814268,0.0027887945,0.99163973,0.00015824233],"about_ca_topic_score_codex":0.0000069278067,"about_ca_topic_score_gemma":0.0000036395347,"teacher_disagreement_score":0.022840727,"about_ca_system_score_codex":0.000060314473,"about_ca_system_score_gemma":0.00046541978,"threshold_uncertainty_score":0.99605805},"labels":[],"label_agreement":null},{"id":"W4232628733","doi":"10.1207/s15328015tlm1901_7","title":"Expert-Type Knowledge Structure in Medical Students is Associated With Increased Odds of Diagnostic Success","year":2007,"lang":"en","type":"article","venue":"Teaching and Learning in Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Odds; Medicine; Odds ratio; Psychology; Medical education; Family medicine; Internal medicine; Logistic regression","score_opus":0.017147574943745116,"score_gpt":0.38419507119419877,"score_spread":0.36704749625045363,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4232628733","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9951196,0.0022184802,0.00007929682,0.0010715995,0.0001328091,0.00020076316,9.557741e-7,0.000056598255,0.0011199038],"genre_scores_gemma":[0.99800295,0.00026225302,0.0001295706,0.0011075736,0.0002026359,0.0000030439742,0.000050674436,0.000028046441,0.00021323198],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9970243,0.0005686021,0.00067491626,0.00036182124,0.0009838429,0.00038654418],"domain_scores_gemma":[0.9533167,0.045810957,0.00019192602,0.00019334337,0.00010568305,0.00038140902],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.007723229,0.00021911357,0.0008237575,0.00033413092,0.0000890607,0.000009100795,0.00018510976,0.00028293172,0.00024645746],"category_scores_gemma":[0.3246136,0.0001502036,0.00002948022,0.0004772835,0.00030085942,0.000034632318,0.000074672826,0.002162212,0.0000014377517],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00033306677,0.00052119105,0.95795375,0.000046297788,0.000054143664,0.0004952675,0.009558878,0.00001088779,0.0001019508,0.00002603844,0.00036600197,0.030532526],"study_design_scores_gemma":[0.008420246,0.001500063,0.9696973,0.017423479,0.00007999826,0.00006504258,0.0015765814,0.0005725864,0.00003605814,0.000022343896,0.00044738388,0.00015890352],"about_ca_topic_score_codex":0.003141624,"about_ca_topic_score_gemma":0.00091266795,"teacher_disagreement_score":0.31689036,"about_ca_system_score_codex":0.000088914014,"about_ca_system_score_gemma":0.00015051245,"threshold_uncertainty_score":0.93938553},"labels":[],"label_agreement":null},{"id":"W4235713443","doi":"10.7202/1084730ar","title":"Développement d’un modèle de raisonnement clinique par son explicitation auprès d’optométristes de deux niveaux d’expertise contrastants","year":2012,"lang":"fr","type":"article","venue":"Recherches qualitatives","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Humanities; Philosophy; Physics","score_opus":0.2789526601052997,"score_gpt":0.4650307938533565,"score_spread":0.18607813374805682,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4235713443","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.78544885,0.05333947,0.1315923,0.025550403,0.0014135832,0.0010158917,0.00007640735,0.0001463914,0.0014167133],"genre_scores_gemma":[0.8275043,0.010038034,0.15597628,0.0025582737,0.001691663,0.00027629582,0.00006713038,0.00010477013,0.0017832875],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.98856485,0.0066573997,0.0016533453,0.0006187284,0.0006594647,0.0018462177],"domain_scores_gemma":[0.90411353,0.09264431,0.0007237014,0.0005263468,0.0004917132,0.0015003782],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.017401055,0.00063316035,0.0011183148,0.0001853987,0.00019208429,0.00007894601,0.00028565933,0.0008557152,0.00050800265],"category_scores_gemma":[0.18487588,0.00060609746,0.00043033413,0.00044159126,0.00045345278,0.00041202537,0.000099300276,0.00095230073,0.00038479004],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001744563,0.0073056123,0.1750723,0.0007001877,0.0014696383,0.000111999936,0.47953355,0.000065607084,0.014836177,0.017457124,0.010260285,0.29144296],"study_design_scores_gemma":[0.020097677,0.0058739022,0.57381463,0.027145945,0.0026477897,0.00013524137,0.1556818,0.0040662377,0.13741592,0.05570675,0.014452725,0.0029613634],"about_ca_topic_score_codex":0.00068224425,"about_ca_topic_score_gemma":0.000037393605,"teacher_disagreement_score":0.39874235,"about_ca_system_score_codex":0.0029528083,"about_ca_system_score_gemma":0.0012300247,"threshold_uncertainty_score":0.99963903},"labels":[],"label_agreement":null},{"id":"W4236185239","doi":"10.1007/s12630-014-0119-8","title":"And the winner is…","year":2014,"lang":"en","type":"letter","venue":"Canadian Journal of Anesthesia/Journal canadien d anesthésie","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Computer science","score_opus":0.01591469756027899,"score_gpt":0.24817166561317275,"score_spread":0.23225696805289375,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4236185239","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.056846112,0.0048618093,0.000007165142,0.9357719,0.00069154584,0.00027802846,0.000008020229,0.000008659322,0.0015267725],"genre_scores_gemma":[0.1518654,0.00090009766,0.00030998458,0.8156784,0.022027252,0.0000069512803,0.000033622346,0.00023968129,0.008938598],"study_design_codex":"not_applicable","study_design_gemma":"case_report","domain_scores_codex":[0.99542916,0.00043257867,0.0015563513,0.0003762413,0.0009762131,0.0012294321],"domain_scores_gemma":[0.9889483,0.00434376,0.0015325233,0.0006061303,0.0011848354,0.0033844907],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0031230666,0.0006331211,0.0018875173,0.00082482916,0.00056047365,0.00047315285,0.0007549215,0.0011516296,0.0005870057],"category_scores_gemma":[0.007740847,0.00039478997,0.00089412177,0.00033003505,0.0013438424,0.00013227478,0.000013392735,0.0063450197,0.000050231756],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010259595,0.0000058494097,0.000629217,0.000034149387,0.00017198785,0.42979044,0.00039285372,0.0000020154225,9.0793534e-8,0.00006754823,0.56764823,0.0011550243],"study_design_scores_gemma":[0.0021301096,0.00051125186,0.0005601918,0.0016057981,0.00067401165,0.55247265,0.0000773938,0.0000093532035,0.0000011516129,0.00059968035,0.44111833,0.00024008066],"about_ca_topic_score_codex":0.009583791,"about_ca_topic_score_gemma":0.0013597893,"teacher_disagreement_score":0.1265299,"about_ca_system_score_codex":0.0005294723,"about_ca_system_score_gemma":0.0070638387,"threshold_uncertainty_score":0.9998504},"labels":[],"label_agreement":null},{"id":"W4236397395","doi":"10.7202/1084633ar","title":"Explicitation du raisonnement clinique : méthodologie novatrice menant à l’identification de deux étapes de développement au cours des 12 mois entourant la fin de la formation préclinique et le début de la formation clinique d’étudiants en médecine","year":2013,"lang":"fr","type":"article","venue":"Recherches qualitatives","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières; Université de Sherbrooke","funders":"","keywords":"Humanities; Physics; Philosophy","score_opus":0.24620202616054365,"score_gpt":0.5007693361992113,"score_spread":0.2545673100386676,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4236397395","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5255133,0.0011454852,0.4440929,0.027437955,0.00020317346,0.0010680723,0.000035410354,0.00010994064,0.00039376816],"genre_scores_gemma":[0.71741897,0.02054245,0.25911328,0.0010090119,0.000451618,0.0008953429,0.00020043446,0.00007374672,0.000295132],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.9562455,0.03843869,0.0028132363,0.0006665664,0.000603411,0.0012325932],"domain_scores_gemma":[0.6992237,0.29686555,0.0017841869,0.0005210936,0.0009264082,0.000679074],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.08701459,0.0005891042,0.00091710815,0.00037835544,0.0002749257,0.00029037267,0.00039744837,0.0013443165,0.000055557586],"category_scores_gemma":[0.29136607,0.0005537744,0.00035081242,0.0005340204,0.0006408244,0.0012365093,0.00012413066,0.001309978,0.00009109839],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012557148,0.006932099,0.1495625,0.0024886332,0.00076738803,0.00009882486,0.57843655,0.0016838628,0.034634903,0.026212212,0.015889619,0.18203771],"study_design_scores_gemma":[0.007045366,0.0012632349,0.71246547,0.007885067,0.0005751741,0.00020378707,0.079765886,0.011227542,0.036279786,0.14106752,0.0013860872,0.00083510205],"about_ca_topic_score_codex":0.00262618,"about_ca_topic_score_gemma":0.0004600101,"teacher_disagreement_score":0.562903,"about_ca_system_score_codex":0.0050341687,"about_ca_system_score_gemma":0.0047311783,"threshold_uncertainty_score":0.99995214},"labels":[],"label_agreement":null},{"id":"W4237165340","doi":"10.15766/mep_2374-8265.1118","title":"Structured Oral Examinations in Internal Medicine - Case F","year":2009,"lang":"en","type":"article","venue":"MedEdPORTAL","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Saskatchewan","funders":"","keywords":"Oral medicine; Medicine; Oral examination; Family medicine; Dentistry; Oral health","score_opus":0.032095240291587286,"score_gpt":0.37501655266529843,"score_spread":0.34292131237371115,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4237165340","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9765568,0.00013971709,0.000078662415,0.002843494,0.00027809516,0.00015782146,0.0000036169454,0.00005246895,0.019889342],"genre_scores_gemma":[0.9961705,0.000011230562,0.00048562387,0.0012686364,0.0002663778,0.000004595187,0.000051719835,0.000008340217,0.0017329406],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99890685,0.000027512844,0.00039235977,0.0002203009,0.00024076378,0.00021222798],"domain_scores_gemma":[0.9985291,0.00092108385,0.00006587564,0.00020696977,0.00005465604,0.00022235773],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0002905362,0.00012014187,0.00031501392,0.00014180852,0.000029124405,0.00000706605,0.000054002983,0.00008795534,0.0013572739],"category_scores_gemma":[0.014386763,0.00008969621,0.00005532402,0.00022280384,0.00009093441,0.00003805532,0.000011872457,0.0002574137,0.000028992163],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003285189,0.0011357032,0.41972214,0.000021962347,0.00008062956,0.118568316,0.0020508904,0.000011495454,0.0010066931,0.0048289746,0.039060403,0.4131843],"study_design_scores_gemma":[0.003854244,0.0008795248,0.97979456,0.00072267157,0.00009451612,0.010504688,0.00038319215,0.0003356988,0.00024764476,0.0019658755,0.0010730322,0.00014435487],"about_ca_topic_score_codex":0.00069774146,"about_ca_topic_score_gemma":0.00057731644,"teacher_disagreement_score":0.5600724,"about_ca_system_score_codex":0.000042610027,"about_ca_system_score_gemma":0.000057079982,"threshold_uncertainty_score":0.99955565},"labels":[],"label_agreement":null},{"id":"W4238096937","doi":"10.1007/978-94-007-0753-5_102374","title":"Margin of Error","year":2014,"lang":"en","type":"book-chapter","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Brandon University; University of Northern British Columbia","funders":"","keywords":"Margin (machine learning); Geology; Computer science; Machine learning","score_opus":0.05262100174366941,"score_gpt":0.3407232302365637,"score_spread":0.2881022284928943,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4238096937","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00015763413,0.00022111041,0.00020120005,0.00076073024,0.00017479823,0.00014226741,0.000009944492,0.000040237475,0.9982921],"genre_scores_gemma":[0.012047078,0.0001223315,0.0009505936,0.0009779116,0.00027025657,0.0000016358256,0.00005567629,0.000033316242,0.9855412],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.99915063,0.000003643626,0.0003357707,0.00019032079,0.0002210589,0.00009857844],"domain_scores_gemma":[0.99551296,0.003755771,0.00012798396,0.0003674051,0.0000924067,0.00014346684],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00014492354,0.0001435996,0.00057165936,0.00005720438,0.000006673628,0.000001957943,0.000053520307,0.00029993508,0.008228568],"category_scores_gemma":[0.007189725,0.000104564366,0.00022627143,0.000007645387,0.00009091177,0.00000267654,0.000031695174,0.00021208377,0.00074788136],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00016281768,0.00008888811,0.0007788449,0.00019587581,0.00024280231,0.0000798154,0.000018433875,6.167202e-7,0.000016137963,0.6663622,0.2543371,0.0777164],"study_design_scores_gemma":[0.001690044,0.0007884269,0.002033003,0.009127652,0.0006187197,0.000045034,0.0000038040464,0.000034729175,0.00010814205,0.012737682,0.972521,0.00029177175],"about_ca_topic_score_codex":0.000020208989,"about_ca_topic_score_gemma":0.0000044795406,"teacher_disagreement_score":0.7181839,"about_ca_system_score_codex":0.000013224876,"about_ca_system_score_gemma":0.00007847891,"threshold_uncertainty_score":0.99267805},"labels":[],"label_agreement":null},{"id":"W4238220947","doi":"10.32920/ryerson.14664381","title":"Effects of experience-based frequency learning on posterior probability judgments","year":2021,"lang":"en","type":"preprint","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University; University of Waterloo","funders":"","keywords":"Posterior probability; Conditional probability; Bayesian probability; Bayesian inference; Psychology; Inference; Statistics; Bayes' theorem; Econometrics; Computer science; Cognitive psychology; Artificial intelligence; Mathematics","score_opus":0.03257109495512512,"score_gpt":0.3358894852740631,"score_spread":0.303318390318938,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4238220947","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9930029,0.00022755157,0.000271488,0.00039540837,0.0005879666,0.0009993133,0.0000033552296,0.000105372106,0.0044066664],"genre_scores_gemma":[0.99319243,0.0000305236,0.005141887,0.0007217126,0.00008326751,0.0001828195,0.000104871346,0.000026188089,0.00051629904],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9977294,0.0001882418,0.00060079026,0.0007106312,0.000514944,0.0002559729],"domain_scores_gemma":[0.9901387,0.008476464,0.00025721797,0.00065005274,0.00023847233,0.0002390735],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00028657776,0.0002843287,0.0008470214,0.00007319857,0.000034204662,0.000028426466,0.00013902406,0.0003829798,0.0005107214],"category_scores_gemma":[0.07332313,0.00022614119,0.00036308306,0.00011105892,0.00013276163,0.000015283274,0.0002193786,0.00085713295,0.000018689878],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012570643,0.009707987,0.8944419,0.008905962,0.0004841121,0.0009798415,0.0028944432,0.0010241946,0.02307364,0.0001755009,0.00030546973,0.056749888],"study_design_scores_gemma":[0.006649777,0.0042046453,0.66117615,0.06849618,0.0006575659,0.000015134825,0.0003541342,0.00065491424,0.25620604,0.0007161005,0.000052549494,0.00081683253],"about_ca_topic_score_codex":0.00037869404,"about_ca_topic_score_gemma":0.000005591657,"teacher_disagreement_score":0.23326576,"about_ca_system_score_codex":0.00012325257,"about_ca_system_score_gemma":0.00056825916,"threshold_uncertainty_score":0.93448263},"labels":[],"label_agreement":null},{"id":"W4238475577","doi":"10.1007/978-3-319-11821-5","title":"Approach to Internal Medicine","year":2015,"lang":"en","type":"book","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta; University of Calgary","funders":"","keywords":"Psychology","score_opus":0.06633283083902951,"score_gpt":0.3672769692264565,"score_spread":0.30094413838742695,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4238475577","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00001567883,0.00050451374,0.0014062218,0.002642696,0.00056674716,0.00040819473,0.00000835703,0.00012805438,0.99431956],"genre_scores_gemma":[0.00024378127,0.000048836464,0.0028440112,0.008103392,0.0029124531,0.000017796629,0.00033391305,0.000054286684,0.9854415],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9981643,0.000015460368,0.00046029646,0.0004514596,0.0006717778,0.00023673553],"domain_scores_gemma":[0.9954157,0.0026998562,0.000088142355,0.00052170275,0.00028055135,0.0009940453],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00051383366,0.00027262562,0.0008847449,0.00015948368,0.0000132733485,0.000008345473,0.00016973045,0.00036616274,0.0013705506],"category_scores_gemma":[0.037526652,0.0001743799,0.00013476727,0.00006805198,0.00012061563,0.0000093860535,0.00011208751,0.0005429179,0.0017272661],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000942063,0.00010297307,0.00006222319,0.00003285154,0.00007895924,0.00004955361,0.00009730099,7.2524733e-7,6.105945e-7,0.0045667496,0.9857003,0.009213519],"study_design_scores_gemma":[0.0017322978,0.0010696758,0.00024528764,0.0060935775,0.00034645016,0.00018563528,0.000033122393,0.000034804496,0.0000019102495,0.0037493177,0.98627996,0.00022793145],"about_ca_topic_score_codex":0.00010361857,"about_ca_topic_score_gemma":0.0000033900415,"teacher_disagreement_score":0.03701282,"about_ca_system_score_codex":0.00022070842,"about_ca_system_score_gemma":0.00068080495,"threshold_uncertainty_score":0.99954236},"labels":[],"label_agreement":null},{"id":"W4239558931","doi":"10.1503/cmaj.191634","title":"Deconstructing the diagnostic reasoning of human versus artificial intelligence","year":2020,"lang":"en","type":"erratum","venue":"Canadian Medical Association Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Computer science; Artificial intelligence; Data science; Human intelligence","score_opus":0.03399220037504716,"score_gpt":0.33218568083504335,"score_spread":0.29819348045999616,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4239558931","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.049598802,0.011115279,0.0008047025,0.41908196,0.29352844,0.00211125,0.00037832523,0.00024314961,0.22313808],"genre_scores_gemma":[0.75118804,0.008755327,0.0014811947,0.04216229,0.17211002,0.00007695756,0.0012573611,0.00045475838,0.022514079],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9952575,0.00028406648,0.0013454233,0.00030420147,0.002185108,0.0006237202],"domain_scores_gemma":[0.9685841,0.026163602,0.0013142971,0.0002892537,0.0006724339,0.0029762967],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0025005976,0.00026350666,0.0007942944,0.00017733277,0.00045011434,0.00011125964,0.00044091226,0.0013438363,0.0030106679],"category_scores_gemma":[0.53801453,0.00020300764,0.00040453236,0.0004152039,0.00021824423,0.000040474737,0.0000491247,0.0058904435,0.00013825379],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003241676,0.00003669792,0.0060638823,0.000027206825,0.0003908123,0.0009825017,0.00028462472,0.0000017273253,6.182204e-7,0.0010352331,0.88859415,0.10255014],"study_design_scores_gemma":[0.0026361654,0.0019161198,0.020926785,0.02613506,0.0026558787,0.0020704367,0.005579542,0.0015817394,0.000045727655,0.0041128006,0.9311546,0.0011851402],"about_ca_topic_score_codex":0.002952964,"about_ca_topic_score_gemma":0.015502737,"teacher_disagreement_score":0.7015892,"about_ca_system_score_codex":0.00096344587,"about_ca_system_score_gemma":0.010583892,"threshold_uncertainty_score":0.9999526},"labels":[],"label_agreement":null},{"id":"W4242345717","doi":"10.1197/aemj.9.9.889","title":"Predictive Validity of the Global Assessment Form Used in a Final‐year Undergraduate Rotation in Emergency Medicine","year":2002,"lang":"en","type":"article","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; Sunnybrook Health Science Centre; University of Toronto; St. Michael's Hospital","funders":"","keywords":"Medicine; Physical examination; Global Assessment of Functioning; Emergency department; Family medicine; Surgery; Psychiatry","score_opus":0.12326561605463089,"score_gpt":0.4224168155497112,"score_spread":0.29915119949508034,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4242345717","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9478854,0.00229214,0.0003030809,0.031609423,0.002578385,0.0011383771,0.00002093225,0.00004987586,0.014122365],"genre_scores_gemma":[0.9769242,0.02108285,0.000055634056,0.00019647671,0.0005925745,0.00008317125,0.0000395595,0.000024576302,0.0010009743],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9947892,0.00028508593,0.0026355607,0.00054576114,0.0012148434,0.0005295566],"domain_scores_gemma":[0.99680084,0.0014163451,0.00074709964,0.0005488782,0.00021694176,0.00026992077],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001899206,0.00033757763,0.00089933083,0.00021217135,0.000058405538,2.7598944e-7,0.00036302645,0.0003601429,0.0034294785],"category_scores_gemma":[0.032370284,0.00021854287,0.00016612168,0.0020001696,0.00032129863,0.000111393456,0.00010182682,0.0011759588,0.00001837319],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009203121,0.00029783277,0.9776147,0.000096697964,0.00004504609,0.0000083216655,0.0011699698,0.00004368568,0.00019695766,0.0012009002,0.018560184,0.00067366555],"study_design_scores_gemma":[0.0054732617,0.001128801,0.9625599,0.004013362,0.00027463402,0.000011273915,0.0008223349,0.0040049404,0.000038251583,0.02106831,0.00041719255,0.00018773002],"about_ca_topic_score_codex":0.0011305845,"about_ca_topic_score_gemma":0.00038702195,"teacher_disagreement_score":0.031412948,"about_ca_system_score_codex":0.00039503296,"about_ca_system_score_gemma":0.000118289834,"threshold_uncertainty_score":0.9974815},"labels":[],"label_agreement":null},{"id":"W4242748982","doi":"10.5489/cuaj.1442","title":"Unmoderated Posters: Training and Evaluation","year":2013,"lang":"en","type":"article","venue":"Canadian Urological Association Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Training (meteorology); Computer science; Medical education; Medicine; Geography","score_opus":0.044103967223639656,"score_gpt":0.3052695921745987,"score_spread":0.2611656249509591,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4242748982","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94743484,0.00017874564,0.00002255501,0.048614766,0.00016805454,0.00019555195,0.0000035297135,0.000016436086,0.003365513],"genre_scores_gemma":[0.99029195,0.00003509971,0.00024358819,0.008912819,0.00026083915,0.000010319704,0.00001938384,0.000005907279,0.0002200627],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9988698,0.00014999816,0.0002475626,0.0001295558,0.0003020651,0.00030105177],"domain_scores_gemma":[0.9968294,0.0016510081,0.0001255517,0.00005157337,0.0004800065,0.00086246926],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0010625181,0.000076252356,0.00018269707,0.00008655327,0.00015698193,0.0001300599,0.000032558517,0.0001882055,0.0016582042],"category_scores_gemma":[0.035671383,0.000058923008,0.000051936368,0.00010083108,0.000023070343,0.00008429498,0.000005436249,0.0003693387,0.00009924947],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000020309477,0.000050846695,0.81283754,0.0000016110978,0.00011600115,0.00004039495,0.0005130301,0.00004613927,0.0001004295,0.00014087654,0.024073122,0.1620597],"study_design_scores_gemma":[0.0012292757,0.0002685284,0.98780507,0.000054890446,0.000092467206,0.00024486647,0.00017490686,0.0045711594,0.0000031321858,0.0029730515,0.0025015979,0.00008104944],"about_ca_topic_score_codex":0.0009808523,"about_ca_topic_score_gemma":0.00015058741,"teacher_disagreement_score":0.17496754,"about_ca_system_score_codex":0.00041646764,"about_ca_system_score_gemma":0.00051261566,"threshold_uncertainty_score":0.9992544},"labels":[],"label_agreement":null},{"id":"W4243422411","doi":"10.1093/med/9780190088743.001.0001","title":"The Cognitive Autopsy","year":2020,"lang":"en","type":"book","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":23,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Cognition; Covert; Root cause; Medicine; Autopsy; Root cause analysis; Disease; Common cause and special cause; Cause of death; Medical emergency; Psychology; Intensive care medicine; Psychiatry; Pathology; Forensic engineering","score_opus":0.03138802087942346,"score_gpt":0.33698417866419456,"score_spread":0.3055961577847711,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4243422411","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000011818922,0.00093579624,0.00006036372,0.008084535,0.00030351957,0.00030540477,0.000018597266,0.00012237683,0.9901576],"genre_scores_gemma":[0.0004971959,0.0005959676,0.000064330445,0.007475048,0.0009424841,0.000013509168,0.00017662841,0.000030826464,0.99020404],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9990892,0.000015273326,0.00023607117,0.00024013936,0.00026433246,0.00015501805],"domain_scores_gemma":[0.96719456,0.032186404,0.000076312244,0.00019259223,0.000115240815,0.00023491611],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00012405698,0.00015238005,0.00033734686,0.000012702256,0.000067521076,0.000023738505,0.00007435251,0.00022343606,0.0008613012],"category_scores_gemma":[0.061050355,0.00008095924,0.00020461656,0.000031182775,0.00015626458,0.0000048160477,0.00005554714,0.0005421655,0.002614618],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009115003,0.000024734076,0.00005178658,0.000013032162,0.00015520837,0.0001640447,0.000027861166,1.2012056e-8,1.6699356e-7,0.009760757,0.93038625,0.05932502],"study_design_scores_gemma":[0.0008929812,0.0003528066,0.00085087307,0.003052432,0.0005099978,0.0000341623,0.000033762935,0.000023195213,0.0000076185206,0.0045159766,0.98958784,0.00013837189],"about_ca_topic_score_codex":0.000004929025,"about_ca_topic_score_gemma":0.000003589532,"teacher_disagreement_score":0.060926296,"about_ca_system_score_codex":0.000039819744,"about_ca_system_score_gemma":0.0007711458,"threshold_uncertainty_score":0.998162},"labels":[],"label_agreement":null},{"id":"W4244378158","doi":"10.1016/j.jogc.2020.02.009","title":"Clinical Conundrums","year":2020,"lang":"en","type":"article","venue":"Journal of Obstetrics and Gynaecology Canada","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"British Columbia Centre of Excellence for Women's Health; University of British Columbia","funders":"","keywords":"Medicine; Presentation (obstetrics); Dilemma; Clinical trial; Publication; Protocol (science); Task (project management); Medical physics; Alternative medicine; Surgery; Pathology","score_opus":0.03504478168735309,"score_gpt":0.3142665289883046,"score_spread":0.2792217473009515,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4244378158","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9741134,0.004206206,0.000054860873,0.011208419,0.008292895,0.00006941678,0.0000059544477,0.000004241787,0.0020445758],"genre_scores_gemma":[0.9875762,0.00039193046,0.000614459,0.011204334,0.0001297295,2.5132036e-7,0.0000012969261,0.0000065559725,0.000075247226],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9988379,0.000046661644,0.00064332125,0.00010807124,0.00021963735,0.00014439723],"domain_scores_gemma":[0.9630202,0.035457205,0.00033924216,0.00007948584,0.00029560964,0.00080823805],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00029039834,0.00007239459,0.00047351382,0.00003841141,0.000029565483,0.000008133188,0.000068171525,0.0001117603,0.00026135537],"category_scores_gemma":[0.2516217,0.000054585445,0.000092271155,0.00017323022,0.00007429407,0.000018566408,0.000028660766,0.00044921334,0.000002105778],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000014449642,0.00007762295,0.79915035,0.000012466895,0.00010665364,0.004984883,0.000022683325,0.0000049131163,0.0000013500498,0.000109537854,0.027169975,0.16834512],"study_design_scores_gemma":[0.0037196924,0.0015458737,0.92336786,0.000021050364,0.00016534717,0.00011045926,0.0000983873,0.00011976268,0.0000065329,0.000061274994,0.070712544,0.00007123239],"about_ca_topic_score_codex":0.0014950309,"about_ca_topic_score_gemma":0.0010291592,"teacher_disagreement_score":0.2513313,"about_ca_system_score_codex":0.00011494492,"about_ca_system_score_gemma":0.0020168754,"threshold_uncertainty_score":0.75468224},"labels":[],"label_agreement":null},{"id":"W4245108909","doi":"10.7287/peerj.preprints.1859","title":"Rushing Roulette – how do learners perform routine tasks under time pressure?","year":2016,"lang":"en","type":"preprint","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Computer science; Task (project management); Adaptation (eye); Process (computing); Resource (disambiguation); Coaching; Artificial intelligence; Time management; Machine learning; Human–computer interaction; Psychology","score_opus":0.027236613081984547,"score_gpt":0.3159742843145293,"score_spread":0.28873767123254473,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4245108909","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.47019342,0.0040237834,0.0059263376,0.23743288,0.002906314,0.0023647642,0.00023032195,0.0015313516,0.27539083],"genre_scores_gemma":[0.8682028,0.00022341745,0.0011986778,0.0033063102,0.0016143359,0.000042773747,0.00021391947,0.00009526894,0.12510246],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9970364,0.00009548867,0.0005141353,0.0009915308,0.0007288483,0.0006335775],"domain_scores_gemma":[0.9919338,0.0058075027,0.0002919984,0.0012519769,0.00018608509,0.00052864186],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00094767107,0.00051939226,0.0010369447,0.0001312153,0.00009001348,0.00019357719,0.00033701034,0.0008998245,0.00493427],"category_scores_gemma":[0.02129506,0.0003364452,0.0004791431,0.00008668585,0.0001742809,0.00007102681,0.000727164,0.0015178056,0.0011027394],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009870712,0.0011612349,0.07712723,0.0007480124,0.003470794,0.00038017044,0.0006894826,0.00065684016,0.00080676045,0.0053876615,0.819041,0.089543745],"study_design_scores_gemma":[0.05422949,0.004161548,0.27239895,0.1431506,0.020166961,0.0009951055,0.0018882152,0.03981688,0.006202275,0.051190536,0.39303252,0.012766912],"about_ca_topic_score_codex":0.00010649988,"about_ca_topic_score_gemma":0.0000031748136,"teacher_disagreement_score":0.42600846,"about_ca_system_score_codex":0.00012403601,"about_ca_system_score_gemma":0.00036615354,"threshold_uncertainty_score":0.99990875},"labels":[],"label_agreement":null},{"id":"W4245809478","doi":"10.1503/cmaj.1040811","title":"References please","year":2004,"lang":"en","type":"article","venue":"Canadian Medical Association Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"University of Saskatchewan","funders":"","keywords":"sort; Feeling; Computer science; Preceptor; World Wide Web; Medical education; Multimedia; Psychology; Medicine; Information retrieval; Social psychology","score_opus":0.016678760569763744,"score_gpt":0.3033216567915787,"score_spread":0.28664289622181494,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4245809478","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.76903564,0.00031735684,0.00011900291,0.20246704,0.0012217048,0.0000977483,0.000013854024,0.00003813551,0.026689501],"genre_scores_gemma":[0.96691084,0.00022653819,0.00033817295,0.0297338,0.0018912187,0.0000024370306,0.000011815735,0.000009298643,0.00087587186],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.99802446,0.000046127003,0.00033312224,0.000114079805,0.0011198523,0.00036235063],"domain_scores_gemma":[0.99507684,0.0011587674,0.00011105367,0.00008145366,0.00029865387,0.0032732612],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011167965,0.00007676605,0.00019781642,0.00010218157,0.00014856306,0.000047676447,0.00009429998,0.00030940628,0.0051222956],"category_scores_gemma":[0.09706562,0.00006164097,0.00010537802,0.00016581266,0.000035001165,0.000043512384,0.000006347853,0.0008459861,0.00065295235],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000056375393,0.00032438556,0.4470186,0.000011166111,0.00030713264,0.007403212,0.000478491,0.000022838078,0.0000092301425,0.0051181517,0.33266252,0.20658791],"study_design_scores_gemma":[0.0077941287,0.0004913278,0.48535204,0.00213505,0.00015243057,0.003269869,0.00053122116,0.000045908535,0.0000424847,0.006075646,0.49380037,0.00030954112],"about_ca_topic_score_codex":0.004254564,"about_ca_topic_score_gemma":0.013546321,"teacher_disagreement_score":0.20627837,"about_ca_system_score_codex":0.00085314136,"about_ca_system_score_gemma":0.0067739254,"threshold_uncertainty_score":0.9988567},"labels":[],"label_agreement":null},{"id":"W4246039120","doi":"10.32920/ryerson.14664381.v1","title":"Effects of experience-based frequency learning on posterior probability judgments","year":2021,"lang":"en","type":"preprint","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University; University of Waterloo","funders":"","keywords":"Posterior probability; Conditional probability; Bayesian probability; Bayesian inference; Psychology; Statistics; Inference; Bayes' theorem; Statistical inference; Econometrics; Computer science; Artificial intelligence; Mathematics","score_opus":0.03257109495512512,"score_gpt":0.3358894852740631,"score_spread":0.303318390318938,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4246039120","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9930029,0.00022755157,0.000271488,0.00039540837,0.0005879666,0.0009993133,0.0000033552296,0.000105372106,0.0044066664],"genre_scores_gemma":[0.99319243,0.0000305236,0.005141887,0.0007217126,0.00008326751,0.0001828195,0.000104871346,0.000026188089,0.00051629904],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9977294,0.0001882418,0.00060079026,0.0007106312,0.000514944,0.0002559729],"domain_scores_gemma":[0.9901387,0.008476464,0.00025721797,0.00065005274,0.00023847233,0.0002390735],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00028657776,0.0002843287,0.0008470214,0.00007319857,0.000034204662,0.000028426466,0.00013902406,0.0003829798,0.0005107214],"category_scores_gemma":[0.07332313,0.00022614119,0.00036308306,0.00011105892,0.00013276163,0.000015283274,0.0002193786,0.00085713295,0.000018689878],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012570643,0.009707987,0.8944419,0.008905962,0.0004841121,0.0009798415,0.0028944432,0.0010241946,0.02307364,0.0001755009,0.00030546973,0.056749888],"study_design_scores_gemma":[0.006649777,0.0042046453,0.66117615,0.06849618,0.0006575659,0.000015134825,0.0003541342,0.00065491424,0.25620604,0.0007161005,0.000052549494,0.00081683253],"about_ca_topic_score_codex":0.00037869404,"about_ca_topic_score_gemma":0.000005591657,"teacher_disagreement_score":0.23326576,"about_ca_system_score_codex":0.00012325257,"about_ca_system_score_gemma":0.00056825916,"threshold_uncertainty_score":0.93448263},"labels":[],"label_agreement":null},{"id":"W4248778216","doi":"10.1007/978-3-030-71363-8_9","title":"Bonfire Red Titles","year":2021,"lang":"en","type":"book-chapter","venue":"Innovation and change in professional education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Paratext; Subtext; Context (archaeology); Computer science; World Wide Web; Literature; History; Art","score_opus":0.0880703652698108,"score_gpt":0.4150996395621706,"score_spread":0.32702927429235984,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4248778216","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.11030508,0.00742355,0.0000039823885,0.059257727,0.009196681,0.001481543,0.00003947059,0.00008907262,0.8122029],"genre_scores_gemma":[0.0694801,0.00070898887,0.00021289966,0.006903866,0.0011680013,0.00011265072,0.0022125414,0.0000307388,0.9191702],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.9990831,0.000011493048,0.0003636626,0.00024652094,0.00020825061,0.0000869842],"domain_scores_gemma":[0.9988671,0.0003917024,0.00016743873,0.00014656493,0.00037739772,0.000049793656],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00017381078,0.00013272325,0.0002325934,0.00026942175,0.00003489176,0.000011433699,0.000026237187,0.00033580593,0.0012260088],"category_scores_gemma":[0.002089225,0.00011847062,0.000029421879,0.00013519955,0.00003844127,0.00003888815,0.000031647403,0.0003597724,0.000031284562],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001017219,0.0005213091,0.009190795,0.00026788554,0.000035213598,0.000025680336,0.00078921014,2.2571163e-8,0.00008144841,0.4481362,0.105201535,0.43564898],"study_design_scores_gemma":[0.0012666051,0.00013533451,0.21492435,0.036042165,0.000093531,0.000081379345,0.0003929472,0.000027446888,0.000060474256,0.038615417,0.707905,0.0004553354],"about_ca_topic_score_codex":0.00003407047,"about_ca_topic_score_gemma":0.000020128407,"teacher_disagreement_score":0.6027035,"about_ca_system_score_codex":0.00007023219,"about_ca_system_score_gemma":0.00069212966,"threshold_uncertainty_score":0.999687},"labels":[],"label_agreement":null},{"id":"W4250353735","doi":"10.1016/j.jogc.2020.03.003","title":"Raisonnement clinique","year":2020,"lang":"fr","type":"editorial","venue":"Journal of Obstetrics and Gynaecology Canada","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"British Columbia Centre of Excellence for Women's Health; University of British Columbia","funders":"","keywords":"Medicine; Humanities; Gynecology; Philosophy","score_opus":0.01716486717306511,"score_gpt":0.2900167944099342,"score_spread":0.27285192723686913,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4250353735","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008172245,0.050998453,0.0000478522,0.005479603,0.9334889,0.00023092065,0.00018331387,0.0000046592736,0.0013940636],"genre_scores_gemma":[0.36442795,0.030158339,0.0021598043,0.0076587587,0.5918401,0.000012475435,0.00019038263,0.00016659618,0.0033855883],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99527174,0.00026729456,0.0021723679,0.0004048055,0.0012250588,0.0006587097],"domain_scores_gemma":[0.6668706,0.32658538,0.002439205,0.00030337484,0.0016830597,0.0021183626],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011012715,0.00044570543,0.0019596058,0.00022603077,0.0001345302,0.000048228616,0.0003454571,0.0012510393,0.0016253301],"category_scores_gemma":[0.7082711,0.00038378846,0.00032508525,0.0004456489,0.00023015236,0.000045188775,0.00015419997,0.003124411,0.000011210497],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000027827642,0.0002275751,0.027421076,0.00018094106,0.00061497197,0.019554095,0.00007186497,0.000017598539,6.2009667e-7,0.00013284007,0.8177092,0.13404137],"study_design_scores_gemma":[0.0061540687,0.0034045754,0.050986018,0.0005731137,0.0013158935,0.00007141334,0.00021428737,0.00005529894,0.0000056645035,0.00014136931,0.93676335,0.00031493165],"about_ca_topic_score_codex":0.02650778,"about_ca_topic_score_gemma":0.024250431,"teacher_disagreement_score":0.7071698,"about_ca_system_score_codex":0.0031803204,"about_ca_system_score_gemma":0.02838082,"threshold_uncertainty_score":0.9998614},"labels":[],"label_agreement":null},{"id":"W4250431025","doi":"10.1197/aemj.9.11.1184","title":"Achieving Quality in Clinical Decision Making: Cognitive Strategies and Detection of Bias","year":2002,"lang":"en","type":"article","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":588,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Heuristics; Cornerstone; Cognition; Medicine; Clinical decision making; Quality (philosophy); Emergency department; Cognitive bias; Resource (disambiguation); Cognitive load; Principal (computer security); Quality management; Cognitive psychology; Operations management; Computer science; Intensive care medicine; Psychology; Nursing; Psychiatry; Computer security; Management system","score_opus":0.20829777717127104,"score_gpt":0.4949761063994055,"score_spread":0.2866783292281344,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4250431025","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98465335,0.010234818,0.0009954291,0.00065607147,0.00067988224,0.00027133516,0.0000025024244,0.000036462152,0.0024701343],"genre_scores_gemma":[0.97413576,0.024994668,0.000098784796,0.00021063947,0.00044582324,0.0000090998665,0.0000046393,0.000014939829,0.000085638225],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99645495,0.00025027725,0.0022208898,0.00039697907,0.00043358895,0.0002433046],"domain_scores_gemma":[0.982451,0.016498141,0.00049906137,0.00020224942,0.00015662039,0.00019291209],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00319583,0.00017195927,0.00071367854,0.0001803801,0.000036279904,6.4183774e-7,0.00008064549,0.00036778676,0.0006587207],"category_scores_gemma":[0.17284155,0.00012954454,0.00009793556,0.00045239166,0.00031929262,0.00010996584,0.00005304655,0.00096825627,0.000013617313],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00032049904,0.00014905457,0.86585575,0.000072201336,0.000034401703,0.000010431985,0.00065363606,0.0000012065472,0.0009356578,0.00018056095,0.000384076,0.13140252],"study_design_scores_gemma":[0.0026485967,0.0008803215,0.98566234,0.006722609,0.00015977105,0.000020570575,0.0019464444,0.00038535715,0.00006957038,0.001258618,0.000119545344,0.00012624555],"about_ca_topic_score_codex":0.00013605042,"about_ca_topic_score_gemma":0.00004218319,"teacher_disagreement_score":0.16964573,"about_ca_system_score_codex":0.000021218311,"about_ca_system_score_gemma":0.000032246204,"threshold_uncertainty_score":0.83412594},"labels":[],"label_agreement":null},{"id":"W4250569317","doi":"10.21203/rs.2.22528/v2","title":"Measuring Clinical Uncertainty and Equipoise by Applying the Agreement Study Methodology to Patient Management Decisions","year":2020,"lang":"en","type":"preprint","venue":"Research Square","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre Hospitalier de l’Université de Montréal; Université de Montréal; Alberta Hospital Edmonton; University of Ottawa","funders":"","keywords":"Agreement; Clinical equipoise; Management science; Medical physics; Psychology; Medicine; Clinical trial; Economics; Philosophy; Internal medicine","score_opus":0.5379909112707619,"score_gpt":0.5598600217313185,"score_spread":0.021869110460556618,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4250569317","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93746966,0.0033327786,0.0049422584,0.027376864,0.0008468215,0.023080546,0.00013371072,0.00013304743,0.0026843343],"genre_scores_gemma":[0.9859196,0.002102399,0.0057559735,0.0012640969,0.00037606043,0.004182371,0.000083929386,0.00005251073,0.0002630917],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99003506,0.003962133,0.001220753,0.0015202857,0.0024515889,0.0008102118],"domain_scores_gemma":[0.9284902,0.0676854,0.00018009986,0.001697196,0.0006332693,0.0013138023],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.014438292,0.00034589274,0.0010194433,0.00023656715,0.00037286474,0.0001452593,0.00056764897,0.00028477397,0.00010412424],"category_scores_gemma":[0.14559558,0.00022688179,0.00028900118,0.00048233278,0.00026495164,0.000014227478,0.0062729637,0.0028097485,0.00014493664],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017043422,0.0028497986,0.04887025,0.00035979543,0.0011121441,0.0008199005,0.0027558818,0.00059878855,0.000018312092,0.00030290478,0.038800012,0.90180784],"study_design_scores_gemma":[0.018229155,0.0345164,0.62050354,0.043861825,0.0030291032,0.000060560807,0.0933975,0.00465912,0.0001169419,0.018154334,0.16106789,0.0024036507],"about_ca_topic_score_codex":0.00061847345,"about_ca_topic_score_gemma":0.00008441925,"teacher_disagreement_score":0.8994042,"about_ca_system_score_codex":0.00026426447,"about_ca_system_score_gemma":0.0002719685,"threshold_uncertainty_score":0.9994908},"labels":[],"label_agreement":null},{"id":"W4250849419","doi":"10.33588/fem.82.154","title":"Lo que todo profesor necesita saber sobre el razonamiento clínico","year":2005,"lang":"es","type":"article","venue":"Educación Médica","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Computer science","score_opus":0.018852578767989212,"score_gpt":0.3495219426590169,"score_spread":0.3306693638910277,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4250849419","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7228032,0.014937545,0.000087173314,0.13382077,0.0062700235,0.0016700694,0.00014273821,0.0004055199,0.119862966],"genre_scores_gemma":[0.92795473,0.0022840535,0.0012774417,0.010900302,0.010635193,0.00011229921,0.0001823695,0.00015223566,0.046501372],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99462503,0.0002988218,0.0013025879,0.0012553982,0.0010719033,0.0014462458],"domain_scores_gemma":[0.98726666,0.008955533,0.0004058175,0.0015068315,0.0003277536,0.0015373772],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0011339836,0.0007313475,0.0011947288,0.0002495015,0.00032721963,0.00024359963,0.0006009383,0.00076686154,0.0077883117],"category_scores_gemma":[0.03866404,0.00064558815,0.00057121215,0.0006184469,0.00049205247,0.00023463281,0.00028421058,0.00134917,0.013328347],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005292004,0.0056462185,0.0740345,0.00024804144,0.00042260814,0.0001684425,0.0014179408,0.000018524785,0.00026200176,0.008125454,0.7194041,0.18972297],"study_design_scores_gemma":[0.0058135106,0.0010652897,0.15154755,0.007921428,0.0010845725,0.00013642709,0.0011254108,0.0004961045,0.0007325391,0.0007191126,0.82822865,0.0011294215],"about_ca_topic_score_codex":0.00033560023,"about_ca_topic_score_gemma":0.00004515887,"teacher_disagreement_score":0.20515154,"about_ca_system_score_codex":0.00035745715,"about_ca_system_score_gemma":0.0014905,"threshold_uncertainty_score":0.9995995},"labels":[],"label_agreement":null},{"id":"W4250901566","doi":"10.9778/cmajo.20200028","title":"Correction: Identifying Ontario geographic regions to assess adults who present to hospital with laboratory-defined conditions: a descriptive study","year":2020,"lang":"en","type":"erratum","venue":"CMAJ Open","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Descriptive statistics; Descriptive research; Geography; Medicine; Statistics; Mathematics","score_opus":0.05896783344863329,"score_gpt":0.3559681557802524,"score_spread":0.29700032233161916,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4250901566","genre_codex":"empirical","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.54964143,0.0016609164,0.002155562,0.051104277,0.1794528,0.07922052,0.0018317847,0.0014660468,0.13346663],"genre_scores_gemma":[0.31434646,0.00020667465,0.0023629314,0.009370659,0.0060086655,0.009858191,0.0042073405,0.00048339064,0.6531557],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99617773,0.00019144196,0.00072474655,0.0014596359,0.0009107013,0.0005357293],"domain_scores_gemma":[0.9942594,0.0018756944,0.00034241143,0.0011098403,0.0011490171,0.0012636065],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0003837308,0.0005948703,0.001455772,0.00032933676,0.00034717348,0.0006215707,0.000744343,0.00038404498,0.00044190354],"category_scores_gemma":[0.017365389,0.00051155436,0.00019890608,0.001403373,0.00010343198,0.00021530071,0.00079190737,0.001809233,0.00036139545],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000828573,0.0013539743,0.085960686,0.000045709312,0.00062412716,0.0008762153,0.008954202,0.0000081073185,4.624688e-7,0.000026134558,0.90112156,0.00020027209],"study_design_scores_gemma":[0.004732874,0.011397731,0.65936404,0.02571944,0.001449774,0.000037364895,0.012853397,0.000014029714,0.000004517143,0.00003664573,0.28348902,0.0009011821],"about_ca_topic_score_codex":0.10464136,"about_ca_topic_score_gemma":0.1875009,"teacher_disagreement_score":0.6176325,"about_ca_system_score_codex":0.00062394654,"about_ca_system_score_gemma":0.0030484712,"threshold_uncertainty_score":0.9997336},"labels":[],"label_agreement":null},{"id":"W4250936097","doi":"10.22374/cjgim.v11i1.110","title":"Developing, Maintaining, and Teaching Clinical Diagnostic Expertise","year":2016,"lang":"en","type":"article","venue":"Canadian Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"University of Alberta; M.S.I. Foundation","keywords":"Scripting language; Process (computing); Identification (biology); Clinical Practice; Key (lock); Abstraction; Medicine; Computer science; Knowledge management; Artificial intelligence; Nursing","score_opus":0.05445950420105238,"score_gpt":0.38618587382271824,"score_spread":0.33172636962166585,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4250936097","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93237895,0.0026343702,0.0072024744,0.052388612,0.0021548378,0.00010973206,0.0000046911196,0.000011267809,0.0031150626],"genre_scores_gemma":[0.9834265,0.0008101009,0.003105472,0.0064995647,0.0035133802,0.0000021715139,0.0000016876292,0.000027110078,0.0026139813],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9977621,0.00015970152,0.0011603974,0.00023017802,0.0002835428,0.0004041228],"domain_scores_gemma":[0.9887133,0.008059404,0.0003843587,0.00020205148,0.00029601564,0.0023448383],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0016776418,0.0001947263,0.0006847933,0.00029338856,0.00008364619,0.000021783899,0.00018573426,0.00013848163,0.00046705524],"category_scores_gemma":[0.15067278,0.00010681337,0.00014617175,0.00005072695,0.0005588212,0.000091306974,0.00002705631,0.0005330286,0.000018699971],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00017382059,0.000034269437,0.6072344,0.000010761357,0.00019802492,0.0040653297,0.0004201137,5.412155e-7,0.000274127,0.0055373814,0.07095184,0.31109938],"study_design_scores_gemma":[0.013821399,0.0076164585,0.76602167,0.040932693,0.0005517248,0.01388996,0.00045056109,0.00007676903,0.00040712956,0.0042993263,0.1514088,0.00052349054],"about_ca_topic_score_codex":0.0048899963,"about_ca_topic_score_gemma":0.0017997889,"teacher_disagreement_score":0.3105759,"about_ca_system_score_codex":0.00025606298,"about_ca_system_score_gemma":0.0010853164,"threshold_uncertainty_score":0.85648143},"labels":[],"label_agreement":null},{"id":"W4250954394","doi":"10.1093/oso/9780198716228.003.0008","title":"Confusion","year":2015,"lang":"en","type":"book-chapter","venue":"Oxford University Press eBooks","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Delirium; Psychology; Dementia; Confusion; Cognition; Cognitive impairment; Consciousness; Minimally conscious state; Psychiatry; Cognitive psychology; Medicine; Neuroscience","score_opus":0.0637897457134831,"score_gpt":0.2810875064860178,"score_spread":0.2172977607725347,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4250954394","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00018383046,0.00016550343,0.00004383985,0.00007421788,0.0002188245,0.000314264,0.0000710286,0.00013552235,0.99879295],"genre_scores_gemma":[0.0007514852,0.0003589651,0.00020492643,0.0001679414,0.0001912326,2.415892e-7,0.000114771334,0.000037355745,0.99817306],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.9989235,0.000013228116,0.00016782635,0.00035392164,0.00034904783,0.00019249813],"domain_scores_gemma":[0.997968,0.00076963915,0.00014370382,0.00045882602,0.00024190798,0.00041787705],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00011718314,0.00024390937,0.000490379,0.00009168045,0.000062064944,0.00001180783,0.00015589595,0.0005096658,0.00020771335],"category_scores_gemma":[0.0007267542,0.00024291383,0.00022372957,0.0000040004506,0.00017622882,0.000017883609,0.00021794987,0.00047721245,0.000012015995],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008146321,0.00005611301,0.0000356303,0.000064902524,0.00024830576,0.0015907051,0.000098270844,0.0000017057534,0.000007735828,0.85396236,0.10486971,0.038249914],"study_design_scores_gemma":[0.0016944531,0.00026100499,0.000025238349,0.0015875774,0.00053526904,0.000035959565,0.000012781937,0.000016407994,0.0000144028545,0.0008256684,0.99475914,0.00023208436],"about_ca_topic_score_codex":0.000045511246,"about_ca_topic_score_gemma":0.0000044984495,"teacher_disagreement_score":0.8898894,"about_ca_system_score_codex":0.00013810805,"about_ca_system_score_gemma":0.0002734094,"threshold_uncertainty_score":0.9905739},"labels":[],"label_agreement":null},{"id":"W4251271678","doi":"10.1007/s10459-017-9780-3","title":"Why?","year":2017,"lang":"en","type":"editorial","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Medical education; Psychology","score_opus":0.026938363232125593,"score_gpt":0.48250777643940723,"score_spread":0.4555694132072816,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4251271678","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00013019405,0.037174515,0.000022087637,0.007860295,0.94014907,0.00043812045,0.0000122877655,0.000030641906,0.014182775],"genre_scores_gemma":[0.00084747514,0.076845676,0.0037643625,0.00376791,0.91094166,0.00014932659,0.00029450055,0.000022097309,0.0033669956],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99754417,0.00006903994,0.0005792556,0.00060296356,0.0008023879,0.00040215597],"domain_scores_gemma":[0.98692834,0.011208264,0.00076890225,0.00056416244,0.00022742324,0.00030292245],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0017746917,0.00017044008,0.00049260445,0.00021148019,0.0003041928,0.00006775424,0.0003595984,0.00027927785,0.000050771574],"category_scores_gemma":[0.12122465,0.00014245578,0.00006216122,0.00024322356,0.00044851596,0.00028479178,0.000037813552,0.00060293474,0.000038115126],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000013443409,0.00016595532,0.0029402508,0.0002378509,0.0000010594756,0.0000011135368,0.000055361626,0.000007836564,3.8492757e-8,0.00011868677,0.8802195,0.11623889],"study_design_scores_gemma":[0.00026330072,0.00042160103,0.0013327659,0.010206858,0.000008155484,0.000002931788,0.00007343853,0.000019823618,7.9410006e-7,0.0020352441,0.985513,0.00012213571],"about_ca_topic_score_codex":0.0019541956,"about_ca_topic_score_gemma":0.00070332125,"teacher_disagreement_score":0.11944996,"about_ca_system_score_codex":0.0002874305,"about_ca_system_score_gemma":0.0109725315,"threshold_uncertainty_score":0.99463433},"labels":[],"label_agreement":null},{"id":"W4251601374","doi":"10.1002/9781118445112.stat05252","title":"Predictive Values","year":2014,"lang":"en","type":"other","venue":"Wiley StatsRef: Statistics Reference Online","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Predictive value; Positive predicative value; Sensitivity (control systems); Test (biology); Diagnostic test; Statistics; Medicine; Mathematics; Internal medicine; Biology; Pediatrics; Engineering; Ecology","score_opus":0.038205087488740457,"score_gpt":0.3625229182905345,"score_spread":0.32431783080179405,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4251601374","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00019376834,0.0047065173,0.20669895,0.0006802299,0.0022852959,0.0020590313,0.099470206,0.0017199845,0.682186],"genre_scores_gemma":[0.0010325216,0.01062191,0.17425646,0.0013016545,0.0021419132,0.000061431456,0.030359318,0.0010790918,0.7791457],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99666387,0.00013493313,0.0007729637,0.0009106933,0.00088333566,0.0006342219],"domain_scores_gemma":[0.99270666,0.004727992,0.0005671793,0.0009934169,0.0003297131,0.0006750642],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00024454147,0.00065844133,0.0013663536,0.00030888678,0.000063108855,0.00004121581,0.0002804638,0.0006999425,0.0069440533],"category_scores_gemma":[0.021397583,0.00054484134,0.0001271541,0.00020841134,0.00043990725,0.000019242396,0.00012894523,0.0010069876,0.0011705509],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015606305,0.00068817893,0.00082042435,0.00022125391,0.00029715113,0.00016299465,0.00003654995,0.000003666691,0.000003818496,0.0036911133,0.9601278,0.033791],"study_design_scores_gemma":[0.0033319418,0.0025385306,0.0045361193,0.016967466,0.0013780982,0.00005651513,0.00008020513,0.0018101276,0.000007019937,0.008446508,0.95982486,0.0010226261],"about_ca_topic_score_codex":0.00040243447,"about_ca_topic_score_gemma":0.00023733337,"teacher_disagreement_score":0.09695969,"about_ca_system_score_codex":0.000113937705,"about_ca_system_score_gemma":0.0005780398,"threshold_uncertainty_score":0.9997003},"labels":[],"label_agreement":null},{"id":"W4253417235","doi":"10.1002/9781444320282.ch21","title":"Quantitative Research Methods in Medical Education","year":2010,"lang":"en","type":"other","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Cronbach's alpha; Epistemology; Psychology; Psychometrics; Philosophy; Clinical psychology","score_opus":0.17610675510973506,"score_gpt":0.6407240440998964,"score_spread":0.4646172889901613,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4253417235","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00011097764,0.0012013866,0.00058796356,0.0050685443,0.0007642877,0.00036271618,0.0000018127514,0.0000734823,0.99182886],"genre_scores_gemma":[0.00005792802,0.0006143498,0.16137318,0.0010717334,0.0005035126,0.00005850114,0.000052439664,0.00015640036,0.83611196],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.998096,0.00038645312,0.00027229416,0.00033497735,0.0006692053,0.00024105703],"domain_scores_gemma":[0.956548,0.0415075,0.000113948365,0.0007899796,0.00026174245,0.00077887066],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0034169925,0.00011804938,0.00040894593,0.0005377167,0.000012631671,0.000009185277,0.00012814632,0.00092955166,0.054852832],"category_scores_gemma":[0.3331322,0.00008574391,0.000064722175,0.00036106768,0.00024239963,0.0000061164837,0.000057235793,0.0017359846,0.0006908761],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000027300865,0.0005477583,0.0009376724,0.00003043179,0.000016996382,0.00002601111,0.00003607024,2.8876952e-9,0.00001079464,0.011504019,0.81413335,0.17272958],"study_design_scores_gemma":[0.0007898133,0.0002754026,0.0026019688,0.008310984,0.000027016857,0.00002387071,0.00019769833,0.00006616156,0.000032214353,0.001990082,0.9855453,0.00013952248],"about_ca_topic_score_codex":0.0036840593,"about_ca_topic_score_gemma":0.0022395893,"teacher_disagreement_score":0.32971522,"about_ca_system_score_codex":0.000038978746,"about_ca_system_score_gemma":0.0026393957,"threshold_uncertainty_score":0.9460112},"labels":[],"label_agreement":null},{"id":"W4253806136","doi":"10.7326/acpjc-2005-142-3-a08","title":"Does clinical experience make up for failure to keep up to date?","year":2005,"lang":"en","type":"article","venue":"ACP Journal Club","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Intensive care medicine","score_opus":0.05915508095768181,"score_gpt":0.44567911875350796,"score_spread":0.38652403779582617,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4253806136","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9257535,0.00011469113,0.0036718585,0.06632943,0.0031663636,0.0004833718,0.000015070491,0.00005125367,0.00041448307],"genre_scores_gemma":[0.9378452,0.00017880663,0.02080131,0.02581608,0.0053275456,0.0000497262,0.000008853038,0.000037993705,0.009934497],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9977568,0.000061711115,0.0008881785,0.00038673886,0.00040838475,0.0004982098],"domain_scores_gemma":[0.992276,0.005525095,0.0001463,0.00041360207,0.00029275604,0.0013462532],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0011270125,0.0001852192,0.0005066456,0.00007339889,0.00016172092,0.00009012109,0.0002897372,0.00014700317,0.00050964317],"category_scores_gemma":[0.06434763,0.0001083238,0.00030079705,0.00017758766,0.00006262025,0.000079611986,0.000105845436,0.0005271705,0.00044024215],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019636948,0.00050612097,0.055848707,0.00001854687,0.0001400922,0.00018986536,0.003454358,0.00005591972,0.00052545394,0.00024950053,0.40413463,0.5329131],"study_design_scores_gemma":[0.0038379498,0.0017486772,0.027583646,0.0011793675,0.000135419,0.00044739238,0.001545585,0.0000931047,0.0013622482,0.0003610944,0.9613993,0.00030617783],"about_ca_topic_score_codex":0.00000774038,"about_ca_topic_score_gemma":0.00006326324,"teacher_disagreement_score":0.5572647,"about_ca_system_score_codex":0.000049687325,"about_ca_system_score_gemma":0.00015705262,"threshold_uncertainty_score":0.9435338},"labels":[],"label_agreement":null},{"id":"W4253868507","doi":"10.21203/rs.2.130/v2","title":"Measuring Clinical Uncertainty and Equipoise by Applying the Agreement Study Methodology to Patient Management Decisions","year":2019,"lang":"en","type":"preprint","venue":"Research Square","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre Hospitalier de l’Université de Montréal; Université de Montréal; University of Alberta Hospital","funders":"","keywords":"Clinical equipoise; Agreement; Management science; Psychology; Medical physics; Medicine; Clinical trial; Economics; Internal medicine","score_opus":0.4614602333211095,"score_gpt":0.5527308301765036,"score_spread":0.09127059685539407,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4253868507","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96588826,0.002712308,0.002642525,0.005692608,0.00090182637,0.019324323,0.00008289077,0.00006730582,0.0026879697],"genre_scores_gemma":[0.9886592,0.0022328722,0.0035351985,0.0006857372,0.0002730379,0.0035947403,0.000077226134,0.000051574334,0.00089040335],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9900436,0.0039121266,0.0011978693,0.0014655543,0.002488956,0.0008919136],"domain_scores_gemma":[0.91788256,0.07807396,0.00019086893,0.0022340147,0.0007301494,0.00088846916],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.02003097,0.00035056673,0.0010379049,0.00032142285,0.00032288083,0.00014273738,0.00057272887,0.00033711974,0.000119644974],"category_scores_gemma":[0.104536474,0.00022475896,0.00029112468,0.0003897812,0.0002442258,0.000016333419,0.0057483735,0.0026268244,0.00023980503],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001472616,0.0034199657,0.12158939,0.00038373857,0.0010500066,0.00035986598,0.0019027377,0.0019155957,0.0000102405875,0.00021562907,0.030813513,0.8368667],"study_design_scores_gemma":[0.01568233,0.024562754,0.7235496,0.04436366,0.002064477,0.000050404233,0.06287027,0.0037517378,0.00006609903,0.0067091817,0.11442165,0.0019078424],"about_ca_topic_score_codex":0.0007393427,"about_ca_topic_score_gemma":0.00008955321,"teacher_disagreement_score":0.83495885,"about_ca_system_score_codex":0.0003167585,"about_ca_system_score_gemma":0.00029285945,"threshold_uncertainty_score":0.99967414},"labels":[],"label_agreement":null},{"id":"W4280567834","doi":"10.1186/s12909-022-03410-x","title":"The effectiveness of using virtual patient educational tools to improve medical students’ clinical reasoning skills: a systematic review","year":2022,"lang":"en","type":"review","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":135,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"School for Public Health Research; University College London; Health Foundation","keywords":"Medical education; Virtual patient; Educational measurement; MEDLINE; Psychology; Medicine; Curriculum; Pedagogy","score_opus":0.062375974357238885,"score_gpt":0.4892368153921224,"score_spread":0.4268608410348835,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4280567834","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005934975,0.9852992,0.0001920966,0.00041349206,0.0053041275,0.007949343,0.000028344193,0.00003536485,0.00018453077],"genre_scores_gemma":[0.00006495406,0.9913359,0.00045802278,0.0023148577,0.0010238022,0.0039773104,0.00051605573,0.000086106025,0.00022294754],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.98039466,0.007596466,0.0046914774,0.0009919622,0.0058186958,0.0005067253],"domain_scores_gemma":[0.5124017,0.47607028,0.003794728,0.0030280282,0.000979578,0.0037256724],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.027364511,0.000565497,0.0046678423,0.0001846943,0.00024837683,0.00007786089,0.0013702553,0.00066602626,0.0014348339],"category_scores_gemma":[0.838064,0.00034846057,0.00142124,0.0009277428,0.00031817777,0.000055667282,0.00070335873,0.0016224924,0.00009976538],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000071792,0.0035372025,0.00038791125,0.48653713,0.00041815484,0.00000899746,0.000036196307,2.4273476e-7,1.1056717e-8,0.00071481976,0.002340854,0.5059467],"study_design_scores_gemma":[0.00054291915,0.0006133676,0.00030879967,0.9295971,0.004945412,0.00031761522,0.000106931606,0.000021051943,8.808862e-8,0.000028667326,0.06320335,0.00031469687],"about_ca_topic_score_codex":0.00006104389,"about_ca_topic_score_gemma":0.000008374906,"teacher_disagreement_score":0.8106995,"about_ca_system_score_codex":0.0007844289,"about_ca_system_score_gemma":0.0435814,"threshold_uncertainty_score":0.99989676},"labels":[],"label_agreement":null},{"id":"W4280635242","doi":"10.1016/j.jogc.2022.02.047","title":"Diagnostic Error in Surgery: Analysis of a Comprehensive National Database of Closed Medicolegal Cases","year":2022,"lang":"en","type":"article","venue":"Journal of Obstetrics and Gynaecology Canada","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Canadian Medical Protective Association","funders":"","keywords":"Medicine; Harm; National database; MEDLINE; Surgery; Database","score_opus":0.040121088133694124,"score_gpt":0.31248626453690714,"score_spread":0.272365176403213,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4280635242","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9932796,0.0038386215,0.00000427348,0.00048137776,0.0018376665,0.00008224524,0.0004059488,8.756049e-7,0.000069442336],"genre_scores_gemma":[0.99910605,0.00019793416,0.000101053025,0.0004884314,0.000012934287,0.0000037919542,0.000074228614,0.0000050330605,0.000010569005],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9979735,0.00014566192,0.0008925806,0.000115910894,0.0007273561,0.00014500375],"domain_scores_gemma":[0.77851504,0.21986279,0.0007372144,0.00010160844,0.0006048153,0.00017854762],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0006323865,0.00008362286,0.00086282525,0.000945842,0.000038285092,0.0000026422124,0.00008187533,0.00004663751,0.00051220536],"category_scores_gemma":[0.26815355,0.00007436834,0.00014713382,0.0015113936,0.000093883194,0.00002920461,0.00006484887,0.00031779887,3.6540712e-8],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000048289636,0.0005308897,0.97704923,0.000045235905,0.000852569,0.00807662,0.000059748585,0.0017167605,0.000011098875,0.00009808292,0.0046127425,0.006898735],"study_design_scores_gemma":[0.0018100734,0.00043078902,0.9945625,0.000043391607,0.00088234304,0.000111226414,0.00056948146,0.0005702569,0.000013582125,0.000019801404,0.00092607073,0.000060449853],"about_ca_topic_score_codex":0.0255987,"about_ca_topic_score_gemma":0.024850978,"teacher_disagreement_score":0.26752117,"about_ca_system_score_codex":0.0004688619,"about_ca_system_score_gemma":0.004012433,"threshold_uncertainty_score":0.9929429},"labels":[],"label_agreement":null},{"id":"W4281260576","doi":"10.1186/s12909-022-03471-y","title":"Embracing uncertainty: medical student perceptions of a pediatric bootcamp developed in response to mandated changes during the pandemic","year":2022,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Rigour; Medical education; Thematic analysis; Medicine; Preparedness; Credibility; Qualitative research; Focus group","score_opus":0.0320326641349985,"score_gpt":0.407586663195007,"score_spread":0.3755539990600085,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4281260576","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98379767,0.0002440901,0.00009474397,0.014538206,0.000630092,0.0005292189,0.000005499788,0.000041016025,0.00011949009],"genre_scores_gemma":[0.9949114,0.00023338413,0.00016055252,0.0029910577,0.00048908446,0.00045641468,0.000051460756,0.000016055596,0.0006906352],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9966366,0.0006504635,0.0005619995,0.00030784443,0.0015763072,0.00026672977],"domain_scores_gemma":[0.98845166,0.010308215,0.00013673343,0.00035367697,0.00012254578,0.0006271769],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.003898954,0.00011916086,0.00029954067,0.00028503453,0.00016036247,0.000008791988,0.0002756696,0.00011516628,0.0031932062],"category_scores_gemma":[0.14124423,0.000087902976,0.00006366353,0.00086370984,0.00007487884,0.00001679168,0.00023967263,0.00057250756,0.000024665975],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0026967684,0.002277393,0.96274644,0.00013129158,0.000030844687,0.000032469405,0.008043179,0.0001715682,0.00023917547,0.000072320625,0.0034896294,0.020068908],"study_design_scores_gemma":[0.0017471712,0.00023900098,0.9894522,0.0005833373,0.00006158188,0.0000952309,0.0066751568,0.00030040502,0.0000051409797,0.00003940878,0.0006952967,0.000106076666],"about_ca_topic_score_codex":0.00025155954,"about_ca_topic_score_gemma":0.0008148332,"teacher_disagreement_score":0.13734527,"about_ca_system_score_codex":0.00033918954,"about_ca_system_score_gemma":0.0067261965,"threshold_uncertainty_score":0.99890476},"labels":[],"label_agreement":null},{"id":"W4281492908","doi":"10.1097/acm.0000000000004758","title":"Development and Pilot Testing of a Data-Rich Clinical Reasoning Training and Assessment Tool","year":2022,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Formative assessment; Medical education; Computer science; Workload; Summative assessment; Psychology; Medicine; Mathematics education","score_opus":0.3150575699023435,"score_gpt":0.4788030286876891,"score_spread":0.16374545878534558,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4281492908","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9923203,0.002176109,0.00031580866,0.0026185678,0.0002082899,0.00031701056,0.000009771002,0.000052438285,0.0019816775],"genre_scores_gemma":[0.9518253,0.00056824397,0.0448064,0.001986384,0.00040671197,0.000032194555,0.000109306944,0.000023023293,0.00024243614],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9971226,0.00015773198,0.0012192004,0.00057381863,0.0006339174,0.00029277525],"domain_scores_gemma":[0.9819313,0.016919773,0.00037191057,0.00039688614,0.00007018221,0.0003099355],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0066205314,0.00017026765,0.0007593665,0.000102853905,0.00017615063,0.000003545277,0.00020463008,0.00009238309,0.00011647947],"category_scores_gemma":[0.0671001,0.00013596808,0.000015205567,0.0002977036,0.000300818,0.000062713574,0.0006754921,0.0014406439,9.34713e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015280086,0.00009409039,0.4668455,0.00008931738,0.00013548638,0.000091858485,0.003431067,0.0000021637788,0.0004671543,0.0008944495,0.00336472,0.5244314],"study_design_scores_gemma":[0.01603606,0.0067605097,0.8668679,0.008748429,0.0011328291,0.0016884678,0.016671302,0.009083469,0.00006137231,0.00069771265,0.07164503,0.0006069019],"about_ca_topic_score_codex":0.000046545527,"about_ca_topic_score_gemma":0.0000012314523,"teacher_disagreement_score":0.52382445,"about_ca_system_score_codex":0.000046691996,"about_ca_system_score_gemma":0.00056504423,"threshold_uncertainty_score":0.9407581},"labels":[],"label_agreement":null},{"id":"W4281624910","doi":"10.21203/rs.3.rs-1666441/v1","title":"Development of algorithms to identify individuals with Neurofibromatosis type 1 within administrative data and electronic medical records in Ontario, Canada.","year":2022,"lang":"en","type":"preprint","venue":"Research Square","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Hospital for Sick Children; Institute for Clinical Evaluative Sciences; University of Toronto","funders":"","keywords":"Medical record; Type (biology); Algorithm; Computer science; Genealogy; Medicine; History; Geology; Paleontology","score_opus":0.14312564584748744,"score_gpt":0.47024771007385224,"score_spread":0.3271220642263648,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4281624910","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99543715,0.00030904508,0.000015139753,0.0019720083,0.00015642372,0.0012213797,0.00026122015,0.00001569971,0.0006119601],"genre_scores_gemma":[0.9898692,0.00017072618,0.006467858,0.00019755878,0.000054882184,0.00023060429,0.0022984312,0.000048520564,0.00066226406],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9937627,0.00042576747,0.0007979271,0.0010102113,0.003334795,0.0006686239],"domain_scores_gemma":[0.9918829,0.0056592748,0.00017407966,0.0011309432,0.00036559734,0.0007871597],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0045802346,0.0002508004,0.0007390588,0.00033424734,0.00010997535,0.000045275458,0.00074514415,0.0002331304,0.0011817872],"category_scores_gemma":[0.031983517,0.00021083467,0.000025685315,0.00066538196,0.00014466995,0.000032100703,0.003149336,0.0033214553,0.000004198325],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0015203438,0.0009972065,0.960085,0.0009122305,0.0006385083,0.0019057526,0.0067265932,0.00003221137,0.000025248017,0.00011044095,0.015714822,0.011331639],"study_design_scores_gemma":[0.0018664352,0.002398104,0.96298563,0.012241401,0.00008372475,0.00011703841,0.0032611466,0.00016444214,0.00023551138,0.00014621553,0.016075611,0.0004247465],"about_ca_topic_score_codex":0.7971392,"about_ca_topic_score_gemma":0.9821508,"teacher_disagreement_score":0.1850116,"about_ca_system_score_codex":0.0012354376,"about_ca_system_score_gemma":0.07251492,"threshold_uncertainty_score":0.99973124},"labels":[],"label_agreement":null},{"id":"W4281635259","doi":"10.1038/s41591-022-01833-z","title":"Evaluating and reducing cognitive load should be a priority for machine learning in healthcare","year":2022,"lang":"en","type":"letter","venue":"Nature Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":60,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Vector Institute; University of Toronto; Artificial Intelligence in Medicine (Canada); Hospital for Sick Children","funders":"Hospital for Sick Children","keywords":"Health care; Cognition; Cognitive load; Computer science; Medicine; Psychology; Neuroscience; Economics","score_opus":0.09693438312235235,"score_gpt":0.4651619580729726,"score_spread":0.36822757495062025,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4281635259","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.029221397,0.03272075,0.00001564527,0.9341468,0.0011074689,0.00201784,0.00015723187,0.0001176913,0.00049515243],"genre_scores_gemma":[0.08418925,0.0012584447,0.0008495761,0.89834446,0.0074082725,0.00032411577,0.0045799143,0.00013839641,0.0029075646],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99539965,0.00041557362,0.0008534375,0.0010915146,0.0015775474,0.00066224963],"domain_scores_gemma":[0.9626589,0.035763785,0.00044303032,0.00033794832,0.00054531085,0.0002510473],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0037764092,0.00049478834,0.0015698463,0.0003470405,0.00022241418,0.000014519033,0.00013660501,0.0018513678,0.00035358558],"category_scores_gemma":[0.26867706,0.00039477364,0.00016167424,0.00045456196,0.00021853423,0.000030260353,0.00014609938,0.016692977,0.0000013119644],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0029978596,0.00023436607,0.025793986,0.003983284,0.0003735498,0.0030825785,0.0032330952,0.000009063432,0.00008052517,0.000048198446,0.76131845,0.19884504],"study_design_scores_gemma":[0.033474453,0.017034661,0.013428973,0.062151257,0.0030128881,0.0010206156,0.0009735103,0.0058194567,0.00004682746,0.0008849877,0.8608084,0.0013439932],"about_ca_topic_score_codex":0.0019254468,"about_ca_topic_score_gemma":0.00012367414,"teacher_disagreement_score":0.26490065,"about_ca_system_score_codex":0.0004985996,"about_ca_system_score_gemma":0.00087905716,"threshold_uncertainty_score":0.9998504},"labels":[],"label_agreement":null},{"id":"W4281715478","doi":"10.3899/jrheum.220502","title":"Fear of the Unknown: Can We Help Individuals With a Systemic Autoimmune Rheumatic Disease Deal With Uncertainty?","year":2022,"lang":"en","type":"letter","venue":"The Journal of Rheumatology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Medicine; Disease; Intensive care medicine; Autoimmune disease; Ambiguity; Rheumatology; Internal medicine","score_opus":0.014544466708015177,"score_gpt":0.26654570608726413,"score_spread":0.25200123937924895,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4281715478","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5396879,0.0045667216,0.00003452459,0.45449176,0.0004247025,0.0006343998,0.000067538356,0.000019968951,0.00007251566],"genre_scores_gemma":[0.9441314,0.004013643,0.0002470442,0.050048638,0.00012903211,0.000043885528,0.00006469843,0.00013632704,0.0011853531],"study_design_codex":"not_applicable","study_design_gemma":"case_report","domain_scores_codex":[0.9951541,0.0012987346,0.0014766518,0.00024156619,0.0013240448,0.0005048956],"domain_scores_gemma":[0.98034555,0.01509875,0.0030354548,0.0009635612,0.00035663636,0.00020002655],"candidate_categories":["research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0014719112,0.00043217387,0.0019969598,0.0002249657,0.00020387751,0.00001881142,0.0009732623,0.0003139605,0.00029130527],"category_scores_gemma":[0.004785178,0.00018229868,0.00036570104,0.0004191307,0.00078319653,0.00002862276,0.00020550944,0.0034446479,0.000006153142],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0058190404,0.00056497945,0.45176506,0.0034007665,0.0056199455,0.0017571747,0.0043535517,0.002436812,0.0000047714634,0.0003917607,0.5216999,0.0021862676],"study_design_scores_gemma":[0.030994225,0.01421415,0.06675108,0.26928428,0.022189612,0.36164507,0.005750704,0.00045158897,0.000012258687,0.0022080585,0.22458361,0.0019153622],"about_ca_topic_score_codex":0.00072293746,"about_ca_topic_score_gemma":0.000053379994,"teacher_disagreement_score":0.40444353,"about_ca_system_score_codex":0.000205473,"about_ca_system_score_gemma":0.002693933,"threshold_uncertainty_score":0.99885446},"labels":[],"label_agreement":null},{"id":"W4281718566","doi":"10.5014/ajot.2022.048074","title":"Clinical Reasoning in Occupational Therapy Practice: A Scoping Review of Qualitative and Conceptual Peer-Reviewed Literature","year":2022,"lang":"en","type":"review","venue":"American Journal of Occupational Therapy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":20,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University Health Centre","funders":"Coordenação de Aperfeiçoamento de Pessoal de Nível Superior","keywords":"Occupational therapy; PsycINFO; Qualitative research; Psychology; MEDLINE; Systematic review; Conceptual framework; Management science; Psychiatry; Sociology; Social science","score_opus":0.34401486375619433,"score_gpt":0.6089871821955033,"score_spread":0.26497231843930896,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4281718566","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000775829,0.99650085,0.000051067607,0.0007607045,0.00029892562,0.0013561728,0.00013297547,0.000008601013,0.00011489552],"genre_scores_gemma":[0.000076660224,0.98743784,0.007646768,0.0037352743,0.00050589035,0.00009028171,0.00041871745,0.000054031694,0.000034507255],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9886334,0.0047574453,0.0041759117,0.00049190765,0.0016665683,0.0002747775],"domain_scores_gemma":[0.903896,0.084105276,0.008955472,0.0004790868,0.0021707544,0.0003934125],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0113692,0.0004955396,0.0049987966,0.0004116983,0.00007356281,0.00003184563,0.00030365383,0.0001818319,0.00055704976],"category_scores_gemma":[0.10599513,0.00035244308,0.0012631265,0.0013700074,0.000773213,0.00021703458,0.00007583699,0.0017381062,0.0000030439767],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002079718,0.00071072835,0.0011667437,0.0063782386,0.0006885687,0.00010244706,0.0013185224,5.934333e-7,6.41614e-8,0.00022523779,0.0010322792,0.98629683],"study_design_scores_gemma":[0.0021966584,0.0033090254,0.001106982,0.60935444,0.00040472636,0.00070378266,0.00061208755,0.0000010140797,8.647199e-8,0.000028550303,0.3820156,0.00026704185],"about_ca_topic_score_codex":0.0000237172,"about_ca_topic_score_gemma":8.0553565e-7,"teacher_disagreement_score":0.9860298,"about_ca_system_score_codex":0.000164396,"about_ca_system_score_gemma":0.004532932,"threshold_uncertainty_score":0.9998928},"labels":[],"label_agreement":null},{"id":"W4282923585","doi":"10.1016/j.ijnurstu.2022.104284","title":"Cognitive and implicit biases in nurses' judgment and decision-making: A scoping review","year":2022,"lang":"en","type":"review","venue":"International Journal of Nursing Studies","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":98,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta; Athabasca University","funders":"Athabasca University","keywords":"Psychology; Cognition; Cognitive bias; MEDLINE; Applied psychology; Clinical decision making; Clinical judgment; Cognitive psychology; Social psychology; Medicine; Psychiatry; Political science","score_opus":0.23673339802698454,"score_gpt":0.5737490832863,"score_spread":0.33701568525931547,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4282923585","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000749362,0.99663407,0.000012305727,0.0006206466,0.0012842491,0.00051213586,0.000017767077,0.0000052554283,0.00016418462],"genre_scores_gemma":[0.00040705397,0.9983573,0.00038952136,0.0005025023,0.00028047548,0.000026582595,0.0000063552766,0.00001989608,0.000010329532],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99769604,0.00010393385,0.0012011419,0.00024328595,0.00059939834,0.00015619115],"domain_scores_gemma":[0.96140814,0.037094403,0.0009411058,0.000080589714,0.00037377063,0.00010197613],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0007859416,0.00023287655,0.0020109254,0.00039481823,0.00005582189,0.000031514453,0.00011760817,0.00006023034,0.00004803536],"category_scores_gemma":[0.077044636,0.00016903348,0.00032256942,0.00016277163,0.00016996206,0.00005411461,0.00011269745,0.0004795035,0.0000011435577],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007801362,0.00021269829,0.000106388026,0.003656985,0.00074424566,0.00054936536,0.00024614268,3.343214e-7,1.0670146e-8,0.000013661616,0.0008628482,0.9935293],"study_design_scores_gemma":[0.0007540722,0.00032746777,0.0002450602,0.9790378,0.0022438492,0.0039819283,0.0005146685,0.0000012956389,3.6832347e-8,0.00036831084,0.012395879,0.00012960876],"about_ca_topic_score_codex":0.000002028624,"about_ca_topic_score_gemma":0.00000198324,"teacher_disagreement_score":0.9933997,"about_ca_system_score_codex":0.00038231627,"about_ca_system_score_gemma":0.00032677353,"threshold_uncertainty_score":0.9307298},"labels":[],"label_agreement":null},{"id":"W4283749302","doi":"10.21474/ijar01/14875","title":"STUDENT OPINION ABOUT SCRIPT CONCORDANCE TEST, JEDDAH, KSA","year":2022,"lang":"en","type":"article","venue":"International Journal of Advanced Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Concordance; Test (biology); Medical education; Psychology; Mathematics education; Medicine","score_opus":0.07165328073081269,"score_gpt":0.5030576859300949,"score_spread":0.43140440519928225,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4283749302","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.963871,0.004413948,0.000297502,0.017890159,0.0056148716,0.0003547905,0.000052989282,0.00002404371,0.0074807107],"genre_scores_gemma":[0.9923495,0.0016327226,0.0015373336,0.00073842425,0.0011486846,0.00002374316,0.00001555011,0.000019148014,0.002534873],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.99572366,0.0001726679,0.00060780486,0.0001919166,0.003027815,0.00027614183],"domain_scores_gemma":[0.9890363,0.008174305,0.00026496488,0.00019263219,0.002067277,0.00026449008],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00220507,0.00009050344,0.0002618531,0.00035331762,0.00015024855,0.000046478322,0.0006039638,0.00003138539,0.0007329689],"category_scores_gemma":[0.03570255,0.00007676964,0.00016305683,0.000278216,0.000143388,0.000114058654,0.0003402241,0.0012615885,0.000031320844],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008313833,0.0092387255,0.3979635,0.00005322437,0.0010524918,0.007444365,0.0020946204,0.0087910965,0.033813674,0.005660106,0.13234538,0.39322898],"study_design_scores_gemma":[0.016056275,0.009407277,0.4068578,0.0035233998,0.000043987187,0.0039277826,0.0047135563,0.0004325322,0.0033282237,0.004569193,0.54680467,0.00033530977],"about_ca_topic_score_codex":0.000017256178,"about_ca_topic_score_gemma":0.0000013712464,"teacher_disagreement_score":0.4144593,"about_ca_system_score_codex":0.00051656034,"about_ca_system_score_gemma":0.0005190044,"threshold_uncertainty_score":0.97242016},"labels":[],"label_agreement":null},{"id":"W4288447045","doi":"10.3390/computation10080130","title":"Understanding Clinical Reasoning through Visual Scanpath and Brain Activity Analysis","year":2022,"lang":"en","type":"article","venue":"Computation","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; McGill University","funders":"","keywords":"Eye tracking; Gaze; Computer science; Electroencephalography; Workload; Cognition; Amnesia; Task (project management); Protocol (science); Artificial intelligence; Psychology; Human–computer interaction; Cognitive psychology; Medicine","score_opus":0.15917755681141077,"score_gpt":0.44635861398857235,"score_spread":0.2871810571771616,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4288447045","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.78399146,0.0000533095,0.21339206,0.0016657942,0.00014613722,0.00011263399,0.0000039517513,0.000057489742,0.00057714933],"genre_scores_gemma":[0.996757,0.00001744167,0.0021264635,0.0008282082,0.00010136438,0.0000075862736,0.000077050405,0.000010190186,0.00007473268],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9987851,0.0002353377,0.00025869635,0.00030793902,0.00027346303,0.000139466],"domain_scores_gemma":[0.9918022,0.007829711,0.00013134153,0.000095862786,0.000031667256,0.00010917891],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00089741446,0.00008280525,0.00031955788,0.00009448832,0.0002464285,0.0000270259,0.000029222345,0.000045860324,0.00006992181],"category_scores_gemma":[0.007335753,0.000082889535,0.00013992618,0.0005712603,0.00007105814,0.000053972235,0.00009054937,0.00027438047,0.000003016309],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00064740813,0.0009637123,0.90871567,0.00002457053,0.0009826659,0.00012093395,0.0013724663,0.013203791,0.000072069175,0.0030814416,0.00513516,0.06568012],"study_design_scores_gemma":[0.0021153442,0.00087034184,0.6550331,0.000100257384,0.0009024318,0.000035163048,0.0008654205,0.33710065,0.0000103944385,0.0026459526,0.00016465585,0.00015629087],"about_ca_topic_score_codex":0.00007082005,"about_ca_topic_score_gemma":0.000007693099,"teacher_disagreement_score":0.32389686,"about_ca_system_score_codex":0.0001492183,"about_ca_system_score_gemma":0.00008341301,"threshold_uncertainty_score":0.87821114},"labels":[],"label_agreement":null},{"id":"W4289523369","doi":"10.1111/medu.14887","title":"Exploring the development of adaptive expertise through the lens of threshold concepts","year":2022,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Mindset; Framing (construction); Psychology; Grounded theory; Liminality; Exploratory research; Dreyfus model of skill acquisition; Qualitative research; Medical education; Medicine; Computer science; Sociology","score_opus":0.14047279478889663,"score_gpt":0.39929326784144314,"score_spread":0.2588204730525465,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4289523369","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97889286,0.0007634587,0.00027815648,0.015933864,0.0008342495,0.000264335,0.0000015759915,0.000011662921,0.0030198395],"genre_scores_gemma":[0.99612886,0.00008029373,0.00045127602,0.0027010904,0.00014079183,0.0003079371,0.0000134503525,0.0000071280065,0.00016914927],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99857414,0.00007759918,0.00038524665,0.00012244063,0.0007297229,0.00011085198],"domain_scores_gemma":[0.9967872,0.0026789738,0.00012977958,0.0002556298,0.00009242852,0.000055969726],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0006271406,0.00006277493,0.0001662437,0.000017387392,0.00013797896,0.0000019583733,0.00016025889,0.00002297442,0.00067769503],"category_scores_gemma":[0.014986965,0.00003385047,0.00005951299,0.0001798906,0.00021839101,0.00003259691,0.00009220199,0.0002480865,0.000004471789],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005487319,0.0049349386,0.01026552,0.00007908915,0.00025402434,0.000005493964,0.24329303,0.000087554516,0.00044524862,0.018485997,0.049507897,0.67209244],"study_design_scores_gemma":[0.0069427625,0.0026660792,0.43543443,0.007423304,0.0006442018,0.00013441332,0.28816226,0.0022378555,0.02181978,0.0035079378,0.23029734,0.000729622],"about_ca_topic_score_codex":0.000093469665,"about_ca_topic_score_gemma":0.00000939148,"teacher_disagreement_score":0.6713628,"about_ca_system_score_codex":0.00006869172,"about_ca_system_score_gemma":0.0017992313,"threshold_uncertainty_score":0.9933102},"labels":[],"label_agreement":null},{"id":"W4289717307","doi":"10.1051/pmed/2022005","title":"Former et évaluer par concordance : des modalités éducatives complémentaires","year":2022,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Humanities; Concordance; Philosophy; Medicine; Internal medicine","score_opus":0.10967497506132828,"score_gpt":0.41607020406104006,"score_spread":0.3063952289997118,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4289717307","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.74740857,0.13511631,0.0010485625,0.06906858,0.0065830788,0.00095552794,0.0006008174,0.00024449878,0.03897405],"genre_scores_gemma":[0.8678225,0.006813137,0.0032120615,0.020797757,0.0010955859,0.000421665,0.00027424516,0.000106367916,0.09945668],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99556416,0.00085203483,0.00080256484,0.000738619,0.00091102,0.0011315801],"domain_scores_gemma":[0.9834259,0.014766737,0.00023403927,0.00066379865,0.00011468414,0.0007948225],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0017403766,0.00040156528,0.00077057333,0.00009311958,0.0005941075,0.00006168522,0.00040778622,0.00014378941,0.013719954],"category_scores_gemma":[0.020009395,0.00039355754,0.0003643023,0.0004040611,0.00162858,0.00015591415,0.0005647572,0.0010884848,0.0005896801],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006232784,0.0071201907,0.17316328,0.00035436315,0.0006989925,0.00062984583,0.014749763,0.0005866398,0.00038063608,0.012245159,0.2755941,0.5138538],"study_design_scores_gemma":[0.00448702,0.0019934697,0.13386674,0.0024617864,0.00063045527,0.0002440153,0.010875308,0.0014096809,0.0007303064,0.014845291,0.8277161,0.00073984073],"about_ca_topic_score_codex":0.0009152246,"about_ca_topic_score_gemma":0.000029329185,"teacher_disagreement_score":0.552122,"about_ca_system_score_codex":0.0004949745,"about_ca_system_score_gemma":0.00088733155,"threshold_uncertainty_score":0.99985164},"labels":[],"label_agreement":null},{"id":"W4289752914","doi":"10.31234/osf.io/7xj2h","title":"Assessing and adjusting for publication bias in the relationship between anxiety and the error-related negativity","year":2018,"lang":"en","type":"preprint","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Anxiety; Worry; Moderation; Publication bias; Psychology; Error-related negativity; Clinical psychology; Meta-analysis; Medicine; Anterior cingulate cortex; Psychiatry; Cognition; Social psychology; Internal medicine","score_opus":0.25371219023198455,"score_gpt":0.44028350983837433,"score_spread":0.18657131960638978,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4289752914","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9503788,0.00027405916,0.0020563083,0.040593058,0.000103255756,0.0014119548,0.000011095864,0.00005378255,0.0051177284],"genre_scores_gemma":[0.9957817,0.000044384302,0.002656073,0.00068300456,0.0002582917,0.0001238424,0.00019538519,0.000015162768,0.00024211057],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982756,0.0003953849,0.00054840365,0.00040047683,0.00020226439,0.00017789651],"domain_scores_gemma":[0.84134954,0.1571411,0.0004923631,0.0006244761,0.00027779947,0.00011468182],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0061523807,0.00016529529,0.00039294973,0.00006957987,0.00021274578,0.00028188757,0.00011256137,0.00037150306,0.000013883067],"category_scores_gemma":[0.28656086,0.00008531108,0.000087354485,0.00017937188,0.00038705528,0.00007709349,0.000191658,0.00084609934,0.0000028471482],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000033278757,0.000033927914,0.9777996,0.00007844432,0.000029309727,3.1488128e-7,0.0007833962,0.000002456921,3.038625e-7,0.0030712057,0.0012053206,0.016962452],"study_design_scores_gemma":[0.0020840159,0.000031883017,0.9708697,0.0014392708,0.0003153463,0.000011051481,0.00040675327,0.0033966713,0.0000024311905,0.021179324,0.00016389227,0.00009964069],"about_ca_topic_score_codex":0.00035011774,"about_ca_topic_score_gemma":0.00003996308,"teacher_disagreement_score":0.28040847,"about_ca_system_score_codex":0.00003204611,"about_ca_system_score_gemma":0.00012922661,"threshold_uncertainty_score":0.71944875},"labels":[],"label_agreement":null},{"id":"W4292542962","doi":"10.1111/jan.15414","title":"Scoping review: Diagnostic reasoning as a component of clinical reasoning in the U.S. primary care nurse practitioner education","year":2022,"lang":"en","type":"article","venue":"Journal of Advanced Nursing","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":43,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Regina","funders":"","keywords":"Nurse practitioners; Component (thermodynamics); Primary care; Nursing; Psychology; Medicine; Medical education; Family medicine; Health care; Political science","score_opus":0.02558062976951597,"score_gpt":0.43009583382395683,"score_spread":0.40451520405444086,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4292542962","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8482037,0.13647366,0.00015600477,0.0067596724,0.002237927,0.0012659432,0.000004886235,0.000014975482,0.0048832283],"genre_scores_gemma":[0.9677565,0.015342369,0.0118753575,0.0041745068,0.00066841434,0.000056999193,0.00006171286,0.000033120148,0.000030994215],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99634653,0.00065446406,0.0017004925,0.00021944812,0.0008544576,0.00022462859],"domain_scores_gemma":[0.9808803,0.016221516,0.0019113611,0.00034966654,0.00046082924,0.0001762869],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0026783447,0.00014599419,0.00084387534,0.00016798962,0.0001578936,0.000014054568,0.00019419358,0.00005486133,0.00005679831],"category_scores_gemma":[0.08110958,0.000113359274,0.00031821075,0.00044642307,0.00009916716,0.00018249915,0.000027288088,0.001124884,0.0000020884347],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0027076094,0.006930412,0.084848836,0.003752596,0.00011137116,0.0006156907,0.0073660114,0.004012686,0.0003473315,0.0006393143,0.006328846,0.8823393],"study_design_scores_gemma":[0.0038651468,0.0023901875,0.18867724,0.77649784,0.00068798976,0.0031274075,0.022626337,0.000038042923,0.00008841217,0.00031741126,0.001454088,0.00022989773],"about_ca_topic_score_codex":0.000017173574,"about_ca_topic_score_gemma":0.0000016534674,"teacher_disagreement_score":0.8821094,"about_ca_system_score_codex":0.00043080817,"about_ca_system_score_gemma":0.0019096937,"threshold_uncertainty_score":0.9266306},"labels":[],"label_agreement":null},{"id":"W4292860082","doi":"10.1111/tct.13523","title":"Ambiguity and uncertainty tolerance and psychological needs of medical students: A cross‐sectional survey","year":2022,"lang":"en","type":"article","venue":"The Clinical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Alberta","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"Ambiguity; Competence (human resources); Autonomy; Ambiguity tolerance; Psychology; Cross-sectional study; Medical education; Social psychology; Medicine; Computer science","score_opus":0.1735748173096371,"score_gpt":0.5050104130447053,"score_spread":0.3314355957350682,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4292860082","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9951156,0.00039713024,0.000024321507,0.0034034352,0.0003902877,0.00021759162,0.000026881016,0.000034937482,0.00038982014],"genre_scores_gemma":[0.9951435,0.00029542207,0.00003144747,0.0030986338,0.00026744083,0.000026972026,0.00001966777,0.000010930766,0.0011060403],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99641436,0.0010768541,0.00080865336,0.00031996018,0.001176487,0.00020365269],"domain_scores_gemma":[0.9804067,0.01861466,0.00016921957,0.00038706025,0.000084413594,0.00033794896],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.013871931,0.0001232643,0.00048016728,0.00002801397,0.00016591811,0.00002882662,0.00037609748,0.0002021507,0.0014448778],"category_scores_gemma":[0.06392829,0.00007641281,0.0001372886,0.00019812689,0.0011097614,0.000015845131,0.00062369055,0.0012334171,0.0000063630796],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00092212437,0.0017871136,0.9869607,0.000005344805,0.00006701057,0.000012256317,0.000109025525,0.0000073781825,0.0000011193379,0.0001485642,0.0025261175,0.007453248],"study_design_scores_gemma":[0.0033184371,0.0011337915,0.99326205,0.000053777996,0.000037020513,0.000091996306,0.00010306758,0.0002814403,3.0297593e-7,0.00053497683,0.0011003648,0.00008277442],"about_ca_topic_score_codex":0.00028778857,"about_ca_topic_score_gemma":0.000017167908,"teacher_disagreement_score":0.050056364,"about_ca_system_score_codex":0.000022174057,"about_ca_system_score_gemma":0.000077156714,"threshold_uncertainty_score":0.9994679},"labels":[],"label_agreement":null},{"id":"W4293661949","doi":"10.2196/39234","title":"Electronic Diagnostic Support in Emergency Physician Triage: Qualitative Study With Thematic Analysis of Interviews","year":2022,"lang":"en","type":"article","venue":"JMIR Human Factors","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Impact; McMaster University","funders":"","keywords":"Triage; Emergency department; Thematic analysis; Medical diagnosis; Context (archaeology); Medical emergency; Qualitative research; Medicine; Psychology; Nursing; Sociology","score_opus":0.053818724461425,"score_gpt":0.43144597573748966,"score_spread":0.37762725127606467,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4293661949","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99793434,0.00007149601,0.000016277198,0.000077804114,0.000045471043,0.0012426722,0.00003169411,0.000030240524,0.00054998475],"genre_scores_gemma":[0.99870354,0.000011729376,0.000005068961,0.0000663637,0.00001963918,0.0004684689,0.00021458203,0.000025423216,0.00048516135],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99719936,0.0005586235,0.0008990462,0.0003837453,0.00061495963,0.00034426973],"domain_scores_gemma":[0.99328107,0.0056453906,0.0004114419,0.00048174686,0.000072318486,0.000108043016],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0010627331,0.00021947027,0.0010674191,0.0006181227,0.000090044414,0.000006960078,0.00019964861,0.00003192645,0.0027789467],"category_scores_gemma":[0.004141122,0.00016157664,0.00032559797,0.002007918,0.000058601512,0.000042786854,0.00009093381,0.00044133246,0.000007647067],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00029601346,0.010455124,0.765,0.000072113944,0.0027039545,0.00007359979,0.21903434,0.00011611146,0.00006759937,0.00041626827,0.0007624629,0.0010024157],"study_design_scores_gemma":[0.0031626136,0.010314314,0.8085442,0.00019452737,0.0027230892,0.0000010124991,0.1741794,0.00003308718,0.00004190124,0.00046593166,0.00006987204,0.0002700096],"about_ca_topic_score_codex":0.00027302364,"about_ca_topic_score_gemma":0.00081249117,"teacher_disagreement_score":0.04485495,"about_ca_system_score_codex":0.00018357506,"about_ca_system_score_gemma":0.00010972215,"threshold_uncertainty_score":0.99813265},"labels":[],"label_agreement":null},{"id":"W4294026051","doi":"10.3899/jrheum.220666","title":"Desk Rejections: Not Without Due Deliberation","year":2022,"lang":"en","type":"letter","venue":"The Journal of Rheumatology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Arthritis Research Centre of Canada","funders":"","keywords":"Desk; Deliberation; Medicine; Law","score_opus":0.03059017520692151,"score_gpt":0.319206104073742,"score_spread":0.28861592886682047,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4294026051","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.104850695,0.0014525114,0.0005671673,0.8896338,0.0025790986,0.00020893145,0.000009915684,0.00002899978,0.00066892046],"genre_scores_gemma":[0.11778464,0.005950165,0.0011242194,0.86882025,0.0011233755,0.000021085714,0.0001121118,0.00010758976,0.004956586],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99755806,0.0005793265,0.0009766136,0.00013972794,0.0004938848,0.00025237567],"domain_scores_gemma":[0.98342806,0.014774392,0.0010468974,0.0003880724,0.0002817418,0.00008084907],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012644735,0.00019238458,0.0009149133,0.0002030154,0.00017211732,0.00001544433,0.0002594689,0.00044543153,0.0010626843],"category_scores_gemma":[0.022371246,0.000119027594,0.0002798368,0.00014617176,0.00016235729,0.000035757388,0.00006597517,0.0036413479,0.000056762972],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00050621276,0.000058926664,0.005914859,0.000036770005,0.00027917288,0.00035785735,0.00016575058,0.000017300199,0.0000046717932,0.000046933605,0.9910682,0.0015433459],"study_design_scores_gemma":[0.0024438945,0.0018584462,0.0020246918,0.0020737324,0.001464377,0.1918152,0.0001409702,0.00008108443,0.0000219347,0.0010382956,0.79680276,0.00023459083],"about_ca_topic_score_codex":0.00016558237,"about_ca_topic_score_gemma":0.000013379364,"teacher_disagreement_score":0.19426543,"about_ca_system_score_codex":0.00011596211,"about_ca_system_score_gemma":0.0004914692,"threshold_uncertainty_score":0.99985045},"labels":[],"label_agreement":null},{"id":"W4294043912","doi":"10.21203/rs.3.rs-2004458/v1","title":"Pathologists aren’t Pigeons: Exploring the neural basis of visual recognition and perceptual expertise in pathology","year":2022,"lang":"en","type":"preprint","venue":"Research Square","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria; University of Calgary","funders":"","keywords":"Categorization; Perception; Task (project management); Psychology; Electroencephalography; Visual perception; Perceptual learning; Event-related potential; Cognitive psychology; Computer science; Artificial intelligence; Neuroscience","score_opus":0.2857144027475815,"score_gpt":0.47042346909011423,"score_spread":0.18470906634253276,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4294043912","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99391985,0.0026128269,0.000010392454,0.0017780556,0.00026769537,0.0009721092,0.00013475375,0.000037852096,0.0002664861],"genre_scores_gemma":[0.99330956,0.0045132674,0.00010193643,0.00011628551,0.00018719104,0.0013199197,0.00034380227,0.00003356082,0.000074508935],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.995736,0.0016025302,0.00057174003,0.00069240667,0.00085966993,0.00053766096],"domain_scores_gemma":[0.9866284,0.012206713,0.000113446775,0.0005744965,0.00028296091,0.00019398468],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0030653381,0.00021091232,0.00059331645,0.00042286524,0.00014358119,0.000032297336,0.0002129195,0.0002482821,0.0006595956],"category_scores_gemma":[0.066619135,0.00015664207,0.00016722726,0.0003465956,0.0006485715,0.00004127417,0.001365314,0.00260363,0.000015968508],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0015327918,0.0017782568,0.30170184,0.0009441427,0.000060628612,0.0031716472,0.017776154,0.0001118027,0.00061667414,0.000029610675,0.00080636767,0.6714701],"study_design_scores_gemma":[0.001596232,0.002294725,0.9715789,0.004561028,0.000053180636,0.00014003279,0.01727785,0.0010247853,0.00024646329,0.0008931176,0.00008243757,0.00025127848],"about_ca_topic_score_codex":0.00088337844,"about_ca_topic_score_gemma":0.000098541524,"teacher_disagreement_score":0.6712188,"about_ca_system_score_codex":0.00020131467,"about_ca_system_score_gemma":0.00027150498,"threshold_uncertainty_score":0.9996974},"labels":[],"label_agreement":null},{"id":"W4294351363","doi":"10.36834/cmej.72429","title":"Nomenclature des patients dans la formation des professionnels de la santé en fonction du rôle et de l'implication : une revue de littérature narrative","year":2022,"lang":"fr","type":"review","venue":"Canadian Medical Education Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"CINAHL; Psychology; MEDLINE; Health care; Medicine; Humanities; Nursing; Philosophy; Political science","score_opus":0.015548094894925323,"score_gpt":0.3836632831778018,"score_spread":0.3681151882828765,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4294351363","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.1516428,0.7177791,0.010142781,0.06792614,0.011739765,0.004003634,0.00037551846,0.00015040065,0.036239903],"genre_scores_gemma":[0.055354588,0.9218009,0.0038549872,0.008379663,0.0067808544,0.00058848405,0.001727767,0.000120228266,0.0013925197],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9917208,0.005102072,0.0011519637,0.0004037769,0.0007254878,0.0008959228],"domain_scores_gemma":[0.9728484,0.014251186,0.0010849762,0.0004354638,0.0007538348,0.010626142],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0070144786,0.00044192697,0.0007727092,0.00043400278,0.00117392,0.00018532915,0.0004580126,0.0015447905,0.010107731],"category_scores_gemma":[0.3260852,0.00039750664,0.00034888773,0.0010486449,0.000549002,0.00035754818,0.00005042637,0.0046911915,0.000035821322],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.000036209593,0.0008356426,0.0069039133,0.0018692495,0.00013198685,0.0001295519,0.07722358,0.000013799985,0.0000040550294,0.0030743252,0.058989853,0.8507878],"study_design_scores_gemma":[0.0012196142,0.00036571512,0.069389015,0.064714335,0.0007844884,0.008781773,0.028364351,0.00018361726,0.0000026540301,0.007473601,0.81826717,0.0004536721],"about_ca_topic_score_codex":0.0041657933,"about_ca_topic_score_gemma":0.0068308813,"teacher_disagreement_score":0.85033417,"about_ca_system_score_codex":0.014069826,"about_ca_system_score_gemma":0.11171659,"threshold_uncertainty_score":0.9998477},"labels":[],"label_agreement":null},{"id":"W4294405937","doi":"10.7759/cureus.28730","title":"Diagnosis: Fundamental Principles and Methods","year":2022,"lang":"en","type":"review","venue":"Cureus","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ivanhoe Energy (Canada)","funders":"","keywords":"Medicine; Medical diagnosis; Terminology; Process (computing); Health care; Harm; Risk analysis (engineering); Management science; Intensive care medicine; Computer