{"meta":{"query_hash":"d263e7495388","filters":{"venue":"BMJ evidence-based medicine"},"cohort_total":125,"direct_labels_cover":115,"predictions_cover":125,"exported":125,"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/d263e7495388","api":"https://metacan.xera.ac/api/v1/cohort?venue=BMJ+evidence-based+medicine"},"results":[{"id":"W2791660821","doi":"10.1136/bmjebm-2017-110852","title":"Assessing the validity of surrogate endpoints in the context of a controversy about the measurement of effectiveness of hepatitis C virus treatment","year":2018,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Hepatitis C virus research","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":"McMaster University; Impact","funders":"National Health and Medical Research Council","keywords":"Surrogate endpoint; Medicine; Clinical endpoint; Intensive care medicine; Context (archaeology); Endpoint Determination; Randomized controlled trial; Clinical trial; Internal medicine","score_opus":0.2545075038471375,"score_gpt":0.4540353686146979,"score_spread":0.1995278647675604,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2791660821","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.02737172,0.23060225,0.20340094,0.49530357,0.012236424,0.0012032329,0.0007551425,0.0001285765,0.028998112],"genre_scores_gemma":[0.6480633,0.065167256,0.209051,0.058204256,0.014567224,0.0022710832,0.00046170317,0.00020550749,0.002008714],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.33331063,0.58824086,0.03959506,0.004374207,0.033126976,0.001352265],"domain_scores_gemma":[0.13150969,0.8105372,0.02789977,0.013070288,0.015451023,0.0015321019],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.5097702,0.001523533,0.005246573,0.009132975,0.0026987328,0.017810076,0.004374332,0.014395604,0.0017715633],"category_scores_gemma":[0.8031639,0.0010534151,0.0038640262,0.0063680615,0.028030291,0.0123480335,0.008928608,0.019054793,0.00057238125],"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":[0.0033047886,0.00017844002,0.015031497,0.01554599,0.004469305,0.0013989885,0.014898976,0.012858991,0.0009576795,0.58951133,0.03655569,0.30528834],"study_design_scores_gemma":[0.00029580333,0.000677554,0.0048438516,0.021967815,0.0010267728,0.0008951726,0.0027729676,0.0077973832,0.001120993,0.8989614,0.05932599,0.0003142798],"about_ca_topic_score_codex":0.0018720797,"about_ca_topic_score_gemma":0.0018798127,"teacher_disagreement_score":0.4902298,"about_ca_system_score_codex":0.0065544187,"about_ca_system_score_gemma":0.0109123485,"threshold_uncertainty_score":0.60454065},"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":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2792407982","doi":"10.1136/bmjebm-2017-110838","title":"Procalcitonin-guided decision-making results in a significant reduction of antibiotic therapy and hospital stay in neonates with suspected early-onset sepsis","year":2018,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Sepsis Diagnosis and Treatment","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":"Procalcitonin; Medicine; Sepsis; Antibiotic therapy; Antibiotics; Intensive care medicine; Pediatrics; Randomized controlled trial; Population; Clinical trial; Internal medicine","score_opus":0.1723921777231203,"score_gpt":0.43128935713986133,"score_spread":0.25889717941674106,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2792407982","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00034629952,0.96720225,0.000121388155,0.024021044,0.006909843,0.0000612452,0.00028061002,0.000014027116,0.0010432373],"genre_scores_gemma":[0.010207479,0.9487878,0.00041852155,0.029550433,0.008992812,0.00016888179,0.0003262425,0.000019882775,0.0015278944],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9974502,0.0011781455,0.00049896096,0.00018211131,0.00059005135,0.00010058024],"domain_scores_gemma":[0.98627996,0.010361883,0.0013113191,0.0001226074,0.0017197293,0.00020451496],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0040211356,0.00092265086,0.003750192,0.0014440652,0.00024481822,0.0013546261,0.002246112,0.0040521976,0.008267507],"category_scores_gemma":[0.034623366,0.00035129473,0.002773887,0.0017260975,0.00043155023,0.0013946161,0.0005378372,0.003563843,0.001392315],"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":[0.0026427456,0.00006969982,0.00050248334,0.3186988,0.0042558564,0.0004613881,0.00021839974,0.00032563217,0.00041812676,0.0027818338,0.34015408,0.32947087],"study_design_scores_gemma":[0.0039329515,0.0006859925,0.005543461,0.43580315,0.012923626,0.0011824301,0.0003026592,0.00042583724,0.0007644841,0.0055463347,0.53273976,0.00014927352],"about_ca_topic_score_codex":0.0033625273,"about_ca_topic_score_gemma":0.005875991,"teacher_disagreement_score":0.008267507,"about_ca_system_score_codex":0.001456194,"about_ca_system_score_gemma":0.002987758,"threshold_uncertainty_score":0.027657509},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2800629548","doi":"10.1136/bmjebm-2018-110943","title":"Screening