science; Pathology","score_opus":0.2271965757935339,"score_gpt":0.509734047040829,"score_spread":0.2825374712472951,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4294405937","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000014595863,0.9943495,0.00001091944,0.00007643261,0.00060950185,0.00066353485,0.000055410314,0.00007763701,0.004142503],"genre_scores_gemma":[0.0000011429998,0.9918734,0.0053712446,0.00021903265,0.00033213347,0.0007178007,0.00022124077,0.0000578498,0.0012061836],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9982087,0.0002810883,0.0005475385,0.0004890265,0.00022584372,0.00024780925],"domain_scores_gemma":[0.9778075,0.021174524,0.00022335058,0.0004748733,0.000015121334,0.00030461233],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0007308151,0.00030836937,0.0016665789,0.00006267617,0.00009147849,0.000030199071,0.00013704569,0.00022888508,0.003017907],"category_scores_gemma":[0.031387735,0.00023041987,0.00044035475,0.00015917562,0.000097780714,0.000020339503,0.0003228915,0.00057871436,0.00004766656],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000038906865,0.00018643463,0.00023579109,0.002403132,0.00021244635,0.00008594332,0.000027363927,1.6750713e-8,9.06255e-9,0.00021068574,0.0022088937,0.9944254],"study_design_scores_gemma":[0.00027472366,0.00021887872,0.00014768084,0.019750468,0.002289579,0.00027764443,0.000032887612,0.0000013901922,2.1344036e-7,0.000044603792,0.9767501,0.0002118509],"about_ca_topic_score_codex":0.000032337295,"about_ca_topic_score_gemma":0.000002608362,"teacher_disagreement_score":0.9942135,"about_ca_system_score_codex":0.00017466611,"about_ca_system_score_gemma":0.00027643974,"threshold_uncertainty_score":0.99789345},"labels":[],"label_agreement":null},{"id":"W4294578992","doi":"10.1177/0272989x221121343","title":"The Fallacy of a Single Diagnosis","year":2022,"lang":"en","type":"article","venue":"Medical Decision Making","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; University of Toronto; Institute for Clinical Evaluative Sciences; Sunnybrook Health Science Centre","funders":"Canadian Institutes of Health Research; Canada Research Chairs; Alfred P. Sloan Foundation","keywords":"Fallacy; Medicine; Single use; Intensive care medicine; Psychology; Computer science; Epistemology; Philosophy; Engineering","score_opus":0.03809602560556307,"score_gpt":0.37673629229054345,"score_spread":0.3386402666849804,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4294578992","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9754458,0.0019432119,0.00446939,0.0100604985,0.0016192505,0.00030304017,0.00000886242,0.000069241956,0.006080717],"genre_scores_gemma":[0.996163,0.00018957446,0.001134569,0.0020589603,0.00018671335,0.00006508448,0.0000033231327,0.000014217922,0.00018452929],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9971217,0.000111575115,0.00052469195,0.00021278417,0.0018032216,0.00022602107],"domain_scores_gemma":[0.9353791,0.06366225,0.00015550251,0.0004611834,0.000089246125,0.00025273766],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0017007779,0.00008431654,0.000270232,0.000056741545,0.00027867052,0.000014852237,0.00027731297,0.00005985902,0.0033529138],"category_scores_gemma":[0.20136365,0.00005221428,0.00016979755,0.0003378679,0.00014676989,0.00001245724,0.00039923686,0.0003736773,0.000027952066],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00026387646,0.00047569748,0.030811315,0.000005200767,0.00002363615,0.00013261862,0.00007736686,0.000019096775,0.000021122032,0.00079876423,0.03141137,0.93595994],"study_design_scores_gemma":[0.00783268,0.004820041,0.20062512,0.008714167,0.00031679097,0.0007924282,0.0011591137,0.006391895,0.00034452748,0.029402953,0.7390939,0.000506415],"about_ca_topic_score_codex":0.000013365852,"about_ca_topic_score_gemma":0.000005589637,"teacher_disagreement_score":0.93545353,"about_ca_system_score_codex":0.000061892504,"about_ca_system_score_gemma":0.0001601389,"threshold_uncertainty_score":0.9975582},"labels":[],"label_agreement":null},{"id":"W4295778983","doi":"10.1101/2022.09.12.22279750","title":"Debiasing and Educational Interventions in Medical Diagnosis: A Systematic Review","year":2022,"lang":"en","type":"review","venue":"medRxiv","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Guelph; North York General Hospital","funders":"","keywords":"Debiasing; PsycINFO; Psychological intervention; Critical appraisal; MEDLINE; Meta-analysis; Intervention (counseling); Medicine; Cognitive bias; Psychology; Cognition; Psychiatry; Social psychology; Alternative medicine","score_opus":0.13288528457509105,"score_gpt":0.4661215688067668,"score_spread":0.33323628423167573,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4295778983","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000025177498,0.99539334,0.0000024724034,0.0020100498,0.0002890081,0.0018970732,0.000016105887,0.000024994333,0.00034176957],"genre_scores_gemma":[0.000012462302,0.99493307,0.000052343497,0.0009916832,0.00011744584,0.0033082787,0.0002083321,0.000037343863,0.00033904242],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99635094,0.0006748464,0.0017117783,0.000427873,0.00061775645,0.0002168075],"domain_scores_gemma":[0.9591127,0.039313003,0.0005349857,0.0005657219,0.000041334348,0.00043225285],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002579809,0.0002678481,0.0030820444,0.00022779286,0.000052916548,0.000017669661,0.00020751772,0.00017985057,0.004164828],"category_scores_gemma":[0.2631415,0.0001943939,0.00078278664,0.0005134655,0.00007507487,0.000022181075,0.00016402446,0.000783688,0.000081389044],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[7.87055e-7,0.00034920702,0.00080135796,0.8723588,0.00010710047,0.00013275885,0.000013742469,3.5651966e-9,1.9503738e-10,0.00010270105,0.00094187754,0.12519167],"study_design_scores_gemma":[0.00017939124,0.000052078358,0.00020542365,0.8754635,0.003827458,0.00070965703,0.00000656246,0.0000015323942,2.0413544e-9,0.000030244033,0.11938829,0.00013588808],"about_ca_topic_score_codex":0.00002345756,"about_ca_topic_score_gemma":0.00001762068,"teacher_disagreement_score":0.2605617,"about_ca_system_score_codex":0.00015196504,"about_ca_system_score_gemma":0.00076826045,"threshold_uncertainty_score":0.9967455},"labels":[],"label_agreement":null},{"id":"W4296785885","doi":"10.1111/medu.14938","title":"Is uncertainty tolerance an epiphenomenon?","year":2022,"lang":"en","type":"letter","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Epiphenomenon; Psychology; Medicine; Epistemology; Philosophy","score_opus":0.02457601478095018,"score_gpt":0.37444579791188626,"score_spread":0.3498697831309361,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4296785885","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008139273,0.00066581427,0.000018498085,0.97820354,0.0036476457,0.00039071907,0.000033935117,0.00010734421,0.008793218],"genre_scores_gemma":[0.0029384154,0.00023521582,0.00022457617,0.9624536,0.011994008,0.00025820953,0.0041708555,0.000056853227,0.017668247],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9966761,0.0001765979,0.00052881724,0.00066247006,0.0015670677,0.0003889558],"domain_scores_gemma":[0.9952216,0.0031779779,0.00019071465,0.0008240779,0.00015760025,0.00042803786],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00058628747,0.00026091156,0.0005612426,0.00012373549,0.00011529792,0.000027310745,0.0003080405,0.00077285635,0.051663723],"category_scores_gemma":[0.036220223,0.00022923648,0.00018463765,0.00024614867,0.00016575467,0.000037946502,0.00006667686,0.00293634,0.00025972433],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000036269423,0.00064867595,0.00070998835,0.00008292323,0.000027148651,0.00011020461,0.00034570202,0.0000015958454,1.7629307e-7,0.000020581712,0.9584499,0.03956682],"study_design_scores_gemma":[0.00057065213,0.00043473652,0.0017827387,0.0011605829,0.00015076823,0.00009813512,0.00008923736,0.00021597579,7.669392e-7,0.0004124332,0.99485284,0.0002311185],"about_ca_topic_score_codex":0.0005743339,"about_ca_topic_score_gemma":0.000006055016,"teacher_disagreement_score":0.051404,"about_ca_system_score_codex":0.000315028,"about_ca_system_score_gemma":0.0053651338,"threshold_uncertainty_score":0.9993639},"labels":[],"label_agreement":null},{"id":"W4297570957","doi":"10.1097/acm.0000000000004899","title":"An Ecological Account of Clinical Reasoning","year":2022,"lang":"en","type":"review","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":26,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Affordance; Ecological psychology; Context (archaeology); Psychology; Competence (human resources); Engineering ethics; Cognitive psychology; Social psychology","score_opus":0.2692485472713973,"score_gpt":0.5496478088068679,"score_spread":0.2803992615354706,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4297570957","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00058501057,0.991237,0.000021478249,0.0003565071,0.0016717868,0.0010008853,0.000030134359,0.00012997148,0.0049672453],"genre_scores_gemma":[0.00027250953,0.99216354,0.00020584736,0.0015020279,0.003929283,0.00013859083,0.00080408243,0.00009009515,0.0008940188],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9919143,0.00096124556,0.004375887,0.001055098,0.0011486743,0.00054483785],"domain_scores_gemma":[0.94721603,0.048694827,0.001914535,0.0011452168,0.00010887196,0.00092051737],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.007810686,0.0005768091,0.007106584,0.0002911043,0.00008439651,0.0000031971836,0.0007624772,0.001807743,0.008404695],"category_scores_gemma":[0.22491926,0.00036309435,0.00084725657,0.0006728298,0.0007308116,0.000048308142,0.00025608134,0.0058148913,0.00008474474],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006685068,0.00024178208,0.0010794911,0.0026127598,0.00025736468,0.00022639435,0.00010373385,2.7545528e-7,1.3108334e-7,0.0006692662,0.017614039,0.9771279],"study_design_scores_gemma":[0.0014968131,0.002626249,0.0006370598,0.06165545,0.0043169037,0.00045687845,0.00016528384,0.000022782851,8.8052275e-8,0.0001267141,0.9282152,0.00028059445],"about_ca_topic_score_codex":0.000054000706,"about_ca_topic_score_gemma":7.1175504e-7,"teacher_disagreement_score":0.9768473,"about_ca_system_score_codex":0.00014396964,"about_ca_system_score_gemma":0.0012965542,"threshold_uncertainty_score":0.9998821},"labels":[],"label_agreement":null},{"id":"W4299571350","doi":"","title":"Confusion reigns.","year":2008,"lang":"en","type":"article","venue":"PubMed","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Canadian Public Health Association","funders":"","keywords":"Confusion; Psychology; Psychoanalysis","score_opus":0.06052637316828681,"score_gpt":0.285020317759112,"score_spread":0.22449394459082522,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4299571350","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.915874,0.0002079798,0.000029103101,0.0016059921,0.00020695188,0.00033022187,0.0000014393796,0.00009432084,0.08165],"genre_scores_gemma":[0.9892152,0.00020322845,0.00015525921,0.0018582364,0.00021430451,0.0001874251,0.00001037208,0.000008280456,0.008147695],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9993687,0.00001187455,0.00013548696,0.00013358325,0.00015588815,0.0001944682],"domain_scores_gemma":[0.9982171,0.0013101638,0.000026352967,0.00017374105,0.00004119196,0.00023142586],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00014875727,0.00005371022,0.00014529735,0.00002704571,0.00004239147,0.000003007518,0.00003184451,0.00005614221,0.00012308054],"category_scores_gemma":[0.019642983,0.000041381245,0.000061013892,0.00008077239,0.00006558402,0.000014358709,0.000018867731,0.00009534807,0.00012595746],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00028183547,0.0004057551,0.25970265,0.000014264248,0.00005003967,0.0007979195,0.00013815176,0.0000022141562,0.000036286066,0.00095058867,0.12150236,0.61611795],"study_design_scores_gemma":[0.0010340635,0.0000340956,0.98032504,0.000043433014,0.000022927588,0.0001793859,0.0000057575826,0.000012037448,0.00036390824,0.00012351884,0.01780333,0.000052483276],"about_ca_topic_score_codex":0.000025387248,"about_ca_topic_score_gemma":5.472629e-7,"teacher_disagreement_score":0.7206224,"about_ca_system_score_codex":0.000026674774,"about_ca_system_score_gemma":0.00004191903,"threshold_uncertainty_score":0.988615},"labels":[],"label_agreement":null},{"id":"W4300457022","doi":"","title":"Desarrollo del Razonamiento Clínico en Medicina","year":2012,"lang":"es","type":"article","venue":"DOAJ (DOAJ: Directory of Open Access Journals)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"College of Family Physicians of Canada","funders":"","keywords":"Computer science; Philosophy","score_opus":0.23059477791572405,"score_gpt":0.5858651346957339,"score_spread":0.35527035678000984,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4300457022","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7990133,0.17997132,0.00030121944,0.0014350069,0.0043302625,0.0010976455,0.00011055143,0.00007940878,0.013661285],"genre_scores_gemma":[0.9094366,0.08271146,0.0003427228,0.0022426995,0.0031036634,0.000064520886,0.00006370979,0.00016624757,0.0018683675],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99141765,0.0010171962,0.0026728201,0.00089870783,0.002333005,0.0016606231],"domain_scores_gemma":[0.9692044,0.024261236,0.0018812312,0.0013684983,0.0007252795,0.0025593273],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0075785746,0.00090478035,0.002715812,0.0011708416,0.00038475715,0.0010259274,0.0035019466,0.00066124677,0.033898838],"category_scores_gemma":[0.073079966,0.0007541923,0.0007938924,0.0016702247,0.00056557346,0.0022318296,0.0019155919,0.0016780162,0.0006076932],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00074075744,0.002040908,0.81569034,0.00021981385,0.00077077845,0.00015123749,0.0005517347,0.000023973531,0.0056973333,0.00036385068,0.11589615,0.057853118],"study_design_scores_gemma":[0.003522668,0.000093666786,0.93669426,0.012781238,0.0015775227,0.000110120774,0.00031903962,0.00010356375,0.0037970815,0.0013919945,0.03880383,0.0008050303],"about_ca_topic_score_codex":0.0020003968,"about_ca_topic_score_gemma":0.000022150758,"teacher_disagreement_score":0.121003896,"about_ca_system_score_codex":0.0004005309,"about_ca_system_score_gemma":0.00070874614,"threshold_uncertainty_score":0.9994909},"labels":[],"label_agreement":null},{"id":"W4300617518","doi":"","title":"Investigating a self-scoring interview simulation for learning and assessment in the medical consultation","year":2017,"lang":"en","type":"article","venue":"DOAJ (DOAJ: Directory of Open Access Journals)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Self-assessment; Psychology; Medical education; Applied psychology; Medicine; Social psychology","score_opus":0.4126784906179353,"score_gpt":0.6568945478342135,"score_spread":0.2442160572162782,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4300617518","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99073684,0.002676849,0.0016144997,0.0023148188,0.0001966611,0.0008117558,0.0000027981844,0.000019574407,0.0016261851],"genre_scores_gemma":[0.99365354,0.0037550102,0.0017452285,0.0005624464,0.0001580873,0.000069008616,0.000010508551,0.000017573131,0.000028581413],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99794114,0.00030789815,0.0007245688,0.00025472417,0.0005789494,0.00019274108],"domain_scores_gemma":[0.9807777,0.017686807,0.0008323448,0.00024294783,0.00025192727,0.00020831775],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0055357506,0.00014106081,0.0004927489,0.00016194915,0.00036963195,0.0008410301,0.00056368404,0.000098174736,0.0003571832],"category_scores_gemma":[0.12594123,0.00010067231,0.00008547592,0.00015232785,0.00013161074,0.00058223034,0.0003248182,0.0006078094,0.0000010968542],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000053308882,0.00015858418,0.866106,0.00014987134,0.000067718014,0.000027804419,0.0005502905,0.00020437186,0.00037657164,0.000109263936,0.00031271644,0.13188352],"study_design_scores_gemma":[0.002663187,0.000037137255,0.96126896,0.0068785623,0.00012668078,0.000021142912,0.00026860033,0.02486559,0.00011663319,0.002441604,0.0011890525,0.00012287182],"about_ca_topic_score_codex":0.00029972807,"about_ca_topic_score_gemma":0.000038171318,"teacher_disagreement_score":0.13176064,"about_ca_system_score_codex":0.0000508374,"about_ca_system_score_gemma":0.00024195673,"threshold_uncertainty_score":0.8814213},"labels":[],"label_agreement":null},{"id":"W4300772998","doi":"10.1590/1516-3180.2021.1028.r1.13072022","title":"Effectiveness of a SNAPPS in psychiatric residents assessed using objective structured teaching encounters: a case-control study","year":2022,"lang":"en","type":"article","venue":"Sao Paulo Medical Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; University of Toronto","funders":"","keywords":"Medicine; Presentation (obstetrics); Intervention (counseling); Consistency (knowledge bases); Randomized controlled trial; Family medicine; Psychiatry; Nursing","score_opus":0.015281891912309154,"score_gpt":0.36262055834825424,"score_spread":0.3473386664359451,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4300772998","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9958373,0.0004251179,0.0015736454,0.00018075689,0.0009923704,0.0008429487,0.000023956034,0.000020177136,0.000103731545],"genre_scores_gemma":[0.9989254,0.000010442176,0.00038824082,0.00026976486,0.00033640044,0.000031887615,0.0000034328025,0.000029797682,0.000004654648],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.992882,0.0033585504,0.001027127,0.00038068587,0.0019245363,0.0004270794],"domain_scores_gemma":[0.9861321,0.012526396,0.0004261123,0.0002661188,0.00014178586,0.00050751306],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.009226476,0.00022495627,0.0008869133,0.0004841353,0.0003278886,0.00003245904,0.00024913703,0.00013688832,0.00063648075],"category_scores_gemma":[0.041255515,0.0001813018,0.00023947509,0.0005365723,0.00008025824,0.000087553715,0.00013070028,0.0028708824,0.0000010384148],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004607023,0.005843929,0.92730606,0.00012489379,0.0006627958,0.046090443,0.0025806236,0.00049855554,0.0004764324,0.000013097125,0.000126601,0.011669527],"study_design_scores_gemma":[0.024653884,0.004150786,0.9233061,0.0022647218,0.00064750016,0.032592442,0.008448448,0.0029873755,0.000020165558,0.0007081381,0.0000032131195,0.00021718831],"about_ca_topic_score_codex":0.0012260014,"about_ca_topic_score_gemma":0.00012676604,"teacher_disagreement_score":0.032029036,"about_ca_system_score_codex":0.0004915341,"about_ca_system_score_gemma":0.0012150647,"threshold_uncertainty_score":0.9994295},"labels":[],"label_agreement":null},{"id":"W4300787271","doi":"","title":"Internal Medicine residents use heuristics to estimate disease probability","year":2015,"lang":"en","type":"article","venue":"DOAJ (DOAJ: Directory of Open Access Journals)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Heuristics; Computer science; Disease; Medicine; Statistics; Mathematics; Mathematical optimization; Internal medicine","score_opus":0.4269495714128823,"score_gpt":0.6293339094999636,"score_spread":0.20238433808708128,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4300787271","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98692155,0.0040860446,0.0012430141,0.0028083068,0.001220201,0.001000158,0.00006632191,0.00007531499,0.0025790858],"genre_scores_gemma":[0.9930472,0.0011590593,0.0017319813,0.0018807291,0.0005582992,0.00004553188,0.000039811286,0.0000642077,0.0014732121],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9957404,0.0002815845,0.0014043672,0.00063405867,0.0014582593,0.00048129042],"domain_scores_gemma":[0.9888496,0.005529089,0.0006314154,0.00093921187,0.0011367962,0.0029138832],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0026067155,0.0003680116,0.001189254,0.00047787247,0.00010573368,0.0005317805,0.001450928,0.000110953464,0.00285359],"category_scores_gemma":[0.2137794,0.00028078578,0.0001936761,0.0007525963,0.0003069606,0.0008031486,0.0011167951,0.00053307286,0.00007819234],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019694534,0.00060695014,0.8028306,0.00007872237,0.000111676585,0.00042703626,0.00012755532,0.0001752846,0.00041960098,0.00006411707,0.1886335,0.004555549],"study_design_scores_gemma":[0.0020235216,0.00010033771,0.9733668,0.006993876,0.00037512276,0.000035486573,0.00002204025,0.00017797935,0.00038687236,0.012892014,0.0033382378,0.00028770525],"about_ca_topic_score_codex":0.0036345238,"about_ca_topic_score_gemma":0.00007014727,"teacher_disagreement_score":0.2111727,"about_ca_system_score_codex":0.00023038025,"about_ca_system_score_gemma":0.00053181074,"threshold_uncertainty_score":0.9999644},"labels":[],"label_agreement":null},{"id":"W4300866056","doi":"","title":"The Reliability and Validity of a Paediatric Script Concordance Test with Medical Students, Paediatric Residents and Experienced Paediatricians","year":2010,"lang":"en","type":"article","venue":"DOAJ (DOAJ: Directory of Open Access Journals)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; University of Calgary","funders":"","keywords":"Concordance; Test (biology); Reliability (semiconductor); Psychology; Validity; Test validity; Medicine; Family medicine; Pediatrics; Clinical psychology; Medical education; Psychometrics","score_opus":0.09860655417740863,"score_gpt":0.517512119749738,"score_spread":0.4189055655723294,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4300866056","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9848528,0.0124787055,0.00004331877,0.00083454495,0.0006357996,0.0007553037,0.00003463532,0.000027628854,0.00033728158],"genre_scores_gemma":[0.95417607,0.044535495,0.00031716021,0.0003698093,0.00042807232,0.000055043813,0.0000049200294,0.00003391787,0.0000794964],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9950106,0.00032334946,0.0012886835,0.000629832,0.002291197,0.0004563292],"domain_scores_gemma":[0.9673805,0.02932342,0.0010274943,0.0007389047,0.00063396175,0.0008956707],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005574852,0.00033126443,0.0010488614,0.00029997606,0.000339515,0.00054084527,0.0017486889,0.00024243964,0.00064357405],"category_scores_gemma":[0.13335854,0.00020660757,0.00014064877,0.0011562271,0.0010729956,0.0005332573,0.0011525195,0.0010383896,0.000002430908],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004425922,0.0006405091,0.98001176,0.000073903735,0.000058084188,0.00007702859,0.00017965441,0.00000112121,0.0007252925,0.0000184426,0.006893654,0.01087796],"study_design_scores_gemma":[0.0023715703,0.00016696438,0.9936736,0.00068323844,0.0002944751,0.00007325747,0.00019632389,0.00003798466,0.0007678733,0.0010099122,0.00048725842,0.0002375001],"about_ca_topic_score_codex":0.0010469704,"about_ca_topic_score_gemma":0.00020327653,"teacher_disagreement_score":0.12778369,"about_ca_system_score_codex":0.000040268198,"about_ca_system_score_gemma":0.0005766281,"threshold_uncertainty_score":0.87394154},"labels":[],"label_agreement":null},{"id":"W4302313112","doi":"","title":"EXPERIENCED PHYSICIANS BENEFIT FROM ANALYZING INITIAL DIAGNOSTIC HYPOTHESES","year":2013,"lang":"en","type":"article","venue":"Data Archiving and Networked Services (DANS)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Calgary","funders":"","keywords":"Psychology; Medicine; Family medicine","score_opus":0.027261384410501053,"score_gpt":0.30610317129935427,"score_spread":0.2788417868888532,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4302313112","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9958191,0.00088808127,0.00073874026,0.00035250562,0.00023161189,0.0002945375,0.00027783032,0.00013908898,0.0012585017],"genre_scores_gemma":[0.99257016,0.0008224647,0.0019011974,0.0013924658,0.0010611248,0.00005068605,0.0020882941,0.000039300365,0.00007433332],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99802846,0.000077545104,0.00041402777,0.00072711066,0.00026752817,0.0004853308],"domain_scores_gemma":[0.9811088,0.01730278,0.00014131106,0.0010773311,0.0000450331,0.00032473906],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00018500655,0.00027484997,0.00049825734,0.0000613078,0.00022951447,0.00016289276,0.00057237677,0.0000846325,0.00015308522],"category_scores_gemma":[0.0023538333,0.00022429884,0.00007134587,0.0002049335,0.00017168086,0.00031861788,0.00053642166,0.00028248498,0.00008753036],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018711203,0.00050293957,0.541452,0.000118134776,0.00034339554,0.00012789732,0.009173797,0.0003204324,0.0017235838,0.000080205435,0.0016479322,0.4443226],"study_design_scores_gemma":[0.0011598051,0.00019219115,0.91310376,0.005152324,0.00033081346,0.0000189281,0.0025811242,0.07450882,0.00015943398,0.0015687434,0.0008050755,0.00041899458],"about_ca_topic_score_codex":0.0058434526,"about_ca_topic_score_gemma":0.00069484866,"teacher_disagreement_score":0.44390357,"about_ca_system_score_codex":0.000013902868,"about_ca_system_score_gemma":0.000029068899,"threshold_uncertainty_score":0.91466415},"labels":[],"label_agreement":null},{"id":"W4308254372","doi":"10.1007/s10459-022-10175-x","title":"Productive struggle and failing safely: implications for developing adaptive expertise in communication","year":2022,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Holland Bloorview Kids Rehabilitation Hospital; University of Toronto","funders":"","keywords":"Subspecialty; Grounded theory; Conceptual framework; Medical education; Psychology; Medicine; Knowledge management; Qualitative research; Computer science; Sociology","score_opus":0.06334905909458728,"score_gpt":0.45887238957376775,"score_spread":0.3955233304791805,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4308254372","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.88224494,0.04631693,0.0032072398,0.06459853,0.00052682374,0.0024235984,0.000014199542,0.000032351574,0.00063538604],"genre_scores_gemma":[0.952673,0.0024699266,0.04249433,0.0011328094,0.000036497255,0.001123867,0.00003224325,0.000004560661,0.000032726555],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9989206,0.00010261755,0.00035890148,0.00031392538,0.000112267,0.00019169175],"domain_scores_gemma":[0.99725,0.00228995,0.00016724349,0.00016182437,0.00007116351,0.00005976905],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012118901,0.00006103829,0.00015122429,0.00016358189,0.00040143394,0.000010980267,0.000102006234,0.000015597581,0.00000430804],"category_scores_gemma":[0.005271257,0.0000609796,0.000012784074,0.0006188812,0.00014083221,0.00022745045,0.000047503174,0.0001272728,2.7686164e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009189326,0.00037236317,0.23822998,0.00008086759,0.0000017424383,8.1075264e-8,0.006465579,0.0014239496,0.000029115306,0.02971394,0.00016411324,0.7234264],"study_design_scores_gemma":[0.0014684544,0.0013638595,0.8591447,0.0024509162,0.000008735106,0.00003504474,0.03961792,0.0058269114,0.0001459256,0.0818633,0.007765349,0.00030890675],"about_ca_topic_score_codex":0.00045098687,"about_ca_topic_score_gemma":0.00034561596,"teacher_disagreement_score":0.7231175,"about_ca_system_score_codex":0.00046278146,"about_ca_system_score_gemma":0.0014706486,"threshold_uncertainty_score":0.6310567},"labels":[],"label_agreement":null},{"id":"W4308787474","doi":"10.36660/abc.20220405","title":"Dedução, Indução e a Arte do Raciocínio Clínico na Educação Médica: Revisão Sistemática e Proposta Bayesiana","year":2022,"lang":"pt","type":"review","venue":"Arquivos Brasileiros de Cardiologia","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Discovery Air (Canada)","funders":"Universidade Federal de Minas Gerais","keywords":"Deductive reasoning; Analytic reasoning; Inductive reasoning; Probabilistic logic; Psychology of reasoning; Epistemology; Systematic review; Bayes' theorem; Psychology; Cognition; Computer science; Verbal reasoning; Artificial intelligence; Bayesian probability; MEDLINE; Philosophy","score_opus":0.07410000496011111,"score_gpt":0.37036168656902674,"score_spread":0.29626168160891564,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4308787474","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.009708926,0.9625715,0.0005095097,0.00069254305,0.005849847,0.007989687,0.0012723649,0.00072899956,0.010676631],"genre_scores_gemma":[0.094671026,0.88407093,0.000696919,0.0038801695,0.004211075,0.0026828619,0.0027531262,0.0006880653,0.006345817],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9780687,0.005764187,0.0051504234,0.004577031,0.0024041445,0.0040355106],"domain_scores_gemma":[0.9401251,0.047849793,0.0027693287,0.0056161294,0.00054182525,0.0030978194],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"category_scores_codex":[0.008057313,0.003230902,0.011432317,0.00089641014,0.0011394012,0.00044750696,0.0027397925,0.0035926865,0.009436199],"category_scores_gemma":[0.17164578,0.0027793283,0.0061585438,0.0026441615,0.0015780999,0.00022002794,0.002024158,0.008304988,0.0032310276],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007727445,0.002286317,0.0842683,0.017000765,0.007446955,0.006841946,0.001072655,0.00002975478,0.000025628548,0.0011228716,0.10270629,0.7764258],"study_design_scores_gemma":[0.005082013,0.0043157414,0.030597439,0.098129034,0.021191048,0.0075909477,0.0019601162,0.00009658175,0.000012956673,0.0002030103,0.82615983,0.004661278],"about_ca_topic_score_codex":0.000115701085,"about_ca_topic_score_gemma":0.000006216033,"teacher_disagreement_score":0.7717645,"about_ca_system_score_codex":0.0027797974,"about_ca_system_score_gemma":0.007825126,"threshold_uncertainty_score":0.9980418},"labels":[],"label_agreement":null},{"id":"W4309346676","doi":"10.1016/j.cptl.2022.10.011","title":"Pharmaceutical care- urgency: Proposing a practical clinical framework for pharmacy students","year":2022,"lang":"en","type":"article","venue":"Currents in Pharmacy Teaching and Learning","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Sunnybrook Health Science Centre; Health Sciences Centre","funders":"","keywords":"Pharmacy; Experiential learning; Medical education; Focus group; Preparedness; Psychology; Preceptor; Pharmacist; Medicine; Nursing; Pedagogy","score_opus":0.14097664783003452,"score_gpt":0.5619632013375743,"score_spread":0.4209865535075398,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4309346676","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.986681,0.0028311636,0.0038518426,0.0023672713,0.002459091,0.0011797782,0.0000146329585,0.0001902303,0.00042500885],"genre_scores_gemma":[0.984558,0.0005540667,0.012080884,0.0013956926,0.00082402385,0.0002538476,0.00008674929,0.00006678255,0.00017998462],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.9947692,0.0020250916,0.0009380871,0.00081895187,0.000771344,0.00067732565],"domain_scores_gemma":[0.9812061,0.017687483,0.000245316,0.00020957214,0.0000751115,0.00057640095],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0055911248,0.00030412167,0.00063130964,0.00018435155,0.0011678068,0.00009650162,0.0002224199,0.00007085568,0.00030662623],"category_scores_gemma":[0.06362528,0.00029896852,0.0002562779,0.00021531389,0.00014109397,0.00012911292,0.00046449085,0.00816504,0.000011525006],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011760149,0.001762344,0.65222377,0.00011467031,0.000100889,0.000102047255,0.002739602,0.000107514614,0.000026269274,0.00035757103,0.00082114653,0.34046814],"study_design_scores_gemma":[0.037544828,0.0016452844,0.033840682,0.003136948,0.0015099301,0.00051461277,0.0074003837,0.058415428,0.000094900606,0.00075374544,0.8539424,0.0012008562],"about_ca_topic_score_codex":0.00006701429,"about_ca_topic_score_gemma":4.3151925e-7,"teacher_disagreement_score":0.8531213,"about_ca_system_score_codex":0.0001928449,"about_ca_system_score_gemma":0.0003673883,"threshold_uncertainty_score":0.99994624},"labels":[],"label_agreement":null},{"id":"W4309931614","doi":"10.1016/j.jcjq.2022.11.009","title":"Providers’ and Patients’ Perspectives on Diagnostic Errors in the Acute Care Setting","year":2022,"lang":"en","type":"article","venue":"The Joint Commission Journal on Quality and Patient Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Agency for Healthcare Research and Quality; Mallinckrodt Pharmaceuticals; Boston Scientific Corporation","keywords":"Focus group; Thematic analysis; Acute care; Health care; Context (archaeology); Qualitative research; Grounded theory; Nursing; Patient safety; Medicine; Ambulatory care; Psychology; Business","score_opus":0.04527727317213637,"score_gpt":0.34638539912968186,"score_spread":0.3011081259575455,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4309931614","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9558228,0.0008113971,0.0000076049287,0.04152679,0.00016699456,0.0005074212,0.00003561339,0.000011570171,0.0011097976],"genre_scores_gemma":[0.99433035,0.0005045252,0.000041645937,0.004993847,0.000048475395,0.00001768924,0.000016327509,0.000010246626,0.000036881578],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.996772,0.001578675,0.00056688895,0.00021781387,0.0006611012,0.00020355756],"domain_scores_gemma":[0.9837597,0.01535053,0.00032684495,0.00028099265,0.00008931252,0.00019263195],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0017812902,0.00015853593,0.0002842429,0.000073202726,0.0010296493,0.0000392285,0.00010832895,0.000041359872,0.00008813468],"category_scores_gemma":[0.017683107,0.00007977149,0.000105173305,0.00013373308,0.00011686074,0.00003813756,0.0001262263,0.0012104298,0.0000024384676],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008685461,0.004261822,0.30806702,0.000112780486,0.00021275444,0.00032221907,0.38110682,0.0008335256,0.00002134339,0.005625281,0.011860813,0.27889016],"study_design_scores_gemma":[0.0033758483,0.0054362016,0.84922725,0.0013180154,0.000104652754,0.00014331767,0.12950812,0.000056423993,0.000013643405,0.00068840216,0.009948991,0.00017915631],"about_ca_topic_score_codex":0.0000350062,"about_ca_topic_score_gemma":0.0000024862632,"teacher_disagreement_score":0.5411602,"about_ca_system_score_codex":0.0001259956,"about_ca_system_score_gemma":0.00006658134,"threshold_uncertainty_score":0.99059135},"labels":[],"label_agreement":null},{"id":"W4310376777","doi":"10.1007/s10459-022-10181-z","title":"To a fault","year":2022,"lang":"en","type":"letter","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Advancing Health Outcomes; University of British Columbia","funders":"","keywords":"Cognition; Process (computing); Psychology; Analytic reasoning; Cognitive psychology; Epistemology; Cognitive science; Engineering ethics; Computer science; Deductive reasoning; Engineering; Psychiatry; Philosophy","score_opus":0.03483840850970616,"score_gpt":0.4583873754210918,"score_spread":0.42354896691138566,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4310376777","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0017833993,0.010783682,0.000029408522,0.9742197,0.0043213274,0.0008282011,0.000010620932,0.000038082188,0.007985614],"genre_scores_gemma":[0.0017160033,0.0037318787,0.004917319,0.9839579,0.0020851379,0.00040785637,0.00017637401,0.000015170755,0.0029923804],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99756455,0.00012092383,0.00055931206,0.00061744306,0.00066939264,0.0004683618],"domain_scores_gemma":[0.9949478,0.0042317538,0.00025202282,0.0003100393,0.000061981526,0.00019642295],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010750857,0.00014945745,0.00038006363,0.0003999114,0.00019276427,0.00002358554,0.00025464813,0.00012129647,0.0005096984],"category_scores_gemma":[0.014598771,0.00013267466,0.000055420853,0.0011340516,0.00013368033,0.00010842838,0.000053608157,0.0009998847,0.000067308996],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000012398766,0.00012362462,0.0072430368,0.00021230194,0.000001124096,0.000009432742,0.00026223718,0.00016102701,8.458219e-8,0.000117778596,0.8763961,0.11546086],"study_design_scores_gemma":[0.00012124061,0.0008524079,0.0022734923,0.0021749379,0.000004944489,0.000025491558,0.00024988974,0.000024331324,6.37071e-7,0.0011755475,0.99297225,0.00012480267],"about_ca_topic_score_codex":0.00086189416,"about_ca_topic_score_gemma":0.00006773244,"teacher_disagreement_score":0.11657619,"about_ca_system_score_codex":0.00048360057,"about_ca_system_score_gemma":0.004022342,"threshold_uncertainty_score":0.9937017},"labels":[],"label_agreement":null},{"id":"W4310565747","doi":"10.1002/alz.12830","title":"Feasibility of a standard cognitive assessment in European academic memory clinics","year":2022,"lang":"en","type":"article","venue":"Alzheimer s & Dementia","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Douglas Mental Health University Institute","funders":"Deutsches Zentrum für Neurodegenerative Erkrankungen; Alzheimer's Association","keywords":"Unavailability; Reliability (semiconductor); Cognition; Medicine; Psychology; Medical education; Psychiatry","score_opus":0.07457743373379339,"score_gpt":0.4088498539208839,"score_spread":0.3342724201870905,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4310565747","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9694119,0.018624086,0.000119916636,0.0006109422,0.00037672181,0.0007611952,0.000121023746,0.000045246015,0.009928957],"genre_scores_gemma":[0.99786395,0.00015548477,0.00074608694,0.0009901301,0.000060643164,0.000039306557,0.00010393862,0.000025865013,0.000014595898],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9974658,0.00051317504,0.0008013239,0.0003931622,0.00056232593,0.00026422166],"domain_scores_gemma":[0.99615437,0.003030171,0.0002499381,0.00030318898,0.00011075701,0.00015156016],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0027059691,0.00014061578,0.00042061674,0.00009178394,0.0000735884,0.00000529451,0.00014862148,0.00005029838,0.0010730923],"category_scores_gemma":[0.0042109434,0.00013887652,0.0001510844,0.00026381642,0.00012257617,0.000036768033,0.00030198932,0.00085629633,0.000020208903],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000812642,0.00096049655,0.88791853,0.000009091466,0.0026921644,0.000121400495,0.00047904122,0.000028339351,0.000106586485,0.00014523973,0.0018643788,0.1048621],"study_design_scores_gemma":[0.007927794,0.0015415527,0.97422373,0.00042675802,0.011421401,0.000021300651,0.0014227697,0.00014718021,0.0007319466,0.00041841343,0.0014679438,0.00024918487],"about_ca_topic_score_codex":0.000051526997,"about_ca_topic_score_gemma":0.000008151179,"teacher_disagreement_score":0.10461292,"about_ca_system_score_codex":0.00002801376,"about_ca_system_score_gemma":0.0003485117,"threshold_uncertainty_score":0.9998401},"labels":[],"label_agreement":null},{"id":"W4310949926","doi":"10.2196/38427","title":"Developing and Integrating Asynchronous Web-Based Cases for Discussing and Learning Clinical Reasoning: Repeated Cross-sectional Study","year":2022,"lang":"en","type":"article","venue":"JMIR Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Asynchronous communication; Computer science; Asynchronous learning; Cross-sectional study; Psychology; Medicine; Mathematics education; Teaching method; Synchronous learning; Cooperative learning","score_opus":0.04639467073170852,"score_gpt":0.45193106345828965,"score_spread":0.40553639272658115,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4310949926","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99489653,0.000227273,0.00077079795,0.0024120617,0.0006571575,0.0008301922,0.0000035457588,0.000099698445,0.00010272123],"genre_scores_gemma":[0.993074,0.000015235269,0.004208642,0.0011950291,0.0005014648,0.0005832256,0.0001695195,0.000025549421,0.00022732017],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99760973,0.0003682106,0.00073311705,0.0005254827,0.0005350272,0.00022841888],"domain_scores_gemma":[0.98743504,0.011571979,0.00023732956,0.00015538794,0.00018037738,0.00041989452],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0031500368,0.0001478465,0.000333377,0.00010190869,0.00075828156,0.00008707913,0.00006590833,0.00012639383,0.00019962643],"category_scores_gemma":[0.15814224,0.000121589335,0.0000782987,0.00020007277,0.00019582495,0.00005138263,0.00009730126,0.0008836633,0.0000012407706],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00027288974,0.0009911066,0.91548103,0.000038981532,0.000029787449,0.00002077914,0.0004464288,0.0000055623095,0.0000043986756,0.0001457284,0.00029169425,0.0822716],"study_design_scores_gemma":[0.0037883811,0.001801352,0.9762023,0.0006753381,0.00006943808,0.00031648076,0.0037187499,0.011234101,0.0000031832071,0.00006765612,0.001953905,0.00016912207],"about_ca_topic_score_codex":0.00012836867,"about_ca_topic_score_gemma":0.00002308577,"teacher_disagreement_score":0.15499221,"about_ca_system_score_codex":0.00018278995,"about_ca_system_score_gemma":0.0032391176,"threshold_uncertainty_score":0.8489491},"labels":[],"label_agreement":null},{"id":"W4311542542","doi":"10.23970/ahrqepccer258","title":"Diagnostic Errors in the Emergency Department: A Systematic Review","year":2022,"lang":"en","type":"review","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":165,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Emergency department; Confidence interval; Emergency medicine; CINAHL; MEDLINE; Harm; Stroke (engine); Systematic review; Aortic dissection; Myocardial infarction; Internal medicine; Psychological intervention; Psychiatry","score_opus":0.08761665898628994,"score_gpt":0.4176135578297489,"score_spread":0.32999689884345895,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4311542542","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[6.464073e-7,0.9835802,0.0000018784685,0.0004362527,0.00037061982,0.008653897,0.000025967112,0.000051477877,0.0068791136],"genre_scores_gemma":[0.0000013829667,0.9886709,0.000018616454,0.0019299175,0.000086768436,0.006847524,0.00047033743,0.000049246348,0.0019253219],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99513805,0.0011545386,0.0021528525,0.00048089752,0.0007130746,0.0003605853],"domain_scores_gemma":[0.92015254,0.0773541,0.00070191995,0.0015565067,0.00003499442,0.00019993098],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0024714791,0.00044973404,0.004342619,0.00012883163,0.00006767378,0.000011180601,0.0004922943,0.00015001248,0.0066484017],"category_scores_gemma":[0.26152146,0.00021791177,0.0012622296,0.0010047536,0.00002984887,0.00001712756,0.00011756393,0.00081071275,0.0006822945],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[8.4402205e-7,0.000313215,0.000014983943,0.9402459,0.0001384326,0.00042668695,0.000018447026,2.7422145e-8,5.988902e-11,0.00017203868,0.020724494,0.037944898],"study_design_scores_gemma":[0.0000955386,0.00008595694,0.000005264929,0.56489676,0.0077850036,0.00029286544,0.000019661531,3.3998668e-7,4.4984869e-10,0.000014362652,0.42667064,0.00013361262],"about_ca_topic_score_codex":0.000034357163,"about_ca_topic_score_gemma":0.000016689204,"teacher_disagreement_score":0.40594617,"about_ca_system_score_codex":0.00015253786,"about_ca_system_score_gemma":0.00031317622,"threshold_uncertainty_score":0.99425966},"labels":[],"label_agreement":null},{"id":"W4312041072","doi":"","title":"Les tests de concordance de script pour l’évaluation des internes de médecine générale. Une expérience en ligne entre la Faculté de Médecine de Nancy et l’Université de Montréal.","year":2010,"lang":"fr","type":"preprint","venue":"HAL (Le Centre pour la Communication Scientifique Directe)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Humanities; Philosophy","score_opus":0.018545553854453806,"score_gpt":0.28979012459135456,"score_spread":0.2712445707369008,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4312041072","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.837062,0.0057171783,0.1265153,0.019943243,0.00022589846,0.00057299034,0.000101802245,0.00024612958,0.009615486],"genre_scores_gemma":[0.8426546,0.011331292,0.1236929,0.0005668458,0.00015816011,0.000107242464,0.00034325098,0.00009598326,0.021049747],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9867305,0.008992799,0.0010361603,0.0011526981,0.0006617335,0.0014261111],"domain_scores_gemma":[0.95982563,0.03371347,0.00097470375,0.0020234091,0.0021562225,0.0013065845],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.015128245,0.0006970568,0.00092284335,0.00022825706,0.00062690966,0.00029060943,0.0015732477,0.0010118734,0.00046938987],"category_scores_gemma":[0.102934405,0.00077353575,0.0004367493,0.0005911054,0.0009642102,0.00019475045,0.0012503371,0.0019348799,0.000082019586],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00022500067,0.0023889206,0.7378846,0.00027928417,0.0002003332,0.000263851,0.054599166,0.0010099522,0.0615562,0.019523298,0.0023174859,0.1197519],"study_design_scores_gemma":[0.004155035,0.000012706627,0.86860543,0.022318,0.0005204288,0.00068859826,0.001916813,0.033036325,0.04983926,0.014032712,0.0041600056,0.0007146913],"about_ca_topic_score_codex":0.040525805,"about_ca_topic_score_gemma":0.031979553,"teacher_disagreement_score":0.13072082,"about_ca_system_score_codex":0.0027352634,"about_ca_system_score_gemma":0.0028167593,"threshold_uncertainty_score":0.99947155},"labels":[],"label_agreement":null},{"id":"W4313197997","doi":"10.1097/acm.0000000000004918","title":"Modeling Diagnostic Expertise in Cases of Irreducible Uncertainty: The Decision-Aligned Response Model","year":2022,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia; Surrey Memorial Hospital","funders":"","keywords":"Medical diagnosis; Conceptualization; Medicine; Myocardial infarction; Scale (ratio); Pericarditis; Cardiology; Artificial intelligence; Computer science; Pathology","score_opus":0.07634501769727314,"score_gpt":0.39526875690652735,"score_spread":0.3189237392092542,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4313197997","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9736959,0.008014058,0.0025439002,0.014646641,0.00036093738,0.0005479434,0.000013734998,0.000041182106,0.00013573187],"genre_scores_gemma":[0.9942881,0.0011426933,0.00014352195,0.0035557752,0.00018513296,0.00021167858,0.000026618081,0.000031695352,0.00041482048],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99670875,0.0003638408,0.0011031808,0.0004615274,0.0009651401,0.00039756106],"domain_scores_gemma":[0.9352182,0.063571796,0.00016547988,0.000681849,0.00010468258,0.00025800173],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004310462,0.00020931232,0.0006698959,0.00027723555,0.00014595967,0.000002284651,0.00036505508,0.00014364716,0.00036988634],"category_scores_gemma":[0.28833696,0.00013442342,0.000096991505,0.00069394486,0.0002617179,0.00003897662,0.00022439264,0.00084424496,0.0000068552445],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.011355708,0.00026399476,0.007611986,0.000042811538,0.00007070345,0.0009309755,0.01837241,0.90515333,0.004147524,0.00084500096,0.03124341,0.019962145],"study_design_scores_gemma":[0.0052565574,0.001140174,0.00095891865,0.0033700669,0.00019959465,0.00052397675,0.0068912385,0.9745536,0.00016034486,0.0065055178,0.00024390128,0.00019614118],"about_ca_topic_score_codex":0.0009348091,"about_ca_topic_score_gemma":0.0000113712495,"teacher_disagreement_score":0.28402647,"about_ca_system_score_codex":0.00021166765,"about_ca_system_score_gemma":0.00047133386,"threshold_uncertainty_score":0.7176577},"labels":[],"label_agreement":null},{"id":"W4313236596","doi":"10.1111/medu.15014","title":"Identifying validity evidence for uncertainty tolerance scales: A systematic review","year":2022,"lang":"en","type":"review","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":23,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Monash University; Australian Government; University of Victoria","keywords":"Scale (ratio); Psychology; Construct validity; Ambiguity; External validity; Incremental validity; Clinical psychology; Test validity; Criterion validity; Psychometrics; Predictive validity; Internal validity; Population; Applied psychology; Social psychology; Medicine","score_opus":0.253194578719911,"score_gpt":0.5262285176071129,"score_spread":0.27303393888720184,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4313236596","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000001662353,0.9890569,0.0003134098,0.0017831997,0.0022387782,0.006403415,0.000018888872,0.00008051766,0.00010325743],"genre_scores_gemma":[0.0000035686237,0.9809936,0.0003155563,0.007552368,0.0008176072,0.008524269,0.0007681558,0.000053354946,0.00097149704],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9948903,0.0007303621,0.0019171368,0.0007111276,0.0014072077,0.00034390055],"domain_scores_gemma":[0.9003486,0.09410702,0.0019917195,0.002046633,0.0004562612,0.0010497352],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004917528,0.0003863012,0.003727404,0.000118104545,0.00015509044,0.00003690225,0.00047742308,0.00033324517,0.001713903],"category_scores_gemma":[0.62068176,0.00027803116,0.0011684591,0.00056997855,0.000102707454,0.000058476293,0.00010294778,0.0008206873,0.00012750705],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000004946308,0.00027109962,0.000002307429,0.691641,0.000053882562,0.0000054309676,0.00002020211,3.0107625e-8,1.4888583e-9,0.00003125547,0.030462956,0.2775069],"study_design_scores_gemma":[0.000120514946,0.00007726284,0.0000017297128,0.6303961,0.0047739893,0.00015329033,0.00001917504,0.000018880186,9.470284e-9,0.000036424055,0.36424628,0.00015633242],"about_ca_topic_score_codex":0.00004468434,"about_ca_topic_score_gemma":0.000005646107,"teacher_disagreement_score":0.6157642,"about_ca_system_score_codex":0.0006255686,"about_ca_system_score_gemma":0.0076876986,"threshold_uncertainty_score":0.99996716},"labels":[],"label_agreement":null},{"id":"W4313518196","doi":"10.4000/questionsvives.6722","title":"Comment des scientifiques en formation envisagent-ils leur propre participation dans la gestion de la pandémie de Covid-19 ?","year":2022,"lang":"fr","type":"article","venue":"Questions vives recherches en éducation","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Université du Québec à Trois-Rivières","keywords":"Political science; Coronavirus disease 2019 (COVID-19); 2019-20 coronavirus outbreak; Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); Humanities; Philosophy; Medicine; Virology; Infectious disease (medical specialty)","score_opus":0.1646189396121201,"score_gpt":0.46531317769510305,"score_spread":0.30069423808298296,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4313518196","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7425373,0.0027070434,0.18243591,0.06550643,0.0010832852,0.0013503577,0.00015387927,0.0003304889,0.0038952567],"genre_scores_gemma":[0.9694212,0.002914959,0.017284447,0.0014246729,0.00041735993,0.0013172092,0.0008804338,0.000049727463,0.006289987],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.98252904,0.015225996,0.00067670527,0.00047282028,0.0004461565,0.0006492609],"domain_scores_gemma":[0.952229,0.045843918,0.00037700956,0.0005220096,0.00022097894,0.00080706336],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.016720317,0.0002734295,0.0002906328,0.00026850883,0.0008893115,0.00018627699,0.00022835302,0.00044880222,0.00036399896],"category_scores_gemma":[0.16443901,0.0003061597,0.00014876631,0.0005875493,0.00062609103,0.00040714623,0.0001315988,0.001057911,0.000058728667],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003780936,0.008345656,0.093739614,0.0014249247,0.00025711948,0.000069600675,0.52989167,0.028942132,0.015316571,0.07091884,0.024404507,0.22631128],"study_design_scores_gemma":[0.0038428642,0.0019029056,0.5669887,0.0071089705,0.0023234398,0.0010574684,0.064540036,0.08061125,0.012488956,0.0838761,0.17396621,0.0012930774],"about_ca_topic_score_codex":0.003543363,"about_ca_topic_score_gemma":0.00023007693,"teacher_disagreement_score":0.4732491,"about_ca_system_score_codex":0.012301292,"about_ca_system_score_gemma":0.0036712443,"threshold_uncertainty_score":0.999939},"labels":[],"label_agreement":null},{"id":"W4313527990","doi":"10.1051/pmed/2022031","title":"Favoriser le développement du raisonnement clinique des étudiants en médecine : une démarche d’amélioration","year":2022,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"Université de Genève","keywords":"Humanities; Philosophy; Psychology","score_opus":0.07877288414749802,"score_gpt":0.3690228503403709,"score_spread":0.2902499661928729,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4313527990","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5865924,0.028206017,0.008840106,0.36312917,0.006982199,0.0018166328,0.00027355584,0.0002866374,0.0038732896],"genre_scores_gemma":[0.9485696,0.016658645,0.00528917,0.004333595,0.003003503,0.000639562,0.0004643676,0.00012628587,0.020915283],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9939737,0.0010738327,0.0014305833,0.0009845154,0.0013168777,0.0012204904],"domain_scores_gemma":[0.9868403,0.010801079,0.00035068468,0.0008134292,0.00023545134,0.0009590304],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0039474405,0.00050571514,0.00091943855,0.00020788144,0.00092409237,0.00007219675,0.0004913696,0.00028929082,0.0074345684],"category_scores_gemma":[0.025963899,0.00051703065,0.0003916751,0.001123413,0.0006317623,0.00016774907,0.00086881453,0.0012837681,0.0006409911],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00069142017,0.010911099,0.4103913,0.00022097465,0.00045352746,0.0007856643,0.0028734612,0.0007854321,0.00042316358,0.0024708621,0.22322077,0.34677234],"study_design_scores_gemma":[0.017661113,0.0043969364,0.5755713,0.002952267,0.0008740176,0.00014484176,0.009446453,0.0027130742,0.00094599696,0.008179872,0.37597004,0.0011440613],"about_ca_topic_score_codex":0.0006779126,"about_ca_topic_score_gemma":0.00015008837,"teacher_disagreement_score":0.3619772,"about_ca_system_score_codex":0.0011978558,"about_ca_system_score_gemma":0.0020502221,"threshold_uncertainty_score":0.99972814},"labels":[],"label_agreement":null},{"id":"W4315646575","doi":"10.3390/ime2010001","title":"What Do Clinicians Mean by “Good Clinical Judgment”: A Qualitative Study","year":2023,"lang":"en","type":"article","venue":"International Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Toronto East General Hospital; Queen's University; Western University; Joseph Brant Hospital; McMaster University","funders":"","keywords":"Conceptualization; Judgement; Clinical judgement; Excellence; Curiosity; Psychology; Clinical judgment; Nice; Medical education; Clinical Practice; Qualitative research; Medicine; Nursing; Social psychology; Family medicine; Epistemology; Medical physics","score_opus":0.07070865741515937,"score_gpt":0.5418585826838602,"score_spread":0.47114992526870086,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4315646575","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.91504186,0.0003065074,0.00007522576,0.067837566,0.012758759,0.0005713566,0.000016771583,0.00018588427,0.003206096],"genre_scores_gemma":[0.9792894,0.001802785,0.00012054361,0.009242908,0.0018471659,0.000178584,0.0006362908,0.000029347088,0.00685299],"study_design_codex":"not_applicable","study_design_gemma":"qualitative","domain_scores_codex":[0.9959378,0.00041483666,0.0011363615,0.00056031736,0.0016915282,0.00025910672],"domain_scores_gemma":[0.9838866,0.014238926,0.0002593409,0.0003736777,0.00047101002,0.00077047147],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0038829676,0.0001701548,0.00039760626,0.00017377987,0.00006353238,0.00009953941,0.00030892427,0.00023338922,0.0020213202],"category_scores_gemma":[0.11903844,0.00014410815,0.0002043092,0.0003872041,0.00018403935,0.00018634391,0.00010127344,0.00057294976,0.0013835842],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00031003967,0.014317603,0.091823585,0.000014494266,0.00052903855,0.00006783766,0.024477355,7.14093e-7,0.000005194103,0.0017404198,0.45149142,0.41522232],"study_design_scores_gemma":[0.015526963,0.0049631214,0.20690306,0.009665957,0.0006522462,0.00010284718,0.560517,0.001697501,0.000041647007,0.010568116,0.18843342,0.00092811417],"about_ca_topic_score_codex":0.00013374945,"about_ca_topic_score_gemma":0.000038734266,"teacher_disagreement_score":0.53603965,"about_ca_system_score_codex":0.00011770137,"about_ca_system_score_gemma":0.0012826289,"threshold_uncertainty_score":0.99939394},"labels":[],"label_agreement":null},{"id":"W4315698769","doi":"10.37051/mir-00108","title":"La cochrane en médecine d’urgence et en réanimation : activités et adaptation pendant la pandémie COVID-19","year":2022,"lang":"fr","type":"article","venue":"Médecine Intensive Réanimation","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University; McMaster University","funders":"","keywords":"Humanities; Coronavirus disease 2019 (COVID-19); Political science; Philosophy; Medicine","score_opus":0.04751260877184605,"score_gpt":0.39485132680535734,"score_spread":0.3473387180335113,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4315698769","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5127516,0.0075088856,0.13857995,0.32554793,0.005040816,0.0028638253,0.0009073193,0.00056286884,0.0062368126],"genre_scores_gemma":[0.9757814,0.0028635787,0.0019476803,0.014350689,0.00070007663,0.00030952934,0.0016474426,0.00012665293,0.0022729782],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"observational","domain_scores_codex":[0.99014163,0.00493229,0.0015043067,0.001068125,0.0016124537,0.0007411895],"domain_scores_gemma":[0.9248176,0.07146552,0.00093724194,0.0007147462,0.0013103161,0.00075457693],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.007559372,0.000658592,0.0011608097,0.0005261453,0.00041212767,0.00014745696,0.0003456297,0.0004589668,0.0039633703],"category_scores_gemma":[0.23266156,0.0006823782,0.00037800387,0.0009666284,0.000281531,0.00058480055,0.00042030108,0.0019484246,0.00044174946],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.012136955,0.011332687,0.028173039,0.00274513,0.0017828458,0.004509688,0.08867176,0.09757767,0.05036426,0.41084307,0.10426129,0.18760161],"study_design_scores_gemma":[0.033126026,0.007474482,0.4512842,0.011077289,0.0028669762,0.0067594056,0.07978401,0.20464048,0.0011815549,0.022442738,0.17681555,0.002547271],"about_ca_topic_score_codex":0.0025203647,"about_ca_topic_score_gemma":0.00021330819,"teacher_disagreement_score":0.46302977,"about_ca_system_score_codex":0.0019024037,"about_ca_system_score_gemma":0.0021188417,"threshold_uncertainty_score":0.99956274},"labels":[],"label_agreement":null},{"id":"W4317463340","doi":"10.1200/op.22.00551","title":"Impact of an Accelerated Diagnostic Assessment Program on the Timeliness of Cancer Diagnosis and Treatment","year":2023,"lang":"en","type":"article","venue":"JCO Oncology Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Kingston Health Sciences Centre; Queen's University","funders":"","keywords":"Medicine; Referral; Biopsy; Malignancy; Concomitant; Demographics; Medical record; Retrospective cohort study; Cancer; Emergency medicine; Pediatrics; Internal medicine; Family medicine","score_opus":0.13158885496006317,"score_gpt":0.5543047025085833,"score_spread":0.42271584754852015,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4317463340","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9843751,0.00018954164,0.0000014183495,0.012318355,0.0001454714,0.0010841662,0.000052936615,0.000069615286,0.0017633985],"genre_scores_gemma":[0.9916094,0.005458049,0.0007753695,0.00047881732,0.00012755583,0.0013656786,0.000061589184,0.000019782876,0.000103754646],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9984946,0.0003761823,0.0004052196,0.00026878374,0.00022617666,0.00022903235],"domain_scores_gemma":[0.9039868,0.094887234,0.0003604501,0.0003197425,0.00028973105,0.0001560155],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0007683174,0.00016241519,0.00051870185,0.000084967454,0.000061907594,0.00001292939,0.00008214967,0.00015565687,0.0002610495],"category_scores_gemma":[0.063667655,0.0000902769,0.000112183734,0.0004461175,0.00018252785,0.00008145638,0.000044394255,0.00022800703,0.000014460462],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009781644,0.011312249,0.32186893,0.000019332856,0.0007644234,0.00021351165,0.00029973243,0.00019058108,0.00013767194,0.000100135,0.0039049683,0.6602103],"study_design_scores_gemma":[0.0030327921,0.021446558,0.96441025,0.00038881804,0.0010360926,0.000029520237,0.00032475818,0.0017976128,0.0006205965,0.000031654785,0.006782484,0.00009886221],"about_ca_topic_score_codex":0.0018108088,"about_ca_topic_score_gemma":0.00011222402,"teacher_disagreement_score":0.6601114,"about_ca_system_score_codex":0.00023558113,"about_ca_system_score_gemma":0.0007390858,"threshold_uncertainty_score":0.94421947},"labels":[],"label_agreement":null},{"id":"W4317910598","doi":"10.1093/tbm/ibac114","title":"Snakes and ladders: A qualitative study understanding the active ingredients of social interaction around the use of audit and feedback","year":2023,"lang":"en","type":"article","venue":"Translational Behavioral Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Michael's Hospital; Women's College Hospital; Trillium Health Centre; University of Toronto","funders":"","keywords":"Facilitator; Psychology; Focus group; Conversation; Thematic analysis; Qualitative research; Qualitative property; Applied psychology; Action (physics); Audit; Social psychology; Action plan; Medical education; Medicine; Computer science; Sociology","score_opus":0.37243430719992937,"score_gpt":0.48576624242343674,"score_spread":0.11333193522350737,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4317910598","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99059993,0.00004732853,0.00012518515,0.008511551,0.00013624411,0.00047689994,0.000038196944,0.000014606604,0.00005008291],"genre_scores_gemma":[0.9996352,0.000030284835,0.000032545388,0.00007008017,0.00009823898,0.000014928503,0.00004917118,0.000008525684,0.000061013914],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.99881256,0.00015877403,0.00034964044,0.0001558588,0.00042020483,0.00010296828],"domain_scores_gemma":[0.9907189,0.008884855,0.00015569202,0.00008310795,0.000109003835,0.000048446847],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005323401,0.000100024125,0.0002837795,0.0000871047,0.00012346926,0.000007499862,0.0000373156,0.00004664322,0.00004342111],"category_scores_gemma":[0.001598187,0.000054121047,0.000041149113,0.00029009106,0.00066715846,0.000087758155,0.000015934653,0.00018556669,8.313641e-7],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0028937142,0.0011558107,0.15035532,0.00009520483,0.00050362246,0.00002221009,0.8188532,0.000040941813,0.0012489747,0.002993835,0.0021757653,0.01966138],"study_design_scores_gemma":[0.0035601305,0.0013756684,0.60456246,0.00074045185,0.0007311228,0.000009141473,0.38718414,0.00024727496,0.000027561564,0.0014443952,0.00004756869,0.000070079666],"about_ca_topic_score_codex":0.00027816446,"about_ca_topic_score_gemma":0.00009631447,"teacher_disagreement_score":0.45420712,"about_ca_system_score_codex":0.000022287777,"about_ca_system_score_gemma":0.00003717284,"threshold_uncertainty_score":0.24581732},"labels":[],"label_agreement":null},{"id":"W4318203699","doi":"10.3390/jintelligence11020026","title":"Clinical Reasoning: A Missing Piece for Improving Evidence-Based Assessment in Psychology","year":2023,"lang":"en","type":"article","venue":"Journal of Intelligence","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Conceptualization; Process (computing); Psychology; Psychological testing; Clinical judgment; Applied psychology; Management science; Computer science; Clinical psychology; Artificial intelligence; Medicine","score_opus":0.19849592711459815,"score_gpt":0.5305744018488668,"score_spread":0.33207847473426866,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4318203699","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6507211,0.0018805987,0.3213384,0.023063691,0.0020694502,0.00051197014,0.0000027312865,0.000055461234,0.00035665688],"genre_scores_gemma":[0.9351259,0.00091973227,0.061223544,0.0021083022,0.0005169248,0.000009862592,0.0000020129994,0.000019623976,0.00007410397],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9974465,0.0001456791,0.0014838281,0.0002618281,0.00034383114,0.00031832786],"domain_scores_gemma":[0.9695729,0.028855415,0.0006093316,0.0002563643,0.00042273515,0.00028326543],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005892998,0.00012589592,0.000570647,0.00028381456,0.00004080833,0.000028577379,0.0002092008,0.00016489998,0.000047255508],"category_scores_gemma":[0.16256292,0.000098217155,0.0003492134,0.00049356674,0.00011093694,0.00009518058,0.0000295799,0.00070806243,0.00001873955],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00106435,0.00052705494,0.236772,0.00009629554,0.000040724983,0.000504062,0.00007550208,0.00036626574,0.00083607726,0.000219564,0.0025623422,0.7569358],"study_design_scores_gemma":[0.005065702,0.01186987,0.80441755,0.043929473,0.00046148108,0.0006469823,0.0006277686,0.11837006,0.003108128,0.007712815,0.003266117,0.0005240373],"about_ca_topic_score_codex":0.00001572168,"about_ca_topic_score_gemma":0.000003654523,"teacher_disagreement_score":0.75641173,"about_ca_system_score_codex":0.0000902375,"about_ca_system_score_gemma":0.0008994034,"threshold_uncertainty_score":0.8444912},"labels":[],"label_agreement":null},{"id":"W4318212280","doi":"10.7202/1095681ar","title":"Cultural adaptation and assessment of validity evidence for scores obtained using a French version of the Tolerance of Ambiguity in Medical Students and Doctors scale","year":2021,"lang":"en","type":"article","venue":"Mesure et évaluation en éducation","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Agence Nationale de la Recherche","keywords":"Ambiguity; Intraclass correlation; Scale (ratio); Reliability (semiconductor); Psychology; Ambiguity tolerance; Test (biology); Adaptation (eye); Internal consistency; Clinical psychology; Medical education; Psychometrics; Applied psychology; Social psychology; Medicine; Computer science; Geography","score_opus":0.24282450460983504,"score_gpt":0.5198436519154241,"score_spread":0.2770191473055891,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4318212280","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99627197,0.0002218799,0.001713493,0.0012493543,0.00016982286,0.00035014952,0.0000065926433,0.0000025645793,0.000014190546],"genre_scores_gemma":[0.9958627,0.00024404173,0.0037653875,0.000045889054,0.000022686196,0.000010960731,0.000032247546,0.0000030632455,0.000012978646],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99805075,0.0006642636,0.000385713,0.00015257821,0.0006913125,0.00005540321],"domain_scores_gemma":[0.9943727,0.0047254707,0.00024199553,0.00012542252,0.0004976465,0.00003676669],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005098714,0.0000568258,0.00020802282,0.000035748453,0.000025339397,0.0000061580954,0.000050421593,0.00006976898,0.000015650654],"category_scores_gemma":[0.04072616,0.000042183652,0.000039429335,0.00017714182,0.00006920062,0.00008486343,0.000038385322,0.00008092594,3.3630343e-8],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004653868,0.00025268365,0.9851487,0.00015064793,0.000012027723,1.6297238e-7,0.0066484855,0.0002881581,0.0040291203,0.000029951616,0.000042572417,0.003350902],"study_design_scores_gemma":[0.0012612115,0.000101580794,0.96743417,0.002824939,0.00010648591,0.0000026039115,0.0006264866,0.023702435,0.003717857,0.00018662713,0.0000020369896,0.000033584653],"about_ca_topic_score_codex":0.00041114024,"about_ca_topic_score_gemma":0.00027739757,"teacher_disagreement_score":0.035627443,"about_ca_system_score_codex":0.00007031816,"about_ca_system_score_gemma":0.0007866392,"threshold_uncertainty_score":0.96735424},"labels":[],"label_agreement":null},{"id":"W4318274797","doi":"10.1016/j.pec.2023.107650","title":"Patient generated research priorities to improve diagnostic safety: A systematic prioritization exercise","year":2023,"lang":"en","type":"article","venue":"Patient Education and Counseling","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute for Work & Health; Toronto East General Hospital; University of Toronto","funders":"VA National Center for Patient Safety; Agency for Healthcare Research and Quality; Gordon and Betty Moore Foundation","keywords":"Prioritization; Ranking (information retrieval); Identification (biology); Medicine; MEDLINE; Diagnostic accuracy; Psychology; Applied psychology; Medical education; Computer science; Management science; Artificial intelligence; Political science","score_opus":0.028475485110185095,"score_gpt":0.35685650084352083,"score_spread":0.3283810157333357,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4318274797","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.992151,0.0027165236,0.00015557809,0.0011501423,0.0010040598,0.002314623,0.000013864697,0.00015981686,0.00033434748],"genre_scores_gemma":[0.9946609,0.002574123,0.00037723317,0.0006961969,0.00017544726,0.00073949125,0.00014214145,0.000030643783,0.0006038047],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99786323,0.00014580619,0.00060823926,0.00039706862,0.0006359232,0.00034974155],"domain_scores_gemma":[0.9909684,0.0072667534,0.00010205897,0.00031195706,0.0009789793,0.00037186363],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.000705413,0.00014969325,0.0003429999,0.00035061294,0.00027158362,0.000102114216,0.00004912071,0.00009009017,0.000024705047],"category_scores_gemma":[0.04459389,0.00013271793,0.000047573827,0.0009077682,0.000045225886,0.00005906902,0.000044928645,0.00018615105,0.000232565],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00054417574,0.0021109292,0.017101005,0.012870679,0.00012751517,0.00005192034,0.06605909,0.0006643986,0.0011743119,0.0011192191,0.029812995,0.86836374],"study_design_scores_gemma":[0.006777755,0.011270322,0.047492668,0.55560577,0.001982853,0.00020189027,0.2848487,0.054718856,0.0073558334,0.0057039754,0.020502528,0.0035388519],"about_ca_topic_score_codex":0.000104775216,"about_ca_topic_score_gemma":0.0000069877583,"teacher_disagreement_score":0.8648249,"about_ca_system_score_codex":0.00014597925,"about_ca_system_score_gemma":0.0005813847,"threshold_uncertainty_score":0.9634539},"labels":[],"label_agreement":null},{"id":"W4319454867","doi":"10.1101/2023.02.04.23285478","title":"ChatGPT for Clinical Vignette Generation, Revision, and Evaluation","year":2023,"lang":"en","type":"preprint","venue":"medRxiv","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":82,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta Hospital; University of Alberta","funders":"","keywords":"Vignette; Rewriting; Triage; Set (abstract data type); Medicine; Perspective (graphical); Literacy; Psychology; Psychiatry; Artificial intelligence; Computer science; Social psychology; Programming language","score_opus":0.25384643831666226,"score_gpt":0.49229167596309076,"score_spread":0.2384452376464285,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4319454867","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9763666,0.0029349672,0.002564113,0.012669358,0.0029877266,0.0021464396,0.00004578567,0.00015052153,0.00013453717],"genre_scores_gemma":[0.95762545,0.013898486,0.00913489,0.0035837889,0.00932608,0.0007272733,0.0024583342,0.00014506717,0.0031006306],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9973259,0.00021833809,0.0009408001,0.00080731226,0.0004943525,0.0002133111],"domain_scores_gemma":[0.9887364,0.009330944,0.00029258803,0.0006902917,0.00067154755,0.0002782171],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.006236599,0.00021527446,0.00067022344,0.00008944834,0.00007643037,0.00004968935,0.00009897412,0.0004978248,0.00008473972],"category_scores_gemma":[0.15782814,0.00017904214,0.0002618621,0.00009112646,0.00009558378,0.00001795727,0.00020526553,0.00046980925,0.00008280275],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021097335,0.00035963228,0.4353725,0.00046215235,0.00034197015,0.00002877732,0.00022547263,0.00011568071,0.00009936231,0.0002824083,0.29394507,0.268556],"study_design_scores_gemma":[0.0053597474,0.0008086019,0.8406899,0.0062095104,0.0021979187,0.000018447816,0.000023518283,0.10068965,0.00013197435,0.0063104206,0.03701798,0.00054232596],"about_ca_topic_score_codex":0.000014780438,"about_ca_topic_score_gemma":0.000009007306,"teacher_disagreement_score":0.40531743,"about_ca_system_score_codex":0.000037681137,"about_ca_system_score_gemma":0.00034757427,"threshold_uncertainty_score":0.8492658},"labels":[],"label_agreement":null},{"id":"W4319457821","doi":"10.1164/rccm.202302-0241ed","title":"Introducing “Viewpoint: Turning the Air Blue”","year":2023,"lang":"en","type":"letter","venue":"American Journal of Respiratory and Critical Care Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Muscular Dystrophy Canada; University of Toronto; St. Michael's Hospital","funders":"","keywords":"Medicine; Intensive care medicine","score_opus":0.027213032880384985,"score_gpt":0.35463800196001694,"score_spread":0.3274249690796319,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4319457821","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.033169318,0.016135305,0.00016814709,0.9483456,0.0015663045,0.00017497774,0.000012989591,0.00004724416,0.00038013508],"genre_scores_gemma":[0.09899557,0.0006031176,0.00020986129,0.8733169,0.026615432,0.000008510856,0.000018406325,0.000088483874,0.00014374759],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9961196,0.00053378165,0.0012287736,0.00043481265,0.0011480695,0.00053493393],"domain_scores_gemma":[0.9647855,0.032655008,0.0005533319,0.000524513,0.0010036184,0.00047799703],"candidate_categories":["metaresearch","sts","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0017933528,0.00039778094,0.0016904752,0.00030659532,0.00018618739,0.00002577907,0.00028057644,0.0002989286,0.000118173746],"category_scores_gemma":[0.12099062,0.00021791161,0.00034486866,0.00051524024,0.0030848593,0.000049433544,0.000094070245,0.0048628203,0.00002223952],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00019018099,0.000026467089,0.0015489608,0.0004146208,0.00019727477,0.006962087,0.0014606032,0.000004028816,0.000035926114,0.000069685004,0.8569936,0.13209656],"study_design_scores_gemma":[0.0016209644,0.012595925,0.0045470716,0.015967472,0.0023483606,0.0022546295,0.009639598,0.0000054138586,0.000035764737,0.00022737768,0.9503953,0.0003621544],"about_ca_topic_score_codex":0.000038613514,"about_ca_topic_score_gemma":0.0000019397557,"teacher_disagreement_score":0.1317344,"about_ca_system_score_codex":0.00009202771,"about_ca_system_score_gemma":0.00039256684,"threshold_uncertainty_score":0.9996282},"labels":[],"label_agreement":null},{"id":"W4319460520","doi":"10.53738/revmed.2023.19.813.264","title":"Pédagogie de l’incertitude en médecine","year":2023,"lang":"fr","type":"article","venue":"Revue Médicale Suisse","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Humanities; Philosophy; Political science","score_opus":0.046021824258455954,"score_gpt":0.37726493699871355,"score_spread":0.3312431127402576,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4319460520","genre_codex":"empirical","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5646771,0.17168482,0.00093035016,0.22628757,0.012627621,0.0011274004,0.00021420613,0.0013438636,0.02110706],"genre_scores_gemma":[0.27020028,0.032768656,0.0015474095,0.0034860023,0.014552603,0.00009693311,0.00024199905,0.00016942709,0.6769367],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99625367,0.00030185032,0.0008121062,0.00069059647,0.00047710715,0.0014646782],"domain_scores_gemma":[0.97285837,0.024616145,0.00013201669,0.00085288816,0.000118242664,0.0014223227],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0020773194,0.00036421322,0.0008306176,0.00019348419,0.00013721523,0.00005535395,0.0003666577,0.0005650922,0.0036248376],"category_scores_gemma":[0.1390374,0.00036457548,0.00040259882,0.0010802075,0.00040023995,0.00006985063,0.00024334602,0.00094065565,0.014429053],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00022344897,0.0018379103,0.1546172,0.0009004461,0.00035915792,0.007880554,0.0008147499,0.00038644412,0.00096971163,0.0046745376,0.36137757,0.46595827],"study_design_scores_gemma":[0.0045502316,0.0008425582,0.4469046,0.020422304,0.0009991984,0.00050407415,0.00035768337,0.006484053,0.0004846745,0.0061994116,0.5115582,0.0006930553],"about_ca_topic_score_codex":0.0002230344,"about_ca_topic_score_gemma":0.000026731443,"teacher_disagreement_score":0.65582967,"about_ca_system_score_codex":0.00042189652,"about_ca_system_score_gemma":0.00071624795,"threshold_uncertainty_score":0.9998806},"labels":[],"label_agreement":null},{"id":"W4319460877","doi":"10.1093/bjs/znad004","title":"Cognitive biases in surgery: systematic review","year":2023,"lang":"en","type":"review","venue":"British journal of surgery","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":45,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Milton District Hospital; St. Michael's Hospital; North York General Hospital; University of Toronto","funders":"","keywords":"PsycINFO; Cognitive bias; Medicine; Cognition; MEDLINE; Meta-analysis; Publication bias; Overconfidence effect; Cochrane Library; Anchoring; Health care; Systematic review; Clinical psychology; Psychiatry; Psychology; Social psychology","score_opus":0.25691563641796245,"score_gpt":0.43295455221907053,"score_spread":0.1760389158011081,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4319460877","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0001029473,0.9976926,0.0000011279187,0.00008129873,0.0008856233,0.0010483798,0.00009824643,0.000032879143,0.000056902638],"genre_scores_gemma":[0.000021404623,0.99833536,0.0000072486146,0.0005165681,0.00046868285,0.00008610837,0.00019197725,0.00010379887,0.00026883304],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9922159,0.0010260723,0.005182239,0.00030755662,0.00079836405,0.0004698818],"domain_scores_gemma":[0.68135476,0.31427756,0.0031437662,0.00023587553,0.0005345953,0.00045342962],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.009594674,0.00037398498,0.01245204,0.00081519113,0.00003386624,0.00006764386,0.0001403124,0.0003222772,0.00014115767],"category_scores_gemma":[0.72289044,0.00032248028,0.00366785,0.0013546223,0.000089160356,0.000085646745,0.000048534916,0.001137818,0.00021348416],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000068940526,0.00016286642,0.000108132655,0.62313724,0.00041106253,0.02762667,0.0000017762441,1.5696967e-8,5.6338473e-10,1.900732e-7,0.033112288,0.31543288],"study_design_scores_gemma":[0.00016051468,0.00002515414,0.00022194885,0.9464638,0.007944409,0.03744138,0.000016225542,1.497311e-7,2.4995751e-8,0.0000058518103,0.007466066,0.000254432],"about_ca_topic_score_codex":0.000063893705,"about_ca_topic_score_gemma":0.000010363715,"teacher_disagreement_score":0.71329576,"about_ca_system_score_codex":0.00015161015,"about_ca_system_score_gemma":0.0023633814,"threshold_uncertainty_score":0.99992275},"labels":[],"label_agreement":null},{"id":"W4319723444","doi":"10.1186/s41235-022-00460-z","title":"The challenge of cognitive science for medical diagnosis","year":2023,"lang":"en","type":"article","venue":"Cognitive Research Principles and Implications","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":33,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Cognition; Process (computing); Judgement; Yardstick; Psychology; Heuristics; Confirmation bias; Management science; Cognitive psychology; Computer science; Social psychology; Epistemology; Engineering","score_opus":0.2709228380340374,"score_gpt":0.5347549932407556,"score_spread":0.2638321552067182,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4319723444","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.90356237,0.0021144804,0.00068928965,0.07544534,0.000067297966,0.0028625291,0.00035958845,0.000084464424,0.014814617],"genre_scores_gemma":[0.9850286,0.012527289,0.000033604574,0.0001038689,0.00009819458,0.0018291506,0.000052361087,0.000011345184,0.0003155893],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99813116,0.0000772389,0.00028495368,0.00034930292,0.00069540116,0.00046193218],"domain_scores_gemma":[0.83731794,0.15844916,0.00008208895,0.00030742935,0.0032301918,0.0006131918],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.004221427,0.000073257674,0.00017401653,0.00018498546,0.00088425545,0.0000406239,0.0001829973,0.00006846929,0.00003513463],"category_scores_gemma":[0.4080109,0.00004792537,0.00006595643,0.0010185273,0.00283026,0.000043002907,0.0003107892,0.0002643563,0.00003450157],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00026994472,0.000681945,0.07224992,0.000104627754,0.00012603396,0.000007710848,0.00078275794,1.3187498e-7,0.00014274176,0.11665894,0.0019749042,0.80700034],"study_design_scores_gemma":[0.0017224052,0.0009939639,0.973627,0.0029955392,0.00006700736,0.000011524565,0.0023727994,0.0004888008,0.0013458132,0.011504276,0.004773252,0.000097628625],"about_ca_topic_score_codex":0.000018931329,"about_ca_topic_score_gemma":0.000021512795,"teacher_disagreement_score":0.9013771,"about_ca_system_score_codex":0.00002419427,"about_ca_system_score_gemma":0.0008707731,"threshold_uncertainty_score":0.9998835},"labels":[],"label_agreement":null},{"id":"W4320023223","doi":"10.36834/cmej.75843","title":"Clinical Reasoning in CanMEDS 2025","year":2022,"lang":"en","type":"article","venue":"Canadian Medical Education Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"McMaster University; University of Saskatchewan; NOSM University; Queen's University; McGill University","funders":"Royal College of Physicians and Surgeons of Canada","keywords":"Computer science","score_opus":0.025026423766792694,"score_gpt":0.39670783267002535,"score_spread":0.37168140890323265,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4320023223","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.79438347,0.002053499,0.000022460606,0.16672097,0.008153166,0.00026756595,0.000009405438,0.000023991208,0.028365448],"genre_scores_gemma":[0.8840826,0.00033885433,0.00066841155,0.10916754,0.001990673,0.000041575553,0.000043665146,0.000021923637,0.0036447523],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.997477,0.00030962646,0.00073601014,0.00022500538,0.00082995865,0.00042238014],"domain_scores_gemma":[0.98014635,0.0024841041,0.0001872892,0.00032550746,0.00021615255,0.016640572],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0034267483,0.000098536446,0.00030015747,0.0003673051,0.00026540313,0.000027922148,0.00021801054,0.00014758087,0.07881912],"category_scores_gemma":[0.3116081,0.000095216295,0.00013494928,0.00046037268,0.000107660366,0.00003614068,0.000035185625,0.0023623651,0.00006849818],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000037146063,0.00042103708,0.28958315,0.0000039743536,0.00002149254,0.0008403592,0.00023787946,0.000008224887,2.2112879e-7,0.0006444465,0.45288813,0.25531393],"study_design_scores_gemma":[0.0020209625,0.00031895234,0.32790437,0.00097175234,0.000046525856,0.004390049,0.0018619895,0.00045874316,4.089355e-7,0.00046568268,0.66139287,0.00016770241],"about_ca_topic_score_codex":0.017701808,"about_ca_topic_score_gemma":0.015819144,"teacher_disagreement_score":0.30818135,"about_ca_system_score_codex":0.00085006567,"about_ca_system_score_gemma":0.11003549,"threshold_uncertainty_score":0.9999392},"labels":[],"label_agreement":null},{"id":"W4320075904","doi":"10.2196/38941","title":"Using Clinical Data Visualizations in Electronic Health Record User Interfaces to Enhance Medical Student Diagnostic Reasoning: Randomized Experiment","year":2023,"lang":"en","type":"article","venue":"JMIR Human Factors","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"University of Pittsburgh","keywords":"Heuristics; Medical diagnosis; Cognition; Medical record; Medical education; Psychology; Visualization; Process (computing); Computer science; Medicine; Artificial intelligence; Psychiatry","score_opus":0.13232122373258198,"score_gpt":0.5630996249995397,"score_spread":0.4307784012669577,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4320075904","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9941004,0.00037389598,0.0009809133,0.0016750846,0.00055293355,0.0020047103,0.000016393087,0.00022699412,0.000068655565],"genre_scores_gemma":[0.9964153,0.0009653376,0.00021200156,0.0010956044,0.00027835777,0.00020421502,0.00044608043,0.000057124067,0.00032596829],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99479985,0.00077326404,0.0017082257,0.0009424759,0.00096869003,0.0008074988],"domain_scores_gemma":[0.97476214,0.022945521,0.0003066551,0.001023208,0.000084419386,0.0008780458],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004358759,0.00030297443,0.0013948268,0.0003373873,0.00015332954,0.00006555869,0.000578408,0.00021603199,0.0005388251],"category_scores_gemma":[0.10872547,0.00024488874,0.00018040555,0.0007765158,0.00019218786,0.0000966391,0.0006184616,0.00065785623,0.00012911436],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0050276504,0.004875204,0.930723,0.000100561054,0.0005202323,0.00021432263,0.0075766253,0.00011518798,0.000100479,0.0016871806,0.035993364,0.013066162],"study_design_scores_gemma":[0.15806928,0.005569601,0.77477527,0.026246557,0.0005052977,0.000029518587,0.0061336337,0.010254382,0.0008110442,0.0005106845,0.015059313,0.0020354313],"about_ca_topic_score_codex":0.000658915,"about_ca_topic_score_gemma":0.0005896716,"teacher_disagreement_score":0.15594777,"about_ca_system_score_codex":0.00034476237,"about_ca_system_score_gemma":0.00074726745,"threshold_uncertainty_score":0.99862725},"labels":[],"label_agreement":null},{"id":"W4321460978","doi":"10.22605/rrh8099","title":"The Prince Edward Island way - does it work?","year":2023,"lang":"en","type":"review","venue":"Rural and Remote Health","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Audit; Medicine; Welfare; Medical diagnosis; Health care; Nursing; District nurse; Family medicine; Psychology","score_opus":0.05415224863341492,"score_gpt":0.4088137105326357,"score_spread":0.35466146189922076,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4321460978","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000024325971,0.9808677,0.000007759809,0.016645169,0.00088219304,0.0008385713,0.000024604271,0.00009201609,0.00061764964],"genre_scores_gemma":[0.0000041337066,0.9729216,0.00009555523,0.0009997871,0.00085798133,0.000006295084,0.00008129937,0.00004315971,0.024990166],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99796283,0.00016786075,0.00076376845,0.0003337394,0.00027419924,0.0004975847],"domain_scores_gemma":[0.9888303,0.009958358,0.0003104341,0.0004125292,0.00003664105,0.00045169852],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.000950424,0.00029478484,0.001512085,0.00004801239,0.00031205098,0.00004892872,0.0001282356,0.00025641892,0.000016982516],"category_scores_gemma":[0.0063537722,0.00011377869,0.0003148584,0.00028245576,0.00014727788,0.000011228017,0.00010858188,0.00070338807,0.00022676529],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000015171552,0.000015005737,0.000015604592,0.005077844,0.00006508176,0.000022800272,0.00003891734,1.654749e-8,1.0395594e-9,0.000042668133,0.013419744,0.9812871],"study_design_scores_gemma":[0.0002173605,0.00013082773,0.00023984193,0.13125992,0.00018766196,0.00007633656,0.000033160228,0.0000041373896,9.089175e-9,0.00017140515,0.867552,0.00012732377],"about_ca_topic_score_codex":0.00029795867,"about_ca_topic_score_gemma":0.00022407396,"teacher_disagreement_score":0.9811598,"about_ca_system_score_codex":0.00007803676,"about_ca_system_score_gemma":0.0005661375,"threshold_uncertainty_score":0.76065177},"labels":[],"label_agreement":null},{"id":"W4323030608","doi":"10.1016/j.jfop.2023.100005","title":"Conversational AI Models for ophthalmic diagnosis: Comparison of ChatGPT and the Isabel Pro Differential Diagnosis Generator","year":2023,"lang":"en","type":"article","venue":"JFO Open Ophthalmology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":116,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta; University of Toronto","funders":"","keywords":"Medical diagnosis; Differential diagnosis; Generator (circuit theory); Medicine; Set (abstract data type); Differential (mechanical device); Computer science; Artificial intelligence; Medical physics; Radiology; Pathology; Engineering","score_opus":0.11807619070082148,"score_gpt":0.42541467460019156,"score_spread":0.3073384838993701,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4323030608","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9831918,0.0005739789,0.00013058577,0.012214925,0.0005500677,0.0025016814,0.00015394596,0.00003491473,0.0006481015],"genre_scores_gemma":[0.99414724,0.0002159083,0.00040109272,0.0005950711,0.00023811191,0.003462532,0.00024488504,0.000036983653,0.00065818586],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99800026,0.00018016368,0.00067414,0.00049962,0.00027002927,0.00037576162],"domain_scores_gemma":[0.9833765,0.015578035,0.0002394881,0.00040743832,0.00021841239,0.00018015083],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0006681544,0.00022672334,0.0009992836,0.00010360001,0.00016736273,0.00005645,0.00036168392,0.00023636097,0.00045281948],"category_scores_gemma":[0.008476087,0.00015615881,0.00019081391,0.00024037893,0.00053642294,0.00010282645,0.00039526654,0.0002312725,0.000031491316],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0036644854,0.0019937058,0.9183761,0.00024347528,0.0007068925,0.00013795785,0.0016612569,0.00043457779,0.00011446954,0.023716236,0.044117067,0.0048337784],"study_design_scores_gemma":[0.06288207,0.0057961415,0.6649717,0.0021849747,0.0020090046,0.0015453249,0.001502461,0.19270875,0.0066130795,0.056681156,0.0019517266,0.0011535845],"about_ca_topic_score_codex":0.00032239643,"about_ca_topic_score_gemma":0.000003885962,"teacher_disagreement_score":0.25340438,"about_ca_system_score_codex":0.0000307958,"about_ca_system_score_gemma":0.00017540876,"threshold_uncertainty_score":0.99987596},"labels":[],"label_agreement":null},{"id":"W4323653269","doi":"10.2196/43988","title":"Proposal of a Method for Transferring High-Quality Scientific Literature Data to Virtual Patient Cases Using Categorical Data Generated by Bernoulli-Distributed Random Values: Development and Prototypical Implementation","year":2023,"lang":"en","type":"article","venue":"JMIR Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Virtual patient; Categorical variable; Computer science; USable; Nausea; Medicine; Medical physics; Psychiatry; Machine learning; Surgery","score_opus":0.1122308950662889,"score_gpt":0.4949122097877513,"score_spread":0.3826813147214624,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4323653269","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9514286,0.0001679884,0.03897883,0.0041732322,0.0004302304,0.0032789116,0.0014812037,0.000060015856,9.869527e-7],"genre_scores_gemma":[0.85227376,0.000018868584,0.05344335,0.0005581668,0.00027869272,0.00080662814,0.09253699,0.000030617506,0.000052936204],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9969679,0.00024719568,0.00086103857,0.0008001844,0.00082923524,0.00029440876],"domain_scores_gemma":[0.9948635,0.0033696515,0.00013834411,0.00068949576,0.00035815395,0.00058081874],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0026019765,0.00016951613,0.00041468407,0.00013350032,0.00017572242,0.00007861355,0.00026931142,0.0001780874,0.0000479955],"category_scores_gemma":[0.03031248,0.000134851,0.000033048498,0.00075418927,0.00010512326,0.0001394976,0.00021413308,0.00019994455,0.0000026122989],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0035983466,0.0038868007,0.003287437,0.0008714036,0.0002656486,0.00003155126,0.00782754,0.000017724573,0.012784801,0.0014727808,0.1358307,0.8301253],"study_design_scores_gemma":[0.10028357,0.012375085,0.13820592,0.02677256,0.0038830724,0.0011898943,0.028977722,0.21556905,0.14580798,0.0032581757,0.31857497,0.0051020198],"about_ca_topic_score_codex":0.00021272324,"about_ca_topic_score_gemma":0.0000413848,"teacher_disagreement_score":0.82502323,"about_ca_system_score_codex":0.000099276265,"about_ca_system_score_gemma":0.0033339194,"threshold_uncertainty_score":0.9778556},"labels":[],"label_agreement":null},{"id":"W4323653652","doi":"10.1186/s41235-023-00472-3","title":"Learning strategy impacts medical diagnostic reasoning in early learners","year":2023,"lang":"en","type":"article","venue":"Cognitive Research Principles and Implications","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University Health Centre; University of Toronto; McGill University","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Medical diagnosis; Test (biology); Psychology; Schizophrenia (object-oriented programming); Task (project management); Transfer of learning; Artificial intelligence; Cognitive psychology; Natural language processing; Medicine; Computer science; Developmental psychology; Psychiatry","score_opus":0.16778733423100403,"score_gpt":0.48355232576966894,"score_spread":0.3157649915386649,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4323653652","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9841589,0.00047290736,0.000047850306,0.0046324925,0.000019264036,0.00057845825,0.0000140453185,0.0001259013,0.009950191],"genre_scores_gemma":[0.9929252,0.0053315666,0.000032544984,0.00010952197,0.000115684445,0.00026900053,0.00013587618,0.000027334241,0.0010532491],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9974805,0.00030437612,0.00036419832,0.00047377506,0.0006457637,0.0007313934],"domain_scores_gemma":[0.94216436,0.056140948,0.00006070102,0.00023478814,0.00047307476,0.00092612515],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002658402,0.0001326258,0.0002754963,0.00043082674,0.00031530825,0.00008503133,0.00011674724,0.00017591995,0.00013867319],"category_scores_gemma":[0.38609746,0.00011687868,0.000060524468,0.0013827113,0.000447141,0.00008459028,0.00025182124,0.001242302,0.00027914322],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000062795705,0.00018332325,0.9037318,0.000036666213,0.000037021433,0.00017697478,0.0006385651,0.000018067201,0.00010432695,0.0032482503,0.00036281825,0.091399394],"study_design_scores_gemma":[0.0012359235,0.00059285836,0.99057543,0.0030441612,0.000020307876,0.00005319196,0.0021188725,0.000994265,0.000044925546,0.00065994635,0.0005417331,0.00011836286],"about_ca_topic_score_codex":0.00035664893,"about_ca_topic_score_gemma":0.0001014703,"teacher_disagreement_score":0.38343906,"about_ca_system_score_codex":0.000068146735,"about_ca_system_score_gemma":0.0006447854,"threshold_uncertainty_score":0.6190737},"labels":[],"label_agreement":null},{"id":"W4327660727","doi":"10.2196/42552","title":"Home Health Care Clinicians’ Use of Judgment Language for Black and Hispanic Patients: Natural Language Processing Study","year":2023,"lang":"en","type":"article","venue":"JMIR Nursing","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Health care; Medicine; Documentation; Affect (linguistics); Family medicine; Psychology; Computer science","score_opus":0.03682922128576604,"score_gpt":0.4178243242119259,"score_spread":0.38099510292615985,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4327660727","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9956889,0.0015991039,0.000010223587,0.0007315872,0.0002797808,0.0015195345,0.00003851018,0.000106148065,0.000026223679],"genre_scores_gemma":[0.9985563,0.000023603374,0.0004237316,0.0004381729,0.00013516394,0.000038512815,0.00018282907,0.000032362943,0.00016931578],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9986816,0.000048575337,0.00041989275,0.00029968665,0.00024947248,0.0003007424],"domain_scores_gemma":[0.99775225,0.0015973109,0.00017964208,0.00020840038,0.000120616816,0.00014180751],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00021972241,0.00013432282,0.00041942153,0.00011371481,0.000080339276,0.000027346816,0.000047334903,0.00006156879,0.0000037870395],"category_scores_gemma":[0.003185343,0.00011276806,0.000081011836,0.0002444955,0.000083607876,0.000064014064,0.000025226753,0.00016208141,0.000004308172],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000558747,0.0016206853,0.17161067,0.00069530983,0.000053275555,0.000026945228,0.12250683,0.000006971485,0.00007998012,0.000006360706,0.005089055,0.69774514],"study_design_scores_gemma":[0.006054287,0.0033147156,0.8985901,0.007551036,0.00014669512,0.0000027711333,0.08277508,0.0011819587,0.000046849353,0.000008132812,0.00013090411,0.00019750533],"about_ca_topic_score_codex":0.00007132696,"about_ca_topic_score_gemma":0.000020734924,"teacher_disagreement_score":0.7269794,"about_ca_system_score_codex":0.00012456674,"about_ca_system_score_gemma":0.000102782935,"threshold_uncertainty_score":0.4598548},"labels":[],"label_agreement":null},{"id":"W4360620752","doi":"10.33137/utmj.v100i1.38937","title":"Debiasing and Educational Interventions in Medical Diagnosis: A Systematic Review","year":2023,"lang":"en","type":"review","venue":"University of Toronto Medical Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"University of Guelph; University of Toronto","funders":"","keywords":"Debiasing; PsycINFO; Psychological intervention; Critical appraisal; MEDLINE; Meta-analysis; Intervention (counseling); Medicine; Cognitive bias; Cognition; Psychology; Psychiatry; Social psychology; Alternative medicine","score_opus":0.08266968188997068,"score_gpt":0.4229625706977679,"score_spread":0.3402928888077972,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4360620752","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000069933008,0.99564415,0.00003763202,0.002893766,0.00026483423,0.00062468875,0.000012902084,0.00001564886,0.00049936696],"genre_scores_gemma":[0.000005449969,0.9988176,0.0001498542,0.00025387315,0.00018895364,0.000004546956,0.00003986074,0.000018285557,0.00052157853],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9964095,0.0005326513,0.001291928,0.00025376878,0.0012817173,0.00023043525],"domain_scores_gemma":[0.9795686,0.017564394,0.00092656683,0.00029702284,0.00019213157,0.0014512581],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0029048203,0.0002128312,0.0028287333,0.0001271351,0.000074381525,0.0000119570295,0.00037308014,0.0004454932,0.0056899246],"category_scores_gemma":[0.13338557,0.00016880888,0.0009602288,0.00014396786,0.00022442816,0.00007196211,0.00018635021,0.00094108254,0.0000412792],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000065933446,0.00038409035,0.00007363898,0.7051531,0.00040987218,0.0011175051,0.00008158373,5.2883e-9,3.7193967e-10,0.00008194412,0.00710359,0.2855881],"study_design_scores_gemma":[0.0005972918,0.00009877007,0.00020948522,0.9718648,0.004484126,0.0023513837,0.0001217978,0.000006439367,9.575418e-10,0.000025923224,0.020119471,0.00012048417],"about_ca_topic_score_codex":0.00084800756,"about_ca_topic_score_gemma":0.0015804529,"teacher_disagreement_score":0.2854676,"about_ca_system_score_codex":0.00044495368,"about_ca_system_score_gemma":0.0023775606,"threshold_uncertainty_score":0.995219},"labels":[],"label_agreement":null},{"id":"W4360979518","doi":"10.1002/9781119616870.ch4","title":"Clinical Decision Rules","year":2023,"lang":"en","type":"other","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Warrant; Test (biology); Decision rule; Physical exam; Computer science; Action (physics); Diagnostic test; Operations research; Medicine; Artificial intelligence; Pediatrics; Engineering; Surgery","score_opus":0.055052353768025385,"score_gpt":0.4388138073725363,"score_spread":0.3837614536045109,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4360979518","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0002394035,0.0004892921,0.00031531413,0.00039511098,0.0018176566,0.00020927118,0.000020754336,0.0017456888,0.9947675],"genre_scores_gemma":[0.000067113346,0.0037959728,0.008241823,0.0009878657,0.0017707185,0.000007311198,0.000043929416,0.0008612112,0.9842241],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99875027,0.000021999787,0.00041934647,0.00036308472,0.0002747519,0.00017055015],"domain_scores_gemma":[0.97671604,0.022311196,0.000104761726,0.0005355849,0.000027718277,0.00030471833],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00038566903,0.00015274726,0.00059126434,0.00011007182,0.000009833423,0.0000086066675,0.0000776224,0.0005924846,0.01326232],"category_scores_gemma":[0.08674035,0.000102413884,0.00029608776,0.0000860043,0.00007918935,0.0000036468768,0.000059539234,0.0003047445,0.023412315],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000020622278,0.00009626344,0.009178504,0.0000059396907,0.000050494884,0.000095785894,6.449704e-7,1.210396e-8,2.032918e-8,0.0002220554,0.76585954,0.22447012],"study_design_scores_gemma":[0.0011185725,0.00017256408,0.029054128,0.006805219,0.00014438988,0.000009386148,0.00000379746,0.000010017745,3.1877167e-7,0.00038668883,0.9621675,0.00012741753],"about_ca_topic_score_codex":0.00019895646,"about_ca_topic_score_gemma":0.00016786212,"teacher_disagreement_score":0.2243427,"about_ca_system_score_codex":0.000012286126,"about_ca_system_score_gemma":0.00009355606,"threshold_uncertainty_score":0.98763967},"labels":[],"label_agreement":null},{"id":"W4361018975","doi":"10.1136/bmjqs-2022-015038","title":"Diagnostic error among vulnerable populations presenting to the emergency department with cardiovascular and cerebrovascular or neurological symptoms: a systematic review","year":2023,"lang":"en","type":"review","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Emergency department; MEDLINE; Medical diagnosis; Stroke (engine); Odds ratio; Myocardial infarction; Meta-analysis; Emergency medicine; Intensive care medicine; Internal medicine; Pathology; Psychiatry","score_opus":0.16877980283864755,"score_gpt":0.45338710390360926,"score_spread":0.2846073010649617,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4361018975","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00041676668,0.97860575,0.00028757253,0.0011284762,0.0003257117,0.018841129,0.00012074612,0.0002111913,0.00006267798],"genre_scores_gemma":[0.00015258267,0.9921087,0.00015132006,0.00041070604,0.00035422487,0.0055133337,0.0004251773,0.000110687535,0.00077329384],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.98878753,0.0042640506,0.0037969153,0.0011452091,0.0013592383,0.00064705935],"domain_scores_gemma":[0.96167547,0.033807576,0.00093453407,0.0026871362,0.0002699332,0.0006253737],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.010620338,0.0007244451,0.0060502905,0.000099394136,0.0004416263,0.00005280459,0.00039008007,0.00031333583,0.000108687986],"category_scores_gemma":[0.2651227,0.00034222536,0.002028481,0.001115706,0.0001220359,0.000052857176,0.00036247552,0.00089092756,0.00019447856],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003081152,0.0002215124,0.0042184796,0.97208184,0.0048118713,0.00028706682,0.00008502949,0.000078241035,1.2080116e-9,0.00012742508,0.0033113966,0.014746295],"study_design_scores_gemma":[0.0005657449,0.00027056586,0.0131582245,0.87842983,0.041408498,0.00039889597,0.00005266735,0.00003131143,6.0345537e-9,0.000019690568,0.06504615,0.0006184182],"about_ca_topic_score_codex":0.00031731362,"about_ca_topic_score_gemma":0.00018826494,"teacher_disagreement_score":0.2545024,"about_ca_system_score_codex":0.000100717116,"about_ca_system_score_gemma":0.00038862266,"threshold_uncertainty_score":0.99990296},"labels":[],"label_agreement":null},{"id":"W4361199889","doi":"10.1136/bmjoq-2022-002062","title":"Types of diagnostic errors reported by paediatric emergency providers in a global paediatric emergency care research network","year":2023,"lang":"en","type":"article","venue":"BMJ Open Quality","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Medicine; Emergency department; Pediatrics; Abdominal pain; Constipation; Harm; Vomiting; Family medicine; Emergency medicine; Psychiatry; Internal medicine; Psychology","score_opus":0.17855506982401612,"score_gpt":0.5271385786246078,"score_spread":0.34858350880059163,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4361199889","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9797108,0.0059372084,0.000014549555,0.0025003878,0.0012094808,0.0045684376,0.00017666075,0.00011315693,0.005769319],"genre_scores_gemma":[0.99305433,0.0036852837,0.00017924272,0.00007591011,0.0004856724,0.00065069355,0.00048433163,0.000037254773,0.0013472867],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9933009,0.0013036035,0.002268622,0.0008369233,0.0012817114,0.0010082154],"domain_scores_gemma":[0.9824,0.01429947,0.00076929916,0.0012182867,0.00080877484,0.00050416705],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.010009162,0.00026195755,0.000931768,0.00018781183,0.00014747836,0.00002660284,0.00055093865,0.0002974098,0.00068633765],"category_scores_gemma":[0.34054822,0.00023566697,0.00025184176,0.00601919,0.00012214965,0.00013080891,0.0007462776,0.00051968225,0.00023466484],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002777338,0.00033711785,0.7417461,0.0002909905,0.000059820635,0.00015838524,0.0004266447,0.00012477058,0.000004410453,0.00014098966,0.24320035,0.013232677],"study_design_scores_gemma":[0.0019524283,0.00047158886,0.9890752,0.0005928235,0.00016000181,0.000007259139,0.0022020808,0.000099273566,0.000010902782,0.0036414058,0.001416675,0.00037039907],"about_ca_topic_score_codex":0.0123700565,"about_ca_topic_score_gemma":0.0019693428,"teacher_disagreement_score":0.33053905,"about_ca_system_score_codex":0.00020708256,"about_ca_system_score_gemma":0.0009599116,"threshold_uncertainty_score":0.99420667},"labels":[],"label_agreement":null},{"id":"W4362522431","doi":"10.1111/opn.12538","title":"Fast thinking: How unconscious bias and binary language contribute to rationing of care to older persons","year":2023,"lang":"en","type":"article","venue":"International Journal of Older People Nursing","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Unconscious mind; Psychology; Rationing; Nursing; Cognitive psychology; Social psychology; Medicine; Psychoanalysis; Health care","score_opus":0.020601185777763636,"score_gpt":0.35472826361307874,"score_spread":0.3341270778353151,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4362522431","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9570011,0.00044024657,0.00265603,0.03840095,0.0010998357,0.0001660714,0.00003752262,0.000025406292,0.00017285947],"genre_scores_gemma":[0.99666893,0.000030935265,0.0017380369,0.00080417085,0.00044223384,0.0000025734155,0.000029975996,0.000017055525,0.00026608148],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.99863535,0.000040863,0.00035205504,0.00015338334,0.00064306904,0.00017530221],"domain_scores_gemma":[0.9966952,0.0019464535,0.00021364016,0.00009858641,0.0007462081,0.00029991698],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0003680351,0.00010958695,0.00032784475,0.0004318725,0.000062215484,0.00008014778,0.00014633678,0.000060252347,0.000036251626],"category_scores_gemma":[0.010853972,0.00009466418,0.00012475949,0.00029904034,0.000038159847,0.000092039394,0.000046234338,0.00018712459,0.000007374364],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013271803,0.0006838236,0.3032659,0.00011619339,0.0007888768,0.0018152993,0.46206304,0.0027039198,0.018445617,0.0003862049,0.05406568,0.15433827],"study_design_scores_gemma":[0.0050002867,0.0013042415,0.90672266,0.021741211,0.0002874539,0.001103397,0.05752647,0.00065358577,0.004942777,0.00010178522,0.00034966212,0.00026645084],"about_ca_topic_score_codex":0.000036482703,"about_ca_topic_score_gemma":0.00003457364,"teacher_disagreement_score":0.6034568,"about_ca_system_score_codex":0.0001349929,"about_ca_system_score_gemma":0.00018082555,"threshold_uncertainty_score":0.997478},"labels":[],"label_agreement":null},{"id":"W4362704597","doi":"10.1016/j.acap.2023.04.002","title":"A Team-based Learning Approach During Pediatric Clerkship to Promote Clinical Reasoning","year":2023,"lang":"en","type":"article","venue":"Academic Pediatrics","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Medical education; Team-based learning; Psychology; Medicine","score_opus":0.052488396630875345,"score_gpt":0.3766847694457155,"score_spread":0.32419637281484015,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4362704597","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9936991,0.00061228854,0.00060363545,0.0011498735,0.0005421029,0.0007697733,0.0000096847225,0.0012013707,0.0014121964],"genre_scores_gemma":[0.9805022,0.0029685835,0.006562089,0.0011810815,0.0065637217,0.00011418539,0.00013874768,0.00012667675,0.0018427392],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99540734,0.00021798372,0.0012934424,0.0010386846,0.000946159,0.0010964124],"domain_scores_gemma":[0.9908761,0.006849003,0.00033692608,0.0005288716,0.00018114198,0.0012280025],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0036819358,0.00037287988,0.0007759884,0.00067942304,0.00025968146,0.00003667962,0.000381354,0.0008929927,0.000030143985],"category_scores_gemma":[0.15596214,0.00035165562,0.0003800212,0.0034129173,0.00006666978,0.00008623651,0.0002812191,0.0034127939,0.0011499794],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00023744223,0.00017552153,0.9744414,0.00019235429,0.000020994561,0.000100080135,0.00021439511,0.0025719977,0.00002275829,0.000032276792,0.015687358,0.0063034026],"study_design_scores_gemma":[0.005022394,0.00074817205,0.9710244,0.00046624165,0.000847443,0.00007417674,0.00015077603,0.015368064,0.000054456676,0.0000932382,0.0053743953,0.00077623304],"about_ca_topic_score_codex":0.000014824244,"about_ca_topic_score_gemma":2.645123e-7,"teacher_disagreement_score":0.15228021,"about_ca_system_score_codex":0.00012633568,"about_ca_system_score_gemma":0.00053878414,"threshold_uncertainty_score":0.99989355},"labels":[],"label_agreement":null},{"id":"W4366083192","doi":"10.1007/s00134-023-07039-2","title":"Variants of concern and clinical outcomes in critically ill COVID-19 patients","year":2023,"lang":"en","type":"letter","venue":"Intensive Care Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Faculty of Medicine and Dentistry, University of Alberta; Universitat Pompeu Fabra; Pontifícia Universidade Católica do Rio de Janeiro; Fundação Carlos Chagas Filho de Amparo à Pesquisa do Estado do Rio de Janeiro; Universidade Federal do Rio de Janeiro; Minderoo Foundation; Conselho Nacional de Desenvolvimento Científico e Tecnológico; Coordenação de Aperfeiçoamento de Pessoal de Nível Superior; Fundação Oswaldo Cruz; University of Alberta; Bill and Melinda Gates Foundation; Instituto D'Or de Pesquisa e Ensino","keywords":"Medicine; Anesthesiology; Coronavirus disease 2019 (COVID-19); Critically ill; Pain medicine; Intensive care medicine; 2019-20 coronavirus outbreak; Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); Critical illness; Coronavirus Infections; Intensive care; MEDLINE; Betacoronavirus; Emergency medicine; Internal medicine; Virology; Anesthesia; Outbreak; Infectious disease (medical specialty); Disease","score_opus":0.0845492154534833,"score_gpt":0.43511760785382847,"score_spread":0.35056839240034515,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4366083192","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06811019,0.0007534734,0.000034671735,0.92548704,0.0034571672,0.0009493358,0.00035623467,0.00009407109,0.00075782073],"genre_scores_gemma":[0.04425536,0.0007189517,0.0000688262,0.9506938,0.0018164733,0.000024417617,0.001277746,0.00008624437,0.0010582157],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9956986,0.00025366613,0.0018194951,0.0008481846,0.00087795715,0.0005021429],"domain_scores_gemma":[0.9182855,0.067191325,0.00044894483,0.00075758994,0.012602147,0.0007144903],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.00058846787,0.00045208735,0.0026901658,0.00036599388,0.000029258872,0.0000065896493,0.00020056538,0.0011954617,0.00020852676],"category_scores_gemma":[0.80604655,0.00031930115,0.00030220448,0.00026529795,0.0016333939,0.00002065878,0.00019051104,0.0023207352,0.000051338255],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000116979616,0.000043003736,0.44887954,0.000330213,0.00015446231,0.0030508707,0.0005125716,6.945663e-8,4.2745768e-7,0.0000057050056,0.5451877,0.0017184605],"study_design_scores_gemma":[0.015120597,0.004308694,0.86737144,0.015495355,0.0013810159,0.0000716897,0.005040054,0.000013198468,0.0000012690762,0.00039081785,0.09034997,0.0004559271],"about_ca_topic_score_codex":0.0014311392,"about_ca_topic_score_gemma":0.000078954996,"teacher_disagreement_score":0.80545807,"about_ca_system_score_codex":0.00019787409,"about_ca_system_score_gemma":0.00055140007,"threshold_uncertainty_score":0.9999809},"labels":[],"label_agreement":null},{"id":"W4366351590","doi":"10.1016/j.annemergmed.2023.03.018","title":"How Emergency Medicine Residents Have Conversations About Life-Sustaining Treatments in Critical Illness: A Qualitative Study Using Inductive Thematic Analysis","year":2023,"lang":"en","type":"article","venue":"Annals of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Public Health; Ottawa Hospital; Dalhousie University; University of Ottawa","funders":"University of Ottawa","keywords":"Thematic analysis; Medicine; Qualitative research; Emergency department; Duty; Distress; Nursing; Family medicine; Medical emergency; Clinical psychology","score_opus":0.40326021208966395,"score_gpt":0.5736119249454514,"score_spread":0.17035171285578748,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4366351590","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97665715,0.001560617,0.00032596235,0.018083699,0.0011643615,0.0010555118,0.000030785468,0.0000762071,0.0010456876],"genre_scores_gemma":[0.9944723,0.0024178447,0.0000624431,0.0001303729,0.00047371653,0.00011674652,0.00014109351,0.000046198416,0.0021392785],"study_design_codex":"observational","study_design_gemma":"qualitative","domain_scores_codex":[0.9940213,0.0007705193,0.0022883918,0.00077505776,0.0013858214,0.00075892097],"domain_scores_gemma":[0.98902327,0.0073196525,0.0007183845,0.0007639397,0.0015012489,0.00067350024],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00341371,0.00045495734,0.0019520451,0.0017791139,0.00018046878,0.00000406497,0.0002360343,0.00018158619,0.00281562],"category_scores_gemma":[0.20388009,0.0003373509,0.0003785077,0.005165874,0.00045796589,0.00020914865,0.00010413671,0.0003890626,0.000015688855],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00054315693,0.002808969,0.5612287,0.00040273124,0.006807617,0.0002436953,0.41197592,0.00010888405,0.00029442972,0.0008190271,0.01407693,0.000689934],"study_design_scores_gemma":[0.0035961915,0.0025922398,0.37346277,0.0029350526,0.005190759,0.0000017869476,0.60559183,0.00200037,0.00003596022,0.004241787,0.000019907005,0.00033131486],"about_ca_topic_score_codex":0.0056664525,"about_ca_topic_score_gemma":0.0004021698,"teacher_disagreement_score":0.20046636,"about_ca_system_score_codex":0.00007885151,"about_ca_system_score_gemma":0.00020002028,"threshold_uncertainty_score":0.99990785},"labels":[],"label_agreement":null},{"id":"W4366371924","doi":"10.1016/j.soin.2023.03.012","title":"“Briser la barrière du silence”, un programme de simulation pour adresser la détresse psychologique","year":2023,"lang":"fr","type":"article","venue":"Soins","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre Hospitalier de l’Université de Montréal","funders":"","keywords":"Psychology; Silence; Psychological distress; Burnout; Distress; Humanities; Psychotherapist; Psychoanalysis; Clinical psychology; Philosophy; Mental health","score_opus":0.039236501683961374,"score_gpt":0.3868196267105061,"score_spread":0.34758312502654476,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4366371924","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.942176,0.0026995721,0.0018021214,0.024467383,0.001787874,0.0006517429,0.00007957877,0.00060251803,0.025733214],"genre_scores_gemma":[0.97669315,0.002219675,0.0006806041,0.001107989,0.0019238307,0.00008320203,0.00009023046,0.00007205561,0.01712927],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9971414,0.0005430676,0.00047412334,0.0005209242,0.00034169454,0.0009787792],"domain_scores_gemma":[0.9686268,0.029984277,0.00013429145,0.00047012596,0.0001589963,0.0006255396],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0020777823,0.00028802166,0.00040272105,0.000112756,0.00013516074,0.00008632179,0.00019791622,0.0006955031,0.0013875498],"category_scores_gemma":[0.07304081,0.00026250328,0.00024110617,0.00053747295,0.00028708394,0.000084101746,0.00009509947,0.0007179411,0.0024385047],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004191025,0.0029923322,0.4613513,0.00032664443,0.00034068417,0.004683782,0.0039313426,0.01995873,0.00021275526,0.01480524,0.09880437,0.3921737],"study_design_scores_gemma":[0.005637708,0.00070170045,0.7689595,0.006995527,0.00052270194,0.00017046471,0.0020666248,0.075029165,0.00046471006,0.0136293275,0.12517732,0.0006452469],"about_ca_topic_score_codex":0.00017450536,"about_ca_topic_score_gemma":0.00002311096,"teacher_disagreement_score":0.39152846,"about_ca_system_score_codex":0.0000871672,"about_ca_system_score_gemma":0.0006730746,"threshold_uncertainty_score":0.9999827},"labels":[],"label_agreement":null},{"id":"W4366599098","doi":"10.1186/s12909-023-04233-0","title":"Preparing future physicians for complexity: a post-graduate elective in HIV psychiatry","year":2023,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; Sinai Health System; University of Toronto","funders":"Ontario HIV Treatment Network","keywords":"Thematic analysis; Medical education; Psychology; Medicine; Standardization; Nursing; Ambiguity; Perspective (graphical); Qualitative research; Computer science","score_opus":0.05150194449362368,"score_gpt":0.4058862032556512,"score_spread":0.35438425876202756,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4366599098","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94019246,0.00013842958,0.00090623833,0.053873546,0.0024337906,0.0007706556,0.000013199359,0.00017107773,0.0015005983],"genre_scores_gemma":[0.9682298,0.00013470434,0.010405137,0.012536988,0.0053172763,0.0006014368,0.0013682016,0.000047913498,0.0013585569],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99865514,0.00005545277,0.0003094086,0.00031074867,0.00039331292,0.00027594273],"domain_scores_gemma":[0.99637914,0.0029423614,0.00007721782,0.00020902848,0.00014565296,0.0002465744],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0005231457,0.00010874519,0.00026349156,0.00012797774,0.000054193064,0.0000113826045,0.00008618254,0.00014085189,0.00005928444],"category_scores_gemma":[0.04826277,0.00009666478,0.00011366242,0.00047870382,0.00006964219,0.000033366392,0.000023183666,0.00026136538,0.000097065604],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00094140734,0.0041410984,0.24620412,0.0006228206,0.000079244164,0.0000050763992,0.003044325,0.00001567165,0.000042519183,0.015858812,0.49860173,0.23044318],"study_design_scores_gemma":[0.0028683036,0.00035214712,0.96538204,0.00405757,0.000095977215,0.000020837386,0.0034933304,0.0075331912,0.000028036593,0.011856905,0.0041103316,0.00020133659],"about_ca_topic_score_codex":0.00009643761,"about_ca_topic_score_gemma":0.0008427724,"teacher_disagreement_score":0.7191779,"about_ca_system_score_codex":0.0000778569,"about_ca_system_score_gemma":0.003231176,"threshold_uncertainty_score":0.9597541},"labels":[],"label_agreement":null},{"id":"W4367834674","doi":"10.1007/s43678-023-00495-1","title":"Using cognitive load theory to develop an emergency airway management curriculum: the Queen’s University Mastery Airway Course (QUMAC)","year":2023,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; Kingston Health Sciences Centre; Queen's University","funders":"","keywords":"Airway management; Cognition; Curriculum; Medicine; Humanities; Psychological intervention; Psychology; Medical education; Airway; Nursing; Pedagogy; Anesthesia; Philosophy; Psychiatry","score_opus":0.08852420331708154,"score_gpt":0.3747317296231493,"score_spread":0.28620752630606777,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4367834674","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95472294,0.0011417568,0.0029906852,0.016628772,0.011887692,0.00080127106,0.000053113545,0.00005068414,0.011723107],"genre_scores_gemma":[0.98129416,0.0020978316,0.00031863083,0.00085116475,0.0017730866,0.0000038927365,0.00003924248,0.000057791574,0.013564224],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99712026,0.00029129046,0.0009153545,0.00033056457,0.0006801946,0.0006623374],"domain_scores_gemma":[0.994623,0.00071133254,0.0003642741,0.0003757818,0.0017079546,0.002217642],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0024182356,0.00028258242,0.00053773244,0.0004998646,0.00034437637,0.000005941152,0.00040443093,0.0001137586,0.006859983],"category_scores_gemma":[0.013124382,0.00019785689,0.00017064971,0.00220841,0.00018889745,0.00012603724,0.000053987278,0.00047986305,0.0001507935],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007612756,0.00044503395,0.2294317,0.00019607996,0.0017977573,0.008953279,0.014114301,0.00039421258,0.000090154725,0.0037753566,0.66475934,0.07528149],"study_design_scores_gemma":[0.010605137,0.005681153,0.6318779,0.017275918,0.007779868,0.00042225645,0.14169477,0.0009432708,0.000071440576,0.0028100002,0.17914851,0.0016897877],"about_ca_topic_score_codex":0.0064406167,"about_ca_topic_score_gemma":0.006732961,"teacher_disagreement_score":0.48561084,"about_ca_system_score_codex":0.00038281977,"about_ca_system_score_gemma":0.0019973433,"threshold_uncertainty_score":0.9951885},"labels":[],"label_agreement":null},{"id":"W4376618681","doi":"10.1007/s10459-023-10232-z","title":"Pathologists aren’t pigeons: exploring the neural basis of visual recognition and perceptual expertise in pathology","year":2023,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria; University of Calgary","funders":"Canada Foundation for Innovation","keywords":"Categorization; Perception; Task (project management); Electroencephalography; Psychology; Visual perception; Cognitive psychology; Event-related potential; Perceptual learning; Audiology; Computer science; Medicine; Artificial intelligence; Neuroscience","score_opus":0.121836229001862,"score_gpt":0.4526430243115631,"score_spread":0.33080679530970114,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4376618681","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99039185,0.0042172447,0.000014141493,0.0043168454,0.00069437776,0.00026131337,0.0000022998615,0.00001949731,0.00008241414],"genre_scores_gemma":[0.985021,0.013461516,0.0004528591,0.00078323757,0.00006799072,0.00018218473,0.000016727849,0.0000035796186,0.000010926035],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9987652,0.00016557924,0.00040334812,0.00026805146,0.00016144715,0.00023640622],"domain_scores_gemma":[0.99674046,0.0029333634,0.00012257499,0.00009512014,0.000038008617,0.00007047927],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0013708316,0.00006675782,0.00018155889,0.00021541877,0.00008508605,0.0000073798037,0.00006317729,0.000030468853,0.000011637488],"category_scores_gemma":[0.0124681825,0.000047374804,0.000018258577,0.0007857302,0.000427654,0.00020880094,0.000025891411,0.0001267878,0.0000063184543],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000021999562,0.00012606064,0.2929553,0.000045366345,2.5265254e-7,0.0000031031293,0.0041352888,0.0000797658,0.00002374425,0.00003364871,0.000024632527,0.7025508],"study_design_scores_gemma":[0.0002910534,0.0006430797,0.9808247,0.0011649646,0.0000026075602,0.000028485347,0.013985751,0.0014195036,0.000054991127,0.0014651507,0.00005923144,0.000060495335],"about_ca_topic_score_codex":0.00046782644,"about_ca_topic_score_gemma":0.00031606437,"teacher_disagreement_score":0.7024903,"about_ca_system_score_codex":0.000056009238,"about_ca_system_score_gemma":0.00028186955,"threshold_uncertainty_score":0.9958502},"labels":[],"label_agreement":null},{"id":"W4377378862","doi":"10.5334/pme.898","title":"What can Designing Learning-by-Concordance Clinical Reasoning Cases Teach Us about Instruction in the Health Sciences?","year":2023,"lang":"en","type":"article","venue":"Perspectives on Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Experiential learning; Thematic analysis; Cognitive apprenticeship; Psychology; Dialogic; Medical education; Computer science; Mathematics education; Pedagogy; Qualitative research; Medicine","score_opus":0.044199472605088025,"score_gpt":0.44678943093382345,"score_spread":0.40258995832873545,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4377378862","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9536201,0.0021481258,0.000096163414,0.041629825,0.0013070288,0.00038675554,0.0000012154287,0.00011902625,0.00069180835],"genre_scores_gemma":[0.98348635,0.008333368,0.00035233865,0.006493835,0.0008280272,0.000087127155,0.000057905465,0.000018194165,0.00034283358],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9965377,0.000835237,0.0005840179,0.0006396583,0.0009799863,0.0004234071],"domain_scores_gemma":[0.98686486,0.011970323,0.000271897,0.00028984982,0.00013285961,0.00047022323],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005813283,0.00016409968,0.00036749398,0.00020731588,0.00037754737,0.0001124899,0.00021177041,0.00016160675,0.00009587489],"category_scores_gemma":[0.21173863,0.00011827899,0.000105331754,0.0011387884,0.0006013516,0.00014312868,0.000029387551,0.0012267395,0.000051521394],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00026121325,0.0022235434,0.49244946,0.00004537196,0.00003405667,0.000063510546,0.034316327,0.000076881,0.000007499708,0.0032376521,0.027223619,0.44006085],"study_design_scores_gemma":[0.0021554355,0.0031077284,0.7448632,0.008181732,0.000039533097,0.00029936427,0.2355159,0.0024322267,0.000016434777,0.00038323985,0.0027365235,0.00026864652],"about_ca_topic_score_codex":0.0013088987,"about_ca_topic_score_gemma":0.00014203115,"teacher_disagreement_score":0.43979222,"about_ca_system_score_codex":0.0003160781,"about_ca_system_score_gemma":0.0028716484,"threshold_uncertainty_score":0.79490125},"labels":[],"label_agreement":null},{"id":"W4380883880","doi":"10.1002/alz.066185","title":"The correction of health misconceptions in older adults with cognitive impairment","year":2023,"lang":"en","type":"article","venue":"Alzheimer s & Dementia","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Cognition; Cognitive impairment; Psychology; Neuropsychology; Surprise; Montreal Cognitive Assessment; Clinical psychology; Medicine; Audiology; Psychiatry; Social psychology","score_opus":0.02635496717699193,"score_gpt":0.33737057925907976,"score_spread":0.31101561208208783,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4380883880","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9749679,0.019177292,0.000097090335,0.0032693655,0.00055275834,0.0010872976,0.000010298681,0.000074256335,0.00076374825],"genre_scores_gemma":[0.99888146,0.00050867366,0.000060517028,0.00030611185,0.000034462028,0.00008413449,0.00007282208,0.000010879545,0.000040952058],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.999022,0.00006799976,0.00032200138,0.00017070203,0.00019526397,0.00022203373],"domain_scores_gemma":[0.9974592,0.0021245312,0.0001215539,0.00013700419,0.00008229361,0.00007542374],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004331012,0.00008023591,0.00017892914,0.000054756078,0.00008920139,0.000007127816,0.000037595062,0.000031365034,0.0000714084],"category_scores_gemma":[0.00087553007,0.000051185947,0.000048918635,0.00034028047,0.00009917632,0.000021235643,0.00002206352,0.00011119236,0.000060397393],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009011531,0.0010286178,0.2553393,0.000014227346,0.004148233,0.000026461681,0.004241175,0.00004469297,0.000008447938,0.0000634685,0.027208101,0.7069761],"study_design_scores_gemma":[0.0034807464,0.0008992226,0.98655355,0.002359609,0.0017915629,0.0000071845297,0.0037000421,0.00041950992,0.00037036004,0.000017050936,0.00032543816,0.00007569903],"about_ca_topic_score_codex":0.00053019094,"about_ca_topic_score_gemma":0.00029138723,"teacher_disagreement_score":0.7312143,"about_ca_system_score_codex":0.000006546691,"about_ca_system_score_gemma":0.00013780863,"threshold_uncertainty_score":0.20873024},"labels":[],"label_agreement":null},{"id":"W4382343985","doi":"10.1136/bmjopen-2023-073337","title":"Impact of emotional competence on physicians’ clinical reasoning: a scoping review protocol","year":2023,"lang":"en","type":"review","venue":"BMJ Open","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Université de Liège; University of Toronto; Faculty of Medicine, McGill University; McGill University","keywords":"Medicine; Competence (human resources); Protocol (science); Medical education; Alternative medicine; Family medicine; Pathology; Social psychology","score_opus":0.42532102768298835,"score_gpt":0.6540488917419658,"score_spread":0.22872786405897744,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4382343985","genre_codex":"protocol","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000014039749,0.37024203,0.0000049028745,0.00016650742,0.00018960497,0.6237467,0.00009868999,0.00006803638,0.0054821107],"genre_scores_gemma":[1.00403476e-7,0.7476451,0.00049650494,0.00042247746,0.0005132195,0.24994306,0.00022564904,0.00007585151,0.0006780488],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99439186,0.00075130403,0.0029255343,0.00081372855,0.00074105326,0.0003765151],"domain_scores_gemma":[0.9768235,0.019095408,0.0021300812,0.0013764888,0.0002594536,0.0003150462],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004162016,0.0004903795,0.006000789,0.00010237991,0.000055761982,0.00004879108,0.00078212644,0.00033935765,0.00045328797],"category_scores_gemma":[0.113670595,0.00032402604,0.0023075223,0.00070158445,0.00016742742,0.00005495194,0.0005705707,0.0007235231,0.00080298976],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00017076565,0.0008208171,0.0004836984,0.24752828,0.0005028796,0.0001183462,0.000003933938,6.890684e-7,6.1345498e-9,0.00011446977,0.02823181,0.7220243],"study_design_scores_gemma":[0.0012195222,0.0017498432,0.001535936,0.9491644,0.0009583716,0.00006890705,0.0000010387439,0.000006137741,3.0877732e-8,0.000021400952,0.04497296,0.00030146833],"about_ca_topic_score_codex":0.00006414296,"about_ca_topic_score_gemma":0.0000049948376,"teacher_disagreement_score":0.72172284,"about_ca_system_score_codex":0.0001281229,"about_ca_system_score_gemma":0.005046142,"threshold_uncertainty_score":0.999975},"labels":[],"label_agreement":null},{"id":"W4382793965","doi":"10.1007/s10459-023-10241-y","title":"Exploring adaptive expertise in residency: the (missed) opportunity of uncertainty","year":2023,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; University of Toronto","funders":"Region Midtjylland; Aarhus Universitetshospital; Aarhus Universitet","keywords":"Medical education; Residency training; MEDLINE; Medicine; Psychology; Computer science; Family medicine; Biology","score_opus":0.2441301742407389,"score_gpt":0.47012408276965467,"score_spread":0.22599390852891577,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4382793965","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96841955,0.0062699276,0.00002101518,0.022259746,0.000901661,0.00044672057,0.0000015119746,0.000027612525,0.0016522612],"genre_scores_gemma":[0.98235095,0.015717944,0.00053775747,0.0010564886,0.00007535694,0.00016279855,0.0000054244892,0.0000037580187,0.000089552814],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9985752,0.00012802969,0.00049353857,0.0002410337,0.00029951314,0.00026267258],"domain_scores_gemma":[0.99451417,0.0048815673,0.00018703277,0.00019932316,0.00006690936,0.00015101596],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0022835988,0.00006761936,0.00019571073,0.00021830833,0.000099348785,0.0000058997102,0.00013755115,0.000022305092,0.00001745155],"category_scores_gemma":[0.02396614,0.000046410907,0.00002588635,0.0016270943,0.00027580507,0.0002776025,0.000024501225,0.00014773352,0.0000062442987],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011291987,0.00039844343,0.22192217,0.00017207912,0.0000018858727,0.0000048900306,0.009984682,0.0074665695,0.000014449005,0.007966648,0.00090854504,0.7510467],"study_design_scores_gemma":[0.0006267916,0.00093386025,0.923831,0.009357851,0.0000052538835,0.00000793265,0.034983728,0.008188836,0.00013809983,0.019548163,0.0022220684,0.00015644798],"about_ca_topic_score_codex":0.003900135,"about_ca_topic_score_gemma":0.00097237714,"teacher_disagreement_score":0.75089025,"about_ca_system_score_codex":0.00014601051,"about_ca_system_score_gemma":0.0029786935,"threshold_uncertainty_score":0.98425543},"labels":[],"label_agreement":null},{"id":"W4384134853","doi":"10.1051/pmed/2023008","title":"Formation par concordance de script en orthophonie : récit de pratique en évaluation des troubles acquis de la communication","year":2023,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Centre Intégré Universitaire de Santé et de Services Sociaux du Centre-Sud-de-l'Île-de-Montréal; Université de Montréal","funders":"","keywords":"Humanities; Political science; Art","score_opus":0.04548392723650688,"score_gpt":0.38099396271812186,"score_spread":0.33551003548161495,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4384134853","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.92841613,0.015116107,0.035386436,0.01386198,0.00039235214,0.0006373492,0.00005323223,0.00039129442,0.005745135],"genre_scores_gemma":[0.9025867,0.051579107,0.031032722,0.0021215042,0.00082394696,0.0003378111,0.00039026028,0.000074664254,0.01105331],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99456966,0.0029281788,0.0007758508,0.00034992708,0.00046759873,0.00090876943],"domain_scores_gemma":[0.96261173,0.03576285,0.0002582415,0.00065285724,0.00020868401,0.00050561875],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0086474195,0.00028157674,0.00046528334,0.0001382219,0.00035499776,0.00013430441,0.0003680166,0.0006381966,0.00038030994],"category_scores_gemma":[0.10936639,0.00029201165,0.00019112176,0.00058832637,0.00076713675,0.00030768034,0.00017372929,0.0007704629,0.0004454283],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00033508457,0.0010129611,0.17048773,0.0006806575,0.0001906726,0.00019293933,0.031168547,0.0016892134,0.008560451,0.0045333914,0.05125171,0.72989666],"study_design_scores_gemma":[0.00648671,0.0007996618,0.6977508,0.013572445,0.000972132,0.00034353213,0.0089949,0.11995614,0.020219937,0.06071916,0.06936072,0.0008238579],"about_ca_topic_score_codex":0.0010841512,"about_ca_topic_score_gemma":0.00018369431,"teacher_disagreement_score":0.72907275,"about_ca_system_score_codex":0.0009657151,"about_ca_system_score_gemma":0.0016724451,"threshold_uncertainty_score":0.9999532},"labels":[],"label_agreement":null},{"id":"W4384210107","doi":"10.1051/pmed/2023009","title":"La physiothérapie française est-elle prête à intégrer des tests de concordance de script dans sa formation initiale ? Étude de faisabilité et d’acceptabilité","year":2023,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institut du Savoir Montfort; University of Ottawa","funders":"","keywords":"Humanities; Medicine; Philosophy","score_opus":0.046666022414629034,"score_gpt":0.3731341309546992,"score_spread":0.3264681085400702,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4384210107","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9654937,0.010213588,0.0077794064,0.0049461783,0.0007835096,0.00079992064,0.00017983171,0.0006353178,0.009168598],"genre_scores_gemma":[0.9774842,0.007928636,0.0058902605,0.002790656,0.001020275,0.000319138,0.00014729849,0.00013370946,0.004285813],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99461883,0.0009982184,0.0010769359,0.0007849266,0.0006910816,0.0018300132],"domain_scores_gemma":[0.97185946,0.025521666,0.00025264674,0.00086316385,0.00024145968,0.0012615966],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0037249913,0.0005175561,0.0008500812,0.00014609279,0.0003538818,0.0001827291,0.0005068707,0.0007037039,0.0012766944],"category_scores_gemma":[0.108202554,0.000551478,0.0004375034,0.0010224882,0.002064983,0.0004398055,0.0003323846,0.0012153432,0.0010115637],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005166659,0.0042083026,0.32166278,0.002290453,0.00025416358,0.0012731458,0.046073187,0.0023404248,0.00874193,0.0013766276,0.02861589,0.5826464],"study_design_scores_gemma":[0.0064868187,0.0011928466,0.8579238,0.0120638,0.0006366836,0.000526124,0.011634298,0.037576124,0.0064843493,0.01110661,0.053256083,0.0011124322],"about_ca_topic_score_codex":0.0036490732,"about_ca_topic_score_gemma":0.0021130384,"teacher_disagreement_score":0.58153397,"about_ca_system_score_codex":0.00074220146,"about_ca_system_score_gemma":0.0016850559,"threshold_uncertainty_score":0.9997663},"labels":[],"label_agreement":null},{"id":"W4384210485","doi":"10.1051/pmed/2023005","title":"Faut-il former plus souvent à distance ?","year":2023,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Université de Sherbrooke","funders":"","keywords":"Humanities; Philosophy; Political science","score_opus":0.07011226078663652,"score_gpt":0.3774530272940189,"score_spread":0.3073407665073824,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4384210485","genre_codex":"empirical","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2987245,0.21050227,0.021401474,0.24443263,0.048727147,0.0030691416,0.0011253025,0.0033491082,0.16866842],"genre_scores_gemma":[0.25965315,0.020558784,0.00083063915,0.004403681,0.004543536,0.0001631311,0.00035854714,0.00015434784,0.7093342],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99573076,0.00012673685,0.00084903534,0.0007704668,0.0009191442,0.0016038539],"domain_scores_gemma":[0.9854151,0.012298474,0.00016412705,0.00082766317,0.00012299257,0.001171657],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0010896623,0.00040355927,0.00081299356,0.00012811003,0.0002159272,0.00007152295,0.0003421383,0.0004695155,0.0040612863],"category_scores_gemma":[0.048403274,0.00037014962,0.00046215698,0.0009468357,0.00071127596,0.00010084341,0.0002728504,0.0007791178,0.020215927],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00023805581,0.0009454352,0.012845963,0.00041387594,0.00022170112,0.0016847199,0.0012356074,0.0001544106,0.00007390626,0.0071693272,0.5615835,0.4134335],"study_design_scores_gemma":[0.0034702157,0.00050170015,0.020612584,0.006024235,0.00038571554,0.00004508595,0.0011100838,0.008897969,0.00026841444,0.002786368,0.95538473,0.0005128719],"about_ca_topic_score_codex":0.00017446354,"about_ca_topic_score_gemma":0.00003079386,"teacher_disagreement_score":0.54066575,"about_ca_system_score_codex":0.00035296133,"about_ca_system_score_gemma":0.0007179286,"threshold_uncertainty_score":0.99987507},"labels":[],"label_agreement":null},{"id":"W4384296888","doi":"10.2196/preprints.50797","title":"Transformation and Articulation of Clinical Data to Understand Students’ and Health Professionals’ Clinical Reasoning: Protocol for a Scoping Review (Preprint)","year":2023,"lang":"en","type":"review","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières; Université Laval; Université de Montréal","funders":"","keywords":"PsycINFO; CINAHL; Protocol (science); Checklist; Data extraction; MEDLINE; Systematic review; Computer science; Psychology; Medical education; Data science; Medicine; Alternative medicine; Political science","score_opus":0.7373878383015707,"score_gpt":0.6990918526696918,"score_spread":0.03829598563187897,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4384296888","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000016592601,0.5442817,0.0017982807,0.0025791468,0.00013058606,0.45100778,0.00006167316,0.000075294105,0.000048907637],"genre_scores_gemma":[0.0000012061769,0.9432236,0.00393642,0.0012688014,0.00017547056,0.050850812,0.00037882858,0.000042255386,0.00012258592],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99275976,0.0005811943,0.004938556,0.000921862,0.00053056475,0.0002680771],"domain_scores_gemma":[0.97725797,0.019668832,0.0013406194,0.00092695915,0.00018601936,0.00061963184],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.016783802,0.0002968512,0.0038781743,0.000099164296,0.00007717392,0.000030529816,0.00022662368,0.00034541392,0.0000256724],"category_scores_gemma":[0.068930626,0.0001986383,0.00034180487,0.00027786344,0.00009903302,0.000086730746,0.00035596677,0.00044362963,0.000011399993],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011471651,0.00042353084,0.0010786422,0.31150794,0.00016538917,0.0000014946371,0.000047796835,6.809266e-8,1.8407667e-9,0.00017180329,0.005485922,0.6810027],"study_design_scores_gemma":[0.0025387902,0.0009830715,0.002089147,0.9533833,0.0008133895,0.000016351236,0.00006205405,0.000087986744,2.8965156e-8,0.000036023815,0.039812304,0.00017755518],"about_ca_topic_score_codex":0.000017246479,"about_ca_topic_score_gemma":0.00003644201,"teacher_disagreement_score":0.6808251,"about_ca_system_score_codex":0.00004151126,"about_ca_system_score_gemma":0.0016798843,"threshold_uncertainty_score":0.93891215},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4384834592","doi":"10.2196/49034","title":"Effects of Combinational Use of Additional Differential Diagnostic Generators on the Diagnostic Accuracy of the Differential Diagnosis List Developed by an Artificial Intelligence–Driven Automated History–Taking System: Pilot Cross-Sectional Study","year":2023,"lang":"en","type":"article","venue":"JMIR Formative Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Japan Society for the Promotion of Science","keywords":"Medical diagnosis; Index (typography); Diagnostic accuracy; Medicine; Data mining; McNemar's test; Medical history; Computer science; Information retrieval; Statistics; World Wide Web; Pathology; Internal medicine; Mathematics","score_opus":0.17040622831726587,"score_gpt":0.4385155011283928,"score_spread":0.26810927281112695,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4384834592","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9949357,0.000024439367,0.00009238144,0.00013397449,0.00077466044,0.0025581964,0.0013446931,0.00010759486,0.000028353126],"genre_scores_gemma":[0.9970808,0.00003518956,0.000011820407,0.000022425458,0.00016235228,0.0017721636,0.0008456098,0.000036760754,0.000032847507],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99370134,0.0015409731,0.0013044812,0.00042211983,0.002547673,0.00048342926],"domain_scores_gemma":[0.818862,0.1786266,0.00066355744,0.00042904718,0.0012407931,0.00017795648],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011180701,0.00028617136,0.0006012666,0.00041702468,0.00043763255,0.00007571122,0.0005275831,0.0001523224,0.0010296713],"category_scores_gemma":[0.15322088,0.00017904893,0.00023782565,0.0010777931,0.001069165,0.0002121908,0.00034666262,0.0007396451,0.0000427966],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017763319,0.029723683,0.86486435,0.0015304716,0.0012036357,0.000069450885,0.006882107,0.000831198,0.004437011,0.013982428,0.07110869,0.0035906294],"study_design_scores_gemma":[0.0010264516,0.0028219686,0.9726142,0.0021317445,0.00007873926,0.0000039388847,0.00039983643,0.010812087,0.009777011,0.00014521611,0.000028980718,0.00015982249],"about_ca_topic_score_codex":0.0003865395,"about_ca_topic_score_gemma":0.000045322544,"teacher_disagreement_score":0.17708562,"about_ca_system_score_codex":0.00049656146,"about_ca_system_score_gemma":0.00053296646,"threshold_uncertainty_score":0.99988353},"labels":[],"label_agreement":null},{"id":"W4384925962","doi":"10.1097/jxx.0000000000000902","title":"An international comparison of student nurse practitioner diagnostic reasoning skills","year":2023,"lang":"en","type":"article","venue":"Journal of the American Association of Nurse Practitioners","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Emergent BioSolutions (Canada)","funders":"","keywords":"Curriculum; Medical education; Nurse practitioners; Clinical Practice; Flexibility (engineering); Critical thinking; Health care; Medicine; Psychology; Nursing; Mathematics education; Pedagogy","score_opus":0.01784388425195408,"score_gpt":0.41149209922436947,"score_spread":0.39364821497241537,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4384925962","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9825926,0.000022448261,0.000121511286,0.014440761,0.0013025893,0.00017424511,0.000033706343,0.00002402638,0.001288158],"genre_scores_gemma":[0.9972049,0.00031198972,0.0012665173,0.00040821958,0.00035256095,0.000006091831,0.00003587162,0.000019637177,0.00039420932],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9969454,0.00034918834,0.0010312661,0.00014389862,0.0013501623,0.00018012972],"domain_scores_gemma":[0.9797315,0.012238037,0.0064203157,0.00028107135,0.0011694013,0.00015969109],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0014902208,0.0001239575,0.00060958584,0.00026874317,0.00007276689,0.000026828437,0.00026939408,0.00006843136,0.000087427725],"category_scores_gemma":[0.0834796,0.00009444462,0.00032995146,0.000716551,0.0001459633,0.00036849617,0.000021902457,0.00044853953,0.00002203533],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00024945466,0.002417432,0.9703478,0.000007694125,0.00054627133,0.000013239758,0.0012355853,0.0027594762,0.0013157377,0.00024320894,0.017349426,0.0035146768],"study_design_scores_gemma":[0.0013945737,0.00077488046,0.98734146,0.0009393901,0.00051413925,0.000040014314,0.006326582,0.0002962545,0.00091158616,0.000095819014,0.0012739319,0.00009134008],"about_ca_topic_score_codex":0.000054791282,"about_ca_topic_score_gemma":0.000005155402,"teacher_disagreement_score":0.08198938,"about_ca_system_score_codex":0.00044330416,"about_ca_system_score_gemma":0.000308165,"threshold_uncertainty_score":0.92424065},"labels":[],"label_agreement":null},{"id":"W4385230119","doi":"10.1007/s10459-023-10267-2","title":"The impact of prior performance information on subsequent assessment: is there evidence of retaliation in an anonymous multisource assessment system?","year":2023,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University Health Centre","funders":"","keywords":"Logistic regression; Odds ratio; Medicine; Odds; Association (psychology); Standard score; Scoring system; Raw score; Psychology; Clinical psychology; Internal medicine; Statistics","score_opus":0.04877002521627438,"score_gpt":0.5003763224678314,"score_spread":0.451606297251557,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4385230119","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99606353,0.0011266774,0.000047877573,0.0011572281,0.00032441117,0.0007218826,0.0000045077986,0.00001775578,0.00053610676],"genre_scores_gemma":[0.99322665,0.0057590185,0.0007541261,0.00010987529,0.000028829476,0.00008223781,0.000017859227,0.0000035788485,0.000017813854],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9980492,0.00016438171,0.0008316436,0.00018783168,0.0005302734,0.00023665644],"domain_scores_gemma":[0.99587876,0.0029231606,0.0006812103,0.00026125176,0.00017172296,0.00008389368],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0033507627,0.0000908927,0.00023136898,0.0002520907,0.00014022195,0.000023366376,0.00016698582,0.00004083633,0.000005229738],"category_scores_gemma":[0.0027117135,0.000058691945,0.000040348285,0.0011930503,0.00015221859,0.0008741182,0.000016689637,0.00014712627,0.0000037836824],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000056266614,0.00016771753,0.8541824,0.00046106044,0.0000017562847,1.325354e-7,0.0012452941,0.011434235,0.000018280369,0.00078601093,0.000027489685,0.13161933],"study_design_scores_gemma":[0.0002772263,0.0017528153,0.92560256,0.007289744,0.000002917521,0.0000018057474,0.0022999786,0.06252902,0.00005647589,0.000107161366,0.00003158823,0.00004871343],"about_ca_topic_score_codex":0.0017300438,"about_ca_topic_score_gemma":0.00012773192,"teacher_disagreement_score":0.13157062,"about_ca_system_score_codex":0.00059696136,"about_ca_system_score_gemma":0.0022014196,"threshold_uncertainty_score":0.3905224},"labels":[],"label_agreement":null},{"id":"W4385235591","doi":"10.1212/wnl.0000000000207734","title":"Toward a More Inclusive Definition of Brain Health","year":2023,"lang":"en","type":"article","venue":"Neurology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Robarts Clinical Trials; Western University","funders":"","keywords":"Psychology; Cognitive science; Neuroscience","score_opus":0.061353863097104465,"score_gpt":0.3845840402472493,"score_spread":0.32323017715014485,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4385235591","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.81240875,0.00004988456,0.000028333217,0.18563867,0.00016685326,0.00011340991,0.0000113218575,0.00007698547,0.0015058267],"genre_scores_gemma":[0.9583031,0.00012361199,0.00013372028,0.041223146,0.00008736958,0.000009486256,0.00006838,0.000010018974,0.00004116514],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99926364,0.00006912212,0.0002156151,0.00015895783,0.000114635855,0.00017799759],"domain_scores_gemma":[0.993841,0.0057914075,0.00007562107,0.00014622159,0.000040974086,0.00010481],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00022385392,0.000054875512,0.0002501967,0.00008445554,0.000022045133,0.0000010447753,0.000038708564,0.000082706734,0.000055100445],"category_scores_gemma":[0.019496953,0.0000482022,0.00005842073,0.00021614331,0.00008426849,0.000006872795,0.0000579971,0.00016251316,0.000117619704],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.006284275,0.0019642934,0.30898115,0.0006672018,0.0002450352,0.0036082426,0.009410175,0.00013846692,0.0031759767,0.050916027,0.46496165,0.14964752],"study_design_scores_gemma":[0.0034576398,0.008380279,0.9646551,0.00024642007,0.000040463638,0.00030449487,0.00006878333,0.00042971142,0.00016027379,0.014794133,0.007357811,0.000104895225],"about_ca_topic_score_codex":0.00015771131,"about_ca_topic_score_gemma":0.000010958976,"teacher_disagreement_score":0.655674,"about_ca_system_score_codex":0.0000069428547,"about_ca_system_score_gemma":0.0001324327,"threshold_uncertainty_score":0.98876226},"labels":[],"label_agreement":null},{"id":"W4385484172","doi":"10.1080/03080188.2023.2234216","title":"Choosing between prediction and explanation in geological engineering: lessons from psychology","year":2023,"lang":"en","type":"article","venue":"Interdisciplinary Science Reviews","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Field (mathematics); Engineering; Artificial intelligence; Engineering ethics; Psychology; Sociology; Computer science; Mathematics","score_opus":0.11348986759608912,"score_gpt":0.46172692136405513,"score_spread":0.348237053767966,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4385484172","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98717,0.0010662547,0.0012643387,0.009393921,0.00031104463,0.0002899124,0.000010464866,0.00008466489,0.00040940568],"genre_scores_gemma":[0.9954704,0.002640367,0.0014297777,0.00014578708,0.00018262684,0.00003611204,0.000056990415,0.0000058270352,0.000032087253],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9987056,0.000041301988,0.00042038268,0.00043262413,0.00015155424,0.000248535],"domain_scores_gemma":[0.99810725,0.0014158201,0.00006817505,0.00021829749,0.000024462877,0.00016596666],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0014636599,0.00010121215,0.00035064013,0.00026046016,0.000077035686,0.000022150256,0.00012274295,0.000080611506,0.00002295072],"category_scores_gemma":[0.00729001,0.00007518104,0.000050904397,0.00083341455,0.00019997718,0.000103347506,0.00025445802,0.00021881424,0.000102179205],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000032180826,0.00008243311,0.6299639,0.000038177743,0.000008547877,0.000081571205,0.0009594012,0.000043479864,0.0034365407,0.00012682354,0.0014581172,0.36376885],"study_design_scores_gemma":[0.00034125737,0.00017245278,0.9921939,0.0018101394,0.000020880952,0.000020441077,0.000052785766,0.0027762821,0.00005715586,0.0010792704,0.0014052466,0.00007022539],"about_ca_topic_score_codex":0.000010437975,"about_ca_topic_score_gemma":0.000004557439,"teacher_disagreement_score":0.36369863,"about_ca_system_score_codex":0.000051629708,"about_ca_system_score_gemma":0.00003651542,"threshold_uncertainty_score":0.8727349},"labels":[],"label_agreement":null},{"id":"W4386055806","doi":"10.4103/singaporemedj.smj-2021-230","title":"Modified endocrinology script concordance test: evaluating the reliability and construct validity for assessing clinical reasoning","year":2023,"lang":"en","type":"article","venue":"Singapore Medical Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Concordance; Medicine; Construct validity; Test (biology); Reliability (semiconductor); Construct (python library); Validity; Psychometrics; Clinical psychology; Medical physics; Internal medicine","score_opus":0.16908707732220105,"score_gpt":0.48181717960260545,"score_spread":0.31273010228040443,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4386055806","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9737494,0.00034738609,0.003590141,0.018973533,0.0021058328,0.00041935584,0.0000105371855,0.00013837512,0.0006654583],"genre_scores_gemma":[0.98701364,0.000561983,0.0076285033,0.0023982823,0.0021750946,0.000023451692,0.000020958854,0.000032417534,0.00014569084],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9954466,0.0007594746,0.0014302089,0.0005361451,0.0010889331,0.00073860085],"domain_scores_gemma":[0.8753927,0.12133939,0.00071904715,0.0005626336,0.00071502215,0.0012712321],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.015952751,0.00023268259,0.0007932613,0.00007941099,0.00073083,0.000148217,0.00025057414,0.00031293553,0.00009883958],"category_scores_gemma":[0.6471915,0.00014409248,0.0003183037,0.00030898314,0.0015353161,0.00009797614,0.00017204345,0.0020082057,0.000011134354],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00037627434,0.00025224045,0.5101995,0.00007036572,0.00009776429,0.0005899791,0.00015682765,0.000031030442,0.00011336239,0.00069673767,0.009590926,0.47782502],"study_design_scores_gemma":[0.014950115,0.003915953,0.73320884,0.007778825,0.0009813644,0.014107254,0.0013690031,0.1947941,0.00011990261,0.024992164,0.0032255389,0.00055692386],"about_ca_topic_score_codex":0.000019163544,"about_ca_topic_score_gemma":0.0000028682443,"teacher_disagreement_score":0.63123876,"about_ca_system_score_codex":0.000078204,"about_ca_system_score_gemma":0.0014947547,"threshold_uncertainty_score":0.87247664},"labels":[],"label_agreement":null},{"id":"W4386305507","doi":"10.1111/tct.13613","title":"Mitigating diagnostic performance bias in a skin‐tone balanced dermatology curriculum","year":2023,"lang":"en","type":"article","venue":"The Clinical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Calgary General Hospital; University of Calgary","funders":"Center for Global Health; Cumming School of Medicine, University of Calgary","keywords":"Medicine; Curriculum; Dermatology; Formative assessment; Physical therapy; Psychology","score_opus":0.08792048454297709,"score_gpt":0.42308083999317814,"score_spread":0.33516035545020106,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4386305507","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9846239,0.0001956174,0.00000889699,0.011462782,0.00048665568,0.00039851142,0.0000021895803,0.0002529815,0.0025684827],"genre_scores_gemma":[0.9950421,0.0008215673,0.00017968955,0.0018849943,0.00051983894,0.00008744681,0.00003222323,0.00003275049,0.001399403],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99724466,0.00043963798,0.0010543553,0.000419142,0.0002876024,0.0005546187],"domain_scores_gemma":[0.94663787,0.052290525,0.00018232167,0.00057844736,0.00006325255,0.0002476064],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002504758,0.00019338245,0.00067440304,0.0000778354,0.000087185246,0.000016325122,0.00023732381,0.0002705612,0.00020719784],"category_scores_gemma":[0.20755789,0.00012284587,0.00024460643,0.000697788,0.00038447094,0.00003955432,0.00016008365,0.001113883,0.0016833126],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000044021483,0.00045120678,0.9585871,0.00002349997,0.000036121768,0.00018546953,0.0002443457,0.000034869827,0.000004879117,0.00006434105,0.009980922,0.030343203],"study_design_scores_gemma":[0.0026637793,0.00023904993,0.9819257,0.0014121119,0.00008857025,0.00009496044,0.00032004688,0.0108837625,0.000040404462,0.00039322564,0.0017630722,0.00017526885],"about_ca_topic_score_codex":0.00010614879,"about_ca_topic_score_gemma":0.000018123565,"teacher_disagreement_score":0.20505312,"about_ca_system_score_codex":0.00003264706,"about_ca_system_score_gemma":0.00011100308,"threshold_uncertainty_score":0.999094},"labels":[],"label_agreement":null},{"id":"W4386305553","doi":"10.1007/s40596-023-01849-8","title":"Uncertainty in the Clinic: the Importance of Adaptive Expertise for Dealing with Complexity","year":2023,"lang":"en","type":"editorial","venue":"Academic Psychiatry","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"The Wilson Centre; Sinai Health System","funders":"Ontario HIV Treatment Network","keywords":"Psychology; Medical education; Computer science; Medicine","score_opus":0.07727597283658412,"score_gpt":0.4075323429824317,"score_spread":0.3302563701458476,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4386305553","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.025932994,0.007842312,0.0001554275,0.0201789,0.94125605,0.0027987543,0.0007323514,0.000118850614,0.000984363],"genre_scores_gemma":[0.029783925,0.0053359447,0.0019654005,0.004218738,0.9563968,0.000769776,0.00081245427,0.00017994797,0.0005370198],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99696803,0.00013986476,0.0010916026,0.00054681126,0.0008236713,0.00043002242],"domain_scores_gemma":[0.93464994,0.06354294,0.00076132297,0.0007347312,0.00020333857,0.00010775365],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0025629986,0.0003208701,0.0008407261,0.000054540567,0.00010147474,0.000008219679,0.00064602407,0.0010904651,0.000022941716],"category_scores_gemma":[0.048915822,0.00016046078,0.00029448513,0.00026729362,0.00046391416,0.000023422182,0.00006371846,0.0030559262,0.00001267555],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016038052,0.00009265723,0.014422641,0.00009497816,0.00012509819,0.0000055473665,0.00048781597,0.00001811053,3.7094887e-7,0.0010183795,0.9814735,0.0006570728],"study_design_scores_gemma":[0.022467615,0.005996532,0.051643737,0.048668724,0.0029975,0.000040303905,0.010150196,0.0017847514,0.0000047800836,0.06282583,0.7917499,0.0016701544],"about_ca_topic_score_codex":0.0004525186,"about_ca_topic_score_gemma":0.0007648178,"teacher_disagreement_score":0.18972366,"about_ca_system_score_codex":0.00006965555,"about_ca_system_score_gemma":0.001044859,"threshold_uncertainty_score":0.99924403},"labels":[],"label_agreement":null},{"id":"W4386333182","doi":"10.36834/cmej.75409","title":"Exploring medical students’ use of principles of self-explanation and structured reflection during clerkship","year":2023,"lang":"en","type":"article","venue":"Canadian Medical Education Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Université de Sherbrooke","funders":"","keywords":"Context (archaeology); Medical education; Focus group; Psychology; Theme (computing); Mathematics education; Medicine; Computer science","score_opus":0.11617926405167467,"score_gpt":0.38635591016380066,"score_spread":0.270176646112126,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4386333182","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9909429,0.000089497444,0.000008594138,0.007798304,0.0009097819,0.0001093621,0.000003373516,0.00002426727,0.0001139538],"genre_scores_gemma":[0.9963033,0.0016718428,0.00033368944,0.0009682977,0.0005380883,0.000010215796,0.000023558994,0.00001392356,0.00013705494],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9975106,0.00008925872,0.00059866474,0.0001510103,0.0014263446,0.00022408146],"domain_scores_gemma":[0.9939174,0.0011773338,0.00019656168,0.00013704831,0.00031817958,0.004253465],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001242067,0.000088913235,0.0002490356,0.00048370805,0.00010010349,0.000030871914,0.00013292085,0.00017405715,0.0012553398],"category_scores_gemma":[0.18337022,0.000077982484,0.000057808054,0.00041599502,0.000094964154,0.00013272866,0.0000275276,0.0004817915,0.0000047764306],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018262822,0.0005571854,0.85203266,0.00035946822,0.0003342785,0.0003084608,0.0046378234,0.000033149085,0.00021299315,0.0015146623,0.009967289,0.12985939],"study_design_scores_gemma":[0.0010182227,0.00010942814,0.9904757,0.0023903563,0.00006686084,0.0006537404,0.0007730303,0.00028309767,0.00020367143,0.000067994224,0.0038793795,0.0000784752],"about_ca_topic_score_codex":0.002137223,"about_ca_topic_score_gemma":0.0031748796,"teacher_disagreement_score":0.18212815,"about_ca_system_score_codex":0.00017146484,"about_ca_system_score_gemma":0.010166367,"threshold_uncertainty_score":0.99965763},"labels":[],"label_agreement":null},{"id":"W4386758252","doi":"10.1177/08404704231200113","title":"Addressing the ethical problem of underdiagnosis in the post-pandemic Canadian healthcare system","year":2023,"lang":"en","type":"article","venue":"Healthcare Management Forum","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Regina","funders":"","keywords":"Health care; Medicine; Medical diagnosis; Meaning (existential); Indigenous; Pandemic; Population; Healthcare system; Medical emergency; Nursing; Psychology; Coronavirus disease 2019 (COVID-19); Political science; Environmental health; Law; Pathology","score_opus":0.12144659214709928,"score_gpt":0.4078471990062772,"score_spread":0.2864006068591779,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4386758252","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.11802852,0.0010530356,0.000012943421,0.87190306,0.0003598293,0.0028286374,0.00006573966,0.00019100045,0.005557237],"genre_scores_gemma":[0.9629552,0.0006390898,0.00010033506,0.035617046,0.00008590729,0.00029872244,0.00012107132,0.000028377397,0.0001542555],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99693596,0.0004753634,0.00069350755,0.00037530353,0.00065349427,0.0008663384],"domain_scores_gemma":[0.9941506,0.0045573576,0.00015073497,0.0006525704,0.00015744723,0.0003312904],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0026961456,0.00018379375,0.00040732193,0.00032585114,0.00033047478,0.000035743433,0.00033385956,0.00029166206,0.000013167503],"category_scores_gemma":[0.001641765,0.000110718276,0.00014771389,0.001169472,0.00011896051,0.000024304843,0.000120200384,0.0009884165,0.000070355556],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013841696,0.000117736905,0.7139863,0.004365843,0.00012504973,0.000708125,0.0030256405,0.000067305475,9.544358e-7,0.122908,0.04563461,0.108922035],"study_design_scores_gemma":[0.0016656853,0.0006180283,0.9121979,0.0143111115,0.00014997601,0.00008812973,0.0534801,0.00060547347,0.0000030557421,0.0019100001,0.014727047,0.00024350677],"about_ca_topic_score_codex":0.20396563,"about_ca_topic_score_gemma":0.29816592,"teacher_disagreement_score":0.84492666,"about_ca_system_score_codex":0.00039113333,"about_ca_system_score_gemma":0.0005253831,"threshold_uncertainty_score":0.8013352},"labels":[],"label_agreement":null},{"id":"W4386927947","doi":"10.1186/s12909-023-04541-5","title":"Deliberate practice of diagnostic clinical reasoning reveals low performance and improvement of diagnostic justification in pre-clerkship students","year":2023,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Formative assessment; Medical diagnosis; Medical education; Medicine; Observational study; Clinical Practice; Clinical clerkship; Medical school; Diagnostic test; Medical physics; Curriculum; Psychology; Family medicine; Mathematics education; Pediatrics; Radiology; Internal medicine; Pedagogy","score_opus":0.035036643683278544,"score_gpt":0.43133718239188157,"score_spread":0.39630053870860305,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4386927947","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99616605,0.0005028837,0.00015486157,0.0014707304,0.00067674887,0.0008228879,0.0000041321805,0.00003725189,0.00016443917],"genre_scores_gemma":[0.98885393,0.008336424,0.0011965374,0.0007451494,0.0002620672,0.00019007418,0.000120134086,0.000021432179,0.00027426513],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9965061,0.0002918601,0.0014960906,0.00042959032,0.0010029185,0.00027347],"domain_scores_gemma":[0.8705925,0.12734312,0.000647947,0.00047477442,0.00043114764,0.00051051826],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005174276,0.00015581229,0.00051860884,0.0001728868,0.000037225123,0.000014937821,0.00017694569,0.00024355919,0.00008824796],"category_scores_gemma":[0.7648895,0.00013864815,0.00007386759,0.0005928644,0.00021653867,0.000121736244,0.00010446657,0.00037958237,0.00003089734],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002442465,0.002351936,0.8537474,0.0007800866,0.000030483074,0.000006347117,0.00050810835,0.000031500433,0.00004806272,0.000096037766,0.0019055969,0.14025015],"study_design_scores_gemma":[0.0018857227,0.00067533884,0.9834283,0.010746307,0.00020456783,0.000014114519,0.00039164416,0.0021760296,0.00014279371,0.00008460287,0.00013472814,0.00011585446],"about_ca_topic_score_codex":0.00019102747,"about_ca_topic_score_gemma":0.00002230252,"teacher_disagreement_score":0.7597152,"about_ca_system_score_codex":0.00006242055,"about_ca_system_score_gemma":0.0015810961,"threshold_uncertainty_score":0.5653907},"labels":[],"label_agreement":null},{"id":"W4387407479","doi":"10.1007/s40670-023-01909-6","title":"The Relationship Between Self-Regulated Learning Competency and Clinical Reasoning Tendency in Medical Students","year":2023,"lang":"en","type":"article","venue":"Medical Science Educator","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"Peking University Health Science Center; Peking University; Social Sciences and Humanities Research Council of Canada; Fonds de Recherche du Québec-Société et Culture; China Medical Board","keywords":"Psychology; Virtual patient; Self-regulated learning; Medical education; Mathematics education; Medicine","score_opus":0.05199614830957926,"score_gpt":0.45486171560262184,"score_spread":0.40286556729304257,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4387407479","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97145474,0.00018362008,0.000008308173,0.026302148,0.0009847624,0.00025304046,5.716633e-7,0.00020285614,0.00060994097],"genre_scores_gemma":[0.9977108,0.00042271908,0.00016439441,0.0004907877,0.00068359234,0.000021618805,0.000010898943,0.000015511052,0.00047966317],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99298894,0.00037378838,0.0009575436,0.0006523603,0.0042493595,0.00077802973],"domain_scores_gemma":[0.9017671,0.09361966,0.0002395404,0.0006096932,0.00020203296,0.003562015],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.02082141,0.00016464239,0.00043085622,0.00020836061,0.0006794404,0.000126047,0.0008653395,0.00033184755,0.00017580653],"category_scores_gemma":[0.5497555,0.00010389986,0.00009036374,0.0023455732,0.0019651148,0.00009353134,0.00050332886,0.0017034373,0.0002238898],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000013003786,0.00021677872,0.9737031,0.000007484342,0.00001682083,0.00009311115,0.0002962403,1.7050237e-7,4.4327675e-7,0.0008134292,0.0008035401,0.024035865],"study_design_scores_gemma":[0.001200382,0.00022474586,0.9936134,0.0015114604,0.000049018654,0.000046840556,0.00062660675,0.0010286034,0.0000011172873,0.0005804188,0.0009965844,0.00012081531],"about_ca_topic_score_codex":0.000110507186,"about_ca_topic_score_gemma":0.000053721284,"teacher_disagreement_score":0.5289341,"about_ca_system_score_codex":0.000119941826,"about_ca_system_score_gemma":0.0038678881,"threshold_uncertainty_score":0.74006826},"labels":[],"label_agreement":null},{"id":"W4387445954","doi":"10.3928/01484834-20230815-05","title":"Scoring Methods in Script Concordance Tests: An Exploratory Psychometric Study","year":2023,"lang":"en","type":"article","venue":"Journal of Nursing Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec","funders":"","keywords":"Concordance; Spearman's rank correlation coefficient; Test (biology); Psychology; Clinical psychology; Medicine; Statistics; Mathematics; Internal medicine","score_opus":0.1869778306894005,"score_gpt":0.549873423595076,"score_spread":0.36289559290567547,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4387445954","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9927437,0.0008174794,0.0004171472,0.0011356687,0.004257588,0.00016774438,2.090971e-7,0.000025772517,0.00043471795],"genre_scores_gemma":[0.98533785,0.00012385736,0.013430172,0.00019029221,0.0007567468,0.000007872664,0.0000024657777,0.000016916587,0.00013383378],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9986546,0.00023133814,0.0005453911,0.00014356204,0.00027311646,0.00015201175],"domain_scores_gemma":[0.9967013,0.002332374,0.00028661554,0.00019956716,0.00027907558,0.00020107752],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0023225287,0.00008120167,0.0002886043,0.00094058365,0.000036183763,0.000028313378,0.00008719521,0.000045859637,0.000011856291],"category_scores_gemma":[0.030050369,0.00007067157,0.000055742596,0.0018299583,0.00003450773,0.0001968226,0.0000052904165,0.00031522816,0.0000112731705],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000126718,0.005983666,0.31459612,0.000011633571,0.000017915521,0.00002087798,0.0075312853,0.00003045972,0.0012515476,0.0000173267,0.005961327,0.6644511],"study_design_scores_gemma":[0.0012345529,0.0014378575,0.95946383,0.0045704967,0.00008576967,0.0001047704,0.031558946,0.0000953485,0.00015962563,0.0010344249,0.00017885651,0.00007551384],"about_ca_topic_score_codex":0.000011848818,"about_ca_topic_score_gemma":0.0000025738345,"teacher_disagreement_score":0.6643756,"about_ca_system_score_codex":0.00028491035,"about_ca_system_score_gemma":0.00058264943,"threshold_uncertainty_score":0.9781199},"labels":[],"label_agreement":null},{"id":"W4387604898","doi":"10.56068/clez2974","title":"Paramedicine Contents","year":2023,"lang":"en","type":"article","venue":"International Journal of Paramedicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Nova Scotia Health Authority","funders":"","keywords":"Psychology","score_opus":0.0531289560136395,"score_gpt":0.41767920384001733,"score_spread":0.36455024782637785,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4387604898","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9438387,0.00051376096,0.00064327236,0.04698328,0.004625795,0.00013339109,0.000008495996,0.00006875797,0.0031845458],"genre_scores_gemma":[0.99010724,0.0009613206,0.0004178581,0.0043187086,0.002395161,0.0000028015475,0.00004252692,0.000017836435,0.0017365492],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99752724,0.000030506668,0.0008349494,0.00014095305,0.0012487129,0.00021761056],"domain_scores_gemma":[0.9946712,0.0034375156,0.00038602602,0.00016015796,0.000975987,0.0003690701],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00074066635,0.00013085677,0.00048886874,0.00041380094,0.000020937208,0.0000149908365,0.00025812283,0.00008720123,0.0003972774],"category_scores_gemma":[0.03796029,0.00008897084,0.00020477967,0.00030345412,0.0001646001,0.00006605027,0.00004586555,0.00039079934,0.00030263202],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0039845468,0.0013070722,0.39485216,0.00004268479,0.0023501918,0.023059582,0.00085511315,0.00021061822,0.005124624,0.002025673,0.3866018,0.17958593],"study_design_scores_gemma":[0.028538648,0.0043596355,0.8473875,0.011838891,0.0007140595,0.008821208,0.0010601025,0.0020480435,0.0015793808,0.0047046407,0.0886068,0.00034110885],"about_ca_topic_score_codex":0.000024170913,"about_ca_topic_score_gemma":9.2700094e-7,"teacher_disagreement_score":0.45253533,"about_ca_system_score_codex":0.000076215634,"about_ca_system_score_gemma":0.00013614497,"threshold_uncertainty_score":0.9701434},"labels":[],"label_agreement":null},{"id":"W4387766983","doi":"10.5430/jnep.v14n2p5","title":"The association between nursing students’ perceptions of their clinical reasoning ability and their background of clinical experience: A cross-sectional correlation study","year":2023,"lang":"en","type":"article","venue":"Journal of Nursing Education and Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Nursing; Statistic; Medicine; Perception; Attendance; Correlation; Association (psychology); Cross-sectional study; Psychology","score_opus":0.2033036932555048,"score_gpt":0.5831565853364417,"score_spread":0.37985289208093687,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4387766983","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99481297,0.00028696487,0.00022990312,0.0019615365,0.0019830875,0.0002462112,0.0000041825087,0.000010136401,0.00046500133],"genre_scores_gemma":[0.99776864,0.00075315725,0.0004655793,0.00006023987,0.00079685956,0.0000035474775,0.000007786405,0.00000942445,0.00013478358],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9961999,0.0012104085,0.0017314649,0.00020035883,0.0005241969,0.00013368968],"domain_scores_gemma":[0.92584157,0.070294745,0.002192681,0.00020091467,0.0012765593,0.00019355063],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.013577637,0.00010316377,0.00046100115,0.00009622778,0.00035949078,0.00008530894,0.00009538516,0.00017205549,0.0000103548755],"category_scores_gemma":[0.0976965,0.00006882831,0.00016804867,0.00030186796,0.00037282825,0.0003078953,0.000025151321,0.0006327448,0.0000011815082],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005247566,0.003781003,0.92330426,0.00000640693,0.00015447568,3.8892645e-7,0.010190951,0.000004925004,0.000012885713,0.00003676288,0.00035784623,0.061625365],"study_design_scores_gemma":[0.001252653,0.0009860764,0.91996646,0.0009936843,0.00029583278,0.00010607682,0.07569325,0.00012051248,0.0000027547815,0.00034103607,0.00018964999,0.000051991658],"about_ca_topic_score_codex":0.00001763717,"about_ca_topic_score_gemma":0.0000012055275,"teacher_disagreement_score":0.084118865,"about_ca_system_score_codex":0.00014593198,"about_ca_system_score_gemma":0.0005362275,"threshold_uncertainty_score":0.909904},"labels":[],"label_agreement":null},{"id":"W4388129276","doi":"10.1503/cjs.015422","title":"Considering human cognitive architecture in stressful medical prehospital interventions might benefit care providers","year":2023,"lang":"en","type":"article","venue":"Canadian Journal of Surgery","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Alberta Medical Association; University of Alberta","funders":"Purdue University; National Science Foundation","keywords":"Psychological intervention; Medicine; Intervention (counseling); Cognition; Medical emergency; Psychiatry","score_opus":0.05559079012319228,"score_gpt":0.34215305887374137,"score_spread":0.2865622687505491,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4388129276","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99385494,0.0017555184,0.000011092521,0.0033307793,0.0005036139,0.00012899334,0.000042049433,0.000015292835,0.00035769603],"genre_scores_gemma":[0.9991329,0.00006596997,0.000019913055,0.00033543384,0.00024035134,0.0000060252046,0.000057633388,0.000020559168,0.000121213445],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.998388,0.00006569172,0.00069186866,0.0001528883,0.00032499345,0.00037655633],"domain_scores_gemma":[0.9884367,0.0097577935,0.00017376155,0.00010459107,0.00024164839,0.0012854962],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00069546537,0.00012184403,0.00046057944,0.00081389345,0.00007961485,0.00003332801,0.00007993122,0.00014314559,0.0005904984],"category_scores_gemma":[0.068611674,0.000108022345,0.00040792808,0.00037577358,0.0001819596,0.000056001783,0.000016183574,0.000610977,0.00001577002],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004718671,0.000050893,0.951616,0.00016917067,0.00012378988,0.013766645,0.0022548535,0.00003982247,0.000004397344,0.000044796652,0.007083067,0.024799414],"study_design_scores_gemma":[0.0014878211,0.00023509283,0.9479325,0.04102056,0.00014535183,0.0006920826,0.006628811,0.0000066564357,0.00008540523,0.00041452565,0.0011471489,0.0002040542],"about_ca_topic_score_codex":0.0028285843,"about_ca_topic_score_gemma":0.056053102,"teacher_disagreement_score":0.06791621,"about_ca_system_score_codex":0.00012543699,"about_ca_system_score_gemma":0.0022660112,"threshold_uncertainty_score":0.96117145},"labels":[],"label_agreement":null},{"id":"W4388281474","doi":"10.1016/j.rcsop.2023.100357","title":"Managing common ambulatory conditions: Exploring clinical decision making performance between pharmacists and family physicians","year":2023,"lang":"en","type":"article","venue":"Exploratory Research in Clinical and Social Pharmacy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Alberta; University of Calgary","funders":"","keywords":"Medicine; Ambulatory; Family medicine; Pharmacy; Medical diagnosis; Ambulatory care; Clinical pharmacy; MEDLINE; Clinical Practice; Health care; Internal medicine","score_opus":0.5275714017556576,"score_gpt":0.5877603533277831,"score_spread":0.06018895157212545,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4388281474","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9954058,0.0005151556,0.000012369703,0.0022376222,0.00065404415,0.0005263774,0.000030902844,0.0001919024,0.00042580348],"genre_scores_gemma":[0.97737193,0.01866534,0.00008561411,0.0013525944,0.0022334293,0.00014644217,0.000046911235,0.00005161812,0.00004612162],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99473816,0.0011471874,0.0015575935,0.0008273932,0.0009137922,0.00081587286],"domain_scores_gemma":[0.96616,0.032763325,0.00015454873,0.0002704415,0.00020040988,0.0004512914],"candidate_categories":["research_integrity"],"consensus_categories":[],"category_scores_codex":[0.011022052,0.00024803658,0.000948658,0.00039551864,0.00064812537,0.00009738942,0.00021879574,0.00018766885,0.000023964109],"category_scores_gemma":[0.0058983774,0.00024122596,0.00021894487,0.00095938693,0.0014291113,0.0004252947,0.0005270662,0.002401595,0.00019294424],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00040875367,0.0002192862,0.581747,0.00008440246,0.000054610082,0.00023342053,0.00042935277,0.0000022054546,0.000019557361,0.000060168946,0.0035903454,0.41315088],"study_design_scores_gemma":[0.0064989715,0.0003159097,0.9761846,0.0023623845,0.00007853538,0.0000048120987,0.0014204857,0.002314994,0.0000268708,0.005936041,0.0045487494,0.0003076594],"about_ca_topic_score_codex":0.000010710843,"about_ca_topic_score_gemma":0.000002641713,"teacher_disagreement_score":0.41284323,"about_ca_system_score_codex":0.00005963702,"about_ca_system_score_gemma":0.00020484424,"threshold_uncertainty_score":0.9998999},"labels":[],"label_agreement":null},{"id":"W4388408786","doi":"10.1212/ne9.0000000000200096","title":"Education Research: Neuroradiology Curriculum and Competencies Among Canadian Adult Neurology Residency Programs","year":2023,"lang":"en","type":"article","venue":"Neurology Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université de Montréal","funders":"Royal College of Physicians and Surgeons of Canada","keywords":"Neuroradiology; Neurology; Curriculum; Medicine; Modalities; Medical education; Neuroimaging; Radiology; Medical physics; Psychology; Family medicine; Psychiatry; Pedagogy","score_opus":0.04538162289221758,"score_gpt":0.3760240725966368,"score_spread":0.3306424497044192,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4388408786","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9181159,0.00018818246,8.2300534e-7,0.07539829,0.0024384116,0.0006733883,0.0000017808937,0.00014666784,0.0030365593],"genre_scores_gemma":[0.9871286,0.00053092546,0.00011331529,0.010167763,0.00064970076,0.00032111735,0.00026154367,0.000032782293,0.00079420925],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99751836,0.0005800521,0.00035227116,0.0006507317,0.00021332475,0.0006852632],"domain_scores_gemma":[0.99540967,0.002794394,0.000092029426,0.0004497096,0.0006196948,0.00063448714],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00062671513,0.00016751501,0.00029971218,0.0007425498,0.00026776132,0.000031979278,0.00015256323,0.00034999673,0.00005460835],"category_scores_gemma":[0.02733701,0.00016062615,0.0000648798,0.0008960052,0.00067108986,0.00007158222,0.00006285689,0.0009203888,0.00029107687],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004804524,0.0003277576,0.9122076,0.000022862154,0.000010515151,0.000038771843,0.00021171269,0.0000011995407,0.000029553517,0.0043167807,0.048054356,0.034730818],"study_design_scores_gemma":[0.00030503792,0.0014875599,0.98048186,0.00008837518,0.00004200235,0.00050849613,0.0002589616,0.00029894267,0.0000040829527,0.002631486,0.013785103,0.00010810292],"about_ca_topic_score_codex":0.095761515,"about_ca_topic_score_gemma":0.055172894,"teacher_disagreement_score":0.069012746,"about_ca_system_score_codex":0.000034819805,"about_ca_system_score_gemma":0.0033473477,"threshold_uncertainty_score":0.9808561},"labels":[],"label_agreement":null},{"id":"W4388428327","doi":"10.1521/jsyt.2023.42.2.79","title":"We Tried to Push the River","year":2023,"lang":"en","type":"article","venue":"Journal of Systemic Therapies","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Process (computing); Psychology; Epistemology; Sociology; Computer science; Philosophy","score_opus":0.038687984517536134,"score_gpt":0.3373425917441968,"score_spread":0.29865460722666065,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4388428327","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9644153,0.003353724,0.0000779431,0.031041693,0.0007742146,0.00016955471,0.0000026515636,0.000032743308,0.00013215186],"genre_scores_gemma":[0.99258965,0.0025984908,0.00009963959,0.00090716576,0.0007183475,0.000004234231,5.180173e-7,0.000013728629,0.003068201],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9988629,0.00007715974,0.00047739717,0.000076563265,0.00035136277,0.00015460762],"domain_scores_gemma":[0.9944933,0.004788826,0.00021673016,0.00019087398,0.00016449862,0.00014574929],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010823406,0.00008376864,0.0003463532,0.00011526042,0.00005066096,0.000021136582,0.0001463661,0.0000545026,0.00005421695],"category_scores_gemma":[0.0069845384,0.00004129687,0.00018683416,0.00031759308,0.00005294763,0.000035636804,0.000022320039,0.00019766914,0.00025659808],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0028569994,0.00040421533,0.07323087,0.00027495253,0.0021870884,0.0014355561,0.17347631,0.0003659313,0.035062384,0.0014192703,0.58830184,0.120984614],"study_design_scores_gemma":[0.021163955,0.0064054267,0.38888097,0.07700411,0.0014812597,0.014920302,0.17970438,0.00012894637,0.002138344,0.005972273,0.30128208,0.0009179586],"about_ca_topic_score_codex":0.000009213217,"about_ca_topic_score_gemma":6.6982636e-7,"teacher_disagreement_score":0.3156501,"about_ca_system_score_codex":0.000037683443,"about_ca_system_score_gemma":0.00010002325,"threshold_uncertainty_score":0.83616495},"labels":[],"label_agreement":null},{"id":"W4388930027","doi":"10.2196/50797","title":"Transformation and Articulation of Clinical Data to Understand Students’ and Health Professionals’ Clinical Reasoning: Protocol for a Scoping Review","year":2023,"lang":"en","type":"review","venue":"JMIR Research Protocols","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières; Université Laval; Université de Montréal","funders":"Social Sciences and Humanities Research Council of Canada; Fonds de Recherche du Québec-Société et Culture","keywords":"PsycINFO; CINAHL; Data extraction; Protocol (science); Checklist; MEDLINE; Systematic review; Medical education; Psychology; Computer science; Medicine; Alternative medicine","score_opus":0.9349515214085985,"score_gpt":0.8169460549232435,"score_spread":0.11800546648535504,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4388930027","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000011678801,0.066549234,0.00010370746,0.0023986886,0.00001383643,0.93076354,0.00009353999,0.00005237517,0.000023878702],"genre_scores_gemma":[3.464485e-8,0.34577134,0.00082465186,0.00026840065,0.00020781616,0.6525105,0.00025688653,0.000050350965,0.00010999085],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.98918116,0.0025032114,0.004886607,0.0012099531,0.0015844255,0.00063465995],"domain_scores_gemma":[0.9717572,0.024330145,0.0011266302,0.0012587211,0.00056477217,0.00096258096],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.04373074,0.00034837893,0.00380864,0.00027496673,0.00020354542,0.00009327352,0.00050505326,0.00050044665,0.000019721221],"category_scores_gemma":[0.08139083,0.00024132385,0.0003388919,0.0007630235,0.00030148934,0.00014843393,0.00074145506,0.0013019068,0.000017036375],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004456052,0.00094436924,0.00048477243,0.4677795,0.000094285766,0.000004338977,0.000059164402,1.7429208e-8,1.0243503e-8,0.000100802514,0.023915047,0.5061721],"study_design_scores_gemma":[0.0033054464,0.0030723626,0.00072072155,0.72646546,0.000074451986,0.0000099816525,0.000035911493,0.000025990008,6.3664956e-8,0.000043426917,0.26610175,0.00014441943],"about_ca_topic_score_codex":0.000010938655,"about_ca_topic_score_gemma":0.000020998552,"teacher_disagreement_score":0.5060277,"about_ca_system_score_codex":0.000094879295,"about_ca_system_score_gemma":0.004271077,"threshold_uncertainty_score":0.9846804},"labels":[],"label_agreement":null},{"id":"W4389046786","doi":"10.1177/08404704231215750","title":"The impact of cognitive biases, mental models, and mindsets on leadership and change in the health system","year":2023,"lang":"en","type":"article","venue":"Healthcare Management Forum","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia; Dalhousie University","funders":"","keywords":"Psychology; Mental health; Cognition; Metacognition; Cognitive bias; Systems thinking; Debiasing; Cognitive psychology; Social psychology; Applied psychology; Psychotherapist; Computer science; Psychiatry","score_opus":0.20667587155047382,"score_gpt":0.43396711969635093,"score_spread":0.22729124814587712,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4389046786","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.840546,0.0051144413,0.0000042093197,0.15046224,0.00013455354,0.0027827756,0.00011973216,0.000056794484,0.00077923795],"genre_scores_gemma":[0.9926351,0.003471504,0.0000051721395,0.0035417175,0.00003691427,0.0001334417,0.00009368349,0.000012096017,0.00007038622],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9986123,0.00017743488,0.00028565744,0.0002252316,0.0002740458,0.00042536302],"domain_scores_gemma":[0.99469906,0.004849332,0.000102195525,0.00018879694,0.000023197064,0.00013743821],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001101546,0.00012072172,0.00024840928,0.00012086507,0.00016700297,0.000019186393,0.000060515253,0.000044299893,0.0000013639421],"category_scores_gemma":[0.0008884462,0.00006539329,0.000055295917,0.00028671324,0.0000849085,0.000026437212,0.00006594948,0.00015799452,0.000005131602],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017450376,0.0004632935,0.16644216,0.002048157,0.0002800053,0.00026941407,0.01748355,0.0000132979285,1.9629043e-7,0.019104667,0.02367872,0.7684715],"study_design_scores_gemma":[0.0028726147,0.0033984398,0.88098377,0.017657317,0.000044473116,0.000033397522,0.08707632,0.007368171,0.0000010921493,0.00033659558,0.00011105222,0.000116747215],"about_ca_topic_score_codex":0.004068212,"about_ca_topic_score_gemma":0.000764058,"teacher_disagreement_score":0.7683548,"about_ca_system_score_codex":0.00009433187,"about_ca_system_score_gemma":0.000036783356,"threshold_uncertainty_score":0.61499465},"labels":[],"label_agreement":null},{"id":"W4389057877","doi":"10.1007/s10459-023-10302-2","title":"Implicit versus explicit first impressions in performance-based assessment: will raters overcome their first impressions when learner performance changes?","year":2023,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital; Medical Council of Canada; University of Alberta; University of Ottawa","funders":"Royal College of Physicians and Surgeons of Canada","keywords":"Psychology; Impression formation; Context (archaeology); Social psychology; Affect (linguistics); Social perception; Perception; Communication","score_opus":0.04751449494962311,"score_gpt":0.41607243429832397,"score_spread":0.36855793934870085,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4389057877","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97549385,0.0013651219,0.000030428757,0.01854926,0.0022632717,0.0008200651,0.000012420071,0.00010601722,0.0013595565],"genre_scores_gemma":[0.9804096,0.016133625,0.00092699536,0.0014102015,0.00022450459,0.00056635577,0.00009412161,0.000020287618,0.0002142804],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9972773,0.00009044988,0.0006575837,0.00065986154,0.000488283,0.0008264828],"domain_scores_gemma":[0.9940376,0.0047957967,0.00030798925,0.00043516504,0.00008035612,0.00034307883],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001474815,0.00024212853,0.00042810783,0.0005978885,0.0006455859,0.000046393707,0.00033071855,0.00011208472,0.00012601232],"category_scores_gemma":[0.002620433,0.0001860586,0.00005829293,0.0017724966,0.00023183781,0.0007125594,0.00008860548,0.00043803806,0.00003350824],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021601433,0.0005337016,0.8687294,0.0005538727,0.000004573205,0.0000019731401,0.0018925617,0.014165749,0.000037970854,0.00020516175,0.00119819,0.112460814],"study_design_scores_gemma":[0.0025396727,0.0015481629,0.905895,0.0094427625,0.000010619051,0.000005020758,0.0021848846,0.055142704,0.00012223513,0.0002202823,0.022562953,0.00032568665],"about_ca_topic_score_codex":0.0007784645,"about_ca_topic_score_gemma":0.0012867667,"teacher_disagreement_score":0.11213513,"about_ca_system_score_codex":0.0004642287,"about_ca_system_score_gemma":0.0016712329,"threshold_uncertainty_score":0.7587249},"labels":[],"label_agreement":null},{"id":"W4389545477","doi":"10.1515/dx-2023-0099","title":"Assessing clinical reasoning skills following a virtual patient dizziness curriculum","year":2023,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital","funders":"","keywords":"Curriculum; Virtual patient; Psychology; Medical education; Medicine; Pedagogy","score_opus":0.03751000390727223,"score_gpt":0.40504917770445886,"score_spread":0.36753917379718665,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4389545477","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9929899,0.00022927695,0.00011190915,0.0016462125,0.0024263472,0.00042171695,0.000010629677,0.000621076,0.0015429266],"genre_scores_gemma":[0.9952426,0.0005678559,0.0010743696,0.001715348,0.00067785534,0.00027693782,0.00015806507,0.00007084239,0.00021615799],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9966491,0.00020638737,0.0010046908,0.0007495251,0.0006983075,0.0006919951],"domain_scores_gemma":[0.9781467,0.020281995,0.00022933805,0.0005767016,0.00015434256,0.00061095523],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001273458,0.00030384213,0.00083096937,0.0002032412,0.00022089844,0.00016184118,0.00016295697,0.00028736913,0.00015095671],"category_scores_gemma":[0.121633135,0.00025752766,0.0007625665,0.0011347779,0.00010689488,0.00022671405,0.00021120522,0.0005204148,0.0008692542],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000013007392,0.0011246416,0.8142036,0.000011494752,0.00012225342,0.0007594793,0.00016038849,0.000026247119,0.000006629913,0.00011079189,0.02003157,0.16342987],"study_design_scores_gemma":[0.0026686464,0.0008286817,0.98023015,0.0061120624,0.00046333592,0.000029459821,0.0011306808,0.0013307076,0.00023043931,0.00011337714,0.006401886,0.0004605433],"about_ca_topic_score_codex":0.000111145724,"about_ca_topic_score_gemma":0.0000058606533,"teacher_disagreement_score":0.16602655,"about_ca_system_score_codex":0.00008076766,"about_ca_system_score_gemma":0.00018426587,"threshold_uncertainty_score":0.9999877},"labels":[],"label_agreement":null},{"id":"W4389733788","doi":"10.3399/bjgpo.2023.0191","title":"Learning to navigate uncertainty in primary care: a scoping literature review","year":2023,"lang":"en","type":"article","venue":"BJGP Open","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Queen's University","keywords":"Ambiguity; Context (archaeology); Constructive; Psychology; MEDLINE; Medical education; Primary care; Medicine; Computer science; Process (computing); Family medicine","score_opus":0.0352992802809504,"score_gpt":0.4006081720588464,"score_spread":0.365308891777896,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4389733788","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8203348,0.08723876,0.0000064530723,0.028722582,0.0005773427,0.0072018867,0.000029573432,0.000406779,0.05548184],"genre_scores_gemma":[0.6304287,0.2107376,0.0041582985,0.113702856,0.0010814096,0.0018238815,0.0047752047,0.00024248761,0.033049576],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99879247,0.00008965029,0.00031611026,0.00033503352,0.00019765826,0.00026908854],"domain_scores_gemma":[0.99769145,0.0016929187,0.00004969813,0.00028256266,0.0000850332,0.00019834554],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00075479713,0.00011523821,0.0004709057,0.00006853189,0.000049403643,0.000073524076,0.00020531571,0.00007390352,0.000059268106],"category_scores_gemma":[0.019939542,0.0000938055,0.00006863128,0.0010438869,0.000015237423,0.000058210222,0.00034293078,0.0003748496,0.00047805242],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007179449,0.00027152686,0.094118625,0.02028703,0.000087632536,0.0045982827,0.009912718,0.0013836406,0.00034195543,0.00016547437,0.09175856,0.77635664],"study_design_scores_gemma":[0.0019213002,0.00039514917,0.07587851,0.89253175,0.0000561316,0.00003812506,0.0002375709,0.00005572271,0.000022931677,0.00004479953,0.028578777,0.0002392219],"about_ca_topic_score_codex":0.0001503379,"about_ca_topic_score_gemma":0.000029678497,"teacher_disagreement_score":0.8722447,"about_ca_system_score_codex":0.000104399005,"about_ca_system_score_gemma":0.00023089867,"threshold_uncertainty_score":0.98831594},"labels":[],"label_agreement":null},{"id":"W4390033366","doi":"10.52358/mm.vi13.380","title":"Accompagner au changement des pratiques de médecins généralistes en formation en santé : Une approche orientée par l’activité et l’acceptabilité des usages du numérique en contexte de crise","year":2023,"lang":"fr","type":"article","venue":"Médiations et médiatisations","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Humanities; Political science; Philosophy","score_opus":0.04367734264798373,"score_gpt":0.3604374784349467,"score_spread":0.316760135786963,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4390033366","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8682606,0.0018464844,0.085409686,0.033535548,0.0007370113,0.0020084344,0.00084832765,0.00061277085,0.006741096],"genre_scores_gemma":[0.96442556,0.011085279,0.014677677,0.0021570025,0.0013727277,0.0017787801,0.0016219387,0.000120794524,0.0027602264],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99287796,0.0031565682,0.0013030256,0.00070996425,0.0006346308,0.0013178797],"domain_scores_gemma":[0.95118934,0.046334006,0.00051032193,0.00075159856,0.00052483217,0.0006898718],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0057405233,0.0005986542,0.0007719884,0.0005442606,0.0009885475,0.0003330736,0.00037461947,0.00058317155,0.0007162726],"category_scores_gemma":[0.08952876,0.0006273998,0.00036637054,0.0017334654,0.0005593621,0.0013596656,0.00024008282,0.0007887948,0.0003961469],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009023466,0.0026503992,0.503734,0.0006482165,0.0004967691,0.00005383791,0.34255818,0.011766454,0.001034822,0.054416005,0.023418872,0.059132215],"study_design_scores_gemma":[0.0022935148,0.00041949196,0.9214936,0.0018531649,0.0005039042,0.000056538447,0.033742964,0.014809525,0.0015517992,0.013384032,0.009272632,0.0006188217],"about_ca_topic_score_codex":0.005503884,"about_ca_topic_score_gemma":0.011511862,"teacher_disagreement_score":0.4177596,"about_ca_system_score_codex":0.0023006313,"about_ca_system_score_gemma":0.0020428493,"threshold_uncertainty_score":0.99961776},"labels":[],"label_agreement":null},{"id":"W4390262114","doi":"10.1016/j.jns.2023.121375","title":"Virtual cognitive testing in the BHSP- neuropsychology, MyCogHealth and Cogniciti","year":2023,"lang":"en","type":"article","venue":"Journal of the Neurological Sciences","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Parkwood Institute; Lawson Health Research Institute; University of British Columbia Hospital; Institut Universitaire de Gériatrie de Montréal; ECW Press (Canada); Baycrest Hospital; University of Victoria","funders":"","keywords":"Neuropsychology; Psychology; Neuropsychological testing; Cognitive psychology; Cognition; Cognitive neuropsychology; Cognitive science; Neuroscience","score_opus":0.11912575563777239,"score_gpt":0.4070333950838926,"score_spread":0.2879076394461202,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4390262114","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96109194,0.00007844664,0.0000025794404,0.03739199,0.00033212587,0.00014695934,0.000001270139,0.000010283471,0.000944398],"genre_scores_gemma":[0.98170376,0.00012478893,0.000037237558,0.01795381,0.00015669128,0.0000021541484,1.05842375e-7,0.000002360163,0.000019080007],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9984396,0.00042782418,0.00033328286,0.00016485195,0.00039978523,0.00023464144],"domain_scores_gemma":[0.9645759,0.03496483,0.00021496363,0.0000729932,0.000069133486,0.000102197926],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002855712,0.00007145168,0.0002010469,0.000066798246,0.0001750434,0.00003350677,0.00028355705,0.00005146443,0.00001003228],"category_scores_gemma":[0.114732765,0.000027437518,0.00006774608,0.00096235226,0.0007986658,0.000038244747,0.00008931686,0.0005956673,0.000008138077],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00022314346,0.00021079314,0.963585,0.0000030026717,0.000005778951,0.0006333731,0.0002428311,0.0000756521,0.0001755804,0.00021686488,0.0012184067,0.033409577],"study_design_scores_gemma":[0.00063742494,0.004136025,0.9910639,0.0001984684,0.000027792219,0.0012906743,0.00026023024,0.00047647688,0.0000056836984,0.0017694451,0.00010299752,0.000030881187],"about_ca_topic_score_codex":0.0000069915204,"about_ca_topic_score_gemma":0.0000022867607,"teacher_disagreement_score":0.11187705,"about_ca_system_score_codex":0.000003066061,"about_ca_system_score_gemma":0.000074443946,"threshold_uncertainty_score":0.8927242},"labels":[],"label_agreement":null},{"id":"W4390262187","doi":"10.1016/j.jns.2023.121402","title":"The autonomous cognitive examination: Preliminary clinical trial results","year":2023,"lang":"en","type":"article","venue":"Journal of the Neurological Sciences","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Winnipeg; University of Manitoba","funders":"","keywords":"Cognition; Medicine; Psychology; Psychiatry","score_opus":0.15170171164452176,"score_gpt":0.4309654843332705,"score_spread":0.2792637726887487,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4390262187","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9430347,0.00009360023,0.00000382687,0.052140284,0.0024591638,0.00023161709,0.0000020590658,0.0000222981,0.0020124416],"genre_scores_gemma":[0.99600565,0.00036709328,0.000037743168,0.0019279981,0.00097266014,0.0000033828076,3.769091e-7,0.0000028478557,0.0006822617],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9972817,0.00073054084,0.000855996,0.00019873782,0.0006995558,0.00023348352],"domain_scores_gemma":[0.92029923,0.07854787,0.0006264394,0.00016172658,0.00018722759,0.00017750284],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0093811415,0.00007983408,0.0002550256,0.00003771048,0.00043400374,0.00006385418,0.00047933962,0.00009905531,0.000013926928],"category_scores_gemma":[0.29053897,0.000028684855,0.0003232747,0.0005707873,0.0012077349,0.000048754042,0.00016817018,0.00059113675,0.00003616615],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.14934872,0.0020301421,0.086927705,0.0000061308583,0.00013418146,0.0012894623,0.0003014113,0.0003152434,0.000022611952,0.0007396783,0.058585905,0.7002988],"study_design_scores_gemma":[0.012902556,0.020698825,0.9601009,0.00014396262,0.000094547984,0.00024822197,0.00008501121,0.00083606463,0.000014603153,0.0016639996,0.003165439,0.000045901656],"about_ca_topic_score_codex":0.0000010083588,"about_ca_topic_score_gemma":6.808606e-7,"teacher_disagreement_score":0.8731732,"about_ca_system_score_codex":0.000007710376,"about_ca_system_score_gemma":0.00021727252,"threshold_uncertainty_score":0.7154371},"labels":[],"label_agreement":null},{"id":"W4390265350","doi":"10.21449/ijate.1406304","title":"A dialectic on validity: Explanation-focused and the many ways of being human","year":2023,"lang":"en","type":"article","venue":"International Journal of Assessment Tools in Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"Social Sciences and Humanities Research Council of Canada; Canada Research Chairs; University of Cambridge","keywords":"Test (biology); External validity; Classical test theory; Test validity; Epistemology; Construct validity; Psychology; Variation (astronomy); Dialectic; Predictive validity; Cognitive psychology; Item response theory; Social psychology; Psychometrics","score_opus":0.10254702649635265,"score_gpt":0.44019930065982515,"score_spread":0.3376522741634725,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4390265350","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95731854,0.00001967547,0.000231523,0.03894414,0.00095365435,0.00014025498,0.0000017388722,0.000006872956,0.0023835858],"genre_scores_gemma":[0.99710524,0.0001796507,0.0004955107,0.0015895814,0.0005075996,0.0000137068155,0.00002255016,0.0000067252195,0.000079444675],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9987117,0.00008475195,0.00050524325,0.00008572622,0.0005383704,0.00007420889],"domain_scores_gemma":[0.99284613,0.006334102,0.00034606768,0.00008928182,0.00034454366,0.00003986158],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0011588655,0.0000626771,0.00018975312,0.00031433656,0.000029105768,0.00004413916,0.00012301617,0.000041569263,0.000032737793],"category_scores_gemma":[0.009446448,0.000043545962,0.000075199896,0.0001649813,0.000058080022,0.00010330923,0.000021545098,0.0002544397,0.0000029532093],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013104354,0.0025637697,0.5657342,0.00006549699,0.000550053,0.00016204652,0.0038356676,0.00032624463,0.0017146745,0.21957794,0.019562136,0.18459736],"study_design_scores_gemma":[0.004426183,0.00051884213,0.95576,0.0019905153,0.00009262128,0.0000764872,0.00072220265,0.0005209153,0.00048220856,0.03468309,0.0006538078,0.000073098316],"about_ca_topic_score_codex":0.000034504337,"about_ca_topic_score_gemma":0.0000034948396,"teacher_disagreement_score":0.39002585,"about_ca_system_score_codex":0.00014401494,"about_ca_system_score_gemma":0.00032051996,"threshold_uncertainty_score":0.99889743},"labels":[],"label_agreement":null},{"id":"W4390957117","doi":"10.5334/ijic.icic23534","title":"Improving Diagnostic Safety through Integrated Care","year":2023,"lang":"en","type":"article","venue":"International Journal of Integrated Care","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto East General Hospital","funders":"","keywords":"Health care; Patient safety; Context (archaeology); Diagnostic test; Medicine; Medical emergency; Risk analysis (engineering); Pediatrics; Political science","score_opus":0.01817632925709825,"score_gpt":0.3394153191695875,"score_spread":0.3212389899124892,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4390957117","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9766324,0.0037481766,0.0029296663,0.0034004848,0.006410457,0.0003417055,0.00045955318,0.0002470954,0.005830434],"genre_scores_gemma":[0.9945297,0.0011008681,0.001414018,0.00089606707,0.00075618573,0.000008242243,0.00088199327,0.000045685858,0.0003671954],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.9973985,0.00008586353,0.0009967951,0.00026510505,0.0009479583,0.0003057914],"domain_scores_gemma":[0.9888667,0.0053204116,0.0004626089,0.00022486392,0.0048850286,0.0002403777],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00029632845,0.00025969723,0.0005173766,0.00036567918,0.00006886912,0.0000911006,0.0004521569,0.00021007037,0.00039812995],"category_scores_gemma":[0.07209869,0.00018719106,0.00044351837,0.0006040908,0.00010871008,0.00018404321,0.0000879871,0.0008664096,0.000178784],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0036344547,0.0005103055,0.23516965,0.00018579111,0.0020330837,0.013373088,0.012667475,0.0013924914,0.0034253013,0.0012581279,0.04088742,0.68546283],"study_design_scores_gemma":[0.036340803,0.007912039,0.20355393,0.05821053,0.0026209836,0.00705275,0.38488302,0.002824515,0.041580383,0.0019263921,0.25067374,0.0024209253],"about_ca_topic_score_codex":0.00076142885,"about_ca_topic_score_gemma":0.00006841609,"teacher_disagreement_score":0.6830419,"about_ca_system_score_codex":0.0011430703,"about_ca_system_score_gemma":0.0015221302,"threshold_uncertainty_score":0.9357174},"labels":[],"label_agreement":null},{"id":"W4391288612","doi":"10.1016/j.nedt.2024.106117","title":"Proposal of a digital tool to develop clinical reasoning of novice nurses in psychiatry: An exploratory study in France","year":2024,"lang":"en","type":"article","venue":"Nurse Education Today","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval","funders":"","keywords":"Interview; Psychology; Feeling; Medical education; Exploratory research; Forensic psychiatry; Process (computing); Medicine; Psychiatry; Computer science; Social psychology","score_opus":0.025022715892269246,"score_gpt":0.41289213268583413,"score_spread":0.3878694167935649,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4391288612","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9922825,0.00032560708,0.000020400304,0.0025191987,0.0033493566,0.0008213325,0.000011436918,0.000038047932,0.0006321289],"genre_scores_gemma":[0.9963821,0.000009003645,0.0025103036,0.00023971687,0.00031175246,0.00010966183,0.000031341235,0.000024008825,0.00038207034],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99806195,0.0001132226,0.0009606199,0.00040979907,0.00028134702,0.0001730665],"domain_scores_gemma":[0.9975611,0.0015441898,0.00011116185,0.00035513585,0.0002688244,0.00015960855],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.000912447,0.00013227657,0.00042293905,0.0002677782,0.000011352861,0.000024189658,0.000116298426,0.00010533079,0.00004194409],"category_scores_gemma":[0.023156399,0.00011817095,0.0000524349,0.0012224419,0.00006964111,0.00018552976,0.000016768752,0.0002825158,0.000038428618],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015464418,0.008007537,0.87939215,0.000067897454,0.000019826399,0.000007375609,0.008433242,0.0000074258483,0.000039813138,0.0010246756,0.0024223484,0.10042308],"study_design_scores_gemma":[0.0011960649,0.0013786175,0.9729481,0.0057284674,0.000049231654,0.0000058828073,0.017063465,0.00008582705,0.00007456979,0.00038388607,0.00092761574,0.00015831359],"about_ca_topic_score_codex":0.00008871321,"about_ca_topic_score_gemma":0.00020948405,"teacher_disagreement_score":0.10026477,"about_ca_system_score_codex":0.00005710239,"about_ca_system_score_gemma":0.003899539,"threshold_uncertainty_score":0.98507196},"labels":[],"label_agreement":null},{"id":"W4391450505","doi":"10.1111/jep.13969","title":"The inseparability of context and clinical reasoning","year":2024,"lang":"en","type":"article","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Canadian Medical Protective Association; University of Toronto","funders":"","keywords":"Context (archaeology); Medical diagnosis; Cognition; Process (computing); Psychology; Health care; Cognitive science; Knowledge management; Management science; Cognitive psychology; Computer science; Medicine; Psychiatry","score_opus":0.12383862511434322,"score_gpt":0.5554030938006129,"score_spread":0.4315644686862697,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4391450505","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9532138,0.014973755,0.0002786493,0.02576339,0.002387677,0.00036900202,0.0000016506839,0.00000911658,0.0030029686],"genre_scores_gemma":[0.9889175,0.0067853983,0.0026641868,0.0009099056,0.0006485569,0.0000031600648,9.873621e-7,0.000009994034,0.000060258353],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99254286,0.0023997359,0.003709855,0.00023935501,0.00096788834,0.00014027492],"domain_scores_gemma":[0.41709068,0.5774187,0.0020035403,0.00048637317,0.002628316,0.00037242452],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.07838779,0.000094111616,0.00058015226,0.00007150939,0.000044728175,0.00005485925,0.000089448935,0.00020634018,0.000062721374],"category_scores_gemma":[0.84196234,0.000056641107,0.00034429316,0.00025818316,0.00041981158,0.00026262156,0.000046117537,0.0014407173,0.000011099039],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00254446,0.0009928385,0.2668256,0.000034444714,0.00031012893,0.0001079787,0.00013830706,0.000022259379,0.0000027560395,0.0015751823,0.004742343,0.7227037],"study_design_scores_gemma":[0.0049899523,0.0032329662,0.92088336,0.0041267476,0.0017999657,0.0006607337,0.00091461564,0.023902096,0.000005764807,0.005180891,0.03420127,0.00010162915],"about_ca_topic_score_codex":0.000027932036,"about_ca_topic_score_gemma":0.000016899521,"teacher_disagreement_score":0.76357454,"about_ca_system_score_codex":0.00006909804,"about_ca_system_score_gemma":0.001087311,"threshold_uncertainty_score":0.9489937},"labels":[],"label_agreement":null},{"id":"W4391576736","doi":"10.1503/cjs.003523","title":"Characteristics and contributing factors of diagnostic error in surgery: analysis of closed medico-legal cases and complaints in Canada","year":2024,"lang":"en","type":"article","venue":"Canadian Journal of Surgery","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"Canadian Medical Protective Association; Queensway-Carleton Hospital; University of Toronto; Canadian Association of General Surgeons; University of Ottawa; Canadian Medical Association","funders":"VA National Center for Patient Safety; Agency for Healthcare Research and Quality; Canadian Institutes of Health Research; U.S. Department of Veterans Affairs","keywords":"Medicine; Complaint; Harm; Health care; Orthopedic surgery; Patient safety; Surgical team; General surgery; Medical emergency; Surgery","score_opus":0.04363269456097407,"score_gpt":0.29178418905645803,"score_spread":0.24815149449548396,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4391576736","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9954917,0.0030545602,0.0000024803066,0.000589583,0.00044106756,0.000057456953,0.00035286075,0.0000014153635,0.00000886866],"genre_scores_gemma":[0.9992892,0.0005001288,0.0000059151057,0.00009918029,0.000044437762,6.564388e-7,0.000047139776,0.000010057735,0.000003283734],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99805707,0.00009677602,0.0011669105,0.000135189,0.00023173579,0.0003123321],"domain_scores_gemma":[0.88625485,0.112501994,0.00031954426,0.00009088138,0.00015486949,0.0006778414],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0013563245,0.00012461228,0.0013110515,0.0012968498,0.000020880765,0.00001620826,0.000037482896,0.00007130616,0.00006915087],"category_scores_gemma":[0.15261963,0.00010680811,0.00014864067,0.00078894966,0.00014507143,0.000051095325,0.000008911262,0.00027493888,6.865416e-8],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000032733617,0.000018505854,0.98567164,0.00010636443,0.00027260478,0.00892765,0.00014452508,0.000007024293,0.0000060737466,0.000012205507,0.0005064691,0.004294219],"study_design_scores_gemma":[0.00021714496,0.000027897695,0.9932932,0.004785251,0.00051248726,0.00020651256,0.00039419672,0.00027886115,0.00003307152,0.00000573063,0.00016167754,0.00008395058],"about_ca_topic_score_codex":0.86241996,"about_ca_topic_score_gemma":0.9469945,"teacher_disagreement_score":0.15126331,"about_ca_system_score_codex":0.00028786765,"about_ca_system_score_gemma":0.0073510855,"threshold_uncertainty_score":0.9982763},"labels":[],"label_agreement":null},{"id":"W4391878439","doi":"10.1136/bjo-2023-325053","title":"Assessing the medical reasoning skills of GPT-4 in complex ophthalmology cases","year":2024,"lang":"en","type":"article","venue":"British Journal of Ophthalmology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":54,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; University of Waterloo; McGill University; Centre Hospitalier de l’Université de Montréal; Institut Universitaire en Santé Mentale de Québec; Université de Montréal; Hôpital du Sacré-Cœur de Montréal; Institut universitaire en santé mentale de Montréal; Hôpital Maisonneuve-Rosemont","funders":"","keywords":"Medicine; Optometry; Ophthalmology","score_opus":0.060085503448622866,"score_gpt":0.42406343293523874,"score_spread":0.3639779294866159,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4391878439","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9813557,0.009875014,0.000047133686,0.0019182904,0.0008597608,0.00012574925,0.000008736615,0.0000117702675,0.0057978183],"genre_scores_gemma":[0.9975375,0.00034008382,0.0012288973,0.00017219794,0.00047795728,0.00000419162,0.000011689534,0.000029844809,0.00019762092],"study_design_codex":"case_report","study_design_gemma":"case_report","domain_scores_codex":[0.9969342,0.00048872543,0.0013231466,0.00026308565,0.00059832056,0.0003925754],"domain_scores_gemma":[0.97942835,0.019499319,0.00033594755,0.00019843623,0.0002599555,0.0002779799],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0027857635,0.00016189505,0.0009050403,0.00022881305,0.00006208109,0.000088802844,0.00030860427,0.00032095407,0.0025956158],"category_scores_gemma":[0.07489123,0.00012888045,0.00031467574,0.0004316614,0.00056930305,0.0001302955,0.00010787482,0.0011978302,0.000012219782],"study_design_candidate":"case_report","study_design_consensus":"case_report","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011467081,0.0008897771,0.14935568,0.00012294535,0.00021361571,0.8118492,0.00014277881,0.000032431122,0.00016478388,0.00032872852,0.002518486,0.03426696],"study_design_scores_gemma":[0.0007900648,0.0005626087,0.35728636,0.0060166023,0.00008051741,0.6341794,0.00015701403,0.00018790536,0.00003227792,0.00046151373,0.0001736469,0.000072057665],"about_ca_topic_score_codex":0.00040804487,"about_ca_topic_score_gemma":0.0000031755524,"teacher_disagreement_score":0.20793068,"about_ca_system_score_codex":0.000072179704,"about_ca_system_score_gemma":0.00076715153,"threshold_uncertainty_score":0.99831617},"labels":[],"label_agreement":null},{"id":"W4392171073","doi":"10.2196/56933","title":"Definitions