ECGs in low-risk patients are associated with increased risk of downstream cardiac testing","year":2018,"lang":"en","type":"letter","venue":"BMJ evidence-based medicine","topic":"Healthcare cost, quality, practices","field":"Health Professions","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":"Downstream (manufacturing); Internal medicine; Medicine; Cardiology; Risk assessment; Computer science; Engineering; Operations management; Computer security","score_opus":0.4956000741268031,"score_gpt":0.49399743443416266,"score_spread":0.0016026396926404618,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2800629548","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0016232296,0.0024967117,0.00006722378,0.9766492,0.015477116,0.000017836395,0.00056324364,0.000027336584,0.0030780125],"genre_scores_gemma":[0.030552616,0.00586272,0.0002882159,0.83717275,0.12020859,0.0000883257,0.00046523096,0.00003991054,0.005321692],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9972633,0.0008041314,0.00043644445,0.00030055377,0.00094367703,0.00025183486],"domain_scores_gemma":[0.97792083,0.013215527,0.0016073238,0.0003511511,0.005749666,0.0011553471],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0026094909,0.00045715465,0.0011056014,0.0013896637,0.0013295935,0.0013682455,0.0015243406,0.017545316,0.011888062],"category_scores_gemma":[0.046892226,0.0003912538,0.0009067493,0.0016641795,0.0011028711,0.0016375553,0.00050554785,0.0096460255,0.0066996412],"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":[0.00012682789,0.00002712649,0.0049441596,0.00025379844,0.000028224686,0.00075408444,0.0001192851,0.000031543084,0.00006765315,0.00057507097,0.98297817,0.010093937],"study_design_scores_gemma":[0.0013859867,0.00037367095,0.060473446,0.012094672,0.0005229434,0.009495078,0.0026866887,0.0019243585,0.00075705146,0.014199625,0.8957779,0.00030856224],"about_ca_topic_score_codex":0.018766623,"about_ca_topic_score_gemma":0.024308357,"teacher_disagreement_score":0.018766623,"about_ca_system_score_codex":0.002270288,"about_ca_system_score_gemma":0.0029212697,"threshold_uncertainty_score":0.03976953},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2809086820","doi":"10.1136/bmjebm-2018-110979","title":"Essential medicines and the challenges in the Evidence-Based Manifesto","year":2018,"lang":"en","type":"editorial","venue":"BMJ evidence-based medicine","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","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; St. Michael's Hospital","funders":"","keywords":"Manifesto; Data science; Computer science; Political science; Law","score_opus":0.517188658566685,"score_gpt":0.4876869785974194,"score_spread":0.02950167996926556,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2809086820","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00009305693,0.064586125,0.0005078772,0.33032542,0.60259056,0.000023414454,0.000084951156,0.00007063247,0.0017179658],"genre_scores_gemma":[0.0021518187,0.030915,0.00065281853,0.103564076,0.85921985,0.00004742293,0.00003648533,0.000067819,0.0033447633],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9694668,0.01324049,0.0031920325,0.001732086,0.011543132,0.0008255769],"domain_scores_gemma":[0.7710573,0.20130484,0.0045178984,0.0032063883,0.016503327,0.0034101426],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.035976596,0.002852008,0.0032587857,0.0037368273,0.0027672022,0.010157488,0.00416427,0.02077877,0.004939637],"category_scores_gemma":[0.14623183,0.0012098056,0.0025222804,0.002353918,0.009399878,0.01267493,0.002418515,0.039669644,0.0025891894],"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":[0.000050542843,0.000017154076,0.000049537375,0.0010744313,0.000045804794,0.000091885246,0.00015890252,0.00008902393,0.000032448697,0.013291525,0.97368354,0.011415186],"study_design_scores_gemma":[0.00014659074,0.000048432554,0.00029462876,0.003879942,0.000074893,0.00020495836,0.00020687933,0.00041981554,0.00009381422,0.03151854,0.96306247,0.000049128354],"about_ca_topic_score_codex":0.0026536942,"about_ca_topic_score_gemma":0.0041593653,"teacher_disagreement_score":0.035976596,"about_ca_system_score_codex":0.0060666013,"about_ca_system_score_gemma":0.008991974,"threshold_uncertainty_score":0.19026476},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2809195898","doi":"10.1136/bmjebm-2018-110968","title":"Reflections on the history of systematic reviews","year":2018,"lang":"en","type":"editorial","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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":"Queen's University","funders":"","keywords":"Psychological intervention; Systematic review; MEDLINE; Medicine; Value (mathematics); Evidence-based