and Measurements for Atypical Presentations at Risk for Diagnostic Errors in Internal Medicine: Protocol for a Scoping Review","year":2024,"lang":"en","type":"review","venue":"JMIR Research Protocols","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Agency for Healthcare Research and Quality","keywords":"Protocol (science); Presentation (obstetrics); Medicine; Medical physics; Alternative medicine; Pathology; Surgery","score_opus":0.7516703503714239,"score_gpt":0.704179764691803,"score_spread":0.047490585679620945,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4392171073","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[7.489181e-8,0.23455626,0.00022795293,0.0008130019,0.000019604855,0.7638594,0.000336536,0.00004780536,0.00013938086],"genre_scores_gemma":[1.805688e-8,0.34905556,0.0009429961,0.00007486671,0.00035686573,0.6488268,0.00029717598,0.00009718964,0.00034851625],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9942043,0.0007161404,0.0020157923,0.0011466368,0.000977425,0.0009397012],"domain_scores_gemma":[0.9191514,0.07851858,0.00042683363,0.00059233647,0.0007916392,0.0005191684],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.006120629,0.0005268845,0.0025057588,0.0005733965,0.0003145973,0.000089630514,0.00035340048,0.00042919652,0.00010667292],"category_scores_gemma":[0.29153508,0.0003587824,0.0007319405,0.0006866617,0.0004321306,0.000068698006,0.000301691,0.0011618979,0.00003974359],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010609935,0.00064428453,0.0002121636,0.6720784,0.00016222238,0.000014926998,0.000038639453,9.285663e-8,2.2095622e-7,0.00026675838,0.17330039,0.15222096],"study_design_scores_gemma":[0.0043012616,0.0016388749,0.000027506094,0.56056434,0.0003777329,0.000023768258,0.0000050303065,0.000012040101,6.4051176e-7,0.00088142726,0.4319964,0.00017097556],"about_ca_topic_score_codex":0.000023920004,"about_ca_topic_score_gemma":0.00011790659,"teacher_disagreement_score":0.28541443,"about_ca_system_score_codex":0.000413176,"about_ca_system_score_gemma":0.0017344079,"threshold_uncertainty_score":0.9998864},"labels":[],"label_agreement":null},{"id":"W4392549803","doi":"10.1007/s11019-024-10198-8","title":"Discovering clinical phronesis","year":2024,"lang":"en","type":"article","venue":"Medicine Health Care and Philosophy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; McGill University Health Centre","funders":"McGill University","keywords":"Phronesis; Medical law; Philosophy of medicine; Philosophy of biology; Epistemology; Philosophy; Medicine; Pathology; Philosophy of science; Alternative medicine","score_opus":0.06265100706831563,"score_gpt":0.444450809274521,"score_spread":0.3817998022062054,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4392549803","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.27692503,0.27409732,0.0004542242,0.3904228,0.0049103466,0.0008952855,0.000038220325,0.00056654663,0.051690243],"genre_scores_gemma":[0.96527445,0.009108254,0.0003047561,0.020516852,0.0045373095,0.000014002372,0.00007783344,0.00003020817,0.00013630823],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9983414,0.00006759486,0.00060725206,0.00043635923,0.0002822874,0.0002651234],"domain_scores_gemma":[0.9951222,0.003938694,0.000050066254,0.00025568696,0.000038984275,0.00059437176],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005992093,0.00016288483,0.00059164286,0.00010324896,0.00008898758,0.000014650913,0.00004388631,0.00010538713,0.0000928631],"category_scores_gemma":[0.0042307107,0.00010979616,0.00011808453,0.00020721708,0.00024411702,0.00003928816,0.00003619729,0.00040584258,0.00005722889],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002645688,0.0001475612,0.049531627,0.0026007164,0.00015918213,0.0004947867,0.0042862655,5.015911e-7,0.0000047847075,0.009443962,0.11965389,0.81341213],"study_design_scores_gemma":[0.008587253,0.019601999,0.13894437,0.068184294,0.00092204585,0.0007253259,0.002115865,0.00045765954,0.000014742918,0.0209528,0.7387186,0.00077498285],"about_ca_topic_score_codex":0.00017487966,"about_ca_topic_score_gemma":0.000008005352,"teacher_disagreement_score":0.81263715,"about_ca_system_score_codex":0.00005590919,"about_ca_system_score_gemma":0.00029955848,"threshold_uncertainty_score":0.5064861},"labels":[],"label_agreement":null},{"id":"W4392947622","doi":"10.1136/bmjqs-2023-016695","title":"Crowdsourcing a diagnosis? Exploring the accuracy of the size and type of group diagnosis: an experimental study","year":2024,"lang":"en","type":"article","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"Royal College of Physicians and Surgeons of Canada","keywords":"Medical diagnosis; Medicine; Differential diagnosis; Nominal group technique; Crowdsourcing; Group (periodic table); Diagnostic accuracy; Medical physics; Radiology; Artificial intelligence; Pathology; Computer science","score_opus":0.16479901184735743,"score_gpt":0.4631113570271014,"score_spread":0.2983123451797439,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4392947622","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99333453,0.0026206893,0.000008408475,0.002196211,0.00052533956,0.0011101816,0.000020816064,0.000051760424,0.00013207985],"genre_scores_gemma":[0.998513,0.00049818214,0.000070140384,0.00028423883,0.00023228537,0.00034426127,0.000004136313,0.000023573582,0.00003022201],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9974472,0.00052257464,0.001003071,0.00034823304,0.00048120928,0.00019773324],"domain_scores_gemma":[0.95525473,0.04366554,0.0001873203,0.0006841107,0.000083756735,0.00012452336],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0027654998,0.00017013094,0.00046453188,0.000026839174,0.00012362786,0.000034043805,0.00019673495,0.000056756406,0.00009416226],"category_scores_gemma":[0.05539755,0.00009364016,0.00017556016,0.0003297267,0.00022939716,0.00013050581,0.0002259941,0.00027211744,0.000004112281],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006460752,0.0032954467,0.9528488,0.0002390664,0.00025633755,0.000027088317,0.021158788,0.000014172455,0.0012395908,0.0012642844,0.0005675661,0.018442817],"study_design_scores_gemma":[0.0012569693,0.0011056742,0.9682068,0.0028937457,0.00026256437,0.00001056923,0.021126242,0.0000407381,0.0043345005,0.00013203417,0.00048884033,0.00014128325],"about_ca_topic_score_codex":0.00091189484,"about_ca_topic_score_gemma":0.00006400756,"teacher_disagreement_score":0.052632052,"about_ca_system_score_codex":0.000043197662,"about_ca_system_score_gemma":0.00012215928,"threshold_uncertainty_score":0.95255923},"labels":[],"label_agreement":null},{"id":"W4392960329","doi":"10.1051/pmed/2024004","title":"Formation par concordance : la théorie des scripts et son application en enseignement revisitées","year":2024,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Centre Hospitalier Universitaire Sainte-Justine","funders":"","keywords":"Humanities; Philosophy","score_opus":0.033160548744561424,"score_gpt":0.36362473000696444,"score_spread":0.330464181262403,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4392960329","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.13630296,0.5029137,0.18744846,0.09507244,0.0056463927,0.0023371154,0.00026188634,0.0010838406,0.06893319],"genre_scores_gemma":[0.93339163,0.044846687,0.0037265844,0.002424725,0.0018473087,0.0002380744,0.00025517092,0.00006829477,0.013201532],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9971458,0.00039148337,0.0007537841,0.0005664811,0.0005560394,0.00058643374],"domain_scores_gemma":[0.9873813,0.01145862,0.00012301671,0.00047823173,0.00013387243,0.00042495498],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0020266338,0.00030824263,0.00047095123,0.00008949305,0.00015611605,0.0001949,0.00018824027,0.00035041623,0.0006017764],"category_scores_gemma":[0.017696166,0.0002776566,0.00021940797,0.00037942932,0.00047033062,0.00030014865,0.00011647744,0.00064039027,0.0015743882],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006975924,0.00046676293,0.002056301,0.0013638568,0.000102590435,0.00018287469,0.0023679952,0.00002326867,0.0010743642,0.023406897,0.07315666,0.89572865],"study_design_scores_gemma":[0.0018557917,0.0008226018,0.023419654,0.0318067,0.0008836555,0.00020968671,0.0011966672,0.01543989,0.0044638203,0.037787247,0.8814932,0.0006210634],"about_ca_topic_score_codex":0.00014664473,"about_ca_topic_score_gemma":0.000019020252,"teacher_disagreement_score":0.8951076,"about_ca_system_score_codex":0.00037050544,"about_ca_system_score_gemma":0.000472789,"threshold_uncertainty_score":0.9999676},"labels":[],"label_agreement":null},{"id":"W4393032519","doi":"10.2196/44805","title":"Decision Support for Managing Common Musculoskeletal Pain Disorders: Development of a Case-Based Reasoning Application","year":2024,"lang":"en","type":"article","venue":"JMIR Formative Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Norwegian Fund for Post-Graduate Training in Physiotherapy; Norges Forskningsråd","keywords":"Preprint; Decision support system; Computer science; Medicine; Psychology; Artificial intelligence; World Wide Web","score_opus":0.04416145428266958,"score_gpt":0.4619681906288346,"score_spread":0.417806736346165,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4393032519","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.53983194,0.00030153303,0.45441967,0.0009471066,0.000089884474,0.0020593896,0.000026022699,0.00010960776,0.0022148076],"genre_scores_gemma":[0.9796615,0.000020141468,0.01901821,0.00003444724,0.000052846375,0.000896434,0.00020509338,0.000026376525,0.000084905456],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9980084,0.00015496381,0.0005222156,0.00028184068,0.0006409648,0.0003916651],"domain_scores_gemma":[0.9801095,0.019131547,0.000061324325,0.0002701202,0.00025502054,0.00017244619],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00546072,0.00012460502,0.00027526836,0.00036955508,0.00020535392,0.000043836026,0.00010379888,0.00009668391,0.00004769859],"category_scores_gemma":[0.0062780045,0.00009883974,0.00015498533,0.0005672722,0.00013004271,0.00010212765,0.00007630734,0.00035820281,0.0000507829],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00027060832,0.00033219502,0.0008882003,0.0009798203,0.000055923407,0.000055361492,0.0021944947,0.000014965838,0.00008596999,0.0011848608,0.0031212217,0.99081635],"study_design_scores_gemma":[0.0091664605,0.010058818,0.025724556,0.03123885,0.0002024761,0.00031011505,0.011119185,0.78375757,0.0044282763,0.015123362,0.107913055,0.0009572748],"about_ca_topic_score_codex":0.000040753508,"about_ca_topic_score_gemma":0.000059271766,"teacher_disagreement_score":0.9898591,"about_ca_system_score_codex":0.0001840488,"about_ca_system_score_gemma":0.0004169688,"threshold_uncertainty_score":0.75158113},"labels":[],"label_agreement":null},{"id":"W4393133727","doi":"10.2196/54507","title":"Evidence-Based Learning Strategies in Medicine Using AI","year":2024,"lang":"en","type":"article","venue":"JMIR Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":27,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Preprint; Artificial intelligence; Computer science; Psychology; Cognitive science; World Wide Web","score_opus":0.07546521640034537,"score_gpt":0.47828446217453163,"score_spread":0.40281924577418626,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4393133727","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9019838,0.009909837,0.0018575928,0.08206434,0.0015903253,0.00039547507,2.9094318e-7,0.00019902652,0.0019993135],"genre_scores_gemma":[0.992527,0.00033043558,0.00041762416,0.0050319773,0.0011890965,0.000065898486,0.000033237306,0.000019078183,0.00038564438],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.998289,0.00009803117,0.0004112641,0.00030957776,0.00067717,0.00021498046],"domain_scores_gemma":[0.9942387,0.005070651,0.00004169193,0.00016503346,0.00010063561,0.00038331904],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0009461286,0.00012250262,0.000272341,0.00024147739,0.00003509546,0.000042228978,0.00007518648,0.00019762949,0.002391203],"category_scores_gemma":[0.0657507,0.0000910452,0.00006121251,0.00057611865,0.00017634276,0.00015892707,0.000015233193,0.0007526437,0.00007081114],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00036554624,0.0019691861,0.14543511,0.0015196224,0.000067045126,0.00052529504,0.004141385,0.0004235747,0.0011260281,0.010119049,0.055357978,0.7789502],"study_design_scores_gemma":[0.0037796155,0.0019389257,0.31635663,0.39975226,0.00039644894,0.0003529409,0.005439294,0.21403377,0.00019362873,0.0054975883,0.051575165,0.00068373256],"about_ca_topic_score_codex":0.00042667336,"about_ca_topic_score_gemma":0.000016085698,"teacher_disagreement_score":0.77826643,"about_ca_system_score_codex":0.00016888884,"about_ca_system_score_gemma":0.0064039766,"threshold_uncertainty_score":0.9992288},"labels":[],"label_agreement":null},{"id":"W4393201478","doi":"10.3138/jvme-2023-0166","title":"Evaluation and Comparison of Tolerance of Ambiguity in Veterinary Pathology Professionals and Trainees","year":2024,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"University of Saskatchewan","funders":"","keywords":"Ambiguity; Medicine; Health professionals; Psychological intervention; Anatomical pathology; Scale (ratio); Psychology; Pathology; Nursing; Health care","score_opus":0.1368143229578685,"score_gpt":0.5218540976771762,"score_spread":0.38503977471930767,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4393201478","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98214227,0.013635484,0.00003866557,0.0033212537,0.00065827195,0.00013323156,0.0000016885012,0.0000028858062,0.00006622758],"genre_scores_gemma":[0.9982005,0.00078355917,0.00070373266,0.00013240616,0.00015199099,0.0000066623256,0.000004773171,0.000005268377,0.000011148478],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9983319,0.00023092468,0.00071333145,0.00010863599,0.0005414329,0.00007375115],"domain_scores_gemma":[0.9972875,0.0020317656,0.00022019712,0.00007224303,0.00024034783,0.00014794407],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002477458,0.00006705723,0.0003823083,0.0001562737,0.0000102108725,0.0000029394257,0.00004182394,0.00011480954,0.000098919736],"category_scores_gemma":[0.021949658,0.00005034988,0.000044866003,0.0001191173,0.00014730573,0.00006526887,0.000025472968,0.00026738233,2.8968805e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00091937353,0.0027534387,0.03840745,0.0010041341,0.000056825567,0.00018643486,0.0049554664,0.0000056511863,0.011709155,0.00009439121,0.0014087624,0.9384989],"study_design_scores_gemma":[0.0019465653,0.007134755,0.9581377,0.020062303,0.00019747268,0.0034983736,0.0015033635,0.0056379926,0.00027904165,0.0009768524,0.00054411934,0.00008148118],"about_ca_topic_score_codex":0.000010339082,"about_ca_topic_score_gemma":0.0000010074735,"teacher_disagreement_score":0.93841743,"about_ca_system_score_codex":0.000033877866,"about_ca_system_score_gemma":0.0011405133,"threshold_uncertainty_score":0.98628885},"labels":[],"label_agreement":null},{"id":"W4393316449","doi":"10.1111/jep.13986","title":"ReThinking clinical reasoning: A paradigm shift","year":2024,"lang":"en","type":"article","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University; McMaster University","funders":"","keywords":"Paradigm shift; Equity (law); Health care; Medical education; Health equity; Engineering ethics; Psychology; Public relations; Medicine; Sociology; Political science; Epistemology; Law","score_opus":0.17356922555481635,"score_gpt":0.5474817057273106,"score_spread":0.37391248017249423,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4393316449","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.75215054,0.01646872,0.005919277,0.18503334,0.015890392,0.0011122702,0.000004264305,0.00011995693,0.023301253],"genre_scores_gemma":[0.974046,0.0028707653,0.014371906,0.004907999,0.0036460843,0.000007698515,0.000006563241,0.000034309855,0.000108676795],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9894315,0.0029239445,0.0049898312,0.00046621522,0.0018998223,0.00028870942],"domain_scores_gemma":[0.6085997,0.3857148,0.0026032496,0.00078368775,0.0013976806,0.00090087956],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.086036645,0.00018340303,0.00092458865,0.00025862033,0.00004855592,0.00014161265,0.00016416622,0.0004975318,0.00046897327],"category_scores_gemma":[0.7778668,0.00014106542,0.0007688786,0.00059404614,0.00019270471,0.0006333248,0.00005654834,0.0034568799,0.00018153149],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008388573,0.0057626073,0.28505242,0.00018336352,0.0017309412,0.008750638,0.0022038594,0.00042687132,0.0000057973143,0.02796814,0.07050996,0.58901685],"study_design_scores_gemma":[0.013319941,0.0067546065,0.6077011,0.020866547,0.0057241214,0.0042855386,0.0003495076,0.042868145,0.0000067902424,0.15811118,0.13952439,0.0004881338],"about_ca_topic_score_codex":0.000019427172,"about_ca_topic_score_gemma":0.0000075624057,"teacher_disagreement_score":0.6918301,"about_ca_system_score_codex":0.00019565564,"about_ca_system_score_gemma":0.0022695048,"threshold_uncertainty_score":0.9988422},"labels":[],"label_agreement":null},{"id":"W4393425754","doi":"10.2196/55357","title":"Identifying Interventions to Improve Diagnostic Safety in Emergency Departments: Protocol for a Participatory Design Study","year":2024,"lang":"en","type":"article","venue":"JMIR Research Protocols","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Agency for Healthcare Research and Quality","keywords":"Preprint; Protocol (science); Psychological intervention; Medicine; Citizen journalism; Participatory design; Medical emergency; Computer science; Medical education; Nursing; Engineering; Alternative medicine; World Wide Web; Operations management","score_opus":0.6091992603387657,"score_gpt":0.6790283495437524,"score_spread":0.06982908920498665,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4393425754","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0014096213,0.0000069157095,0.0056799445,0.00063641445,0.000039272563,0.9918543,0.00003095521,0.00015659975,0.00018596176],"genre_scores_gemma":[0.010580637,6.2195437e-7,0.0011209698,0.000038203132,0.00024185507,0.9865159,0.000013549165,0.00006922275,0.0014190748],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99519116,0.00081216975,0.0012484341,0.00087993656,0.000894642,0.00097366545],"domain_scores_gemma":[0.9847781,0.013524143,0.00006695786,0.00064683973,0.00041499513,0.00056897436],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.006741391,0.00024063165,0.00045290482,0.00056207104,0.00018781441,0.00018939313,0.00029599117,0.00012569648,0.0006188414],"category_scores_gemma":[0.06618923,0.00019758969,0.00026486302,0.0011273965,0.00007552228,0.00016371044,0.00028465554,0.00071231544,0.00039313332],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.027489005,0.09483488,0.2769638,0.042352416,0.00088949717,0.0036607175,0.010489319,0.00023239154,0.0030195494,0.0025410498,0.3272648,0.21026257],"study_design_scores_gemma":[0.05518575,0.106136866,0.3346323,0.27961951,0.00020370034,0.000037010526,0.0031768181,0.0034601546,0.003409388,0.0148732085,0.19738936,0.0018759292],"about_ca_topic_score_codex":0.000049180315,"about_ca_topic_score_gemma":0.0001178673,"teacher_disagreement_score":0.23726709,"about_ca_system_score_codex":0.00037717228,"about_ca_system_score_gemma":0.00055745675,"threshold_uncertainty_score":0.9416767},"labels":[],"label_agreement":null},{"id":"W4394008818","doi":"10.1016/j.jaad.2024.01.091","title":"Cognitive bias in the patient encounter: Part I. Background and significance","year":2024,"lang":"en","type":"review","venue":"Journal of the American Academy of Dermatology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Cognitive bias; Cognition; Heuristics; Medicine; Affect (linguistics); Clinical judgment; Confirmation bias; Cognitive psychology; Dual process theory (moral psychology); Process (computing); Clinical psychology; Psychology; Psychiatry; Social psychology; Medical physics; Computer science","score_opus":0.12304003657404039,"score_gpt":0.43827730684533717,"score_spread":0.3152372702712968,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4394008818","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.017408969,0.9727069,0.0000023852601,0.0090178745,0.00024989643,0.00040907247,0.00003260072,0.000003635735,0.00016867282],"genre_scores_gemma":[0.037346296,0.959518,0.000037616697,0.0028673941,0.00016768272,0.000015403324,0.000001912879,0.000022089986,0.000023585691],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9968619,0.0008078095,0.0014984421,0.00020703713,0.00039204024,0.00023278414],"domain_scores_gemma":[0.98044956,0.01677813,0.0024668328,0.0001409303,0.000059379043,0.00010518953],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005654527,0.00026315136,0.0024679943,0.00020021401,0.000027472062,0.00001544432,0.00035576688,0.00018671616,0.0000099801855],"category_scores_gemma":[0.0074509676,0.00011607634,0.000616807,0.00052029605,0.0010612422,0.00002856578,0.00011031174,0.0018968713,0.00001243006],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001905112,0.00034808737,0.0019945966,0.0037971253,0.0009465429,0.00076066516,0.0005352648,4.1067673e-7,6.6505635e-7,0.00020495105,0.026298206,0.96492296],"study_design_scores_gemma":[0.0005989939,0.0005306477,0.0012265955,0.09831135,0.0053577223,0.031940825,0.0009735349,0.0000044030626,0.000004149716,0.0003710508,0.8604857,0.00019506479],"about_ca_topic_score_codex":0.0000136024955,"about_ca_topic_score_gemma":0.0000011418336,"teacher_disagreement_score":0.96472794,"about_ca_system_score_codex":0.00004973643,"about_ca_system_score_gemma":0.0003348288,"threshold_uncertainty_score":0.8920042},"labels":[],"label_agreement":null},{"id":"W4394015458","doi":"10.1016/j.jaad.2024.02.061","title":"Cognitive bias in the patient encounter: Part II. Debiasing using an adaptive toolbox","year":2024,"lang":"en","type":"review","venue":"Journal of the American Academy of Dermatology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Debiasing; Cognitive bias; Cognition; Toolbox; Confirmation bias; Cognitive psychology; Heuristics; Metacognition; Psychology; Situational ethics; Computer science; Social psychology","score_opus":0.1764372603810211,"score_gpt":0.45214755548332863,"score_spread":0.27571029510230755,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4394015458","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.07651727,0.91903746,0.000013931484,0.003330345,0.00041417044,0.00050373917,0.000041117994,0.000007388192,0.00013458473],"genre_scores_gemma":[0.09746248,0.8987845,0.00026632642,0.0030985076,0.00032272725,0.000009045378,0.0000038106214,0.000043769942,0.000008838295],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99558914,0.0013866605,0.0018982854,0.00025671092,0.0005362986,0.0003329202],"domain_scores_gemma":[0.98505425,0.011005282,0.003485856,0.00019779778,0.00011235468,0.00014442977],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.00071986753,0.00034824634,0.002913866,0.00034152385,0.00007276659,0.00001715552,0.00051242183,0.00025937267,0.000010145255],"category_scores_gemma":[0.011131058,0.00016455077,0.0009898564,0.0007446487,0.0008241041,0.0000690351,0.00014839043,0.0023900054,0.000006514335],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00041538078,0.00087998837,0.0008667497,0.0014474402,0.0013136269,0.0011471482,0.0018790792,0.000011388566,0.0000015926264,0.00013554582,0.004242525,0.9876595],"study_design_scores_gemma":[0.0016785209,0.003492941,0.00078999205,0.37538466,0.022822397,0.12831593,0.005614298,0.0002663776,0.000031496,0.000749295,0.46011636,0.00073771464],"about_ca_topic_score_codex":0.00003682291,"about_ca_topic_score_gemma":0.0000023456544,"teacher_disagreement_score":0.98692185,"about_ca_system_score_codex":0.00013088206,"about_ca_system_score_gemma":0.00064066524,"threshold_uncertainty_score":0.99991155},"labels":[],"label_agreement":null},{"id":"W4394031394","doi":"10.5281/zenodo.7031181","title":"Data set of Sepsis awareness at the university hospital level: a survey-based cross-sectional study","year":2022,"lang":"en","type":"dataset","venue":"Zenodo (CERN European Organization for Nuclear Research)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Regina","funders":"","keywords":"Cross-sectional study; Set (abstract data type); Data set; Medicine; Computer science; Artificial intelligence; Pathology; Programming language","score_opus":0.1804356878912624,"score_gpt":0.3735442642698541,"score_spread":0.1931085763785917,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4394031394","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.043203045,0.000028954917,0.000036006375,0.00022641651,0.00022927341,0.00070929516,0.95516497,0.00011653018,0.0002855084],"genre_scores_gemma":[0.026513925,0.000055830373,0.000008033039,0.00007449124,0.000101546044,3.4305277e-8,0.97199756,0.00042239332,0.00082618534],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9964322,0.001187504,0.00039600828,0.00074654276,0.0009670669,0.0002706958],"domain_scores_gemma":[0.9948876,0.0019277434,0.00030064114,0.0019412584,0.0007536412,0.00018909859],"candidate_categories":["metaresearch","sts","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.002828816,0.00019889545,0.00034797203,0.00019156562,0.0021925576,0.00019524834,0.0021473183,0.00012501627,0.120167255],"category_scores_gemma":[0.031075582,0.00017566966,0.00009856319,0.0007186361,0.0004050818,0.00007353699,0.0056495457,0.0006558878,0.0012597558],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004906799,0.0008992107,0.0062177298,0.00006367591,0.00018650979,0.00008576597,0.00006454841,0.0000600675,0.0000015835328,0.0000017073534,0.9913248,0.00060371537],"study_design_scores_gemma":[0.0012931279,0.00077304005,0.22335465,0.000053932214,0.00011390653,0.000023685532,0.000149615,0.000036676076,0.0000017157333,8.2913385e-7,0.77406496,0.00013385923],"about_ca_topic_score_codex":0.001508499,"about_ca_topic_score_gemma":0.00003776531,"teacher_disagreement_score":0.21725984,"about_ca_system_score_codex":0.0003364105,"about_ca_system_score_gemma":0.000080621714,"threshold_uncertainty_score":0.99951786},"labels":[],"label_agreement":null},{"id":"W4394259035","doi":"10.6084/m9.figshare.5123776","title":"Supplementary Material for: Alternate-Form Reliability of the Montreal Cognitive Assessment Screening Test in a Clinical Setting","year":2012,"lang":"en","type":"dataset","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Reliability (semiconductor); Test (biology); Montreal Cognitive Assessment; Reliability engineering; Cognition; Psychology; Cognitive impairment; Computer science; Engineering; Psychiatry; Geology","score_opus":0.05882305298346468,"score_gpt":0.40961095730496866,"score_spread":0.350787904321504,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4394259035","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004106086,0.000024010491,3.7570726e-7,0.00020180462,0.00035385208,0.0016084908,0.99363244,0.000011898909,0.000061035775],"genre_scores_gemma":[0.012142014,0.000016794482,0.00025789847,0.00049722486,0.0010570021,0.0004573901,0.9855242,0.000027562022,0.00001993905],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99719065,0.00015050938,0.0012950762,0.0004767059,0.0004425168,0.0004445603],"domain_scores_gemma":[0.9530667,0.044801697,0.00086962664,0.0007363791,0.00028548046,0.00024008195],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0009311621,0.00029637286,0.0008229458,0.00005858519,0.00005862602,0.000023643988,0.00030715627,0.00042051548,0.07863105],"category_scores_gemma":[0.18188818,0.00020639284,0.00041961123,0.00010303577,0.000042352323,0.000040282488,0.00045849098,0.000807696,0.00003593016],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00022726388,0.0006741298,0.042037666,0.00046248364,0.000057607074,0.0000176852,0.000013785851,5.668624e-7,2.5132312e-7,6.8058874e-8,0.95253116,0.00397732],"study_design_scores_gemma":[0.006008376,0.00088912656,0.5828046,0.07824998,0.0007400214,0.000024587098,0.000099840494,0.00010036913,0.00006511438,0.000028186589,0.33060154,0.00038822167],"about_ca_topic_score_codex":0.0013823083,"about_ca_topic_score_gemma":0.0008005045,"teacher_disagreement_score":0.62192965,"about_ca_system_score_codex":0.00010043358,"about_ca_system_score_gemma":0.00038854388,"threshold_uncertainty_score":0.9222112},"labels":[],"label_agreement":null},{"id":"W4394364008","doi":"10.6084/m9.figshare.16531192.v2","title":"The Student-Generated Reasoning Tool (SGRT): Linking medical knowledge and clinical reasoning in preclinical education","year":2021,"lang":"en","type":"dataset","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Model-based reasoning; Computer science; Medicine; Artificial intelligence; Knowledge representation and reasoning","score_opus":0.06643700012178563,"score_gpt":0.4496779015046737,"score_spread":0.383240901382888,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4394364008","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.002353979,0.018901682,1.9502816e-7,0.0006114347,0.0013868597,0.00078332494,0.9754914,0.00007162173,0.0003995194],"genre_scores_gemma":[0.00034188238,0.0051537664,0.00012914841,0.0012409379,0.0034417133,0.00036743053,0.98872674,0.00006080951,0.0005375477],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9941472,0.0008958816,0.002044976,0.0012459854,0.0010072966,0.0006586946],"domain_scores_gemma":[0.9087912,0.08608173,0.00095394166,0.0020013938,0.00092711434,0.001244652],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0025171032,0.0005224378,0.0012910995,0.00013620457,0.0002791884,0.00029542015,0.000605098,0.0018415324,0.016764808],"category_scores_gemma":[0.598889,0.00039398202,0.0003801112,0.00052397355,0.000105181854,0.0000675886,0.00089907955,0.0037234593,0.0007216405],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007633037,0.0010679297,0.005118302,0.00020830025,0.00014576629,0.00032096455,0.000027084865,4.2910762e-7,6.703889e-8,0.000004423833,0.9371718,0.05585856],"study_design_scores_gemma":[0.0016197921,0.00023299818,0.046098437,0.1317,0.00025502592,0.00017080514,0.000073648975,0.00025083462,0.0000012395366,0.0000070073174,0.81920683,0.00038337556],"about_ca_topic_score_codex":0.00006912159,"about_ca_topic_score_gemma":0.00077383564,"teacher_disagreement_score":0.5963719,"about_ca_system_score_codex":0.00017391627,"about_ca_system_score_gemma":0.005858131,"threshold_uncertainty_score":0.9998512},"labels":[],"label_agreement":null},{"id":"W4394484798","doi":"10.6084/m9.figshare.16531192","title":"The Student-Generated Reasoning Tool (SGRT): Linking medical knowledge and clinical reasoning in preclinical education","year":2022,"lang":"en","type":"dataset","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Psychology; Medical education; Computer science; Medicine","score_opus":0.06553516645562873,"score_gpt":0.44733857227300095,"score_spread":0.3818034058173722,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4394484798","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.001963243,0.011084536,9.4937924e-8,0.00056754635,0.0014983048,0.0010219866,0.98335326,0.0000874611,0.0004235774],"genre_scores_gemma":[0.00030642728,0.0034971735,0.000058402045,0.0012787641,0.002290141,0.00075380795,0.99117076,0.00006662952,0.0005778944],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99387205,0.0010357255,0.0020166372,0.0012063694,0.0012041997,0.00066500314],"domain_scores_gemma":[0.89852977,0.09653613,0.0011509996,0.0020355622,0.00053534226,0.0012121771],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0032654817,0.00050985534,0.0011705713,0.00017664919,0.00044716132,0.00021179029,0.0007889294,0.0012758198,0.1587707],"category_scores_gemma":[0.51733506,0.00039359686,0.0003595888,0.000539456,0.000101054815,0.000066241504,0.0012832996,0.0045512165,0.00046762612],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012936974,0.0011248416,0.004927845,0.00015524765,0.00012904576,0.00020944508,0.00003850441,9.805833e-7,3.0749515e-8,0.000006439092,0.9416276,0.051650602],"study_design_scores_gemma":[0.0017456763,0.00043146368,0.037446715,0.048136726,0.00023942717,0.0001479226,0.00008837339,0.00024090873,3.4520536e-7,0.000009953966,0.9111169,0.00039560164],"about_ca_topic_score_codex":0.000098536424,"about_ca_topic_score_gemma":0.0005122161,"teacher_disagreement_score":0.51406956,"about_ca_system_score_codex":0.0002543889,"about_ca_system_score_gemma":0.0047018216,"threshold_uncertainty_score":0.9998516},"labels":[],"label_agreement":null},{"id":"W4394844254","doi":"10.1111/medu.15380","title":"The influence of viewing time on visual diagnostic accuracy: Less is more","year":2024,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Impact; McMaster University","funders":"Royal College of Physicians and Surgeons of Canada; McMaster University","keywords":"Psychology; Optometry; Computer science; Audiology; Medicine","score_opus":0.01644225987006724,"score_gpt":0.40511310564160724,"score_spread":0.38867084577154,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4394844254","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9416001,0.0019722919,0.000018790824,0.054984808,0.00060475466,0.00019003838,0.0000023240975,0.000048388232,0.0005785208],"genre_scores_gemma":[0.9881948,0.0008517245,0.000026135463,0.00933926,0.00043971871,0.000047451962,0.00002448089,0.000013584558,0.0010628251],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9985848,0.00004489537,0.00034754697,0.00021075875,0.00064899813,0.00016299865],"domain_scores_gemma":[0.9524311,0.046847116,0.000066908295,0.0002780263,0.0001350922,0.00024177233],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0006156077,0.0000955305,0.00018383456,0.000048453796,0.00006251278,0.000028275217,0.000117498625,0.000111108304,0.00044096806],"category_scores_gemma":[0.3093381,0.00005738353,0.00008193065,0.0002504612,0.00019109722,0.000038428032,0.000030255978,0.00031402818,0.0003065235],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000053543394,0.00085540337,0.0067627593,0.00015925527,0.00006155289,0.000028641765,0.0015011991,0.000012527097,0.0000979517,0.0006140712,0.14572534,0.8441278],"study_design_scores_gemma":[0.0015813597,0.0014826406,0.6835955,0.077933505,0.00067874586,0.00017804881,0.0010617579,0.0110510355,0.0020866236,0.0022257194,0.21759588,0.0005292166],"about_ca_topic_score_codex":0.000068578534,"about_ca_topic_score_gemma":0.0000011334467,"teacher_disagreement_score":0.84359854,"about_ca_system_score_codex":0.00004943836,"about_ca_system_score_gemma":0.0015812572,"threshold_uncertainty_score":0.6964797},"labels":[],"label_agreement":null},{"id":"W4396582209","doi":"10.1186/s12909-024-05399-x","title":"Use of the International IFOMPT Cervical Framework to inform clinical reasoning in postgraduate level physiotherapy students: a qualitative study using think aloud methodology","year":2024,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Medical education; Qualitative research; Think aloud protocol; Psychology; Medicine; Physical therapy; Computer science; Sociology","score_opus":0.3674010047126138,"score_gpt":0.6113837766209341,"score_spread":0.24398277190832035,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4396582209","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9761252,0.00012966518,0.011126218,0.008313798,0.0034355575,0.0007817572,0.00000887771,0.000036359073,0.000042590986],"genre_scores_gemma":[0.9379194,0.00006266163,0.054414894,0.006506495,0.00073437183,0.000035259374,0.000021390686,0.000026970441,0.00027855116],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9955273,0.0012178314,0.0011517347,0.00044686368,0.0014277834,0.0002284458],"domain_scores_gemma":[0.9402129,0.05809568,0.00023893859,0.0005806232,0.0003937732,0.0004780592],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005434402,0.00017195557,0.00053163694,0.00018291538,0.000048227343,0.00006144734,0.0004101202,0.00026955953,0.0002597916],"category_scores_gemma":[0.3102635,0.00011452029,0.00020922454,0.0007621454,0.00017165505,0.00010019718,0.00020924979,0.0008321142,0.0000258657],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013690002,0.010466505,0.70961714,0.00017423593,0.00058807334,0.000025056368,0.1920797,0.000073753,0.00006177549,0.011010481,0.0038024588,0.07073181],"study_design_scores_gemma":[0.0015963111,0.00070617546,0.9389564,0.011855312,0.00018170822,0.000034003435,0.03697631,0.005073117,0.000027956345,0.0038866268,0.0004954912,0.00021059466],"about_ca_topic_score_codex":0.0013090763,"about_ca_topic_score_gemma":0.00032628948,"teacher_disagreement_score":0.3048291,"about_ca_system_score_codex":0.00019862833,"about_ca_system_score_gemma":0.003156523,"threshold_uncertainty_score":0.69554645},"labels":[],"label_agreement":null},{"id":"W4396659779","doi":"10.7202/1110632ar","title":"La formation à l’offre active : design pédagogique, défis et avenues pour le futur","year":2024,"lang":"fr","type":"article","venue":"Minorités linguistiques et société","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"Université de Moncton; Montfort Hospital; University of Ottawa","funders":"","keywords":"Humanities; Philosophy","score_opus":0.06081204719384903,"score_gpt":0.4208744109954627,"score_spread":0.36006236380161366,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4396659779","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.13333781,0.39022526,0.09791422,0.064795814,0.025723029,0.0026533906,0.0011279177,0.002326158,0.28189638],"genre_scores_gemma":[0.8899123,0.013066438,0.018019341,0.0033929476,0.009547028,0.00015766008,0.00042541235,0.00017839129,0.06530045],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9956701,0.001659165,0.00079396757,0.0006831117,0.00047043734,0.00072325196],"domain_scores_gemma":[0.9624463,0.03602457,0.00020882618,0.0004388363,0.0004586594,0.00042284536],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0030079219,0.0005250602,0.0007296636,0.00016980126,0.0001610839,0.00032039575,0.00023224196,0.0009570896,0.00044618567],"category_scores_gemma":[0.07777785,0.0005279339,0.00047417148,0.00029987443,0.00027768474,0.00029995255,0.00012870072,0.0017582193,0.00055721786],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00028635887,0.00094869564,0.00022817643,0.0011348238,0.000497652,0.0019946431,0.06759293,0.00008772862,0.00016954284,0.06324756,0.08136242,0.7824495],"study_design_scores_gemma":[0.004716268,0.0019061817,0.009155143,0.047008995,0.002225323,0.00076027337,0.038162332,0.015650528,0.01376592,0.08257361,0.7820477,0.0020277344],"about_ca_topic_score_codex":0.0024313116,"about_ca_topic_score_gemma":0.00024816004,"teacher_disagreement_score":0.78042173,"about_ca_system_score_codex":0.0005345496,"about_ca_system_score_gemma":0.0021219747,"threshold_uncertainty_score":0.99971724},"labels":[],"label_agreement":null},{"id":"W4396742649","doi":"10.1016/j.jcjq.2024.05.004","title":"The Quest for Diagnostic Excellence in the Emergency Department","year":2024,"lang":"en","type":"letter","venue":"The Joint Commission Journal on Quality and Patient Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Excellence; Emergency department; Medical emergency; Medicine; Political science; Nursing; Law","score_opus":0.09695379264845647,"score_gpt":0.38662686414886194,"score_spread":0.28967307150040544,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4396742649","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0060784905,0.006240315,0.00009925172,0.9827256,0.0021239957,0.0011825903,0.000088780616,0.000017973884,0.0014429762],"genre_scores_gemma":[0.11205161,0.050373238,0.00009443362,0.82731724,0.005391572,0.0003106051,0.00023495533,0.0000951888,0.004131149],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9953669,0.0015692016,0.0014753232,0.00031271626,0.00086764735,0.00040822814],"domain_scores_gemma":[0.9279403,0.07040398,0.0006015612,0.0007174827,0.0001465521,0.000190091],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.00506814,0.0003266394,0.0005308704,0.0000588604,0.00086668227,0.0000973915,0.00032121048,0.00028692745,0.0001056026],"category_scores_gemma":[0.036132537,0.00012219549,0.00040502712,0.00014086634,0.00017314596,0.000023319279,0.000100919686,0.0033294284,0.000037885962],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00049841194,0.00020063325,0.0003248583,0.00012714308,0.000068741305,0.00017256741,0.0009412184,0.000009894455,6.3973425e-7,0.0007887263,0.97235423,0.02451295],"study_design_scores_gemma":[0.00068576855,0.0013011994,0.010315151,0.00563393,0.00021345977,0.00019399193,0.00032361504,0.000039863728,0.0000045808515,0.009884845,0.97122,0.00018360022],"about_ca_topic_score_codex":0.000030762585,"about_ca_topic_score_gemma":0.000010683465,"teacher_disagreement_score":0.15540838,"about_ca_system_score_codex":0.00009843933,"about_ca_system_score_gemma":0.0001480807,"threshold_uncertainty_score":0.9989699},"labels":[],"label_agreement":null},{"id":"W4397005719","doi":"10.1051/pmed/2024011","title":"Didactique professionnelle, analyse du travail, analyse de l’activité des étudiants et des professionnels de santé : un large champ de recherche en éducation à investir","year":2024,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Centre for Interdisciplinary Research in Rehabilitation","funders":"","keywords":"Sociology; Humanities; Political science; Art","score_opus":0.20900058212287906,"score_gpt":0.47653901097542223,"score_spread":0.2675384288525432,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4397005719","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.90661156,0.050383322,0.019193722,0.018705357,0.0012402524,0.0007611592,0.00010653497,0.00039755573,0.002600511],"genre_scores_gemma":[0.9294566,0.0400739,0.014554638,0.0027593218,0.001959137,0.0003061956,0.00011025751,0.00016009036,0.010619838],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.98636657,0.008970466,0.0010541489,0.0010770984,0.00058207376,0.0019496514],"domain_scores_gemma":[0.9431258,0.053810224,0.0002568048,0.0007780178,0.00027151857,0.0017576523],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.01886544,0.00065601483,0.0009948955,0.00039539993,0.00049122766,0.00019022191,0.0004533669,0.0018070304,0.0017866299],"category_scores_gemma":[0.2240908,0.0005574272,0.00050384394,0.0017519881,0.0007872071,0.00050512573,0.00026496753,0.003532088,0.00037689696],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00040115177,0.0074338177,0.55642503,0.0019952867,0.0012572969,0.0017734842,0.10121452,0.00009072651,0.010977959,0.0006061136,0.0437673,0.2740573],"study_design_scores_gemma":[0.004620804,0.0012134878,0.65324324,0.059039474,0.0036437912,0.0007788395,0.15634419,0.038420893,0.0385898,0.037474122,0.0048284344,0.0018029071],"about_ca_topic_score_codex":0.0008299045,"about_ca_topic_score_gemma":0.00068031176,"teacher_disagreement_score":0.2722544,"about_ca_system_score_codex":0.003558415,"about_ca_system_score_gemma":0.009073411,"threshold_uncertainty_score":0.99968773},"labels":[],"label_agreement":null},{"id":"W4398144941","doi":"10.1101/2024.05.20.24307611","title":"Temporal trends and practice variation of paediatric diagnostic tests in primary care","year":2024,"lang":"en","type":"preprint","venue":"medRxiv","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hospital for Sick Children; University of Toronto","funders":"","keywords":"Variation (astronomy); Primary care; Diagnostic test; Medicine; Pediatrics; Family medicine; Physics","score_opus":0.02285852208135061,"score_gpt":0.33796981550246663,"score_spread":0.31511129342111605,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398144941","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97313607,0.016187338,0.000013480672,0.0031769706,0.0008073333,0.000343087,0.00006749565,0.00007359625,0.0061946497],"genre_scores_gemma":[0.99531126,0.001848593,0.0016006335,0.00022546519,0.00036648117,0.00005079179,0.00039292904,0.00003904571,0.00016478433],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9980895,0.00016762687,0.00063454034,0.00054164574,0.00037614338,0.00019052374],"domain_scores_gemma":[0.9624066,0.036518637,0.0002806079,0.0004208419,0.00021331156,0.0001600463],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0008118672,0.00023171325,0.0006349429,0.00043507112,0.000014405365,0.000026864689,0.00009209119,0.00038397725,0.000028597475],"category_scores_gemma":[0.17077994,0.00020078839,0.00011338231,0.0005510174,0.0000703783,0.000029959045,0.00044781526,0.0008948113,0.000017275464],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000111670764,0.00037277112,0.94736403,0.001954899,0.00007901739,0.00056484743,0.001471632,0.000039770483,0.00003259315,0.00006734859,0.00039611352,0.047545284],"study_design_scores_gemma":[0.0008994463,0.00020951725,0.99185973,0.005165192,0.0007853114,0.000051475614,0.00007883255,0.000054683765,0.000011985088,0.00046923305,0.00024716146,0.0001674227],"about_ca_topic_score_codex":0.00043104682,"about_ca_topic_score_gemma":0.000029447465,"teacher_disagreement_score":0.16996807,"about_ca_system_score_codex":0.00013875676,"about_ca_system_score_gemma":0.0004966442,"threshold_uncertainty_score":0.83620495},"labels":[],"label_agreement":null},{"id":"W4398166126","doi":"10.1152/physiol.2024.39.s1.2307","title":"Physiological Knowledge Loss in Fourth-Year Nursing Students","year":2024,"lang":"en","type":"article","venue":"Physiology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"MacEwan University","funders":"","keywords":"Nursing; Medicine; Psychology; Biology; Physiology","score_opus":0.04672305002397088,"score_gpt":0.4274549887547786,"score_spread":0.38073193873080774,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398166126","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99410456,0.0011326933,0.00002288264,0.0007053266,0.0006699914,0.00015325747,0.000003285022,0.00011380205,0.0030941758],"genre_scores_gemma":[0.99831134,0.00021503886,0.00014430798,0.000277434,0.0005607118,0.000020720174,0.000028525634,0.00001641098,0.00042550176],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9988745,0.00008435862,0.00023643947,0.00039144693,0.00010672347,0.00030657108],"domain_scores_gemma":[0.99703526,0.002604152,0.00001897327,0.0001968011,0.00002915761,0.00011564791],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00016246596,0.00012807474,0.0003740685,0.00008100789,0.000022431543,0.0000119139295,0.00010952479,0.00014736979,0.00023164944],"category_scores_gemma":[0.0031825048,0.000093922325,0.000119960794,0.00022429401,0.00017994943,0.000023940771,0.00006123976,0.00035678066,0.0010005494],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0037461638,0.028589427,0.34071532,0.0011030134,0.0009534035,0.00393126,0.007250205,0.0002749781,0.19640519,0.050878868,0.12104193,0.24511024],"study_design_scores_gemma":[0.0012044603,0.0005415721,0.98445934,0.0030694867,0.000050261613,0.000034776276,0.0000630978,0.00066887075,0.00023599136,0.008576529,0.0009556733,0.00013992483],"about_ca_topic_score_codex":0.000011662408,"about_ca_topic_score_gemma":0.0000020506614,"teacher_disagreement_score":0.64374405,"about_ca_system_score_codex":0.000073216346,"about_ca_system_score_gemma":0.000074693686,"threshold_uncertainty_score":0.99977726},"labels":[],"label_agreement":null},{"id":"W4398400406","doi":"10.7910/dvn/ghnwka/kvhlcb","title":"Dataset 1. Raw data of the response of subjects.pdf","year":2020,"lang":"en","type":"dataset","venue":"Harvard Dataverse","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Raw data; Raw score; Psychology; Medical education; Audiology; Applied psychology; Computer science; Medicine","score_opus":0.05325184402993398,"score_gpt":0.3370809017011248,"score_spread":0.2838290576711908,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398400406","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00027360738,0.0000056329077,0.000004913431,0.00013299585,0.0005472711,0.00049678073,0.9984917,0.000012278846,0.000034803543],"genre_scores_gemma":[0.0001892764,0.0005688637,0.00021515702,0.0014998397,0.00021871256,0.000006336421,0.9971297,0.000025304034,0.00014684974],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9967915,0.00041166603,0.00093093875,0.0007541461,0.00083153317,0.0002802227],"domain_scores_gemma":[0.97974384,0.0107732145,0.0007429171,0.008320884,0.00011680577,0.00030233205],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0013642762,0.00033040895,0.0010340515,0.00010278753,0.00004849966,0.00001705265,0.002289093,0.00034282228,0.01411412],"category_scores_gemma":[0.16802044,0.00023139149,0.00019209653,0.00038568853,0.0005008669,0.00010675094,0.0030454844,0.00072851573,0.03624595],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0033833089,0.00036904795,0.00012987352,0.00033690193,0.00022871824,0.00019269827,0.000010761119,0.0000010603407,0.00011732254,0.0000026275354,0.99511266,0.00011500375],"study_design_scores_gemma":[0.0018946123,0.00029632886,0.0017485823,0.00251354,0.0014353037,0.000050632127,0.000041442396,0.00002164766,0.0003414634,0.0000042396714,0.9914594,0.00019278005],"about_ca_topic_score_codex":0.0007668221,"about_ca_topic_score_gemma":0.000070936854,"teacher_disagreement_score":0.16665617,"about_ca_system_score_codex":0.00003466713,"about_ca_system_score_gemma":0.0010689145,"threshold_uncertainty_score":0.98678714},"labels":[],"label_agreement":null},{"id":"W4399123121","doi":"10.2196/58157","title":"A Symptom-Checker for Adult Patients Visiting an Interdisciplinary Emergency Care Center and the Safety of Patient Self-Triage: Real-Life Prospective Evaluation","year":2024,"lang":"en","type":"article","venue":"Journal of Medical Internet Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Immunovaccine (Canada)","funders":"","keywords":"Triage; Center (category theory); Preprint; Trauma center; Medical emergency; Medicine; Psychology; Computer science; World Wide Web; Pathology","score_opus":0.038009561518310066,"score_gpt":0.4803377487015386,"score_spread":0.4423281871832285,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4399123121","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.990836,0.0015402408,0.0002890764,0.0048091155,0.0006822578,0.001074718,0.00001823032,0.0000120935765,0.00073823286],"genre_scores_gemma":[0.9978052,0.0011582552,0.00018521743,0.00009253877,0.0006137347,0.00005377637,0.000029379511,0.000021192614,0.000040725496],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99469745,0.00065922795,0.001145089,0.00027123457,0.002946071,0.00028094856],"domain_scores_gemma":[0.98795456,0.006139663,0.00026535097,0.00021922126,0.0048508886,0.0005703416],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0075392406,0.00012493799,0.00043762755,0.00021077413,0.00007648166,0.00004333535,0.00023684275,0.00017372318,0.00033269753],"category_scores_gemma":[0.08287747,0.000067709865,0.00024440596,0.00022758913,0.0002545229,0.00010710261,0.00034083345,0.0010153016,0.000003645966],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.03296406,0.004737337,0.34538054,0.0024249356,0.0025834374,0.000279541,0.14163165,0.000008213578,0.000025549016,0.011135735,0.05319959,0.40562943],"study_design_scores_gemma":[0.0903882,0.043558784,0.44233608,0.08053559,0.002374608,0.0005576901,0.07511196,0.2513743,0.00071057456,0.0101425005,0.0021914484,0.00071826327],"about_ca_topic_score_codex":0.00007157298,"about_ca_topic_score_gemma":0.000022107055,"teacher_disagreement_score":0.40491116,"about_ca_system_score_codex":0.00027085454,"about_ca_system_score_gemma":0.00063596526,"threshold_uncertainty_score":0.92484784},"labels":[],"label_agreement":null},{"id":"W4399238455","doi":"10.1111/jep.14023","title":"The influence of context on diagnostic reasoning: A narrative synthesis of experimental findings","year":2024,"lang":"en","type":"review","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"McMaster University","keywords":"Narrative; Context (archaeology); Psychology; Medicine; History; Linguistics; Philosophy","score_opus":0.1384235414192977,"score_gpt":0.5515688638123217,"score_spread":0.41314532239302404,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4399238455","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.017893454,0.9779721,0.0000020323325,0.0013716702,0.00096826215,0.0009979825,0.0000110947485,0.0000048129737,0.0007786254],"genre_scores_gemma":[0.11121605,0.88779485,0.00032593895,0.0001965492,0.00031545383,0.00006951267,0.0000036425956,0.000032900105,0.000045069202],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.98871917,0.0031311244,0.005752301,0.00034982048,0.0018472456,0.00020034937],"domain_scores_gemma":[0.38526022,0.60619694,0.005900017,0.0004966784,0.0019325186,0.00021364828],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.027441587,0.00030688036,0.0026527988,0.000268265,0.00005372019,0.000036304875,0.0002851223,0.00045341675,0.00007310237],"category_scores_gemma":[0.932072,0.00017452879,0.0011847566,0.00054261705,0.00040527072,0.0001782959,0.00007967613,0.0021249128,0.000035307476],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001321249,0.002862417,0.0003002189,0.0016250862,0.0015082197,0.00024485032,0.0009496337,0.00010157173,7.691907e-7,0.0006093587,0.005862309,0.9846143],"study_design_scores_gemma":[0.003465483,0.008009363,0.0025742191,0.60327697,0.025949158,0.00096206355,0.0043383474,0.00022082518,0.000028039718,0.00037296276,0.3503887,0.00041388161],"about_ca_topic_score_codex":0.000021536485,"about_ca_topic_score_gemma":0.000002570907,"teacher_disagreement_score":0.9842004,"about_ca_system_score_codex":0.00029900996,"about_ca_system_score_gemma":0.0022435247,"threshold_uncertainty_score":0.95107615},"labels":[],"label_agreement":null},{"id":"W4399295615","doi":"10.1111/jep.13998","title":"Dual process models of clinical reasoning: The central role of knowledge in diagnostic expertise","year":2024,"lang":"en","type":"review","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":39,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Cognition; Process (computing); Dual process theory (moral psychology); Computer science; Cognitive science; Cognitive psychology; Experiential knowledge; Experiential learning; Psychology; Cognitive bias; Dual (grammatical number); Epistemology; Neuroscience","score_opus":0.24489551615781588,"score_gpt":0.5910585093743066,"score_spread":0.3461629932164907,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4399295615","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.006118555,0.9869744,0.000041285366,0.00091742625,0.0022008666,0.0014429177,0.00001469044,0.000007659457,0.002282204],"genre_scores_gemma":[0.05822778,0.9390729,0.00074202055,0.00013201525,0.001647587,0.00006124065,0.000016119877,0.00005761455,0.000042688713],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.97581536,0.0065049226,0.014669912,0.00059554557,0.0020099715,0.00040428055],"domain_scores_gemma":[0.35326687,0.6256781,0.0147976475,0.0011774292,0.0043745665,0.0007053678],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.056524195,0.000410489,0.004670611,0.0004175841,0.000024590465,0.0000306437,0.00040710566,0.0009984958,0.00009693598],"category_scores_gemma":[0.85864776,0.00024507433,0.00193075,0.0010932289,0.0004026748,0.00034859614,0.00014451783,0.004257251,0.000024858096],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009498695,0.006336323,0.003290992,0.0032263238,0.00070357445,0.00025902083,0.000804309,0.00035437528,5.863058e-8,0.00036341161,0.0019003848,0.98181134],"study_design_scores_gemma":[0.01459474,0.0065954593,0.014900953,0.61187893,0.05195695,0.0019456352,0.0030501306,0.03538899,0.0000019755103,0.008878647,0.24988554,0.0009220545],"about_ca_topic_score_codex":0.00004838773,"about_ca_topic_score_gemma":0.000024959843,"teacher_disagreement_score":0.9808893,"about_ca_system_score_codex":0.00024291324,"about_ca_system_score_gemma":0.0077451933,"threshold_uncertainty_score":0.9993841},"labels":[],"label_agreement":null},{"id":"W4399494727","doi":"10.1037/xge0001610","title":"The exaggerated benefits of failure.","year":2024,"lang":"en","type":"article","venue":"Journal of Experimental Psychology General","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Kellogg's (Canada)","funders":"","keywords":"PsycINFO; Psychology; Addiction; Medical education; MEDLINE; Medicine; Political science; Psychiatry; Law","score_opus":0.03830303880381286,"score_gpt":0.41457452922548427,"score_spread":0.3762714904216714,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4399494727","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9603012,0.030800218,0.000028752886,0.004692232,0.0020547677,0.000067545894,0.000003190662,0.000011663966,0.0020403955],"genre_scores_gemma":[0.9962123,0.0005510463,0.0010134407,0.00083978334,0.00079385727,0.0000026451967,0.0000022735126,0.000015043973,0.00056960067],"study_design_codex":"bench_or_experimental","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.9988531,0.000068576286,0.0005718668,0.00012655053,0.0002159318,0.00016397657],"domain_scores_gemma":[0.9979285,0.0015187241,0.00014685014,0.00016751907,0.00010573662,0.0001326713],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00047567356,0.00009961607,0.00027721573,0.000080376776,0.00004961502,0.000020760928,0.00011748733,0.00009859797,0.00019504949],"category_scores_gemma":[0.0015024972,0.00005706708,0.00022238694,0.00014889415,0.00016914301,0.000039871866,0.000020464086,0.00034605458,0.000022806042],"study_design_candidate":"bench_or_experimental","study_design_consensus":"bench_or_experimental","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019644122,0.0016247486,0.0069485838,0.000018058188,0.00094369496,0.000598829,0.0010044093,0.000057370333,0.56384957,0.0071205734,0.38293055,0.032939192],"study_design_scores_gemma":[0.013072821,0.018562907,0.14461976,0.006691172,0.00072094053,0.017546715,0.0018578522,0.0005884638,0.6832425,0.0017766092,0.11064627,0.0006739686],"about_ca_topic_score_codex":0.000003750729,"about_ca_topic_score_gemma":8.068826e-7,"teacher_disagreement_score":0.2722843,"about_ca_system_score_codex":0.000028167611,"about_ca_system_score_gemma":0.00008414252,"threshold_uncertainty_score":0.2327128},"labels":[],"label_agreement":null},{"id":"W4399760306","doi":"10.1111/medu.15466","title":"Personal career decisions during medical training are not complicated, they are complex","year":2024,"lang":"en","type":"review","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Subconscious; Business decision mapping; Computer science; Inference; Happiness; Fallacy; Psychology; Management science; Decision support system; Social psychology; Artificial intelligence; Medicine; Economics","score_opus":0.22065173863325252,"score_gpt":0.47209518296784597,"score_spread":0.2514434443345934,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4399760306","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0007833374,0.9747374,0.000036525405,0.019221416,0.0023448837,0.0008436532,0.00012235044,0.00028334354,0.0016270903],"genre_scores_gemma":[0.0045689065,0.9823502,0.00016515786,0.005445556,0.0038413794,0.0003845195,0.0016185099,0.00014458445,0.0014811953],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99356574,0.00026358184,0.0015003192,0.0010729983,0.002963564,0.00063378486],"domain_scores_gemma":[0.9844708,0.011740683,0.00053785596,0.0007608456,0.00031922423,0.0021706263],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity","insufficient_payload"],"category_scores_codex":[0.001273009,0.0006323514,0.00266669,0.0003594858,0.0001969296,0.00008005776,0.00055553176,0.0013594769,0.003989831],"category_scores_gemma":[0.19650495,0.00043666107,0.0010030132,0.0005438726,0.00033257323,0.00002737262,0.0002435805,0.0027191753,0.0012890395],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000017365586,0.000696734,0.00010900947,0.0041686264,0.00023561358,0.0003388947,0.00047931037,3.5668e-8,6.0858575e-8,0.000080359,0.06325897,0.930615],"study_design_scores_gemma":[0.0005418403,0.000051190502,0.0023344865,0.2726348,0.0020145199,0.0019176044,0.0006144432,0.00014898278,5.489025e-8,0.00005991546,0.71930844,0.00037373535],"about_ca_topic_score_codex":0.0001109681,"about_ca_topic_score_gemma":0.00010431895,"teacher_disagreement_score":0.9302413,"about_ca_system_score_codex":0.0004294518,"about_ca_system_score_gemma":0.009552492,"threshold_uncertainty_score":0.99993694},"labels":[],"label_agreement":null},{"id":"W4399819980","doi":"10.1515/9782760535053-022","title":"Divergences de vues sur l’échec d’un stagiaire","year":2012,"lang":"fr","type":"book-chapter","venue":"Presses de l'Université du Québec eBooks","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Geography","score_opus":0.018775866390372657,"score_gpt":0.22903431381325914,"score_spread":0.2102584474228865,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4399819980","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.32711583,0.06215917,0.0009492013,0.0024916304,0.0017819036,0.00092591305,0.00031947333,0.00048803265,0.6037688],"genre_scores_gemma":[0.8610239,0.009799287,0.00061241,0.0007666858,0.0016929039,0.000012214898,0.00007711637,0.00014547845,0.12586997],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99646455,0.00016965819,0.00065801654,0.0007957572,0.0006242728,0.0012877322],"domain_scores_gemma":[0.98474354,0.012147138,0.00050991616,0.0008188535,0.00038373336,0.0013968451],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00059481774,0.0008497505,0.0011346035,0.00019006938,0.00047018306,0.00006929459,0.00077668944,0.0010258887,0.043349937],"category_scores_gemma":[0.007815259,0.0008601871,0.0006920789,0.000037378923,0.0010952716,0.00014722317,0.0006855094,0.0011647918,0.0028852099],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00066287187,0.00041738135,0.34043986,0.00054745417,0.0017031456,0.0023681296,0.008832312,0.000042146574,0.00027876522,0.014082803,0.011285664,0.61933947],"study_design_scores_gemma":[0.003973259,0.00092765584,0.092589945,0.012444693,0.0075693876,0.0005868184,0.0014589283,0.0003144544,0.0025223633,0.0031429871,0.8724822,0.0019873409],"about_ca_topic_score_codex":0.06391929,"about_ca_topic_score_gemma":0.012492766,"teacher_disagreement_score":0.8611965,"about_ca_system_score_codex":0.0008926557,"about_ca_system_score_gemma":0.0021345369,"threshold_uncertainty_score":0.9993849},"labels":[],"label_agreement":null},{"id":"W4399994530","doi":"10.1016/j.soin.2024.05.013","title":"Éclairer le raisonnement clinique des étudiants en pratique avancée par les théories infirmières","year":2024,"lang":"fr","type":"article","venue":"Soins","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval","funders":"","keywords":"Psychology; Philosophy","score_opus":0.03535612643756663,"score_gpt":0.3623309880694228,"score_spread":0.3269748616318562,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4399994530","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7529326,0.10417154,0.0034185466,0.105667286,0.010009438,0.0011918375,0.00026082672,0.00059204595,0.021755856],"genre_scores_gemma":[0.9612008,0.018583637,0.0014168435,0.00081575813,0.003360693,0.00011959334,0.000065484644,0.00008982298,0.014347403],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.9970854,0.0002836896,0.0008313458,0.00066433253,0.00036032838,0.0007749362],"domain_scores_gemma":[0.9814856,0.017288065,0.00010272078,0.0004858062,0.0001482097,0.0004896167],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0011770162,0.0004078809,0.00064549694,0.00012555358,0.00029548546,0.00022323216,0.00022502302,0.0005063661,0.0012076113],"category_scores_gemma":[0.031414133,0.000354342,0.00033904877,0.00037590764,0.00088615523,0.0003413347,0.00020748739,0.0008646028,0.0010225152],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00017298457,0.0019044955,0.39308664,0.001743803,0.0009877433,0.0018038523,0.005283414,0.00012552697,0.0001640287,0.12764099,0.21643017,0.25065634],"study_design_scores_gemma":[0.0029007155,0.0018714718,0.2708598,0.04534012,0.0009091957,0.00016301082,0.0026472362,0.0013785163,0.0041603697,0.041335206,0.62747127,0.00096309755],"about_ca_topic_score_codex":0.0017484259,"about_ca_topic_score_gemma":0.0034291788,"teacher_disagreement_score":0.41104108,"about_ca_system_score_codex":0.00022974687,"about_ca_system_score_gemma":0.002988497,"threshold_uncertainty_score":0.99989086},"labels":[],"label_agreement":null},{"id":"W4400002127","doi":"10.1136/bmj-2023-076814","title":"Reducing unnecessary investigations in paediatric seizures in the emergency department","year":2024,"lang":"en","type":"article","venue":"BMJ","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Izaak Walton Killam Health Centre; Dalhousie University; Hospital for Sick Children; University of Toronto","funders":"","keywords":"Emergency department; Medicine; Medical emergency; Pediatrics; Computer science; Psychiatry","score_opus":0.041347066188719625,"score_gpt":0.3771664859660487,"score_spread":0.33581941977732904,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4400002127","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95294136,0.004688267,0.000013588414,0.034561142,0.0005114362,0.00041470717,0.0000055586665,0.000043400425,0.006820547],"genre_scores_gemma":[0.9975286,0.0004565722,0.0003001074,0.0007023621,0.0004817358,0.000119094555,0.00003469061,0.000008551032,0.0003683025],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9990688,0.00008255343,0.000344408,0.00017511277,0.00018038882,0.00014876778],"domain_scores_gemma":[0.9965127,0.0031934755,0.000024235223,0.000191748,0.000013457643,0.00006434419],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0007975958,0.00006874089,0.00011773855,0.00012140759,0.000021839556,0.000013959132,0.00006184538,0.000052400206,0.00011814746],"category_scores_gemma":[0.016124217,0.00004462562,0.000053976673,0.00066774414,0.000023916986,0.000029047691,0.00002080423,0.0002282083,0.00008080871],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000028721477,0.00046021634,0.47940502,0.00013636383,0.00003487636,0.0015315758,0.0030142905,0.00066113874,0.00004038262,0.0032845025,0.485182,0.026220942],"study_design_scores_gemma":[0.000724687,0.00016117102,0.9715551,0.0023023316,0.00011022199,0.00011731853,0.00058305974,0.003249859,0.000047108668,0.0077757426,0.013208046,0.00016535043],"about_ca_topic_score_codex":0.00021478315,"about_ca_topic_score_gemma":0.000121185825,"teacher_disagreement_score":0.4921501,"about_ca_system_score_codex":0.00003589103,"about_ca_system_score_gemma":0.0001618539,"threshold_uncertainty_score":0.99216336},"labels":[],"label_agreement":null},{"id":"W4400045905","doi":"10.1007/s40279-024-02054-9","title":"Judgement and Decision Making in Clinical and Return-to-Sports Decision Making: A Narrative Review","year":2024,"lang":"en","type":"review","venue":"Sports Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Judgement; Context (archaeology); Narrative; Sports medicine; Heuristics; Quality (philosophy); Decision quality; Applied psychology; Psychology; Management science; Computer science; Data science; Knowledge management","score_opus":0.07485606500098263,"score_gpt":0.4885229922717845,"score_spread":0.41366692727080184,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4400045905","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0011104874,0.991245,0.0001039046,0.0015493742,0.0019112122,0.003064937,0.000010225366,0.00007732011,0.0009275926],"genre_scores_gemma":[0.0014041688,0.9898907,0.0029392163,0.0042440146,0.0009990915,0.00011379437,0.00007328458,0.00010783155,0.00022788983],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99210465,0.00014364756,0.004068997,0.0017258349,0.0014248575,0.0005320165],"domain_scores_gemma":[0.9851661,0.012112158,0.0007870753,0.0010072987,0.00018212807,0.0007452713],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0065999627,0.00084661995,0.0059534637,0.0007071175,0.00006706351,0.000027377104,0.00018087398,0.00063323224,0.0005064458],"category_scores_gemma":[0.061515503,0.0005121209,0.00044596003,0.0011330319,0.00036963972,0.0000460785,0.00040595882,0.0016056051,0.000031875155],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013911728,0.00010862933,0.0019332807,0.015918955,0.00007719183,0.0040909643,0.00023660326,4.0811557e-8,5.403399e-9,0.000036870624,0.045464057,0.93199426],"study_design_scores_gemma":[0.0006478212,0.00038415167,0.0045411414,0.57939523,0.0027633586,0.00066465104,0.00006769741,0.0000058778082,2.8603362e-9,0.0007975522,0.41046745,0.00026508796],"about_ca_topic_score_codex":0.000008017602,"about_ca_topic_score_gemma":0.000013619656,"teacher_disagreement_score":0.9317292,"about_ca_system_score_codex":0.00014317532,"about_ca_system_score_gemma":0.00046220672,"threshold_uncertainty_score":0.99973303},"labels":[],"label_agreement":null},{"id":"W4400408157","doi":"10.1371/journal.pone.0297019","title":"Herding-like behaviour in medical decision making: An experimental study investigating general practitioners’ prescription behaviour","year":2024,"lang":"en","type":"article","venue":"PLoS ONE","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Saskatoon City Hospital Foundation","keywords":"Vignette; Herding; Medical prescription; Medicine; Observational study; Family medicine; Psychology; Internal medicine; Nursing; Social psychology","score_opus":0.09454815411337907,"score_gpt":0.39880432677717265,"score_spread":0.30425617266379357,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4400408157","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9975748,0.00043491204,0.000050208306,0.00043709995,0.0003321533,0.0007091376,0.0000069482317,0.0002267288,0.00022801057],"genre_scores_gemma":[0.9921897,0.000019311645,0.0062755174,0.00050863693,0.00050389953,0.00016191868,0.000079361336,0.0000499348,0.00021168645],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99703217,0.00015658702,0.00059631397,0.00058335566,0.0013320055,0.0002995427],"domain_scores_gemma":[0.9975857,0.0015322705,0.000072163566,0.00032113367,0.000073022566,0.00041567674],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0009015897,0.0001959397,0.00038727635,0.00019699614,0.00007848676,0.000112317786,0.00011982176,0.00020632801,0.00058181514],"category_scores_gemma":[0.010615339,0.00017923959,0.00007663027,0.00029714074,0.0000758863,0.000250483,0.00010424872,0.0006577617,0.00006716558],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018103317,0.031147745,0.94788563,0.000041411997,0.00011223,0.0017403238,0.0031294737,0.000007410941,0.011531224,0.0000929645,0.00076631515,0.0033642408],"study_design_scores_gemma":[0.0048296014,0.005488659,0.9387728,0.022183895,0.0009540833,0.00018075689,0.0046238233,0.015566253,0.0065823672,0.0002592613,0.000015057906,0.0005434533],"about_ca_topic_score_codex":0.00027212003,"about_ca_topic_score_gemma":0.000103708655,"teacher_disagreement_score":0.025659086,"about_ca_system_score_codex":0.00019758273,"about_ca_system_score_gemma":0.00021181251,"threshold_uncertainty_score":0.9977187},"labels":[],"label_agreement":null},{"id":"W4400458788","doi":"10.5430/jnep.v14n11p1","title":"Nursing explanation skills in education and practice: Development skills and influence on incident occurrence","year":2024,"lang":"en","type":"article","venue":"Journal of Nursing Education and Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Compassion; Nursing; Scale (ratio); Psychology; Explanatory model; Empathy; Patient safety; Mental health; Explanatory power; Nurse education; Empirical research; Health care; Medicine; Social psychology; Psychiatry","score_opus":0.04506082885316422,"score_gpt":0.4775102241044313,"score_spread":0.4324493952512671,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4400458788","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93418944,0.011142224,0.00029045017,0.04853598,0.0020664025,0.0002267216,0.0000012478916,0.000014759349,0.0035327955],"genre_scores_gemma":[0.97913074,0.004809494,0.012829225,0.0027695647,0.00027439775,0.0000068273316,0.000010337227,0.000009662853,0.00015972344],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99858713,0.00023520188,0.0005074152,0.00021502061,0.00033026346,0.0001249875],"domain_scores_gemma":[0.9829014,0.015880506,0.00040651325,0.00009415288,0.00044496031,0.0002724949],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001810637,0.00012057798,0.00019218132,0.00036542118,0.0001193553,0.00017383884,0.000034482655,0.00008704274,0.000008667009],"category_scores_gemma":[0.09659055,0.00010531573,0.000019702178,0.00025232113,0.00008700282,0.000947207,0.000008881069,0.00043063983,0.0000037169875],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00025279715,0.004707933,0.0023740423,0.000061654,0.000024459716,0.000020168764,0.022122454,0.000009622748,0.000097933196,0.00061578886,0.004526594,0.96518654],"study_design_scores_gemma":[0.0018381422,0.0012784583,0.7356645,0.08612735,0.00085419626,0.03598721,0.036022082,0.00012646511,0.0003417953,0.0031726928,0.098124705,0.00046240402],"about_ca_topic_score_codex":0.000016393587,"about_ca_topic_score_gemma":0.0000011594634,"teacher_disagreement_score":0.9647241,"about_ca_system_score_codex":0.0003972854,"about_ca_system_score_gemma":0.0020957724,"threshold_uncertainty_score":0.91101927},"labels":[],"label_agreement":null},{"id":"W4400857314","doi":"10.7759/cureus.64954","title":"Mock Wards: Incorporating a Theoretical Framework to Create a Blended Virtual and In-Person Clinical Reasoning Education Platform","year":2024,"lang":"en","type":"article","venue":"Cureus","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"McMaster University","keywords":"Interactivity; Medical education; Medicine; Qualitative research; Virtual patient; Modalities; Multimedia; Psychology; Computer science","score_opus":0.0340249971665975,"score_gpt":0.3913075326848056,"score_spread":0.3572825355182081,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4400857314","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9821638,0.0015351295,0.0010398406,0.0035828091,0.0009903795,0.00040289178,0.0000064490355,0.00015091432,0.010127825],"genre_scores_gemma":[0.9888555,0.00018363824,0.008463687,0.0011807911,0.00082771305,0.00005027873,0.000020331394,0.00003385857,0.0003842497],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"observational","domain_scores_codex":[0.9983819,0.000063968306,0.0004962159,0.00052829215,0.00024516435,0.0002844489],"domain_scores_gemma":[0.99016154,0.008935473,0.00005230264,0.00030239407,0.00006692061,0.00048139048],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010152888,0.00018064922,0.00041027172,0.00009894035,0.00005185666,0.00009018858,0.0000797466,0.0002929025,0.00013907929],"category_scores_gemma":[0.066295296,0.00014556738,0.0001046026,0.00032548484,0.00017719243,0.00008392746,0.00006950441,0.00072162534,0.00009186738],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010659294,0.00107969,0.13334565,0.00012888525,0.00016772041,0.00041317518,0.004454018,0.000015959742,0.0000914964,0.5728567,0.0054261335,0.2809546],"study_design_scores_gemma":[0.0061076693,0.009903567,0.46062747,0.1043024,0.0013634217,0.0011169583,0.018015582,0.04375611,0.00059004547,0.3476773,0.004421783,0.002117701],"about_ca_topic_score_codex":0.000060386883,"about_ca_topic_score_gemma":0.000012647403,"teacher_disagreement_score":0.32728183,"about_ca_system_score_codex":0.00010498917,"about_ca_system_score_gemma":0.0005807789,"threshold_uncertainty_score":0.9415697},"labels":[],"label_agreement":null},{"id":"W4401029491","doi":"10.1111/2047-3095.12484","title":"Clinical reasoning and clinical judgment in nursing research: A bibliometric analysis","year":2024,"lang":"en","type":"article","venue":"International Journal of Nursing Knowledge","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières","funders":"","keywords":"Clinical judgment; Thematic analysis; Competence (human resources); Nursing; Psychology; Nursing process; Nurse education; Critical thinking; MEDLINE; Medicine; Medical education; Qualitative research; Pedagogy; Social psychology; Sociology","score_opus":0.19113135818044688,"score_gpt":0.5874985959632787,"score_spread":0.3963672377828318,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4401029491","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9558897,0.019817064,0.0014637997,0.0055509303,0.0051798304,0.00012462548,0.0000045355355,0.000028502813,0.0119410185],"genre_scores_gemma":[0.99108326,0.0039508794,0.0023859923,0.00007677318,0.0021627853,0.0000014512028,0.000006661982,0.000021582575,0.00031059954],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9959498,0.0005457089,0.001794607,0.00038999642,0.0010075932,0.00031229726],"domain_scores_gemma":[0.97421277,0.023492778,0.0002509034,0.00019701393,0.0013483049,0.0004982287],"candidate_categories":["metaresearch","bibliometrics"],"consensus_categories":["bibliometrics"],"category_scores_codex":[0.009917001,0.00014152135,0.00070038217,0.043900274,0.000044902317,0.0002130962,0.0002626255,0.00021328978,0.00006165816],"category_scores_gemma":[0.05089252,0.00011551012,0.0006192264,0.030001182,0.00049357145,0.00016032677,0.000052343512,0.0015535557,0.000034669833],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00035044324,0.0020037317,0.3039543,0.0000075582157,0.000788884,0.0006744666,0.00048119604,0.000006827763,0.0000059392605,0.00062823755,0.0054527,0.6856457],"study_design_scores_gemma":[0.0021510439,0.0009939363,0.94747615,0.03365878,0.0011326813,0.0006778654,0.0003291514,0.009122889,0.000008637564,0.0026224453,0.0017016771,0.00012472915],"about_ca_topic_score_codex":0.000010214786,"about_ca_topic_score_gemma":0.000005594855,"teacher_disagreement_score":0.685521,"about_ca_system_score_codex":0.0005961909,"about_ca_system_score_gemma":0.00055265555,"threshold_uncertainty_score":0.9906166},"labels":[{"model":"gemma","categories":["bibliometrics"],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["bibliometrics"],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4401280181","doi":"10.1111/medu.15479","title":"Understanding cultural dynamics shaping clinical reasoning skills: A dialogical exploration","year":2024,"lang":"en","type":"article","venue":"Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Dialogical self; Psychology; Collectivism; Hofstede's cultural dimensions theory; Cultural diversity; Individualism; Dynamics (music); Cultural competence; Social psychology; Medical education; Pedagogy; Sociology; Medicine","score_opus":0.17385429714245293,"score_gpt":0.4576839562993882,"score_spread":0.28382965915693525,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4401280181","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5413395,0.004176103,0.22741918,0.17004693,0.018685544,0.001289435,0.000013439313,0.0016186396,0.03541123],"genre_scores_gemma":[0.9861115,0.0010744544,0.002013493,0.0068766954,0.0024972558,0.000055289252,0.00046673682,0.0000305053,0.00087402033],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99755,0.00014927481,0.0007116207,0.00049663393,0.0007613814,0.0003310885],"domain_scores_gemma":[0.99269956,0.0058987997,0.00009000265,0.000267029,0.00014128644,0.0009033029],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0017557816,0.00017165212,0.0003515225,0.000111862304,0.00011530285,0.00012521964,0.000103096216,0.0003837362,0.00055267976],"category_scores_gemma":[0.18896244,0.00012778505,0.00022008798,0.00041522694,0.00022429017,0.00022932362,0.000047701807,0.0007707769,0.00017735477],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015470505,0.0029882966,0.050824072,0.00024614067,0.00020516795,0.00025640818,0.001923538,0.000010729754,0.000010686166,0.14065264,0.25481373,0.54791385],"study_design_scores_gemma":[0.006612847,0.0028470873,0.07686698,0.0798788,0.0018083997,0.0018673877,0.014303355,0.6573842,0.00004836709,0.06510414,0.091187306,0.0020910895],"about_ca_topic_score_codex":0.00005576887,"about_ca_topic_score_gemma":0.000033723885,"teacher_disagreement_score":0.6573735,"about_ca_system_score_codex":0.0005925405,"about_ca_system_score_gemma":0.0025937343,"threshold_uncertainty_score":0.81786925},"labels":[],"label_agreement":null},{"id":"W4401758086","doi":"10.3138/jvme-2024-0017","title":"Teaching and Assessment of Clinical Reasoning Skills in a Case-Based Veterinary Cardiology Elective","year":2024,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Curriculum; Medical education; Medicine; Peer assessment; Internal medicine; Veterinary medicine; Psychology; Pedagogy","score_opus":0.08119777807201806,"score_gpt":0.5147142406355621,"score_spread":0.4335164625635441,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4401758086","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9918735,0.0016603632,0.00081571477,0.003211513,0.0014859603,0.00013378417,0.0000018788892,0.00001184276,0.0008054598],"genre_scores_gemma":[0.9870927,0.0003724164,0.011046407,0.0005278625,0.0008872674,0.00000880008,0.000006126244,0.000015802996,0.000042616222],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9971686,0.0007662041,0.0012427719,0.00022896063,0.0004217537,0.00017168936],"domain_scores_gemma":[0.9879666,0.01102732,0.00029746746,0.0001463528,0.000151559,0.00041067688],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005263454,0.00012661287,0.0007242743,0.0002870022,0.000034130557,0.000013621909,0.00007499951,0.00023525525,0.000060981605],"category_scores_gemma":[0.054064494,0.000098167075,0.00021274056,0.00016331759,0.00016862621,0.000086847715,0.00004185507,0.0015822988,9.451596e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007935489,0.0023500356,0.13036333,0.0004538622,0.00023960815,0.029928248,0.0005963566,0.000015191151,0.00036310486,0.00027041725,0.0015035598,0.83312273],"study_design_scores_gemma":[0.0037202637,0.020300116,0.772234,0.025376154,0.00038497738,0.15560278,0.0008929002,0.00974056,0.000011365353,0.00027990964,0.011236273,0.00022065424],"about_ca_topic_score_codex":0.00011981203,"about_ca_topic_score_gemma":0.0000022342076,"teacher_disagreement_score":0.8329021,"about_ca_system_score_codex":0.00012549729,"about_ca_system_score_gemma":0.0039829835,"threshold_uncertainty_score":0.9539035},"labels":[],"label_agreement":null},{"id":"W4401879307","doi":"10.2196/50935","title":"Evaluation of a Natural Language Processing Approach to Identify Diagnostic Errors and Analysis of Safety Learning System Case Review Data: Retrospective Cohort Study","year":2024,"lang":"en","type":"article","venue":"Journal of Medical Internet Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto East General Hospital; University of Toronto","funders":"Agency for Healthcare Research and Quality","keywords":"Medicine; Logistic regression; Retrospective cohort study; Patient safety; Diagnosis code; Artificial intelligence; Emergency medicine; Health care; Medical emergency; Surgery; Computer science; Population; Internal medicine","score_opus":0.13353218047613308,"score_gpt":0.5479332831673962,"score_spread":0.4144011026912631,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4401879307","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.92420524,0.07331349,0.0006282243,0.00058014836,0.00013762117,0.00082009594,0.00001190904,0.000009824427,0.00029346254],"genre_scores_gemma":[0.9977566,0.0018091875,0.00011498271,0.00002330253,0.00015058921,0.000015446432,0.000031711872,0.0000128996035,0.00008525088],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99009424,0.0016376313,0.0012522767,0.00038017263,0.0064246487,0.00021101761],"domain_scores_gemma":[0.9828174,0.013269479,0.00034575298,0.00040392496,0.0026946447,0.0004687661],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.06387843,0.00011177986,0.0010624767,0.0007266533,0.000034229317,0.000049145012,0.00034181867,0.00010292226,0.00012960697],"category_scores_gemma":[0.37552488,0.000072801784,0.0001551616,0.0017456929,0.00017682227,0.000106703796,0.00035153035,0.0015776289,0.0000023889106],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006237385,0.002554096,0.6362677,0.009903414,0.012380214,0.013825946,0.010448179,0.000116945186,0.000057871417,0.0000721538,0.0066458397,0.30710387],"study_design_scores_gemma":[0.0016236783,0.0016225544,0.4922284,0.09145498,0.017306006,0.003745365,0.0133549385,0.37847933,0.000015703123,0.0000040328773,0.000043584503,0.0001214495],"about_ca_topic_score_codex":0.0006697769,"about_ca_topic_score_gemma":0.000078882535,"teacher_disagreement_score":0.3783624,"about_ca_system_score_codex":0.00034211765,"about_ca_system_score_gemma":0.0007213386,"threshold_uncertainty_score":0.9639342},"labels":[],"label_agreement":null},{"id":"W4401890417","doi":"10.1016/j.gpeds.2024.100229","title":"Knowledge, confidence and clinical reasoning of dental students for diagnosing pediatric periodontal diseases","year":2024,"lang":"en","type":"article","venue":"Global Pediatrics","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Medicine; Dentistry","score_opus":0.026533888417830123,"score_gpt":0.4203708772641076,"score_spread":0.3938369888462775,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4401890417","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9571085,0.039130483,0.0005835626,0.000100467936,0.0016607017,0.0005105399,0.00034223363,0.00011984158,0.0004436678],"genre_scores_gemma":[0.98899376,0.0074052424,0.001047691,0.000094212264,0.0022572712,0.00002317763,0.000055819488,0.000023138859,0.00009967089],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9979062,0.000056452747,0.000728352,0.0005368006,0.00043082846,0.0003413841],"domain_scores_gemma":[0.98789924,0.011084791,0.00012501676,0.00021493813,0.00017980798,0.0004962308],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00057407684,0.00021168741,0.00053878187,0.000091553586,0.000080880025,0.00010852644,0.00015245096,0.00018959277,0.000027466798],"category_scores_gemma":[0.034137476,0.0001871889,0.00034151506,0.0004948758,0.00016416625,0.000092081435,0.00018274081,0.00021052573,0.000025932783],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014557212,0.0005955438,0.9817422,0.00038486696,0.000052292668,0.00008623055,0.000047580652,0.0000010179872,4.1574452e-7,0.0009547018,0.0065703564,0.009419245],"study_design_scores_gemma":[0.0021255328,0.0008895708,0.9920728,0.0007064913,0.002102363,0.00008924856,0.00017548274,0.00054094987,0.000002642599,0.00042352147,0.0006405186,0.00023088713],"about_ca_topic_score_codex":0.000038280545,"about_ca_topic_score_gemma":0.0000118328035,"teacher_disagreement_score":0.0335634,"about_ca_system_score_codex":0.00009647617,"about_ca_system_score_gemma":0.00044884527,"threshold_uncertainty_score":0.97399837},"labels":[],"label_agreement":null},{"id":"W4401945506","doi":"10.1007/s40596-024-02033-2","title":"Exploring Perceived Bias at Psychiatry Grand Rounds","year":2024,"lang":"en","type":"letter","venue":"Academic Psychiatry","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"Dalhousie University","keywords":"Psychology; Grand Challenges; MEDLINE; Medical education; Psychiatry; Medicine; Computer science; Political science","score_opus":0.14438080041176327,"score_gpt":0.3542076965172675,"score_spread":0.20982689610550426,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4401945506","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.14520513,0.018255316,0.000005160499,0.79466254,0.032782562,0.00058510003,0.00011750646,0.00067866786,0.007708044],"genre_scores_gemma":[0.0046646968,0.01162555,0.0011657628,0.8853515,0.06048492,0.00028170887,0.0011447388,0.00044199004,0.034839127],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99418724,0.00014542934,0.0015107719,0.0017106847,0.0012093751,0.0012365088],"domain_scores_gemma":[0.9939883,0.0037147983,0.00039430722,0.0012829653,0.00009377641,0.0005258992],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0006349898,0.0010059616,0.001488314,0.00052277173,0.00022136582,0.000073382675,0.0006016222,0.0034841916,0.00062678003],"category_scores_gemma":[0.004608858,0.00085323903,0.0011482526,0.00059745833,0.00026169416,0.00016489527,0.00029789266,0.013258834,0.003906488],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013362967,0.000068330985,0.045865912,0.00079353125,0.0004219244,0.00022879615,0.00016729566,7.579602e-7,0.000007611867,0.0001243034,0.9510433,0.0011446156],"study_design_scores_gemma":[0.0021230243,0.0003324949,0.02333061,0.022503339,0.002098211,0.0009162987,0.000101832135,0.000017164906,0.0000033131805,0.0053645526,0.9421982,0.0010109298],"about_ca_topic_score_codex":0.000063370935,"about_ca_topic_score_gemma":0.000017622064,"teacher_disagreement_score":0.14054042,"about_ca_system_score_codex":0.0003607225,"about_ca_system_score_gemma":0.0010667038,"threshold_uncertainty_score":0.99939185},"labels":[],"label_agreement":null},{"id":"W4402146967","doi":"10.1515/dx-2024-0102","title":"Root cause analysis of cases involving diagnosis","year":2024,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"Gordon and Betty Moore Foundation","keywords":"Root cause analysis; Root cause; Root (linguistics); Process (computing); Perspective (graphical); Psychological intervention; Medicine; Intensive care medicine; Medical emergency; Computer science; Risk analysis (engineering); Psychology; Forensic engineering; Artificial intelligence; Engineering; Operations management; Nursing","score_opus":0.05419566207748881,"score_gpt":0.3744757925957886,"score_spread":0.3202801305182998,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402146967","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98161423,0.014491088,0.000045564328,0.002121641,0.00038185366,0.00022384702,0.00013417867,0.00019775738,0.0007898457],"genre_scores_gemma":[0.99295413,0.00540564,0.0002313274,0.0005076285,0.00017688074,0.0003353759,0.00014060296,0.000036663427,0.0002117687],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99812037,0.000060016293,0.00059018546,0.00049443624,0.00041674505,0.00031822102],"domain_scores_gemma":[0.9614889,0.037482526,0.00008346295,0.0005045792,0.00014490167,0.00029560475],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00036254403,0.00021509951,0.0007881975,0.0007003111,0.000050700964,0.000057558314,0.00011592234,0.00015646114,0.001715687],"category_scores_gemma":[0.042049818,0.00017905283,0.0007006084,0.0023241816,0.000121846126,0.00008382821,0.00007417845,0.00022473451,0.000113955335],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000023433931,0.00054696004,0.95505214,0.00011823108,0.0029762762,0.001249587,0.00027862124,0.00006699032,0.000019309513,0.00079603423,0.02039683,0.018475596],"study_design_scores_gemma":[0.0006166923,0.0005537898,0.9554064,0.005350851,0.020800618,0.000048838105,0.00015722634,0.0026700494,0.005035451,0.00018031964,0.008813489,0.00036624414],"about_ca_topic_score_codex":0.0011826606,"about_ca_topic_score_gemma":0.0004113712,"teacher_disagreement_score":0.041687276,"about_ca_system_score_codex":0.000081788065,"about_ca_system_score_gemma":0.0001208919,"threshold_uncertainty_score":0.9991969},"labels":[],"label_agreement":null},{"id":"W4402221982","doi":"10.1001/jamanetworkopen.2024.31123","title":"Post Hoc Bias in Treatment Decisions","year":2024,"lang":"en","type":"article","venue":"JAMA Network Open","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; University of Toronto; Institute for Clinical Evaluative Sciences; Sunnybrook Health Science Centre","funders":"Canadian Institutes of Health Research; Physicians' Services Incorporated Foundation; Alfred P. Sloan Foundation; National Science Foundation","keywords":"Medicine; Sore throat; Post-hoc analysis; Health care; Family medicine; Odds ratio; Odds; Demography; Physical therapy; Logistic regression; Surgery; Internal medicine","score_opus":0.10537938873011837,"score_gpt":0.410092777746629,"score_spread":0.3047133890165106,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402221982","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.92779344,0.0065523745,0.000039480346,0.02197401,0.0007903416,0.0008502521,0.000022557275,0.0000893217,0.041888192],"genre_scores_gemma":[0.98211336,0.0024171714,0.0020072365,0.0026002617,0.0009866947,0.00008948239,0.00010606824,0.000029399109,0.009650349],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99884707,0.00007272099,0.00031916847,0.00032487587,0.00015294783,0.0002832032],"domain_scores_gemma":[0.9819115,0.017472185,0.000025932526,0.0003399232,0.000036642872,0.00021385148],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00067473546,0.00012626701,0.00033624386,0.000056984252,0.00004217858,0.00023001137,0.0001591495,0.00011786702,0.00060825533],"category_scores_gemma":[0.012568875,0.00008516218,0.000111431604,0.00041411805,0.000029213403,0.000077573015,0.00014944884,0.00019330067,0.0008211205],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00076819153,0.00046428517,0.07976924,0.000006461034,0.00011801127,0.0018587639,0.00032350133,0.00029945187,0.000004785937,0.0026399847,0.10924529,0.804502],"study_design_scores_gemma":[0.0056426525,0.002614851,0.578829,0.018108113,0.00022617602,0.0001538681,0.00015978536,0.004257097,0.000021244314,0.0034561313,0.38623285,0.00029825055],"about_ca_topic_score_codex":0.0005709063,"about_ca_topic_score_gemma":0.00035716593,"teacher_disagreement_score":0.8042038,"about_ca_system_score_codex":0.00012609488,"about_ca_system_score_gemma":0.0002629998,"threshold_uncertainty_score":0.99995685},"labels":[],"label_agreement":null},{"id":"W4402332012","doi":"10.1002/acp.4236","title":"The effect of calibration training on the calibration of intelligence analysts' judgments","year":2024,"lang":"en","type":"article","venue":"Applied Cognitive Psychology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"York University; Defence Research and Development Canada","funders":"","keywords":"Psychology; Calibration; Training (meteorology); Applied psychology; Cognitive psychology; Social psychology; Statistics","score_opus":0.04862864532436076,"score_gpt":0.3927501088103708,"score_spread":0.34412146348601,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402332012","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9257112,0.00042974902,0.037490312,0.0052564023,0.00049021863,0.0010409435,0.000023936893,0.000055213568,0.029502034],"genre_scores_gemma":[0.9986565,0.00009680462,0.000023091074,0.0009546421,0.00008234007,0.00009122288,0.000032993234,0.000011962765,0.00005044025],"study_design_codex":"design_other","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.9989471,0.0001508567,0.00035049865,0.00023300288,0.00018258867,0.00013598967],"domain_scores_gemma":[0.9718473,0.027794395,0.00009882887,0.00017692211,0.00004186419,0.000040721596],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00078839407,0.0001035525,0.00024606672,0.000058304682,0.000052269286,0.000011614238,0.00007810203,0.00011837218,0.000073243726],"category_scores_gemma":[0.0055647446,0.000052527594,0.00009437854,0.0002934979,0.00029392223,0.000014377562,0.000015389145,0.00028605742,0.000018929743],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.003457226,0.00023446708,0.0038590173,0.0000910473,0.00083136256,0.000025163461,0.0020572098,0.000031886433,0.009900846,0.07407093,0.002391966,0.9030489],"study_design_scores_gemma":[0.008775525,0.021534849,0.09613948,0.012669487,0.0042292853,0.00013221912,0.006696275,0.032344874,0.76173294,0.053158216,0.0016102891,0.00097653666],"about_ca_topic_score_codex":0.000006876871,"about_ca_topic_score_gemma":0.0000023722623,"teacher_disagreement_score":0.9020723,"about_ca_system_score_codex":0.00000825624,"about_ca_system_score_gemma":0.000034704837,"threshold_uncertainty_score":0.66619205},"labels":[],"label_agreement":null},{"id":"W4402440264","doi":"10.1097/01.nep.0000000000001320","title":"Preliminary Validation of a Clinical Reasoning Theory-Based Assessment Rubric: An e-Delphi Study","year":2024,"lang":"en","type":"article","venue":"Nursing Education Perspectives","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institut du Savoir Montfort; Centre for Interdisciplinary Research in Rehabilitation","funders":"","keywords":"Rubric; CLARITY; Delphi method; Concordance; Psychology; Content validity; Educational measurement; Computer science; Medical education; Psychometrics; Mathematics education; Curriculum; Artificial intelligence; Medicine; Pedagogy; Clinical psychology","score_opus":0.06250483122584286,"score_gpt":0.5092190363425525,"score_spread":0.4467142051167097,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402440264","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98584235,0.0007169528,0.0030754257,0.0020655685,0.0014007073,0.00062431855,0.000004485608,0.00018190582,0.006088291],"genre_scores_gemma":[0.99086434,0.000019008781,0.007828313,0.00007858603,0.0006429001,0.00005637897,0.000061548235,0.00003282274,0.00041610672],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99785876,0.0005050028,0.0005772513,0.0005320886,0.00037198333,0.00015491057],"domain_scores_gemma":[0.9932541,0.005600345,0.00015064805,0.00042248107,0.00038036378,0.00019201385],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0020083578,0.00014586163,0.00035660382,0.00021853778,0.00008098141,0.00006268449,0.00008418499,0.000105175364,0.00020177137],"category_scores_gemma":[0.014103674,0.0001265926,0.00016561394,0.0003596682,0.00019905575,0.00013845248,0.000008775251,0.0003407944,0.000013180071],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0021515833,0.12714343,0.28339812,0.0002077675,0.00047956622,0.000032948195,0.081664935,0.00023798877,0.00033629203,0.05495593,0.0038891302,0.44550228],"study_design_scores_gemma":[0.0010431183,0.0046809963,0.8795495,0.006072457,0.00068293296,0.000018146737,0.10094076,0.0033803384,0.00025737297,0.0031376977,0.00006194535,0.00017473844],"about_ca_topic_score_codex":0.00004125358,"about_ca_topic_score_gemma":7.896611e-7,"teacher_disagreement_score":0.59615135,"about_ca_system_score_codex":0.00023199432,"about_ca_system_score_gemma":0.0020571435,"threshold_uncertainty_score":0.99420094},"labels":[],"label_agreement":null},{"id":"W4402455903","doi":"10.1097/pts.0000000000001273","title":"“What Else Could It Be?” A Scoping Review of Questions for Patients to Ask Throughout the Diagnostic Process","year":2024,"lang":"en","type":"review","venue":"Journal of Patient Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Mary's University; Institute for Work & Health; Toronto East General Hospital; University of Toronto","funders":"Agency for Healthcare Research and Quality","keywords":"MEDLINE; Ask price; Medicine; Inter-rater reliability; Process (computing); Grey literature; Crowdsourcing; Psychology; Medical education; Computer science; World Wide Web; Political science","score_opus":0.07469388895765824,"score_gpt":0.4901281008419663,"score_spread":0.41543421188430807,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402455903","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00003162887,0.98741007,0.00011238206,0.0043698866,0.002498161,0.0052731596,0.00020891885,0.000012697929,0.00008310176],"genre_scores_gemma":[0.000045854373,0.99547553,0.00033029795,0.003150917,0.00045162727,0.000262719,0.00012569595,0.000067625726,0.00008972448],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9947832,0.00021394683,0.003381145,0.00033405673,0.0009727081,0.00031492743],"domain_scores_gemma":[0.965203,0.029767133,0.0025249592,0.00053273805,0.0015820216,0.00039018088],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0013666551,0.00042939855,0.003019357,0.00018337372,0.0000900226,0.00004228565,0.00032984995,0.00023318667,0.000059450656],"category_scores_gemma":[0.16224474,0.0002256916,0.0015349542,0.0005746009,0.000087694854,0.00013252335,0.00010703947,0.0007999649,0.000028509854],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010528734,0.0008648357,0.00006838793,0.3325686,0.000531263,0.000035043027,0.00040647935,0.000007855597,9.083686e-9,0.000046291105,0.024142444,0.6412235],"study_design_scores_gemma":[0.00040203115,0.0015862838,0.000013661031,0.68532264,0.006237623,0.000056720415,0.000048960446,9.931597e-7,2.8008537e-7,0.000043565364,0.3061426,0.00014466138],"about_ca_topic_score_codex":0.0000038665125,"about_ca_topic_score_gemma":0.0000037113964,"teacher_disagreement_score":0.6410788,"about_ca_system_score_codex":0.00023930274,"about_ca_system_score_gemma":0.0017131459,"threshold_uncertainty_score":0.92034364},"labels":[],"label_agreement":null},{"id":"W4402564857","doi":"10.1016/j.sapharm.2024.09.005","title":"A graphical model to make explicit pharmacist clinical reasoning during medication review","year":2024,"lang":"en","type":"article","venue":"Research in Social and Administrative Pharmacy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Pharmacist; Computer science; Medicine; Nursing; Pharmacy","score_opus":0.48904964403379275,"score_gpt":0.6321546394710681,"score_spread":0.14310499543727534,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402564857","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.87457925,0.01801208,0.00028367317,0.096379496,0.00023664114,0.0016641828,0.00008963301,0.00015699588,0.008598035],"genre_scores_gemma":[0.9635667,0.032993857,0.00031582912,0.0015515252,0.0008574898,0.00025330862,0.000028610682,0.000020397374,0.00041225232],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9969085,0.00045336856,0.00068859156,0.00064584904,0.00078402535,0.00051965285],"domain_scores_gemma":[0.99482995,0.003986666,0.000042644424,0.0001427472,0.00018860387,0.000809374],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.003698443,0.00017343668,0.0004930599,0.00014747829,0.00022665002,0.00007428833,0.00012901939,0.00008802277,0.00010179213],"category_scores_gemma":[0.010785277,0.0001464062,0.00016777027,0.0010616017,0.00033339442,0.00007649456,0.00013404056,0.0012435685,0.00006697797],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004525289,0.003714577,0.15541056,0.010145176,0.00065562705,0.00614808,0.0037161254,0.0000030555989,0.0031516594,0.052565232,0.060038015,0.6999266],"study_design_scores_gemma":[0.039809342,0.0047567175,0.34616888,0.27815676,0.0018284626,0.0016910278,0.003250941,0.09067984,0.0041482877,0.046481203,0.17913471,0.0038938401],"about_ca_topic_score_codex":0.000036571797,"about_ca_topic_score_gemma":0.000010444596,"teacher_disagreement_score":0.69603276,"about_ca_system_score_codex":0.00011728398,"about_ca_system_score_gemma":0.00081925147,"threshold_uncertainty_score":0.9975473},"labels":[],"label_agreement":null},{"id":"W4402580158","doi":"10.1097/pts.0000000000001283","title":"Using Patient Experience Surveys to Identify Potential Diagnostic Safety Breakdowns: A Mixed Methods Study","year":2024,"lang":"en","type":"article","venue":"Journal of Patient Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute for Work & Health; Toronto East General Hospital; University of Toronto","funders":"Agency for Healthcare Research and Quality","keywords":"Patient safety; MEDLINE; Medicine; Computer science; Medical emergency; Data science; Health care","score_opus":0.04955137819355488,"score_gpt":0.45133558386809525,"score_spread":0.4017842056745404,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402580158","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9109921,0.0009877591,0.08175184,0.0005157341,0.004776279,0.0007578214,0.000032225827,0.0000440987,0.00014212998],"genre_scores_gemma":[0.9824084,0.00015398266,0.01667861,0.00030063523,0.0003609751,0.000011572838,0.000007731282,0.000045892488,0.000032184154],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9947446,0.0012197528,0.0020365135,0.00045213025,0.0011223048,0.00042469543],"domain_scores_gemma":[0.9843867,0.013410907,0.00047428985,0.00042739062,0.00056909746,0.0007316032],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0035014867,0.00031166335,0.00085089286,0.00035351794,0.00015595579,0.0000941208,0.00019516742,0.00012640479,0.0003894776],"category_scores_gemma":[0.050788227,0.00023385003,0.00048223167,0.0006034641,0.00007866712,0.00019245659,0.00020859981,0.0005804134,0.000051093506],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002094289,0.0047773016,0.057927888,0.00011420484,0.0010392093,0.0069755553,0.018399443,0.00852108,0.003704798,0.00007778666,0.0018493582,0.8945191],"study_design_scores_gemma":[0.004529799,0.013632656,0.9436614,0.010870814,0.0017940965,0.0023618822,0.011768931,0.0023625772,0.002087795,0.00027208516,0.005842636,0.0008153292],"about_ca_topic_score_codex":0.00014898744,"about_ca_topic_score_gemma":0.000011774236,"teacher_disagreement_score":0.89370376,"about_ca_system_score_codex":0.000337406,"about_ca_system_score_gemma":0.00029531348,"threshold_uncertainty_score":0.9572074},"labels":[],"label_agreement":null},{"id":"W4402580678","doi":"10.1097/j.pain.0000000000003367","title":"Defining pain-related suffering requires partnership with people living with pain and careful critical thought: a commentary on the proposed definition by Noe-Steinmüller et al.","year":2024,"lang":"en","type":"article","venue":"Pain","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Kelowna General Hospital; Ethica (Canada); Centre Hospitalier de l’Université de Montréal; McGill University; Université de Montréal; Centre Intégré Universitaire de Santé et de Services Sociaux du Centre-Sud-de-l'Île-de-Montréal; Centre for Interdisciplinary Research in Rehabilitation","funders":"H2020 Marie Skłodowska-Curie Actions; Chronic Pain Centre of Excellence for Canadian Veterans; Social Sciences and Humanities Research Council of Canada; Canadian Institutes of Health Research; HORIZON EUROPE Marie Sklodowska-Curie Actions; Réseau québécois de recherche sur la douleur; Social Sciences and Humanities Research Council; Louise and Alan Edwards Foundation","keywords":"General partnership; Medicine; Psychology; Psychotherapist; Neuroscience; Physical medicine and rehabilitation; Political science","score_opus":0.022968238663852135,"score_gpt":0.30584986357108485,"score_spread":0.28288162490723273,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402580678","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6626246,0.0044754534,0.010362854,0.31603518,0.00014938893,0.001238585,0.000041444422,0.00071686023,0.004355618],"genre_scores_gemma":[0.94708186,0.00006702347,0.0008567011,0.051608726,0.000052911808,0.00009276216,0.00009924,0.000061747836,0.00007904721],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9960166,0.0023608007,0.00032051266,0.00047739135,0.00041463386,0.0004100694],"domain_scores_gemma":[0.8747518,0.12466894,0.00004727298,0.00026115385,0.000056531597,0.00021430399],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.011549586,0.00026384345,0.0003361856,0.00006571956,0.00018813912,0.00014147241,0.00007041114,0.00010180574,0.00008769651],"category_scores_gemma":[0.06358761,0.00015100886,0.000065911714,0.00028065036,0.00020941977,0.00007032484,0.000042143383,0.00060808007,0.000015186743],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0015478671,0.0015837236,0.3762756,0.0014834026,0.0009618143,0.0011433731,0.016451085,0.00006731654,0.0005614505,0.04510124,0.4914305,0.06339265],"study_design_scores_gemma":[0.016290713,0.05886408,0.08850012,0.43603674,0.0040227165,0.001062374,0.066404134,0.23077025,0.001868764,0.007847358,0.08207418,0.0062585706],"about_ca_topic_score_codex":0.00018054247,"about_ca_topic_score_gemma":0.00014795635,"teacher_disagreement_score":0.43455333,"about_ca_system_score_codex":0.000066494926,"about_ca_system_score_gemma":0.00010051626,"threshold_uncertainty_score":0.9443002},"labels":[],"label_agreement":null},{"id":"W4402582762","doi":"10.1111/jep.14126","title":"Exploring context and culture in clinical reasoning medical education: A qualitative exploratory study","year":2024,"lang":"en","type":"article","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University; NOSM University","funders":"Northern Ontario Academic Medicine Association","keywords":"Context (archaeology); Reflexivity; Thematic analysis; Psychology; Exploratory research; Qualitative research; Pedagogy; Sociology; Social science","score_opus":0.36911512987262884,"score_gpt":0.616729820500783,"score_spread":0.2476146906281541,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402582762","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96295863,0.009761828,0.0001343929,0.02185579,0.0034160328,0.00066941255,9.2867066e-7,0.00001860727,0.0011843913],"genre_scores_gemma":[0.9867831,0.0056118504,0.0025981423,0.0031118654,0.0017182333,0.0000706537,0.000003763045,0.000022793205,0.000079606754],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9859845,0.0067957444,0.0045039584,0.00049369305,0.0020083012,0.00021377305],"domain_scores_gemma":[0.6711015,0.32019886,0.0024936928,0.000696212,0.0039675017,0.0015422326],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.090867326,0.0001809147,0.0009032045,0.00030473538,0.00004012459,0.00008781219,0.000120929624,0.00029144133,0.00020515393],"category_scores_gemma":[0.8171123,0.00013880436,0.0002450497,0.0006290225,0.000195448,0.0011286674,0.00007191252,0.002992724,0.00003227268],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0033006123,0.013843373,0.19683787,0.00009609198,0.0005213155,0.0019404893,0.058477294,0.00001983893,0.0000010694716,0.0012730374,0.013579482,0.71010953],"study_design_scores_gemma":[0.023774112,0.0114393085,0.4701065,0.03173518,0.0024460861,0.0017049855,0.4185902,0.0078917565,0.0000023053983,0.005069723,0.026702516,0.00053733645],"about_ca_topic_score_codex":0.000045779267,"about_ca_topic_score_gemma":0.000118153446,"teacher_disagreement_score":0.726245,"about_ca_system_score_codex":0.00020130219,"about_ca_system_score_gemma":0.0044900444,"threshold_uncertainty_score":0.9993074},"labels":[],"label_agreement":null},{"id":"W4402732704","doi":"10.1001/jamainternmed.2024.4748","title":"Direct-to-Consumer Educational Brochures to Promote Gabapentinoid Deprescribing in Older Adults","year":2024,"lang":"en","type":"letter","venue":"JAMA Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University Health Centre; Université de Sherbrooke","funders":"","keywords":"Deprescribing; Medicine; Tapering; Polypharmacy; Pharmacology","score_opus":0.017359013901129147,"score_gpt":0.3279668792661606,"score_spread":0.31060786536503143,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402732704","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.092535615,0.0024629582,0.000038998653,0.8800261,0.004655799,0.0014305133,0.000052102783,0.00011099749,0.018686898],"genre_scores_gemma":[0.115236446,0.00011821976,0.0006462662,0.818987,0.022130964,0.0003848835,0.00050045236,0.00016921188,0.041826572],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9954719,0.00009247329,0.0012177906,0.001267386,0.001192255,0.0007581878],"domain_scores_gemma":[0.9917257,0.006435835,0.00018152206,0.00069353584,0.00032499814,0.00063840335],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0005841412,0.0006015105,0.0012506361,0.001010484,0.000034646357,0.000079518395,0.00045036434,0.00069910695,0.003086068],"category_scores_gemma":[0.06667446,0.00045104363,0.00025519019,0.00058033946,0.000121472396,0.000055817425,0.00026260287,0.003164806,0.002250125],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00045412607,0.00018247026,0.006921818,0.00049800065,0.00016832667,0.0026653784,0.0012231378,0.0000011876363,0.000082115126,0.000005614545,0.96057767,0.02722017],"study_design_scores_gemma":[0.003681522,0.00082803593,0.07158734,0.19274336,0.00046351613,0.0005761491,0.00008363165,0.000036755293,0.00030003843,0.00009410191,0.7290452,0.000560355],"about_ca_topic_score_codex":0.0022783703,"about_ca_topic_score_gemma":0.000116164745,"teacher_disagreement_score":0.23153245,"about_ca_system_score_codex":0.0003139024,"about_ca_system_score_gemma":0.00029568817,"threshold_uncertainty_score":0.9997941},"labels":[],"label_agreement":null},{"id":"W4402786428","doi":"10.2196/58426","title":"Artificial Intelligence–Powered Training Database for Clinical Thinking: App Development Study","year":2024,"lang":"en","type":"article","venue":"JMIR Formative Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Preprint; Training (meteorology); Artificial intelligence; Computer science; Training set; Psychology; Database; Data science; World Wide Web; Geography","score_opus":0.44141795971950354,"score_gpt":0.5880468305314368,"score_spread":0.14662887081193327,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402786428","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9528065,0.0007536912,0.030839333,0.0029092382,0.0017228312,0.0054513556,0.000084452935,0.0004430569,0.004989578],"genre_scores_gemma":[0.9933176,0.000079761296,0.0043269084,0.00012266406,0.0006056817,0.0007439258,0.00017918358,0.00003710602,0.0005871942],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9960503,0.0003214431,0.0011455035,0.0005557184,0.0012219169,0.00070514844],"domain_scores_gemma":[0.97055906,0.028132329,0.000059002065,0.00039666408,0.0004338365,0.00041909254],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.011571878,0.0001875132,0.00048291343,0.0003527656,0.0003520239,0.00020025179,0.0002712195,0.00014623621,0.00021736274],"category_scores_gemma":[0.02651481,0.0001385998,0.00019370287,0.00068889034,0.0002748242,0.0001805836,0.0002508615,0.0012401006,0.0006166502],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009561858,0.0026656091,0.0014504635,0.0002492896,0.00036056404,0.00021452548,0.05439833,0.0000013330558,0.000018453964,0.007845212,0.015601136,0.9162389],"study_design_scores_gemma":[0.008542829,0.048728824,0.11295133,0.03713606,0.00065397855,0.00026697927,0.30620158,0.04484631,0.005975683,0.113184184,0.31848466,0.0030275688],"about_ca_topic_score_codex":0.000009292367,"about_ca_topic_score_gemma":0.000012534975,"teacher_disagreement_score":0.91321135,"about_ca_system_score_codex":0.00013512169,"about_ca_system_score_gemma":0.0011958019,"threshold_uncertainty_score":0.9816853},"labels":[],"label_agreement":null},{"id":"W4402789086","doi":"10.53300/001c.123911","title":"The Education and Implementation of Clinical Reasoning Education in Post-Graduate Emergency Medicine Training: An Integrative Review","year":2024,"lang":"en","type":"article","venue":"Australian Journal of Clinical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Training (meteorology); Medical education; Graduate education; Psychology; Medicine","score_opus":0.19015460031633674,"score_gpt":0.5971728784596656,"score_spread":0.40701827814332886,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402789086","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9306481,0.018846517,0.000022128379,0.04058321,0.008503023,0.0005697509,0.0000039196643,0.0000118856915,0.00081148924],"genre_scores_gemma":[0.92117053,0.071713045,0.0014155037,0.0010263487,0.0027975007,0.000026470598,0.00009572318,0.000023894332,0.0017309674],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99343354,0.00077696086,0.0047737854,0.00034235997,0.00046707952,0.00020630128],"domain_scores_gemma":[0.98902494,0.0065537863,0.0019126935,0.000352895,0.0015888788,0.0005668243],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.009447712,0.00019864533,0.0009286245,0.00020966631,0.000060346527,0.000025349804,0.00017793299,0.00016775826,0.00031616175],"category_scores_gemma":[0.07201685,0.00012339037,0.0003254108,0.0004903695,0.00028515526,0.00033591586,0.000016525286,0.001024417,0.000005786172],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012948285,0.0008978244,0.070992164,0.00024274181,0.000091410686,0.0000018556132,0.001592378,1.9647494e-7,0.000013596958,0.001308252,0.04054406,0.884186],"study_design_scores_gemma":[0.0008742659,0.0029359069,0.9092029,0.038918704,0.00079692475,0.0001760342,0.029848566,0.000018559198,0.000020571664,0.0033895883,0.013667154,0.00015086275],"about_ca_topic_score_codex":0.00038155916,"about_ca_topic_score_gemma":0.00013542782,"teacher_disagreement_score":0.88403517,"about_ca_system_score_codex":0.000060069062,"about_ca_system_score_gemma":0.011149505,"threshold_uncertainty_score":0.99445635},"labels":[],"label_agreement":null},{"id":"W4402840152","doi":"10.1155/2024/8385716","title":"The Place of Intuition in the Clinical Reasoning of Occupational Therapists: A Multiple‐Case Study","year":2024,"lang":"en","type":"article","venue":"Health & Social Care in the Community","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"GDG Environnement; Université de Montréal; Centre for Interdisciplinary Research in Rehabilitation","funders":"","keywords":"Intuition; Occupational therapy; Psychology; Psychotherapist; Medicine; Clinical psychology; Psychiatry; Cognitive science","score_opus":0.1768799621607277,"score_gpt":0.5268684546333243,"score_spread":0.3499884924725966,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402840152","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99157035,0.0024012122,0.000013521687,0.004584674,0.00018921882,0.0009571952,0.000033490738,0.00001564071,0.0002347221],"genre_scores_gemma":[0.99811,0.00029528933,0.00004081658,0.0012667495,0.00016624996,0.00007300308,0.000035961595,0.000009344407,0.0000025912777],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.9897044,0.008710643,0.00088191644,0.00010913779,0.0004084347,0.00018549057],"domain_scores_gemma":[0.9045661,0.09454984,0.00020616877,0.00048681808,0.0001509497,0.000040101993],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.015976664,0.00009873237,0.00034658724,0.000044564047,0.0005041987,0.000023953062,0.00032645618,0.00010110156,0.0000051261077],"category_scores_gemma":[0.030599998,0.000051130988,0.00014563468,0.00045246736,0.0003061889,0.000023570126,0.000077626304,0.0017937052,0.0000012624812],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013875219,0.002881171,0.43063352,0.00034256445,0.00006808141,0.00019874956,0.45727795,0.000009324564,5.3353585e-7,0.0022103016,0.0016893629,0.103300914],"study_design_scores_gemma":[0.0016259471,0.00149331,0.6034262,0.0013615303,0.000037181882,0.00008040017,0.39076218,0.00016968332,6.085128e-7,0.0007156479,0.00027439202,0.000052890122],"about_ca_topic_score_codex":0.0147299385,"about_ca_topic_score_gemma":0.026344908,"teacher_disagreement_score":0.1727927,"about_ca_system_score_codex":0.00015654764,"about_ca_system_score_gemma":0.00073964114,"threshold_uncertainty_score":0.99183106},"labels":[],"label_agreement":null},{"id":"W4402854766","doi":"","title":"Exploring strategies to improve clinical decision making in a chiropractic office: a case series.","year":2024,"lang":"en","type":"article","venue":"PubMed","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Disability Prevention and Rehabilitation; University of Toronto; Université du Québec à Trois-Rivières; Toronto and Region Conservation Authority; Canadian Memorial Chiropractic College","funders":"","keywords":"Chiropractic; Computer science; Data science; Series (stratigraphy); Alternative medicine; Bioinformatics; Medicine; Pathology; Biology","score_opus":0.15967455844536113,"score_gpt":0.41010122735311544,"score_spread":0.25042666890775434,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402854766","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99218595,0.00048193286,0.00044052055,0.0015257539,0.0018461886,0.00083288125,0.0000052442956,0.00018219321,0.0024993336],"genre_scores_gemma":[0.995115,0.00020262922,0.0020921552,0.00044610616,0.00061266887,0.0012836265,0.000002749581,0.000033363493,0.00021172638],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9980714,0.000048985657,0.0006669711,0.0005152103,0.00025445697,0.00044299406],"domain_scores_gemma":[0.9892452,0.009972819,0.000047491336,0.0003377207,0.000036453097,0.0003603257],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0015540275,0.00015392686,0.00037502023,0.00018937758,0.0000354782,0.00017111666,0.00006853678,0.00010693864,0.000027996844],"category_scores_gemma":[0.06496848,0.00012524186,0.00014035143,0.00050741405,0.000042054195,0.00036712148,0.00009318062,0.0005374315,0.00011122094],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007465585,0.00018894268,0.0031920688,0.000048901,0.00003850403,0.020574179,0.00026399276,0.000016975126,0.00000374239,0.00034142908,0.00043084458,0.9741539],"study_design_scores_gemma":[0.002856218,0.0007273901,0.958665,0.005458355,0.00031233215,0.009850883,0.0022602403,0.0007576878,0.00004428012,0.001999656,0.016539283,0.0005286521],"about_ca_topic_score_codex":0.000078944766,"about_ca_topic_score_gemma":0.000121343524,"teacher_disagreement_score":0.9736252,"about_ca_system_score_codex":0.00010228293,"about_ca_system_score_gemma":0.00017756772,"threshold_uncertainty_score":0.9429077},"labels":[],"label_agreement":null},{"id":"W4403184224","doi":"10.1007/s11606-024-09087-w","title":"To Err Is Human: A Quarter Century of Progress","year":2024,"lang":"en","type":"editorial","venue":"Journal of General Internal Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Quarter (Canadian coin); MEDLINE; Gerontology","score_opus":0.020172620582687796,"score_gpt":0.39927746921921264,"score_spread":0.3791048486365248,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403184224","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.025117846,0.014296693,0.00006132631,0.014865359,0.94211173,0.00033556562,0.000093346294,0.00002442446,0.0030936997],"genre_scores_gemma":[0.012734505,0.0012853179,0.0008509229,0.0014009903,0.96830493,0.000008036724,0.000054320044,0.000091816226,0.015269136],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99447167,0.000080530524,0.0023473883,0.0003897199,0.002336222,0.000374456],"domain_scores_gemma":[0.9928293,0.0030159832,0.0012285901,0.00042456752,0.0016791678,0.00082233927],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0011738249,0.00044841314,0.0020906301,0.00069549866,0.000024807521,0.000029269433,0.00047621774,0.00063443335,0.00081415474],"category_scores_gemma":[0.01964714,0.0002765734,0.000677877,0.00028225253,0.00027201616,0.0000454621,0.00012997272,0.002265525,0.00006153822],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010752256,0.00029428516,0.0005701148,0.00041562633,0.00084039307,0.001646052,0.0007332239,5.8789806e-7,0.00056162564,0.000023041544,0.9880619,0.005777894],"study_design_scores_gemma":[0.003454379,0.012843182,0.0005833574,0.06838324,0.0017952358,0.00043285638,0.00012340207,0.000015064363,0.00041269744,0.00023421591,0.9114835,0.00023886967],"about_ca_topic_score_codex":0.0002610806,"about_ca_topic_score_gemma":0.000008247421,"teacher_disagreement_score":0.07657843,"about_ca_system_score_codex":0.00020957821,"about_ca_system_score_gemma":0.00050089415,"threshold_uncertainty_score":0.99996865},"labels":[],"label_agreement":null},{"id":"W4403271274","doi":"10.1017/cjn.2024.296","title":"Avoiding Cognitive Bias in Radiology: New Brain Lesions in Homeopathically Treated Breast Cancer Patient","year":2024,"lang":"en","type":"article","venue":"Canadian Journal of Neurological Sciences / Journal Canadien des Sciences Neurologiques","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Sunnybrook Health Science Centre; University of Toronto; Health Sciences Centre; University Health Network","funders":"","keywords":"Medicine; Breast cancer; Cognition; Radiology; Cancer; Medical physics; Oncology; Internal medicine; Psychiatry","score_opus":0.06799239475659959,"score_gpt":0.34196954344074443,"score_spread":0.27397714868414486,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403271274","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9621561,0.005602782,0.000012065336,0.030614423,0.0007876492,0.00020260329,0.00002713129,0.000022227154,0.0005749757],"genre_scores_gemma":[0.99121255,0.0016320403,0.00021785217,0.0065954383,0.00029262374,0.0000046378423,6.665965e-7,0.000013587381,0.000030627278],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99462163,0.0015842811,0.0011888021,0.0006578025,0.00062797364,0.0013195068],"domain_scores_gemma":[0.98391366,0.012798384,0.00036014005,0.000094913616,0.0002749517,0.0025579666],"candidate_categories":["metaresearch","sts","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.006016296,0.00032323587,0.00072772516,0.0018104939,0.00088135543,0.00044294,0.0006840304,0.00030424804,0.000283005],"category_scores_gemma":[0.04887502,0.0002069396,0.00026249787,0.0027222903,0.0037343286,0.0004925214,0.00005056592,0.003068256,0.0000054893967],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021739211,0.00006231396,0.8913414,0.000007324331,0.000014414674,0.041760847,0.0007653176,0.001516736,0.00002514975,0.00024720246,0.00088162237,0.0631603],"study_design_scores_gemma":[0.0008792774,0.035619583,0.86977035,0.0024746482,0.0000659625,0.080621324,0.0005449973,0.002874338,0.000031626467,0.0062913145,0.00050227117,0.000324338],"about_ca_topic_score_codex":0.009621594,"about_ca_topic_score_gemma":0.12760511,"teacher_disagreement_score":0.11798352,"about_ca_system_score_codex":0.0004517994,"about_ca_system_score_gemma":0.008813115,"threshold_uncertainty_score":0.9992317},"labels":[],"label_agreement":null},{"id":"W4403307412","doi":"10.2196/58666","title":"Facilitating Trust Calibration in Artificial Intelligence–Driven Diagnostic Decision Support Systems for Determining Physicians’ Diagnostic Accuracy: Quasi-Experimental Study","year":2024,"lang":"en","type":"article","venue":"JMIR Formative Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Japan Society for the Promotion of Science","keywords":"Preprint; Medical diagnosis; Calibration; Diagnostic accuracy; Computer science; Artificial intelligence; Medical physics; Telepathology; Data mining; Machine learning; Medicine; Statistics; World Wide Web; Pathology; Mathematics; Radiology; Telemedicine","score_opus":0.12644244543756114,"score_gpt":0.49338956533156525,"score_spread":0.3669471198940041,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403307412","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9791453,0.0005560119,0.014444583,0.00024011193,0.00075629965,0.004491973,0.00009820798,0.00011218315,0.00015535581],"genre_scores_gemma":[0.99588346,0.00007870244,0.000322834,0.00003353714,0.0003663335,0.0030612354,0.00015350812,0.00004710794,0.000053295113],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9958466,0.00044213413,0.0011746158,0.0006238016,0.0011421769,0.0007706376],"domain_scores_gemma":[0.8230491,0.17609636,0.00009206604,0.00030227003,0.0002518307,0.00020836128],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0027192098,0.0002726756,0.0005952222,0.00059321,0.00027310284,0.0003676168,0.00022462473,0.00015561482,0.00007629798],"category_scores_gemma":[0.12495479,0.00022976409,0.00017778062,0.00096268835,0.00019857418,0.0005399731,0.000165253,0.0007766318,0.00020800547],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0015476121,0.020120608,0.03841735,0.0013072031,0.00028200887,0.0013478665,0.18310331,0.0023727848,0.0005916609,0.0030885527,0.005810861,0.7420102],"study_design_scores_gemma":[0.0020192154,0.027416734,0.019940805,0.015398562,0.00008815136,0.000049864757,0.17701401,0.7526324,0.0024606434,0.0020130693,0.0003481285,0.00061843],"about_ca_topic_score_codex":0.00014513315,"about_ca_topic_score_gemma":0.000045666107,"teacher_disagreement_score":0.7502596,"about_ca_system_score_codex":0.00043006643,"about_ca_system_score_gemma":0.00036779078,"threshold_uncertainty_score":0.93695074},"labels":[],"label_agreement":null},{"id":"W4403323578","doi":"10.1007/978-3-031-70718-6_5","title":"Design","year":2024,"lang":"en","type":"book-chapter","venue":"Studies in systems, decision and control","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Osta Bio Technologies (Canada)","funders":"","keywords":"Computer science","score_opus":0.1007374484721203,"score_gpt":0.39593560384883664,"score_spread":0.29519815537671634,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403323578","genre_codex":"review","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00015147697,0.6270928,0.0027755538,0.0009327341,0.0054572374,0.0021777584,0.000048970705,0.0001472941,0.36121622],"genre_scores_gemma":[0.21284688,0.038061555,0.00024062114,0.00090553414,0.0010845939,0.00014485157,0.000009019624,0.00010042582,0.7466065],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9979591,0.000032137454,0.00083581964,0.0005312224,0.0004312755,0.00021046346],"domain_scores_gemma":[0.9767433,0.022456806,0.0001313226,0.0003388653,0.00018271814,0.0001469967],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.000978625,0.00033043828,0.0014625846,0.00020299073,0.00004703922,0.000038229857,0.00006849852,0.00033866693,0.00004956199],"category_scores_gemma":[0.010902104,0.00022324902,0.0001518133,0.00003699278,0.00019161621,0.000012572197,0.00008740899,0.00043885555,0.0004036833],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004854241,0.00024830044,0.0009647378,0.0023105124,0.0061549814,0.0130749615,0.001845012,0.00011660731,0.0000121856165,0.2039598,0.41456434,0.35189432],"study_design_scores_gemma":[0.015223862,0.0019333232,0.00034497,0.1728294,0.0026934757,0.00086063,0.0013582652,0.0016613987,0.0000012314986,0.1038717,0.69807357,0.0011481703],"about_ca_topic_score_codex":0.000009619195,"about_ca_topic_score_gemma":0.000004403437,"teacher_disagreement_score":0.5890312,"about_ca_system_score_codex":0.00010547779,"about_ca_system_score_gemma":0.000075462354,"threshold_uncertainty_score":0.9974295},"labels":[],"label_agreement":null},{"id":"W4403329619","doi":"10.1016/j.lers.2024.10.001","title":"Exploring a novel approach to assessing surgical team collaboration: Evidence of brain activity synchronization","year":2024,"lang":"en","type":"article","venue":"Laparoscopic Endoscopic and Robotic Surgery","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Synchronization (alternating current); Process management; Business; Knowledge management; Computer science; Psychology; Computer network","score_opus":0.135634453600115,"score_gpt":0.3702471278942359,"score_spread":0.23461267429412092,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403329619","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96442777,0.0031959903,0.027181778,0.002800308,0.0011891904,0.00036048447,0.000002540312,0.00012439025,0.0007175359],"genre_scores_gemma":[0.9943638,0.0005907434,0.0041595004,0.000143501,0.00032896505,0.00007150823,0.000012343624,0.00002355412,0.000306072],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9983987,0.0000821366,0.00043704515,0.00047313224,0.00033624415,0.00027273616],"domain_scores_gemma":[0.9840309,0.015299053,0.00007376695,0.00025750988,0.000095220596,0.00024356568],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00060760556,0.00018437313,0.00065282895,0.00018263659,0.00009498649,0.00010776731,0.000050337658,0.00010580353,0.000022845988],"category_scores_gemma":[0.010403969,0.0001532559,0.00009256055,0.00076606276,0.0001255056,0.00036128736,0.000058557565,0.00020009006,0.0000069996986],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000724098,0.0036209763,0.7471197,0.013085398,0.0008187506,0.00037841356,0.006201197,0.005813853,0.014243659,0.010887681,0.008821597,0.18828465],"study_design_scores_gemma":[0.005604658,0.0017404399,0.70811796,0.18773982,0.0016374289,0.0005865514,0.0027112938,0.061721906,0.02488656,0.00021644915,0.0031666881,0.0018702378],"about_ca_topic_score_codex":0.000051370454,"about_ca_topic_score_gemma":0.0000025888362,"teacher_disagreement_score":0.1864144,"about_ca_system_score_codex":0.00007937087,"about_ca_system_score_gemma":0.00043736162,"threshold_uncertainty_score":0.99793184},"labels":[],"label_agreement":null},{"id":"W4403420321","doi":"10.1111/tme.13104","title":"Pitfalls of reasoning in hospital‐based transfusion medicine","year":2024,"lang":"en","type":"review","venue":"Transfusion Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University; Canadian Blood Services; University of Toronto","funders":"Canadian Blood Services","keywords":"Transfusion medicine; Cognitive bias; Cognition; Satisficing; Psychology; Framing effect; Psychological intervention; Health care; Confirmation bias; Fallacy; Causal inference; Medicine; Social psychology; Blood transfusion; Computer science; Psychiatry; Persuasion; Artificial intelligence; Economics","score_opus":0.0501356519299426,"score_gpt":0.40057466429036104,"score_spread":0.3504390123604184,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403420321","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005324852,0.98378015,0.00044754654,0.0064202663,0.0018088847,0.0026997381,0.000045045148,0.00021185186,0.004054047],"genre_scores_gemma":[0.00045584532,0.9958491,0.00037799936,0.00044762218,0.0011802352,0.0001550935,0.00061756355,0.00021352315,0.000703038],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.992313,0.00036421823,0.0036449449,0.0013613745,0.00161167,0.0007047533],"domain_scores_gemma":[0.9878743,0.009615001,0.00051075453,0.001102757,0.00022848092,0.00066865655],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002564562,0.0011238104,0.0071797916,0.0015675409,0.00005996401,0.000007788025,0.00041629953,0.0010913048,0.0020731487],"category_scores_gemma":[0.016039195,0.00067832705,0.0010444608,0.0024578173,0.00083743664,0.0000615304,0.000041746956,0.0021123367,0.00009985916],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002686506,0.0010763712,0.00017758101,0.049077522,0.0002960654,0.0012465694,0.0012281309,0.0000023003163,0.000044776276,0.00046845782,0.01074014,0.9353734],"study_design_scores_gemma":[0.005659622,0.004153832,0.00015577543,0.65014917,0.008988058,0.00012157448,0.00015865374,0.00009442093,0.0000074989025,0.00012238801,0.32985383,0.0005351865],"about_ca_topic_score_codex":0.001943675,"about_ca_topic_score_gemma":0.00016826179,"teacher_disagreement_score":0.93483824,"about_ca_system_score_codex":0.0002366618,"about_ca_system_score_gemma":0.0008645023,"threshold_uncertainty_score":0.9995668},"labels":[],"label_agreement":null},{"id":"W4403601070","doi":"10.1111/acem.14968","title":"Diagnostic reasoning and cognitive error in emergency medicine: Implications for teaching and learning","year":2024,"lang":"en","type":"review","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Interpretation (philosophy); Cognition; Process (computing); Experiential learning; Dual process theory (moral psychology); Medicine; Dual (grammatical number); Medical diagnosis; Thinking processes; Cognitive psychology; Cognitive science; Epistemology; Mathematics education; Computer science; Psychology; Pathology; Psychiatry","score_opus":0.10669703520297408,"score_gpt":0.48805845261250563,"score_spread":0.38136141740953156,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403601070","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0016011471,0.9862339,0.00019738125,0.0046687857,0.0019530618,0.0026267234,0.000056484245,0.00016729403,0.0024952204],"genre_scores_gemma":[0.0015853386,0.98888755,0.0001223088,0.0001363235,0.0030054345,0.0011186878,0.0007230471,0.00018771939,0.0042336215],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9934789,0.0004922599,0.003425261,0.0014368197,0.000394948,0.0007718342],"domain_scores_gemma":[0.9634842,0.034287006,0.00093933014,0.00037350372,0.00020102473,0.000714881],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0043257317,0.0008998511,0.003234971,0.00077466725,0.00032235685,0.0000017550813,0.00023032693,0.0010100822,0.00042254926],"category_scores_gemma":[0.39945745,0.0006320629,0.00030716907,0.00087664434,0.00050304405,0.00009205132,0.00021046499,0.0059816916,0.000025066716],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004902588,0.000100691475,0.029198062,0.027906826,0.0007051187,0.000034176406,0.0035250243,3.4085988e-7,0.0000046705077,0.0025655048,0.05043464,0.88547593],"study_design_scores_gemma":[0.0018755657,0.0011906569,0.0077799456,0.39433306,0.018198274,0.00032104805,0.001542711,0.00009759067,6.322875e-8,0.0025413886,0.5713144,0.0008052775],"about_ca_topic_score_codex":0.00031841226,"about_ca_topic_score_gemma":0.000024231076,"teacher_disagreement_score":0.8846706,"about_ca_system_score_codex":0.00010815546,"about_ca_system_score_gemma":0.00023305621,"threshold_uncertainty_score":0.99961305},"labels":[],"label_agreement":null},{"id":"W4403645728","doi":"10.5334/pme.1394","title":"How Do Paediatricians Manage Comfort with Uncertainty in Clinical Decision-Making","year":2024,"lang":"en","type":"article","venue":"Perspectives on Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Clinical decision making; Medical education; Computer science; Medicine; Psychology; Family medicine","score_opus":0.014230183272292307,"score_gpt":0.39829212202124015,"score_spread":0.38406193874894784,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403645728","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9361853,0.005174009,0.0032209947,0.040898476,0.0022131733,0.0006015446,0.0000044150497,0.00020141856,0.011500641],"genre_scores_gemma":[0.99287987,0.0014866523,0.0013946737,0.0020665664,0.0016494956,0.00005120857,0.00001731011,0.000028881945,0.00042536788],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99766546,0.0001036019,0.00042571608,0.0006562462,0.0008780505,0.00027094595],"domain_scores_gemma":[0.98234624,0.016545357,0.0000764629,0.00040057042,0.0001376865,0.0004936932],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0012271327,0.0001754812,0.00037906194,0.00029920656,0.000047351372,0.00010651714,0.00013304669,0.00023374156,0.0003236151],"category_scores_gemma":[0.09954538,0.00011996774,0.00015607926,0.00069846865,0.00023963884,0.00007583177,0.00003012279,0.00092602865,0.00005962127],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005937259,0.002794954,0.054071568,0.000062068284,0.00008016022,0.00038884504,0.003235845,0.000024632793,2.3873656e-7,0.011695586,0.011318717,0.91573364],"study_design_scores_gemma":[0.0054853815,0.0042072115,0.79307383,0.05280464,0.0006140735,0.00040293176,0.074327506,0.012656411,0.0000018951708,0.016932191,0.038519606,0.0009743316],"about_ca_topic_score_codex":0.00005535939,"about_ca_topic_score_gemma":0.000053338375,"teacher_disagreement_score":0.91475934,"about_ca_system_score_codex":0.00030821442,"about_ca_system_score_gemma":0.0015861721,"threshold_uncertainty_score":0.9080395},"labels":[],"label_agreement":null},{"id":"W4403731047","doi":"10.1016/j.revmed.2024.10.009","title":"Le raisonnement clinique : que nous apprend la théorie des scripts ?","year":2024,"lang":"fr","type":"letter","venue":"La Revue de Médecine Interne","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Centre Hospitalier Universitaire Sainte-Justine","funders":"","keywords":"Philosophy; Humanities","score_opus":0.02131286078126664,"score_gpt":0.31436146665959797,"score_spread":0.2930486058783313,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403731047","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.060773686,0.02100995,0.010102105,0.7272897,0.013531236,0.0016468047,0.00046972831,0.00064733386,0.16452949],"genre_scores_gemma":[0.73652774,0.0051165023,0.0038869968,0.103574336,0.03234343,0.0004751887,0.0010060788,0.0005702193,0.11649948],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99358195,0.00082844496,0.0019708942,0.0015938106,0.0005714132,0.0014535086],"domain_scores_gemma":[0.96756154,0.029647673,0.00045489494,0.0012985786,0.0002675376,0.0007697575],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity","insufficient_payload"],"category_scores_codex":[0.0022142266,0.0010788995,0.0020273614,0.00035592818,0.000115575094,0.0003156741,0.00079859246,0.0021095946,0.0016786589],"category_scores_gemma":[0.037950557,0.001017747,0.0011069028,0.00044031328,0.0009134342,0.0000889672,0.0005807979,0.005087575,0.0023661263],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00042075888,0.0022947155,0.044323314,0.003991859,0.0014931401,0.030564142,0.0014244036,0.00004562557,0.00016757532,0.047480147,0.82528883,0.042505503],"study_design_scores_gemma":[0.004770236,0.0012920186,0.012404584,0.09242085,0.0018021803,0.0037722145,0.00027136534,0.00070500554,0.00030530637,0.016022548,0.86517495,0.0010587127],"about_ca_topic_score_codex":0.0015624105,"about_ca_topic_score_gemma":0.0003901074,"teacher_disagreement_score":0.6757541,"about_ca_system_score_codex":0.0005874509,"about_ca_system_score_gemma":0.001338847,"threshold_uncertainty_score":0.99923396},"labels":[],"label_agreement":null},{"id":"W4403788255","doi":"10.2196/55709","title":"Impact of Clinical Decision Support Systems on Medical Students’ Case-Solving Performance: Comparison Study with a Focus Group","year":2024,"lang":"en","type":"article","venue":"JMIR Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Preprint; Psychology; Computer science; Medical education; Medicine; World Wide Web","score_opus":0.03915948217529472,"score_gpt":0.4977967545484347,"score_spread":0.45863727237314,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403788255","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9944347,0.0004118864,0.00024337503,0.00060782785,0.002201809,0.0013370216,0.000004313098,0.00012332582,0.00063571637],"genre_scores_gemma":[0.99772,0.00012851931,0.0001020435,0.00022415545,0.0013326512,0.00023266942,0.00006358957,0.000042252508,0.00015411296],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9928955,0.0002581067,0.0017580882,0.0006396925,0.004095324,0.00035331867],"domain_scores_gemma":[0.9858307,0.011631065,0.00024929928,0.000567903,0.00024821973,0.001472764],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0048931357,0.00027413573,0.0010117764,0.00024040006,0.00007660071,0.00008485484,0.00032200365,0.00044141343,0.00154349],"category_scores_gemma":[0.036610417,0.00016451746,0.0002724709,0.00053137465,0.00021892617,0.000088411376,0.000103404986,0.0012283415,0.00013478695],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010484139,0.012891847,0.66862667,0.0001248548,0.00023879738,0.0011757604,0.0007549184,0.0000041335024,2.4750085e-7,0.000037966867,0.01634515,0.29875124],"study_design_scores_gemma":[0.005575975,0.02589288,0.92867076,0.018583797,0.00041589345,0.0031676083,0.0022664045,0.014016282,0.000001213143,0.000009607844,0.0011445198,0.00025507333],"about_ca_topic_score_codex":0.0006656807,"about_ca_topic_score_gemma":0.00007105046,"teacher_disagreement_score":0.29849616,"about_ca_system_score_codex":0.00019804928,"about_ca_system_score_gemma":0.004619459,"threshold_uncertainty_score":0.99936926},"labels":[],"label_agreement":null},{"id":"W4403967626","doi":"10.1016/b978-0-443-13552-1.00029-1","title":"Clinical reasoning in psychology: from assessment to intervention","year":2024,"lang":"en","type":"book-chapter","venue":"Elsevier eBooks","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; University of Calgary","funders":"","keywords":"Intervention (counseling); Psychology; Psychotherapist; Applied psychology; Clinical psychology; Cognitive psychology; Cognitive science; Psychiatry","score_opus":0.0467551743237819,"score_gpt":0.4354114017877657,"score_spread":0.3886562274639838,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403967626","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0027845192,0.0021840357,0.00003151239,0.0012163532,0.0023309058,0.0007760262,0.000058574446,0.00012392453,0.99049413],"genre_scores_gemma":[0.0061243684,0.00045148138,0.0042145625,0.0037223152,0.0014166853,0.0000776578,0.00021401011,0.0001328456,0.9836461],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9966964,0.00006554122,0.0014783853,0.001055207,0.00038884842,0.00031560083],"domain_scores_gemma":[0.99556273,0.0029130024,0.00022058783,0.00080193346,0.00008332252,0.00041843424],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012680801,0.0004031172,0.0012012465,0.00026896212,0.000023982206,0.000042678814,0.00018860276,0.0007908906,0.0008648704],"category_scores_gemma":[0.0036551822,0.00035803235,0.0007349281,0.000030230422,0.000114435985,0.000013861755,0.00021350254,0.0018506801,0.0016765763],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001207746,0.00010754973,0.0025677804,0.00004234309,0.00022524665,0.0006521293,0.000049669397,1.6106554e-7,0.0000021593128,0.0026228048,0.0019270384,0.99168235],"study_design_scores_gemma":[0.0017856928,0.00079570594,0.020500075,0.043284997,0.00063226506,0.00002683638,0.000009335717,0.000024145798,0.0000013855073,0.033823516,0.8987252,0.00039084538],"about_ca_topic_score_codex":0.000009158028,"about_ca_topic_score_gemma":0.00009626916,"teacher_disagreement_score":0.9912915,"about_ca_system_score_codex":0.0001655853,"about_ca_system_score_gemma":0.00017154636,"threshold_uncertainty_score":0.99988717},"labels":[],"label_agreement":null},{"id":"W4403967647","doi":"10.1016/b978-0-443-13552-1.00018-7","title":"Intellectual disabilities: clinical reasoning and decision-making","year":2024,"lang":"en","type":"book-chapter","venue":"Elsevier eBooks","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Intellectual disability; Psychology; Management science; Computer science; Cognitive science; Psychiatry; Engineering","score_opus":0.034610678816274124,"score_gpt":0.3646722290155776,"score_spread":0.3300615501993035,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403967647","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0045312527,0.0073821545,0.000025072992,0.00014677805,0.0013275256,0.0005548794,0.000034614088,0.00019938182,0.98579836],"genre_scores_gemma":[0.015393798,0.0015165496,0.002904759,0.0010073844,0.0017792161,0.00002277746,0.000023531251,0.00018570008,0.9771663],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9967547,0.000026563854,0.0012963664,0.0010360709,0.0005111466,0.00037512215],"domain_scores_gemma":[0.9265529,0.07194749,0.00017759007,0.00074642664,0.00013038274,0.00044520947],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.001014192,0.0005329989,0.0013508566,0.00016558985,0.00009493659,0.000101633836,0.00015329661,0.000803043,0.0011536063],"category_scores_gemma":[0.09267767,0.000423481,0.0006622221,0.000022846303,0.000745104,0.000020295474,0.00037135044,0.0016960835,0.0009531722],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018591036,0.000025838852,0.0001965197,0.00013121104,0.00025477802,0.00023255573,0.0005136314,1.13102594e-7,1.1348239e-7,0.008356206,0.0020143746,0.9880887],"study_design_scores_gemma":[0.0006480757,0.00060502905,0.00010028605,0.059743296,0.0010684093,0.00034744214,0.00014719737,0.000101873586,4.870148e-7,0.116714,0.82002157,0.00050234166],"about_ca_topic_score_codex":5.184623e-7,"about_ca_topic_score_gemma":0.000008237415,"teacher_disagreement_score":0.9875864,"about_ca_system_score_codex":0.00010894308,"about_ca_system_score_gemma":0.00030626942,"threshold_uncertainty_score":0.9998247},"labels":[],"label_agreement":null},{"id":"W4403967659","doi":"10.1016/b978-0-443-13552-1.00002-3","title":"Clinical reasoning in psychology: a conceptual model","year":2024,"lang":"en","type":"book-chapter","venue":"Elsevier eBooks","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Psychology; Conceptual model; Cognitive science; Epistemology; Cognitive psychology; Philosophy","score_opus":0.06588539579615463,"score_gpt":0.3992233207150236,"score_spread":0.33333792491886893,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403967659","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0006948476,0.00574844,0.000010905947,0.0006268572,0.0009738462,0.00061491184,0.000041319876,0.00017276272,0.9911161],"genre_scores_gemma":[0.0038805322,0.0013180008,0.0018288719,0.0048524863,0.0010350231,0.00004589197,0.00006970442,0.00017036468,0.9867991],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99655384,0.000039103194,0.0014389746,0.0011025091,0.0004133077,0.00045226896],"domain_scores_gemma":[0.99521965,0.0031504603,0.0002283486,0.0008645986,0.00009993104,0.0004370331],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0010582377,0.00050940184,0.001510643,0.00022749345,0.00003183223,0.000025530799,0.00020753797,0.0013062578,0.00032091973],"category_scores_gemma":[0.0054462394,0.00044041727,0.000722279,0.00002656572,0.0006957816,0.000015302718,0.00016478707,0.0026299509,0.0013187098],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00016962613,0.000068355715,0.00039656728,0.00005155388,0.00018303175,0.0011003388,0.00016666744,0.0000029128619,0.0000010878184,0.030486096,0.0028399087,0.96453387],"study_design_scores_gemma":[0.0026366129,0.00054389203,0.00028041255,0.023124665,0.0007874586,0.00017499871,0.0000144279775,0.0010350198,8.8132833e-7,0.07382619,0.89696103,0.00061438774],"about_ca_topic_score_codex":8.1332314e-7,"about_ca_topic_score_gemma":0.000016054626,"teacher_disagreement_score":0.96391946,"about_ca_system_score_codex":0.00008991524,"about_ca_system_score_gemma":0.0004960643,"threshold_uncertainty_score":0.9999903},"labels":[],"label_agreement":null},{"id":"W4403997665","doi":"10.1016/j.psychres.2024.116255","title":"Exploring the interplay of clinical reasoning and artificial intelligence in psychiatry: Current insights and future directions","year":2024,"lang":"en","type":"review","venue":"Psychiatry Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Mila - Quebec Artificial Intelligence Institute; Centre Hospitalier Universitaire Sainte-Justine","funders":"Fonds de Recherche du Québec - Santé; Institut de Valorisation des Données; Fondation Brain Canada; Canadian Institute for Advanced Research","keywords":"Scrutiny; Realm; Artificial intelligence; Psychology; GRASP; Cognitive science; Computer science; Data science; Management science","score_opus":0.3373176168244787,"score_gpt":0.5595564091629815,"score_spread":0.22223879233850286,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403997665","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0051492164,0.98087734,0.000007148623,0.002470215,0.010069227,0.0009351517,0.000018512832,0.00003547576,0.00043772138],"genre_scores_gemma":[0.00053365645,0.99367243,0.0005264855,0.000019937868,0.004882514,0.000247302,0.00002122004,0.000056525278,0.000039938226],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99530685,0.0008849464,0.0018205136,0.00089648267,0.00063241256,0.0004587885],"domain_scores_gemma":[0.9892434,0.009381904,0.00021887058,0.00065404526,0.00015411162,0.00034767803],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0040154355,0.00035293412,0.0016866211,0.00069765584,0.00016245183,0.00009732509,0.0002693128,0.00031615776,0.000021106427],"category_scores_gemma":[0.010696992,0.00020552406,0.00043237122,0.0016865099,0.00059582555,0.000066146706,0.00035733703,0.0044572693,0.000042271182],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000069881935,0.00038842086,0.0024968188,0.006211063,0.000134669,0.00000843699,0.00028447754,3.4035484e-8,8.79087e-9,0.011172607,0.0010775818,0.97815603],"study_design_scores_gemma":[0.00016073632,0.0004974464,0.0030396823,0.1526805,0.0008630318,0.000100851765,0.0009702914,0.00007092545,4.9619576e-8,0.0050925463,0.83627987,0.00024404701],"about_ca_topic_score_codex":0.00008407996,"about_ca_topic_score_gemma":0.0003363802,"teacher_disagreement_score":0.97791195,"about_ca_system_score_codex":0.000048135527,"about_ca_system_score_gemma":0.001289716,"threshold_uncertainty_score":0.9978395},"labels":[],"label_agreement":null},{"id":"W4404026051","doi":"10.2196/51446","title":"The Potential of Artificial Intelligence Tools for Reducing Uncertainty in Medicine and Directions for Medical Education","year":2024,"lang":"en","type":"article","venue":"JMIR Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"St. Michael's Hospital; Canada Research Chairs; Public Health Ontario; University of Toronto","funders":"Wellcome Trust","keywords":"Certainty; Plan (archaeology); Artificial intelligence; Field (mathematics); Patient care; Computer science; Psychology; Medicine; Nursing","score_opus":0.04061356468283583,"score_gpt":0.44393570877975586,"score_spread":0.40332214409692,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4404026051","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.53871286,0.013580434,0.018261347,0.41111934,0.012816244,0.0043320414,0.000021041029,0.00014342253,0.0010132556],"genre_scores_gemma":[0.9910881,0.0017416965,0.0008994712,0.0014515072,0.0028140827,0.001400486,0.00018709611,0.000022002658,0.000395565],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9981492,0.000056613022,0.0007251659,0.00031936108,0.00054030796,0.00020933889],"domain_scores_gemma":[0.98350096,0.015522329,0.00009321222,0.00020310498,0.0002702524,0.00041015266],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0018190317,0.00011297711,0.00027418943,0.00014814197,0.00010820439,0.000037914855,0.000108168875,0.00022405908,0.00019721927],"category_scores_gemma":[0.1980859,0.00007414153,0.00009169834,0.00034206326,0.00039142166,0.000066428736,0.000022285669,0.00028861154,0.0000026120292],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018029995,0.00059287203,0.0000770111,0.00020417346,0.000022043965,9.69524e-7,0.0006446852,0.000003364899,0.000056171415,0.021418484,0.01653353,0.9602664],"study_design_scores_gemma":[0.003478843,0.004514963,0.029912371,0.08451149,0.0012085825,0.00062686874,0.03171765,0.3209716,0.00076466287,0.25543976,0.265856,0.0009972155],"about_ca_topic_score_codex":0.00028227505,"about_ca_topic_score_gemma":0.00011000134,"teacher_disagreement_score":0.95926917,"about_ca_system_score_codex":0.00009515382,"about_ca_system_score_gemma":0.00539506,"threshold_uncertainty_score":0.9570605},"labels":[],"label_agreement":null},{"id":"W4404093069","doi":"10.1080/0142159x.2024.2425027","title":"Commentary to concerns regarding the assessment of clinical reasoning in medical students","year":2024,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Research Foundation","funders":"","keywords":"Psychology; Medical education; Medicine; Engineering ethics; Engineering","score_opus":0.07281500698948865,"score_gpt":0.5195902160097986,"score_spread":0.44677520902031,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4404093069","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.85267156,0.00055743137,0.00062014064,0.13598092,0.0012273084,0.00041419122,0.0000027952824,0.000083809,0.008441808],"genre_scores_gemma":[0.978694,0.00028149012,0.00040311637,0.018244902,0.0014063497,0.000047313024,0.00001837011,0.000029202689,0.0008752195],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9940449,0.0005346926,0.0012931665,0.00046587552,0.0032957683,0.00036559592],"domain_scores_gemma":[0.986281,0.012001579,0.00008461016,0.00049357815,0.000055261396,0.0010839683],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.012850738,0.00016946644,0.00073788053,0.00008631702,0.000037581423,0.000035937574,0.00057822233,0.00035632474,0.004662776],"category_scores_gemma":[0.099803716,0.000099909135,0.00026257223,0.00037538758,0.00033673205,0.000030231811,0.0004423241,0.0018657427,0.000056333116],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000052081257,0.00071905035,0.8173506,0.000035066532,0.0001486616,0.0008712944,0.0002643588,0.0000012310959,7.4559955e-7,0.0004249593,0.114899695,0.0652322],"study_design_scores_gemma":[0.005615638,0.00091383694,0.81513333,0.025619842,0.000285696,0.00012975345,0.0007155521,0.028542662,0.000007570925,0.00015604355,0.12260055,0.0002795343],"about_ca_topic_score_codex":0.0003238863,"about_ca_topic_score_gemma":0.000072617986,"teacher_disagreement_score":0.12602244,"about_ca_system_score_codex":0.00011374503,"about_ca_system_score_gemma":0.00067021087,"threshold_uncertainty_score":0.9962471},"labels":[],"label_agreement":null},{"id":"W4404162510","doi":"10.1002/ffo2.199","title":"Calibration Feedback With the Practical Scoring Rule Does Not Improve Calibration of Confidence","year":2024,"lang":"en","type":"article","venue":"Futures & Foresight Science","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Defence Research and Development Canada; Université du Québec à Montréal","funders":"","keywords":"Calibration; Computer science; Confidence interval; Artificial intelligence; Statistics; Machine learning; Mathematics","score_opus":0.019307429337904563,"score_gpt":0.32805272709700095,"score_spread":0.30874529775909637,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4404162510","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9318054,0.00049688533,0.030507792,0.032804616,0.0014151991,0.0007787393,0.000027524335,0.00016780062,0.0019960578],"genre_scores_gemma":[0.9961412,0.000034897515,0.0020415792,0.00062655733,0.00045447386,0.000020176527,0.0000048386446,0.000011083318,0.00066517934],"study_design_codex":"bench_or_experimental","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.998102,0.00003396846,0.0002545322,0.00040162474,0.00095608435,0.0002518063],"domain_scores_gemma":[0.99558175,0.0036243182,0.00009233335,0.0003434816,0.00020239892,0.00015571693],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0009255787,0.000119835095,0.0001761216,0.00007873857,0.00020969046,0.00020282004,0.00017520151,0.00006450522,0.00007906881],"category_scores_gemma":[0.009128929,0.00005289452,0.000056530585,0.0005286289,0.00088846456,0.0005031114,0.000058361893,0.00025132,0.000007697248],"study_design_candidate":"bench_or_experimental","study_design_consensus":"bench_or_experimental","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019145571,0.00047993654,0.011658517,0.0005202367,0.00014290467,0.00038045485,0.006185783,0.00043276334,0.55896354,0.37467638,0.01811623,0.026528694],"study_design_scores_gemma":[0.0010167387,0.001366089,0.046236128,0.0030089952,0.0003055403,0.00018278458,0.0012390285,0.07517798,0.8651105,0.004318492,0.0016676522,0.0003700386],"about_ca_topic_score_codex":0.00013687211,"about_ca_topic_score_gemma":0.000071047514,"teacher_disagreement_score":0.3703579,"about_ca_system_score_codex":0.00004695315,"about_ca_system_score_gemma":0.00094769744,"threshold_uncertainty_score":0.99921757},"labels":[],"label_agreement":null},{"id":"W4404200940","doi":"10.1007/978-3-031-71452-8_18","title":"Limitations and Future Directions","year":2024,"lang":"en","type":"book-chapter","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Jewish General Hospital","funders":"","keywords":"Computer science","score_opus":0.054946091441230194,"score_gpt":0.3212307039659588,"score_spread":0.2662846125247286,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4404200940","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000022608436,0.009932212,0.000005036807,0.01365299,0.0007345191,0.00013232841,0.000024620465,0.00016944608,0.97532624],"genre_scores_gemma":[0.0002464942,0.022337181,0.000491234,0.0007713209,0.0016550317,0.0000071569934,0.00008175788,0.000034246477,0.9743756],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.99942595,0.0000014188497,0.00016059412,0.00022835424,0.000109314475,0.00007437407],"domain_scores_gemma":[0.99662036,0.002984297,0.000022904109,0.00016138493,0.000055697292,0.00015533534],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.000042738135,0.00012995562,0.00022849285,0.00008433676,0.000035296056,0.000020947484,0.000015833271,0.00025069105,0.0011980261],"category_scores_gemma":[0.0015034684,0.00009583154,0.000117247466,0.00001984898,0.000059867165,0.000008564677,0.00002358213,0.000344007,0.000741717],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000007836519,0.000020922274,0.000019082518,0.00003288087,0.00015251742,0.00007376402,0.000056601,2.558909e-8,5.1974087e-7,0.7859259,0.06886188,0.14484808],"study_design_scores_gemma":[0.00014595664,0.00009555511,0.00044078933,0.0011197026,0.00052799215,0.000086279266,0.000020890648,0.0000066887624,5.423892e-7,0.029401142,0.9680583,0.00009617367],"about_ca_topic_score_codex":0.000007145238,"about_ca_topic_score_gemma":0.000025888447,"teacher_disagreement_score":0.8991964,"about_ca_system_score_codex":0.000017973429,"about_ca_system_score_gemma":0.000059840117,"threshold_uncertainty_score":0.99971503},"labels":[],"label_agreement":null},{"id":"W4404200976","doi":"10.1007/978-3-031-71452-8_2","title":"Clinical Judgement and Decision-Making","year":2024,"lang":"en","type":"book-chapter","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Jewish General Hospital","funders":"","keywords":"Judgement; Clinical decision making; Psychology; Computer science; Medicine; Epistemology; Philosophy; Intensive care medicine","score_opus":0.04726000258387454,"score_gpt":0.40837853344509145,"score_spread":0.3611185308612169,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4404200976","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005748973,0.003180139,0.000323628,0.00096342154,0.0012309816,0.00026771837,0.000011489498,0.0001326294,0.9933151],"genre_scores_gemma":[0.011427783,0.004133069,0.007860218,0.003709426,0.0012888538,0.0000049944683,0.000022283017,0.00008140392,0.97147197],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9982028,0.0000036252698,0.00073021447,0.0005707136,0.00033635317,0.00015629128],"domain_scores_gemma":[0.98182666,0.017372223,0.00008279541,0.00038343822,0.00007110849,0.0002637849],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00050254737,0.00023685857,0.0006672767,0.00009354332,0.000025180427,0.000037706806,0.00005683845,0.00046481306,0.0037511145],"category_scores_gemma":[0.010219994,0.0001660595,0.00031276085,0.000015772055,0.00013695007,0.0000096556805,0.0001991746,0.000651033,0.0019363171],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008407325,0.000043751035,0.00053579354,0.000046232584,0.00021604924,0.0005811972,0.000007248346,1.17628986e-7,5.502709e-8,0.17275287,0.09140583,0.7343268],"study_design_scores_gemma":[0.0007171994,0.00050116266,0.0014100733,0.025467897,0.0009456175,0.00010669833,0.0000055402156,0.00010480022,1.7777887e-7,0.2742371,0.69624263,0.00026113167],"about_ca_topic_score_codex":0.0000026389437,"about_ca_topic_score_gemma":0.000004225559,"teacher_disagreement_score":0.73406565,"about_ca_system_score_codex":0.00003143321,"about_ca_system_score_gemma":0.0001015039,"threshold_uncertainty_score":0.9988408},"labels":[],"label_agreement":null},{"id":"W4404814639","doi":"10.36834/cmej.77080","title":"Génération d’hypothèses diagnostiques par l’observation engagée des pairs: une étude descriptive quantitative en contexte de simulation clinique","year":2024,"lang":"fr","type":"article","venue":"Canadian Medical Education Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"Francophone University Association; NOSM University; University of Ottawa","funders":"","keywords":"Humanities; Philosophy; Physics","score_opus":0.05285696192317268,"score_gpt":0.3950567536181107,"score_spread":0.342199791694938,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4404814639","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5935426,0.09508023,0.056402113,0.22989617,0.020927947,0.0012134464,0.00014478361,0.00020345951,0.002589247],"genre_scores_gemma":[0.9502367,0.016502416,0.0073026973,0.017418157,0.007043611,0.00008468347,0.00020472791,0.00008373641,0.0011233216],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9944526,0.0018379347,0.0013489896,0.00052661856,0.00089129375,0.0009425743],"domain_scores_gemma":[0.8996856,0.07648708,0.00061113836,0.0005372662,0.002914163,0.019764757],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.006285249,0.0003962309,0.00055496587,0.00050480076,0.00060743606,0.00048008392,0.00025761928,0.00093103177,0.008433199],"category_scores_gemma":[0.62181395,0.000393604,0.00029068495,0.00088732503,0.0008112572,0.0007833467,0.000021145675,0.0024341033,0.00037354606],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001533269,0.0008857396,0.02827624,0.0003752567,0.0005021774,0.0011046251,0.044059735,0.0017730935,0.000035469613,0.036324136,0.08083253,0.80567765],"study_design_scores_gemma":[0.0025292488,0.0026699584,0.44464862,0.0982572,0.0016274003,0.0020283838,0.028776811,0.15643233,0.0002094062,0.09788599,0.16363543,0.0012992109],"about_ca_topic_score_codex":0.035730876,"about_ca_topic_score_gemma":0.061144352,"teacher_disagreement_score":0.80437845,"about_ca_system_score_codex":0.0028985941,"about_ca_system_score_gemma":0.08110626,"threshold_uncertainty_score":0.9998673},"labels":[],"label_agreement":null},{"id":"W4404839993","doi":"10.32920/27926799","title":"Improving diagnostic performance through feedback: the Diagnosis Learning Cycle","year":2024,"lang":"en","type":"preprint","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Learning cycle; Computer science; Psychology; Mathematics education","score_opus":0.027607589411831675,"score_gpt":0.32280554169756787,"score_spread":0.2951979522857362,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4404839993","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9378764,0.0077840616,0.00027526042,0.010998623,0.0025651434,0.0011349551,0.00001771132,0.00075721345,0.038590644],"genre_scores_gemma":[0.9793044,0.010260842,0.0010537963,0.0017280178,0.0014514821,0.00064219406,0.00009526531,0.000115700794,0.0053483043],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9970034,0.00010465658,0.0007236532,0.000955727,0.0006066231,0.00060598314],"domain_scores_gemma":[0.9630292,0.03541323,0.00022130826,0.0009518473,0.00015813988,0.00022625833],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00072917424,0.0005119639,0.0007622498,0.00007682087,0.00024353889,0.00024517035,0.00037427765,0.0005199655,0.0008210613],"category_scores_gemma":[0.08954765,0.00031153992,0.00051466667,0.00026636806,0.00023316072,0.000047539008,0.001947121,0.0037535874,0.0011363307],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020448107,0.0010858332,0.33259508,0.003960223,0.0011942341,0.0007203669,0.003905066,0.01554943,0.000028802839,0.0024413136,0.051065985,0.5872492],"study_design_scores_gemma":[0.008029395,0.004393153,0.41339463,0.15720099,0.014646924,0.0008584084,0.00436279,0.24923882,0.0051956675,0.03449725,0.102228574,0.005953419],"about_ca_topic_score_codex":0.0012883043,"about_ca_topic_score_gemma":0.000019743815,"teacher_disagreement_score":0.5812958,"about_ca_system_score_codex":0.00015393976,"about_ca_system_score_gemma":0.0005238991,"threshold_uncertainty_score":0.99993366},"labels":[],"label_agreement":null},{"id":"W4404839994","doi":"10.32920/27926799.v1","title":"Improving diagnostic performance through feedback: the Diagnosis Learning Cycle","year":2024,"lang":"en","type":"preprint","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Computer science","score_opus":0.027607589411831675,"score_gpt":0.32280554169756787,"score_spread":0.2951979522857362,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4404839994","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9378764,0.0077840616,0.00027526042,0.010998623,0.0025651434,0.0011349551,0.00001771132,0.00075721345,0.038590644],"genre_scores_gemma":[0.9793044,0.010260842,0.0010537963,0.0017280178,0.0014514821,0.00064219406,0.00009526531,0.000115700794,0.0053483043],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9970034,0.00010465658,0.0007236532,0.000955727,0.0006066231,0.00060598314],"domain_scores_gemma":[0.9630292,0.03541323,0.00022130826,0.0009518473,0.00015813988,0.00022625833],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00072917424,0.0005119639,0.0007622498,0.00007682087,0.00024353889,0.00024517035,0.00037427765,0.0005199655,0.0008210613],"category_scores_gemma":[0.08954765,0.00031153992,0.00051466667,0.00026636806,0.00023316072,0.000047539008,0.001947121,0.0037535874,0.0011363307],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020448107,0.0010858332,0.33259508,0.003960223,0.0011942341,0.0007203669,0.003905066,0.01554943,0.000028802839,0.0024413136,0.051065985,0.5872492],"study_design_scores_gemma":[0.008029395,0.004393153,0.41339463,0.15720099,0.014646924,0.0008584084,0.00436279,0.24923882,0.0051956675,0.03449725,0.102228574,0.005953419],"about_ca_topic_score_codex":0.0012883043,"about_ca_topic_score_gemma":0.000019743815,"teacher_disagreement_score":0.5812958,"about_ca_system_score_codex":0.00015393976,"about_ca_system_score_gemma":0.0005238991,"threshold_uncertainty_score":0.99993366},"labels":[],"label_agreement":null},{"id":"W4404840600","doi":"10.32920/27931722.v1","title":"ReThinking clinical reasoning: A paradigm shift","year":2024,"lang":"en","type":"preprint","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University; McMaster University","funders":"","keywords":"Paradigm shift; Computer science; Cognitive science; Epistemology; Psychology; Philosophy","score_opus":0.08533649227511052,"score_gpt":0.4162741384911465,"score_spread":0.33093764621603594,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4404840600","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.61874264,0.009755141,0.0014758091,0.08617194,0.0141297765,0.0019071779,0.000064290994,0.0029245643,0.26482862],"genre_scores_gemma":[0.975362,0.0010695562,0.008227257,0.0046259295,0.0033280004,0.000067636545,0.00021293896,0.000119889606,0.006986815],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99592507,0.00016325449,0.0014338897,0.0013615767,0.00062495086,0.00049125054],"domain_scores_gemma":[0.98189265,0.015593631,0.00025217034,0.0014542756,0.0000638474,0.0007434372],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0023924343,0.00046424064,0.0013802177,0.00014586172,0.00005378153,0.0001694004,0.0002839377,0.0014856721,0.0008030107],"category_scores_gemma":[0.06708722,0.00035343732,0.0011446076,0.00017771633,0.00023326905,0.0000136983535,0.0015913164,0.004986196,0.0016984433],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011119996,0.0017955558,0.16556819,0.0021480003,0.0033245648,0.008187608,0.0029870176,0.00004623929,0.0000025166528,0.3866196,0.34360817,0.08460055],"study_design_scores_gemma":[0.0015600299,0.00054757105,0.047475304,0.03848424,0.0020167879,0.00016863365,0.000024079878,0.002547894,0.000020387944,0.89238185,0.0140401255,0.0007330868],"about_ca_topic_score_codex":0.00024423262,"about_ca_topic_score_gemma":0.000026482758,"teacher_disagreement_score":0.5057623,"about_ca_system_score_codex":0.00010326239,"about_ca_system_score_gemma":0.0012116614,"threshold_uncertainty_score":0.99989176},"labels":[],"label_agreement":null},{"id":"W4404840923","doi":"10.32920/27931722","title":"ReThinking clinical reasoning: A paradigm shift","year":2024,"lang":"en","type":"preprint","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University; McMaster University","funders":"","keywords":"Paradigm shift; Epistemology; Computer science; Cognitive science; Psychology; Philosophy","score_opus":0.08533649227511052,"score_gpt":0.4162741384911465,"score_spread":0.33093764621603594,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4404840923","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.61874264,0.009755141,0.0014758091,0.08617194,0.0141297765,0.0019071779,0.000064290994,0.0029245643,0.26482862],"genre_scores_gemma":[0.975362,0.0010695562,0.008227257,0.0046259295,0.0033280004,0.000067636545,0.00021293896,0.000119889606,0.006986815],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99592507,0.00016325449,0.0014338897,0.0013615767,0.00062495086,0.00049125054],"domain_scores_gemma":[0.98189265,0.015593631,0.00025217034,0.0014542756,0.0000638474,0.0007434372],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity"],"category_scores_codex":[0.0023924343,0.00046424064,0.0013802177,0.00014586172,0.00005378153,0.0001694004,0.0002839377,0.0014856721,0.0008030107],"category_scores_gemma":[0.06708722,0.00035343732,0.0011446076,0.00017771633,0.00023326905,0.0000136983535,0.0015913164,0.004986196,0.0016984433],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011119996,0.0017955558,0.16556819,0.0021480003,0.0033245648,0.008187608,0.0029870176,0.00004623929,0.0000025166528,0.3866196,0.34360817,0.08460055],"study_design_scores_gemma":[0.0015600299,0.00054757105,0.047475304,0.03848424,0.0020167879,0.00016863365,0.000024079878,0.002547894,0.000020387944,0.89238185,0.0140401255,0.0007330868],"about_ca_topic_score_codex":0.00024423262,"about_ca_topic_score_gemma":0.000026482758,"teacher_disagreement_score":0.5057623,"about_ca_system_score_codex":0.00010326239,"about_ca_system_score_gemma":0.0012116614,"threshold_uncertainty_score":0.99989176},"labels":[],"label_agreement":null},{"id":"W4405073865","doi":"10.46747/cfp.701112701","title":"How to decrease routine and repetitive blood tests in hospitalized patients","year":2024,"lang":"en","type":"article","venue":"Canadian Family Physician","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Sunnybrook Health Science Centre","funders":"","keywords":"Medicine; Data science; Computer science","score_opus":0.011291705336750927,"score_gpt":0.2645602700351001,"score_spread":0.25326856469834913,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4405073865","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9778076,0.0006893366,0.0000021266997,0.0042703566,0.00020310903,0.00041152388,0.000118497,0.000047213463,0.016450278],"genre_scores_gemma":[0.9539797,0.00003173592,0.00012576903,0.04502218,0.00017367557,0.00003291295,0.0000889069,0.000029872368,0.0005152532],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9988659,0.000024021183,0.00016527002,0.00041607465,0.00015061143,0.00037810148],"domain_scores_gemma":[0.9972189,0.0007741703,0.000019006564,0.00022713433,0.000077683726,0.0016831137],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00009646861,0.00015941724,0.00028010397,0.00022204407,0.000035827557,0.00009237048,0.00005524135,0.00007519582,0.000003772384],"category_scores_gemma":[0.008112552,0.00014904047,0.00006141625,0.00042815233,0.000053786534,0.000061749844,0.000021126743,0.0001731526,0.000044712822],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00024740968,0.0013286516,0.47242972,0.00025519973,0.00038557156,0.006476532,0.0015206115,0.0000045580177,0.0008618476,0.0055493796,0.07745313,0.4334874],"study_design_scores_gemma":[0.001808153,0.0005233752,0.97610986,0.003233826,0.00013657763,0.0000020460836,0.00017724706,0.00005028642,0.000087286484,0.0003690942,0.017252378,0.00024987356],"about_ca_topic_score_codex":0.01311492,"about_ca_topic_score_gemma":0.0074553667,"teacher_disagreement_score":0.5036801,"about_ca_system_score_codex":0.0001434971,"about_ca_system_score_gemma":0.000601906,"threshold_uncertainty_score":0.99345684},"labels":[],"label_agreement":null},{"id":"W4405373825","doi":"10.2147/amep.s478193","title":"Assessing the Difficulty and Long-Term Retention of Factual and Conceptual Knowledge Through Multiple-Choice Questions: A Longitudinal Study","year":2024,"lang":"en","type":"article","venue":"Advances in Medical Education and Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Office of the Chief Medical Examiner","funders":"","keywords":"Term (time); Psychology; Computer science; Data science; Information retrieval","score_opus":0.0860952689917893,"score_gpt":0.5197427523275574,"score_spread":0.4336474833357681,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4405373825","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.936236,0.056229535,0.00078283006,0.0052204668,0.0005450705,0.0003236946,0.0000015281925,0.000021966309,0.0006388966],"genre_scores_gemma":[0.98729175,0.011358681,0.0006188251,0.0003971043,0.00019959042,0.000037169928,0.000013266482,0.0000068633944,0.000076729935],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99868447,0.00031853703,0.00033088156,0.00028760853,0.00027199733,0.0001064862],"domain_scores_gemma":[0.93925345,0.060218982,0.00010747702,0.00013180444,0.00013175882,0.00015651638],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010435353,0.00010046646,0.00019826346,0.00004039418,0.00008491884,0.00007223974,0.00004042404,0.00008048334,0.000048844307],"category_scores_gemma":[0.22361043,0.000064695494,0.000020118961,0.00018423876,0.00050439953,0.0007017604,0.000044547054,0.00040174497,0.0000015192736],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000087709,0.0017356555,0.58859515,0.00016775103,0.000042977514,0.000023985098,0.0073625715,3.7411374e-7,0.000010777992,0.001484103,0.00014198424,0.40034696],"study_design_scores_gemma":[0.0010806217,0.00032981942,0.9741793,0.0030755943,0.00024910728,0.00033586298,0.015496876,0.000311596,0.0000036795116,0.00024358553,0.0046081357,0.000085806685],"about_ca_topic_score_codex":0.00012035375,"about_ca_topic_score_gemma":0.000137565,"teacher_disagreement_score":0.40026113,"about_ca_system_score_codex":0.000022827177,"about_ca_system_score_gemma":0.00035517613,"threshold_uncertainty_score":0.7829294},"labels":[],"label_agreement":null},{"id":"W4405420222","doi":"10.1016/j.jpeds.2024.114436","title":"Rash Decisions: Improving Pediatrician Skills in Dermatologic Diagnosis","year":2025,"lang":"en","type":"article","venue":"The Journal of Pediatrics","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"SickKids Foundation; University of Toronto; Hospital for Sick Children","funders":"Hospital for Sick Children; Canadian Dermatology Foundation","keywords":"Medicine; Rash; Dermatology; Pediatrics; MEDLINE; Intensive care medicine","score_opus":0.0157089325447769,"score_gpt":0.32819248105918153,"score_spread":0.31248354851440463,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4405420222","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9752076,0.0111413915,0.0011846591,0.010677827,0.00090150366,0.00022811345,0.0000035607604,0.000017300008,0.00063801114],"genre_scores_gemma":[0.96981525,0.024907418,0.0011548232,0.0031826948,0.0006922899,0.0000049761425,0.0000014427951,0.000012753812,0.0002283513],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99793273,0.0001749224,0.0010317762,0.00012583684,0.00043541292,0.00029934078],"domain_scores_gemma":[0.9544352,0.0442985,0.00048522945,0.00030076198,0.00030757298,0.00017272144],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0020627175,0.00015094643,0.00050980167,0.0005413255,0.00008252152,0.000025311138,0.00032142785,0.000166561,0.000030522093],"category_scores_gemma":[0.14357834,0.000090556096,0.00020822491,0.0011870051,0.00006279421,0.000073296935,0.00011131986,0.0007387079,0.000021645861],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012117618,0.000572034,0.92587095,0.000043969772,0.000010789746,0.00023534177,0.00014891534,0.00010146049,0.000004918081,0.000083508174,0.03974324,0.033063717],"study_design_scores_gemma":[0.0054647843,0.0004891648,0.9826847,0.0011341537,0.0018198304,0.00045123903,0.00064783386,0.00009563835,0.00017910507,0.0045048147,0.0022950685,0.00023369046],"about_ca_topic_score_codex":0.00004214597,"about_ca_topic_score_gemma":0.000017242148,"teacher_disagreement_score":0.14151561,"about_ca_system_score_codex":0.000111546215,"about_ca_system_score_gemma":0.0005462231,"threshold_uncertainty_score":0.86363566},"labels":[],"label_agreement":null},{"id":"W4406242736","doi":"10.1080/0142159x.2024.2445058","title":"Twelve tips on applying AI tools in HPE scholarship using Boyer’s model","year":2025,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University; University of Toronto","funders":"","keywords":"Scholarship; Computer science; Engineering ethics; Artificial intelligence; Medical education; Psychology; Data science; Medicine; Political science; Engineering; Law","score_opus":0.12206730239741184,"score_gpt":0.4380363458022013,"score_spread":0.3159690434047895,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406242736","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95343244,0.00036122598,0.0047640675,0.010911707,0.00029850483,0.00046050717,0.0000018583886,0.00011452486,0.02965514],"genre_scores_gemma":[0.9825116,0.0000375022,0.000815811,0.012221213,0.000238094,0.000054663207,0.000016170445,0.000024542787,0.0040804283],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99775934,0.00012675519,0.00050048134,0.00042626692,0.00076852884,0.0004186222],"domain_scores_gemma":[0.99673784,0.0023414798,0.000048591213,0.0004009949,0.000053079973,0.0004180191],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0017206967,0.00018665806,0.00047250823,0.00016648996,0.000059838814,0.000037361082,0.0001767253,0.00044233544,0.00074255484],"category_scores_gemma":[0.11065908,0.00015087546,0.0001371388,0.00034493982,0.00011922883,0.00006649477,0.000096695614,0.0016382281,0.00010329052],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00068077917,0.0030077577,0.61355543,0.00013720035,0.00015682068,0.0007419481,0.00041123375,0.000811725,0.00017088954,0.005482646,0.021794481,0.3530491],"study_design_scores_gemma":[0.015427116,0.0004632933,0.08560354,0.03078669,0.0004159141,0.00008904547,0.0003185082,0.8362924,0.00038523795,0.014957389,0.014381731,0.00087915186],"about_ca_topic_score_codex":0.00007105293,"about_ca_topic_score_gemma":0.000013093622,"teacher_disagreement_score":0.83548063,"about_ca_system_score_codex":0.00020231,"about_ca_system_score_gemma":0.0005819646,"threshold_uncertainty_score":0.8968322},"labels":[],"label_agreement":null},{"id":"W4406288649","doi":"10.1186/s12909-025-06644-7","title":"Transformation and articulation of clinical data to understand students’ clinical reasoning: a scoping review","year":2025,"lang":"en","type":"review","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières; Montreal Heart Institute; Université Laval; Université de Montréal; Centre for Interdisciplinary Research in Rehabilitation","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"PsycINFO; CINAHL; Rubric; Think aloud protocol; MEDLINE; Curriculum; Medical education; Psychology; Computer science; Mathematics education; Medicine; Pedagogy; Usability","score_opus":0.3679992449776397,"score_gpt":0.6262531680665957,"score_spread":0.258253923088956,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406288649","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0002324521,0.9908001,0.0022475605,0.0010893269,0.0012503046,0.003931172,0.000021553511,0.000038146285,0.00038938443],"genre_scores_gemma":[0.00002977771,0.99228895,0.0031049522,0.002133796,0.000633686,0.000092607304,0.0015047932,0.000023941147,0.00018747845],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9921794,0.0010752267,0.004469152,0.0008452853,0.0011977168,0.0002332413],"domain_scores_gemma":[0.9411802,0.05264574,0.0016835235,0.0022202616,0.0004670758,0.0018031653],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01016424,0.0003075783,0.0033316098,0.00017884387,0.000049994636,0.000029373583,0.00047100478,0.0007118297,0.00018660442],"category_scores_gemma":[0.457141,0.00023562797,0.00045591375,0.00058195513,0.00018661455,0.00009657196,0.0002528008,0.0007191273,0.000031173106],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003449075,0.00083330314,0.0019318791,0.14345723,0.000091833266,0.0000014507497,0.000032811735,3.779943e-8,1.0419463e-9,0.00015375257,0.006833518,0.8466297],"study_design_scores_gemma":[0.000841416,0.00022474359,0.0022008235,0.9475825,0.0036462385,0.00003273619,0.000060212773,0.00005805633,1.7651791e-8,0.00001936089,0.045149673,0.0001842346],"about_ca_topic_score_codex":0.000025467156,"about_ca_topic_score_gemma":0.00005750161,"teacher_disagreement_score":0.84644544,"about_ca_system_score_codex":0.00006637411,"about_ca_system_score_gemma":0.019576486,"threshold_uncertainty_score":0.9859816},"labels":[],"label_agreement":null},{"id":"W4406346651","doi":"10.3389/frhs.2024.1474195","title":"Engaging older adults in diagnostic safety: implementing a diagnostic communication note sheet in a primary care setting","year":2025,"lang":"en","type":"article","venue":"Frontiers in Health Services","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto East General Hospital; University of Toronto","funders":"","keywords":"Primary care; Medicine; Medical emergency; Psychology; Family medicine","score_opus":0.0060118717553014345,"score_gpt":0.3172028478456321,"score_spread":0.31119097609033064,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406346651","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9396769,0.046491172,0.00054930995,0.008247123,0.0007392992,0.002588159,0.000035722187,0.0001074741,0.0015648315],"genre_scores_gemma":[0.97505045,0.0063584396,0.01031545,0.007401052,0.000058593705,0.00018794897,0.0005807477,0.000030457339,0.000016851993],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9961435,0.0005193185,0.0016354155,0.0005689372,0.00028493346,0.00084789743],"domain_scores_gemma":[0.9693017,0.029463913,0.00035212687,0.0006084583,0.00009220255,0.00018158341],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0022526337,0.00026021508,0.000785957,0.00064594194,0.0001791261,0.00003643344,0.00030079263,0.00017813988,0.000013868622],"category_scores_gemma":[0.016022451,0.00027770427,0.00007540202,0.0010039676,0.000049181588,0.00013987266,0.0002660939,0.00088475575,0.000003992872],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00016708026,0.00022437838,0.8378281,0.0047029136,0.0000110448555,0.00004682459,0.014919952,0.00022941292,3.6679918e-7,0.000020426352,0.00025784422,0.14159168],"study_design_scores_gemma":[0.0055015236,0.00007248308,0.87749624,0.08867,0.000027535363,0.0000025200127,0.022630284,0.004583056,0.00000500218,0.00021721766,0.00061079214,0.00018331439],"about_ca_topic_score_codex":0.01119396,"about_ca_topic_score_gemma":0.009592825,"teacher_disagreement_score":0.14140835,"about_ca_system_score_codex":0.0012349727,"about_ca_system_score_gemma":0.00069412554,"threshold_uncertainty_score":0.9999675},"labels":[],"label_agreement":null},{"id":"W4406503468","doi":"10.1212/cpj.0000000000200428","title":"Practice Current—Surveys Navigating Equipoise in Clinical Neurology Worldwide","year":2025,"lang":"en","type":"editorial","venue":"Neurology Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Clinical equipoise; Neurology; Current (fluid); Medicine; Geography; Clinical trial; Psychiatry; Engineering; Pathology","score_opus":0.051619467156027686,"score_gpt":0.5003234001777321,"score_spread":0.44870393302170447,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406503468","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.003402435,0.003928759,0.00027937905,0.094504125,0.88290054,0.0016045991,0.00012809233,0.00037539043,0.012876693],"genre_scores_gemma":[0.0028158773,0.034609172,0.0031110381,0.0992099,0.85741603,0.00021786081,0.0012197631,0.0003072218,0.0010931082],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9435283,0.038526136,0.009646034,0.0043800883,0.0020310502,0.0018884236],"domain_scores_gemma":[0.11679837,0.8710949,0.0051600747,0.0030421289,0.0025815777,0.0013229502],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch","research_integrity"],"category_scores_codex":[0.06115123,0.0012617216,0.0049022706,0.00038928952,0.000213826,0.00013595908,0.0013075925,0.008381583,0.00017045114],"category_scores_gemma":[0.9749555,0.0012430991,0.0015688319,0.0012927607,0.0015708532,0.00057723385,0.0014504136,0.05070171,0.0008892956],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0112289945,0.0064976783,0.028701399,0.0001412707,0.00037492195,0.0066244313,0.00001976127,0.000012182783,1.5867379e-7,0.00005272193,0.7411893,0.20515718],"study_design_scores_gemma":[0.010430596,0.007033895,0.024240222,0.002055481,0.002884861,0.00028595809,0.000013501104,0.00032001038,2.0196977e-7,0.00033018336,0.95168376,0.0007213037],"about_ca_topic_score_codex":0.00069693406,"about_ca_topic_score_gemma":0.00020867337,"teacher_disagreement_score":0.91380423,"about_ca_system_score_codex":0.00009226124,"about_ca_system_score_gemma":0.005362108,"threshold_uncertainty_score":0.9998886},"labels":[],"label_agreement":null},{"id":"W4406509938","doi":"10.1016/j.yped.2014.08.089","title":"10.1016/j.yped.2014.08.089","year":2000,"lang":"en","type":"article","venue":"Time to knit","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Medicine; Constipation; General surgery; Radiology; Surgery","score_opus":0.009965800788594142,"score_gpt":0.2479061946936016,"score_spread":0.23794039390500746,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406509938","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0065954183,0.00013535943,7.798242e-7,0.0017057504,0.0000033075298,0.0001992939,0.00001081921,0.00016409846,0.9911852],"genre_scores_gemma":[0.00087442546,0.0000021396113,0.00023104007,0.00044823004,0.000357787,0.000014675673,0.00003662284,0.000027377724,0.9980077],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99896204,0.000026195039,0.00023941268,0.00026220697,0.00022864931,0.00028148948],"domain_scores_gemma":[0.9972477,0.0018779964,0.000025663712,0.000386511,0.00005155017,0.00041058616],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00019647625,0.00013136632,0.00028878337,0.0000482862,0.00004120456,0.000018172454,0.00009805548,0.000097704156,0.9984422],"category_scores_gemma":[0.0056293514,0.00011066687,0.00010702337,0.00012946401,0.000043832715,0.000025021951,0.000025790825,0.00015792031,0.9985154],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015010472,0.00012641762,0.000004071955,0.0000030700141,0.000016266386,0.000027458012,0.000008382227,0.000006640275,0.000008093309,5.8938934e-7,0.5034914,0.49615747],"study_design_scores_gemma":[0.0008556604,0.0004894298,0.0009197286,0.00029653593,0.0000747282,0.00003552438,7.1545537e-7,0.00007346964,0.000042137137,0.000011144378,0.9970685,0.00013237407],"about_ca_topic_score_codex":0.000031551637,"about_ca_topic_score_gemma":1.0217691e-7,"teacher_disagreement_score":0.49602512,"about_ca_system_score_codex":0.000027896342,"about_ca_system_score_gemma":0.000061676066,"threshold_uncertainty_score":0.6739266},"labels":[],"label_agreement":null},{"id":"W4406546528","doi":"10.1016/s1541-9800(10)70473-7","title":"10.1016/s1541-9800(10)70473-7","year":2000,"lang":"en","type":"article","venue":"Time to knit","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Psychology; Medicine; History","score_opus":0.011261220650148337,"score_gpt":0.24798245971188035,"score_spread":0.236721239061732,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406546528","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004043007,0.00011796738,6.2379775e-7,0.0017168813,0.0000031207628,0.00022107159,0.000013993761,0.00017843672,0.9937049],"genre_scores_gemma":[0.00067405263,0.0000013127515,0.00023706342,0.0004049192,0.00037973167,0.000019575928,0.000049979382,0.00003121108,0.99820215],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9987992,0.000030619485,0.00027827985,0.00029976282,0.00026741362,0.00032469354],"domain_scores_gemma":[0.9969168,0.0020909824,0.000030179033,0.0004277395,0.0000717807,0.00046255632],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00021556782,0.00015447939,0.00032659216,0.0000559563,0.000050595932,0.000021788135,0.00011556674,0.00011193028,0.999087],"category_scores_gemma":[0.008642551,0.00013155177,0.00012441566,0.00020544472,0.000052815507,0.00002858126,0.000031162428,0.00016847406,0.99816376],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002168882,0.0001513969,0.0000015902715,0.000003722752,0.00001938752,0.000034726327,0.000007680726,0.0000078149915,0.000008097993,5.8410205e-7,0.55344707,0.44610104],"study_design_scores_gemma":[0.0010330945,0.00059670524,0.00070355594,0.0003841783,0.00008717563,0.00003647392,9.2956117e-7,0.00010727259,0.00006181328,0.000012458516,0.9968208,0.0001555222],"about_ca_topic_score_codex":0.000036387293,"about_ca_topic_score_gemma":1.2887925e-7,"teacher_disagreement_score":0.4459455,"about_ca_system_score_codex":0.00003673808,"about_ca_system_score_gemma":0.00007202747,"threshold_uncertainty_score":0.99970806},"labels":[],"label_agreement":null},{"id":"W4406554129","doi":"10.1016/s1558-0164(11)70468-0","title":"10.1016/s1558-0164(11)70468-0","year":2000,"lang":"en","type":"article","venue":"Time to knit","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Medicine; Virology; Family medicine","score_opus":0.010554185534486728,"score_gpt":0.2461396038921146,"score_spread":0.23558541835762786,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406554129","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00793194,0.00009269858,9.448536e-7,0.0018619049,0.0000033773038,0.00022434007,0.000011683725,0.00017935874,0.98969376],"genre_scores_gemma":[0.0009631023,0.0000012330463,0.00029466423,0.00044875356,0.00036773496,0.000018779143,0.000038528495,0.000030218118,0.997837],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99884593,0.00002896214,0.0002658098,0.00028900805,0.0002581475,0.0003121419],"domain_scores_gemma":[0.9969554,0.0020974758,0.000028272474,0.0004124337,0.000060368446,0.00044603404],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00021046681,0.0001470715,0.00031304956,0.000054293367,0.000048201513,0.000020718891,0.000111589536,0.00010703315,0.99803716],"category_scores_gemma":[0.0074263625,0.0001248269,0.00011937011,0.00019579528,0.00005077415,0.000027459588,0.000029387207,0.00015336802,0.9988225],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018108894,0.00020287205,0.0000020806883,0.0000050584936,0.00002103016,0.000050671515,0.0000111369645,0.000012485264,0.000009994402,0.0000013128844,0.77027667,0.22922558],"study_design_scores_gemma":[0.0009310164,0.00043427217,0.0011265568,0.00037689708,0.000080925325,0.0000421576,9.295639e-7,0.00008996035,0.00006249409,0.000011995367,0.99669355,0.00014926435],"about_ca_topic_score_codex":0.000052613228,"about_ca_topic_score_gemma":2.3809032e-7,"teacher_disagreement_score":0.22907631,"about_ca_system_score_codex":0.00003452492,"about_ca_system_score_gemma":0.00006704511,"threshold_uncertainty_score":0.8890586},"labels":[],"label_agreement":null},{"id":"W4406646374","doi":"10.1016/b978-84-8086-648-4.00027-3","title":"10.1016/b978-84-8086-648-4.00027-3","year":2000,"lang":"en","type":"book-chapter","venue":"Time to knit","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Computer science","score_opus":0.01554826204753287,"score_gpt":0.24026709769796148,"score_spread":0.2247188356504286,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406646374","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000042503434,0.0006753204,6.4610543e-7,0.00089254114,0.000012286497,0.00058848114,0.00013103172,0.00029803085,0.9973974],"genre_scores_gemma":[0.0000040521018,0.000009159268,0.00014707464,0.000499042,0.0011958893,0.000024551646,0.00038894906,0.00017414751,0.99755716],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9975231,0.000019327195,0.00066010345,0.00072043587,0.00059806707,0.00047891552],"domain_scores_gemma":[0.9945338,0.0034359382,0.0001560262,0.00095270557,0.00015058089,0.00077094836],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00024253351,0.00054578413,0.0010723349,0.00021822573,0.000065927634,0.000040169944,0.00025483203,0.0006962695,0.9982729],"category_scores_gemma":[0.0051668957,0.0004958105,0.0004504385,0.000063148276,0.00013573901,0.00002803692,0.00011131554,0.0007579135,0.9974894],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000267962,0.000076463686,8.9182734e-8,0.000017374256,0.00011542179,0.00019018001,0.000005650175,0.0000026586833,0.0000010292823,0.00005081999,0.438301,0.5609714],"study_design_scores_gemma":[0.0010062149,0.00080408645,0.00001932435,0.003526002,0.00052292325,0.000098070945,2.6726173e-7,0.000012308485,0.0000057576435,0.00035382595,0.99314296,0.0005082366],"about_ca_topic_score_codex":0.000030094083,"about_ca_topic_score_gemma":2.921204e-7,"teacher_disagreement_score":0.56046313,"about_ca_system_score_codex":0.00011708669,"about_ca_system_score_gemma":0.00024833667,"threshold_uncertainty_score":0.99974936},"labels":[],"label_agreement":null},{"id":"W4406857194","doi":"10.57598/r182a","title":"Profylactisch verwijderen van verstandskiezen in afwezigheid van pathologie","year":2012,"lang":"nl","type":"book","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Philosophy","score_opus":0.05581120156350845,"score_gpt":0.3429513530981764,"score_spread":0.287140151534668,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406857194","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0065292493,0.0054658586,0.0002182904,0.0016562636,0.0038082672,0.0020481325,0.00017127211,0.0003270124,0.97977567],"genre_scores_gemma":[0.217458,0.004294364,0.0008920527,0.0014984214,0.0031397797,0.00012602583,0.0010035268,0.00024186194,0.771346],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99296385,0.00039926358,0.0018724151,0.0015460628,0.0012064427,0.0020119916],"domain_scores_gemma":[0.9690429,0.027040606,0.0006201046,0.001718675,0.00024519564,0.0013325139],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["research_integrity","insufficient_payload"],"category_scores_codex":[0.002156695,0.0012305862,0.0025266188,0.0005595094,0.00014507558,0.00009239898,0.00063750835,0.0025493104,0.036912125],"category_scores_gemma":[0.05579983,0.0010219027,0.0006532198,0.00044015917,0.00052044936,0.0002098099,0.00048896554,0.0031460312,0.031276032],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0027354835,0.00577152,0.22812834,0.0010503568,0.001005712,0.0055335006,0.0012026496,0.000014029087,0.000092320166,0.01802032,0.6430625,0.09338326],"study_design_scores_gemma":[0.034985907,0.0051958137,0.18999335,0.04402101,0.005065696,0.0004790238,0.0020891766,0.00010790501,0.00051147514,0.023682633,0.6882347,0.0056333253],"about_ca_topic_score_codex":0.00024729277,"about_ca_topic_score_gemma":0.00005383687,"teacher_disagreement_score":0.21092875,"about_ca_system_score_codex":0.0012867078,"about_ca_system_score_gemma":0.0020573114,"threshold_uncertainty_score":0.9992231},"labels":[],"label_agreement":null},{"id":"W4406986850","doi":"10.2196/57331","title":"Collaborative Development of Feedback Concept Maps for Virtual Patient–Based Clinical Reasoning Education: Mixed Methods Study","year":2025,"lang":"en","type":"article","venue":"JMIR Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Erasmus+; Deutscher Akademischer Austauschdienst; European Commission","keywords":"Concept map; Scope (computer science); Focus group; Presentation (obstetrics); Content analysis; Process (computing); Qualitative research; Field (mathematics); Psychology; Medical education; Computer science; Mathematics education; Medicine; Mathematics; Sociology","score_opus":0.023275964946026715,"score_gpt":0.4851960831809061,"score_spread":0.4619201182348794,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406986850","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9660027,0.00064205116,0.014215397,0.005817859,0.006638463,0.0040277503,0.000014992278,0.00008075444,0.0025600328],"genre_scores_gemma":[0.83730304,0.000012643532,0.15324739,0.004966627,0.00048111298,0.0020482729,0.00047629178,0.000025808538,0.0014388376],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9966316,0.00057488214,0.0014517028,0.0005135418,0.0005883435,0.00023992418],"domain_scores_gemma":[0.9839404,0.013286209,0.00044719398,0.0004526593,0.0013245103,0.00054901047],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0021559047,0.0002048197,0.00067776,0.00016899026,0.0001262709,0.000019279101,0.00017175193,0.0003012932,0.00025270358],"category_scores_gemma":[0.12782992,0.00017517002,0.00015272881,0.00071001484,0.00024605682,0.00004386276,0.00006004505,0.00034323617,0.000010397971],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004753982,0.012942751,0.018846722,0.0000649241,0.00014303243,7.034981e-7,0.0029385812,0.0000010836005,0.000009604794,0.001232284,0.07857608,0.88476884],"study_design_scores_gemma":[0.018411161,0.006744377,0.63507277,0.015895521,0.0010028641,0.000007897249,0.10045724,0.00091028115,0.0040739435,0.0010151837,0.21560492,0.0008038326],"about_ca_topic_score_codex":0.000016482198,"about_ca_topic_score_gemma":0.000010862928,"teacher_disagreement_score":0.883965,"about_ca_system_score_codex":0.00016982006,"about_ca_system_score_gemma":0.027890377,"threshold_uncertainty_score":0.9776206},"labels":[],"label_agreement":null},{"id":"W4407230972","doi":"10.1111/acem.15120","title":"Diagnosis through prisms: Unraveling its complexity","year":2025,"lang":"en","type":"article","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Medicine; Metaphor; Percept; Variety (cybernetics); Beauty; Epistemology; Artificial intelligence; Computer science; Linguistics","score_opus":0.14569336631466673,"score_gpt":0.45034624759034536,"score_spread":0.30465288127567863,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4407230972","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6736543,0.07505441,0.004501453,0.099813506,0.008449079,0.0015312656,0.00002380059,0.0006108812,0.13636126],"genre_scores_gemma":[0.883373,0.09299095,0.00097458827,0.008230797,0.0019350938,0.00017882505,0.000118222415,0.000046022473,0.012152495],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9966718,0.000088033106,0.0015196763,0.000611686,0.0005623323,0.0005464907],"domain_scores_gemma":[0.99479413,0.0038367186,0.0002540858,0.0005179976,0.00029156829,0.00030548943],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0009065529,0.00031964178,0.0008072045,0.00016105217,0.000157117,6.8051725e-7,0.00030150745,0.0004289259,0.0048832246],"category_scores_gemma":[0.07489836,0.00024667277,0.00018084727,0.0009911025,0.00030426192,0.0000869978,0.00013395539,0.001225967,0.00021412387],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00016053188,0.00027684335,0.5119724,0.0003049562,0.00028284616,0.000033264398,0.0010417986,0.0000057331104,0.00088863866,0.025263932,0.45051214,0.009256882],"study_design_scores_gemma":[0.01009712,0.0011499467,0.4554307,0.031177461,0.0027458565,0.000060788912,0.0019394676,0.00094652025,0.005906042,0.06897276,0.42050114,0.0010722058],"about_ca_topic_score_codex":0.00023331358,"about_ca_topic_score_gemma":0.0000074727436,"teacher_disagreement_score":0.20971866,"about_ca_system_score_codex":0.00009167954,"about_ca_system_score_gemma":0.0001677193,"threshold_uncertainty_score":0.99999857},"labels":[],"label_agreement":null},{"id":"W4407633314","doi":"10.29333/ajqr/15880","title":"Is Anecdotal Evidence Science?","year":2025,"lang":"en","type":"article","venue":"American Journal of Qualitative Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Windsor","funders":"","keywords":"Astrobiology; Physics","score_opus":0.43811351225851985,"score_gpt":0.6859147150484807,"score_spread":0.24780120278996082,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4407633314","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95660895,0.0013251877,0.00041480307,0.03463683,0.000118096905,0.00012154833,0.0000025612362,0.0000060950038,0.006765946],"genre_scores_gemma":[0.9936618,0.00056053215,0.0032851622,0.0012120641,0.0001024315,0.0000031040875,2.1220525e-7,0.0000060411394,0.0011686126],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9960616,0.0010080965,0.0005422939,0.0002227,0.0017395008,0.00042578788],"domain_scores_gemma":[0.9085892,0.08537388,0.0003103157,0.0003671764,0.00479005,0.00056941016],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.01875708,0.000082683655,0.00044727387,0.0008151051,0.00017183379,0.000051806644,0.00033401026,0.000024087736,0.00012629249],"category_scores_gemma":[0.27406293,0.000058767986,0.00015069116,0.0031903198,0.0050688614,0.00015686575,0.00010575464,0.0007945494,0.000050731986],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0049353037,0.0016662023,0.07580773,0.0001483111,0.0006294071,0.0007392984,0.08874415,0.000006562757,0.0133894095,0.017566947,0.07795676,0.7184099],"study_design_scores_gemma":[0.0057023256,0.03625092,0.43667424,0.04414301,0.00031529184,0.0007016935,0.41050658,0.0002906772,0.022617787,0.033090923,0.009145621,0.0005609443],"about_ca_topic_score_codex":0.00016930082,"about_ca_topic_score_gemma":9.825023e-7,"teacher_disagreement_score":0.71784896,"about_ca_system_score_codex":0.00023871889,"about_ca_system_score_gemma":0.0030848822,"threshold_uncertainty_score":0.99763876},"labels":[],"label_agreement":null},{"id":"W4407795764","doi":"10.1111/eje.13079","title":"Learning by Concordance as a Tool for Paediatric Dental Traumatology Education","year":2025,"lang":"en","type":"article","venue":"European Journal Of Dental Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Likert scale; Concordance; Modality (human–computer interaction); Modalities; Medical education; Traumatology; Feeling; Scale (ratio); Psychology; Mathematics education; Medicine; Computer science; Social psychology; Artificial intelligence; Surgery; Developmental psychology","score_opus":0.009505796058305427,"score_gpt":0.33923507153445454,"score_spread":0.3297292754761491,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4407795764","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9868003,0.0025995488,0.0007828053,0.0015131386,0.0024390598,0.00022885267,0.0000018926626,0.000013941679,0.0056204624],"genre_scores_gemma":[0.9919683,0.00024046241,0.0013368363,0.0014211137,0.00055799034,0.000007679207,0.000052320436,0.00001758679,0.004397738],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99886686,0.00017504748,0.0005448471,0.0001398814,0.00014436868,0.00012898006],"domain_scores_gemma":[0.99797106,0.0012123247,0.00036888925,0.00009269287,0.00022817936,0.00012685588],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0005571099,0.00009835782,0.00022718766,0.00015924049,0.000075442884,0.00003326691,0.000102848346,0.0000322466,0.000052097977],"category_scores_gemma":[0.014564427,0.00009037191,0.00015337615,0.00018275112,0.000043996704,0.000082704515,0.0000164918,0.0002553971,0.00005476023],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00052400073,0.0018789042,0.23428676,0.00007814132,0.0000853869,0.000015194761,0.00039740803,0.0000064089936,0.001926796,0.00062965485,0.11402965,0.6461417],"study_design_scores_gemma":[0.014619761,0.007822508,0.6845591,0.00837313,0.0017261477,0.005912912,0.02702776,0.00012383332,0.007955561,0.0010218836,0.24013333,0.0007240922],"about_ca_topic_score_codex":0.000009710001,"about_ca_topic_score_gemma":7.757705e-7,"teacher_disagreement_score":0.6454176,"about_ca_system_score_codex":0.00011871631,"about_ca_system_score_gemma":0.0011027779,"threshold_uncertainty_score":0.9937363},"labels":[],"label_agreement":null},{"id":"W4408073568","doi":"10.1515/dx-2024-0160","title":"Equity-Driven Diagnostic Excellence framework: An upstream approach to minimize risk of diagnostic inequity","year":2025,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Nipissing University; Hamilton Health Sciences","funders":"","keywords":"Excellence; Equity (law); Upstream (networking); Business; Economics; Computer science; Political science","score_opus":0.03943722389982272,"score_gpt":0.37143058238515614,"score_spread":0.3319933584853334,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4408073568","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9499262,0.0016267152,0.029712519,0.0028509195,0.0008857842,0.002014252,0.00032505597,0.00029180068,0.012366776],"genre_scores_gemma":[0.96772283,0.004672824,0.023588978,0.0022340606,0.00029627752,0.00108962,0.00011503314,0.000056950965,0.00022341972],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99582475,0.00034181168,0.0011252344,0.0011098542,0.0007407771,0.0008575832],"domain_scores_gemma":[0.83241045,0.16420618,0.00035075826,0.0016733371,0.0004058272,0.00095346116],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0010527053,0.0005065958,0.0012936711,0.0003910262,0.00018839262,0.00006819431,0.00069050904,0.0004853668,0.00025518512],"category_scores_gemma":[0.51531637,0.0004545213,0.00042652898,0.0012694962,0.00033515188,0.00012061577,0.00065341825,0.0007706304,0.00014317053],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00017791368,0.0052223476,0.9002608,0.00021731401,0.00024218453,0.000049597395,0.0005621185,0.0007827712,0.000013190868,0.0026735559,0.0090695135,0.08072872],"study_design_scores_gemma":[0.0026625253,0.0012933125,0.9722977,0.009688969,0.0019353696,0.000011907915,0.0005075863,0.0009759935,0.002240977,0.006165493,0.0015552611,0.0006648925],"about_ca_topic_score_codex":0.000917845,"about_ca_topic_score_gemma":0.000029973002,"teacher_disagreement_score":0.51426363,"about_ca_system_score_codex":0.00015342174,"about_ca_system_score_gemma":0.00036169135,"threshold_uncertainty_score":0.99979067},"labels":[],"label_agreement":null},{"id":"W4408139285","doi":"10.1093/med:psych/9780190271442.003.0170","title":"Letter to Dr W. Clifford M. Scott, 5 December 1961","year":2016,"lang":"en","type":"letter","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Art; History; Medicine","score_opus":0.032912038881494174,"score_gpt":0.33849111926587466,"score_spread":0.3055790803843805,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4408139285","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0004230334,0.00006807106,0.0005554704,0.8515055,0.0014300783,0.0005565312,0.000037434347,0.00018666692,0.1452372],"genre_scores_gemma":[0.0001726496,0.000032404692,0.00068651617,0.7242164,0.026459467,0.00005689794,0.00015327836,0.00009176406,0.2481306],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9969993,0.000060292637,0.0006253181,0.0008009095,0.00068857323,0.0008256076],"domain_scores_gemma":[0.9881904,0.010043839,0.000113578724,0.0010642323,0.00016742563,0.000420537],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0002765293,0.0004618194,0.0009018983,0.00018896176,0.000049240796,0.000048255017,0.00023846855,0.001345491,0.006102271],"category_scores_gemma":[0.016269704,0.00027958615,0.00043050983,0.0001470695,0.000104299004,0.000030461004,0.00013899783,0.0015571442,0.016708361],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005241392,0.000036913298,0.00065459334,0.000039533214,0.0000943432,0.0011125717,0.000012016009,4.7184063e-8,0.000028181254,0.000029654564,0.98940074,0.008539011],"study_design_scores_gemma":[0.001167713,0.0002721564,0.00045602745,0.0039651566,0.00022368792,0.00007852321,9.39413e-7,9.0416523e-7,0.0001856304,0.00019480077,0.9930354,0.00041909178],"about_ca_topic_score_codex":0.000086671935,"about_ca_topic_score_gemma":0.000016379976,"teacher_disagreement_score":0.1272891,"about_ca_system_score_codex":0.00011060377,"about_ca_system_score_gemma":0.00020930653,"threshold_uncertainty_score":0.9999656},"labels":[],"label_agreement":null},{"id":"W4408312537","doi":"10.2196/68891","title":"Designing Health Care Provider–Centered Emergency Department Interventions: Participatory Design Study","year":2025,"lang":"en","type":"article","venue":"JMIR Formative Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Preprint; Psychological intervention; Emergency department; Citizen journalism; Participatory design; Computer science; Engineering; Medicine; World Wide Web; Nursing; Operations management","score_opus":0.3003982668833761,"score_gpt":0.5679071365705929,"score_spread":0.26750886968721677,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4408312537","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.90886486,0.012019142,0.051090844,0.0061684153,0.0010958572,0.016313743,0.000056102876,0.00034948776,0.0040415223],"genre_scores_gemma":[0.9961366,0.00016458675,0.0007556393,0.0001850916,0.000047164864,0.0020780838,0.000052886633,0.00001682845,0.0005630834],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99583644,0.0014814135,0.00086926407,0.00036264086,0.00074085704,0.0007093658],"domain_scores_gemma":[0.9952743,0.003122275,0.00016035262,0.0004470787,0.0006568969,0.00033909996],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0031453576,0.00016816592,0.0004308904,0.00038937756,0.00044143715,0.000045074128,0.00018165213,0.0000686945,0.00021378217],"category_scores_gemma":[0.0100790365,0.00013564309,0.00018842923,0.00069981295,0.00011170215,0.00012402788,0.0002111066,0.0005931846,0.0001644716],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0027950096,0.020488905,0.25568953,0.003776605,0.0013453106,0.00018288488,0.10267665,0.000053393807,0.00008093963,0.00088898203,0.38094112,0.23108068],"study_design_scores_gemma":[0.01706043,0.03627589,0.7171646,0.025734223,0.0002965014,0.000014187856,0.19119458,0.000966637,0.0024374723,0.0016019676,0.00656494,0.00068857754],"about_ca_topic_score_codex":0.00008428621,"about_ca_topic_score_gemma":0.0000511878,"teacher_disagreement_score":0.46147504,"about_ca_system_score_codex":0.00056059886,"about_ca_system_score_gemma":0.00080455863,"threshold_uncertainty_score":0.9982595},"labels":[],"label_agreement":null},{"id":"W4408345546","doi":"10.1016/j.kine.2024.12.095","title":"Impact d’une formation par concordance sur les performances des étudiants en physiothérapie lors des examens cliniques objectifs structurés (ECOS) : une étude randomisée","year":2025,"lang":"fr","type":"article","venue":"Kinésithérapie la Revue","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Medicine","score_opus":0.01985262478607013,"score_gpt":0.3252545564345671,"score_spread":0.30540193164849694,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4408345546","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94844395,0.03823434,0.001870337,0.0047697076,0.0019425835,0.0012547063,0.00017583469,0.00026651946,0.003042052],"genre_scores_gemma":[0.94196373,0.051761348,0.0034134206,0.00021504668,0.0013935274,0.00013430708,0.00012543777,0.00009403619,0.00089913554],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9946708,0.00079560373,0.0017539643,0.0009474203,0.00047991663,0.0013523102],"domain_scores_gemma":[0.97403735,0.023169694,0.00070208916,0.0008826773,0.00061259145,0.0005955772],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0018749915,0.0009237745,0.0019708658,0.00034884314,0.00064119545,0.00022384395,0.0005090424,0.00074977,0.00055276015],"category_scores_gemma":[0.019548627,0.00080258714,0.0008300875,0.0015153036,0.0017793368,0.0010243187,0.00019648879,0.0011646874,0.00016051646],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0038717568,0.0016464046,0.42523143,0.0046476983,0.0021491249,0.00043781285,0.010786785,0.0011998588,0.0015905891,0.0009742861,0.021425717,0.5260385],"study_design_scores_gemma":[0.017647866,0.0012771287,0.93349195,0.029844541,0.0011827652,0.00019901797,0.0014986614,0.0039031215,0.0031332981,0.0060168332,0.0010142665,0.00079052855],"about_ca_topic_score_codex":0.0036880637,"about_ca_topic_score_gemma":0.002481986,"teacher_disagreement_score":0.525248,"about_ca_system_score_codex":0.0005999277,"about_ca_system_score_gemma":0.0007905002,"threshold_uncertainty_score":0.9994425},"labels":[],"label_agreement":null},{"id":"W4408561604","doi":"10.46747/cfp.7103214","title":"Des