medicine; Alternative medicine; Intensive care medicine; Computer science; Political science; Pathology; Psychiatry","score_opus":0.9165977472242489,"score_gpt":0.6280865975200167,"score_spread":0.2885111497042322,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2809195898","genre_codex":"commentary","genre_gemma":"editorial","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000050408195,0.17056993,0.00060488016,0.6171408,0.2102184,0.000025979014,0.00007562433,0.000054173935,0.0012597681],"genre_scores_gemma":[0.0020180636,0.14458211,0.0016562603,0.49897233,0.35006845,0.00018218864,0.000055331253,0.00017606773,0.002289219],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.76615214,0.11352998,0.035603963,0.010687983,0.07038541,0.0036405323],"domain_scores_gemma":[0.20996419,0.65077555,0.014715179,0.014717678,0.0998072,0.010020291],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.23697731,0.0030622447,0.0064977156,0.013525213,0.0052527934,0.02712596,0.007894434,0.035872713,0.010231908],"category_scores_gemma":[0.5847899,0.0029837452,0.0058051157,0.010238344,0.027102772,0.042070966,0.009709366,0.08315832,0.005791347],"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":[0.00014829701,0.000030745927,0.000109227185,0.005669051,0.00025488282,0.00011767485,0.0010206082,0.00012338182,0.00004255938,0.024305137,0.9280829,0.040095553],"study_design_scores_gemma":[0.00014987131,0.00004750093,0.00022442635,0.022228723,0.00015113769,0.0002180567,0.0005030933,0.00012047668,0.00007497945,0.026941828,0.949244,0.00009591563],"about_ca_topic_score_codex":0.006849435,"about_ca_topic_score_gemma":0.010297637,"teacher_disagreement_score":0.76302266,"about_ca_system_score_codex":0.020166164,"about_ca_system_score_gemma":0.023958722,"threshold_uncertainty_score":0.9409428},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","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":"split"},{"id":"W2809788330","doi":"10.1136/bmjebm-2018-111014","title":"Don’t just blame the evidence: considering the role of medical education in the poor uptake of evidence-based medicine in clinical practice","year":2018,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","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":"Blame; Critical appraisal; Evidence-based medicine; Medical education; Curriculum; Manifesto; Quality (philosophy); Psychology; Health care; Evidence-based practice; Medicine; Alternative medicine; Nursing; Pedagogy; Political science; Epistemology; Social psychology","score_opus":0.5347146743972018,"score_gpt":0.6382019501777273,"score_spread":0.10348727578052552,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2809788330","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0009898279,0.019459575,0.00025629764,0.97602934,0.0025946144,0.0000074746695,0.0000085240235,0.000005955735,0.0006484631],"genre_scores_gemma":[0.09778902,0.049871963,0.0035983147,0.82063025,0.026801698,0.0000987376,0.0000308663,0.000090500835,0.0010886922],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.866812,0.101452455,0.0074081137,0.0049297176,0.01520429,0.0041934797],"domain_scores_gemma":[0.43243232,0.50020355,0.017682103,0.0052580843,0.025949435,0.018474553],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.15486568,0.00067742146,0.00201039,0.0039394074,0.009578258,0.02283265,0.0068309032,0.042334527,0.005386491],"category_scores_gemma":[0.3714734,0.0013504159,0.0016557588,0.003858313,0.03549133,0.036828782,0.010260329,0.04215114,0.0009474665],"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":[0.00048169805,0.00018054324,0.0100408085,0.00839892,0.00060536066,0.0043099383,0.04354727,0.00083901454,0.0003744177,0.109209165,0.60867643,0.21333642],"study_design_scores_gemma":[0.00030942404,0.00036345783,0.009450306,0.032055717,0.00044030458,0.004211589,0.07003175,0.0018946667,0.00048689678,0.31114143,0.56923604,0.00037843766],"about_ca_topic_score_codex":0.017162772,"about_ca_topic_score_gemma":0.050490163,"teacher_disagreement_score":0.8451343,"about_ca_system_score_codex":0.01730749,"about_ca_system_score_gemma":0.06350183,"threshold_uncertainty_score":0.81901777},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"incentives","study_design":"theoretical_or_conceptual","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2837347473","doi":"10.1136/bmjebm-2018-110981","title":"One of the proposed solutions of the EBM Manifesto Educate the public in evidence-based healthcare to make informed decisions","year":2018,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Patient-Provider Communication in Healthcare","field":"Health Professions","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":"Manifesto; Health care; Evidence-based medicine; Public relations; Political science; Psychology; Computer science; Medical