soins primaires caractérisés par l’incertitude","year":2025,"lang":"fr","type":"article","venue":"Canadian Family Physician","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Computer science; Data science","score_opus":0.025845591197792058,"score_gpt":0.30656521605660675,"score_spread":0.2807196248588147,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4408561604","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.16274583,0.08792086,0.00013459621,0.015425612,0.006123235,0.0007164554,0.00064094615,0.00012403615,0.7261684],"genre_scores_gemma":[0.78391355,0.0020490848,0.00032414126,0.19851492,0.002577709,0.00005020968,0.00016637055,0.000070149,0.012333841],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9968864,0.00012438945,0.0005490041,0.0006444482,0.00027628356,0.0015194864],"domain_scores_gemma":[0.9942415,0.0030830153,0.000103143444,0.0006451256,0.00022197432,0.0017052446],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00028759948,0.00042753908,0.00072488416,0.00029374412,0.00045751163,0.000115286246,0.00034460507,0.0003291092,0.00004056741],"category_scores_gemma":[0.004865402,0.00046611726,0.00033334797,0.0008215106,0.000981774,0.00014026226,0.000059455193,0.0006393216,0.0010855487],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000030994644,0.0002994337,0.00852253,0.00017445303,0.00032626608,0.00047491127,0.00013498115,0.00005098245,0.00035637792,0.008898088,0.4166257,0.5641053],"study_design_scores_gemma":[0.0009858594,0.00016957452,0.5362786,0.009772324,0.0005560877,0.000001793072,0.00042424843,0.00020871557,0.00031652008,0.004456003,0.44645256,0.00037770346],"about_ca_topic_score_codex":0.21949035,"about_ca_topic_score_gemma":0.1585249,"teacher_disagreement_score":0.7138346,"about_ca_system_score_codex":0.0015846135,"about_ca_system_score_gemma":0.008437132,"threshold_uncertainty_score":0.99977905},"labels":[],"label_agreement":null},{"id":"W4408688552","doi":"10.2196/65263","title":"Large Language Models for Pediatric Differential Diagnoses in Rural Health Care: Multicenter Retrospective Cohort Study Comparing GPT-3 With Pediatrician Performance","year":2025,"lang":"en","type":"article","venue":"JMIRx Med","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Medical diagnosis; Medicine; Retrospective cohort study; Pediatrics; Cohort; Health care; Cohort study; Family medicine; Surgery; Internal medicine; Radiology","score_opus":0.011303112770279573,"score_gpt":0.3343810915467513,"score_spread":0.3230779787764717,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4408688552","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99372864,0.0006364286,0.00060741,0.00067055627,0.0003386197,0.003504581,0.000033974276,0.00009219031,0.0003875921],"genre_scores_gemma":[0.9976985,0.00018911932,0.00020430038,0.00034521386,0.00036917324,0.0006305287,0.00012609613,0.000028389457,0.00040870195],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99806714,0.000074007745,0.0005035916,0.00043614194,0.00037919154,0.00053993776],"domain_scores_gemma":[0.9975245,0.0016487909,0.00014951051,0.00033550293,0.00016719803,0.00017450986],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0003646343,0.00023590583,0.00080531114,0.00025047144,0.00012613552,0.000029129807,0.00012344448,0.00008597571,0.000014469822],"category_scores_gemma":[0.0024873675,0.00018250002,0.00009828447,0.00043953257,0.000029209148,0.00007568466,0.00007493232,0.00035329844,0.0000045128168],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00079627853,0.0016080652,0.9909644,0.00016710228,0.000063488325,0.00001871204,0.0042361314,0.000046918165,7.3884786e-7,0.00003344403,0.0013324972,0.0007322136],"study_design_scores_gemma":[0.012473674,0.0010679777,0.97649837,0.00042605045,0.0002858546,0.0000011077719,0.0047584595,0.0042785774,0.000010613518,0.000006033663,0.00002356873,0.0001696904],"about_ca_topic_score_codex":0.00051480887,"about_ca_topic_score_gemma":0.0011468759,"teacher_disagreement_score":0.014466018,"about_ca_system_score_codex":0.00034203182,"about_ca_system_score_gemma":0.00023642633,"threshold_uncertainty_score":0.7442135},"labels":[],"label_agreement":null},{"id":"W4408978095","doi":"10.1016/j.soin.2025.02.008","title":"L’approche par concordance en formation infirmière : une innovation à découvrir","year":2025,"lang":"fr","type":"article","venue":"Soins","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Interdisciplinary Research in Rehabilitation","funders":"","keywords":"Philosophy; Psychology; Humanities","score_opus":0.02553720361311921,"score_gpt":0.3547302280210149,"score_spread":0.32919302440789566,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4408978095","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.64219356,0.008851272,0.04166412,0.10998769,0.009049434,0.0012270838,0.00008364495,0.00026978093,0.18667343],"genre_scores_gemma":[0.9546759,0.00087634637,0.0011621533,0.004401346,0.0009916939,0.00004975273,0.00023367068,0.000019092358,0.037590038],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9982092,0.000079830825,0.0007634156,0.0002908082,0.0002443489,0.00041241586],"domain_scores_gemma":[0.99329835,0.0056724544,0.0002047706,0.0003803573,0.00033584543,0.0001082266],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0007251539,0.00020452289,0.0003835293,0.00017768914,0.00011845996,0.00003805659,0.00012459178,0.0003341884,0.00015325687],"category_scores_gemma":[0.043134376,0.00020117025,0.00009283198,0.0017852181,0.0001951389,0.00018814369,0.00007053388,0.00048596947,0.00049044687],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00017587344,0.00067834347,0.034846764,0.00071098807,0.0001639277,0.000046039582,0.0015292902,0.0001381725,0.0008442019,0.25547528,0.19194509,0.51344603],"study_design_scores_gemma":[0.0057704095,0.0006015933,0.19423336,0.020540278,0.0005695836,0.000035483252,0.0019099454,0.0076575982,0.0064495727,0.030338315,0.7314152,0.00047866802],"about_ca_topic_score_codex":0.0003950668,"about_ca_topic_score_gemma":0.00012185103,"teacher_disagreement_score":0.5394701,"about_ca_system_score_codex":0.00022201658,"about_ca_system_score_gemma":0.00078834547,"threshold_uncertainty_score":0.9649257},"labels":[],"label_agreement":null},{"id":"W4409240838","doi":"10.2196/56384","title":"Differential Diagnosis Assessment in Ambulatory Care With a Digital Health History Device: Pseudorandomized Study","year":2025,"lang":"en","type":"article","venue":"JMIR Formative Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Preprint; Ambulatory; Randomized controlled trial; Digital health; Medicine; Health care; Computer science; Political science; World Wide Web; Surgery","score_opus":0.05364371653089382,"score_gpt":0.47394203008451896,"score_spread":0.4202983135536251,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4409240838","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9896528,0.0009344381,0.00010922309,0.0010381879,0.00014531176,0.0029401032,0.000018325354,0.000051943705,0.005109674],"genre_scores_gemma":[0.99737376,0.00013647799,0.00006818778,0.00021963086,0.00004446852,0.0017087045,0.00007972021,0.000018270099,0.00035075934],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9969503,0.00047086953,0.0005780925,0.00038421276,0.0010411079,0.0005754001],"domain_scores_gemma":[0.9909993,0.007790193,0.00009772836,0.0004075943,0.00042430047,0.00028091588],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009877323,0.0001950299,0.0007491975,0.0004888338,0.00014384672,0.000064205764,0.00016322194,0.00008606606,0.00012993021],"category_scores_gemma":[0.0021956475,0.00013702283,0.00013290416,0.0005960881,0.00023505752,0.00016848618,0.00018166812,0.0011021418,0.000025274858],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.005187823,0.0076222327,0.8976703,0.00057977583,0.0003143077,0.0001134585,0.027114777,0.0000048188804,0.0000033644776,0.00022583507,0.01790938,0.043253925],"study_design_scores_gemma":[0.045365818,0.004362212,0.9167893,0.005226548,0.000031399966,0.0000036914407,0.024913844,0.00026594868,0.000015852127,0.000045135475,0.002801264,0.00017900935],"about_ca_topic_score_codex":0.00037220767,"about_ca_topic_score_gemma":0.00018003338,"teacher_disagreement_score":0.043074917,"about_ca_system_score_codex":0.0022353698,"about_ca_system_score_gemma":0.0022457442,"threshold_uncertainty_score":0.58454144},"labels":[],"label_agreement":null},{"id":"W4409341463","doi":"10.2196/66126","title":"Feasibility of a Randomized Controlled Trial of Large AI-Based Linguistic Models for Clinical Reasoning Training of Physical Therapy Students: Pilot Randomized Parallel-Group Study","year":2025,"lang":"en","type":"article","venue":"JMIR Formative Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Preprint; Training (meteorology); Randomized controlled trial; Effi; Artificial intelligence; Computer science; Psychology; Physical therapy; Natural language processing; Medicine; World Wide Web; Surgery","score_opus":0.1683317366160802,"score_gpt":0.5576376589262059,"score_spread":0.3893059223101257,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4409341463","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93658173,0.00032303494,0.019258149,0.0001965448,0.00023796696,0.042225495,0.000050949955,0.00003262855,0.0010935242],"genre_scores_gemma":[0.9919318,0.00016581088,0.000648442,0.000078583704,0.00016619478,0.006871343,0.000037216043,0.000025972966,0.00007464905],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.98416513,0.008505718,0.004167376,0.00053793413,0.0020032122,0.0006206111],"domain_scores_gemma":[0.7461681,0.24916367,0.0011112437,0.0006786872,0.0026695782,0.00020874705],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.0636164,0.0003274981,0.008769188,0.0004896888,0.0001320294,0.000028137118,0.0004410902,0.00017680845,0.000019998572],"category_scores_gemma":[0.32420298,0.00020275652,0.0027056183,0.0006403922,0.0013820172,0.00008126047,0.00014861414,0.00088178425,9.721559e-7],"study_design_candidate":"randomized_trial","study_design_consensus":"randomized_trial","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.94416314,0.037537206,0.0017241339,0.00024810334,0.0024877943,0.0000027261192,0.006804832,0.00006816724,0.0000150795995,0.0060732146,0.00007285175,0.00080272934],"study_design_scores_gemma":[0.8994532,0.014883368,0.000749795,0.0024301698,0.000648628,1.747799e-7,0.0014149676,0.07334807,0.0000412206,0.0069062556,0.0000032722198,0.00012085075],"about_ca_topic_score_codex":0.00006396139,"about_ca_topic_score_gemma":0.000005460385,"teacher_disagreement_score":0.26058656,"about_ca_system_score_codex":0.000080877115,"about_ca_system_score_gemma":0.001108298,"threshold_uncertainty_score":0.96420395},"labels":[],"label_agreement":null},{"id":"W4409428311","doi":"10.46747/cfp.7104e68","title":"Prevalence and management of symptom diagnoses in children in general practice","year":2025,"lang":"en","type":"article","venue":"Canadian Family Physician","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Medical diagnosis; General practice; Data science; Medicine; Family medicine; Computer science; Pediatrics; Pathology","score_opus":0.010619120488500706,"score_gpt":0.2963312293514619,"score_spread":0.28571210886296117,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4409428311","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9367337,0.0025406093,0.0000020456123,0.001310356,0.000054989214,0.00043182564,0.000019331988,0.0000073771703,0.058899723],"genre_scores_gemma":[0.9665608,0.0021372826,0.0003083576,0.030565241,0.000034036417,0.000035032765,0.000014218804,0.000008349196,0.00033671953],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99915373,0.000038143324,0.00023196005,0.00023941716,0.000098725825,0.00023802494],"domain_scores_gemma":[0.9986073,0.0009421599,0.00004298123,0.00021365007,0.000032963133,0.00016099721],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00015455899,0.00009705414,0.00022784062,0.00021377177,0.000017688702,0.0000092607115,0.00007554544,0.000054463442,0.0000011461583],"category_scores_gemma":[0.00087801827,0.00010034461,0.00003050469,0.0003816429,0.00007037043,0.000052469673,0.00002714941,0.00013116776,0.0000048746733],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003774204,0.00026173465,0.91993874,0.00016341121,0.00009335412,0.00015114745,0.000058953578,0.000029430776,0.000027874687,0.0069813747,0.002970587,0.06928564],"study_design_scores_gemma":[0.0010418925,0.000051859577,0.99393445,0.003368359,0.00008611529,8.779737e-7,0.00012754837,0.00004433375,0.000035925303,0.00025992253,0.00096929405,0.00007944387],"about_ca_topic_score_codex":0.028082581,"about_ca_topic_score_gemma":0.0035215828,"teacher_disagreement_score":0.07399567,"about_ca_system_score_codex":0.00012652644,"about_ca_system_score_gemma":0.00023550299,"threshold_uncertainty_score":0.9783895},"labels":[],"label_agreement":null},{"id":"W4409499964","doi":"10.52358/mm.vi20.429","title":"Ingenium de formation lors de la mise en place de formations transversales en santé Illustration par une séquence d’apprentissage du raisonnement clinique","year":2025,"lang":"fr","type":"article","venue":"Médiations et médiatisations","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Philosophy; Humanities","score_opus":0.019102314582448206,"score_gpt":0.33300349297174886,"score_spread":0.31390117838930065,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4409499964","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3020151,0.003258717,0.614199,0.05530851,0.001076602,0.0017817401,0.00083377573,0.0003471718,0.02117943],"genre_scores_gemma":[0.9613852,0.009152016,0.02239805,0.001989738,0.0005970478,0.0007191231,0.0012712849,0.0000505115,0.0024370372],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"observational","domain_scores_codex":[0.99469227,0.0021488399,0.0014299362,0.00045791344,0.0004227883,0.0008482313],"domain_scores_gemma":[0.96731776,0.030957269,0.0004413082,0.0005508817,0.0002953897,0.00043739166],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0039180513,0.0004301109,0.00050156354,0.00051640254,0.00079006836,0.00026630066,0.00031461398,0.00075994234,0.0005040206],"category_scores_gemma":[0.030785048,0.00051292544,0.00033391692,0.0011618959,0.00040305572,0.0008604564,0.00006731022,0.0009003293,0.00013238679],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00029193723,0.0053711915,0.15859391,0.001383621,0.0011395415,0.000093867086,0.15632784,0.15495929,0.0024171534,0.26757467,0.14547901,0.10636797],"study_design_scores_gemma":[0.009548922,0.00046889382,0.6220674,0.0067036385,0.0034201287,0.00018653653,0.023721574,0.23595218,0.0037399011,0.01735584,0.07560068,0.0012343207],"about_ca_topic_score_codex":0.0010572147,"about_ca_topic_score_gemma":0.0011436337,"teacher_disagreement_score":0.6593701,"about_ca_system_score_codex":0.0018594967,"about_ca_system_score_gemma":0.003561414,"threshold_uncertainty_score":0.99973226},"labels":[],"label_agreement":null},{"id":"W4409512100","doi":"10.1186/s12909-025-07134-6","title":"Quantify difference between physicians and medical students in clinical reasoning: evidence from eye-tracking","year":2025,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"Inner Mongolia Medical University","keywords":"Medical education; Tracking (education); Eye tracking; MEDLINE; Educational measurement; Medicine; Psychology; Curriculum; Computer science; Pedagogy; Artificial intelligence","score_opus":0.06997983774980288,"score_gpt":0.5007527503252808,"score_spread":0.4307729125754779,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4409512100","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98357123,0.0018949697,0.0022651276,0.010516957,0.0011632118,0.0003194865,0.0000023333337,0.000058240457,0.00020841825],"genre_scores_gemma":[0.9906021,0.0017338822,0.001304751,0.004854764,0.0010075772,0.000044738514,0.00006753114,0.000014220358,0.00037044665],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9957588,0.00044038397,0.0010954936,0.00067609025,0.001714094,0.0003151925],"domain_scores_gemma":[0.9369374,0.0607038,0.00027976854,0.00069670245,0.00020862305,0.0011737453],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0031086751,0.00019120077,0.00069662195,0.000121645986,0.00006405242,0.00006213703,0.00043799356,0.00053360994,0.00024674294],"category_scores_gemma":[0.44943428,0.0001630913,0.000116846684,0.00038309826,0.00036503928,0.000070352966,0.00022082403,0.0009796302,0.000033626504],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006927302,0.0009275749,0.7983242,0.00005136813,0.00002899835,0.000011277025,0.00018850515,9.546625e-8,0.0000035144553,0.00015845225,0.0012387175,0.198998],"study_design_scores_gemma":[0.0014875396,0.00010858148,0.9576137,0.03873967,0.0001513408,0.0000024973062,0.00027494854,0.0007357484,0.000016929327,0.00057650864,0.00016336412,0.00012917514],"about_ca_topic_score_codex":0.0011730398,"about_ca_topic_score_gemma":0.00051654375,"teacher_disagreement_score":0.4463256,"about_ca_system_score_codex":0.0000876513,"about_ca_system_score_gemma":0.0043225116,"threshold_uncertainty_score":0.766795},"labels":[],"label_agreement":null},{"id":"W4409592332","doi":"10.1177/12034754251334581","title":"Evaluating the Impact of Novel Educational Modules on Medical Students’ Confidence, Knowledge, and Diagnostic Accuracy in Skin of Colour Dermatology: A National Pilot Study","year":2025,"lang":"en","type":"letter","venue":"Journal of Cutaneous Medicine and Surgery","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Probity Medical Research; Women's College Hospital; McMaster University; University of Toronto","funders":"","keywords":"Medicine; Dermatology; Medical physics","score_opus":0.10150373118855782,"score_gpt":0.4711471223329896,"score_spread":0.3696433911444318,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4409592332","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.83443785,0.0028561498,0.0000070176366,0.16161388,0.00043301142,0.00044460787,0.000018452043,0.0000026638434,0.00018635902],"genre_scores_gemma":[0.96848965,0.0022658987,0.000029819015,0.0275769,0.0014397503,0.000015952764,0.000031724074,0.00001755462,0.00013277052],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99452025,0.00050029205,0.0021046575,0.000269941,0.0023560766,0.00024880114],"domain_scores_gemma":[0.6824621,0.31461266,0.001367416,0.00020049752,0.0011680457,0.0001892781],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0042173946,0.0002999201,0.0018788158,0.00083331694,0.00005391802,0.000014614737,0.00027787217,0.00027590536,0.00014056587],"category_scores_gemma":[0.3445006,0.00016931938,0.0002069639,0.00034895624,0.00053966,0.00003331506,0.000090511574,0.0016933274,6.1245106e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014197619,0.0061220936,0.3854619,0.00051555666,0.0011646922,0.0023583472,0.0009961078,0.000047923793,0.000017152157,0.000039919898,0.5980986,0.0037579592],"study_design_scores_gemma":[0.011813277,0.022905847,0.84441704,0.083675765,0.0020499397,0.02978372,0.0010399055,0.0011573391,0.000009157106,0.0015611758,0.0011406357,0.0004461796],"about_ca_topic_score_codex":0.00027544875,"about_ca_topic_score_gemma":0.000026785572,"teacher_disagreement_score":0.5969579,"about_ca_system_score_codex":0.00010693509,"about_ca_system_score_gemma":0.0037845178,"threshold_uncertainty_score":0.73567593},"labels":[],"label_agreement":null},{"id":"W4410002602","doi":"10.1080/0142159x.2025.2497889","title":"Proceed with caution: Conceptualizing the risk of artificial intelligence scribes in medical education","year":2025,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"MEDLINE; Psychology; Medical education; Computer science; Medicine; Political science","score_opus":0.03024599987352408,"score_gpt":0.38048215906067884,"score_spread":0.3502361591871548,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4410002602","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9529468,0.0012259863,0.004029487,0.032150082,0.00032329437,0.0003903626,8.4687076e-7,0.00004410715,0.008889055],"genre_scores_gemma":[0.9968301,0.00016745154,0.00018918455,0.0017710347,0.0002141308,0.0000492122,0.000011248072,0.000007255203,0.00076036603],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9980711,0.00016461105,0.0005097615,0.00022546551,0.00084591313,0.00018320125],"domain_scores_gemma":[0.99685794,0.0024251416,0.00009985677,0.00023923586,0.00013361542,0.00024422316],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0017471865,0.000103395825,0.00029562428,0.00007007594,0.00004610829,0.000008166508,0.00017065991,0.00021797491,0.0016345824],"category_scores_gemma":[0.13807398,0.000056494267,0.000060075654,0.00047994615,0.00069810444,0.000020466545,0.000047134607,0.00072047143,0.000018737972],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00033441823,0.0020668844,0.504519,0.00006759996,0.00009761358,0.00003569348,0.0017513839,0.0000045448583,0.0000036530175,0.021218382,0.006084734,0.46381614],"study_design_scores_gemma":[0.0074791783,0.0025301354,0.6840885,0.079513244,0.0018727529,0.00036267348,0.04431159,0.103477746,0.0020340076,0.044850472,0.028280575,0.0011990991],"about_ca_topic_score_codex":0.00044159012,"about_ca_topic_score_gemma":0.00020606672,"teacher_disagreement_score":0.46261704,"about_ca_system_score_codex":0.000048612084,"about_ca_system_score_gemma":0.0022140318,"threshold_uncertainty_score":0.99927807},"labels":[],"label_agreement":null},{"id":"W4410188644","doi":"10.1007/s40670-025-02402-y","title":"Twelve Tips to Train Medical Students to Manage Their Uncertainty and to Provide Reassurance to Patients and Their Caregivers","year":2025,"lang":"en","type":"article","venue":"Medical Science Educator","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"University College Dublin; Irish Research eLibrary","keywords":"Medical education; Psychology; Medicine","score_opus":0.011971521463858775,"score_gpt":0.353963015404356,"score_spread":0.3419914939404972,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4410188644","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8932121,0.000055664463,0.00027734996,0.103416204,0.001030895,0.0011371596,0.000020967162,0.000043752483,0.00080592703],"genre_scores_gemma":[0.9525291,0.00002708896,0.00037267964,0.0461163,0.00019731146,0.00013207004,0.0000043463706,0.000011560032,0.0006095875],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9956631,0.00007671625,0.00041287133,0.000990416,0.0021470848,0.0007098239],"domain_scores_gemma":[0.9908905,0.0028159847,0.000039815262,0.00045083402,0.00021524543,0.0055876174],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0026897842,0.00024982353,0.00050015503,0.00029300727,0.00022032816,0.0001136023,0.0008999676,0.00013984733,0.00016396226],"category_scores_gemma":[0.13650917,0.00016333675,0.00005321593,0.0015266612,0.00054875435,0.00006012756,0.00076535216,0.00031121218,0.000062731066],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00031394,0.0010119305,0.21658708,0.00006463178,0.000045991506,0.000064390544,0.0143153025,0.0000026478453,0.00012268136,0.00020177828,0.06723109,0.70003855],"study_design_scores_gemma":[0.0022901085,0.0012531088,0.8908662,0.005050114,0.00004453355,0.000022280372,0.0056049395,0.00009109429,0.0003341091,0.00016975758,0.09385733,0.00041641982],"about_ca_topic_score_codex":0.00077043526,"about_ca_topic_score_gemma":0.00035266008,"teacher_disagreement_score":0.6996221,"about_ca_system_score_codex":0.0002373469,"about_ca_system_score_gemma":0.0013392187,"threshold_uncertainty_score":0.8707644},"labels":[],"label_agreement":null},{"id":"W4410293035","doi":"10.2196/68527","title":"The Advanced Reasoning Capabilities of Large Language Models for Detecting Contraindicated Options in Medical Exams","year":2025,"lang":"en","type":"article","venue":"JMIR Medical Informatics","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Computer science; Clinical Practice; Sample (material); Medical physics; Medical practice; Medicine; Medical education; Family medicine","score_opus":0.015348139348143726,"score_gpt":0.36914523427384355,"score_spread":0.35379709492569983,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4410293035","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9723343,0.00071142503,0.016535766,0.0030741964,0.00024827293,0.0011775832,0.000030015734,0.00010042484,0.00578799],"genre_scores_gemma":[0.9952229,0.0002540568,0.0021503975,0.0016781471,0.000058288715,0.00029580627,0.000048363585,0.0000112291555,0.00028075816],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99736696,0.000060376773,0.00122894,0.0001163558,0.0007879563,0.0004393972],"domain_scores_gemma":[0.97974575,0.019242335,0.00021299903,0.00033897144,0.00014003903,0.0003199242],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0026627805,0.00013890481,0.000476004,0.00011823274,0.00012976964,0.00001698808,0.00025000272,0.00033222363,0.00007818192],"category_scores_gemma":[0.11310435,0.0000913177,0.00014057133,0.00035594098,0.00033443188,0.000087021916,0.00009860833,0.0005855823,0.0000031615216],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013381131,0.0018307491,0.006679756,0.0022143268,0.0004353065,0.00007973249,0.042055946,0.00033441262,0.000057135043,0.13397862,0.0042007696,0.8067951],"study_design_scores_gemma":[0.019946959,0.000632879,0.006694317,0.021523276,0.00016691012,0.00006792896,0.07415674,0.866093,0.0006448735,0.0052254735,0.004472281,0.00037535897],"about_ca_topic_score_codex":0.00003591926,"about_ca_topic_score_gemma":0.000116172625,"teacher_disagreement_score":0.8657586,"about_ca_system_score_codex":0.00006903814,"about_ca_system_score_gemma":0.0007474762,"threshold_uncertainty_score":0.8943663},"labels":[],"label_agreement":null},{"id":"W4410499705","doi":"10.1111/jep.70124","title":"Characterizing Personal Clinical Cognitive Uncertainty and Its Association With Clinical Judgment","year":2025,"lang":"en","type":"article","venue":"Journal of Evaluation in Clinical Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université de Montréal; Centre Hospitalier Universitaire Sainte-Justine; Centre Hospitalier de l’Université de Montréal; Université Laval","funders":"","keywords":"Clinical judgment; Context (archaeology); Psychology; Cognition; Test (biology); Supervisor; Clinical psychology; Medicine; Social psychology; Psychiatry; Medical physics","score_opus":0.1172555181004824,"score_gpt":0.5233957039174945,"score_spread":0.40614018581701206,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4410499705","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9546214,0.000910541,0.0002925085,0.038314704,0.0019317961,0.0006742475,0.0000061961377,0.000012319102,0.0032363127],"genre_scores_gemma":[0.9803568,0.0034563986,0.0015253107,0.013099059,0.0012283956,0.000014105615,0.000014269851,0.000014854985,0.00029076356],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9892694,0.0039016188,0.0045203455,0.00048053492,0.001547524,0.00028062594],"domain_scores_gemma":[0.58978164,0.39564705,0.006154142,0.0002926091,0.007418927,0.00070560846],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.06935172,0.00019457494,0.001158859,0.00018567541,0.00007916521,0.00006741429,0.00008766455,0.00053069316,0.00014268725],"category_scores_gemma":[0.7584794,0.00014690998,0.00039168313,0.00039251318,0.00015429225,0.0004284553,0.00006567245,0.0025482713,0.00002431025],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008916488,0.0033982024,0.85349,0.000028654322,0.0008982069,0.00014973628,0.00013380409,0.000014061791,0.000004516368,0.000101953,0.0026440637,0.13022031],"study_design_scores_gemma":[0.018549632,0.0036333187,0.9573122,0.0043375716,0.003317692,0.00014671947,0.0008542373,0.0073040333,0.0000067959495,0.0002516682,0.004129848,0.0001562611],"about_ca_topic_score_codex":0.000011517336,"about_ca_topic_score_gemma":0.000011079343,"teacher_disagreement_score":0.6891277,"about_ca_system_score_codex":0.0003126204,"about_ca_system_score_gemma":0.0019870205,"threshold_uncertainty_score":0.9997529},"labels":[],"label_agreement":null},{"id":"W4410863666","doi":"10.1016/j.bja.2025.04.033","title":"Incidence of cognitive errors in difficult airway management: an inference human factors study from the Pediatric Difficult Intubation Registry","year":2025,"lang":"en","type":"article","venue":"British Journal of Anaesthesia","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"SickKids Foundation; BC Children's Hospital; Hospital for Sick Children","funders":"","keywords":"Intubation; Incidence (geometry); Cognition; Inference; Airway management; Medicine; Intensive care medicine; Psychology; Computer science; Anesthesia; Psychiatry; Artificial intelligence","score_opus":0.01935137422248517,"score_gpt":0.32903069236875626,"score_spread":0.3096793181462711,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4410863666","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9985778,0.00053000747,0.00007883731,0.00017126418,0.000080482816,0.0003567807,0.0000122180745,0.000011095828,0.00018154019],"genre_scores_gemma":[0.99855083,0.0010642805,0.00007025818,0.000120227814,0.00008728859,0.0000058133164,0.00002705753,0.000009902448,0.00006431539],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99804425,0.00023437195,0.000871904,0.00022329151,0.00045814793,0.00016800807],"domain_scores_gemma":[0.9960313,0.0026160928,0.0005555865,0.000201199,0.0004792357,0.00011662729],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0005940761,0.00014160485,0.00046695065,0.0001587852,0.00009189352,0.000054695156,0.000292912,0.000094077615,0.000029455246],"category_scores_gemma":[0.016258959,0.000112128604,0.00010447373,0.0006922843,0.000114544586,0.00013658527,0.000054842556,0.00048841554,0.0000013813456],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000101762285,0.0016120119,0.9636116,0.000015935997,0.00007008836,0.00069247466,0.0010035541,0.00002319792,0.000008111514,0.000040209343,0.0002762328,0.03254481],"study_design_scores_gemma":[0.0024265333,0.0013812479,0.98220736,0.0045582964,0.00042433385,0.000036160593,0.008645232,0.000014647107,0.000026759682,0.00017912991,0.0000034016616,0.00009689021],"about_ca_topic_score_codex":0.0017999038,"about_ca_topic_score_gemma":0.0008256664,"teacher_disagreement_score":0.03244792,"about_ca_system_score_codex":0.00005940991,"about_ca_system_score_gemma":0.00013916253,"threshold_uncertainty_score":0.9920275},"labels":[],"label_agreement":null},{"id":"W4411039051","doi":"10.21203/rs.3.rs-6725427/v1","title":"Across Generations, Sizes, and Types, Large Language Models Poorly Report Self-Confidence in Gastroenterology Clinical Reasoning Tasks","year":2025,"lang":"en","type":"preprint","venue":"Research Square","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ontario Tech University","funders":"American College of Gastroenterology","keywords":"Confidence interval; Self-confidence; Psychology; Internal medicine; Computer science; Medicine; Social psychology","score_opus":0.0712675693048107,"score_gpt":0.49358168001146924,"score_spread":0.42231411070665853,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411039051","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9864348,0.005476213,0.0010709523,0.0030448365,0.000400426,0.0013526374,0.0003813174,0.0001484568,0.0016903663],"genre_scores_gemma":[0.9864171,0.0041743894,0.004800974,0.00054725894,0.0005702303,0.00023631481,0.0007133656,0.000036416437,0.0025039485],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9944144,0.00109539,0.0012534753,0.0013098735,0.00085860054,0.0010682583],"domain_scores_gemma":[0.9880547,0.009250513,0.00021949313,0.0012926349,0.00068074896,0.000501923],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.009391898,0.0003019251,0.0010372548,0.00028290355,0.00017636432,0.00013506657,0.00030374428,0.0008463552,0.000067642075],"category_scores_gemma":[0.07741781,0.00027952637,0.00022315158,0.00031961553,0.00026607438,0.00006511643,0.0015324898,0.0036923627,0.000020869946],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00086022663,0.0014915623,0.97118205,0.0008849523,0.0002687782,0.007714564,0.0020882874,0.00033583413,0.000011817187,0.0020931582,0.007945632,0.0051231687],"study_design_scores_gemma":[0.010590931,0.002250906,0.8721311,0.035944447,0.00033701002,0.0010768039,0.003360168,0.06464348,0.00006414918,0.0054502157,0.0032937445,0.0008570546],"about_ca_topic_score_codex":0.0018161329,"about_ca_topic_score_gemma":0.0016542509,"teacher_disagreement_score":0.099050924,"about_ca_system_score_codex":0.00021371053,"about_ca_system_score_gemma":0.0014507812,"threshold_uncertainty_score":0.99996567},"labels":[],"label_agreement":null},{"id":"W4411094071","doi":"10.1080/09593985.2025.2515476","title":"The clinical reasoning in neurological physiotherapy: an exploratory qualitative study on the (dis)similarities between Bobath and non-Bobath therapists","year":2025,"lang":"en","type":"article","venue":"Physiotherapy Theory and Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; University of Toronto","funders":"Fundação para a Ciência e a Tecnologia","keywords":"Physical medicine and rehabilitation; Physical therapy; Medicine; Psychology","score_opus":0.06967347746902486,"score_gpt":0.5003993033251856,"score_spread":0.4307258258561607,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411094071","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9825077,0.0018857475,0.00016731348,0.012994347,0.00016640853,0.0009897337,0.000008601591,0.000054917444,0.0012252217],"genre_scores_gemma":[0.978812,0.0037695356,0.000093384435,0.016653962,0.00032260793,0.00011956827,0.000004766336,0.000033167216,0.00019101962],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.98839784,0.009685731,0.0005962076,0.0006532597,0.00032020707,0.00034674827],"domain_scores_gemma":[0.81184554,0.18707627,0.00022585198,0.0006152383,0.00009912861,0.0001379655],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.013161636,0.00032569162,0.00062130584,0.000062059386,0.00056824955,0.0001752693,0.00024190753,0.00015775282,0.000019685936],"category_scores_gemma":[0.027968455,0.00017277442,0.00011346869,0.00030171536,0.00076388934,0.0002546506,0.00007420169,0.0010074436,0.0000037175753],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.09605359,0.015208338,0.18609698,0.00011439335,0.0031635126,0.00019868645,0.28343615,0.000016122394,0.00095685653,0.1726817,0.0026006184,0.23947306],"study_design_scores_gemma":[0.0087230215,0.010446631,0.6997946,0.001088788,0.0003176913,0.000011360023,0.18680772,0.00014985711,0.000088458044,0.087842405,0.004208647,0.0005208022],"about_ca_topic_score_codex":0.000054840908,"about_ca_topic_score_gemma":0.00002601847,"teacher_disagreement_score":0.5136976,"about_ca_system_score_codex":0.000017056655,"about_ca_system_score_gemma":0.00013412033,"threshold_uncertainty_score":0.98021936},"labels":[],"label_agreement":null},{"id":"W4411272806","doi":"10.1073/pnas.2426153122","title":"Human–AI collectives most accurately diagnose clinical vignettes","year":2025,"lang":"en","type":"article","venue":"Proceedings of the National Academy of Sciences","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":30,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Innovation Cluster (Canada)","funders":"Max-Planck-Institut für Bildungsforschung; HORIZON EUROPE Framework Programme; Universität zu Köln; European Commission; Max-Planck-Gesellschaft; Deutsche Forschungsgemeinschaft; HORIZON EUROPE European Research Council; Kaiser Permanente; University of Oxford; Harvey Mudd College","keywords":"Artificial intelligence; Psychology; Computer science; History; Natural language processing","score_opus":0.10426877262262277,"score_gpt":0.46617043520033785,"score_spread":0.36190166257771506,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411272806","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95441264,0.0001841848,6.5932693e-7,0.023480948,0.000048679336,0.00025689905,0.0000069701396,0.000018384908,0.021590648],"genre_scores_gemma":[0.9938175,0.00007503231,0.0004234563,0.0041831406,0.000120264434,0.000014546205,1.9341583e-7,0.0000030470874,0.0013628058],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9981158,0.000012119077,0.00061991304,0.0002918617,0.0008093339,0.00015096678],"domain_scores_gemma":[0.9922116,0.0066603976,0.0004072753,0.000013372027,0.0006347063,0.000072647614],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0020178023,0.000095070885,0.000311959,0.00016099526,0.00022124642,0.000024263569,0.00046058817,0.00014136442,0.000036862453],"category_scores_gemma":[0.07714915,0.000059280625,0.00014936386,0.0013020618,0.0015217011,0.00015009435,0.00015165303,0.00027496406,0.0000024407864],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006633271,0.0006260938,0.8042928,0.000082461214,0.000077455756,5.002068e-8,0.00008023252,0.000015593167,0.015752956,0.10863613,0.06773358,0.0026363072],"study_design_scores_gemma":[0.0007389585,0.00016644424,0.9076524,0.0013638729,0.00006286393,0.0000038954304,0.00010762047,0.00017418366,0.038055275,0.050952718,0.00065622514,0.000065563094],"about_ca_topic_score_codex":0.000012396667,"about_ca_topic_score_gemma":1.585272e-7,"teacher_disagreement_score":0.10335957,"about_ca_system_score_codex":0.000028527398,"about_ca_system_score_gemma":0.00016387939,"threshold_uncertainty_score":0.9306244},"labels":[],"label_agreement":null},{"id":"W4411433553","doi":"10.1016/j.ard.2025.06.818","title":"POS1474-HPR VALIDATION OF THE DANISH TRANSLATION OF THE ILLNESS INTRUSIVENESS RATING SCALE IN PATIENTS WITH RHEUMATIC DISEASE","year":2025,"lang":"en","type":"article","venue":"Annals of the Rheumatic Diseases","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Medicine; Danish; Rating scale; Disease; Severity of illness; Physical therapy; Scale (ratio); Internal medicine; Statistics","score_opus":0.01701024765021691,"score_gpt":0.30371411222644046,"score_spread":0.28670386457622354,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411433553","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99595076,0.000318437,0.000086064996,0.0023040967,0.00018372326,0.0009341884,0.000088137465,0.000008546514,0.00012603306],"genre_scores_gemma":[0.99959326,0.000055384156,0.000036640242,0.00020398207,0.0000054701277,0.000030140895,0.000020761063,0.0000113739925,0.00004298596],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9980988,0.0002912949,0.00073112914,0.00016306098,0.0005679114,0.00014782517],"domain_scores_gemma":[0.9952994,0.0030409065,0.00057683,0.0007121252,0.00029881555,0.000071900104],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0003315785,0.00014153967,0.0004263897,0.000057666985,0.0000945132,0.000011701312,0.00033154778,0.00004070103,0.00002379395],"category_scores_gemma":[0.015760344,0.00006752292,0.000248762,0.0005818339,0.000297821,0.000076712844,0.00009412725,0.00009473569,5.1638426e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00046253207,0.001102867,0.96943855,0.0011406698,0.00010466954,1.1708826e-7,0.0005923095,0.0015675313,0.0000096036765,0.00026160013,0.00013533956,0.02518423],"study_design_scores_gemma":[0.0016423787,0.000057259236,0.96777844,0.025079323,0.00037051563,1.4491361e-7,0.00022953005,0.00067200983,0.0019353295,0.002158643,0.0000026958928,0.000073732204],"about_ca_topic_score_codex":0.00012435946,"about_ca_topic_score_gemma":0.000016006286,"teacher_disagreement_score":0.025110498,"about_ca_system_score_codex":0.000016653968,"about_ca_system_score_gemma":0.0003160706,"threshold_uncertainty_score":0.99253035},"labels":[],"label_agreement":null},{"id":"W4411445091","doi":"10.7202/1118374ar","title":"Narrative of a Practice for the Adaptation and Use of a Training tool for Evaluative Judgment Based on Concordance","year":2023,"lang":"en","type":"article","venue":"Mesure et évaluation en éducation","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"University of Ottawa; Université de Montréal","funders":"","keywords":"Formative assessment; Concordance; Adaptation (eye); Interpretation (philosophy); Narrative; Context (archaeology); Psychology; Test (biology); Pedagogy; Applied psychology; Mathematics education; Cognitive psychology; Computer science; Medicine","score_opus":0.24183210369631894,"score_gpt":0.4766062542430951,"score_spread":0.23477415054677617,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411445091","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8298824,0.000051237996,0.13968895,0.026210979,0.00034245357,0.0036571918,0.00006001008,0.000029610237,0.00007718293],"genre_scores_gemma":[0.98216236,0.000038539907,0.016030071,0.00062925974,0.000055145192,0.00069344556,0.00025548527,0.000012403183,0.0001232784],"study_design_codex":"qualitative","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99834937,0.000632645,0.00036749785,0.0001854528,0.0003778215,0.000087190325],"domain_scores_gemma":[0.9040499,0.09447697,0.00037237172,0.00014993847,0.00092694093,0.000023906112],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.010030432,0.00008294033,0.0002021125,0.00008057953,0.00006470947,0.000011453511,0.000027738613,0.000054130378,0.000010869097],"category_scores_gemma":[0.24536869,0.00006145936,0.00006682505,0.00021631432,0.000048707887,0.000109028595,0.0000049007717,0.00006618704,9.127909e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004000686,0.00079409004,0.0018399643,0.000215506,0.0002954654,3.2521007e-7,0.74294484,0.06831458,0.0028936374,0.004244534,0.006355052,0.1681013],"study_design_scores_gemma":[0.004803886,0.0022544414,0.24349314,0.0011650707,0.0006852735,0.0000014348893,0.033123784,0.71079713,0.001748947,0.0011271989,0.0006889724,0.00011072429],"about_ca_topic_score_codex":0.000034549386,"about_ca_topic_score_gemma":0.000018145945,"teacher_disagreement_score":0.70982105,"about_ca_system_score_codex":0.00006293205,"about_ca_system_score_gemma":0.0007966533,"threshold_uncertainty_score":0.7609879},"labels":[],"label_agreement":null},{"id":"W4411777726","doi":"10.1093/cid/ciaf270","title":"Practicing With Intent: How to Teach an Old Dogma New Tricks","year":2025,"lang":"en","type":"article","venue":"Clinical Infectious Diseases","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Medicine","score_opus":0.03762108200655636,"score_gpt":0.4121296751271532,"score_spread":0.37450859312059687,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411777726","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9632564,0.00030565093,0.00856179,0.01827659,0.0010392538,0.0009423149,0.000015975449,0.0005849871,0.0070170225],"genre_scores_gemma":[0.9730289,0.00011599887,0.0003192642,0.017033096,0.00090064347,0.000048888924,0.00004372396,0.000033519405,0.008475939],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9977811,0.00019643932,0.00060948986,0.00072477316,0.00029304257,0.00039515563],"domain_scores_gemma":[0.9796959,0.017374128,0.00018711871,0.0007973474,0.0002962042,0.0016493063],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00047627088,0.00025745577,0.0007519376,0.00018444748,0.00012723803,0.00013285402,0.00014572749,0.00019785181,0.00013702778],"category_scores_gemma":[0.16831361,0.0001882079,0.0002932753,0.00065727514,0.00016258337,0.00013979687,0.0001004429,0.00057509175,0.00011264003],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006052145,0.00265287,0.85279053,0.00003081921,0.00021960816,0.000113435926,0.000037462076,0.000022227203,0.0000059948015,0.00050625764,0.051078558,0.09193703],"study_design_scores_gemma":[0.0063000573,0.0051150066,0.95421106,0.0023900177,0.0015017971,0.000026971695,0.00019818946,0.00017343265,0.00002992108,0.00085390825,0.028874338,0.00032529898],"about_ca_topic_score_codex":0.0002172212,"about_ca_topic_score_gemma":0.000078588695,"teacher_disagreement_score":0.16783734,"about_ca_system_score_codex":0.00008775843,"about_ca_system_score_gemma":0.00073596585,"threshold_uncertainty_score":0.83869207},"labels":[],"label_agreement":null},{"id":"W4411866731","doi":"10.1371/journal.pone.0327192","title":"Identification of digital clinical decision support systems for supporting diagnosis and triage of patients with shoulder disorders: A scoping review protocol","year":2025,"lang":"en","type":"review","venue":"PLoS ONE","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Mount Royal University","funders":"Alberta Innovates","keywords":"CINAHL; Clinical decision support system; MEDLINE; Triage; Medicine; Protocol (science); Systematic review; Decision support system; Point of care; Medical emergency; Nursing; Psychological intervention; Computer science; Alternative medicine; Data mining; Pathology","score_opus":0.13125271279113593,"score_gpt":0.4718147082813539,"score_spread":0.340561995490218,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411866731","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0001688471,0.7377206,0.00038260673,0.00003160387,0.00005508635,0.26116154,0.00041378546,0.000026156893,0.000039739974],"genre_scores_gemma":[0.000073034586,0.86483854,0.00035959692,0.000019766194,0.000058191024,0.13400811,0.0005314454,0.000040093702,0.00007122907],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.994308,0.00012799325,0.0041111973,0.0005908436,0.00063843053,0.00022355313],"domain_scores_gemma":[0.96329683,0.03213074,0.0030943183,0.00066037284,0.00063878257,0.00017892984],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001705842,0.0003135088,0.0041739657,0.00015597306,0.000033023214,0.000034334542,0.00014811082,0.00028855924,0.000022964334],"category_scores_gemma":[0.1303078,0.00021522549,0.0004992561,0.00032913766,0.000118278884,0.000088160894,0.000094026305,0.0002380747,0.0000021246742],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013392481,0.003302718,0.018357161,0.48908615,0.00037214713,8.779555e-7,0.000004059463,3.1094455e-8,9.655507e-9,0.0000035482499,0.00022128169,0.4885181],"study_design_scores_gemma":[0.0027146712,0.001436624,0.00032200463,0.98467344,0.007575565,7.275918e-7,0.0000039709766,0.000007839507,0.000001792002,0.0000105183935,0.0030650543,0.00018779396],"about_ca_topic_score_codex":0.0000024899705,"about_ca_topic_score_gemma":0.000001245123,"teacher_disagreement_score":0.4955873,"about_ca_system_score_codex":0.000030708936,"about_ca_system_score_gemma":0.0006931463,"threshold_uncertainty_score":0.8776641},"labels":[],"label_agreement":null},{"id":"W4411891485","doi":"10.1186/s12909-025-07501-3","title":"Entry to practice physiotherapy students’ use of the international IFOMPT cervical framework to inform clinical reasoning: a qualitative case study design","year":2025,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Medical education; Medicine; Qualitative research; Physical therapy; MEDLINE; Clinical Practice; Entry Level; Psychology; Medical physics","score_opus":0.10748573812255253,"score_gpt":0.5700374881891258,"score_spread":0.4625517500665733,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411891485","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.91924655,0.000041900395,0.04607849,0.029291082,0.0031781185,0.0018467818,0.0000037939137,0.000035509445,0.00027775558],"genre_scores_gemma":[0.8856032,0.00004319784,0.067507476,0.044288207,0.00060075783,0.00026777506,0.000009503529,0.000018372448,0.0016615174],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.995841,0.00094591617,0.00102279,0.0004167212,0.0015686814,0.0002048371],"domain_scores_gemma":[0.8914589,0.10502541,0.00039251312,0.0009929432,0.0011576402,0.0009725445],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0037656277,0.00016539654,0.0004221483,0.00012504778,0.00009562088,0.00006549006,0.00043529304,0.00022433711,0.00027719248],"category_scores_gemma":[0.62442005,0.00011416034,0.00016226694,0.0007992963,0.00011161011,0.00009640704,0.00024788297,0.00057835394,0.000046999754],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0043033417,0.03289132,0.71061593,0.00008195248,0.0007720309,0.000104815445,0.04992697,0.00007969139,0.0000073892606,0.004509711,0.122125104,0.074581735],"study_design_scores_gemma":[0.0068709305,0.0033817517,0.807174,0.013833584,0.0008419421,0.0003519123,0.13141714,0.0009007803,0.000070050504,0.0010516741,0.033648178,0.00045804045],"about_ca_topic_score_codex":0.00078154844,"about_ca_topic_score_gemma":0.00011206685,"teacher_disagreement_score":0.62065446,"about_ca_system_score_codex":0.00016438091,"about_ca_system_score_gemma":0.0042914865,"threshold_uncertainty_score":0.76129127},"labels":[],"label_agreement":null},{"id":"W4412034537","doi":"10.5334/pme.1658","title":"Learning-by-Concordance Approach in Health Professions Education: A Scoping Review","year":2025,"lang":"en","type":"review","venue":"Perspectives on Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Montréal; Université de Montréal","funders":"Université de Genève","keywords":"CINAHL; Concordance; Comparability; Medical education; Inclusion (mineral); MEDLINE; Variety (cybernetics); Health care; Systematic review; Medicine; Management science; Psychology; Computer science; Psychological intervention; Nursing; Artificial intelligence; Engineering","score_opus":0.042115220964403455,"score_gpt":0.49527048999179685,"score_spread":0.4531552690273934,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4412034537","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000017548947,0.9723202,0.000108821485,0.013614575,0.000832419,0.0042513297,0.000004654358,0.00007769854,0.0087885745],"genre_scores_gemma":[0.000011010421,0.9826672,0.000544682,0.0061773267,0.00057664723,0.002603748,0.00074550835,0.000041607993,0.0066322624],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9953673,0.0010241043,0.0014104652,0.0010266678,0.0007767793,0.00039470164],"domain_scores_gemma":[0.99003536,0.0074548637,0.000761491,0.0006685548,0.00025921338,0.0008205021],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0018180885,0.00043047703,0.0025111868,0.0003698869,0.00013348671,0.000021160382,0.00027944637,0.00043993877,0.0005224613],"category_scores_gemma":[0.18515797,0.00033662922,0.00039453604,0.0012750324,0.0001618561,0.000041879215,0.00006353117,0.0022998527,0.000074713535],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000071971945,0.003421092,0.000030129784,0.14947112,0.000023257202,9.972345e-7,0.0005247798,7.38912e-8,1.3044905e-9,0.0006289214,0.033698466,0.812194],"study_design_scores_gemma":[0.00024577134,0.00019949074,0.000022114074,0.8329229,0.00019725044,0.000031793246,0.0019812146,0.0000034030427,1.1438993e-8,0.000031040956,0.16416222,0.0002027994],"about_ca_topic_score_codex":0.00012479603,"about_ca_topic_score_gemma":0.0000056300355,"teacher_disagreement_score":0.81199116,"about_ca_system_score_codex":0.0010617882,"about_ca_system_score_gemma":0.058475174,"threshold_uncertainty_score":0.99990857},"labels":[],"label_agreement":null},{"id":"W4412558422","doi":"10.36834/cmej.78199","title":"Documenting medical students’ use of self-explanations: tool development and initial validity evidence","year":2025,"lang":"en","type":"article","venue":"Canadian Medical Education Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Université de Sherbrooke","funders":"","keywords":"Computer science; Data science","score_opus":0.058389546976328956,"score_gpt":0.41901184429375904,"score_spread":0.36062229731743006,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4412558422","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9606091,0.0007142143,0.00026104567,0.03642727,0.0012740054,0.00018446223,0.000001764082,0.000015420499,0.00051273435],"genre_scores_gemma":[0.95778716,0.0010849156,0.0044569597,0.035470128,0.00052902824,0.000022772265,0.000017977169,0.00000967341,0.0006214195],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.997233,0.00013327939,0.0007253158,0.00018084346,0.0014768752,0.0002506866],"domain_scores_gemma":[0.98489976,0.0071536275,0.00022403197,0.00023201753,0.00074255763,0.00674798],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0025101828,0.0001035347,0.00025185183,0.0003095487,0.00018915471,0.00011149298,0.00021831936,0.00020942948,0.0045711696],"category_scores_gemma":[0.45122227,0.000090210284,0.000051377378,0.0002575697,0.00012270227,0.00013757129,0.00005041279,0.0005964311,0.0000106896805],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004497159,0.0007903673,0.5335082,0.00012615154,0.00022220475,0.00024320996,0.0012835544,5.665651e-7,0.0000015114005,0.0015098208,0.13476306,0.3275064],"study_design_scores_gemma":[0.0024374013,0.0001716759,0.71261203,0.02599294,0.00031994094,0.0013103989,0.00088180014,0.00024662912,0.000100068384,0.00040791012,0.25524226,0.00027695138],"about_ca_topic_score_codex":0.0020564275,"about_ca_topic_score_gemma":0.0026476786,"teacher_disagreement_score":0.44871208,"about_ca_system_score_codex":0.00031078132,"about_ca_system_score_gemma":0.073412694,"threshold_uncertainty_score":0.9963388},"labels":[],"label_agreement":null},{"id":"W4412618712","doi":"10.1111/tct.70161","title":"Thinking About the Why: A Qualitative Study on Students' Perspectives of Paediatric Team‐Based Learning Discussions","year":2025,"lang":"en","type":"article","venue":"The Clinical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"Women and Children's Health Research Institute; Children's Health Research Institute","keywords":"Medical education; Psychology; Team-based learning; Qualitative research; Mathematics education; Medicine; Sociology","score_opus":0.07939386787825405,"score_gpt":0.5029297359825068,"score_spread":0.42353586810425276,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4412618712","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9702614,0.000549347,0.00038475453,0.021488473,0.0002710178,0.0010192586,0.000002987309,0.00009865297,0.00592414],"genre_scores_gemma":[0.99203396,0.00006667121,0.00010311705,0.0021880688,0.00029419755,0.00006570512,0.0000035898026,0.000020841459,0.005223871],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99441034,0.003212172,0.0009080984,0.00043865808,0.00077644805,0.00025428316],"domain_scores_gemma":[0.92252827,0.07609596,0.00032168548,0.00077599694,0.00016006327,0.0001180247],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.010025793,0.00020106399,0.00063290744,0.000078609286,0.00032546942,0.000040592673,0.0005568814,0.00014034234,0.00012019412],"category_scores_gemma":[0.14877737,0.00008507933,0.0004634591,0.0004982205,0.00052413344,0.000020403466,0.00024257085,0.0017416923,0.000041026375],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009609759,0.011416699,0.834483,0.000014914258,0.00077167386,0.000010343369,0.13056558,0.00007809888,0.000003692716,0.0022642445,0.010052416,0.009378348],"study_design_scores_gemma":[0.003589226,0.0016829944,0.68204194,0.0009384621,0.00078514725,4.063963e-7,0.30790025,0.0001149383,0.000003832828,0.0019836754,0.0008443132,0.00011483068],"about_ca_topic_score_codex":0.00007242189,"about_ca_topic_score_gemma":0.0000076613005,"teacher_disagreement_score":0.17733465,"about_ca_system_score_codex":0.000052222877,"about_ca_system_score_gemma":0.0001750669,"threshold_uncertainty_score":0.85839283},"labels":[],"label_agreement":null},{"id":"W4412629305","doi":"10.2196/70751","title":"Using Natural Language Processing to Identify Symptomatic Adverse Events in Pediatric Oncology: Tutorial for Clinician Researchers","year":2025,"lang":"en","type":"article","venue":"JMIR Bioinformatics and Biotechnology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Pediatric oncology; Medicine; Adverse effect; Medical education; Oncology; Medical physics; Internal medicine; Cancer","score_opus":0.045729921261460346,"score_gpt":0.4622073313749039,"score_spread":0.4164774101134435,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4412629305","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99348325,0.00035653095,0.0008053853,0.0032952768,0.00047294833,0.0012837782,0.0000081457665,0.000086362525,0.00020833653],"genre_scores_gemma":[0.97709155,0.00011326086,0.021617774,0.0008110425,0.00014442259,0.00007269906,0.000020472899,0.00001021987,0.0001185754],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9986363,0.000023584103,0.0006270162,0.00019833326,0.0001424273,0.0003723355],"domain_scores_gemma":[0.99823415,0.0012390319,0.00013938807,0.00020298577,0.00007762964,0.00010679487],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00069057016,0.00013707973,0.00039986672,0.00066416484,0.00008000618,0.000011086303,0.0001340851,0.00051603786,0.0000024735186],"category_scores_gemma":[0.011420104,0.000112405,0.000072409224,0.0007252735,0.00009416921,0.00006693127,0.00016152143,0.00042218424,0.000008024665],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017203529,0.0010806242,0.10051241,0.0031736232,0.00014351674,0.00007913187,0.003914983,0.000029738934,0.0018226405,0.0010305711,0.0020541768,0.8844382],"study_design_scores_gemma":[0.046188906,0.006959791,0.41013774,0.015792219,0.001237797,0.00024578202,0.02547699,0.47437105,0.00329277,0.0034649677,0.010925672,0.0019063152],"about_ca_topic_score_codex":0.000025627236,"about_ca_topic_score_gemma":0.000033230892,"teacher_disagreement_score":0.88253194,"about_ca_system_score_codex":0.00015699836,"about_ca_system_score_gemma":0.0003638274,"threshold_uncertainty_score":0.9969071},"labels":[],"label_agreement":null},{"id":"W4412759760","doi":"10.1007/s10459-025-10461-4","title":"Medical knowledge decline: the role of active usage","year":2025,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"123 Certification (Canada)","funders":"North Carolina Cotton Producers Association","keywords":"Certification; Specialty; Medical knowledge; Closeness; Test (biology); Knowledge level; Psychology; Medicine; Medical education; Family medicine; Political science; Mathematics education","score_opus":0.016383249624036124,"score_gpt":0.4677525693803103,"score_spread":0.45136931975627415,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4412759760","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.62476075,0.14505993,0.00048657384,0.07153081,0.003891373,0.0011228147,0.000003799721,0.00004048615,0.15310349],"genre_scores_gemma":[0.98964965,0.0062125465,0.0008199308,0.0029062892,0.00010368939,0.00004267443,0.0000033787376,0.0000018836114,0.00025993254],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9988421,0.00009755555,0.00040061164,0.00019830496,0.0002863115,0.00017508217],"domain_scores_gemma":[0.9905235,0.008933553,0.00015104083,0.00017447899,0.00009830022,0.00011908435],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0016990447,0.000055656546,0.00017745898,0.00011264253,0.00010771987,0.0000056949852,0.00019807766,0.000043664826,0.00005418343],"category_scores_gemma":[0.030531755,0.000034845732,0.000029916133,0.0009040343,0.00045884846,0.00008993394,0.000044672233,0.00016240822,0.0000060464927],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00002007736,0.00035657806,0.12665273,0.000051714327,0.000001431969,9.22506e-8,0.00044462856,0.000012833892,0.0000021420872,0.018097859,0.00024986162,0.85411006],"study_design_scores_gemma":[0.0010159098,0.00069429656,0.674899,0.011301576,0.00002133318,0.000012848107,0.008811099,0.0030704078,0.000676978,0.12720029,0.17215806,0.00013824279],"about_ca_topic_score_codex":0.0004756655,"about_ca_topic_score_gemma":0.00047392808,"teacher_disagreement_score":0.8539718,"about_ca_system_score_codex":0.00007552114,"about_ca_system_score_gemma":0.004596699,"threshold_uncertainty_score":0.9776345},"labels":[],"label_agreement":null},{"id":"W4412912501","doi":"10.1007/978-3-031-88752-9_3","title":"Remediation of Learners Who Perform Poorly on an OSCE","year":2025,"lang":"en","type":"book-chapter","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Environmental remediation; Psychology; Mathematics education; Environmental science; Contamination; Biology","score_opus":0.027388240493337056,"score_gpt":0.32679802290482063,"score_spread":0.2994097824114836,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4412912501","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004864875,0.00009848025,0.000019597777,0.0010611957,0.00027389263,0.00025114862,0.000016318914,0.00006155441,0.99335295],"genre_scores_gemma":[0.040179007,0.00062660745,0.00031361912,0.0017060209,0.00035024827,0.0000034538712,0.00027742976,0.00002947377,0.9565141],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9987242,0.00000933332,0.0004212194,0.00032145664,0.00039332715,0.00013043405],"domain_scores_gemma":[0.9957368,0.0033207505,0.00016673838,0.00044996158,0.00015850402,0.00016727089],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.000256961,0.0001967074,0.00054232654,0.00016413302,0.000022168859,0.0000067023147,0.00007886216,0.00040944386,0.0018564638],"category_scores_gemma":[0.011086984,0.00015467801,0.00017843864,0.000028225955,0.00007507682,0.000020959871,0.000024050589,0.00043047246,0.00013402376],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00092096365,0.00046422036,0.0032464238,0.0002995208,0.00031359406,0.000061941675,0.00023689868,0.000011132648,0.000029775705,0.407246,0.09454442,0.49262512],"study_design_scores_gemma":[0.02068566,0.034871265,0.065737724,0.1263128,0.00618594,0.000063787826,0.0006361706,0.0040641804,0.003032214,0.039882395,0.69496363,0.0035642027],"about_ca_topic_score_codex":0.00007575473,"about_ca_topic_score_gemma":0.000023922936,"teacher_disagreement_score":0.6004192,"about_ca_system_score_codex":0.00006350506,"about_ca_system_score_gemma":0.00026540185,"threshold_uncertainty_score":0.999056},"labels":[],"label_agreement":null},{"id":"W4413037049","doi":"10.1136/bmjoq-2025-003389","title":"Identifying diagnostic errors in the emergency department using trigger-based strategies","year":2025,"lang":"en","type":"article","venue":"BMJ Open Quality","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"Agency for Healthcare Research and Quality","keywords":"Emergency department; Medicine; Emergency medicine; Retrospective cohort study; Observational study; Intensive care unit; Medical emergency; Patient safety; Health care; Intensive care medicine; Internal medicine","score_opus":0.24793696067859325,"score_gpt":0.5530911256875662,"score_spread":0.30515416500897297,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413037049","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9813518,0.00035251683,0.0022191752,0.0050854385,0.00047536945,0.0028566413,0.0000146386365,0.00002880749,0.00761564],"genre_scores_gemma":[0.996865,0.000038869293,0.00081093976,0.0014975846,0.000057773672,0.0003266033,0.000045158475,0.00000922895,0.00034889072],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9972636,0.0008003508,0.0009104431,0.00035615772,0.00037889802,0.0002905581],"domain_scores_gemma":[0.9787859,0.019956151,0.00026912187,0.00078968896,0.00009920896,0.000099939585],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0051814583,0.00016080112,0.000467839,0.00007956118,0.00013539965,0.0001341948,0.00038852214,0.00009261945,0.00025144272],"category_scores_gemma":[0.07997041,0.00011148216,0.00014574513,0.00054042984,0.000071403745,0.000114675015,0.00013607963,0.00024475742,0.000027159203],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010737658,0.0036233233,0.92580855,0.00064055,0.00020126186,0.00045599323,0.0012690788,0.0009659161,0.00019547965,0.023605281,0.013810707,0.028350102],"study_design_scores_gemma":[0.0028180485,0.000101937774,0.98239803,0.0026826349,0.00019237334,0.0000056677163,0.0031320096,0.0006263463,0.00019554452,0.006885593,0.00073041173,0.00023142407],"about_ca_topic_score_codex":0.003048859,"about_ca_topic_score_gemma":0.0008545256,"teacher_disagreement_score":0.07478895,"about_ca_system_score_codex":0.00009682417,"about_ca_system_score_gemma":0.0007758036,"threshold_uncertainty_score":0.9277794},"labels":[],"label_agreement":null},{"id":"W4413041679","doi":"10.1111/acem.70120","title":"Response to the Commentary on Sheppard et al.'s Study of First Trimester POCUS Behaviors","year":2025,"lang":"en","type":"letter","venue":"Academic Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hospital for Sick Children; Memorial University of Newfoundland","funders":"Memorial University of Newfoundland","keywords":"Medicine; Obstetrics; Gynecology","score_opus":0.057170372735230186,"score_gpt":0.4227067748973601,"score_spread":0.3655364021621299,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413041679","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.1818181,0.0010221284,0.0000054828456,0.8102155,0.0036095467,0.0021699402,0.000073219126,0.00006301885,0.0010230662],"genre_scores_gemma":[0.032504253,0.0030339218,0.000012407216,0.9498218,0.0026688122,0.00030067028,0.0003175817,0.00008088909,0.011259626],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9936612,0.0009218277,0.0022989502,0.00092688174,0.0015993986,0.0005916996],"domain_scores_gemma":[0.9776386,0.019599352,0.0005592658,0.0016914734,0.00022549041,0.00028581885],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0035782752,0.00064112345,0.0015210345,0.0005288339,0.0001266037,0.0000010674459,0.0008561034,0.00085542747,0.0013193161],"category_scores_gemma":[0.046995923,0.0003880488,0.00029530717,0.000823497,0.00017591329,0.000027511222,0.00027153257,0.006029322,0.00007631083],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019050431,0.000503324,0.046375122,0.00005884997,0.0002759023,0.00022158868,0.001968972,0.000005149824,0.000024044428,0.0000011184426,0.9482998,0.00036109012],"study_design_scores_gemma":[0.0037031362,0.00679385,0.026622547,0.007821686,0.0016601479,0.000012764444,0.0006719885,0.000003036521,0.000021072186,0.000018162023,0.9523711,0.00030049108],"about_ca_topic_score_codex":0.0012512715,"about_ca_topic_score_gemma":0.000048560018,"teacher_disagreement_score":0.14931384,"about_ca_system_score_codex":0.00014445202,"about_ca_system_score_gemma":0.00018935034,"threshold_uncertainty_score":0.9998571},"labels":[],"label_agreement":null},{"id":"W4413097679","doi":"10.1016/j.ijnsa.2025.100393","title":"Decision support strategies for bedside nursing clinical reasoning: A scoping review","year":2025,"lang":"en","type":"review","venue":"International Journal of Nursing Studies Advances","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"Fundação para a Ciência e a Tecnologia","keywords":"Nursing; Clinical decision support system; Decision support system; Medicine; Psychology; Management science; Computer science; Artificial intelligence; Engineering","score_opus":0.19818199652296867,"score_gpt":0.6264267020805883,"score_spread":0.4282447055576196,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413097679","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000022356369,0.9798756,0.0050936937,0.0016319985,0.011143431,0.0010667146,0.00002849698,0.000026058739,0.0011317217],"genre_scores_gemma":[0.000004656041,0.9733097,0.023516644,0.00055006833,0.0023231127,0.000043787993,0.000042468295,0.000037956546,0.00017159477],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9944653,0.00017837592,0.0036210392,0.0004727053,0.0009283759,0.00033421587],"domain_scores_gemma":[0.9530324,0.040882427,0.0030615265,0.00029492276,0.0025173307,0.00021142833],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0020538932,0.00046102985,0.004304564,0.00038867045,0.00012207602,0.00008541493,0.000516996,0.00022461613,0.000020273052],"category_scores_gemma":[0.105977565,0.0003260628,0.0020297933,0.00026429933,0.00040844947,0.00028313082,0.000059124635,0.0007126315,0.000005013366],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00022795371,0.00043102662,0.000012872673,0.017032364,0.0013398333,0.00014526973,0.00006473358,0.0000037554871,1.2125039e-8,0.00032723253,0.018601831,0.9618131],"study_design_scores_gemma":[0.00093262224,0.0006587192,0.0000059111007,0.7863972,0.004768798,0.00089731335,0.0003281425,0.0000019193888,1.9033575e-7,0.0028686158,0.20294677,0.00019380539],"about_ca_topic_score_codex":4.4184898e-7,"about_ca_topic_score_gemma":0.000001501021,"teacher_disagreement_score":0.9616193,"about_ca_system_score_codex":0.00059272134,"about_ca_system_score_gemma":0.0028297254,"threshold_uncertainty_score":0.9999191},"labels":[],"label_agreement":null},{"id":"W4413402654","doi":"10.3390/educsci15081077","title":"Score Your Way to Clinical Reasoning Excellence: SCALENEo Online Serious Game in Physiotherapy Education","year":2025,"lang":"en","type":"article","venue":"Education Sciences","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Erasmus+; Université du Québec à Chicoutimi","keywords":"Excellence; Psychology; Medical education; Computer science; Applied psychology; Physical therapy; Medicine; Political science","score_opus":0.056480689669278876,"score_gpt":0.49843920876650954,"score_spread":0.44195851909723066,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413402654","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9688322,0.0013108718,0.0001094047,0.021577748,0.0032372363,0.00045058515,0.0000023731552,0.000044327444,0.004435269],"genre_scores_gemma":[0.97598195,0.00052297424,0.0066096038,0.008291796,0.00067875616,0.000070605834,0.000021048145,0.000008324906,0.007814958],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.998129,0.000101533384,0.0005798511,0.0005731833,0.00034570714,0.0002707158],"domain_scores_gemma":[0.99705595,0.0019912985,0.0001234904,0.00034769517,0.00022293319,0.0002586096],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010109729,0.00013426182,0.00030735054,0.00033221114,0.000111022295,0.000068880676,0.0002531204,0.00010413536,0.0001271154],"category_scores_gemma":[0.020886116,0.0001112265,0.0000876891,0.0015409152,0.00024350455,0.00010631682,0.00004078124,0.0002236805,0.00006981107],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000037997193,0.0019085758,0.4531464,0.000016779984,0.0000055051164,8.940179e-7,0.0004736854,0.000028028027,0.000102425416,0.0006304972,0.008161312,0.5354879],"study_design_scores_gemma":[0.00045393637,0.00027044286,0.98309153,0.0051000332,0.000017784172,0.000006970866,0.0023159883,0.0002110506,0.00014670206,0.0012770218,0.0069754114,0.00013315128],"about_ca_topic_score_codex":0.0005805704,"about_ca_topic_score_gemma":0.00009880752,"teacher_disagreement_score":0.53535473,"about_ca_system_score_codex":0.00010523667,"about_ca_system_score_gemma":0.003784141,"threshold_uncertainty_score":0.9873614},"labels":[],"label_agreement":null},{"id":"W4413685311","doi":"10.1007/s11096-025-01978-0","title":"Community pharmacists’ practices and clinical reasoning towards hospital discharge prescription: a study using simulations and retrospective think-aloud methodology","year":2025,"lang":"en","type":"article","venue":"International Journal of Clinical Pharmacy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"Université de Genève; Hôpitaux Universitaires de Genève","keywords":"Medical prescription; Medicine; Checklist; Pharmacy; Think aloud protocol; Qualitative research; Clinical pharmacy; Simulated patient; Patient safety; Family medicine; Nursing; Psychology; Health care","score_opus":0.30039277875288445,"score_gpt":0.6016528868628128,"score_spread":0.30126010810992837,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413685311","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98719156,0.00057279074,0.0017270519,0.0052538845,0.004319272,0.00041774375,0.000026133981,0.00002653618,0.00046503343],"genre_scores_gemma":[0.98725194,0.0012667223,0.00821551,0.0015403458,0.0015613049,0.0000013654269,0.0000081527705,0.00001640184,0.00013826114],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9938011,0.0026148914,0.0024267496,0.00034229318,0.0006229433,0.00019203249],"domain_scores_gemma":[0.959461,0.03532281,0.0026693481,0.00029138272,0.0017914008,0.0004640914],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0093981745,0.00021000157,0.0010169195,0.00020328617,0.00018928481,0.00012416935,0.0003412261,0.00013722177,0.000113051494],"category_scores_gemma":[0.23303014,0.00016968368,0.00035166022,0.00017295126,0.00042692252,0.00042463382,0.0003760934,0.0024279137,0.0000019092117],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0026450364,0.004624248,0.95627177,0.000013986261,0.0022604612,0.00021413346,0.0011976848,0.000014613498,0.00005273729,0.00022208496,0.00076066144,0.031722594],"study_design_scores_gemma":[0.01732763,0.0017442923,0.964985,0.00085716514,0.001875918,0.0003648832,0.0016031843,0.0077161086,0.00004401672,0.0024969832,0.00080606446,0.00017871804],"about_ca_topic_score_codex":0.00020935229,"about_ca_topic_score_gemma":0.000009099333,"teacher_disagreement_score":0.22363196,"about_ca_system_score_codex":0.0000893136,"about_ca_system_score_gemma":0.0005193168,"threshold_uncertainty_score":0.9998735},"labels":[],"label_agreement":null},{"id":"W4413748974","doi":"10.1136/bmjqs-2025-018723","title":"Bridging diagnostic safety and mental health: a systematic review highlighting inequities in autism spectrum disorder diagnosis","year":2025,"lang":"en","type":"review","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute for Work & Health; Toronto East General Hospital; Public Health Ontario; University of Toronto","funders":"Agency for Healthcare Research and Quality","keywords":"Medicine; Bridging (networking); Autism spectrum disorder; Mental health; Autism; Psychiatry; Computer security","score_opus":0.05234564209854585,"score_gpt":0.43478185899152455,"score_spread":0.3824362168929787,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413748974","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000006137032,0.9493641,0.00015554683,0.036674045,0.00055668876,0.012061904,0.0006075566,0.0001683592,0.0004056226],"genre_scores_gemma":[0.00006878069,0.9925209,0.00011462918,0.0033005471,0.00020231021,0.0016329145,0.0013405684,0.000087000524,0.0007323386],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9852701,0.0030029218,0.0085860975,0.0012749251,0.00086657115,0.0009994037],"domain_scores_gemma":[0.8847242,0.11060824,0.0025365364,0.0014483642,0.00006158596,0.0006210425],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.01040652,0.0010368022,0.011494825,0.00037266576,0.0003167766,0.00007338895,0.00035031495,0.0004613775,0.00012455214],"category_scores_gemma":[0.19916706,0.00080658233,0.0011390524,0.001046091,0.00020481522,0.000103010854,0.0005215219,0.0011395158,0.000062876796],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000030407744,0.00027794615,0.0007301577,0.83658826,0.00026506485,0.0000711795,0.0001577638,2.7545158e-7,1.2259276e-9,0.0020704803,0.0019017603,0.15790671],"study_design_scores_gemma":[0.0011830318,0.00013552871,0.0010106185,0.90827984,0.00229701,0.00014024305,0.00009485998,0.000007745728,1.8239787e-8,0.00008097259,0.08611321,0.0006569067],"about_ca_topic_score_codex":0.0015435006,"about_ca_topic_score_gemma":0.0004166962,"teacher_disagreement_score":0.18876053,"about_ca_system_score_codex":0.0010420256,"about_ca_system_score_gemma":0.0020375855,"threshold_uncertainty_score":0.9994385},"labels":[],"label_agreement":null},{"id":"W4413884223","doi":"10.1016/j.sapharm.2025.08.012","title":"Mapping the Processes of Pharmacist Therapeutic Reasoning: A Scoping Review and Development of the Pharmacist Therapeutic Reasoning Model","year":2025,"lang":"en","type":"review","venue":"Research in Social and Administrative Pharmacy","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Pharmacist; Medicine; Pharmacy; Nursing","score_opus":0.5800440273177146,"score_gpt":0.596887681615359,"score_spread":0.01684365429764434,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413884223","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0002554212,0.9932944,0.00005135799,0.0010601127,0.000054306685,0.0037741545,0.00007606265,0.000015684367,0.0014184723],"genre_scores_gemma":[0.01608125,0.9824097,0.00019023172,0.0003354792,0.00010940972,0.00060449744,0.000023711198,0.00002963109,0.0002161004],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9954152,0.0011700751,0.001409325,0.00061307254,0.0008281822,0.0005641557],"domain_scores_gemma":[0.98821354,0.0100401165,0.0006980781,0.00028744384,0.0005636372,0.00019720598],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0034262137,0.00045816024,0.0021095052,0.00015453836,0.0006067709,0.0000557848,0.00046498628,0.00013113607,0.000026257034],"category_scores_gemma":[0.006367796,0.00026232775,0.00026854244,0.0018136104,0.0012071092,0.00006890501,0.00038227133,0.0013447929,0.000001543312],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011846398,0.0003495941,0.00076164753,0.23182234,0.00050040643,0.000013645191,0.0016690169,2.0771229e-7,0.0000069221473,0.00031167516,0.00010762552,0.76433843],"study_design_scores_gemma":[0.0024058677,0.00009109427,0.00011106723,0.9539902,0.001568715,0.000071174414,0.0006348888,0.00013352103,0.00016986432,0.00036290596,0.040037893,0.00042280045],"about_ca_topic_score_codex":0.000017911552,"about_ca_topic_score_gemma":0.000010310003,"teacher_disagreement_score":0.76391566,"about_ca_system_score_codex":0.0001486112,"about_ca_system_score_gemma":0.010734708,"threshold_uncertainty_score":0.9999829},"labels":[],"label_agreement":null},{"id":"W4413892626","doi":"10.1080/0142159x.2025.2550477","title":"From struggle to strength: Embracing productive failure in clinical learning","year":2025,"lang":"en","type":"article","venue":"Medical Teacher","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"HealthForceOntario; The Wilson Centre; Office of the Chief Medical Examiner; University of Toronto","funders":"","keywords":"Psychology; Medical education; Political science; Medicine","score_opus":0.025127870015511075,"score_gpt":0.38804119988179886,"score_spread":0.36291332986628777,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413892626","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96212333,0.00017647175,0.0003666674,0.023938704,0.0005356675,0.00030315042,0.000001849479,0.00011492109,0.012439242],"genre_scores_gemma":[0.98629963,0.000020147192,0.0011134198,0.0017904912,0.0011834598,0.000032835393,0.00004254116,0.0000196993,0.0094977785],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9973455,0.000322608,0.0007655125,0.000647815,0.000559522,0.00035901525],"domain_scores_gemma":[0.9935009,0.0054038325,0.00006790102,0.00037962245,0.00006887113,0.00057886716],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0017555407,0.00017062342,0.0006074355,0.00012523064,0.00004412477,0.000018978533,0.00016890967,0.0003938702,0.0027668981],"category_scores_gemma":[0.29420015,0.00013773887,0.00014563266,0.00044807812,0.00013276396,0.000032789736,0.00016308422,0.0018384795,0.00021847982],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00033345015,0.0011380943,0.7102239,0.000017578488,0.000112921975,0.0002092714,0.0008708016,0.000019043873,0.000018846376,0.00015510821,0.034422938,0.252478],"study_design_scores_gemma":[0.010035441,0.00093494565,0.8662804,0.013048756,0.00029357223,0.000010932742,0.0041856538,0.006867146,0.0002010511,0.0023824396,0.09517741,0.00058229326],"about_ca_topic_score_codex":0.00044829858,"about_ca_topic_score_gemma":0.00016767671,"teacher_disagreement_score":0.2924446,"about_ca_system_score_codex":0.00008503949,"about_ca_system_score_gemma":0.00042771062,"threshold_uncertainty_score":0.9981447},"labels":[],"label_agreement":null},{"id":"W4413984515","doi":"10.1016/j.msksp.2025.103405","title":"Physiotherapy students' perceptions of how the Cervical Framework influences their clinical reasoning: A qualitative single case study of a Canadian university","year":2025,"lang":"en","type":"article","venue":"Musculoskeletal Science and Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Western University","funders":"","keywords":"Perception; Qualitative research; Medicine; Physical therapy; Medical education; Psychology","score_opus":0.0746272194390735,"score_gpt":0.4692969859226181,"score_spread":0.39466976648354457,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413984515","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9887147,0.00007926939,0.00009067225,0.00932546,0.000092424474,0.00039996413,0.000007436051,0.000009031758,0.0012810288],"genre_scores_gemma":[0.99822736,0.00017801265,0.0008308696,0.0006547478,0.00003230059,0.000002942421,6.0363055e-7,0.0000032587538,0.00006992855],"study_design_codex":"observational","study_design_gemma":"qualitative","domain_scores_codex":[0.99791336,0.00080300076,0.00026570854,0.00032261296,0.00049171015,0.00020358237],"domain_scores_gemma":[0.96669805,0.031717673,0.00021707725,0.00039576675,0.0006947617,0.0002766772],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0043311524,0.000102732374,0.00029913452,0.0002267224,0.00038917686,0.000050832416,0.00026860076,0.000082760496,0.000018973675],"category_scores_gemma":[0.08350444,0.000065436536,0.00009138273,0.0021062756,0.001503734,0.00024871426,0.00010635554,0.0003403962,9.543988e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011202238,0.01079897,0.7401522,0.00011848052,0.00077499374,0.0006512798,0.16973038,0.000013149001,0.0010430535,0.006158613,0.0007361365,0.06870255],"study_design_scores_gemma":[0.0009891406,0.0020939566,0.4624185,0.00042083015,0.00029413414,0.00009312331,0.5321404,0.00004737857,0.000009159599,0.00013946138,0.0012656841,0.000088267174],"about_ca_topic_score_codex":0.08887005,"about_ca_topic_score_gemma":0.01816298,"teacher_disagreement_score":0.36241,"about_ca_system_score_codex":0.0000603601,"about_ca_system_score_gemma":0.00086026586,"threshold_uncertainty_score":0.999753},"labels":[],"label_agreement":null},{"id":"W4414049707","doi":"10.54531/ppra5726","title":"A115 “Chest pain in simulation is always an MI” - Developing diagnostic reasoning and dispelling simulation myths with foundation trainees","year":2024,"lang":"en","type":"article","venue":"Journal of Healthcare Simulation","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medical diagnosis; Foundation (evidence); Debriefing; Clinical judgment; Simulated patient; Complaint; Medical history; Clinical judgement; Judgement; Cognition","score_opus":0.04973328534916665,"score_gpt":0.4003972875251147,"score_spread":0.35066400217594806,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4414049707","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.79509497,0.0012605191,0.1993397,0.0036381588,0.00020040033,0.00038506836,0.0000031783645,0.000048947564,0.0000290244],"genre_scores_gemma":[0.9906385,0.00023129719,0.007946985,0.0005360332,0.0005285338,0.0000035569612,0.000062804065,0.00004298133,0.000009307359],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99762124,0.00026114983,0.0010064638,0.00031932877,0.0005262554,0.000265553],"domain_scores_gemma":[0.9746947,0.023969827,0.00039489518,0.0001589798,0.0005142414,0.0002673248],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0021251373,0.00020741357,0.0004219834,0.00043189555,0.00011280811,0.0001346083,0.00004749002,0.00018962487,0.000023906865],"category_scores_gemma":[0.022385113,0.00017316031,0.0000753051,0.0005100559,0.00004264843,0.00063563784,0.000011192732,0.00045332397,0.0000036706717],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003207768,0.000071757684,0.039422873,0.0002892268,0.000031065723,0.000092061935,0.0024495556,0.7403506,0.000015979707,0.00020847094,0.0000041073263,0.21674353],"study_design_scores_gemma":[0.001175453,0.0012609715,0.09994879,0.009789826,0.00009969851,0.00003907163,0.0005250848,0.88563126,0.00001793572,0.00084916875,0.00048923865,0.00017348824],"about_ca_topic_score_codex":0.00011033986,"about_ca_topic_score_gemma":0.00010086261,"teacher_disagreement_score":0.21657003,"about_ca_system_score_codex":0.00032451694,"about_ca_system_score_gemma":0.0004877166,"threshold_uncertainty_score":0.98584974},"labels":[],"label_agreement":null},{"id":"W4414124985","doi":"10.1016/j.nepr.2025.104553","title":"Learning-by-concordance to support the transition-to-practice process of new graduate nurses in critical care: A descriptive interpretative study","year":2025,"lang":"en","type":"article","venue":"Nurse Education in Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Interdisciplinary Research in Rehabilitation; Université de Montréal","funders":"Fonds de Recherche du Québec - Santé; Fonds de Recherche du Québec-Société et Culture; Université de Montréal","keywords":"Process (computing); Graduate students; Descriptive research; Longitudinal study; Critical thinking; Patient care","score_opus":0.03222899003426125,"score_gpt":0.4813310627937144,"score_spread":0.44910207275945313,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4414124985","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8409538,0.0003480418,0.0019672855,0.13411173,0.0030591395,0.0026386536,0.0000104412275,0.00003993117,0.016870929],"genre_scores_gemma":[0.9878284,0.000021348278,0.0017026148,0.0078635495,0.00008405455,0.0003757084,0.000016984048,0.000016933036,0.0020904094],"study_design_codex":"qualitative","study_design_gemma":"qualitative","domain_scores_codex":[0.9976012,0.0006741761,0.00062102097,0.00045788442,0.0004049057,0.00024079082],"domain_scores_gemma":[0.96876395,0.029177746,0.00018686551,0.00033362655,0.0013130427,0.000224772],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010645358,0.00017166088,0.00035664462,0.00023833431,0.000061374354,0.000045072964,0.00017400713,0.00007237061,0.000042710777],"category_scores_gemma":[0.33726078,0.00014698728,0.000047628113,0.001245534,0.00010554664,0.0004075043,0.000016598167,0.0007568293,0.000042121264],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0067136097,0.013696515,0.038275726,0.0001324538,0.00013820762,0.000048838232,0.78837556,0.0004802235,0.00007771216,0.0013002258,0.048214834,0.1025461],"study_design_scores_gemma":[0.0017265005,0.0031036502,0.025314664,0.0024789392,0.0006146231,0.000031964384,0.95765114,0.0000483205,0.00038401553,0.00032842203,0.008144053,0.00017368021],"about_ca_topic_score_codex":0.0023653393,"about_ca_topic_score_gemma":0.00018832198,"teacher_disagreement_score":0.33619624,"about_ca_system_score_codex":0.00026461916,"about_ca_system_score_gemma":0.0032149716,"threshold_uncertainty_score":0.6683217},"labels":[],"label_agreement":null},{"id":"W4414367144","doi":"10.1515/dx-2025-0120","title":"The Society to Improve Diagnosis in Medicine’s legacy: building a foundation for diagnostic excellence","year":2025,"lang":"en","type":"article","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Sinai Health System","funders":"Australian Research Council; American Heart Association; Shionogi; Agency for Healthcare Research and Quality; University of Toronto; Gordon and Betty Moore Foundation","keywords":"Excellence; Harm; Transformative learning; Health care; Foundation (evidence); Best practice; Do no harm; Patient safety","score_opus":0.026930617238405646,"score_gpt":0.37468171913436626,"score_spread":0.3477511018959606,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4414367144","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7216433,0.0057412954,0.010602106,0.25048274,0.003491098,0.0058668056,0.0000379734,0.0002411633,0.0018934746],"genre_scores_gemma":[0.96757203,0.00729011,0.002992081,0.009973205,0.0005427518,0.010874915,0.00003583356,0.00004111309,0.00067798357],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99745387,0.00007309635,0.000777907,0.00069576903,0.000364437,0.0006349382],"domain_scores_gemma":[0.8725423,0.12622324,0.00013132863,0.00059212913,0.0002574872,0.00025352478],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001386459,0.00029115815,0.0005924242,0.00016008847,0.00035600323,0.00011459819,0.0003127045,0.00018836434,0.00007166627],"category_scores_gemma":[0.22048114,0.00020961545,0.00038257404,0.00093032024,0.0002198193,0.00011776561,0.00014149082,0.00031107827,0.00003970517],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00019001072,0.00062157883,0.43105137,0.00016333346,0.00018328102,0.000015975938,0.00069379556,0.00006509427,0.00012331325,0.0063421032,0.18609695,0.3744532],"study_design_scores_gemma":[0.011297472,0.0023778137,0.5454391,0.021793207,0.0013493109,0.000006791355,0.0022684906,0.0023278692,0.0068089534,0.017293198,0.38809893,0.00093881093],"about_ca_topic_score_codex":0.0008300997,"about_ca_topic_score_gemma":0.0002703617,"teacher_disagreement_score":0.37351435,"about_ca_system_score_codex":0.00029720235,"about_ca_system_score_gemma":0.00023748427,"threshold_uncertainty_score":0.854787},"labels":[],"label_agreement":null},{"id":"W4414389563","doi":"10.1515/dx-2025-0106","title":"The importance of cognition for improving diagnostic safety: Salerno redux?","year":2025,"lang":"en","type":"review","venue":"Diagnosis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Cognition; Rationality; Dual process theory (moral psychology); Psychological intervention; Cognitive bias; Dual (grammatical number); Need for cognition","score_opus":0.043708551821598586,"score_gpt":0.3874983498736052,"score_spread":0.3437897980520066,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4414389563","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000027521071,0.99309087,0.00011897227,0.0006371544,0.00076512387,0.0034566415,0.0007309274,0.0000706677,0.0011021382],"genre_scores_gemma":[0.00013375691,0.99296397,0.00031119006,0.00031932318,0.00044775903,0.0041521676,0.0009342046,0.000069473404,0.00066815084],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9969276,0.00012513493,0.0015456952,0.00060722255,0.00033890738,0.00045543315],"domain_scores_gemma":[0.7595283,0.23843715,0.0008188719,0.0007329562,0.00031116343,0.00017156613],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0009332334,0.00044792405,0.0022477827,0.0001390875,0.00020420495,0.00003815735,0.00030199348,0.00046233673,0.00005545412],"category_scores_gemma":[0.34430477,0.0002943369,0.001306069,0.00046819434,0.00019902902,0.00003168075,0.00011958114,0.0004239367,0.000027028535],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004412794,0.00023127419,0.0013083004,0.01743945,0.00028395298,0.000017078964,0.000008886168,1.5505138e-7,5.6209007e-8,0.0003845434,0.011282349,0.9689998],"study_design_scores_gemma":[0.0010363901,0.00036676528,0.00073080993,0.19053909,0.009158911,0.000016531096,0.00002056298,0.0000049612377,0.000009036625,0.0004041305,0.79739046,0.00032232993],"about_ca_topic_score_codex":0.000025281812,"about_ca_topic_score_gemma":0.000019306326,"teacher_disagreement_score":0.9686775,"about_ca_system_score_codex":0.00012784287,"about_ca_system_score_gemma":0.0009276109,"threshold_uncertainty_score":0.9999509},"labels":[],"label_agreement":null},{"id":"W4414431571","doi":"10.1080/13803395.2025.2561162","title":"The AI inflection point in clinical neuropsychology: a call to action","year":2025,"lang":"en","type":"article","venue":"Journal of Clinical and Experimental Neuropsychology","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Abbotsford Veterinary Clinic","funders":"University of Sydney","keywords":"Leverage (statistics); Neuropsychology; Call to action; Key (lock); Point (geometry); Quality (philosophy)","score_opus":0.06582309699880128,"score_gpt":0.522062404140061,"score_spread":0.4562393071412597,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4414431571","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9510586,0.00046509338,0.00011917501,0.039573647,0.006219589,0.00022861417,5.868285e-7,0.000014704294,0.0023199946],"genre_scores_gemma":[0.9580785,0.0011342418,0.0002027895,0.039341915,0.00066137174,0.000007473645,6.0774215e-7,0.000013499155,0.0005596136],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9958582,0.0007828808,0.0025013294,0.0003958181,0.00018724045,0.00027451976],"domain_scores_gemma":[0.99042207,0.008366432,0.0003905269,0.00030137144,0.00014506547,0.00037453583],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0025489924,0.00016053715,0.00076807634,0.0001780831,0.000087424494,0.000031487507,0.00016765644,0.00027144048,0.000029416538],"category_scores_gemma":[0.027003797,0.00010359546,0.0003460966,0.00033329087,0.0004450536,0.00006302928,0.00009245114,0.0014976207,0.000025613732],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.01635518,0.0039236676,0.73153657,0.00000830967,0.00014698527,0.00063407415,0.000101370555,0.000008795791,0.0028048116,0.0004969552,0.04887932,0.19510397],"study_design_scores_gemma":[0.006480578,0.006424288,0.9502274,0.00024976328,0.000053616586,0.00068861985,0.0000794731,0.00003459144,0.00022382732,0.0004847826,0.034967724,0.00008538151],"about_ca_topic_score_codex":0.000011818621,"about_ca_topic_score_gemma":0.000010496111,"teacher_disagreement_score":0.2186908,"about_ca_system_score_codex":0.000020427726,"about_ca_system_score_gemma":0.00011456268,"threshold_uncertainty_score":0.9811922},"labels":[],"label_agreement":null},{"id":"W4414447213","doi":"10.1542/peds.2025-072981","title":"Enhancing Diagnostic Reasoning in Medical Education Through Patient Stories and Illness Scripts","year":2025,"lang":"en","type":"article","venue":"PEDIATRICS","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Preceptor; Scripting language; Context (archaeology); Cornerstone; Medical history; Disease; Alternative medicine; Graduate medical education","score_opus":0.01011476942999329,"score_gpt":0.3314567752379545,"score_spread":0.32134200580796124,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4414447213","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.982424,0.009632084,0.00070705736,0.0023288878,0.0016328332,0.00026280573,0.0000026715743,0.00005171873,0.002957906],"genre_scores_gemma":[0.99123865,0.0041869027,0.0012683489,0.0026122502,0.0003784363,0.000054461165,0.000029086714,0.000011799445,0.00022007985],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9986372,0.000054444652,0.00040901,0.00029633148,0.0003658466,0.0002371896],"domain_scores_gemma":[0.9881355,0.011243423,0.00008033926,0.00020623104,0.00013473052,0.00019978204],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00037512425,0.00012795013,0.00028639528,0.00013633938,0.00007838952,0.000025153799,0.00005921693,0.00017350446,0.00003948753],"category_scores_gemma":[0.23219162,0.00011537294,0.00003922032,0.00055762235,0.00007153421,0.000068536625,0.000083866245,0.00031601405,0.000008489383],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000061227314,0.0006762347,0.8717465,0.0002491338,0.00001202322,0.000085450614,0.002566916,0.0000040595264,0.000004836623,0.0044463417,0.006102408,0.114044905],"study_design_scores_gemma":[0.0060404968,0.0007302984,0.9307981,0.017542904,0.0007972004,0.0001408384,0.007927939,0.0005191755,0.0009533987,0.008716281,0.024902469,0.00093090604],"about_ca_topic_score_codex":0.0004946825,"about_ca_topic_score_gemma":0.00013460779,"teacher_disagreement_score":0.23181649,"about_ca_system_score_codex":0.00010478981,"about_ca_system_score_gemma":0.0012109311,"threshold_uncertainty_score":0.77427596},"labels":[],"label_agreement":null},{"id":"W4414481220","doi":"10.1097/pts.0000000000001409","title":"Incorporating Machine Learning Driven Factors in the Design of Electronic-triggers to Detect Diagnostic Errors in the Emergency Department","year":2025,"lang":"en","type":"article","venue":"Journal of Patient Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Emergency department; Identification (biology); Construct (python library); Diagnostic accuracy; Medical diagnosis","score_opus":0.019385671926458247,"score_gpt":0.3131289478270114,"score_spread":0.29374327590055316,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4414481220","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99362487,0.0005851788,0.0030651058,0.0017872659,0.00018613196,0.0006118035,0.0000026971363,0.000003871952,0.00013308274],"genre_scores_gemma":[0.9986649,0.00037762057,0.0005742707,0.00031864303,0.00002749916,0.000012871656,0.0000055817186,0.000007777904,0.000010828163],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99735075,0.00066774315,0.0011234516,0.0001283279,0.00047385372,0.00025586458],"domain_scores_gemma":[0.9827193,0.016381677,0.00054496335,0.00017614492,0.00010815987,0.000069751964],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0020302418,0.00014084173,0.00039963538,0.00023596028,0.00006531739,0.0000062554564,0.00022202138,0.000060461774,0.000024227205],"category_scores_gemma":[0.051921554,0.00007367366,0.00016008901,0.0007115461,0.00003001084,0.000036544167,0.000034092187,0.0007151058,0.0000016165436],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006668315,0.00048703208,0.9254826,0.00002288929,0.000073285075,0.000091325186,0.0032424666,0.062092945,0.00014012685,0.00014521983,0.00036627543,0.0071890573],"study_design_scores_gemma":[0.00297892,0.0064123175,0.9790574,0.0029235762,0.00028375123,0.000036900583,0.003682123,0.0017434409,0.00067372125,0.0011540316,0.0008716906,0.00018218382],"about_ca_topic_score_codex":0.00010952719,"about_ca_topic_score_gemma":0.00016591139,"teacher_disagreement_score":0.060349505,"about_ca_system_score_codex":0.0001835931,"about_ca_system_score_gemma":0.00025108762,"threshold_uncertainty_score":0.9560645},"labels":[],"label_agreement":null},{"id":"W4414499338","doi":"10.1056/aidbp2500120","title":"Assessment of Large Language Models in Clinical Reasoning: A Novel Benchmarking Study","year":2025,"lang":"en","type":"article","venue":"NEJM AI","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":28,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Benchmarking; Language model; Quality (philosophy); Identification (biology); Measure (data warehouse)","score_opus":0.037095345024797416,"score_gpt":0.45895407758977547,"score_spread":0.4218587325649781,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4414499338","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97598404,0.00015430199,0.0041954103,0.0005577447,0.00028516655,0.00047571378,0.000009428488,0.000034082463,0.018304097],"genre_scores_gemma":[0.9949498,0.000025606976,0.002928823,0.001478453,0.000106156076,0.000025427427,0.000017320583,0.000010828687,0.00045754146],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982597,0.00009676005,0.0007728121,0.00034451243,0.00028259013,0.00024361476],"domain_scores_gemma":[0.9948824,0.0043813894,0.000113913346,0.0004301701,0.00008817848,0.000103970546],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0018241976,0.000118778815,0.0005705086,0.00013437695,0.000025239138,0.000009989755,0.00010691305,0.0001307134,0.00006355077],"category_scores_gemma":[0.011510427,0.00010152879,0.00014454275,0.00035221962,0.000051654813,0.0000374901,0.00011672905,0.00047158616,0.000002757449],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008877677,0.0057655,0.9790066,0.000030334655,0.000088742054,0.00008244335,0.0005864494,0.00010608541,0.000057850906,0.0059099146,0.0008411314,0.0074362005],"study_design_scores_gemma":[0.0068835816,0.0006166971,0.9599111,0.0020498596,0.00013652132,0.0000025263796,0.0012838815,0.028689492,0.000017366641,0.00017744709,0.00014393147,0.0000875621],"about_ca_topic_score_codex":0.00028033066,"about_ca_topic_score_gemma":0.00013706814,"teacher_disagreement_score":0.028583407,"about_ca_system_score_codex":0.000053174314,"about_ca_system_score_gemma":0.00034271696,"threshold_uncertainty_score":0.99681604},"labels":[],"label_agreement":null},{"id":"W4414529017","doi":"10.1007/978-981-95-0673-6","title":"Landmark Studies in Critical Care","year":2025,"lang":"en","type":"book","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Critical appraisal; Landmark; Relevance (law); Subject (documents); Evidence-based medicine; Clinical Practice; Alternative medicine; MEDLINE; Clinical trial","score_opus":0.10345162748903479,"score_gpt":0.482761130956223,"score_spread":0.37930950346718817,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4414529017","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00009582226,0.01985992,0.000004877808,0.0017522032,0.0006337047,0.00019520108,0.000019148129,0.000059960148,0.97737914],"genre_scores_gemma":[0.0030769482,0.0014004929,0.00035550212,0.0020482838,0.000373464,0.000020389334,0.00012219179,0.0000145101,0.9925882],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9989857,0.000020323958,0.00031662642,0.00030531464,0.00019083396,0.00018118926],"domain_scores_gemma":[0.9721731,0.027223708,0.000020525375,0.00026943785,0.0002098585,0.00010334906],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00012321284,0.00016940897,0.00070446293,0.00011772373,0.000016879016,0.0000061589094,0.000053769745,0.00032452407,0.00030580064],"category_scores_gemma":[0.106602654,0.00012430435,0.00013871735,0.00005296723,0.00016615044,0.0000066254497,0.000080400154,0.00046484036,0.0001185355],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000086198364,0.00010741643,0.0050306697,0.001511193,0.00012467711,0.0010721069,0.00034219,2.8387257e-7,9.0405e-8,0.020066826,0.95878524,0.012873121],"study_design_scores_gemma":[0.0056289835,0.0011091635,0.007284714,0.1405269,0.0018449465,0.00006835858,0.0029355409,0.0000108157155,0.000015127612,0.028707458,0.8109987,0.00086929306],"about_ca_topic_score_codex":0.00002119468,"about_ca_topic_score_gemma":0.00007511713,"teacher_disagreement_score":0.14778653,"about_ca_system_score_codex":0.00021285647,"about_ca_system_score_gemma":0.00073458115,"threshold_uncertainty_score":0.90092283},"labels":[],"label_agreement":null},{"id":"W4414564301","doi":"10.1212/ne9.0000000000200250","title":"Education Research: Creating Online Interactive Case-Based Learning Experiences From Educational Case Reports With Large Language Models","year":2025,"lang":"en","type":"article","venue":"Neurology Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Frontier; Educational technology; Nondeterministic algorithm; Interactive Learning; Process (computing); Language acquisition; Online learning","score_opus":0.039444007374454165,"score_gpt":0.4376303828323319,"score_spread":0.3981863754578777,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4414564301","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98599154,0.00046056873,0.00040431664,0.008104147,0.0010078413,0.00042822497,0.000007873234,0.000057276033,0.0035381927],"genre_scores_gemma":[0.9903764,0.000008623799,0.003189836,0.0033364722,0.0005241774,0.00041202808,0.00066123594,0.00001938046,0.0014718331],"study_design_codex":"observational","study_design_gemma":"qualitative","domain_scores_codex":[0.998178,0.0003658647,0.00038125055,0.00057433365,0.00022529229,0.00027525006],"domain_scores_gemma":[0.9897854,0.008882385,0.0001969968,0.000387988,0.00058864173,0.00015862798],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00047438714,0.0001479933,0.00023407138,0.00032605656,0.00034074284,0.000042018568,0.0000495209,0.00014750804,0.00028120392],"category_scores_gemma":[0.02198901,0.00013003017,0.000049942737,0.00043683616,0.00014623854,0.00012500588,0.000036056234,0.00079026044,0.000007867936],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0033824726,0.029274397,0.6494992,0.00020517322,0.00032494494,0.031183178,0.097526476,0.0016831274,0.000747263,0.0033266102,0.03451152,0.14833564],"study_design_scores_gemma":[0.0064414586,0.00534794,0.19952053,0.0077262865,0.0011574151,0.15392008,0.54042095,0.062049095,0.0019990862,0.007843182,0.012244656,0.0013292964],"about_ca_topic_score_codex":0.0053463127,"about_ca_topic_score_gemma":0.0006908485,"teacher_disagreement_score":0.44997868,"about_ca_system_score_codex":0.00009826897,"about_ca_system_score_gemma":0.0042361985,"threshold_uncertainty_score":0.9862492},"labels":[],"label_agreement":null},{"id":"W4414620189","doi":"10.1093/jmp/jhaf029","title":"Pellegrino and Thomasma’s Anatomy of Clinical Judgments Revisited","year":2025,"lang":"en","type":"article","venue":"The Journal of Medicine and Philosophy A Forum for Bioethics and Philosophy of Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Realm; Situated; Process (computing); Clinical Practice; Work (physics); Clinical judgment","score_opus":0.1286161754053852,"score_gpt":0.45325890423457477,"score_spread":0.32464272882918954,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4414620189","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.124201655,0.07719898,0.000437378,0.7872979,0.00085715577,0.00096380047,0.000057994213,0.000013280359,0.008971848],"genre_scores_gemma":[0.94960594,0.03964594,0.0005609421,0.008566733,0.0015268322,0.0000029303442,0.000015331192,0.000019768966,0.000055607463],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99611825,0.00025809996,0.0023790295,0.00028705038,0.000695621,0.00026196608],"domain_scores_gemma":[0.98148525,0.015498875,0.0013710945,0.00038982977,0.0008417461,0.00041318606],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.007021915,0.00032202934,0.0020317903,0.00041749238,0.00014400372,0.0000044240974,0.00022619047,0.00036129288,0.000022101745],"category_scores_gemma":[0.020816814,0.00017159816,0.00022732245,0.00038100535,0.006800794,0.00006054354,0.000119662895,0.0009952628,1.7626411e-7],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.017385049,0.0010757657,0.08982448,0.0056741307,0.0026387074,0.00007056673,0.002747206,0.0000017265905,0.0005966754,0.75614625,0.031154111,0.09268534],"study_design_scores_gemma":[0.031616736,0.018877871,0.010323275,0.036615085,0.0040839664,0.00042695747,0.001402871,0.00019307907,0.00022535045,0.8877067,0.0082912715,0.0002368647],"about_ca_topic_score_codex":0.00008746731,"about_ca_topic_score_gemma":0.0000033207448,"teacher_disagreement_score":0.8254043,"about_ca_system_score_codex":0.000010101028,"about_ca_system_score_gemma":0.00019188742,"threshold_uncertainty_score":0.9959021},"labels":[],"label_agreement":null},{"id":"W4414741071","doi":"10.1093/clinchem/hvaf086.172","title":"A-178 Evaluating a Patient-Centered Education Initiative in Preventing Pre-analytical Errors and Improving Patient Satisfaction in Clinical Diagnostic Laboratories","year":2025,"lang":"en","type":"article","venue":"Clinical Chemistry","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; Sunnybrook Health Science Centre; Hospital for Sick Children; Sunnybrook Hospital; University of Toronto","funders":"","keywords":"Patient satisfaction; Patient education; Demographics; General partnership; Patient safety; MEDLINE; Telephone survey; Patient experience","score_opus":0.04448833690896816,"score_gpt":0.4336115253167501,"score_spread":0.389123188407782,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4414741071","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9961823,0.00067498477,0.0000177431,0.0009412295,0.0006470317,0.0006369258,0.000011424886,0.00004276397,0.00084554486],"genre_scores_gemma":[0.9976643,0.00020531203,0.000681032,0.0009874481,0.00019172288,0.00011661644,0.00004774936,0.000019479012,0.00008637457],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99556494,0.000405366,0.0025433928,0.00080584706,0.0003155082,0.0003649737],"domain_scores_gemma":[0.9656337,0.032751936,0.000546542,0.00043280565,0.00031136087,0.00032364836],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0018728959,0.0002474951,0.0007987326,0.000082373226,0.00006062668,0.000043453983,0.000075647054,0.00046096114,0.000048409842],"category_scores_gemma":[0.37563562,0.00023820081,0.00018839641,0.0005023524,0.00027625516,0.00010214112,0.00018982134,0.0011772548,0.000005191061],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00023163404,0.001207856,0.8005754,0.00016178147,0.000041260697,0.000022269345,0.00014719501,0.0000040113996,0.000030751835,0.000020269108,0.00022614218,0.19733144],"study_design_scores_gemma":[0.00313776,0.0005052708,0.98591405,0.0055610402,0.00022011287,0.000007759328,0.001034406,0.0024846904,0.0003184045,0.00051184907,0.000093610266,0.00021107197],"about_ca_topic_score_codex":0.00009117703,"about_ca_topic_score_gemma":0.00003845091,"teacher_disagreement_score":0.37376273,"about_ca_system_score_codex":0.00018868847,"about_ca_system_score_gemma":0.0012635549,"threshold_uncertainty_score":0.9713547},"labels":[],"label_agreement":null},{"id":"W4414904528","doi":"10.2196/68981","title":"Comparison of Physician Assistant and Medical Students' Clinical Reasoning Processes Using an Online Patient Simulation Tool to Support Clinical Reasoning (eCREST): Mixed Methods Study","year":2025,"lang":"en","type":"article","venue":"JMIR Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Case-based reasoning; Simulated patient; MEDLINE; Model-based reasoning; Virtual patient; Medical diagnosis; Clinical judgment; Decision support system","score_opus":0.09598065234673224,"score_gpt":0.593249595228026,"score_spread":0.49726894288129375,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4414904528","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9885225,0.00022992489,0.0063889083,0.001292011,0.0016291116,0.0017167919,0.000008045825,0.00009624363,0.00011640517],"genre_scores_gemma":[0.9714718,0.000057290457,0.023159947,0.004120591,0.00080204784,0.00011862417,0.00018190549,0.00003875134,0.000049046117],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9911085,0.0015876039,0.0033116238,0.001039861,0.0025151155,0.00043729265],"domain_scores_gemma":[0.975603,0.019211408,0.0009991797,0.0009611418,0.0012277594,0.0019975554],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.008118952,0.00035061527,0.0015745738,0.0002549359,0.0001688021,0.00007724176,0.00043875456,0.0006053238,0.00014625558],"category_scores_gemma":[0.2888861,0.00029796,0.00019465439,0.0009552577,0.00035287594,0.00014540557,0.00038687838,0.0011030682,0.0000048163392],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004681132,0.023861255,0.51246804,0.00010246516,0.0001242824,0.000011386762,0.0010977669,0.00003179296,0.0000042698794,0.000040946255,0.0007545418,0.46103516],"study_design_scores_gemma":[0.0038818384,0.005592856,0.9404804,0.008007948,0.0006817899,0.000011299509,0.00903508,0.02975177,0.000039936593,0.000068912086,0.0021077842,0.00034038027],"about_ca_topic_score_codex":0.00013104235,"about_ca_topic_score_gemma":0.00010080309,"teacher_disagreement_score":0.4606948,"about_ca_system_score_codex":0.00013417886,"about_ca_system_score_gemma":0.0063933525,"threshold_uncertainty_score":0.99994725},"labels":[],"label_agreement":null},{"id":"W4415300115","doi":"10.1186/s13643-025-02937-6","title":"Conceptualizations of anaesthetists’ clinical reasoning expertise: protocol for a systematic review and qualitative thematic synthesis","year":2025,"lang":"en","type":"review","venue":"Systematic Reviews","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hôpital Maisonneuve-Rosemont","funders":"","keywords":"Protocol (science); Thematic analysis; Qualitative research; Thematic map; MEDLINE; Data collection","score_opus":0.2352444189535031,"score_gpt":0.5720743686130343,"score_spread":0.33682994965953117,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4415300115","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[8.992227e-9,0.56288207,0.001166482,0.00003898605,0.000043650536,0.4355287,0.000049976366,0.000034757068,0.00025535695],"genre_scores_gemma":[7.480076e-8,0.5148154,0.0021793924,0.00016025674,0.00003132521,0.48205346,0.000030994477,0.00004889058,0.00068022683],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9720093,0.011942689,0.014102368,0.00095414737,0.00058107066,0.0004104475],"domain_scores_gemma":[0.796866,0.1854817,0.013177176,0.0030631947,0.0009038632,0.0005080506],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.020318463,0.0010420637,0.031539597,0.0002630598,0.000113116854,0.000049499762,0.00048334428,0.0005483121,0.0000688985],"category_scores_gemma":[0.66145563,0.00055525795,0.0035057287,0.0008438487,0.00032556537,0.00005937065,0.00012283592,0.00038372126,0.000064563195],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000057556326,0.00024221656,5.689642e-7,0.98291194,0.0008403648,0.0000041323738,0.0005182504,5.782318e-9,3.0131921e-9,0.0013998523,0.003357748,0.010719182],"study_design_scores_gemma":[0.00055600196,0.0002700217,8.724111e-8,0.938911,0.04174509,0.00014051965,0.00035933283,0.000012126349,5.760717e-8,0.00005137686,0.017548705,0.00040565728],"about_ca_topic_score_codex":0.000005265768,"about_ca_topic_score_gemma":0.0000012626879,"teacher_disagreement_score":0.6411372,"about_ca_system_score_codex":0.00014967848,"about_ca_system_score_gemma":0.0008520488,"threshold_uncertainty_score":0.9996899},"labels":[],"label_agreement":null},{"id":"W4415526077","doi":"10.2196/85624","title":"Peer Review of “Interactive Evaluation of an Adaptive-Questioning Symptom Checker Using Standardized Clinical Vignettes (Preprint)”","year":2025,"lang":"en","type":"article","venue":"JMIRx Med","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"MEDLINE; Reliability (semiconductor); Data collection; Quality of life (healthcare)","score_opus":0.11259048029433501,"score_gpt":0.5112773889591795,"score_spread":0.3986869086648445,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4415526077","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9794495,0.0036079467,0.0035022544,0.004538385,0.00047540167,0.0015324996,0.000023045159,0.000048441387,0.0068224883],"genre_scores_gemma":[0.985765,0.00072268455,0.011424382,0.0008641083,0.0001267796,0.00005284828,0.00007054072,0.000016913997,0.0009567645],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99658686,0.0006618467,0.0010857888,0.00034132175,0.0011795487,0.00014461915],"domain_scores_gemma":[0.98745114,0.006270525,0.00053800654,0.0005273742,0.005075017,0.00013792001],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.009146905,0.00013181321,0.0008122562,0.00008953255,0.000030298528,0.000007147511,0.00009215411,0.00013065404,0.00037163112],"category_scores_gemma":[0.2332034,0.0001101317,0.0002509665,0.00026670797,0.0001426472,0.00007762144,0.00005867264,0.00027484348,0.0000050046506],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008984725,0.0057758023,0.16203238,0.0046994393,0.0025808418,0.00003761126,0.0013058237,0.0006121723,0.006565115,0.002947445,0.03387862,0.77058],"study_design_scores_gemma":[0.019415567,0.0024014527,0.5545044,0.33619845,0.0066320417,0.000029160457,0.0007192009,0.054560993,0.017590439,0.0021949508,0.0052466034,0.0005067589],"about_ca_topic_score_codex":0.00007394228,"about_ca_topic_score_gemma":0.0000047549065,"teacher_disagreement_score":0.77007324,"about_ca_system_score_codex":0.00012966759,"about_ca_system_score_gemma":0.00075776334,"threshold_uncertainty_score":0.77325565},"labels":[],"label_agreement":null},{"id":"W4415619271","doi":"10.1002/hem3.70232","title":"Choosing words wisely: Language, metaphor, and psychological challenges after pulmonary embolism","year":2025,"lang":"en","type":"article","venue":"HemaSphere","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Queen's University","funders":"Wellcome Trust","keywords":"Hypervigilance; Anxiety; Depression (economics); Mental health; Normative; Population; Cohort; Perspective (graphical)","score_opus":0.03149613265867566,"score_gpt":0.35911287236921224,"score_spread":0.32761673971053656,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4415619271","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.84080046,0.07571125,0.000055503344,0.006683902,0.0002515507,0.00019954589,0.0000037237896,0.00012589026,0.07616819],"genre_scores_gemma":[0.98689985,0.0040650913,0.0011979572,0.0021250292,0.00019082389,0.000047231708,0.0000071677714,0.000015916776,0.0054509044],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9989177,0.000045175177,0.00023095387,0.00040324693,0.00015278261,0.00025014914],"domain_scores_gemma":[0.9979661,0.0014480883,0.00003477815,0.0003343941,0.000039909988,0.00017669245],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00018454743,0.00016338688,0.00038167194,0.00003926426,0.00004260906,0.00002014937,0.000058693116,0.0001942331,0.0005787225],"category_scores_gemma":[0.0037964254,0.00012287407,0.00010322676,0.00014751672,0.00011015962,0.000032020846,0.00007020709,0.00026664202,0.000049992144],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00077882595,0.0014691303,0.03504526,0.00028188239,0.00021124152,0.0016190337,0.0010485334,7.717537e-7,0.0004399109,0.0034841963,0.005819192,0.94980204],"study_design_scores_gemma":[0.002881085,0.00037554442,0.9621759,0.00470212,0.00045355034,0.00033822763,0.0025869103,0.00011317012,0.0003678102,0.0039488766,0.021633666,0.0004231552],"about_ca_topic_score_codex":0.000014829685,"about_ca_topic_score_gemma":0.000009081275,"teacher_disagreement_score":0.94937885,"about_ca_system_score_codex":0.00001675122,"about_ca_system_score_gemma":0.000026480138,"threshold_uncertainty_score":0.63366055},"labels":[],"label_agreement":null},{"id":"W4415926074","doi":"10.4088/jcp.25m15979","title":"Targeting Intolerance of Uncertainty During Pregnancy","year":2025,"lang":"en","type":"article","venue":"The Journal of Clinical Psychiatry","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; St. Joseph’s Healthcare Hamilton","funders":"","keywords":"Pregnancy; MEDLINE; Gestation; Disease","score_opus":0.027798696922693933,"score_gpt":0.4060178783163008,"score_spread":0.3782191813936069,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4415926074","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97262824,0.007764148,0.00048606854,0.01325422,0.0032027622,0.0001418576,0.0000018608755,0.00001145885,0.002509377],"genre_scores_gemma":[0.99361026,0.002179035,0.002119056,0.000871803,0.0007732782,7.334229e-7,6.3265105e-7,0.0000094186125,0.00043578228],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9968858,0.00024738873,0.0022560135,0.00011874345,0.0003162628,0.0001758253],"domain_scores_gemma":[0.9863445,0.011645787,0.0011496736,0.00035356986,0.00035356655,0.0001529128],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0021517195,0.0001124452,0.00075789506,0.0000656439,0.000061135805,0.000005506223,0.0002724218,0.00013964732,0.00006047509],"category_scores_gemma":[0.046906445,0.000064171865,0.00046627544,0.0002554921,0.00022075417,0.000037239082,0.00005494544,0.00089212507,0.0000055880273],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004334342,0.001558572,0.9548032,0.00033702658,0.00054089795,0.000025391872,0.00014644748,0.00092730625,0.00020374404,0.0011060949,0.011199019,0.024817903],"study_design_scores_gemma":[0.009551538,0.0012701075,0.9385874,0.03603242,0.0009924866,0.00011499123,0.00046467606,0.00042488208,0.00014344743,0.011416942,0.00085142656,0.00014965581],"about_ca_topic_score_codex":0.000011480462,"about_ca_topic_score_gemma":0.0000030489234,"teacher_disagreement_score":0.044754725,"about_ca_system_score_codex":0.000014516363,"about_ca_system_score_gemma":0.00053214363,"threshold_uncertainty_score":0.96112186},"labels":[],"label_agreement":null},{"id":"W4415972410","doi":"10.1007/s43678-025-01049-3","title":"Missing the needle in the haystack: diagnostic errors in neurological emergencies within Canadian emergency departments","year":2025,"lang":"en","type":"article","venue":"Canadian Journal of Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Women and Children’s Health Research Institute; Royal Alexandra Hospital; University of Ottawa; Ottawa Hospital; Canadian Medical Protective Association; University of Alberta","funders":"","keywords":"Harm; Emergency department; MEDLINE; Diagnostic test; Emergency medical services; Patient safety","score_opus":0.08492694399614965,"score_gpt":0.37601913616985416,"score_spread":0.2910921921737045,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4415972410","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.877181,0.009229866,0.000022569455,0.096478485,0.009653233,0.00053196255,0.0000105739055,0.000005969971,0.0068863444],"genre_scores_gemma":[0.9933519,0.0015138737,0.0000330208,0.0037360813,0.00052650034,0.000020404666,0.000013433173,0.000018883034,0.0007859288],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99621034,0.00038678784,0.0018485013,0.00028376575,0.00046926434,0.00080136093],"domain_scores_gemma":[0.99457425,0.0030742332,0.0003543859,0.00049489096,0.00030095832,0.0012012788],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0027452777,0.00028854102,0.00062706997,0.0008686172,0.00026976966,0.000010184534,0.0006396545,0.00018020709,0.0027325493],"category_scores_gemma":[0.10039317,0.00015937845,0.00019709954,0.0019205875,0.00025075584,0.00007857986,0.00001793127,0.0011165906,0.000013715714],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000046466077,0.00008267148,0.8643123,0.000025535119,0.000058992326,0.0014281087,0.0031929223,0.00017270756,0.000014031186,0.0005689531,0.129038,0.0010593254],"study_design_scores_gemma":[0.001476337,0.00068197004,0.9588809,0.0013259398,0.00028648844,0.0001098401,0.0038227073,0.00030913958,0.0000054456095,0.0033744685,0.02953502,0.00019175731],"about_ca_topic_score_codex":0.2687177,"about_ca_topic_score_gemma":0.87081665,"teacher_disagreement_score":0.60209894,"about_ca_system_score_codex":0.00019286397,"about_ca_system_score_gemma":0.003165389,"threshold_uncertainty_score":0.9981791},"labels":[],"label_agreement":null},{"id":"W4416221412","doi":"10.1080/1091367x.2025.2579239","title":"Dealing with Expert Uncertainty in Assigning Motor Competence Scores: New Dimensions of Validity Evidence","year":2025,"lang":"en","type":"article","venue":"Measurement in Physical Education and Exercise Science","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Competence (human resources); Construct validity; Motor skill; Dreyfus model of skill acquisition; Predictive validity; External validity; Validity","score_opus":0.08983193852424332,"score_gpt":0.38763708702096744,"score_spread":0.2978051484967241,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4416221412","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9969832,0.0007099427,0.00027258074,0.001222995,0.00015928903,0.00034575278,2.8389744e-7,0.000016345952,0.0002896131],"genre_scores_gemma":[0.9978115,0.0002897092,0.0015279851,0.00027684006,0.000026086766,0.000029500938,4.5759703e-7,0.0000037422526,0.000034209294],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9985404,0.00003187357,0.00025512956,0.0003539441,0.00062632497,0.00019230563],"domain_scores_gemma":[0.9981356,0.0011047443,0.00006993819,0.00023011008,0.00027224488,0.00018735982],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010030644,0.000099548626,0.0002795115,0.00019181162,0.000070588096,0.000024024934,0.00012230495,0.0000238604,0.0000049172245],"category_scores_gemma":[0.012048866,0.00006800332,0.000028091437,0.00091090496,0.00033477755,0.00013320321,0.000043034346,0.0001388941,0.0000015549392],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001229204,0.0063082757,0.5333597,0.00030901917,0.000015074389,0.0000077566,0.009711773,0.0027154728,0.13750234,0.005012006,0.0007035863,0.30312577],"study_design_scores_gemma":[0.0017417305,0.0005186938,0.88015604,0.07549612,0.00007724709,0.000002996547,0.0026291895,0.012645958,0.023125317,0.0032605955,0.000061834755,0.00028426954],"about_ca_topic_score_codex":0.0012700475,"about_ca_topic_score_gemma":0.000057942518,"teacher_disagreement_score":0.34679633,"about_ca_system_score_codex":0.00016638526,"about_ca_system_score_gemma":0.0021998887,"threshold_uncertainty_score":0.99627304},"labels":[],"label_agreement":null},{"id":"W4416321925","doi":"10.2196/73818","title":"Effect of Intolerance of Uncertainty and Resource Consumption on Therapeutic Strategies Chosen by Physiotherapists: Virtual Patient Study","year":2025,"lang":"en","type":"article","venue":"JMIR Rehabilitation and Assistive Technologies","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Centre for Interdisciplinary Research in Rehabilitation","funders":"","keywords":"Consumption (sociology); Resource (disambiguation); Resource consumption; Quality (philosophy); Patient care; Quality of life (healthcare); Resource use","score_opus":0.00905630155948258,"score_gpt":0.3462634684343009,"score_spread":0.33720716687481833,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4416321925","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99700266,0.0010393045,0.00004663274,0.00067018386,0.000018122008,0.00096363405,0.000014245537,0.00010290302,0.00014229208],"genre_scores_gemma":[0.9996184,0.00009470876,0.00006402334,0.000039221646,0.0000018900721,0.00014743289,0.0000057985185,0.000004668891,0.000023842673],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9991237,0.00013970025,0.000276756,0.00024449945,0.00013677686,0.00007854769],"domain_scores_gemma":[0.98910224,0.010482888,0.00013401765,0.00018934158,0.00007548428,0.000016034655],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00024154673,0.00012282476,0.00036094725,0.00011705534,0.00003932673,0.0000099234685,0.000040184845,0.000094612165,0.0000027038534],"category_scores_gemma":[0.0055186385,0.00008494493,0.000041962074,0.00016288903,0.00067794573,0.000027207969,0.00003015333,0.00013350346,3.4044749e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002647728,0.0012965278,0.28199425,0.00066159613,0.00015108405,8.9200046e-7,0.0012639335,0.000027394128,0.007249961,0.0027449133,0.00035004603,0.70161164],"study_design_scores_gemma":[0.0040016044,0.030733254,0.9316802,0.003581136,0.0000754681,5.5007706e-7,0.022157505,0.00015211759,0.006217167,0.0010961633,0.00017572595,0.00012913316],"about_ca_topic_score_codex":0.000021411626,"about_ca_topic_score_gemma":0.000002466703,"teacher_disagreement_score":0.70148253,"about_ca_system_score_codex":0.000030627984,"about_ca_system_score_gemma":0.000021827691,"threshold_uncertainty_score":0.6606725},"labels":[],"label_agreement":null},{"id":"W4416572991","doi":"10.2196/76618","title":"Teaching Clinical Reasoning in Health Care Professions Learners Using AI-Generated Script Concordance Tests: Mixed Methods Formative Evaluation","year":2025,"lang":"en","type":"article","venue":"JMIR Formative Research","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Institut Universitaire en Santé Mentale de Québec; Université de Montréal; Institut universitaire en santé mentale de Montréal; Centre Hospitalier Universitaire Sainte-Justine; Institut national de psychiatrie légale Philippe-Pinel; Université du Québec à Montréal","funders":"","keywords":"Formative assessment; Concordance; Health care; Generative grammar; Expert system; Teaching method; Educational measurement; Health professions","score_opus":0.21211504845246595,"score_gpt":0.6378854026687963,"score_spread":0.42577035421633036,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4416572991","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96993953,0.0018751335,0.009918813,0.006683272,0.00090772193,0.0033115686,0.000024556344,0.00011031199,0.007229065],"genre_scores_gemma":[0.9763485,0.00014764152,0.021221217,0.0012324982,0.00012596457,0.00039125845,0.00016803885,0.00002786084,0.00033702762],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9852519,0.010466812,0.0016289393,0.00054617814,0.001153559,0.00095258356],"domain_scores_gemma":[0.9785194,0.018511955,0.00036983244,0.00053669914,0.0016483887,0.00041376936],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.027653735,0.00025320947,0.00085982285,0.0007668128,0.0007993567,0.000083794526,0.00024109734,0.00029685674,0.00004036849],"category_scores_gemma":[0.10683003,0.000205739,0.00017525077,0.001716859,0.00040887826,0.00042834296,0.00030621022,0.0036814993,0.000024485073],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011833084,0.0015854252,0.061330885,0.0010036931,0.00023833153,0.000026356474,0.06764014,0.00039207505,0.000498626,0.0058743143,0.0204165,0.8398104],"study_design_scores_gemma":[0.016356645,0.0036469195,0.28999934,0.05340022,0.00013182165,0.00002618458,0.24164641,0.38331032,0.0012877247,0.0034805413,0.0059912424,0.00072262617],"about_ca_topic_score_codex":0.001048375,"about_ca_topic_score_gemma":0.00014842271,"teacher_disagreement_score":0.8390877,"about_ca_system_score_codex":0.0018281555,"about_ca_system_score_gemma":0.0042467527,"threshold_uncertainty_score":0.99861705},"labels":[],"label_agreement":null},{"id":"W4416690664","doi":"10.1136/bmjqs-2025-019299","title":"Artificial intelligence chain-of-thought reasoning in nuanced medical scenarios: mitigation of cognitive biases through model intransigence","year":2025,"lang":"en","type":"article","venue":"BMJ Quality & Safety","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; Sunnybrook Health Science Centre; University of Toronto","funders":"Canadian Institutes of Health Research; Canada Research Chairs; Physicians' Services Incorporated Foundation","keywords":"Cognitive bias; Cognition; Diversity (politics); Motivated reasoning; Confirmation bias; Cognitive computing; Model-based reasoning; MEDLINE","score_opus":0.10434124905932249,"score_gpt":0.46361369656425716,"score_spread":0.3592724475049347,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4416690664","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.66192305,0.00041205593,0.3306034,0.004352457,0.00018406997,0.0007417347,0.00006598612,0.000042546242,0.0016746888],"genre_scores_gemma":[0.9930767,0.0005457752,0.0051807608,0.0009101828,0.00007704422,0.00003761898,0.00010549639,0.000011276357,0.00005513732],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99595624,0.00035406725,0.0022633055,0.00043269686,0.00070286886,0.00029082785],"domain_scores_gemma":[0.9598384,0.0386842,0.0004317068,0.0003903812,0.0004872976,0.00016804664],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002962142,0.00018470052,0.0007956727,0.00010210692,0.000053473163,0.000005990132,0.00015639434,0.00027383066,0.00008916492],"category_scores_gemma":[0.29087177,0.00016887314,0.00018820023,0.00075868296,0.0005363983,0.00008188933,0.00006749482,0.00041125907,0.0000050190433],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.012302766,0.0049758228,0.07139349,0.0012371001,0.00040724844,0.000119989854,0.015538102,0.01084463,0.002641616,0.28996074,0.0003682655,0.5902102],"study_design_scores_gemma":[0.0053241234,0.0010890922,0.12575567,0.22210656,0.0007601525,0.000029835443,0.018928643,0.35849023,0.15623105,0.11002771,0.000041685016,0.0012152662],"about_ca_topic_score_codex":0.00074225385,"about_ca_topic_score_gemma":0.00035232358,"teacher_disagreement_score":0.588995,"about_ca_system_score_codex":0.00007476838,"about_ca_system_score_gemma":0.0013086388,"threshold_uncertainty_score":0.71510154},"labels":[],"label_agreement":null},{"id":"W4416768144","doi":"10.1186/s41687-025-00964-4","title":"Beyond diagnostic test performance: two content-validated questionnaires assessing patient and clinician satisfaction with diagnostic tests","year":2025,"lang":"en","type":"article","venue":"Journal of Patient-Reported Outcomes","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Armand Frappier Museum","funders":"","keywords":"Debriefing; Diagnostic test; Test (biology); Patient satisfaction; Medical diagnosis; Cognitive interview; Diagnostic accuracy","score_opus":0.02388479223893791,"score_gpt":0.34381473881333546,"score_spread":0.31992994657439755,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4416768144","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9953327,0.0007505772,0.00015522997,0.001520693,0.0013572704,0.0005440807,0.00000756005,0.000059140577,0.00027273461],"genre_scores_gemma":[0.99653023,0.00045109415,0.0014206676,0.001354595,0.00010263453,0.000020262687,0.00002042923,0.000030898715,0.000069168236],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9964247,0.00012144509,0.0020111853,0.00034569058,0.00073461456,0.00036239394],"domain_scores_gemma":[0.94732463,0.048817515,0.0019216081,0.00037870734,0.0011464125,0.000411137],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0005399566,0.00034486808,0.0009633574,0.00039638096,0.00019799438,0.00012224884,0.0000752768,0.00015126937,0.000029043153],"category_scores_gemma":[0.23961975,0.00023185246,0.00019062514,0.00042572946,0.00021033341,0.00036520694,0.000058879556,0.00061075686,0.0000047243534],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00016946028,0.00063235883,0.9090241,0.000063149215,0.00034650406,0.00077070756,0.00017928086,0.00008428472,0.00011427375,0.00001828096,0.0004587084,0.088138856],"study_design_scores_gemma":[0.005321298,0.0026789,0.9793688,0.00882392,0.0012878891,0.0008642138,0.00050617394,0.00009386792,0.0006029636,0.0001406244,0.00008003675,0.00023131372],"about_ca_topic_score_codex":0.0002531704,"about_ca_topic_score_gemma":0.00010007742,"teacher_disagreement_score":0.23907979,"about_ca_system_score_codex":0.00017591355,"about_ca_system_score_gemma":0.0005718528,"threshold_uncertainty_score":0.9454669},"labels":[],"label_agreement":null},{"id":"W4417023170","doi":"10.30770/2572-1852-111.3.49","title":"Book Review","year":2025,"lang":"en","type":"article","venue":"Journal of Medical Regulation","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Elegance; Event (particle physics); Intervention (counseling); Natural (archaeology); Commission; Power (physics)","score_opus":0.01732633757060913,"score_gpt":0.3958054590306377,"score_spread":0.37847912146002854,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4417023170","genre_codex":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0035658835,0.5698659,0.0098012965,0.40965506,0.00086351775,0.00018289202,3.3586318e-7,0.000017476708,0.006047637],"genre_scores_gemma":[0.08608249,0.40409982,0.0015702347,0.4794518,0.0019606566,0.0000054877273,0.000017039618,0.00002041704,0.026792066],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9983023,0.00004668588,0.00059443066,0.0000616184,0.0009162049,0.000078758036],"domain_scores_gemma":[0.99641067,0.0027578415,0.00021334755,0.0001224608,0.00024977137,0.00024588313],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001454569,0.000050188184,0.00031501596,0.00007239445,0.000016478945,0.000004857313,0.00007822756,0.00011308478,0.003209248],"category_scores_gemma":[0.104760416,0.00003288529,0.0001534366,0.00016889932,0.00004782333,0.000038713344,0.000016817687,0.00027336404,0.00001710218],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004978578,0.00009455523,0.000623053,0.00017230662,0.000034409855,0.000049513674,0.000003914771,8.3876097e-7,0.000009776146,0.0007449,0.903927,0.094289936],"study_design_scores_gemma":[0.0018661338,0.00021693668,0.057158954,0.08631714,0.00033604781,0.00029315858,0.000004064355,0.0002946287,0.00009882033,0.0011832389,0.8521886,0.000042294396],"about_ca_topic_score_codex":0.0000013575481,"about_ca_topic_score_gemma":1.4718276e-7,"teacher_disagreement_score":0.16576608,"about_ca_system_score_codex":0.000041790958,"about_ca_system_score_gemma":0.00046799838,"threshold_uncertainty_score":0.99770194},"labels":[],"label_agreement":null},{"id":"W4417136032","doi":"10.2196/73196","title":"Virtual Standardized Patients for Improving Clinical Thinking Ability Training in Residents: Randomized Controlled Trial","year":2025,"lang":"en","type":"article","venue":"JMIR Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Randomized controlled trial; Promotion (chess); Clinical trial; Critical thinking; Clinical Practice; Training (meteorology); Virtual patient","score_opus":0.0191821282021243,"score_gpt":0.4244807316390945,"score_spread":0.4052986034369702,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4417136032","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9695461,0.00015850565,0.0033351248,0.0061134724,0.005962281,0.012688137,0.000004651296,0.000081506856,0.0021102314],"genre_scores_gemma":[0.98900247,0.000044194923,0.0012305697,0.003981282,0.0009633886,0.0041274657,0.00012409317,0.000017632483,0.00050891197],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.99495524,0.000884609,0.0023963454,0.00053035916,0.0008977159,0.0003357314],"domain_scores_gemma":[0.9086893,0.08886369,0.0006558972,0.00051337027,0.0006735754,0.00060416636],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.013542449,0.00019844178,0.0021797584,0.0002015872,0.0000961888,0.000043653876,0.00016884247,0.00046476375,0.00015648818],"category_scores_gemma":[0.69741756,0.00014385545,0.0007366469,0.0002555128,0.0002820177,0.000070953414,0.000050854207,0.00061678333,0.000004017018],"study_design_candidate":"randomized_trial","study_design_consensus":"randomized_trial","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.80299973,0.0020615412,0.0054846667,0.000050354192,0.00009742145,9.363332e-7,0.0007652018,3.5306124e-7,9.0718726e-7,0.0016608749,0.002299421,0.1845786],"study_design_scores_gemma":[0.9748953,0.0006570135,0.0124958875,0.002486186,0.00032232812,5.4844475e-7,0.0005454395,0.0008430275,0.0000027430747,0.007371177,0.00026020568,0.00012014416],"about_ca_topic_score_codex":0.00009246269,"about_ca_topic_score_gemma":0.000023835732,"teacher_disagreement_score":0.6838751,"about_ca_system_score_codex":0.00022574101,"about_ca_system_score_gemma":0.008019362,"threshold_uncertainty_score":0.99760425},"labels":[],"label_agreement":null},{"id":"W4417199531","doi":"10.36834/cmej.80873","title":"AI that teaches: an evidence-based GPT model to improve medical student understanding of pulmonary function tests","year":2025,"lang":"en","type":"article","venue":"Canadian Medical Education Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Function (biology); Pulmonary medicine; Statement (logic); MEDLINE; Test (biology)","score_opus":0.07498799951974351,"score_gpt":0.41562600248064846,"score_spread":0.34063800296090496,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4417199531","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3044109,0.0020965117,0.07460126,0.6076076,0.006188247,0.00092780427,0.000013617769,0.00006564815,0.0040884004],"genre_scores_gemma":[0.8889634,0.000105939216,0.0003656299,0.10952339,0.0005303348,0.00002477966,0.000016795235,0.000014920961,0.00045479645],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.997009,0.00014300326,0.00062119425,0.00031363894,0.0015212584,0.00039191014],"domain_scores_gemma":[0.9840768,0.0024978681,0.00014932103,0.0003751699,0.00043925477,0.012461617],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0022305907,0.00016118593,0.00035080433,0.0005591444,0.00018290196,0.000056183526,0.00029885664,0.00035388544,0.0039231577],"category_scores_gemma":[0.1840811,0.00013651232,0.00012703551,0.00040284527,0.00017134263,0.00014244382,0.000024001421,0.0010287485,0.000013567002],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005836932,0.0031103473,0.16290651,0.00021809919,0.00025111344,0.00034231777,0.001307725,0.0004274137,0.00013230466,0.0056627337,0.21454774,0.61051],"study_design_scores_gemma":[0.01114104,0.004973209,0.4601826,0.13657342,0.002300472,0.0017672427,0.02945244,0.30860156,0.0003287343,0.023114022,0.019787824,0.0017774325],"about_ca_topic_score_codex":0.0027270368,"about_ca_topic_score_gemma":0.003537656,"teacher_disagreement_score":0.6087326,"about_ca_system_score_codex":0.0012622549,"about_ca_system_score_gemma":0.0934026,"threshold_uncertainty_score":0.9969874},"labels":[],"label_agreement":null},{"id":"W4417304307","doi":"10.1007/s10459-025-10493-w","title":"Can you really infer that? An exploratory study using the reasoning task typology to analyze clinic notes","year":2025,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Typology; Task (project management); Exploratory research; Explication; Ambiguity; Think aloud protocol; Verbal reasoning; Reading (process); Task analysis","score_opus":0.06568179075857844,"score_gpt":0.493443299791533,"score_spread":0.42776150903295457,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4417304307","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9820114,0.0025436373,0.0002115905,0.012167676,0.0015987066,0.0008256876,0.0000017778591,0.000030171152,0.00060940103],"genre_scores_gemma":[0.98691285,0.0008075889,0.003316098,0.008609683,0.00016341238,0.00009542023,0.00000695925,0.0000064568976,0.00008153394],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9978607,0.00036925497,0.0005675751,0.00051378243,0.00032099904,0.00036765102],"domain_scores_gemma":[0.9944305,0.0045596785,0.00022733858,0.00042846857,0.00013097469,0.00022306814],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0035400887,0.0001225983,0.00030714733,0.0002963655,0.00037433047,0.00004495594,0.00029178188,0.00005163805,0.000020203224],"category_scores_gemma":[0.019890541,0.00008719599,0.000030966716,0.0014461284,0.00027226785,0.00022449698,0.00006319993,0.00023741864,0.000005707896],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000039292987,0.0005348848,0.9630665,0.000021337335,0.0000039656607,8.7941083e-7,0.0037233355,0.0005168927,0.0000041384005,0.00062705233,0.00009584053,0.031365868],"study_design_scores_gemma":[0.0005656952,0.0017226645,0.93997604,0.0022389232,0.00004167014,0.000006837436,0.049973544,0.0004687885,0.000019394301,0.0030637744,0.0017564936,0.00016616883],"about_ca_topic_score_codex":0.0037846505,"about_ca_topic_score_gemma":0.009681318,"teacher_disagreement_score":0.04625021,"about_ca_system_score_codex":0.00021784134,"about_ca_system_score_gemma":0.003527355,"threshold_uncertainty_score":0.98836535},"labels":[],"label_agreement":null},{"id":"W4417304527","doi":"10.1136/bmj.r2570","title":"Parallel pressures: the common roots of doctor bullshit and large language model hallucinations","year":2025,"lang":"en","type":"article","venue":"BMJ","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Manitoba; University of Alberta","funders":"","keywords":"Mental model; Language model; Health care; Mental healthcare; Loop (graph theory); Action (physics)","score_opus":0.023121320565561543,"score_gpt":0.3832492770843326,"score_spread":0.36012795651877105,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4417304527","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.90359807,0.0036461272,0.0036545226,0.0361878,0.00014045261,0.00086656475,0.000088981054,0.00006427288,0.05175318],"genre_scores_gemma":[0.98643196,0.0000666138,0.0007005081,0.0010752002,0.00004523963,0.000027622515,0.00001529179,0.0000047187564,0.011632858],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9994434,0.00003199031,0.00019874427,0.00010542098,0.00011847488,0.00010195347],"domain_scores_gemma":[0.9964982,0.003090991,0.000048969054,0.000269444,0.000046477315,0.00004589604],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0003112731,0.000057291323,0.00017544963,0.000028165205,0.00004551343,0.00000810211,0.0000627807,0.000055573608,0.000037447746],"category_scores_gemma":[0.009173561,0.000036311845,0.000046990237,0.00007474859,0.000051839495,0.000012482643,0.000052668016,0.00010446669,0.000005509717],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00069994887,0.0024320895,0.2457107,0.00046712442,0.00065272866,0.00009899218,0.004472336,0.0024369084,0.0008251391,0.18204895,0.5290594,0.031095665],"study_design_scores_gemma":[0.0064271456,0.00022682652,0.8836372,0.0029586283,0.00070259726,0.000018658151,0.0008826711,0.077436216,0.00040791775,0.0064006015,0.020687921,0.00021361023],"about_ca_topic_score_codex":0.00010963301,"about_ca_topic_score_gemma":0.000109585584,"teacher_disagreement_score":0.6379265,"about_ca_system_score_codex":0.0000060438924,"about_ca_system_score_gemma":0.00006637796,"threshold_uncertainty_score":0.99917257},"labels":[],"label_agreement":null},{"id":"W4417311643","doi":"10.1503/cmaj.250663-f","title":"L’IA médicale, la mésinformation et l’importance d’un dialogue ouvert avec la patientèle","year":2025,"lang":"en","type":"article","venue":"Canadian Medical Association Journal","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"MEDLINE; General practice","score_opus":0.005904261768331088,"score_gpt":0.28419749948199746,"score_spread":0.2782932377136664,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4417311643","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4637229,0.0006258528,0.00419683,0.20107867,0.0029710634,0.00050735124,0.000096221374,0.00013876881,0.32666233],"genre_scores_gemma":[0.947621,0.0005859579,0.00032719152,0.048824184,0.0004970494,0.000011654819,0.00010392335,0.000014397791,0.0020146547],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9972719,0.00030544316,0.000726991,0.00015808185,0.0011257557,0.0004118536],"domain_scores_gemma":[0.99177706,0.0058012744,0.00027423777,0.00014759069,0.0003786335,0.0016211886],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002955729,0.00013794607,0.00032870585,0.00020799338,0.0001846777,0.00010832529,0.00015164605,0.0006190857,0.0019571094],"category_scores_gemma":[0.10331218,0.00012074981,0.00016196325,0.0003517067,0.0000751186,0.00012475387,0.000020949867,0.0013816295,0.0001288941],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003245114,0.00017090085,0.37218058,0.000016844957,0.00018338594,0.0007509305,0.000559167,0.000008845017,0.0000016003445,0.0038789916,0.48053104,0.14168528],"study_design_scores_gemma":[0.003112006,0.00008888515,0.49653953,0.0011507542,0.00010606543,0.0003881209,0.0002751837,0.0006316741,0.000014078773,0.0012173781,0.4963124,0.00016390436],"about_ca_topic_score_codex":0.000562296,"about_ca_topic_score_gemma":0.0027883223,"teacher_disagreement_score":0.48389807,"about_ca_system_score_codex":0.0006761283,"about_ca_system_score_gemma":0.0041140076,"threshold_uncertainty_score":0.99895525},"labels":[],"label_agreement":null},{"id":"W4417348466","doi":"10.1108/978-1-62396-597-620251019","title":"Validating Diagnostic Inferences","year":2014,"lang":"en","type":"book-chapter","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"","score_opus":0.0452343452386184,"score_gpt":0.32772908523233746,"score_spread":0.28249473999371905,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4417348466","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0003758652,0.00033558047,0.00064820045,0.0007981388,0.0003565445,0.00024192236,0.0000086040045,0.00016918665,0.99706596],"genre_scores_gemma":[0.05436963,0.00059608475,0.0010827646,0.0020900953,0.0009867728,0.000010676966,0.00017414038,0.000059476293,0.9406304],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.9985533,0.000010859205,0.00047575022,0.00038739535,0.00035970286,0.00021303342],"domain_scores_gemma":[0.95514315,0.043829206,0.00017855478,0.00043171685,0.00013247292,0.0002848909],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0002563103,0.00028397472,0.0006772602,0.00008136844,0.000042505737,0.000030488687,0.00009779686,0.00040919217,0.007706458],"category_scores_gemma":[0.09639344,0.00021094813,0.00023402939,0.000015807294,0.000113421505,0.000010853814,0.00007276287,0.0004559234,0.0025670817],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000031903583,0.000075284515,0.008109893,0.0001491623,0.00026436648,0.00024106631,0.00004162234,0.0000029234652,0.0000034102354,0.66552883,0.085062444,0.24048911],"study_design_scores_gemma":[0.0026483685,0.0018092026,0.0064693163,0.03728051,0.001999751,0.00019747751,0.00001960571,0.0001352656,0.000102880535,0.0952044,0.85273695,0.0013962677],"about_ca_topic_score_codex":0.000038823793,"about_ca_topic_score_gemma":0.000011963089,"teacher_disagreement_score":0.7676745,"about_ca_system_score_codex":0.000032567023,"about_ca_system_score_gemma":0.00015221152,"threshold_uncertainty_score":0.99820954},"labels":[],"label_agreement":null},{"id":"W4417526999","doi":"10.1016/j.nepr.2025.104676","title":"Nursing clinical reasoning cognitive strategies used in a learning-by-concordance modality: A qualitative descriptive study","year":2025,"lang":"en","type":"article","venue":"Nurse Education in Practice","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières; Montreal Heart Institute; Université Laval; Université de Montréal; Centre for Interdisciplinary Research in Rehabilitation","funders":"Fonds de Recherche du Québec - Santé; Social Sciences and Humanities Research Council of Canada; Fonds de recherche du Québec","keywords":"Interactivity; Cognition; Descriptive research; Modality (human–computer interaction); Qualitative research; Structuring","score_opus":0.05573735661764446,"score_gpt":0.5359107297705723,"score_spread":0.48017337315292785,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4417526999","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95762134,0.0006363438,0.0012405526,0.0071327453,0.0010871935,0.0012598818,0.0000038087996,0.00006126135,0.030956862],"genre_scores_gemma":[0.9949255,0.00009166507,0.0014923508,0.00073934573,0.000120739176,0.00026972985,0.00004344059,0.000021239664,0.0022959858],"study_design_codex":"observational","study_design_gemma":"qualitative","domain_scores_codex":[0.9945763,0.0030450434,0.0010303321,0.0006764287,0.0003478074,0.0003241428],"domain_scores_gemma":[0.93956393,0.05874682,0.0005193988,0.0002976843,0.0006955494,0.00017660462],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00411573,0.0002207904,0.0005970952,0.00027147844,0.000092005685,0.000103404745,0.000102254795,0.00018635165,0.000031754775],"category_scores_gemma":[0.2786432,0.00022900976,0.00008110964,0.0009789586,0.00027414976,0.00057965773,0.000017603585,0.0014446549,0.00002950801],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0033832004,0.025337117,0.71959704,0.000028816423,0.00019968339,0.000050510833,0.19358353,0.000031577885,0.000021933109,0.007543404,0.0023151624,0.04790799],"study_design_scores_gemma":[0.0042879856,0.000841225,0.1816162,0.0072855065,0.00034797893,0.00001001843,0.80277,0.00022817975,0.000021009646,0.0022288824,0.00019177972,0.000171226],"about_ca_topic_score_codex":0.0020298571,"about_ca_topic_score_gemma":0.00020581816,"teacher_disagreement_score":0.6091865,"about_ca_system_score_codex":0.0005255281,"about_ca_system_score_gemma":0.0045919884,"threshold_uncertainty_score":0.93387467},"labels":[],"label_agreement":null},{"id":"W560172043","doi":"10.71781/6059","title":"Contribution à l'étude de la construction des concepts scientifiques au cours de l'apprentissage par problèmes en médecine","year":2010,"lang":"fr","type":"dissertation","venue":"Open MIND","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Philosophy; Humanities","score_opus":0.019517390428131602,"score_gpt":0.39081221561026214,"score_spread":0.37129482518213053,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W560172043","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9866259,0.0012457123,0.001599806,0.00034023824,0.0023977098,0.0015440175,0.0002676413,0.000013346156,0.005965614],"genre_scores_gemma":[0.9663474,0.0006240526,0.024908468,0.000038218386,0.00088083313,0.00014303136,0.0035255293,0.00005454887,0.003477896],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99607426,0.0008106273,0.00087599986,0.0008487945,0.00047162082,0.0009186692],"domain_scores_gemma":[0.98617,0.0113587435,0.0006348259,0.00045337304,0.00066392095,0.0007191176],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00458174,0.00048321852,0.00092665444,0.00014559917,0.00052312925,0.0005519107,0.00051649555,0.0013347095,0.0034026224],"category_scores_gemma":[0.066904545,0.00046928864,0.00023215036,0.00030819408,0.0018882122,0.0002586865,0.00012237398,0.001189757,0.00030696756],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007295645,0.0013222949,0.56672466,0.00020553889,0.00035151112,0.00020173052,0.009049657,0.000041534488,0.03909884,0.0015970872,0.00063326414,0.38004434],"study_design_scores_gemma":[0.0068871486,0.00055794977,0.7504303,0.023356492,0.002030176,0.00041238256,0.008310685,0.0006127466,0.19900355,0.0032723122,0.004386009,0.0007402828],"about_ca_topic_score_codex":0.0014953661,"about_ca_topic_score_gemma":0.0024755027,"teacher_disagreement_score":0.37930405,"about_ca_system_score_codex":0.0006445492,"about_ca_system_score_gemma":0.0030612962,"threshold_uncertainty_score":0.9999618},"labels":[],"label_agreement":null},{"id":"W58770535","doi":"","title":"Nurse judgments of hallucinated voice descriptions: relevance for intervention.","year":2004,"lang":"en","type":"article","venue":"PubMed","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Windsor; Chrysler (Canada)","funders":"","keywords":"Concordance; Cronbach's alpha; Psychology; Equivalence (formal languages); Psychological intervention; Consistency (knowledge bases); Sample (material); Relevance (law); Content analysis; Social psychology; Clinical psychology; Applied psychology; Linguistics; Psychometrics; Medicine; Computer science; Psychiatry","score_opus":0.043128578912989934,"score_gpt":0.3216375452975696,"score_spread":0.27850896638457967,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W58770535","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9818629,0.001281666,0.007912616,0.0038049202,0.0007672297,0.0020794813,0.00004089927,0.00010652757,0.002143762],"genre_scores_gemma":[0.99401265,0.00006355336,0.0023546373,0.00030465354,0.000074200325,0.0011473738,0.000038264523,0.00001396096,0.001990694],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9991026,0.000013336943,0.0003446062,0.00017944704,0.00014332426,0.0002167144],"domain_scores_gemma":[0.9981805,0.001191302,0.00011516502,0.00020298247,0.0001671407,0.0001429183],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00032189075,0.00007728255,0.00021335363,0.000054873835,0.000026532149,0.000006765003,0.000062987645,0.00006517005,0.000013380035],"category_scores_gemma":[0.038253807,0.00006801299,0.00015975024,0.00013552807,0.000060267994,0.00004715501,0.000011246839,0.00008013257,0.000017061655],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019499307,0.0071004755,0.03928862,0.00052642333,0.00050474,0.0000409534,0.0003603439,0.00020895695,0.00037132582,0.0065380144,0.012243396,0.93086684],"study_design_scores_gemma":[0.012025186,0.0003740382,0.9699717,0.0018388832,0.00030411623,0.000019847887,0.000114493625,0.000041192696,0.004792991,0.006583394,0.0037894258,0.00014473376],"about_ca_topic_score_codex":0.00006782077,"about_ca_topic_score_gemma":0.00001206097,"teacher_disagreement_score":0.93072206,"about_ca_system_score_codex":0.000081880826,"about_ca_system_score_gemma":0.000030573945,"threshold_uncertainty_score":0.9698474},"labels":[],"label_agreement":null},{"id":"W6884612161","doi":"10.1051/pmed/2023008/pdf","title":"Formation par concordance de script en orthophonie : récit de pratique en évaluation des troubles acquis de la communication","year":2023,"lang":"fr","type":"article","venue":"Springer Link (Chiba Institute of Technology)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Qualitative analysis; Occupational training; Context (archaeology)","score_opus":0.021306004693647605,"score_gpt":0.3154830484240493,"score_spread":0.29417704373040165,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6884612161","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95296985,0.005770915,0.02245587,0.014538393,0.0006709023,0.0006210046,0.00002425316,0.00060470833,0.0023441133],"genre_scores_gemma":[0.85828453,0.019764451,0.12046757,0.00028103992,0.00034765684,0.0001899896,0.000085536994,0.00004396247,0.000535229],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9972769,0.000461968,0.000891669,0.00036243652,0.00027920931,0.0007277859],"domain_scores_gemma":[0.9944011,0.0038086271,0.0004356308,0.0008952092,0.00027057095,0.0001888367],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00454978,0.00030134903,0.0005565434,0.00050545024,0.0002678143,0.000048642538,0.00054988463,0.0012235336,0.000040001287],"category_scores_gemma":[0.06225432,0.00032606698,0.00017483285,0.0010925583,0.001631318,0.00038309552,0.0003032752,0.0010895465,0.00011816414],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018717446,0.0005490722,0.21225871,0.0008566055,0.00030156426,0.00012289676,0.004782292,0.002663622,0.035018757,0.09437945,0.0018925185,0.6469873],"study_design_scores_gemma":[0.0068772295,0.0010517117,0.5057953,0.029015396,0.0012052852,0.00044684592,0.0021989543,0.03806379,0.19917399,0.13698538,0.07820109,0.0009850472],"about_ca_topic_score_codex":0.00083146855,"about_ca_topic_score_gemma":0.00032725945,"teacher_disagreement_score":0.6460023,"about_ca_system_score_codex":0.000773471,"about_ca_system_score_gemma":0.0013364168,"threshold_uncertainty_score":0.9999191},"labels":[],"label_agreement":null},{"id":"W6901902759","doi":"10.6084/m9.figshare.18095993","title":"Types of clinical reasoning in a summative clerkship oral examination","year":2022,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Toronto; University Health Network","funders":"","keywords":"Summative assessment; Objective structured clinical examination; Type (biology); Educational measurement; Clinical clerkship; Descriptive statistics","score_opus":0.14643905489476922,"score_gpt":0.42662026709917145,"score_spread":0.28018121220440223,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6901902759","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.81633484,0.00052668096,0.0000017474047,0.0004958287,0.00024434715,0.00080910674,0.15281256,0.000114645525,0.028660277],"genre_scores_gemma":[0.9471,0.0000025092427,0.00022071826,0.00020836468,0.00006721692,0.00012771202,0.051850285,0.000010802421,0.00041243184],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9988121,0.00021969553,0.00041906894,0.00017963207,0.00024637932,0.00012315488],"domain_scores_gemma":[0.9838346,0.015571556,0.00020579823,0.00019043048,0.00010461777,0.00009296978],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0005257813,0.00006425563,0.00030249593,0.00006834554,0.000030084142,0.0000042170077,0.000075978474,0.000060218954,0.22049506],"category_scores_gemma":[0.2534541,0.000060536757,0.00009298356,0.00026789613,0.00001012827,0.000027461798,0.00010643621,0.0003449108,0.000164216],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00047000535,0.0014063455,0.11370207,0.0001522338,0.00007501051,0.00029390302,0.0009729543,0.00019844716,0.000007910208,0.00023476323,0.6858295,0.19665684],"study_design_scores_gemma":[0.0015513162,0.0005576458,0.97354615,0.006768523,0.000021196394,0.000012407423,0.00028924007,0.0018986162,0.000045985988,0.000054448465,0.01516223,0.000092227885],"about_ca_topic_score_codex":0.00001922019,"about_ca_topic_score_gemma":0.000012183918,"teacher_disagreement_score":0.8598441,"about_ca_system_score_codex":0.000037750226,"about_ca_system_score_gemma":0.00015174648,"threshold_uncertainty_score":0.7802175},"labels":[],"label_agreement":null},{"id":"W6902016736","doi":"10.6084/m9.figshare.24181056","title":"Additional file 1 of Deliberate practice of diagnostic clinical reasoning reveals low performance and improvement of diagnostic justification in pre-clerkship students","year":2023,"lang":"en","type":"article","venue":"Open MIND","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Case-based reasoning; Clinical Practice; Quality (philosophy); Model-based reasoning","score_opus":0.050478786163622356,"score_gpt":0.40947726341721913,"score_spread":0.3589984772535968,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6902016736","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98174864,0.00004801674,0.0000031550246,0.00016724913,0.00007044773,0.0008770058,0.016195454,0.0000026519092,0.00088735623],"genre_scores_gemma":[0.98377526,0.0008220029,0.0045679477,0.00008511516,0.000069167436,0.00030663537,0.009114172,0.000018601306,0.0012411248],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9976056,0.00013672582,0.001227127,0.00036914198,0.00045871493,0.00020267567],"domain_scores_gemma":[0.8078055,0.19082968,0.00065413245,0.0003002795,0.00028253326,0.00012790161],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0017919029,0.00012936015,0.0005446798,0.00010113486,0.000029547391,0.0000206086,0.00022296602,0.00012376752,0.04645256],"category_scores_gemma":[0.52013946,0.00011926778,0.0000628353,0.0004007706,0.00017172359,0.00016128451,0.00024415227,0.00022285296,0.00009650477],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010911922,0.0034176037,0.21580842,0.00037297426,0.00020489028,0.00006963115,0.0008111221,0.00017144572,0.00022855405,0.000008520543,0.31312782,0.46468782],"study_design_scores_gemma":[0.0018891151,0.0011820676,0.97664565,0.013879743,0.00018080647,0.000008641687,0.00022423117,0.00069619546,0.0004902499,0.000018478775,0.004666795,0.00011803335],"about_ca_topic_score_codex":0.000066170636,"about_ca_topic_score_gemma":0.00001304791,"teacher_disagreement_score":0.7608372,"about_ca_system_score_codex":0.000026198131,"about_ca_system_score_gemma":0.00022838192,"threshold_uncertainty_score":0.95441914},"labels":[],"label_agreement":null},{"id":"W6902046395","doi":"10.6084/m9.figshare.19817724","title":"Additional file 1 of Embracing uncertainty: medical student perceptions of a pediatric bootcamp developed in response to mandated changes during the pandemic","year":2022,"lang":"en","type":"article","venue":"Open MIND","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Pandemic; Perception; Coronavirus disease 2019 (COVID-19); MEDLINE; Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2)","score_opus":0.041705202040512084,"score_gpt":0.38400783012183126,"score_spread":0.34230262808131917,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6902046395","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96311736,0.000015067966,0.0000012641492,0.0012043228,0.00003245436,0.00043616214,0.034648985,0.0000018579254,0.0005425236],"genre_scores_gemma":[0.9892584,0.000020513015,0.0007818614,0.0002734344,0.000085567306,0.0006423401,0.0053427084,0.00001210758,0.003583062],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.998451,0.00023454605,0.0003738417,0.00021337777,0.0005686523,0.00015859888],"domain_scores_gemma":[0.9809856,0.018486729,0.00012224303,0.00020007753,0.00007164758,0.00013373428],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0010658111,0.00007877009,0.00026796502,0.00014649724,0.000099476536,0.000009148874,0.00028735236,0.000046143738,0.7749353],"category_scores_gemma":[0.042196617,0.000059580776,0.000044117194,0.00049865973,0.000049438528,0.00001741025,0.00052010576,0.00025933725,0.000065059496],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.011593433,0.0031657638,0.1089957,0.00006443988,0.00024117113,0.0004858111,0.016934443,0.0009114554,0.0015073911,0.0000049984774,0.7669596,0.08913576],"study_design_scores_gemma":[0.0021401495,0.00048452476,0.95123535,0.0009115333,0.000057951132,0.00006907908,0.005427824,0.000113117276,0.000021942691,0.000006648962,0.039411455,0.00012043188],"about_ca_topic_score_codex":0.00006857611,"about_ca_topic_score_gemma":0.00034207516,"teacher_disagreement_score":0.8422396,"about_ca_system_score_codex":0.000103840925,"about_ca_system_score_gemma":0.00072383706,"threshold_uncertainty_score":0.9658714},"labels":[],"label_agreement":null},{"id":"W6902129392","doi":"10.6084/m9.figshare.12096669.v1","title":"Additional file 1 of Mapping clinical reasoning literature across the health professions: a scoping review","year":2020,"lang":"en","type":"article","venue":"Open MIND","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Interdisciplinary Research in Rehabilitation; McGill University","funders":"","keywords":"MEDLINE; Systematic review; Case-based reasoning; Health care; Key (lock)","score_opus":0.15701872277189188,"score_gpt":0.49367587049279876,"score_spread":0.3366571477209069,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6902129392","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004945718,0.07539461,0.00012645854,0.2658006,0.0006097145,0.0134841055,0.61727756,0.000035615907,0.02232565],"genre_scores_gemma":[0.0075242287,0.037139084,0.32438976,0.26032072,0.0055502066,0.003231552,0.34619114,0.00019723052,0.015456079],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9981006,0.00024540228,0.0008431135,0.00033963996,0.00024581968,0.00022544443],"domain_scores_gemma":[0.96860427,0.029998098,0.0005220805,0.00035766006,0.00017089535,0.00034702293],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0010880511,0.00011083512,0.00062888354,0.00000857199,0.00015593541,0.000048830236,0.0002540548,0.00008994471,0.45229954],"category_scores_gemma":[0.19847287,0.00006821218,0.00017102153,0.0003631944,0.00010822517,0.00006647687,0.0002792692,0.0005134621,0.00035381323],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000046831614,0.00008114578,0.00015523889,0.0004930517,0.00002275877,0.00002165835,0.0005640348,1.00097125e-7,3.66103e-7,0.0000021611365,0.8890906,0.10952206],"study_design_scores_gemma":[0.000506899,0.00027228304,0.0027856948,0.68808335,0.000021053153,0.00003434141,0.00034980904,0.000050640174,0.000002487843,0.0000056662884,0.30781415,0.000073645686],"about_ca_topic_score_codex":0.0000049994364,"about_ca_topic_score_gemma":0.0000040674013,"teacher_disagreement_score":0.6875903,"about_ca_system_score_codex":0.000013875221,"about_ca_system_score_gemma":0.0009536569,"threshold_uncertainty_score":0.80827874},"labels":[],"label_agreement":null},{"id":"W6902225828","doi":"10.6084/m9.figshare.28192152","title":"Additional file 3 of Transformation and articulation of clinical data to understand students’ clinical reasoning: a scoping review","year":2025,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières; Montreal Heart Institute; Université Laval; Université de Montréal; Centre for Interdisciplinary Research in Rehabilitation","funders":"","keywords":"Articulation (sociology); Transformation (genetics); Data transformation; Frame (networking)","score_opus":0.29310838999836447,"score_gpt":0.5252879853470106,"score_spread":0.23217959534864613,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6902225828","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00078700366,0.0010097366,0.000017545273,0.0004216022,0.000032393906,0.0007886034,0.9958359,0.000016839356,0.0010903929],"genre_scores_gemma":[0.0061900946,0.00034797026,0.0013431854,0.0012879433,0.000075950775,0.00007905017,0.99055684,0.000006484385,0.00011250877],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9981393,0.00009283801,0.0011394391,0.00025196068,0.00028494405,0.00009150676],"domain_scores_gemma":[0.9563824,0.04228528,0.00035128885,0.0005136086,0.00028686327,0.00018057026],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00047266186,0.000076509314,0.0005233487,0.00004345168,0.000022927525,0.000008398585,0.00012914426,0.000101893675,0.6419701],"category_scores_gemma":[0.40425295,0.00006651586,0.00010101064,0.00020854585,0.00002526382,0.00007179084,0.00012496623,0.00014410539,0.00008119234],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000057456862,0.00016695926,0.00058624125,0.0017662108,0.000042850257,0.0000017149116,0.000014213712,6.409019e-7,1.11263596e-7,0.000008055383,0.98645365,0.010901903],"study_design_scores_gemma":[0.000808437,0.00020983211,0.12615193,0.8487976,0.00011233803,0.0000029693974,0.000036002628,0.00033229325,0.000002941604,0.000015838605,0.023470057,0.000059729817],"about_ca_topic_score_codex":0.0000019312865,"about_ca_topic_score_gemma":0.000008469679,"teacher_disagreement_score":0.9629836,"about_ca_system_score_codex":0.000009468945,"about_ca_system_score_gemma":0.00037034866,"threshold_uncertainty_score":0.6007653},"labels":[],"label_agreement":null},{"id":"W6902349143","doi":"10.6084/m9.figshare.26620024","title":"Additional file 2 of History and physical exam: a retrospective analysis of a clinical opportunity","year":2024,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"NOSM University","funders":"","keywords":"Data collection; MEDLINE; Retrospective cohort study; Data file","score_opus":0.11415430286495376,"score_gpt":0.38410808090987963,"score_spread":0.2699537780449259,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6902349143","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004054443,0.00015779184,1.30543e-7,0.000034813038,0.000024111376,0.00006356655,0.9878752,0.00002849516,0.0077613983],"genre_scores_gemma":[0.11957311,0.0000050652816,0.0001136812,0.00012154582,0.00019865688,0.00009244815,0.87831545,0.000008662378,0.0015714028],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9991031,0.000036763868,0.0003286395,0.00022937886,0.00022293303,0.00007918238],"domain_scores_gemma":[0.9503038,0.04896818,0.00013308348,0.00019085275,0.00018909916,0.00021500349],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00006919013,0.00007338313,0.0005335098,0.00011205401,0.00000800024,0.0000028758495,0.000038125952,0.000085655796,0.98823106],"category_scores_gemma":[0.28693786,0.00006091371,0.0003646346,0.000291203,0.00005398008,0.000024900242,0.00004413149,0.00021614168,0.00014959388],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000020639814,0.00021184431,0.0003216475,0.00006633302,0.0004996736,0.00004384391,0.000058632315,2.3493438e-7,0.0000010938041,0.00001579592,0.9944729,0.004287343],"study_design_scores_gemma":[0.00019313768,0.0003238641,0.6019823,0.011283118,0.0007131786,0.0000051423185,0.000017909271,0.0027425515,0.0000039322854,0.000053969674,0.3826109,0.00007000736],"about_ca_topic_score_codex":0.000008345258,"about_ca_topic_score_gemma":0.000003831346,"teacher_disagreement_score":0.98808146,"about_ca_system_score_codex":0.000043252494,"about_ca_system_score_gemma":0.00048877276,"threshold_uncertainty_score":0.7190685},"labels":[],"label_agreement":null},{"id":"W6902433569","doi":"10.6084/m9.figshare.c.3645986_d1","title":"Additional file 2: of Cognitive biases associated with medical decisions: a systematic review","year":2016,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Cognitive bias; Cognition; Data collection; Cognitive interview","score_opus":0.08673324297405419,"score_gpt":0.3592790496283639,"score_spread":0.2725458066543097,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6902433569","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000126399555,0.0012298424,6.0580214e-7,0.00020996461,0.0000048643747,0.0004707938,0.996865,0.0000363135,0.0011700218],"genre_scores_gemma":[0.0022985537,0.00008401233,0.000026393413,0.0016479365,0.000038864422,0.0015085646,0.99349785,0.000012500311,0.0008853371],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.998587,0.00007859259,0.00038216557,0.00016248635,0.0006609084,0.000128832],"domain_scores_gemma":[0.04788004,0.9505542,0.00034570255,0.00021325506,0.0007030028,0.00030380196],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00009254715,0.00009720615,0.00054876396,0.000032360273,0.000020830292,0.0000032375235,0.000075381715,0.00009323197,0.99713445],"category_scores_gemma":[0.9960355,0.000047015175,0.000118425945,0.00018296838,0.000026135891,0.000025799218,0.000034569905,0.00008499803,0.0033210462],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00001458832,0.00014862153,0.000001844603,0.0018650822,0.00008497286,0.00009570221,0.0000025625163,2.3290276e-9,2.2149091e-8,4.472094e-7,0.99689406,0.00089207],"study_design_scores_gemma":[0.0006366173,0.00023368777,0.0013087039,0.9964064,0.00018737713,0.000066525085,0.000008316936,0.000004441236,0.000003403351,0.000006537318,0.0010679621,0.00007006563],"about_ca_topic_score_codex":6.0994404e-7,"about_ca_topic_score_gemma":0.0000038523376,"teacher_disagreement_score":0.99594295,"about_ca_system_score_codex":0.00002361486,"about_ca_system_score_gemma":0.00042075515,"threshold_uncertainty_score":0.997455},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"not_applicable","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W6902474608","doi":"10.6084/m9.figshare.28192146.v1","title":"Additional file 1 of Transformation and articulation of clinical data to understand students’ clinical reasoning: a scoping review","year":2025,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières; Montreal Heart Institute; Université Laval; Université de Montréal; Centre for Interdisciplinary Research in Rehabilitation","funders":"","keywords":"Articulation (sociology); Transformation (genetics); Data transformation; Frame (networking)","score_opus":0.2973298541209662,"score_gpt":0.5259346920607099,"score_spread":0.22860483793974368,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6902474608","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0007798058,0.0010025229,0.000017411863,0.00042250386,0.000032031137,0.0007868023,0.9958602,0.000016820828,0.0010818621],"genre_scores_gemma":[0.0061070146,0.00034223104,0.0013537847,0.0012847113,0.00007489295,0.00008020263,0.9906406,0.000006463819,0.00011011546],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9981397,0.0000929218,0.0011389568,0.00025206295,0.00028488808,0.00009151295],"domain_scores_gemma":[0.95598704,0.04268069,0.00035165733,0.0005149506,0.00028484242,0.00018080098],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00047286347,0.000076519456,0.0005232984,0.00004346534,0.000022929382,0.000008382374,0.00012922079,0.00010190183,0.6473436],"category_scores_gemma":[0.40431383,0.00006653182,0.00010101671,0.00020861872,0.00002526213,0.00007173887,0.00012505132,0.00014413311,0.00008036532],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005709773,0.00016692851,0.0005851218,0.0017372444,0.000042856598,0.0000016945912,0.0000142040235,6.4030456e-7,1.1093181e-7,0.00000809797,0.9866384,0.010747594],"study_design_scores_gemma":[0.000797628,0.00021000029,0.12672494,0.8486626,0.00011323858,0.000002994185,0.00003614552,0.00033162374,0.0000028812424,0.000015786976,0.023042437,0.000059692793],"about_ca_topic_score_codex":0.0000019108734,"about_ca_topic_score_gemma":0.000008415851,"teacher_disagreement_score":0.963596,"about_ca_system_score_codex":0.000009433712,"about_ca_system_score_gemma":0.00037015602,"threshold_uncertainty_score":0.6007039},"labels":[],"label_agreement":null},{"id":"W6903401555","doi":"10.1051/pmed/2023009/pdf","title":"La physiothérapie française est-elle prête à intégrer des tests de concordance de script dans sa formation initiale ? Étude de faisabilité et d’acceptabilité","year":2023,"lang":"fr","type":"article","venue":"Springer Link (Chiba Institute of Technology)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institut du Savoir Montfort; University of Ottawa","funders":"","keywords":"Population; Medical screening; Concordance","score_opus":0.021281425556380146,"score_gpt":0.3114242393840256,"score_spread":0.29014281382764545,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6903401555","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9755002,0.0045757964,0.0063574095,0.006232043,0.0013407347,0.0008305629,0.00008108902,0.001010987,0.00407118],"genre_scores_gemma":[0.95987254,0.004585851,0.033855464,0.0004821019,0.00053005386,0.00022736813,0.000045823923,0.000093242015,0.00030752533],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99612683,0.00018112893,0.0011796679,0.00074406376,0.00039946215,0.0013688467],"domain_scores_gemma":[0.9948145,0.002912482,0.00041235308,0.0011208582,0.00030021238,0.0004395793],"candidate_categories":["metaresearch","metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.0021317485,0.0005141127,0.00094177754,0.00049969,0.00026582446,0.00006599539,0.0007151581,0.00129004,0.00012744103],"category_scores_gemma":[0.062046867,0.0005700736,0.0003661297,0.0017442635,0.0038428379,0.00052041817,0.0005294213,0.0016086027,0.00025780633],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00028487778,0.0020656092,0.37221923,0.002579425,0.00038736887,0.0007049208,0.0074544554,0.0036162825,0.034284815,0.030106647,0.0012274462,0.5450689],"study_design_scores_gemma":[0.007569133,0.0016562523,0.7680532,0.02819909,0.00089568616,0.00072633935,0.0032607606,0.015375033,0.07358381,0.03149274,0.067772776,0.0014152147],"about_ca_topic_score_codex":0.002554097,"about_ca_topic_score_gemma":0.0030680504,"teacher_disagreement_score":0.54365367,"about_ca_system_score_codex":0.0006055502,"about_ca_system_score_gemma":0.001341192,"threshold_uncertainty_score":0.9996751},"labels":[],"label_agreement":null},{"id":"W6920629959","doi":"10.6084/m9.figshare.16994246.v1","title":"Additional file 3 of Improved preventive care clinical decision-making efficiency: leveraging a point-of-care clinical decision support system","year":2021,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Preventive care; Decision support system; Clinical decision support system; Preventive maintenance; Usability; Decision aids","score_opus":0.03598883780067279,"score_gpt":0.377915765154634,"score_spread":0.3419269273539612,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6920629959","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0032064633,0.000544477,0.00007773124,0.000024517867,0.00037230144,0.00048215216,0.99196607,0.00008149766,0.0032447749],"genre_scores_gemma":[0.30115777,0.000008978699,0.015367507,0.00022085918,0.0005089639,0.0002064801,0.6822812,0.00004657082,0.00020170523],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.99571943,0.00023803668,0.0021022987,0.00081666006,0.000762913,0.00036063325],"domain_scores_gemma":[0.7985145,0.19706853,0.00088001695,0.00083124056,0.0023338844,0.00037186316],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0003586602,0.0002616468,0.0011431412,0.000103253624,0.00009514278,0.00002662359,0.00027179316,0.00043519863,0.8828308],"category_scores_gemma":[0.5670967,0.00023088529,0.0010855945,0.000407757,0.000061864346,0.00007295091,0.00040513565,0.000595919,0.0008631111],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00030138157,0.000344304,0.0005905344,0.0004112097,0.000076607255,0.00030848762,0.00016986992,0.000006474716,8.817605e-7,0.000002182241,0.8601895,0.13759854],"study_design_scores_gemma":[0.006290741,0.0029746261,0.13052893,0.59529465,0.00044401424,0.00032819976,0.007088986,0.0024430125,0.00016237721,0.00007929974,0.25363666,0.00072851905],"about_ca_topic_score_codex":0.0000021555363,"about_ca_topic_score_gemma":0.000008752214,"teacher_disagreement_score":0.8819677,"about_ca_system_score_codex":0.000117008756,"about_ca_system_score_gemma":0.002119451,"threshold_uncertainty_score":0.9999148},"labels":[],"label_agreement":null},{"id":"W6920734996","doi":"10.6084/m9.figshare.28192158.v1","title":"Additional file 4 of Transformation and articulation of clinical data to understand students’ clinical reasoning: a scoping review","year":2025,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières; Montreal Heart Institute; Université Laval; Université de Montréal; Centre for Interdisciplinary Research in Rehabilitation","funders":"","keywords":"Articulation (sociology); Transformation (genetics); Data transformation; Frame (networking)","score_opus":0.2947793660154213,"score_gpt":0.5258061109761897,"score_spread":0.23102674496076836,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6920734996","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00078460795,0.0010145761,0.00001767363,0.00042144046,0.000032163895,0.00079029985,0.99583316,0.000016883385,0.0010891783],"genre_scores_gemma":[0.006336923,0.00034793772,0.001344426,0.0012891432,0.00007522573,0.00007904845,0.990412,0.0000064822666,0.00010882596],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9981389,0.00009282644,0.0011398222,0.0002519842,0.00028492682,0.00009151645],"domain_scores_gemma":[0.95650095,0.042177167,0.0003498304,0.00051164965,0.00028062,0.00017980077],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00047301978,0.000076514334,0.0005234652,0.00004347075,0.00002293829,0.0000084028015,0.0001292339,0.0001019018,0.6383016],"category_scores_gemma":[0.40105814,0.000066527784,0.00010101802,0.00020852072,0.000025268104,0.000071800736,0.00012501118,0.00014410117,0.000079627695],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000058084777,0.00016722368,0.00058812683,0.0017551134,0.000042898442,0.0000017224438,0.000014243009,6.331659e-7,1.09775755e-7,0.000008341189,0.9864575,0.010905971],"study_design_scores_gemma":[0.0008052519,0.00021316041,0.12663455,0.849213,0.00011311208,0.0000029978548,0.000036098416,0.00033331182,0.0000028839388,0.00001598697,0.022569941,0.000059729766],"about_ca_topic_score_codex":0.000001915106,"about_ca_topic_score_gemma":0.000008502458,"teacher_disagreement_score":0.9638876,"about_ca_system_score_codex":0.000009434566,"about_ca_system_score_gemma":0.00036903948,"threshold_uncertainty_score":0.603987},"labels":[],"label_agreement":null},{"id":"W6920805920","doi":"10.6084/m9.figshare.28192149","title":"Additional file 2 of Transformation and articulation of clinical data to understand students’ clinical reasoning: a scoping review","year":2025,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières; Montreal Heart Institute; Université Laval; Université de Montréal; Centre for Interdisciplinary Research in Rehabilitation","funders":"","keywords":"Articulation (sociology); Transformation (genetics); Data transformation; Frame (networking)","score_opus":0.2896117403072626,"score_gpt":0.5251741820055059,"score_spread":0.2355624416982433,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6920805920","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00077728176,0.0010205748,0.000017115108,0.00042175813,0.00003230426,0.0007879092,0.99581796,0.00001686869,0.0011082077],"genre_scores_gemma":[0.006009793,0.0003503489,0.0013249076,0.0012836806,0.00007586266,0.00007869964,0.9907599,0.0000064498736,0.00011036019],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9981395,0.00009289281,0.0011394377,0.00025188777,0.00028482286,0.00009148189],"domain_scores_gemma":[0.95671016,0.041964047,0.00035068914,0.00051194476,0.00028309927,0.00018004353],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00047045527,0.00007649675,0.0005234179,0.000043438788,0.00002293095,0.000008399349,0.00012915368,0.000101902384,0.643281],"category_scores_gemma":[0.40158778,0.00006651153,0.000100998484,0.0002084674,0.000025261805,0.000071784736,0.0001250349,0.00014410214,0.00007956504],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000056797657,0.00016684845,0.0005677,0.0017516008,0.000042849453,0.0000017375218,0.000014212713,6.430627e-7,1.0732465e-7,0.0000082294155,0.9863665,0.011022749],"study_design_scores_gemma":[0.0008063375,0.0002109797,0.12556848,0.84889495,0.00011336452,0.0000030139051,0.000035962497,0.00033104053,0.000002877426,0.000015975898,0.023957156,0.00005984437],"about_ca_topic_score_codex":0.0000019288354,"about_ca_topic_score_gemma":0.000008462318,"teacher_disagreement_score":0.9624094,"about_ca_system_score_codex":0.000009436234,"about_ca_system_score_gemma":0.00036861925,"threshold_uncertainty_score":0.60345286},"labels":[],"label_agreement":null},{"id":"W6920824370","doi":"10.6084/m9.figshare.13544452","title":"Additional file 1 of Understanding the clinical reasoning processes involved in the management of multimorbidity in an ambulatory setting: study protocol of a stimulated recall research","year":2021,"lang":"en","type":"article","venue":"Open MIND","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Protocol (science); Recall; Ambulatory; Interview; Protocol analysis; Ambulatory care; Patient care","score_opus":0.4009102997418652,"score_gpt":0.5381676943352826,"score_spread":0.13725739459341735,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6920824370","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9262558,0.000010952848,0.0000064311,0.00028289625,0.000022249998,0.064092554,0.0048848707,0.0000020484658,0.0044421777],"genre_scores_gemma":[0.9744114,0.0000030308393,0.005597547,0.000030198664,0.0000328317,0.018109458,0.0016473458,0.000012426737,0.00015576051],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9969769,0.0010008252,0.0008808288,0.00033426387,0.0006215449,0.0001856324],"domain_scores_gemma":[0.95855314,0.040143773,0.00029759295,0.0005220116,0.0004177258,0.000065741544],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004910782,0.00008996965,0.0003837269,0.00007510068,0.00005180372,0.000025642952,0.00036065612,0.00008123811,0.0479825],"category_scores_gemma":[0.04609401,0.0000564094,0.000051651205,0.0008613017,0.00024668343,0.00006293834,0.0002803352,0.00045719702,0.000005258221],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008812946,0.057990592,0.63932407,0.0017858767,0.0008409517,0.0018663803,0.025710097,0.0007525278,0.00007641476,0.00014278998,0.20079443,0.06190292],"study_design_scores_gemma":[0.010154662,0.0025570977,0.85642284,0.0269012,0.00011518354,0.0000067376604,0.09707305,0.0023209455,0.00023808409,0.00038784504,0.0036499912,0.00017238897],"about_ca_topic_score_codex":0.00007845002,"about_ca_topic_score_gemma":0.0006017238,"teacher_disagreement_score":0.21709874,"about_ca_system_score_codex":0.00005761769,"about_ca_system_score_gemma":0.000756353,"threshold_uncertainty_score":0.9619411},"labels":[],"label_agreement":null},{"id":"W6920867414","doi":"10.6084/m9.figshare.24172078","title":"Additional file 2 of Influence of research evidence on the use of cardiovascular clinical prediction rules in primary care: an exploratory qualitative interview study","year":2023,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Nipissing University","funders":"","keywords":"Exploratory research; Qualitative research; Interview; Primary care; Data collection; MEDLINE","score_opus":0.5531494764241607,"score_gpt":0.5088264211759751,"score_spread":0.04432305524818558,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6920867414","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.26997367,0.00024508825,4.8098663e-8,0.000029684115,0.000011106912,0.0005170615,0.7291689,0.00001601174,0.00003845669],"genre_scores_gemma":[0.47048208,0.00012918097,0.00008612098,0.00009934422,0.000096668904,0.0017072175,0.5273401,0.000019592075,0.000039643106],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9959724,0.0020360246,0.00067565014,0.00027211977,0.0009105785,0.00013320883],"domain_scores_gemma":[0.8062357,0.1919817,0.00016293245,0.00056206796,0.0009841243,0.00007351895],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0018756939,0.000082571765,0.00044035562,0.00014475315,0.000026360633,0.000005572362,0.00017663739,0.00008737536,0.19501466],"category_scores_gemma":[0.54604155,0.000058907757,0.00021128813,0.0006103625,0.000110451314,0.0001295465,0.00015543422,0.00038680268,0.00021513009],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015027694,0.00057971006,0.000756899,0.00046992695,0.00014415207,0.000028571667,0.0049306257,0.00023145958,0.000001476194,0.0000018574165,0.9867857,0.005919353],"study_design_scores_gemma":[0.0005359927,0.002687158,0.8163291,0.15963714,0.000044717963,8.8930375e-7,0.0087500885,0.00006487966,0.000026109112,0.000022012502,0.011836791,0.00006509293],"about_ca_topic_score_codex":0.000035794183,"about_ca_topic_score_gemma":0.000012705324,"teacher_disagreement_score":0.9749489,"about_ca_system_score_codex":0.000050099272,"about_ca_system_score_gemma":0.0004444781,"threshold_uncertainty_score":0.8057212},"labels":[],"label_agreement":null},{"id":"W6920941696","doi":"10.6084/m9.figshare.25744449","title":"Additional file 2 of Use of the International IFOMPT Cervical Framework to inform clinical reasoning in postgraduate level physiotherapy students: a qualitative study using think aloud methodology","year":2024,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Qualitative research; Think aloud protocol; Photo elicitation; Protocol analysis; MEDLINE; Grounded theory; Qualitative analysis","score_opus":0.3991223934647101,"score_gpt":0.556077493574945,"score_spread":0.15695510011023484,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6920941696","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08917117,0.000025416508,0.000026315009,0.00046470767,0.00019073935,0.0005304346,0.9094916,0.000025312427,0.00007428633],"genre_scores_gemma":[0.3178275,0.000009644167,0.10089293,0.00628315,0.0011327034,0.0005763918,0.5717768,0.0001184198,0.0013824266],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99756956,0.00042881942,0.00080510566,0.00031898622,0.000705695,0.00017186099],"domain_scores_gemma":[0.8371261,0.16173601,0.00025062286,0.00034848717,0.00041481951,0.00012398876],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0005328282,0.00014783199,0.0004898837,0.00012348885,0.000031755742,0.000036120728,0.000317381,0.00016155753,0.6201769],"category_scores_gemma":[0.3958023,0.00010414586,0.00024901598,0.0005400411,0.000039261882,0.00008555044,0.00029747776,0.00055149425,0.00015214583],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00031514614,0.0010305761,0.0015203326,0.000057175894,0.00037835844,0.000034581328,0.017662719,0.00004983985,0.0000059909544,0.000084991574,0.97603583,0.0028244562],"study_design_scores_gemma":[0.0018185392,0.0013979445,0.81019133,0.13117087,0.000104910236,0.00002918091,0.011675102,0.0028337499,0.000037481957,0.0011480032,0.039261412,0.00033150372],"about_ca_topic_score_codex":0.000103051716,"about_ca_topic_score_gemma":0.000040547922,"teacher_disagreement_score":0.93677443,"about_ca_system_score_codex":0.00007960639,"about_ca_system_score_gemma":0.00033692998,"threshold_uncertainty_score":0.60928714},"labels":[],"label_agreement":null},{"id":"W6920952525","doi":"10.6084/m9.figshare.28192152.v1","title":"Additional file 3 of Transformation and articulation of clinical data to understand students’ clinical reasoning: a scoping review","year":2025,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières; Montreal Heart Institute; Université Laval; Université de Montréal; Centre for Interdisciplinary Research in Rehabilitation","funders":"","keywords":"Articulation (sociology); Transformation (genetics); Data transformation; Frame (networking)","score_opus":0.29310838999836447,"score_gpt":0.5252879853470106,"score_spread":0.23217959534864613,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6920952525","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00078700366,0.0010097366,0.000017545273,0.0004216022,0.000032393906,0.0007886034,0.9958359,0.000016839356,0.0010903929],"genre_scores_gemma":[0.0061900946,0.00034797026,0.0013431854,0.0012879433,0.000075950775,0.00007905017,0.99055684,0.000006484385,0.00011250877],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9981393,0.00009283801,0.0011394391,0.00025196068,0.00028494405,0.00009150676],"domain_scores_gemma":[0.9563824,0.04228528,0.00035128885,0.0005136086,0.00028686327,0.00018057026],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00047266186,0.000076509314,0.0005233487,0.00004345168,0.000022927525,0.000008398585,0.00012914426,0.000101893675,0.6419701],"category_scores_gemma":[0.40425295,0.00006651586,0.00010101064,0.00020854585,0.00002526382,0.00007179084,0.00012496623,0.00014410539,0.00008119234],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000057456862,0.00016695926,0.00058624125,0.0017662108,0.000042850257,0.0000017149116,0.000014213712,6.409019e-7,1.11263596e-7,0.000008055383,0.98645365,0.010901903],"study_design_scores_gemma":[0.000808437,0.00020983211,0.12615193,0.8487976,0.00011233803,0.0000029693974,0.000036002628,0.00033229325,0.000002941604,0.000015838605,0.023470057,0.000059729817],"about_ca_topic_score_codex":0.0000019312865,"about_ca_topic_score_gemma":0.000008469679,"teacher_disagreement_score":0.9629836,"about_ca_system_score_codex":0.000009468945,"about_ca_system_score_gemma":0.00037034866,"threshold_uncertainty_score":0.6007653},"labels":[],"label_agreement":null},{"id":"W6921029184","doi":"10.6084/m9.figshare.28192161.v1","title":"Additional file 5 of Transformation and articulation of clinical data to understand students’ clinical reasoning: a scoping review","year":2025,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières; Montreal Heart Institute; Université Laval; Université de Montréal; Centre for Interdisciplinary Research in Rehabilitation","funders":"","keywords":"Articulation (sociology); Transformation (genetics); Data transformation; Frame (networking)","score_opus":0.29298845090034686,"score_gpt":0.5254013398466941,"score_spread":0.23241288894634726,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6921029184","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0007891263,0.0010146402,0.000017643184,0.0004240862,0.00003220454,0.00079109444,0.9958256,0.000016904982,0.001088713],"genre_scores_gemma":[0.0062829666,0.0003481512,0.0013452619,0.0012907706,0.00007528457,0.00007909682,0.9904623,0.0000064874835,0.000109666216],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.99813884,0.00009290034,0.0011397891,0.0002520149,0.0002849431,0.00009151441],"domain_scores_gemma":[0.9566012,0.04207644,0.00034938497,0.00051093346,0.0002824209,0.00017959684],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00047324444,0.00007652148,0.00052355,0.000043469103,0.000022939288,0.0000083993,0.00012918803,0.000101906466,0.63824266],"category_scores_gemma":[0.40196845,0.00006652895,0.00010103089,0.0002085842,0.000025268218,0.000071800634,0.00012501716,0.00014411716,0.0000791224],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005722066,0.00016720021,0.0005860889,0.0017552061,0.000042889606,0.0000017220015,0.000014232801,6.404408e-7,1.1052829e-7,0.000008343622,0.98644686,0.0109195085],"study_design_scores_gemma":[0.0008033929,0.0002101122,0.12816466,0.84791696,0.00011295291,0.000002992897,0.000036052494,0.00033051192,0.0000028999455,0.000015879057,0.022343982,0.00005963302],"about_ca_topic_score_codex":0.000001931517,"about_ca_topic_score_gemma":0.000008531282,"teacher_disagreement_score":0.96410286,"about_ca_system_score_codex":0.000009436514,"about_ca_system_score_gemma":0.00036943075,"threshold_uncertainty_score":0.603069},"labels":[],"label_agreement":null},{"id":"W6921085692","doi":"10.6084/m9.figshare.28192161","title":"Additional file 5 of Transformation and articulation of clinical data to understand students’ clinical reasoning: a scoping review","year":2025,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières; Montreal Heart Institute; Université Laval; Université de Montréal; Centre for Interdisciplinary Research in Rehabilitation","funders":"","keywords":"Articulation (sociology); Transformation (genetics); Data transformation; Frame (networking)","score_opus":0.29298845090034686,"score_gpt":0.5254013398466941,"score_spread":0.23241288894634726,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6921085692","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0007891263,0.0010146402,0.000017643184,0.0004240862,0.00003220454,0.00079109444,0.9958256,0.000016904982,0.001088713],"genre_scores_gemma":[0.0062829666,0.0003481512,0.0013452619,0.0012907706,0.00007528457,0.00007909682,0.9904623,0.0000064874835,0.000109666216],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.99813884,0.00009290034,0.0011397891,0.0002520149,0.0002849431,0.00009151441],"domain_scores_gemma":[0.9566012,0.04207644,0.00034938497,0.00051093346,0.0002824209,0.00017959684],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00047324444,0.00007652148,0.00052355,0.000043469103,0.000022939288,0.0000083993,0.00012918803,0.000101906466,0.63824266],"category_scores_gemma":[0.40196845,0.00006652895,0.00010103089,0.0002085842,0.000025268218,0.000071800634,0.00012501716,0.00014411716,0.0000791224],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005722066,0.00016720021,0.0005860889,0.0017552061,0.000042889606,0.0000017220015,0.000014232801,6.404408e-7,1.1052829e-7,0.000008343622,0.98644686,0.0109195085],"study_design_scores_gemma":[0.0008033929,0.0002101122,0.12816466,0.84791696,0.00011295291,0.000002992897,0.000036052494,0.00033051192,0.0000028999455,0.000015879057,0.022343982,0.00005963302],"about_ca_topic_score_codex":0.000001931517,"about_ca_topic_score_gemma":0.000008531282,"teacher_disagreement_score":0.96410286,"about_ca_system_score_codex":0.000009436514,"about_ca_system_score_gemma":0.00036943075,"threshold_uncertainty_score":0.603069},"labels":[],"label_agreement":null},{"id":"W6921128697","doi":"10.6084/m9.figshare.28192158","title":"Additional file 4 of Transformation and articulation of clinical data to understand students’ clinical reasoning: a scoping review","year":2025,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières; Montreal Heart Institute; Université Laval; Université de Montréal; Centre for Interdisciplinary Research in Rehabilitation","funders":"","keywords":"Articulation (sociology); Transformation (genetics); Data transformation; Frame (networking)","score_opus":0.2947793660154213,"score_gpt":0.5258061109761897,"score_spread":0.23102674496076836,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6921128697","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00078460795,0.0010145761,0.00001767363,0.00042144046,0.000032163895,0.00079029985,0.99583316,0.000016883385,0.0010891783],"genre_scores_gemma":[0.006336923,0.00034793772,0.001344426,0.0012891432,0.00007522573,0.00007904845,0.990412,0.0000064822666,0.00010882596],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9981389,0.00009282644,0.0011398222,0.0002519842,0.00028492682,0.00009151645],"domain_scores_gemma":[0.95650095,0.042177167,0.0003498304,0.00051164965,0.00028062,0.00017980077],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00047301978,0.000076514334,0.0005234652,0.00004347075,0.00002293829,0.0000084028015,0.0001292339,0.0001019018,0.6383016],"category_scores_gemma":[0.40105814,0.000066527784,0.00010101802,0.00020852072,0.000025268104,0.000071800736,0.00012501118,0.00014410117,0.000079627695],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000058084777,0.00016722368,0.00058812683,0.0017551134,0.000042898442,0.0000017224438,0.000014243009,6.331659e-7,1.09775755e-7,0.000008341189,0.9864575,0.010905971],"study_design_scores_gemma":[0.0008052519,0.00021316041,0.12663455,0.849213,0.00011311208,0.0000029978548,0.000036098416,0.00033331182,0.0000028839388,0.00001598697,0.022569941,0.000059729766],"about_ca_topic_score_codex":0.000001915106,"about_ca_topic_score_gemma":0.000008502458,"teacher_disagreement_score":0.9638876,"about_ca_system_score_codex":0.000009434566,"about_ca_system_score_gemma":0.00036903948,"threshold_uncertainty_score":0.603987},"labels":[],"label_agreement":null},{"id":"W6921143530","doi":"10.6084/m9.figshare.28192146","title":"Additional file 1 of Transformation and articulation of clinical data to understand students’ clinical reasoning: a scoping review","year":2025,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières; Montreal Heart Institute; Université Laval; Université de Montréal; Centre for Interdisciplinary Research in Rehabilitation","funders":"","keywords":"Articulation (sociology); Transformation (genetics); Data transformation; Frame (networking)","score_opus":0.2973298541209662,"score_gpt":0.5259346920607099,"score_spread":0.22860483793974368,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6921143530","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0007798058,0.0010025229,0.000017411863,0.00042250386,0.000032031137,0.0007868023,0.9958602,0.000016820828,0.0010818621],"genre_scores_gemma":[0.0061070146,0.00034223104,0.0013537847,0.0012847113,0.00007489295,0.00008020263,0.9906406,0.000006463819,0.00011011546],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9981397,0.0000929218,0.0011389568,0.00025206295,0.00028488808,0.00009151295],"domain_scores_gemma":[0.95598704,0.04268069,0.00035165733,0.0005149506,0.00028484242,0.00018080098],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00047286347,0.000076519456,0.0005232984,0.00004346534,0.000022929382,0.000008382374,0.00012922079,0.00010190183,0.6473436],"category_scores_gemma":[0.40431383,0.00006653182,0.00010101671,0.00020861872,0.00002526213,0.00007173887,0.00012505132,0.00014413311,0.00008036532],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005709773,0.00016692851,0.0005851218,0.0017372444,0.000042856598,0.0000016945912,0.0000142040235,6.4030456e-7,1.1093181e-7,0.00000809797,0.9866384,0.010747594],"study_design_scores_gemma":[0.000797628,0.00021000029,0.12672494,0.8486626,0.00011323858,0.000002994185,0.00003614552,0.00033162374,0.0000028812424,0.000015786976,0.023042437,0.000059692793],"about_ca_topic_score_codex":0.0000019108734,"about_ca_topic_score_gemma":0.000008415851,"teacher_disagreement_score":0.963596,"about_ca_system_score_codex":0.000009433712,"about_ca_system_score_gemma":0.00037015602,"threshold_uncertainty_score":0.6007039},"labels":[],"label_agreement":null},{"id":"W6929805333","doi":"10.5281/zenodo.10306503","title":"Fig. 4 in A revision of the Australian genus Pleuroseta Richards (Diptera, Sphaeroceridae)","year":2023,"lang":"en","type":"other","venue":"Zenodo (CERN European Organization for Nuclear Research)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Guelph","funders":"","keywords":"Paratype; Genus; Holotype; Dorsum; Scale (ratio)","score_opus":0.04774457628095218,"score_gpt":0.31081449412766976,"score_spread":0.26306991784671757,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6929805333","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.011529904,0.00067327387,0.00007268745,0.003999292,0.0008281632,0.0022603134,0.001318482,0.0021002963,0.9772176],"genre_scores_gemma":[0.17486402,0.0026748152,0.00021529685,0.0005697626,0.00085185864,1.15816384e-7,0.0018401225,0.01933804,0.79964596],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9981513,0.00022326161,0.0003973101,0.0004192254,0.00050417,0.00030471096],"domain_scores_gemma":[0.99851143,0.00020056317,0.00022753474,0.0007147721,0.00017537939,0.00017028957],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00055785605,0.00018414717,0.00037355066,0.00026882262,0.00020623219,0.000088166504,0.00061726524,0.00019069729,0.022832187],"category_scores_gemma":[0.018026344,0.00014556867,0.00013153507,0.0007638825,0.00016889932,0.000025713845,0.000709382,0.00048176834,0.0054940837],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000052378786,0.00013711075,0.00009104325,0.0001617541,0.00004056808,0.00004995232,0.00018753493,0.000003370747,0.00007360745,0.00010779654,0.92864585,0.07044903],"study_design_scores_gemma":[0.0008275134,0.00018159572,0.006808554,0.0032874157,0.00005053928,0.000047098627,0.000043890803,0.000019819689,0.000043502583,0.000011043456,0.9885515,0.00012755],"about_ca_topic_score_codex":0.00014780347,"about_ca_topic_score_gemma":0.0000040930445,"teacher_disagreement_score":0.17757162,"about_ca_system_score_codex":0.000104619285,"about_ca_system_score_gemma":0.000011535339,"threshold_uncertainty_score":0.99528027},"labels":[],"label_agreement":null},{"id":"W6930467774","doi":"10.5281/zenodo.12493407","title":"visa application form canada pdf","year":2024,"lang":"en","type":"other","venue":"Zenodo (CERN European Organization for Nuclear Research)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Checklist; Download; Visitor pattern; Immigration; Minor (academic); Section (typography); Schedule; Refugee","score_opus":0.02085679072678947,"score_gpt":0.2782321512359192,"score_spread":0.25737536050912974,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6930467774","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000049793525,0.00045122724,0.0006112732,0.0014271155,0.0002342131,0.000582531,0.00032123533,0.00086820376,0.99545443],"genre_scores_gemma":[0.008532775,0.00031897618,0.00016465997,0.000716212,0.0008994025,1.5046074e-7,0.0065185702,0.008702391,0.97414684],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9984539,0.000058012913,0.00024778364,0.00048438457,0.00046263103,0.0002932551],"domain_scores_gemma":[0.9987222,0.00012420371,0.000108531596,0.0005517056,0.00019437881,0.0002989955],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0002682061,0.00018287596,0.0002633314,0.00018524661,0.00025891632,0.00019646886,0.00037045855,0.00016449459,0.08778076],"category_scores_gemma":[0.0051815533,0.00017391483,0.00007089349,0.0003464092,0.000088691144,0.000020265323,0.0004038308,0.00041273516,0.085500136],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000019103376,0.00006406244,0.0000014403103,0.00015466711,0.000073390846,0.00004327087,0.000040271338,4.2779976e-7,0.000027793629,0.0007332482,0.94000036,0.058841992],"study_design_scores_gemma":[0.00037528833,0.00012220592,0.000082284256,0.0008698784,0.000091798145,0.00011528454,0.000043678432,0.00006563025,0.000017384771,0.00011479215,0.99793667,0.00016508353],"about_ca_topic_score_codex":0.019073263,"about_ca_topic_score_gemma":0.00087656535,"teacher_disagreement_score":0.05867691,"about_ca_system_score_codex":0.00028383356,"about_ca_system_score_gemma":0.000034461038,"threshold_uncertainty_score":0.9874588},"labels":[],"label_agreement":null},{"id":"W6930636901","doi":"10.5281/zenodo.15156731","title":"Table 4 in A revision of Scipopus Enderlein including the subgenera Scipopus s. str., Phaeopterina Frey and Parascipopus subgen. nov. (Diptera, Micropezidae, Taeniapterinae)","year":2023,"lang":"en","type":"dataset","venue":"Zenodo (CERN European Organization for Nuclear Research)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Guelph","funders":"","keywords":"Subgenus; Table (database); Character (mathematics); Genus; Taxonomy (biology)","score_opus":0.0671263964919024,"score_gpt":0.32817142932430426,"score_spread":0.2610450328324019,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6930636901","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.29837716,0.0030670834,0.000082285274,0.0018674675,0.00094362785,0.0027967205,0.6903721,0.0005925367,0.0019010252],"genre_scores_gemma":[0.12917796,0.008969699,0.0003099784,0.00034525397,0.0005083518,5.4834055e-7,0.85829824,0.0015714043,0.0008185418],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99620587,0.0004998508,0.000945789,0.0009201694,0.0007707215,0.00065759674],"domain_scores_gemma":[0.9969294,0.00072314945,0.00041786034,0.0011723961,0.00042046618,0.00033673033],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.002125455,0.0004049956,0.00077863946,0.00057670404,0.0010403997,0.0004615108,0.0010649564,0.0002884834,0.0013923829],"category_scores_gemma":[0.016636262,0.00033730114,0.00012664529,0.001267609,0.00047656617,0.00014245318,0.0026729042,0.00092854764,0.002289599],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00038771963,0.00026088237,0.00006290121,0.00038406864,0.00008787108,0.00020223467,0.00019813527,0.000009688874,0.00032127617,0.000036219226,0.8974841,0.10056486],"study_design_scores_gemma":[0.0014815524,0.0009144198,0.0011576456,0.002325815,0.00012633596,0.00031276367,0.00012205661,0.00006094238,0.00016336706,0.000015475096,0.99300855,0.00031106052],"about_ca_topic_score_codex":0.0016503565,"about_ca_topic_score_gemma":0.00003841808,"teacher_disagreement_score":0.16919918,"about_ca_system_score_codex":0.00018388493,"about_ca_system_score_gemma":0.0000333073,"threshold_uncertainty_score":0.9999079},"labels":[],"label_agreement":null},{"id":"W6931128238","doi":"10.5281/zenodo.15777840","title":"Alliance Federation Cybersecurity - 2025 DRI Connect poster","year":2025,"lang":"en","type":"article","venue":"Zenodo (CERN European Organization for Nuclear Research)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Toronto Dementia Research Alliance","funders":"","keywords":"Alliance; Government (linguistics); Digital government; Espionage; E-Government","score_opus":0.03777323985436637,"score_gpt":0.3184224725353174,"score_spread":0.28064923268095104,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6931128238","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.20532973,0.0004973442,0.0052534435,0.020440875,0.00050968555,0.0012571826,0.00014118008,0.0014178425,0.7651527],"genre_scores_gemma":[0.9884791,0.00018280876,0.00023976577,0.0022003935,0.00015129072,6.2759725e-8,0.0012815867,0.00036281298,0.0071021775],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9986944,0.00016999774,0.00026623692,0.00036299918,0.0002499798,0.00025634616],"domain_scores_gemma":[0.99845874,0.00041226207,0.00006188851,0.00037823382,0.0005234953,0.00016539519],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0005232012,0.0001132102,0.00018722698,0.0000962888,0.001025008,0.00037870908,0.00026026697,0.00008798217,0.0074550267],"category_scores_gemma":[0.018978486,0.00011116698,0.00007083416,0.00038595326,0.00011933015,0.00009566083,0.0003488285,0.00027579756,0.0047042826],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00037418335,0.0005293478,0.00026087786,0.000101169,0.00012359889,0.000054027707,0.0005383369,0.000014470467,0.0018330553,0.013882133,0.8492931,0.13299568],"study_design_scores_gemma":[0.0016237878,0.00033150974,0.010714413,0.00059776136,0.000060175706,0.000057203,0.00015531715,0.00034041374,0.0011174206,0.0005381433,0.9843151,0.00014875575],"about_ca_topic_score_codex":0.000020226647,"about_ca_topic_score_gemma":6.5853953e-7,"teacher_disagreement_score":0.78314936,"about_ca_system_score_codex":0.000116833435,"about_ca_system_score_gemma":0.000012501795,"threshold_uncertainty_score":0.9960707},"labels":[],"label_agreement":null},{"id":"W6931332210","doi":"10.5281/zenodo.6818056","title":"Crostis Casey 1922","year":2022,"lang":"en","type":"article","venue":"Zenodo (CERN European Organization for Nuclear Research)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Canadian Museum of Nature","funders":"","keywords":"Rostrum; Seta; Genus; Key (lock); Ridge; Apex (geometry)","score_opus":0.04612793135312898,"score_gpt":0.30290906145711627,"score_spread":0.2567811301039873,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6931332210","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.39744645,0.0005340631,0.0018820858,0.012190167,0.0009478573,0.0014308991,0.001161266,0.0028308607,0.58157635],"genre_scores_gemma":[0.990329,0.000048783277,0.00018201025,0.0011604595,0.00024159346,8.430059e-8,0.002360164,0.000930272,0.0047475914],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9985009,0.00021762491,0.00021660668,0.00032561139,0.00045184462,0.00028743912],"domain_scores_gemma":[0.9986936,0.00033065665,0.00006473558,0.0004022217,0.00024552626,0.00026324592],"candidate_categories":["metaresearch","sts","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0006696461,0.00009292197,0.00015592123,0.00012773833,0.0022737407,0.00018046728,0.00035890454,0.00003220383,0.06478787],"category_scores_gemma":[0.0147670135,0.00009724428,0.00007180042,0.00045123507,0.00011247756,0.000048521782,0.001087128,0.00039059462,0.0055093323],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002580273,0.00053723756,0.00009760113,0.000025784095,0.00005212411,0.00033544019,0.00065243745,0.00009730524,0.0007680786,0.0024788743,0.87967944,0.11501764],"study_design_scores_gemma":[0.0010339569,0.00070297223,0.0032774117,0.000047057554,0.000030322548,0.0009989563,0.0002696569,0.00020683582,0.000077012424,0.000109288696,0.99313825,0.000108265216],"about_ca_topic_score_codex":0.000036882015,"about_ca_topic_score_gemma":1.0315409e-7,"teacher_disagreement_score":0.5928826,"about_ca_system_score_codex":0.00016311214,"about_ca_system_score_gemma":0.000007987645,"threshold_uncertainty_score":0.99902517},"labels":[],"label_agreement":null},{"id":"W6931456960","doi":"10.5281/zenodo.4342787","title":"Myotis annectans Dobson 1871","year":2013,"lang":"en","type":"article","venue":"Zenodo (CERN European Organization for Nuclear Research)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Environment and Climate Change Canada","funders":"","keywords":"Cave; Distribution (mathematics); Population","score_opus":0.036170937265939035,"score_gpt":0.2853297541746276,"score_spread":0.24915881690868855,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6931456960","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5367533,0.00023079824,0.0014419454,0.0095131,0.00024545332,0.0011842926,0.00009634417,0.001849156,0.44868562],"genre_scores_gemma":[0.99293774,0.00012332277,0.00044087623,0.0008174499,0.000246167,7.49401e-8,0.0008280579,0.00091665104,0.003689641],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99860233,0.00012695961,0.00024630476,0.00034216786,0.00033435677,0.0003478705],"domain_scores_gemma":[0.99823874,0.0003283798,0.00006419071,0.00044166474,0.00055785204,0.00036917103],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00035481315,0.000118650256,0.00018759316,0.00012289408,0.0007444898,0.0003516514,0.000333589,0.00008089127,0.0482],"category_scores_gemma":[0.017943628,0.00010855273,0.00007495578,0.00037870635,0.0001352234,0.00012830914,0.0003692847,0.00027432863,0.043497853],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005845737,0.00035363712,0.00017067377,0.00003525068,0.000053334104,0.000038711583,0.000463095,0.0000061280175,0.002310059,0.0016079887,0.76281303,0.23208964],"study_design_scores_gemma":[0.0012095685,0.0006669013,0.03233127,0.000262117,0.000038773054,0.00023400449,0.0001945817,0.00034865167,0.00042533086,0.00032633808,0.9637885,0.0001739477],"about_ca_topic_score_codex":0.00006331293,"about_ca_topic_score_gemma":9.683266e-8,"teacher_disagreement_score":0.45618448,"about_ca_system_score_codex":0.000086050495,"about_ca_system_score_gemma":0.0000047028157,"threshold_uncertainty_score":0.99032867},"labels":[],"label_agreement":null},{"id":"W6939388591","doi":"10.6084/m9.figshare.18095993.v1","title":"Types of clinical reasoning in a summative clerkship oral examination","year":2022,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Toronto; University Health Network","funders":"","keywords":"Summative assessment; Objective structured clinical examination; Type (biology); Educational measurement; Clinical clerkship; Descriptive statistics","score_opus":0.14643905489476922,"score_gpt":0.42662026709917145,"score_spread":0.28018121220440223,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6939388591","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.81633484,0.00052668096,0.0000017474047,0.0004958287,0.00024434715,0.00080910674,0.15281256,0.000114645525,0.028660277],"genre_scores_gemma":[0.9471,0.0000025092427,0.00022071826,0.00020836468,0.00006721692,0.00012771202,0.051850285,0.000010802421,0.00041243184],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9988121,0.00021969553,0.00041906894,0.00017963207,0.00024637932,0.00012315488],"domain_scores_gemma":[0.9838346,0.015571556,0.00020579823,0.00019043048,0.00010461777,0.00009296978],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0005257813,0.00006425563,0.00030249593,0.00006834554,0.000030084142,0.0000042170077,0.000075978474,0.000060218954,0.22049506],"category_scores_gemma":[0.2534541,0.000060536757,0.00009298356,0.00026789613,0.00001012827,0.000027461798,0.00010643621,0.0003449108,0.000164216],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00047000535,0.0014063455,0.11370207,0.0001522338,0.00007501051,0.00029390302,0.0009729543,0.00019844716,0.000007910208,0.00023476323,0.6858295,0.19665684],"study_design_scores_gemma":[0.0015513162,0.0005576458,0.97354615,0.006768523,0.000021196394,0.000012407423,0.00028924007,0.0018986162,0.000045985988,0.000054448465,0.01516223,0.000092227885],"about_ca_topic_score_codex":0.00001922019,"about_ca_topic_score_gemma":0.000012183918,"teacher_disagreement_score":0.8598441,"about_ca_system_score_codex":0.000037750226,"about_ca_system_score_gemma":0.00015174648,"threshold_uncertainty_score":0.7802175},"labels":[],"label_agreement":null},{"id":"W6939497868","doi":"10.6084/m9.figshare.13698230","title":"Additional file 1 of Teaching clinical reasoning to undergraduate medical students by illness script method: a randomized controlled trial","year":2021,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Randomized controlled trial; MEDLINE; Undergraduate student; Informed consent; Patient care; Qualitative research","score_opus":0.03677306375266611,"score_gpt":0.41585689097651196,"score_spread":0.37908382722384587,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6939497868","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00028999895,0.00013411087,0.00009511452,0.0019324856,0.0003217474,0.0020811216,0.9887329,0.00007976481,0.006332792],"genre_scores_gemma":[0.0014514176,0.000010666989,0.009432847,0.0046016625,0.0009511906,0.004983764,0.9702465,0.000043110027,0.008278877],"study_design_codex":"not_applicable","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9937467,0.0018715928,0.0016200008,0.0005648717,0.0018564601,0.00034036633],"domain_scores_gemma":[0.41179353,0.5841684,0.0009152405,0.00073487865,0.0008790472,0.0015089246],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0025077625,0.0002445268,0.0026226887,0.000071808056,0.00010772069,0.000060510145,0.00031074102,0.0003368872,0.9750073],"category_scores_gemma":[0.94194925,0.00018476097,0.0011083105,0.0002039417,0.00004043036,0.000041989275,0.0002997392,0.00078939484,0.00093977683],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.20630622,0.00082225655,0.000004236127,0.00001925912,0.00040408026,0.0001400112,0.00002222864,7.08307e-7,5.503211e-7,0.000016984908,0.78671,0.0055534462],"study_design_scores_gemma":[0.677008,0.00025824548,0.00006214678,0.027591132,0.00021275654,0.000038056016,0.000045337176,0.00041315396,0.00000891006,0.000059846792,0.29412994,0.00017250363],"about_ca_topic_score_codex":0.000011643659,"about_ca_topic_score_gemma":0.0000048730117,"teacher_disagreement_score":0.97406757,"about_ca_system_score_codex":0.000042317864,"about_ca_system_score_gemma":0.0010673192,"threshold_uncertainty_score":0.9998381},"labels":[],"label_agreement":null},{"id":"W6939536959","doi":"10.6084/m9.figshare.12096669","title":"Additional file 1 of Mapping clinical reasoning literature across the health professions: a scoping review","year":2020,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Interdisciplinary Research in Rehabilitation; McGill University","funders":"","keywords":"MEDLINE; Systematic review; Case-based reasoning; Health care; Key (lock)","score_opus":0.16604901625195273,"score_gpt":0.46514217969229604,"score_spread":0.2990931634403433,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6939536959","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000057529437,0.0189559,7.3212465e-7,0.01015262,0.000029414912,0.0007983292,0.9695876,0.000056652287,0.00041296362],"genre_scores_gemma":[0.0001424615,0.0006945646,0.0010255324,0.03696391,0.00059761864,0.0007114524,0.9596514,0.00001743681,0.00019563476],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.99827963,0.00017356408,0.0007193779,0.00029303177,0.0002869278,0.000247439],"domain_scores_gemma":[0.9334514,0.064447425,0.00072667375,0.00046532138,0.00039540097,0.00051379134],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00023695073,0.000122835,0.00053740386,0.000010080195,0.00013997487,0.000019942147,0.00014663563,0.00011194824,0.9359281],"category_scores_gemma":[0.5971156,0.00007777858,0.000259934,0.00040666485,0.000023983941,0.000043168424,0.00014992128,0.0006158242,0.0009004994],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00001580036,0.00004393922,0.000020387872,0.0037685398,0.000016787666,0.000022999002,0.00019771155,1.13747056e-7,5.563445e-8,0.000001171657,0.9870677,0.008844804],"study_design_scores_gemma":[0.00021065833,0.00013676833,0.0024822687,0.7543318,0.0000065803692,0.000023025825,0.000066441666,0.000050506023,5.15633e-7,0.0000025749391,0.24263193,0.00005690687],"about_ca_topic_score_codex":9.151192e-7,"about_ca_topic_score_gemma":9.218281e-7,"teacher_disagreement_score":0.9350276,"about_ca_system_score_codex":0.000014837348,"about_ca_system_score_gemma":0.0007384509,"threshold_uncertainty_score":0.9998774},"labels":[],"label_agreement":null},{"id":"W6939673608","doi":"10.6084/m9.figshare.16994240","title":"Additional file 1 of Improved preventive care clinical decision-making efficiency: leveraging a point-of-care clinical decision support system","year":2021,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Preventive care; Decision support system; Clinical decision support system; Preventive maintenance; MEDLINE; Health care","score_opus":0.03661142306253171,"score_gpt":0.37843655053314385,"score_spread":0.3418251274706121,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6939673608","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0031811562,0.00054124574,0.00007723483,0.000024597002,0.0003686041,0.00048160378,0.99202067,0.00008150014,0.0032233854],"genre_scores_gemma":[0.29833394,0.000008844067,0.01550467,0.0002205966,0.00050262484,0.00020971554,0.6849754,0.000046484984,0.00019773029],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9957199,0.00023824903,0.0021014283,0.0008169889,0.00076276757,0.00036065828],"domain_scores_gemma":[0.7972338,0.19836785,0.000879725,0.0008322437,0.0023145543,0.00037182347],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00035881103,0.000261681,0.0011430329,0.000103285616,0.00009515037,0.000026572943,0.00027195184,0.00043523294,0.88520116],"category_scores_gemma":[0.56715786,0.00023093988,0.0010856588,0.00040789734,0.000061860264,0.00007289887,0.00040540745,0.00059603195,0.00085445255],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00029937044,0.0003440634,0.00058911834,0.00040434545,0.000076578806,0.00030472816,0.000169668,0.0000064654237,8.787145e-7,0.0000021924766,0.8619615,0.13584109],"study_design_scores_gemma":[0.006227198,0.0029862153,0.13157564,0.5977914,0.00044891128,0.00033192674,0.007138743,0.0024457325,0.00015958697,0.00007929038,0.25008506,0.0007303344],"about_ca_topic_score_codex":0.0000021330804,"about_ca_topic_score_gemma":0.000008697391,"teacher_disagreement_score":0.8843467,"about_ca_system_score_codex":0.00011657971,"about_ca_system_score_gemma":0.0021183665,"threshold_uncertainty_score":0.9999235},"labels":[],"label_agreement":null},{"id":"W6939703834","doi":"10.6084/m9.figshare.25744449.v1","title":"Additional file 2 of Use of the International IFOMPT Cervical Framework to inform clinical reasoning in postgraduate level physiotherapy students: a qualitative study using think aloud methodology","year":2024,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Qualitative research; Think aloud protocol; Photo elicitation; Protocol analysis; MEDLINE; Grounded theory; Qualitative analysis","score_opus":0.3991223934647101,"score_gpt":0.556077493574945,"score_spread":0.15695510011023484,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6939703834","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08917117,0.000025416508,0.000026315009,0.00046470767,0.00019073935,0.0005304346,0.9094916,0.000025312427,0.00007428633],"genre_scores_gemma":[0.3178275,0.000009644167,0.10089293,0.00628315,0.0011327034,0.0005763918,0.5717768,0.0001184198,0.0013824266],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99756956,0.00042881942,0.00080510566,0.00031898622,0.000705695,0.00017186099],"domain_scores_gemma":[0.8371261,0.16173601,0.00025062286,0.00034848717,0.00041481951,0.00012398876],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0005328282,0.00014783199,0.0004898837,0.00012348885,0.000031755742,0.000036120728,0.000317381,0.00016155753,0.6201769],"category_scores_gemma":[0.3958023,0.00010414586,0.00024901598,0.0005400411,0.000039261882,0.00008555044,0.00029747776,0.00055149425,0.00015214583],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00031514614,0.0010305761,0.0015203326,0.000057175894,0.00037835844,0.000034581328,0.017662719,0.00004983985,0.0000059909544,0.000084991574,0.97603583,0.0028244562],"study_design_scores_gemma":[0.0018185392,0.0013979445,0.81019133,0.13117087,0.000104910236,0.00002918091,0.011675102,0.0028337499,0.000037481957,0.0011480032,0.039261412,0.00033150372],"about_ca_topic_score_codex":0.000103051716,"about_ca_topic_score_gemma":0.000040547922,"teacher_disagreement_score":0.93677443,"about_ca_system_score_codex":0.00007960639,"about_ca_system_score_gemma":0.00033692998,"threshold_uncertainty_score":0.60928714},"labels":[],"label_agreement":null},{"id":"W6958209092","doi":"10.6084/m9.figshare.13544452.v1","title":"Additional file 1 of Understanding the clinical reasoning processes involved in the management of multimorbidity in an ambulatory setting: study protocol of a stimulated recall research","year":2021,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Protocol (science); Recall; Ambulatory; Interview; Protocol analysis; Ambulatory care; Patient care","score_opus":0.3977611189845838,"score_gpt":0.5088048321177275,"score_spread":0.11104371313314376,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6958209092","genre_codex":"dataset","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.062363833,0.000036832767,9.547276e-7,0.00022183167,0.000020073594,0.06606217,0.86925703,0.000029721179,0.0020075336],"genre_scores_gemma":[0.7514434,0.0000017503397,0.0007055555,0.00007377641,0.000065786975,0.06575947,0.18187016,0.000020515956,0.000059607282],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99726355,0.00076513493,0.00077793404,0.0002954997,0.0006992644,0.0001986321],"domain_scores_gemma":[0.92246544,0.07572817,0.00035967195,0.00059742905,0.00076766405,0.00008159789],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0014393822,0.00009725851,0.00033766116,0.000085889806,0.00004716846,0.000012442203,0.00023338669,0.000098316275,0.3568687],"category_scores_gemma":[0.18090694,0.00006236244,0.00007216312,0.0009450119,0.000070460344,0.000044272198,0.00017088569,0.0005368407,0.000011327714],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006815091,0.0068261954,0.027310826,0.0019421095,0.000121866644,0.00037998534,0.0020916148,0.00015997932,0.0000029803068,0.000017389388,0.9589756,0.0014899437],"study_design_scores_gemma":[0.0049896706,0.0014478778,0.83933055,0.1219515,0.00004458351,0.000004841292,0.025631372,0.002289314,0.00006193183,0.00022691274,0.003882051,0.00013938692],"about_ca_topic_score_codex":0.000019117813,"about_ca_topic_score_gemma":0.00018647328,"teacher_disagreement_score":0.95509356,"about_ca_system_score_codex":0.000061211445,"about_ca_system_score_gemma":0.0006146427,"threshold_uncertainty_score":0.82599264},"labels":[],"label_agreement":null},{"id":"W6958341153","doi":"10.6084/m9.figshare.26715091.v1","title":"Additional file 3 of Use of the International IFOMPT Cervical Framework to inform clinical reasoning in postgraduate level physiotherapy students: a qualitative study using think aloud methodology","year":2024,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Qualitative research; Think aloud protocol; Photo elicitation; Protocol analysis; MEDLINE; Grounded theory; Qualitative analysis","score_opus":0.40269734729256557,"score_gpt":0.5561591851646095,"score_spread":0.15346183787204393,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6958341153","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.09005055,0.000025134888,0.000026918604,0.00046409594,0.0001910589,0.00053036783,0.90861356,0.000025246127,0.000073077],"genre_scores_gemma":[0.32567412,0.0000095582145,0.101951085,0.006287533,0.0011311702,0.0005774143,0.5628458,0.000118730735,0.0014045864],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9975694,0.00042857855,0.00080510665,0.00031907414,0.000705981,0.0001719055],"domain_scores_gemma":[0.8361666,0.16268942,0.00025080796,0.00034925525,0.00041969714,0.00012422375],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00053520815,0.00014785511,0.00048982206,0.00012352376,0.000031751224,0.0000361176,0.00031735896,0.00016154438,0.61889535],"category_scores_gemma":[0.39832842,0.00010415232,0.00024904453,0.0005402346,0.00003926487,0.00008555736,0.00029732214,0.000551506,0.00015510806],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003211158,0.001039278,0.0015798084,0.000058080583,0.0003813188,0.000034419183,0.017801963,0.000050068134,0.0000062485988,0.000083928906,0.9758273,0.002816481],"study_design_scores_gemma":[0.0018180734,0.0013869053,0.8108023,0.13147569,0.000103721264,0.000028691777,0.01165548,0.0028361902,0.000038173304,0.0011354982,0.038389307,0.00032999594],"about_ca_topic_score_codex":0.000103176426,"about_ca_topic_score_gemma":0.000040581515,"teacher_disagreement_score":0.937438,"about_ca_system_score_codex":0.00007986919,"about_ca_system_score_gemma":0.00033843546,"threshold_uncertainty_score":0.6067397},"labels":[],"label_agreement":null},{"id":"W6958360109","doi":"10.6084/m9.figshare.26715091","title":"Additional file 3 of Use of the International IFOMPT Cervical Framework to inform clinical reasoning in postgraduate level physiotherapy students: a qualitative study using think aloud methodology","year":2024,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Qualitative research; Think aloud protocol; Photo elicitation; Protocol analysis; MEDLINE; Grounded theory; Qualitative analysis","score_opus":0.40269734729256557,"score_gpt":0.5561591851646095,"score_spread":0.15346183787204393,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6958360109","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.09005055,0.000025134888,0.000026918604,0.00046409594,0.0001910589,0.00053036783,0.90861356,0.000025246127,0.000073077],"genre_scores_gemma":[0.32567412,0.0000095582145,0.101951085,0.006287533,0.0011311702,0.0005774143,0.5628458,0.000118730735,0.0014045864],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9975694,0.00042857855,0.00080510665,0.00031907414,0.000705981,0.0001719055],"domain_scores_gemma":[0.8361666,0.16268942,0.00025080796,0.00034925525,0.00041969714,0.00012422375],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00053520815,0.00014785511,0.00048982206,0.00012352376,0.000031751224,0.0000361176,0.00031735896,0.00016154438,0.61889535],"category_scores_gemma":[0.39832842,0.00010415232,0.00024904453,0.0005402346,0.00003926487,0.00008555736,0.00029732214,0.000551506,0.00015510806],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003211158,0.001039278,0.0015798084,0.000058080583,0.0003813188,0.000034419183,0.017801963,0.000050068134,0.0000062485988,0.000083928906,0.9758273,0.002816481],"study_design_scores_gemma":[0.0018180734,0.0013869053,0.8108023,0.13147569,0.000103721264,0.000028691777,0.01165548,0.0028361902,0.000038173304,0.0011354982,0.038389307,0.00032999594],"about_ca_topic_score_codex":0.000103176426,"about_ca_topic_score_gemma":0.000040581515,"teacher_disagreement_score":0.937438,"about_ca_system_score_codex":0.00007986919,"about_ca_system_score_gemma":0.00033843546,"threshold_uncertainty_score":0.6067397},"labels":[],"label_agreement":null},{"id":"W6958489783","doi":"10.6084/m9.figshare.24181056.v1","title":"Additional file 1 of Deliberate practice of diagnostic clinical reasoning reveals low performance and improvement of diagnostic justification in pre-clerkship students","year":2023,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Case-based reasoning; Clinical Practice; Quality (philosophy); Model-based reasoning","score_opus":0.045909498766610504,"score_gpt":0.37624463001771874,"score_spread":0.33033513125110825,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6958489783","genre_codex":"dataset","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3291979,0.00009412978,2.827975e-7,0.000078984005,0.000037603357,0.00054188585,0.66978645,0.00002336634,0.0002393987],"genre_scores_gemma":[0.58256465,0.00030737193,0.00037799918,0.00013389351,0.000088369925,0.00072818104,0.41546103,0.00001989798,0.00031861896],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9977883,0.00010212882,0.0010665794,0.00032051516,0.00050906715,0.00021337435],"domain_scores_gemma":[0.7163915,0.2820945,0.00066211977,0.00028693085,0.00043080395,0.00013409961],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00052386266,0.00013716391,0.00047373952,0.00011474747,0.000026303034,0.000009699984,0.00013985811,0.00014722743,0.3443907],"category_scores_gemma":[0.83406615,0.00012940635,0.00008606305,0.00043040994,0.00004902415,0.000110317764,0.00014633602,0.00025525034,0.0002025228],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015026273,0.00069335254,0.016610133,0.00070024125,0.000052472304,0.00002440214,0.00011849908,0.00006349589,0.000016576592,0.0000017692792,0.95924675,0.022322068],"study_design_scores_gemma":[0.0010649774,0.0007654625,0.9205912,0.07076813,0.000079252015,0.0000069334737,0.00006720054,0.0007942472,0.00015259701,0.000011556074,0.0055901473,0.00010832967],"about_ca_topic_score_codex":0.000016425505,"about_ca_topic_score_gemma":0.0000038624435,"teacher_disagreement_score":0.95365655,"about_ca_system_score_codex":0.000027135073,"about_ca_system_score_gemma":0.00018198481,"threshold_uncertainty_score":0.65620863},"labels":[],"label_agreement":null},{"id":"W6958643567","doi":"10.6084/m9.figshare.26620021","title":"Additional file 1 of History and physical exam: a retrospective analysis of a clinical opportunity","year":2024,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"NOSM University","funders":"","keywords":"Data collection; MEDLINE; Retrospective cohort study; Data file","score_opus":0.11839902393281639,"score_gpt":0.38489797813221455,"score_spread":0.2664989541993982,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6958643567","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004083909,0.00015510853,1.3363503e-7,0.00003500717,0.000023959665,0.000063686726,0.98803365,0.000028501607,0.0075760665],"genre_scores_gemma":[0.121354535,0.000004920123,0.0001163308,0.0001214234,0.0001953427,0.000094311676,0.8765401,0.000008667255,0.0015643451],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99910295,0.000036777332,0.0003284805,0.00022956466,0.00022299327,0.000079213896],"domain_scores_gemma":[0.9494766,0.04979321,0.00013315005,0.00019163407,0.00018993567,0.0002154702],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.000069600355,0.00007340834,0.0005333689,0.00011213328,0.000007999607,0.0000028691243,0.000038148883,0.00008565526,0.9884676],"category_scores_gemma":[0.2896285,0.00006093522,0.00036471075,0.0002914476,0.000053980883,0.000024881829,0.000044138196,0.00021619542,0.00015133648],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000020774425,0.00021204719,0.00033323542,0.00006572965,0.00049997086,0.000042609812,0.000058614394,2.3386256e-7,0.0000011369191,0.000015510528,0.9945855,0.004164623],"study_design_scores_gemma":[0.00019255915,0.0003252236,0.61094314,0.011484356,0.0007191129,0.0000051524926,0.000018188102,0.002774554,0.0000039762085,0.00005374383,0.37340954,0.000070473674],"about_ca_topic_score_codex":0.000008255382,"about_ca_topic_score_gemma":0.0000038070084,"teacher_disagreement_score":0.98831624,"about_ca_system_score_codex":0.000043239117,"about_ca_system_score_gemma":0.00049113156,"threshold_uncertainty_score":0.71635526},"labels":[],"label_agreement":null},{"id":"W6958863103","doi":"10.6084/m9.figshare.26620024.v1","title":"Additional file 2 of History and physical exam: a retrospective analysis of a clinical opportunity","year":2024,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"NOSM University","funders":"","keywords":"Data collection; MEDLINE; Retrospective cohort study; Data file","score_opus":0.11415430286495376,"score_gpt":0.38410808090987963,"score_spread":0.2699537780449259,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6958863103","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004054443,0.00015779184,1.30543e-7,0.000034813038,0.000024111376,0.00006356655,0.9878752,0.00002849516,0.0077613983],"genre_scores_gemma":[0.11957311,0.0000050652816,0.0001136812,0.00012154582,0.00019865688,0.00009244815,0.87831545,0.000008662378,0.0015714028],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9991031,0.000036763868,0.0003286395,0.00022937886,0.00022293303,0.00007918238],"domain_scores_gemma":[0.9503038,0.04896818,0.00013308348,0.00019085275,0.00018909916,0.00021500349],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00006919013,0.00007338313,0.0005335098,0.00011205401,0.00000800024,0.0000028758495,0.000038125952,0.000085655796,0.98823106],"category_scores_gemma":[0.28693786,0.00006091371,0.0003646346,0.000291203,0.00005398008,0.000024900242,0.00004413149,0.00021614168,0.00014959388],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000020639814,0.00021184431,0.0003216475,0.00006633302,0.0004996736,0.00004384391,0.000058632315,2.3493438e-7,0.0000010938041,0.00001579592,0.9944729,0.004287343],"study_design_scores_gemma":[0.00019313768,0.0003238641,0.6019823,0.011283118,0.0007131786,0.0000051423185,0.000017909271,0.0027425515,0.0000039322854,0.000053969674,0.3826109,0.00007000736],"about_ca_topic_score_codex":0.000008345258,"about_ca_topic_score_gemma":0.000003831346,"teacher_disagreement_score":0.98808146,"about_ca_system_score_codex":0.000043252494,"about_ca_system_score_gemma":0.00048877276,"threshold_uncertainty_score":0.7190685},"labels":[],"label_agreement":null},{"id":"W6959940617","doi":"10.1051/pmed/2024004/pdf","title":"Formation par concordance : la théorie des scripts et son application en enseignement revisitées","year":2024,"lang":"fr","type":"article","venue":"Springer Link (Chiba Institute of Technology)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Centre Hospitalier Universitaire Sainte-Justine","funders":"","keywords":"Scapegoat; Context (archaeology); ESPACE","score_opus":0.014955684878630108,"score_gpt":0.29791856979292514,"score_spread":0.28296288491429505,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6959940617","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.64324313,0.13862048,0.098675884,0.08680364,0.008326436,0.002144021,0.00010660515,0.0015953695,0.020484399],"genre_scores_gemma":[0.9560287,0.017385082,0.024724796,0.00029532835,0.00076891965,0.00015175251,0.000061574196,0.000045753623,0.00053808716],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9975673,0.00006740213,0.0009421756,0.00058832025,0.0003353668,0.00049943733],"domain_scores_gemma":[0.9975288,0.0011324773,0.00025366707,0.0007281086,0.00020197828,0.0001549542],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0011983019,0.0003475722,0.00060986547,0.00040850486,0.00012894825,0.00006454514,0.00033057848,0.000779721,0.000052998243],"category_scores_gemma":[0.011371583,0.0003323115,0.00020949554,0.0008293838,0.0011766512,0.0004072967,0.00023409944,0.0009799271,0.00032424575],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000039560826,0.00023268964,0.003800955,0.0014685061,0.00016271125,0.000093773175,0.00035120503,0.000052785173,0.0053554974,0.32343626,0.0017155461,0.6632905],"study_design_scores_gemma":[0.002126328,0.0010595247,0.019213246,0.061788656,0.0010782572,0.00026359406,0.00030294084,0.004831613,0.06000419,0.086223744,0.76237184,0.0007360529],"about_ca_topic_score_codex":0.00014305543,"about_ca_topic_score_gemma":0.000048647817,"teacher_disagreement_score":0.7606563,"about_ca_system_score_codex":0.0003300257,"about_ca_system_score_gemma":0.00044537976,"threshold_uncertainty_score":0.9999129},"labels":[],"label_agreement":null},{"id":"W6963770815","doi":"10.25384/sage.20939002","title":"sj-docx-1-mdm-10.1177_0272989X221121343 – Supplemental material for The Fallacy of a Single Diagnosis","year":2022,"lang":"en","type":"article","venue":"Sage Journals Data","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Institute for Clinical Evaluative Sciences; Sunnybrook Health Science Centre","funders":"","keywords":"Fallacy; Decision theory; Medical diagnosis; Medical decision making","score_opus":0.09044157771233495,"score_gpt":0.38567614452781257,"score_spread":0.2952345668154776,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6963770815","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6686193,0.03328531,0.0018909293,0.061982296,0.006662362,0.004477552,0.22168766,0.0001619599,0.0012326529],"genre_scores_gemma":[0.98330283,0.003174076,0.0018399659,0.0019599113,0.0018035263,0.0002987116,0.006468956,0.00006436709,0.0010876318],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9978006,0.00012647639,0.00072529097,0.0003481158,0.00063463347,0.0003649169],"domain_scores_gemma":[0.98866785,0.009526008,0.0003635034,0.0011652588,0.00009120124,0.00018615987],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0018163084,0.00016470668,0.00044106296,0.0000767766,0.00036439497,0.00007197026,0.000813553,0.00004760787,0.12696676],"category_scores_gemma":[0.0136526935,0.00011597629,0.00019101056,0.0002065962,0.00010418199,0.00011744112,0.0010186031,0.00027067642,0.000018710167],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001051846,0.0019467594,0.026108623,0.000046675406,0.0004663777,0.00008494644,0.00023954989,0.000020448982,0.004150535,0.000014346371,0.84613025,0.11973961],"study_design_scores_gemma":[0.011099183,0.0035589286,0.028106872,0.0022019001,0.0017916593,0.00059608713,0.002793911,0.00061510346,0.007596082,0.0006452727,0.94054294,0.0004520677],"about_ca_topic_score_codex":0.00017546909,"about_ca_topic_score_gemma":0.000040547457,"teacher_disagreement_score":0.3146836,"about_ca_system_score_codex":0.000076094446,"about_ca_system_score_gemma":0.00016457273,"threshold_uncertainty_score":0.9946557},"labels":[],"label_agreement":null},{"id":"W6977072043","doi":"10.6084/m9.figshare.16994246","title":"Additional file 3 of Improved preventive care clinical decision-making efficiency: leveraging a point-of-care clinical decision support system","year":2021,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Preventive care; Decision support system; Clinical decision support system; Preventive maintenance; Usability; Decision aids","score_opus":0.03598883780067279,"score_gpt":0.377915765154634,"score_spread":0.3419269273539612,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6977072043","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0032064633,0.000544477,0.00007773124,0.000024517867,0.00037230144,0.00048215216,0.99196607,0.00008149766,0.0032447749],"genre_scores_gemma":[0.30115777,0.000008978699,0.015367507,0.00022085918,0.0005089639,0.0002064801,0.6822812,0.00004657082,0.00020170523],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.99571943,0.00023803668,0.0021022987,0.00081666006,0.000762913,0.00036063325],"domain_scores_gemma":[0.7985145,0.19706853,0.00088001695,0.00083124056,0.0023338844,0.00037186316],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0003586602,0.0002616468,0.0011431412,0.000103253624,0.00009514278,0.00002662359,0.00027179316,0.00043519863,0.8828308],"category_scores_gemma":[0.5670967,0.00023088529,0.0010855945,0.000407757,0.000061864346,0.00007295091,0.00040513565,0.000595919,0.0008631111],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00030138157,0.000344304,0.0005905344,0.0004112097,0.000076607255,0.00030848762,0.00016986992,0.000006474716,8.817605e-7,0.000002182241,0.8601895,0.13759854],"study_design_scores_gemma":[0.006290741,0.0029746261,0.13052893,0.59529465,0.00044401424,0.00032819976,0.007088986,0.0024430125,0.00016237721,0.00007929974,0.25363666,0.00072851905],"about_ca_topic_score_codex":0.0000021555363,"about_ca_topic_score_gemma":0.000008752214,"teacher_disagreement_score":0.8819677,"about_ca_system_score_codex":0.000117008756,"about_ca_system_score_gemma":0.002119451,"threshold_uncertainty_score":0.9999148},"labels":[],"label_agreement":null},{"id":"W6977095326","doi":"10.6084/m9.figshare.26620021.v1","title":"Additional file 1 of History and physical exam: a retrospective analysis of a clinical opportunity","year":2024,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"NOSM University","funders":"","keywords":"Data collection; MEDLINE; Retrospective cohort study; Data file","score_opus":0.11839902393281639,"score_gpt":0.38489797813221455,"score_spread":0.2664989541993982,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6977095326","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004083909,0.00015510853,1.3363503e-7,0.00003500717,0.000023959665,0.000063686726,0.98803365,0.000028501607,0.0075760665],"genre_scores_gemma":[0.121354535,0.000004920123,0.0001163308,0.0001214234,0.0001953427,0.000094311676,0.8765401,0.000008667255,0.0015643451],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99910295,0.000036777332,0.0003284805,0.00022956466,0.00022299327,0.000079213896],"domain_scores_gemma":[0.9494766,0.04979321,0.00013315005,0.00019163407,0.00018993567,0.0002154702],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.000069600355,0.00007340834,0.0005333689,0.00011213328,0.000007999607,0.0000028691243,0.000038148883,0.00008565526,0.9884676],"category_scores_gemma":[0.2896285,0.00006093522,0.00036471075,0.0002914476,0.000053980883,0.000024881829,0.000044138196,0.00021619542,0.00015133648],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000020774425,0.00021204719,0.00033323542,0.00006572965,0.00049997086,0.000042609812,0.000058614394,2.3386256e-7,0.0000011369191,0.000015510528,0.9945855,0.004164623],"study_design_scores_gemma":[0.00019255915,0.0003252236,0.61094314,0.011484356,0.0007191129,0.0000051524926,0.000018188102,0.002774554,0.0000039762085,0.00005374383,0.37340954,0.000070473674],"about_ca_topic_score_codex":0.000008255382,"about_ca_topic_score_gemma":0.0000038070084,"teacher_disagreement_score":0.98831624,"about_ca_system_score_codex":0.000043239117,"about_ca_system_score_gemma":0.00049113156,"threshold_uncertainty_score":0.71635526},"labels":[],"label_agreement":null},{"id":"W6977179615","doi":"10.6084/m9.figshare.26715094","title":"Additional file 4 of Use of the International IFOMPT Cervical Framework to inform clinical reasoning in postgraduate level physiotherapy students: a qualitative study using think aloud methodology","year":2024,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Qualitative research; Think aloud protocol; Photo elicitation; Protocol analysis; MEDLINE; Grounded theory; Qualitative analysis","score_opus":0.4044008045684645,"score_gpt":0.5566070379831897,"score_spread":0.15220623341472517,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6977179615","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08983962,0.000025254987,0.000027111828,0.00046402373,0.00018980661,0.0005315661,0.90882427,0.000025314306,0.00007301462],"genre_scores_gemma":[0.33023974,0.000009489288,0.10131183,0.006248286,0.0011129001,0.0005732896,0.55903554,0.00011784835,0.0013510742],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99756914,0.00042852762,0.0008053644,0.00031910243,0.00070594024,0.0001719228],"domain_scores_gemma":[0.8365358,0.1623312,0.00024992062,0.000348132,0.00041116544,0.00012377126],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0005355942,0.00014786435,0.00048992597,0.0001235754,0.000031765423,0.000036134872,0.00031756872,0.00016155666,0.6153114],"category_scores_gemma":[0.39530027,0.0001041701,0.00024906185,0.0005401726,0.000039271206,0.0000855686,0.000297424,0.00055149064,0.00015225993],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003253879,0.0010438313,0.0015891927,0.000057897967,0.00038282239,0.000034662324,0.01788776,0.00004963436,0.0000061866867,0.00008701165,0.97571003,0.0028255666],"study_design_scores_gemma":[0.0018086914,0.001405858,0.8121758,0.1315357,0.00010425448,0.000028912733,0.011668411,0.0028404489,0.000037407317,0.0011440059,0.036920954,0.00032952966],"about_ca_topic_score_codex":0.00010235299,"about_ca_topic_score_gemma":0.000040731084,"teacher_disagreement_score":0.93878907,"about_ca_system_score_codex":0.00007959299,"about_ca_system_score_gemma":0.00033729582,"threshold_uncertainty_score":0.60979337},"labels":[],"label_agreement":null},{"id":"W6977374118","doi":"10.6084/m9.figshare.26715094.v1","title":"Additional file 4 of Use of the International IFOMPT Cervical Framework to inform clinical reasoning in postgraduate level physiotherapy students: a qualitative study using think aloud methodology","year":2024,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Qualitative research; Think aloud protocol; Photo elicitation; Protocol analysis; MEDLINE; Grounded theory; Qualitative analysis","score_opus":0.4044008045684645,"score_gpt":0.5566070379831897,"score_spread":0.15220623341472517,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6977374118","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08983962,0.000025254987,0.000027111828,0.00046402373,0.00018980661,0.0005315661,0.90882427,0.000025314306,0.00007301462],"genre_scores_gemma":[0.33023974,0.000009489288,0.10131183,0.006248286,0.0011129001,0.0005732896,0.55903554,0.00011784835,0.0013510742],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99756914,0.00042852762,0.0008053644,0.00031910243,0.00070594024,0.0001719228],"domain_scores_gemma":[0.8365358,0.1623312,0.00024992062,0.000348132,0.00041116544,0.00012377126],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0005355942,0.00014786435,0.00048992597,0.0001235754,0.000031765423,0.000036134872,0.00031756872,0.00016155666,0.6153114],"category_scores_gemma":[0.39530027,0.0001041701,0.00024906185,0.0005401726,0.000039271206,0.0000855686,0.000297424,0.00055149064,0.00015225993],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003253879,0.0010438313,0.0015891927,0.000057897967,0.00038282239,0.000034662324,0.01788776,0.00004963436,0.0000061866867,0.00008701165,0.97571003,0.0028255666],"study_design_scores_gemma":[0.0018086914,0.001405858,0.8121758,0.1315357,0.00010425448,0.000028912733,0.011668411,0.0028404489,0.000037407317,0.0011440059,0.036920954,0.00032952966],"about_ca_topic_score_codex":0.00010235299,"about_ca_topic_score_gemma":0.000040731084,"teacher_disagreement_score":0.93878907,"about_ca_system_score_codex":0.00007959299,"about_ca_system_score_gemma":0.00033729582,"threshold_uncertainty_score":0.60979337},"labels":[],"label_agreement":null},{"id":"W6977514231","doi":"10.6084/m9.figshare.7116356.v1","title":"Additional file 1: of Avoid reinventing the wheel: implementation of the Ottawa Clinic Assessment Tool (OCAT) in Internal Medicine","year":2018,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"MEDLINE; Data collection; Work (physics)","score_opus":0.04684867185005287,"score_gpt":0.39619903889996144,"score_spread":0.3493503670499086,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6977514231","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0047439286,0.000023878161,9.422125e-7,0.0012736991,0.00009565859,0.00034761283,0.980877,0.000011241444,0.01262605],"genre_scores_gemma":[0.4358187,0.0000027444325,0.00037748058,0.0014355169,0.0007715632,0.0004042535,0.5595649,0.000017793316,0.0016070775],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9986897,0.000086172986,0.00059242576,0.00014498168,0.00035524365,0.00013146963],"domain_scores_gemma":[0.9750744,0.023894187,0.0004509456,0.00028985276,0.00024990016,0.000040693412],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00025501393,0.00007987513,0.00020484572,0.000037650592,0.000042909836,0.0000043350465,0.00016702866,0.000050628714,0.98486793],"category_scores_gemma":[0.10993588,0.00004402954,0.00010442418,0.00018845036,0.000070387534,0.000028527184,0.00009847082,0.00021251138,0.0002427899],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000011374564,0.000049192855,0.0021904234,0.000034169894,0.000020144602,0.0000017988858,0.00009124168,2.6275438e-7,0.000011415136,0.000010970249,0.9928651,0.0047139027],"study_design_scores_gemma":[0.0011264905,0.00043388674,0.74115866,0.038882952,0.000032221098,0.00001056499,0.0004889729,0.00015422385,0.00031384797,0.00011594907,0.21722057,0.00006164402],"about_ca_topic_score_codex":0.000095137424,"about_ca_topic_score_gemma":0.00017835895,"teacher_disagreement_score":0.98462516,"about_ca_system_score_codex":0.00003749198,"about_ca_system_score_gemma":0.00026556023,"threshold_uncertainty_score":0.8975615},"labels":[],"label_agreement":null},{"id":"W6977728620","doi":"10.6084/m9.figshare.8062553.v1","title":"Additional file 1: of Just-in-time faculty development: a mobile application helps clinical teachers verify and describe clinical reasoning difficulties","year":2019,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Université de Sherbrooke","keywords":"Interview; Mobile device; Data collection; Case-based reasoning; Semi-structured interview; Qualitative research","score_opus":0.057750692492568284,"score_gpt":0.3608517586754721,"score_spread":0.3031010661829038,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6977728620","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.103669114,0.0000718746,4.293392e-7,0.000030404222,0.00004040754,0.00063258456,0.8940039,0.00006554903,0.001485703],"genre_scores_gemma":[0.029827157,0.000008592619,0.0034094807,0.0002566944,0.00017513525,0.0009997946,0.9622286,0.000026105292,0.003068449],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9980167,0.00008219846,0.000870107,0.00047287924,0.0003210035,0.00023708891],"domain_scores_gemma":[0.97201097,0.027003482,0.00030531944,0.00028139472,0.00014375325,0.000255088],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00026023053,0.00016856035,0.00056034076,0.000052186933,0.000037192116,0.000017787339,0.00012643539,0.00033445607,0.82363904],"category_scores_gemma":[0.10720556,0.0001467159,0.00015991474,0.00014824947,0.000048935217,0.00005942575,0.00012653045,0.0004540998,0.004032907],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007488336,0.00038106993,0.012217306,0.000049552178,0.00003659845,0.000007054419,0.000110208704,0.0000028016632,0.0000022426957,8.8460723e-7,0.9639972,0.023120202],"study_design_scores_gemma":[0.0012592482,0.00023325342,0.5320603,0.012097142,0.000018578872,0.00001136244,0.00012795447,0.00036034317,0.000012395169,0.0000023060752,0.45366853,0.00014858542],"about_ca_topic_score_codex":0.0000053507947,"about_ca_topic_score_gemma":0.0000042413153,"teacher_disagreement_score":0.8196061,"about_ca_system_score_codex":0.00004318705,"about_ca_system_score_gemma":0.00029959812,"threshold_uncertainty_score":0.99674255},"labels":[],"label_agreement":null},{"id":"W6983183436","doi":"","title":"L'incertitude, le coping et l'espoir chez la personne blessee medullaire en phase de readaptation post-aigue","year":2001,"lang":"fr","type":"dissertation","venue":"Papyrus : Institutional Repository (Université de Montréal)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Cognition; Coping (psychology); Psychosocial; Personality; Social environment","score_opus":0.008575472463239534,"score_gpt":0.2516773334040606,"score_spread":0.2431018609408211,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6983183436","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9182373,0.025763955,0.0014134166,0.0046508415,0.0020133173,0.00077952107,0.00017546675,0.00023450097,0.04673165],"genre_scores_gemma":[0.9293016,0.0054687154,0.0014280077,0.00080054795,0.0008813531,0.00006385454,0.003190267,0.00009170489,0.058773942],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.99562865,0.0007009404,0.0008121868,0.0009986657,0.0009894007,0.0008701403],"domain_scores_gemma":[0.9917401,0.0054422026,0.00065901165,0.00052428263,0.0006870177,0.00094736955],"candidate_categories":["metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.0010314349,0.0007223353,0.0009072802,0.00039902862,0.0040394403,0.00010730564,0.00042935924,0.0010945977,0.00017739432],"category_scores_gemma":[0.008299727,0.0008386943,0.00058533554,0.0004940069,0.00048303304,0.00040861513,0.0001305489,0.0012052705,0.00014798298],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.021592418,0.011625154,0.06826271,0.002290266,0.0045604887,0.07225844,0.37956923,0.022512982,0.12453388,0.16401453,0.007424555,0.12135533],"study_design_scores_gemma":[0.03756339,0.0032575738,0.40397143,0.04070604,0.007272119,0.015308876,0.3211601,0.022344124,0.028332328,0.0011142826,0.114680104,0.0042896187],"about_ca_topic_score_codex":0.08267144,"about_ca_topic_score_gemma":0.014574285,"teacher_disagreement_score":0.3357087,"about_ca_system_score_codex":0.0074912435,"about_ca_system_score_gemma":0.010948909,"threshold_uncertainty_score":0.9994064},"labels":[],"label_agreement":null},{"id":"W7000331520","doi":"","title":"Étude rétrospective comparant la qualité de trois types de référentiels utilisés pour préparer les Épreuves Classantes Nationales 2009, 2010 et 2011","year":2012,"lang":"fr","type":"other","venue":"OpenGrey (Institut de l'Information Scientifique et Technique)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Context (archaeology); Nova scotia; Intrusion","score_opus":0.041719275771818054,"score_gpt":0.3420806534312454,"score_spread":0.30036137765942733,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7000331520","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.030517817,0.00539061,0.35540062,0.008204921,0.0036985918,0.012702832,0.001955706,0.0010732319,0.5810557],"genre_scores_gemma":[0.5524167,0.0068986155,0.11501994,0.00617384,0.0015501434,0.004182903,0.0030127345,0.00043049434,0.31031463],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9943964,0.0011150736,0.0015477266,0.00066068576,0.0010518029,0.0012283376],"domain_scores_gemma":[0.9890157,0.0069744657,0.0011847913,0.00091853,0.0010992683,0.0008072476],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.009542185,0.00080967543,0.0011027699,0.00084442453,0.00040319195,0.00066435366,0.0008575337,0.00132792,0.010160755],"category_scores_gemma":[0.028614162,0.00074699597,0.00042967906,0.0005682784,0.0009189087,0.001599903,0.00030366672,0.0016969538,0.0025322426],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00045183263,0.0023352723,0.07751701,0.0017066563,0.00077546295,0.0000924426,0.0143486075,0.0008719706,0.0029345704,0.2672195,0.5498221,0.081924595],"study_design_scores_gemma":[0.0025847077,0.00016252843,0.18757546,0.018153971,0.00063014776,0.00038200512,0.0008734427,0.0031205025,0.014509661,0.003373825,0.76718235,0.0014514326],"about_ca_topic_score_codex":0.0025635466,"about_ca_topic_score_gemma":0.001684953,"teacher_disagreement_score":0.52189887,"about_ca_system_score_codex":0.0010175064,"about_ca_system_score_gemma":0.0036338598,"threshold_uncertainty_score":0.9999686},"labels":[],"label_agreement":null},{"id":"W7007221034","doi":"","title":"معركة سجن غويران ISLAMIC STATE PRISON BREAKING WAR IN GHUWAIRAN ISLAMIC STATE VIDEO اصدار الدولةالإسلامية","year":2020,"lang":"ar","type":"other","venue":"Bulletin of Miscellaneous Information (Royal Gardens Kew)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Islam; State (computer science); Brother; Supporter; Hinduism","score_opus":0.007696618341189313,"score_gpt":0.22567879507908603,"score_spread":0.21798217673789672,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7007221034","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.01808294,0.0016589386,0.000003228656,0.0015318481,0.0011372573,0.0019264237,0.00062025595,0.0003118239,0.9747273],"genre_scores_gemma":[0.08986385,0.0058564637,0.0009038125,0.0020576662,0.00052537746,0.000049366125,0.000734588,0.0003590185,0.89964986],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9918204,0.00047632345,0.0036985066,0.0010496733,0.0015117709,0.00144333],"domain_scores_gemma":[0.990096,0.0056808116,0.0022559965,0.0011175575,0.00037705287,0.00047258663],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["metaepi_narrow","insufficient_payload"],"category_scores_codex":[0.0012838728,0.0013336895,0.0023662744,0.0003560901,0.00016299788,0.00012835617,0.0008289012,0.0010409391,0.2431573],"category_scores_gemma":[0.012846013,0.0013308742,0.0006968933,0.00008247152,0.0005015906,0.0000015119812,0.00037707234,0.0020278837,0.078526184],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011628163,0.00036322488,0.00026037518,0.0016977241,0.00034686108,0.0006173727,0.0037453927,0.0021720582,0.0000051339457,0.000022781353,0.9706706,0.018935677],"study_design_scores_gemma":[0.005705349,0.0014460252,0.00075651915,0.012499114,0.00041546387,0.00019596229,0.0006362694,0.0009312993,0.00005528554,0.00018169267,0.9759409,0.0012361387],"about_ca_topic_score_codex":0.020242855,"about_ca_topic_score_gemma":0.005463349,"teacher_disagreement_score":0.16463111,"about_ca_system_score_codex":0.0004069994,"about_ca_system_score_gemma":0.00038653813,"threshold_uncertainty_score":0.9999414},"labels":[],"label_agreement":null},{"id":"W7009651231","doi":"","title":"Evaluación de las habilidades comunicacionales de los internos rotativos de la carrera de medicina mediante la aplicación de la Guía Calgary Cambridge. Cuenca 2018","year":2019,"lang":"es","type":"dissertation","venue":"Repositorio Institucional (Universidad de Cuenca)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Initial training; Limiting; Archivist","score_opus":0.011384469046403846,"score_gpt":0.34102833315871794,"score_spread":0.3296438641123141,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7009651231","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95737,0.0045348234,0.0031767623,0.001455051,0.0019200521,0.0012388267,0.0003142267,0.00028778205,0.02970251],"genre_scores_gemma":[0.9746308,0.0039624553,0.008558866,0.00076196174,0.0024197667,0.0001789096,0.0012794051,0.0002458044,0.007962065],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.98986256,0.0032194974,0.0014101529,0.0016398722,0.0019636247,0.0019043152],"domain_scores_gemma":[0.9523448,0.04216909,0.0010926917,0.001369534,0.000844956,0.002178977],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":["metaepi_narrow","research_integrity"],"category_scores_codex":[0.0049507264,0.0013211799,0.0019199974,0.00084096077,0.000753161,0.0003603921,0.0015461837,0.0030862573,0.00020489274],"category_scores_gemma":[0.03957167,0.0014206569,0.0010919762,0.00077502383,0.0016731317,0.00039609848,0.00032406521,0.00391008,0.00017077425],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.009297348,0.00547287,0.6379835,0.004480822,0.00530007,0.03231258,0.03478709,0.002837672,0.112497106,0.09380671,0.04216645,0.019057812],"study_design_scores_gemma":[0.016346168,0.002293806,0.7705074,0.06260723,0.008183464,0.018028887,0.019824188,0.023834469,0.012682406,0.0024798822,0.059194878,0.004017241],"about_ca_topic_score_codex":0.00665899,"about_ca_topic_score_gemma":0.00022194692,"teacher_disagreement_score":0.13252391,"about_ca_system_score_codex":0.0056164837,"about_ca_system_score_gemma":0.015698643,"threshold_uncertainty_score":0.9999558},"labels":[],"label_agreement":null},{"id":"W7014276431","doi":"","title":"Photographs exhibited in Britain 1839-1865","year":2008,"lang":"en","type":"other","venue":"DMU Open Research Archive (De Montfort University)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Exhibition; Ephemeral key; Painting; Photography; The arts; Historical record; Fine art; Period (music)","score_opus":0.07721911874540226,"score_gpt":0.3733222805889748,"score_spread":0.29610316184357255,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7014276431","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0053741583,0.00034453804,0.00011017947,0.00031867434,0.000061774415,0.0023925926,0.00024661553,0.00011645897,0.991035],"genre_scores_gemma":[0.006542045,0.007697017,0.0034579732,0.00024879992,0.00019718612,0.000023166991,0.0006823209,0.00034868022,0.98080283],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99602515,0.0005518275,0.00031523494,0.00097127975,0.0008499959,0.0012865349],"domain_scores_gemma":[0.99337494,0.004384587,0.00013101961,0.0009799371,0.00013667661,0.0009928498],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011768439,0.00035212032,0.0009075209,0.0032893906,0.00018909501,0.00007400759,0.0012408434,0.0004232299,0.0018592578],"category_scores_gemma":[0.007818481,0.0003681285,0.00028522208,0.0019222613,0.00081771094,0.00008181888,0.0012255772,0.0018843773,0.00026498013],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013333857,0.0010755614,0.056220725,0.00009130844,0.0002591605,0.020091638,0.00036397262,0.000002692546,0.000082394494,0.0016612416,0.91426075,0.0045571458],"study_design_scores_gemma":[0.006096456,0.00072919385,0.021707775,0.0064190878,0.00006743786,0.00009347853,0.00070597837,0.00016494127,0.000030728108,0.00061146234,0.96289223,0.00048121862],"about_ca_topic_score_codex":0.050519586,"about_ca_topic_score_gemma":0.013228673,"teacher_disagreement_score":0.04863147,"about_ca_system_score_codex":0.00037284632,"about_ca_system_score_gemma":0.0015546416,"threshold_uncertainty_score":0.9998771},"labels":[],"label_agreement":null},{"id":"W7086631865","doi":"","title":"Associations between nicotine vaping uptake and cigarette smoking cessation vary by smokers' plans to quit: longitudinal findings from the International Tobacco Control Four Country Smoking and Vaping Surveys","year":2023,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Smoking cessation; Nicotine; Tobacco control; Odds ratio; Quit smoking; Odds; Cigarette smoking; Population; Pack-year","score_opus":0.06892183189266052,"score_gpt":0.32104908220691475,"score_spread":0.2521272503142542,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7086631865","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8010967,0.00021749293,0.00007080964,0.0047691916,0.00020886016,0.0003220131,0.19301786,0.00013729291,0.0001598068],"genre_scores_gemma":[0.89835966,0.000024151617,0.00001904115,0.0015803825,0.0005730776,0.00004939534,0.09930889,0.000027849503,0.000057566736],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9981492,0.0001018177,0.00043323633,0.00043895518,0.0005476122,0.00032913926],"domain_scores_gemma":[0.9782761,0.02098833,0.00018828368,0.00019212674,0.00016523669,0.00018992614],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0009908593,0.00019275742,0.0003574945,0.000091717295,0.00030325548,0.00022801016,0.00017544876,0.00018592624,0.002096782],"category_scores_gemma":[0.057768267,0.00016768443,0.0000643776,0.00030532043,0.000016035528,0.00013737856,0.00017671887,0.00040096042,0.0000339789],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000012122688,0.000013384709,0.8762845,0.00000897781,0.00029451985,0.000013721467,0.00033059908,0.000017684913,0.000023972747,0.000006138671,0.11844605,0.004548328],"study_design_scores_gemma":[0.0015157895,0.000045396395,0.99015546,0.005172325,0.00019040014,0.000004321595,0.00013544499,0.0003761,0.000036244455,0.00004905735,0.0021428934,0.00017654803],"about_ca_topic_score_codex":0.00037699053,"about_ca_topic_score_gemma":0.00014783989,"teacher_disagreement_score":0.11630315,"about_ca_system_score_codex":0.00012133248,"about_ca_system_score_gemma":0.00007329326,"threshold_uncertainty_score":0.9988154},"labels":[],"label_agreement":null},{"id":"W7095854664","doi":"","title":"Printed in U.S.A. Chen et al. Respond to “Obesity and Asthma ” Chen et al. Chen et al. Respond to “Obesity and Asthma”","year":2001,"lang":"en","type":"article","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Asthma; Chen; Obesity; Disease; Relation (database); Respiratory disease","score_opus":0.024403718914892247,"score_gpt":0.35148212788818695,"score_spread":0.3270784089732947,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7095854664","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7808865,0.00027154273,0.0003720165,0.20968215,0.00013637234,0.0011318295,0.000043838936,0.00016830351,0.0073074717],"genre_scores_gemma":[0.65839845,0.0011858249,0.007074299,0.31857765,0.00007658058,0.00008248984,0.000103864724,0.00009703537,0.014403799],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.995258,0.00058215257,0.00094129995,0.0014732236,0.00082770124,0.0009176546],"domain_scores_gemma":[0.98941165,0.0077019273,0.00016398603,0.0010155267,0.0002724571,0.0014344614],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.005373604,0.0005795538,0.0011616737,0.00048415485,0.00011751383,0.00015049419,0.00029321318,0.00030597104,0.0006667021],"category_scores_gemma":[0.045790344,0.0005009305,0.0001472477,0.0008784047,0.00016745491,0.00022200203,0.00070729986,0.0009959814,0.00025098296],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0074460367,0.0021126103,0.4850756,0.000045423018,0.00016615435,0.0008724435,0.0027549386,0.000023910081,0.004379487,0.0009668293,0.48228788,0.013868696],"study_design_scores_gemma":[0.003799952,0.0010995303,0.6367318,0.0010543158,0.000032119122,0.00013888575,0.000255543,0.00009886102,0.0007652677,0.00016935603,0.35542426,0.00043011163],"about_ca_topic_score_codex":0.0016807725,"about_ca_topic_score_gemma":0.0023530275,"teacher_disagreement_score":0.15165621,"about_ca_system_score_codex":0.00017016537,"about_ca_system_score_gemma":0.000341165,"threshold_uncertainty_score":0.99974424},"labels":[],"label_agreement":null},{"id":"W7096113766","doi":"","title":"Development and application of clinical prediction rules to improve decision making in physical therapist practice","year":2006,"lang":"en","type":"article","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Clinical prediction rule; Psychological intervention; Clinical decision making; Deep vein; Clinical Practice; Ankle; Carpal tunnel syndrome; Neck pain","score_opus":0.02324230196465388,"score_gpt":0.412139848401544,"score_spread":0.3888975464368901,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7096113766","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9617276,0.000040287137,0.034245085,0.00048653074,0.00006065092,0.00032749583,0.0000022513038,0.000024608013,0.0030854982],"genre_scores_gemma":[0.90900624,0.000012730486,0.09040768,0.00031749066,0.00017286009,0.000026990576,0.000010048017,0.000006654927,0.000039307317],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99885136,0.000029530142,0.00056710426,0.00024604233,0.0002095307,0.00009642703],"domain_scores_gemma":[0.9911325,0.008451862,0.000114623275,0.0001464052,0.00008841824,0.00006620937],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0007393713,0.00006983549,0.00024545655,0.000048105652,0.00001859458,0.000007660294,0.00002910118,0.00007240751,0.000004353344],"category_scores_gemma":[0.013361139,0.000052391508,0.000035145764,0.0001051291,0.00003427794,0.000040135823,0.000031292704,0.00011338826,0.000013399976],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00061287853,0.0010724515,0.10332524,0.0000114832455,0.000012346321,0.000003588298,0.00013950655,0.00001792548,0.00063138816,0.0016696871,0.00019489683,0.8923086],"study_design_scores_gemma":[0.0015446962,0.000295596,0.98804635,0.00054195645,0.000033412307,0.000006916397,0.00010382162,0.004310958,0.00081570994,0.0009095995,0.0033234276,0.00006757564],"about_ca_topic_score_codex":0.00010267788,"about_ca_topic_score_gemma":0.000034436696,"teacher_disagreement_score":0.89224106,"about_ca_system_score_codex":0.000026286805,"about_ca_system_score_gemma":0.00006426622,"threshold_uncertainty_score":0.99494976},"labels":[],"label_agreement":null},{"id":"W7104292769","doi":"10.71781/6928","title":"L'incertitude, le coping et l'espoir chez la personne blessee medullaire en phase de readaptation post-aigue","year":2001,"lang":"fr","type":"dissertation","venue":"Papyrus : Institutional Repository (Université de Montréal)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Cognition; Coping (psychology); Psychosocial; Personality; Social environment","score_opus":0.008575472463239534,"score_gpt":0.2516773334040606,"score_spread":0.2431018609408211,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7104292769","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9182373,0.025763955,0.0014134166,0.0046508415,0.0020133173,0.00077952107,0.00017546675,0.00023450097,0.04673165],"genre_scores_gemma":[0.9293016,0.0054687154,0.0014280077,0.00080054795,0.0008813531,0.00006385454,0.003190267,0.00009170489,0.058773942],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.99562865,0.0007009404,0.0008121868,0.0009986657,0.0009894007,0.0008701403],"domain_scores_gemma":[0.9917401,0.0054422026,0.00065901165,0.00052428263,0.0006870177,0.00094736955],"candidate_categories":["metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.0010314349,0.0007223353,0.0009072802,0.00039902862,0.0040394403,0.00010730564,0.00042935924,0.0010945977,0.00017739432],"category_scores_gemma":[0.008299727,0.0008386943,0.00058533554,0.0004940069,0.00048303304,0.00040861513,0.0001305489,0.0012052705,0.00014798298],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.021592418,0.011625154,0.06826271,0.002290266,0.0045604887,0.07225844,0.37956923,0.022512982,0.12453388,0.16401453,0.007424555,0.12135533],"study_design_scores_gemma":[0.03756339,0.0032575738,0.40397143,0.04070604,0.007272119,0.015308876,0.3211601,0.022344124,0.028332328,0.0011142826,0.114680104,0.0042896187],"about_ca_topic_score_codex":0.08267144,"about_ca_topic_score_gemma":0.014574285,"teacher_disagreement_score":0.3357087,"about_ca_system_score_codex":0.0074912435,"about_ca_system_score_gemma":0.010948909,"threshold_uncertainty_score":0.9994064},"labels":[],"label_agreement":null},{"id":"W7105099166","doi":"10.71781/30451","title":"L'intuition au cœur du raisonnement clinique des ergothérapeutes : une investigation constructiviste multimodale","year":2024,"lang":"fr","type":"dissertation","venue":"Open MIND","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Context (archaeology); Validation test; Self evaluation; Agrégation","score_opus":0.0518377172814414,"score_gpt":0.3750431464672184,"score_spread":0.323205429185777,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7105099166","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9648238,0.0015367771,0.0005536832,0.0018248516,0.004796836,0.0021197658,0.00022478109,0.000016893933,0.024102643],"genre_scores_gemma":[0.95823926,0.0009428223,0.016370673,0.00022009977,0.0017127275,0.00029749118,0.0060561313,0.00011676549,0.016044052],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99589735,0.0003015129,0.001451564,0.0011911859,0.00050766737,0.00065069227],"domain_scores_gemma":[0.9894815,0.0083876485,0.0005842536,0.00044210118,0.00047217534,0.00063230447],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0014654462,0.00063857663,0.001046173,0.00021372511,0.0003938906,0.00043659974,0.00036704366,0.00077315595,0.0041245203],"category_scores_gemma":[0.02057981,0.00061992434,0.0003438659,0.0005834965,0.0008190373,0.0002841225,0.00017166538,0.0008571072,0.004242074],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006568305,0.0008673466,0.0606813,0.00084054767,0.00072573987,0.00030207422,0.032142073,0.00006100651,0.002820924,0.0013670601,0.0014630817,0.898072],"study_design_scores_gemma":[0.020132318,0.00522271,0.41979712,0.26493722,0.012664376,0.0005028617,0.050456073,0.033820543,0.10364625,0.03858334,0.045724764,0.004512431],"about_ca_topic_score_codex":0.007977187,"about_ca_topic_score_gemma":0.008204711,"teacher_disagreement_score":0.8935596,"about_ca_system_score_codex":0.000606137,"about_ca_system_score_gemma":0.0023093994,"threshold_uncertainty_score":0.9996252},"labels":[],"label_agreement":null},{"id":"W7111979066","doi":"","title":"GEMINI MedED: Leveraging ‘Big Data’ to Understand Clinical Practice Variation in Resident Physicians","year":2024,"lang":"en","type":"article","venue":"Digital Access to Scholarship at Harvard (DASH) (Harvard University)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Clinical Practice; Variation (astronomy); Medical record; MEDLINE; Descriptive statistics; Cohort; Quality (philosophy); Health care","score_opus":0.16732546279572225,"score_gpt":0.39049868827549633,"score_spread":0.22317322547977408,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7111979066","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94288343,0.00005788794,0.00566041,0.008431454,0.0019843406,0.0013449948,0.0005112621,0.00048323534,0.03864297],"genre_scores_gemma":[0.9867234,0.00018972316,0.0009589076,0.0049723615,0.00077288976,0.0000038678863,0.000410518,0.0000803548,0.0058879755],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.9955128,0.0002651703,0.0008397142,0.0016421204,0.0010191873,0.0007210484],"domain_scores_gemma":[0.9870619,0.010131805,0.00018751486,0.0013400338,0.00035473288,0.0009239679],"candidate_categories":["metaresearch","metaepi_narrow","scholarly_communication","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0018067247,0.0004140313,0.00073451805,0.0008966663,0.00020097592,0.0016959833,0.0013019348,0.00030948024,0.00024176495],"category_scores_gemma":[0.051796608,0.00044615477,0.0002535006,0.0023471967,0.00010809275,0.0047255927,0.0020948246,0.001069866,0.0040853294],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.011616504,0.0044065723,0.47896355,0.0006661117,0.0023843586,0.024931438,0.0073552025,0.0009904208,0.000732514,0.009078947,0.14223458,0.3166398],"study_design_scores_gemma":[0.0020873381,0.0004427286,0.32914147,0.0038253602,0.00050926884,0.000067862224,0.0010310995,0.00021734329,0.000091523696,0.00036081372,0.66156155,0.0006636494],"about_ca_topic_score_codex":0.00016879238,"about_ca_topic_score_gemma":0.00026498383,"teacher_disagreement_score":0.5193269,"about_ca_system_score_codex":0.0011406058,"about_ca_system_score_gemma":0.0006207571,"threshold_uncertainty_score":0.999799},"labels":[],"label_agreement":null},{"id":"W7115889060","doi":"10.2196/85614","title":"Evaluation of Few-Shot AI-Generated Feedback on Case Reports in Physical Therapy Education: Mixed Methods Study","year":2025,"lang":"en","type":"article","venue":"JMIR Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Reflection (computer programming); Visual feedback; Physical activity; Action (physics); Simulated patient","score_opus":0.06719029324216029,"score_gpt":0.5429899308784939,"score_spread":0.4757996376363336,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7115889060","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9853925,0.00045128117,0.000037475504,0.0068730097,0.002183539,0.0018801155,7.544649e-7,0.000032386295,0.0031489166],"genre_scores_gemma":[0.99489653,0.000029701672,0.00042381583,0.0026266582,0.00047595517,0.0010202769,0.00008782289,0.000015096454,0.0004241579],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99610823,0.0011872371,0.0007956666,0.00048117538,0.0012488768,0.00017881594],"domain_scores_gemma":[0.9946938,0.0030886405,0.00022717497,0.0006175218,0.0010891795,0.00028365265],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0052514747,0.00017961071,0.00047977068,0.0002806793,0.00004982831,0.000016206795,0.000068317895,0.00019580258,0.0004580533],"category_scores_gemma":[0.07677448,0.0001426849,0.00010448602,0.00094520394,0.00009458056,0.000045675202,0.00002502641,0.0004291028,0.000009671916],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018882238,0.02959865,0.015988115,0.00003077183,0.00008633181,0.00008682895,0.0017377284,0.000013515764,0.00022495395,0.00021819868,0.016505407,0.9353207],"study_design_scores_gemma":[0.008591242,0.0021786282,0.9542134,0.0048739347,0.0008182074,0.00085899944,0.0098506715,0.0060460707,0.0041466528,0.00575797,0.0022612237,0.00040303648],"about_ca_topic_score_codex":0.00040449118,"about_ca_topic_score_gemma":0.000049010498,"teacher_disagreement_score":0.93822527,"about_ca_system_score_codex":0.00028165933,"about_ca_system_score_gemma":0.010763862,"threshold_uncertainty_score":0.9948442},"labels":[],"label_agreement":null},{"id":"W7116761629","doi":"10.1007/s10459-025-10490-z","title":"The first step in visual diagnosis: a study of novices developing the ability to distinguish normal from abnormal cases","year":2025,"lang":"en","type":"article","venue":"Advances in Health Sciences Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Certainty; Abnormality; Curriculum; Medical diagnosis; Medical imaging; Diagnostic accuracy; Contrast (vision); Diagnostic test","score_opus":0.026934953740659233,"score_gpt":0.4518050355302187,"score_spread":0.4248700817895595,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7116761629","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9799761,0.00470091,0.000031202966,0.012980426,0.0010484284,0.00097248267,0.0000023792584,0.000009766,0.00027829508],"genre_scores_gemma":[0.99585885,0.0009004173,0.00075439725,0.0019884363,0.00009286141,0.0003913572,0.0000026828493,0.0000024579138,0.000008540036],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99816227,0.00014703604,0.00078376214,0.00033062004,0.00030097686,0.0002753485],"domain_scores_gemma":[0.96982324,0.02939047,0.0002819649,0.0002778503,0.00014026355,0.0000862265],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002325794,0.00009167837,0.00022747557,0.00010766114,0.0004020507,0.000035658144,0.0002659035,0.000024981213,0.0000040581294],"category_scores_gemma":[0.040352467,0.000054241835,0.00002110198,0.0013096036,0.0002421887,0.00013277066,0.000076845725,0.00014974992,0.0000013647833],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000052030646,0.000634483,0.8948946,0.00005636455,0.0000016350963,6.564125e-7,0.0021930372,0.00040528396,9.1277805e-8,0.00020756328,0.00006647251,0.10148781],"study_design_scores_gemma":[0.00033281356,0.0005471984,0.9757468,0.0023461918,0.0000062173945,9.791407e-7,0.017636474,0.00021366826,0.000008755317,0.00033859038,0.0027693156,0.000052974425],"about_ca_topic_score_codex":0.025736924,"about_ca_topic_score_gemma":0.0911265,"teacher_disagreement_score":0.101434834,"about_ca_system_score_codex":0.00017927625,"about_ca_system_score_gemma":0.0015513703,"threshold_uncertainty_score":0.9807508},"labels":[],"label_agreement":null},{"id":"W7116926332","doi":"10.1051/pmed/2025003","title":"Étude exploratoire des pratiques de supervision du raisonnement clinique des orthophonistes superviseur·es","year":2024,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières","funders":"","keywords":"General interest; Patient rights","score_opus":0.06546760748959715,"score_gpt":0.36829105021024006,"score_spread":0.3028234427206429,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7116926332","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.69163793,0.2833961,0.004848214,0.014574103,0.0024107967,0.00060070655,0.00010918204,0.0006677021,0.001755307],"genre_scores_gemma":[0.7627219,0.20772913,0.012673819,0.0026704087,0.0038508982,0.00025718883,0.00017080402,0.00020225313,0.009723591],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9943943,0.0007279205,0.0013864917,0.001152147,0.0007891168,0.0015500026],"domain_scores_gemma":[0.9763431,0.020958737,0.000107134205,0.00078082597,0.00036268975,0.0014475393],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.002721097,0.00070332736,0.0010044748,0.0002009855,0.0005620368,0.00052513747,0.000433032,0.0007255943,0.004194406],"category_scores_gemma":[0.052779797,0.0006063065,0.0006566212,0.00072620815,0.0021019408,0.00045622812,0.00038683962,0.0011140864,0.0012568433],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002507464,0.0027529693,0.160745,0.0027045445,0.00045165746,0.0036006807,0.015925735,0.00005041265,0.001281339,0.0044726245,0.030495813,0.77726847],"study_design_scores_gemma":[0.0071310536,0.008756491,0.16190562,0.17957889,0.003372316,0.0012463413,0.024582714,0.023449725,0.026002381,0.029418513,0.5312658,0.0032901948],"about_ca_topic_score_codex":0.0007859131,"about_ca_topic_score_gemma":0.00049459643,"teacher_disagreement_score":0.7739783,"about_ca_system_score_codex":0.00051230995,"about_ca_system_score_gemma":0.0018282023,"threshold_uncertainty_score":0.99963886},"labels":[],"label_agreement":null},{"id":"W7117156966","doi":"10.3138/jvme-2025-0065","title":"Development and Evaluation of the Script Concordance Test to Assess Veterinary Students on Equine Colic","year":2025,"lang":"en","type":"article","venue":"Journal of Veterinary Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Concordance; Test (biology); Focus group; MEDLINE; Alternative medicine; Multiple choice","score_opus":0.1820405676654763,"score_gpt":0.5149897520788489,"score_spread":0.3329491844133726,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7117156966","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9921295,0.00036143194,0.00004409655,0.005125966,0.0015614519,0.00030300423,8.6376906e-7,0.0000036757808,0.00047001633],"genre_scores_gemma":[0.9953568,0.00006648622,0.0010516141,0.0031208205,0.000172178,0.000022620845,0.0000023260498,0.000005903443,0.00020126528],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99765944,0.00015152487,0.0006267249,0.00012727916,0.0013280464,0.00010695072],"domain_scores_gemma":[0.99659383,0.002066693,0.00029262653,0.00018949856,0.00058831414,0.0002690633],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.002325202,0.0000928734,0.00027323485,0.0001217804,0.000051190964,0.000013563316,0.0001904099,0.00007484315,0.00006871258],"category_scores_gemma":[0.06253933,0.00006160483,0.000054133492,0.00023635174,0.000076250246,0.00003919286,0.00010205676,0.00024020637,0.0000028259435],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017698592,0.0073148343,0.14967355,0.00029831985,0.00020587142,0.000059003436,0.001611305,0.000016152357,0.011910531,0.00017007047,0.013939267,0.81303126],"study_design_scores_gemma":[0.0024147069,0.0046063224,0.9634457,0.012962921,0.00018731067,0.0004192306,0.0004329089,0.000109080385,0.00081821,0.00007116829,0.014460089,0.000072316696],"about_ca_topic_score_codex":0.0000059587655,"about_ca_topic_score_gemma":0.000001295113,"teacher_disagreement_score":0.8137722,"about_ca_system_score_codex":0.00020399255,"about_ca_system_score_gemma":0.0032459102,"threshold_uncertainty_score":0.9453573},"labels":[],"label_agreement":null},{"id":"W7118004973","doi":"10.1007/978-3-032-05288-9_2","title":"Clinical Decision Analysis","year":2025,"lang":"en","type":"book-chapter","venue":"Difficult decisions in surgery: an evidence-based approach","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Joseph’s Healthcare Hamilton; Impact; McMaster University; St. Joseph's Hospital; University of Manitoba","funders":"","keywords":"Decision analysis; Process (computing); Clinical decision making; Balance (ability); Decision process; Decision theory; Evidential reasoning approach; Decision support system","score_opus":0.1327620949525191,"score_gpt":0.40582434788565874,"score_spread":0.2730622529331397,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7118004973","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2908216,0.043935984,0.255455,0.005565988,0.009675721,0.012294504,0.0017957876,0.0024905573,0.37796485],"genre_scores_gemma":[0.6071201,0.045564637,0.09001796,0.00898069,0.0025532623,0.00053303863,0.010813668,0.0005411892,0.23387545],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.98948497,0.0004506208,0.004258425,0.0027965752,0.0021799312,0.0008294746],"domain_scores_gemma":[0.78929377,0.20517975,0.0007589224,0.0028869668,0.00090092636,0.0009796529],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.006963153,0.0010777641,0.004746265,0.0032621855,0.00019739327,0.00016876083,0.00077773613,0.002189113,0.0007381136],"category_scores_gemma":[0.22590101,0.000849113,0.003737377,0.0023515818,0.0005733431,0.00019280474,0.00026706234,0.002292034,0.00022192705],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0023448057,0.0032390784,0.1216878,0.00006535285,0.0011858428,0.000887863,0.000028922457,0.0014968365,8.299917e-7,0.002439035,0.06468598,0.80193764],"study_design_scores_gemma":[0.010464706,0.002077043,0.49105838,0.20962422,0.03155911,0.00012877051,0.00024773815,0.09340054,0.000019472474,0.012259791,0.14249086,0.0066693714],"about_ca_topic_score_codex":0.000085306965,"about_ca_topic_score_gemma":0.00007153908,"teacher_disagreement_score":0.7952683,"about_ca_system_score_codex":0.00038981243,"about_ca_system_score_gemma":0.0024575728,"threshold_uncertainty_score":0.99939597},"labels":[],"label_agreement":null},{"id":"W7118043854","doi":"10.1007/978-3-032-10050-4_5","title":"Clinical Decision-Making for Pharmacists","year":2025,"lang":"en","type":"book-chapter","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Clinical pharmacy; Scope (computer science); Scope of practice; Health care; Health professionals; Cognition; Pharmacy practice","score_opus":0.06325596741058244,"score_gpt":0.4630457612783989,"score_spread":0.3997897938678165,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7118043854","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000046287634,0.00079900154,0.013262155,0.0005266584,0.0020052544,0.00085100136,0.00007730142,0.00015627538,0.9822761],"genre_scores_gemma":[0.002318398,0.001255824,0.028139034,0.0069238176,0.0014279344,0.00002411531,0.00010326125,0.000055685217,0.9597519],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9977814,0.000008461542,0.0010600219,0.0006129525,0.000303834,0.00023336257],"domain_scores_gemma":[0.9142319,0.08458042,0.00018760528,0.0004967587,0.00026652266,0.00023681592],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00067528867,0.00029138674,0.001006031,0.00012118102,0.000051597897,0.000020283856,0.00014327181,0.00065613235,0.0031681992],"category_scores_gemma":[0.07981386,0.00022722996,0.0009395757,0.00002299444,0.0001051879,0.000012545552,0.0001047928,0.0005619727,0.00033602532],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00045133836,0.00007265955,0.000260487,0.000046730736,0.00017736264,0.00008197974,0.0000019398942,2.01044e-7,9.669031e-8,0.03287099,0.34962314,0.61641306],"study_design_scores_gemma":[0.0035125616,0.00044403443,0.00023195123,0.01864509,0.0011864965,0.00004624774,0.0000014771867,0.00027292006,0.0000023922225,0.08100437,0.8943202,0.00033227084],"about_ca_topic_score_codex":6.679443e-7,"about_ca_topic_score_gemma":0.0000013500254,"teacher_disagreement_score":0.6160808,"about_ca_system_score_codex":0.00004975754,"about_ca_system_score_gemma":0.0004299727,"threshold_uncertainty_score":0.997743},"labels":[],"label_agreement":null},{"id":"W7118085648","doi":"10.1093/geroni/igaf122.2277","title":"Pilot of a Priorities Aligned Decision-Making Microskills Curriculum","year":2025,"lang":"en","type":"article","venue":"Innovation in Aging","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Alpha Technologies (Canada)","funders":"","keywords":"Curriculum; Baseline (sea); Adaptation (eye); Scale (ratio); Resource (disambiguation); Health care; Usability; Core competency; Work (physics)","score_opus":0.01789541070635135,"score_gpt":0.36968071686225684,"score_spread":0.3517853061559055,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7118085648","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98622936,0.0002542444,0.007971534,0.0007437467,0.0005999059,0.00018388937,0.0000020507111,0.00003778023,0.003977496],"genre_scores_gemma":[0.9916127,0.000013584676,0.007065942,0.0009991139,0.00006492202,0.000014109441,0.00000700944,0.0000068352915,0.00021583062],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99877965,0.000018578134,0.0007165487,0.0001760311,0.00016468704,0.00014449388],"domain_scores_gemma":[0.99428463,0.0050308686,0.00014893993,0.00019463562,0.00032551461,0.000015422942],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0005367083,0.00008104253,0.00024334878,0.0005538367,0.000030009274,0.000015766005,0.000063806794,0.00003981663,0.00004453803],"category_scores_gemma":[0.03607716,0.00007557224,0.000032690004,0.0018580777,0.00005393861,0.000045037472,0.00004768058,0.00010453186,0.0000056004856],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000088778375,0.00038569904,0.87369055,0.00007079671,0.000033136046,0.000026220547,0.00040102113,0.000027117954,0.0030180523,0.0410146,0.0027987175,0.07844533],"study_design_scores_gemma":[0.0039879484,0.00024904584,0.9048811,0.04773421,0.000060137456,0.000018114022,0.0011482368,0.00053142745,0.004619374,0.035751395,0.0007816394,0.00023738123],"about_ca_topic_score_codex":0.0001345904,"about_ca_topic_score_gemma":0.0000091276725,"teacher_disagreement_score":0.07820795,"about_ca_system_score_codex":0.000061918225,"about_ca_system_score_gemma":0.00014503952,"threshold_uncertainty_score":0.9720424},"labels":[],"label_agreement":null},{"id":"W7127627197","doi":"10.1051/pmed/2025023","title":"Modéliser une situation de triage en accès direct : une traduction didactique du raisonnement clinique en kinésithérapie","year":2025,"lang":"fr","type":"article","venue":"Pédagogie médicale","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Triage; Validation test; Poison control; Signs and symptoms","score_opus":0.026251779369517194,"score_gpt":0.3649155197610616,"score_spread":0.33866374039154445,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7127627197","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.49227372,0.023421632,0.06763145,0.358068,0.010630555,0.0027371657,0.00013532933,0.0006893714,0.04441278],"genre_scores_gemma":[0.9015155,0.026746202,0.0027346888,0.004149126,0.0043825773,0.00032328637,0.00024468274,0.00009091945,0.05981301],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9946252,0.0012754111,0.0013925495,0.001000787,0.0006546194,0.0010514145],"domain_scores_gemma":[0.96718997,0.030749459,0.00032986785,0.0008263817,0.00021637634,0.0006879392],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0039157444,0.0005556559,0.001128124,0.0002809775,0.0002721403,0.00011131549,0.00037941686,0.0010548707,0.0013502474],"category_scores_gemma":[0.08296743,0.00054084003,0.00055816333,0.0011939445,0.00038508905,0.00021558687,0.00017633189,0.0012773639,0.00028740568],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0015960131,0.0061576595,0.026076019,0.0005362562,0.0009704764,0.0004283982,0.008095866,0.0010625779,0.0028265778,0.0031433837,0.13908663,0.81002015],"study_design_scores_gemma":[0.020176524,0.0020537986,0.47279924,0.01448072,0.0038943277,0.000095909774,0.0034861946,0.020433202,0.014674062,0.011380507,0.43508008,0.0014454325],"about_ca_topic_score_codex":0.0019998364,"about_ca_topic_score_gemma":0.00017718485,"teacher_disagreement_score":0.8085747,"about_ca_system_score_codex":0.0009181574,"about_ca_system_score_gemma":0.0013994446,"threshold_uncertainty_score":0.9997043},"labels":[],"label_agreement":null},{"id":"W7132892953","doi":"","title":"Applying Clinician Macrocognition in Designing a Clinical Decision Support Tool for Congenital Heart Disease","year":2023,"lang":"","type":"dissertation","venue":"TSpace","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Vector Institute; Toronto Rehabilitation Institute; University of Toronto","funders":"","keywords":"Decision support system; Clinical decision support system; Sociotechnical system; Emergency department; Heart disease; Decision analysis; Cognition; Disease","score_opus":0.07958381629875226,"score_gpt":0.4753132311568088,"score_spread":0.3957294148580565,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7132892953","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9738134,0.0007984635,0.0054644477,0.00091041904,0.007091582,0.010950992,0.00021160331,0.0003137824,0.00044528124],"genre_scores_gemma":[0.9537763,0.0016064391,0.009961215,0.0031297598,0.0033880982,0.0062233848,0.0097751,0.00048615222,0.011653539],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99058264,0.00048240015,0.004154261,0.0021599266,0.0011638665,0.0014569284],"domain_scores_gemma":[0.88332117,0.11276752,0.0010079154,0.0008503721,0.0007315631,0.0013214657],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.006740885,0.0009604621,0.0024351678,0.0006332563,0.00035291773,0.00019026405,0.00034476907,0.0015083334,0.00058308424],"category_scores_gemma":[0.3107923,0.0010220879,0.0014879336,0.0009094636,0.0002305565,0.00014869394,0.00012453287,0.0018549907,0.0019551343],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.045616884,0.004260469,0.37774032,0.0014606044,0.0005963426,0.0017901107,0.004220486,0.00011924779,0.00027189465,0.000087314904,0.027628005,0.53620833],"study_design_scores_gemma":[0.039120372,0.0060838833,0.8371182,0.06323792,0.006281655,0.000036752295,0.02480762,0.009353965,0.00025926525,0.0045547965,0.0059574684,0.0031880767],"about_ca_topic_score_codex":0.00021434588,"about_ca_topic_score_gemma":0.00028442754,"teacher_disagreement_score":0.53302026,"about_ca_system_score_codex":0.00023850541,"about_ca_system_score_gemma":0.002499917,"threshold_uncertainty_score":0.9997879},"labels":[],"label_agreement":null},{"id":"W7133238288","doi":"10.3138/come.2006.008","title":"What healthcare students do with what they don't know: The socializing power of ‘uncertainty’ in the case presentation","year":2006,"lang":"en","type":"article","venue":"Communication & Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; University of Waterloo","funders":"","keywords":"Rhetoric; Presentation (obstetrics); Identity (music); Power (physics); Health care; Empathy","score_opus":0.04329479421904737,"score_gpt":0.4145378900948037,"score_spread":0.3712430958757563,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7133238288","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.82274634,0.024859931,0.000019114399,0.15049087,0.00017300669,0.00089720474,0.000001706504,0.000021207208,0.0007906183],"genre_scores_gemma":[0.9869207,0.009806478,0.000050247774,0.0027753138,0.000122316,0.000080550366,0.00010182455,0.000014770534,0.00012776558],"study_design_codex":"observational","study_design_gemma":"qualitative","domain_scores_codex":[0.9977643,0.00072638824,0.00056297577,0.00017741138,0.000585858,0.0001830742],"domain_scores_gemma":[0.9874049,0.010728497,0.00030139805,0.0012056917,0.0002983006,0.00006124804],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002216972,0.00013621859,0.00033125558,0.00007145927,0.0001913961,0.00008166229,0.0004688566,0.00008081367,0.00005223332],"category_scores_gemma":[0.0036182455,0.00006548559,0.000051987914,0.00033172424,0.00041883189,0.00021421124,0.000072391296,0.00040611622,0.00000376093],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019543245,0.0043008835,0.44549814,0.00035089478,0.00037194826,0.0008401629,0.29051778,0.00084782974,0.00016049517,0.024402022,0.03053246,0.20022306],"study_design_scores_gemma":[0.010903312,0.0013484091,0.3622208,0.026134256,0.0004796942,0.0008737227,0.584766,0.00031161876,0.000036795875,0.0055334237,0.007107838,0.0002841857],"about_ca_topic_score_codex":0.007609356,"about_ca_topic_score_gemma":0.005814009,"teacher_disagreement_score":0.29424816,"about_ca_system_score_codex":0.000071108494,"about_ca_system_score_gemma":0.00009299562,"threshold_uncertainty_score":0.99899906},"labels":[],"label_agreement":null},{"id":"W7135485991","doi":"","title":"Eliciting user decision requirements, designing and developing a diagnostic support tool for general physicians","year":2015,"lang":"en","type":"article","venue":"Research Portal (King's College London)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute of Population and Public Health","funders":"","keywords":"Decision support system; MEDLINE; Key (lock); Context (archaeology)","score_opus":0.12290046381298152,"score_gpt":0.43270929660877777,"score_spread":0.30980883279579624,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7135485991","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9799597,0.00048432537,0.0130195115,0.0013328099,0.00033042187,0.0021658863,0.00006805107,0.00011977297,0.002519485],"genre_scores_gemma":[0.9323913,0.00029389892,0.06270334,0.001102191,0.00096177554,0.00043501594,0.00015868322,0.00010161852,0.0018521639],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9954973,0.00018121896,0.0008231723,0.0007814854,0.0015197088,0.0011971309],"domain_scores_gemma":[0.9718388,0.025755232,0.0001753312,0.0004535703,0.0011490259,0.00062805007],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.004381994,0.00030013747,0.00066569226,0.00033513395,0.00042155787,0.00014064912,0.00021847457,0.00020661722,0.00006029441],"category_scores_gemma":[0.14694394,0.0002763333,0.0001633807,0.0007171496,0.00023683276,0.00025306467,0.00029329435,0.0005157935,0.00005014072],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.003052183,0.0012625647,0.42615744,0.0005341483,0.00058229174,0.0056726485,0.001493823,0.00008580495,0.0030988331,0.020434614,0.17825085,0.3593748],"study_design_scores_gemma":[0.11285778,0.025854781,0.33206853,0.07690018,0.001963832,0.0035701978,0.014963267,0.019039068,0.043329004,0.078618415,0.28325135,0.0075835884],"about_ca_topic_score_codex":0.00015330735,"about_ca_topic_score_gemma":0.000071232345,"teacher_disagreement_score":0.3517912,"about_ca_system_score_codex":0.00021908301,"about_ca_system_score_gemma":0.0012726933,"threshold_uncertainty_score":0.9999689},"labels":[],"label_agreement":null},{"id":"W7135655133","doi":"","title":"Implementation of the TRANSFoRm evidence service supporting diagnosis in primary care","year":2015,"lang":"en","type":"article","venue":"Research Portal (King's College London)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute of Population and Public Health","funders":"","keywords":"Primary care; Service (business); Primary health care; MEDLINE; Health services","score_opus":0.10955370475873111,"score_gpt":0.4530160298050329,"score_spread":0.3434623250463018,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7135655133","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9826474,0.001222343,0.0000033891672,0.006552413,0.00014179578,0.001726058,0.00007037484,0.000026499052,0.0076097087],"genre_scores_gemma":[0.998355,0.0003239721,0.00016340756,0.0005002938,0.0001125919,0.0002664505,0.00005807574,0.000026601338,0.00019362842],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99620104,0.00026461147,0.00081925007,0.00038261487,0.001693249,0.0006392641],"domain_scores_gemma":[0.99155426,0.006395711,0.00017740305,0.00048158548,0.0010381023,0.00035295993],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0031985852,0.00015536611,0.00042834444,0.00021519704,0.00010809499,0.000023883294,0.0003165159,0.00012744516,0.00021842387],"category_scores_gemma":[0.016255246,0.00011565795,0.00015086938,0.0015917923,0.00015138973,0.00019387699,0.00017638208,0.0006042119,0.000020344809],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00028072597,0.00021730506,0.9727372,0.00047734036,0.00003108266,0.00027162596,0.0026428013,0.000012263253,0.00017799532,0.00024267953,0.0031117692,0.019797197],"study_design_scores_gemma":[0.005288904,0.0010100175,0.94655824,0.009063219,0.00009294985,0.00006773283,0.024796454,0.00008613612,0.010883888,0.0004376559,0.0014844863,0.00023032757],"about_ca_topic_score_codex":0.0060896706,"about_ca_topic_score_gemma":0.008595595,"teacher_disagreement_score":0.026178986,"about_ca_system_score_codex":0.00032100113,"about_ca_system_score_gemma":0.0023710947,"threshold_uncertainty_score":0.9920313},"labels":[],"label_agreement":null},{"id":"W7136031779","doi":"10.6084/m9.figshare.4473794","title":"Additional file 2: of Cognitive biases associated with medical decisions: a systematic review","year":2016,"lang":"","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Cognitive bias; Cognition; Data collection; Cognitive interview","score_opus":0.09057787686179139,"score_gpt":0.35974586730922165,"score_spread":0.2691679904474302,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7136031779","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000042637303,0.014266955,0.0000010152108,0.00037003637,0.000030180176,0.001515575,0.9822885,0.000046782578,0.0014767265],"genre_scores_gemma":[0.0023860824,0.0011678014,0.000016869757,0.002442766,0.0001674056,0.0038177706,0.98779714,0.000042931773,0.002161246],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9951598,0.00045060352,0.0014445639,0.0005162708,0.0019876896,0.00044105083],"domain_scores_gemma":[0.012136787,0.98336285,0.0013120571,0.00047455245,0.0018754819,0.0008382549],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00033468162,0.00035725214,0.0018086253,0.00009430261,0.000077325545,0.000014807251,0.00028682113,0.0003778116,0.999649],"category_scores_gemma":[0.9992105,0.00018889584,0.00040030244,0.0005824469,0.000117326395,0.000080223144,0.00014775866,0.00031314904,0.016897172],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005482624,0.0006482271,0.0000014760162,0.01163001,0.0004953325,0.00036528058,0.000011798387,1.0038419e-8,1.6531912e-8,7.3093867e-7,0.98412883,0.0026634673],"study_design_scores_gemma":[0.0017616014,0.00086751557,0.00081317546,0.9930471,0.0012358555,0.00011078588,0.0000433978,0.000020943105,0.0000032571797,0.000010041848,0.001821102,0.00026518464],"about_ca_topic_score_codex":0.0000018626802,"about_ca_topic_score_gemma":0.000010820609,"teacher_disagreement_score":0.9988758,"about_ca_system_score_codex":0.000094641764,"about_ca_system_score_gemma":0.0021342344,"threshold_uncertainty_score":0.9838683},"labels":[],"label_agreement":null},{"id":"W7136085559","doi":"10.6084/m9.figshare.4473794.v1","title":"Additional file 2: of Cognitive biases associated with medical decisions: a systematic review","year":2016,"lang":"","type":"article","venue":"Figshare","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Cognitive bias; Cognition; Data collection; Cognitive interview","score_opus":0.09057787686179139,"score_gpt":0.35974586730922165,"score_spread":0.2691679904474302,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7136085559","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000042637303,0.014266955,0.0000010152108,0.00037003637,0.000030180176,0.001515575,0.9822885,0.000046782578,0.0014767265],"genre_scores_gemma":[0.0023860824,0.0011678014,0.000016869757,0.002442766,0.0001674056,0.0038177706,0.98779714,0.000042931773,0.002161246],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9951598,0.00045060352,0.0014445639,0.0005162708,0.0019876896,0.00044105083],"domain_scores_gemma":[0.012136787,0.98336285,0.0013120571,0.00047455245,0.0018754819,0.0008382549],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00033468162,0.00035725214,0.0018086253,0.00009430261,0.000077325545,0.000014807251,0.00028682113,0.0003778116,0.999649],"category_scores_gemma":[0.9992105,0.00018889584,0.00040030244,0.0005824469,0.000117326395,0.000080223144,0.00014775866,0.00031314904,0.016897172],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005482624,0.0006482271,0.0000014760162,0.01163001,0.0004953325,0.00036528058,0.000011798387,1.0038419e-8,1.6531912e-8,7.3093867e-7,0.98412883,0.0026634673],"study_design_scores_gemma":[0.0017616014,0.00086751557,0.00081317546,0.9930471,0.0012358555,0.00011078588,0.0000433978,0.000020943105,0.0000032571797,0.000010041848,0.001821102,0.00026518464],"about_ca_topic_score_codex":0.0000018626802,"about_ca_topic_score_gemma":0.000010820609,"teacher_disagreement_score":0.9988758,"about_ca_system_score_codex":0.000094641764,"about_ca_system_score_gemma":0.0021342344,"threshold_uncertainty_score":0.9838683},"labels":[],"label_agreement":null},{"id":"W7154584503","doi":"10.2310/6620.2009.00005bb","title":"Physician-Patient Concordance Regarding Relevance of Positive Patch Tests","year":2009,"lang":"en","type":"article","venue":"Dermatitis","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Concordance; Relevance (law); Clinical significance; Patch testing; MEDLINE","score_opus":0.011580130450171896,"score_gpt":0.2978791288296425,"score_spread":0.2862989983794706,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7154584503","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9680682,0.000631093,0.000421219,0.0050019994,0.00015124009,0.00022674234,0.00002982763,0.000055571796,0.025414152],"genre_scores_gemma":[0.98536336,0.00018191253,0.0019979659,0.012208816,0.000086608336,0.000005446431,0.0000238564,0.000008482042,0.00012356811],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99908775,0.000027599026,0.00031837836,0.00016605003,0.00022901416,0.00017122975],"domain_scores_gemma":[0.9970578,0.0022993207,0.00015426929,0.00023009468,0.00015796845,0.000100547266],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.000057856527,0.00010076837,0.00032242452,0.000030317533,0.000032788965,0.000007257251,0.000048819104,0.000049440812,0.000017630786],"category_scores_gemma":[0.0075554526,0.00009030296,0.00009499022,0.00015540629,0.000056566238,0.000058345075,0.000013217443,0.00012134435,0.00004198871],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008596865,0.0009389335,0.055003908,0.000088679706,0.000094472794,0.0005285528,0.0015666238,0.000045953082,0.0033459025,0.0029260647,0.09245496,0.84292],"study_design_scores_gemma":[0.0017610039,0.001176481,0.9425126,0.006912931,0.00009055817,0.00010602826,0.00010645721,0.00043310452,0.04037914,0.002035635,0.0042769806,0.00020909394],"about_ca_topic_score_codex":0.000056950354,"about_ca_topic_score_gemma":0.000004276378,"teacher_disagreement_score":0.8875087,"about_ca_system_score_codex":0.00003512643,"about_ca_system_score_gemma":0.00004412348,"threshold_uncertainty_score":0.9045128},"labels":[],"label_agreement":null},{"id":"W7155082894","doi":"10.2196/88652","title":"Benefits of Chain of Thought Prompting for Clinical Record Rubric Evaluation in Undergraduate Medicine Education: An Experimental Evaluation Study with Medical Faculty (Preprint)","year":2025,"lang":"en","type":"article","venue":"JMIR Medical Education","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Rubric; Chain (unit); Medical record; MEDLINE; Data collection; Identification (biology)","score_opus":0.10370711201607251,"score_gpt":0.5304194780632709,"score_spread":0.4267123660471984,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7155082894","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9804487,0.0007587596,0.00022543599,0.010734672,0.0011148904,0.0056574675,0.0000027184865,0.00002956106,0.0010277809],"genre_scores_gemma":[0.9939499,0.00008046427,0.0013047233,0.0007491395,0.0005921578,0.0026543094,0.0005014073,0.000020735046,0.00014717334],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9938013,0.00079633814,0.0018447422,0.0006934987,0.0026284093,0.00023572368],"domain_scores_gemma":[0.9920196,0.0047547887,0.00061326905,0.00056913384,0.0015827918,0.00046042207],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.011197361,0.00021889698,0.0007720873,0.00033177392,0.000057056346,0.00000808179,0.00024096342,0.00032607382,0.0005340442],"category_scores_gemma":[0.13044997,0.00016629414,0.00010141492,0.0007844808,0.00029166177,0.00010544147,0.000056100835,0.00041349928,0.0000022559063],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00088902423,0.01873043,0.28055033,0.00014468956,0.000079539124,4.803669e-7,0.0017504519,0.000022714597,0.000015092418,0.0010246447,0.0022150204,0.6945776],"study_design_scores_gemma":[0.014034088,0.0052222074,0.9355559,0.015984258,0.00054237974,0.000014487027,0.012105756,0.01465269,0.0002509539,0.0012213658,0.00021073998,0.00020519133],"about_ca_topic_score_codex":0.00037073233,"about_ca_topic_score_gemma":0.0002735643,"teacher_disagreement_score":0.6943724,"about_ca_system_score_codex":0.00029068967,"about_ca_system_score_gemma":0.012398066,"threshold_uncertainty_score":0.9932007},"labels":[],"label_agreement":null},{"id":"W785136403","doi":"10.1097/ncc.0000000000000282","title":"Lived Experiences of “Illness Uncertainty” of Iranian Cancer Patients","year":2015,"lang":"en","type":"article","venue":"Cancer Nursing","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":44,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Brantford Energy (Canada)","funders":"","keywords":"Ambiguity; Ambivalence; Coping (psychology); Stressor; Medicine; Phenomenological method; Sociocultural evolution; Cancer; Social psychology; Psychology; Clinical psychology; Epistemology; Sociology","score_opus":0.057541632496806945,"score_gpt":0.39266584570044283,"score_spread":0.3351242132036359,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W785136403","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99436146,0.0021145756,0.000029194354,0.0002870143,0.0016493956,0.00016418187,0.000008791862,0.000016466169,0.0013689175],"genre_scores_gemma":[0.9992994,0.000120081,0.00011474927,0.00013156097,0.00017375329,0.000041604646,0.00000832332,0.000012318377,0.000098218465],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9989794,0.000027180291,0.00032350293,0.0001790875,0.00031252604,0.00017833756],"domain_scores_gemma":[0.9985674,0.0005011693,0.00018130738,0.00016226475,0.0003803627,0.00020746721],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.000112812755,0.00009723323,0.00036150136,0.000046393925,0.000024335572,0.000003885787,0.00007855476,0.00006864378,0.00013164598],"category_scores_gemma":[0.0034773943,0.000078318066,0.000069779526,0.00018915923,0.00028247584,0.000032437052,0.000011536682,0.00007840684,0.0000016675978],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012277578,0.0012424141,0.36377546,0.00013373523,0.00009837768,0.0000066367143,0.27330932,0.0005211534,0.0005073196,0.000046212972,0.0056620147,0.3534696],"study_design_scores_gemma":[0.027769892,0.0043340195,0.4119937,0.09436459,0.001415881,0.000008955889,0.358835,0.002996303,0.091949485,0.0011814667,0.003775543,0.0013751539],"about_ca_topic_score_codex":0.00086186134,"about_ca_topic_score_gemma":0.000009732779,"teacher_disagreement_score":0.35209444,"about_ca_system_score_codex":0.00016236106,"about_ca_system_score_gemma":0.00036696953,"threshold_uncertainty_score":0.4163017},"labels":[],"label_agreement":null},{"id":"W78801501","doi":"10.1007/978-90-481-2605-7_2","title":"On the Relationship Between Problem-Solving Skills and Professional Practice","year":2009,"lang":"en","type":"book-chapter","venue":"","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":12,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Heuristics; Flexibility (engineering); Psychology; Quality (philosophy); Field (mathematics); Management science; Computer science; Epistemology; Mathematics; Engineering","score_opus":0.053339685536814646,"score_gpt":0.36614942416478313,"score_spread":0.31280973862796846,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W78801501","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0013125185,0.0002661691,0.000036331585,0.041228704,0.00011461942,0.00084508135,0.000012423697,0.00008109095,0.9561031],"genre_scores_gemma":[0.0756485,0.00012730129,0.0018835306,0.008501101,0.0005403061,0.000011413711,0.00009579668,0.00004297574,0.91314906],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9984295,0.000056755387,0.00044192182,0.0003795055,0.00050376606,0.00018855902],"domain_scores_gemma":[0.8509915,0.1481129,0.00024466004,0.00034370908,0.0001302929,0.00017692515],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0008801507,0.00025833462,0.00040397485,0.00007361733,0.00017951052,0.00003359478,0.0000795096,0.00046883282,0.0006088868],"category_scores_gemma":[0.11904441,0.00014209007,0.00011517912,0.000029633178,0.000121392724,0.000045905177,0.0000561313,0.0012829865,0.000246722],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006828737,0.00019271632,0.012521311,0.00003120052,0.00011045478,0.00004968194,0.000120609046,4.1077766e-7,2.4308844e-7,0.9067179,0.05979279,0.020394389],"study_design_scores_gemma":[0.0018970676,0.0013433254,0.29738498,0.029845294,0.0014129011,0.00014474163,0.000058181344,0.000005689362,0.0000041901635,0.52551574,0.14177099,0.0006168788],"about_ca_topic_score_codex":0.0000069586886,"about_ca_topic_score_gemma":0.0000016635055,"teacher_disagreement_score":0.38120213,"about_ca_system_score_codex":0.000047504196,"about_ca_system_score_gemma":0.00020471925,"threshold_uncertainty_score":0.88837624},"labels":[],"label_agreement":null},{"id":"W825601419","doi":"","title":"The influence of task demands and experience on diagnostic accuracy: Investigating the assumptions of a default interventionist dual systems model","year":2014,"lang":"en","type":"dissertation","venue":"MacSphere (McMaster University)","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Dual (grammatical number); Task (project management); Computer science; Cognitive psychology; Psychology; Economics; Management; Linguistics; Philosophy","score_opus":0.023389567733921472,"score_gpt":0.2920291562143497,"score_spread":0.26863958848042824,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W825601419","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9832309,0.00039778205,0.00021102889,0.00013988372,0.00015792603,0.00046260035,0.000031706535,0.00002197446,0.015346198],"genre_scores_gemma":[0.95966184,0.00023968953,0.00006311636,0.00004762547,0.000034449782,0.000009819356,0.00005810073,0.000018323417,0.03986705],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9985472,0.00014587729,0.0004775222,0.00029798923,0.0003448311,0.00018659006],"domain_scores_gemma":[0.9859291,0.012655174,0.00064312475,0.0003828574,0.0002619184,0.00012783363],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00026782887,0.00021161471,0.00041252087,0.000067652014,0.0002725529,0.000050281156,0.00024952032,0.00019006107,0.00012915717],"category_scores_gemma":[0.018755814,0.00014257256,0.00015576914,0.00024316719,0.00040755406,0.00006835841,0.00008042945,0.00036047798,0.000004577086],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.003916002,0.0021438005,0.24129188,0.008238983,0.002729287,0.00027456175,0.03788258,0.15483794,0.0039311755,0.085074574,0.004434235,0.455245],"study_design_scores_gemma":[0.010830413,0.0038620958,0.6666355,0.12562245,0.005502953,0.00010206645,0.070043966,0.08790561,0.0016700333,0.001810837,0.024102656,0.0019114215],"about_ca_topic_score_codex":0.00039910624,"about_ca_topic_score_gemma":0.0003817695,"teacher_disagreement_score":0.45333356,"about_ca_system_score_codex":0.000045964967,"about_ca_system_score_gemma":0.00014624817,"threshold_uncertainty_score":0.98950964},"labels":[],"label_agreement":null},{"id":"W89697325","doi":"10.1097/00152192-200409000-00006","title":"Can Decision Support Improve NursesʼUse of Knowledge?","year":2004,"lang":"en","type":"review","venue":"Journal of Wound Ostomy and Continence Nursing","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Psychology; Medical education; Computer science; Medicine","score_opus":0.03164325837278924,"score_gpt":0.3970773972499388,"score_spread":0.36543413887714954,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W89697325","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0021980302,0.9945227,0.00018691664,0.00025831134,0.0015390797,0.0003319165,0.000014309211,0.000006979505,0.0009417504],"genre_scores_gemma":[0.0029706804,0.9937886,0.0021031378,0.00004152091,0.00060729467,0.0000026954533,0.000009046106,0.000034454246,0.00044257104],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9973346,0.00006798056,0.0016549686,0.0002638867,0.0003913936,0.00028716703],"domain_scores_gemma":[0.99267757,0.0045335637,0.001575237,0.00026977403,0.0005302649,0.0004135876],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.00074687816,0.00033366337,0.002506503,0.00031364622,0.000060771574,0.000037062036,0.00018000018,0.0003683833,0.000053052885],"category_scores_gemma":[0.009379593,0.00023063444,0.0007060064,0.00028362172,0.00031346764,0.000081405145,0.000023897419,0.0008218252,0.0000068119953],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013910062,0.00044024026,0.000082365485,0.0020230154,0.00011046406,0.0001305282,0.000300291,1.7535554e-7,0.0000031859033,0.0000846808,0.00058027083,0.9961057],"study_design_scores_gemma":[0.0027963447,0.001994143,0.00030612884,0.5931504,0.0045826295,0.0039518746,0.00018088079,0.0000024501599,0.000018367757,0.0006379708,0.3920768,0.00030203976],"about_ca_topic_score_codex":0.00001275113,"about_ca_topic_score_gemma":0.0000022135855,"teacher_disagreement_score":0.99580365,"about_ca_system_score_codex":0.00022208042,"about_ca_system_score_gemma":0.0020836587,"threshold_uncertainty_score":0.99896485},"labels":[],"label_agreement":null},{"id":"W972375664","doi":"10.1007/s11739-015-1283-8","title":"Insights into emergency physicians’ minds in the seconds before and into a patient encounter","year":2015,"lang":"en","type":"article","venue":"Internal and Emergency Medicine","topic":"Clinical Reasoning and Diagnostic Skills","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Centre for Health Evaluation and Outcome Sciences; Université de Montréal; Université de Sherbrooke","funders":"H2020 European Research Council; Shandong First Medical University; Université de Sherbrooke","keywords":"Seriousness; Medicine; Perspective (graphical); Thematic analysis; Medical education; Medical emergency; Qualitative research; Artificial intelligence; Computer science","score_opus":0.026941444109913112,"score_gpt":0.33553411287909296,"score_spread":0.30859266876917985,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W972375664","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97856885,0.0058550867,0.000051357125,0.006877026,0.0017224655,0.00021790902,0.0000013909222,0.000016503525,0.006689432],"genre_scores_gemma":[0.9957112,0.0013805946,0.000041643834,0.0010037193,0.0008199681,0.000035457193,0.000018869083,0.000015138064,0.00097342057],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982152,0.000068126785,0.000707929,0.00034528243,0.00043567037,0.00022779997],"domain_scores_gemma":[0.9989301,0.00026602586,0.00012543169,0.0002462764,0.00015368097,0.0002785053],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00033478226,0.00022333588,0.00040528347,0.00012820268,0.000053097934,0.0000046589794,0.00013255236,0.000093251336,0.0004801377],"category_scores_gemma":[0.0034490956,0.00012332243,0.00006649909,0.00024635514,0.000184212,0.00008127921,0.00008180074,0.00032942643,0.00001862762],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00056138035,0.000681374,0.58717006,0.00017279999,0.00015985461,0.00028447536,0.16658549,0.0000016312715,0.00057737925,0.00069352356,0.08567329,0.15743873],"study_design_scores_gemma":[0.00851624,0.013034253,0.80569625,0.0068188882,0.0004959814,0.00025993751,0.033360265,0.0017201072,0.00020873331,0.050227232,0.07895185,0.0007102375],"about_ca_topic_score_codex":0.0024525437,"about_ca_topic_score_gemma":0.0016872179,"teacher_disagreement_score":0.21852618,"about_ca_system_score_codex":0.00003469764,"about_ca_system_score_gemma":0.000029029388,"threshold_uncertainty_score":0.52571714},"labels":[],"label_agreement":null}]}