education; MEDLINE; Medicine; Law","score_opus":0.6084637929433494,"score_gpt":0.5074133762075725,"score_spread":0.10105041673577686,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2837347473","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0013165731,0.0026192938,0.017772991,0.9346545,0.010660094,0.00028128346,0.0001299778,0.00018600622,0.03237939],"genre_scores_gemma":[0.11165024,0.007509875,0.11369934,0.69306225,0.010111638,0.0019464147,0.00031985907,0.00050029723,0.061200052],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.8813733,0.08458451,0.0062064677,0.0042202407,0.018001938,0.0056136115],"domain_scores_gemma":[0.8440029,0.09681767,0.009938141,0.01276922,0.0225662,0.013905906],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.12458691,0.0015192203,0.0010189979,0.0023525816,0.008830499,0.021555016,0.004180207,0.032415725,0.012364047],"category_scores_gemma":[0.14849995,0.0009336651,0.0024621484,0.0021039604,0.032572605,0.030456048,0.015886102,0.043906078,0.0057229297],"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":[0.00005750706,0.00016042226,0.00072720385,0.00089544005,0.000037116493,0.00013606061,0.0108979875,0.0002972133,0.0003422681,0.7693293,0.18024412,0.0368754],"study_design_scores_gemma":[0.0001065181,0.00020013882,0.0007219404,0.0030425773,0.000040131235,0.00026531643,0.006103846,0.00097352196,0.0006723685,0.22652164,0.7611564,0.0001956412],"about_ca_topic_score_codex":0.0068872175,"about_ca_topic_score_gemma":0.008067686,"teacher_disagreement_score":0.12458691,"about_ca_system_score_codex":0.010448823,"about_ca_system_score_gemma":0.034678634,"threshold_uncertainty_score":0.65888643},"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":"W2846027767","doi":"10.1136/bmjebm-2018-110922","title":"Shared decision is the only outcome that matters when it comes to evaluating evidence-based practice","year":2018,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":60,"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":"Population; Argument (complex analysis); Psychological intervention; Medicine; Health care; Incidence (geometry); Psychology; Nursing; Political science","score_opus":0.7238961987099377,"score_gpt":0.5691387085368007,"score_spread":0.154757490173137,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2846027767","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":"methods","domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0063493787,0.026626607,0.091362655,0.8321643,0.015815262,0.0007290162,0.00043397077,0.0003430316,0.026175726],"genre_scores_gemma":[0.44984818,0.028082138,0.22868553,0.24475418,0.033964396,0.004478369,0.00085680344,0.0018420875,0.007488232],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.34320804,0.49565974,0.050772224,0.018632881,0.08565213,0.0060750167],"domain_scores_gemma":[0.24465239,0.5807871,0.033494674,0.0546001,0.068355806,0.018109985],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.47523955,0.0016430073,0.011075414,0.008351098,0.0063873646,0.036550883,0.0065645827,0.018345423,0.0117811],"category_scores_gemma":[0.7170177,0.0017478748,0.003949826,0.0062099993,0.041983213,0.04541993,0.023200922,0.029193766,0.0058615333],"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":[0.0012697587,0.0004263265,0.010841924,0.0060128896,0.0038962446,0.0006551504,0.021411262,0.002177198,0.0008512767,0.42581296,0.1401917,0.38645327],"study_design_scores_gemma":[0.0002915948,0.00041377626,0.0026203156,0.007627339,0.00046654206,0.0005136355,0.005346876,0.0019041316,0.00059098285,0.8787202,0.101163946,0.0003406677],"about_ca_topic_score_codex":0.003104718,"about_ca_topic_score_gemma":0.0019595798,"teacher_disagreement_score":0.47523955,"about_ca_system_score_codex":0.012535544,"about_ca_system_score_gemma":0.03134824,"threshold_uncertainty_score":0.6471231},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"theoretical_or_conceptual","genre":"empirical","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":"low"}],"label_agreement":"split"},{"id":"W2855200104","doi":"10.1136/bmjebm-2018-110972","title":"‘Depletion of the susceptibles’ taught through a story, a table and basic arithmetic","year":2018,"lang":"en","type":"letter","venue":"BMJ evidence-based medicine","topic":"Advanced Causal Inference Techniques","field":"Mathematics","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":"McGill University","funders":"","keywords":"Phrase; Arithmetic; Table (database); Psychology; Statistics; Computer science; Natural language processing; Mathematics; Data mining","score_opus":0.22595200651833938,"score_gpt":0.4243785418092606,"score_spread":0.19842653529092122,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2855200104","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00021560761,0.0013104803,0.0006908819,0.99177295,0.004358878,0.000014378495,0.000034766585,0.000022436165,0.0015796063],"genre_scores_gemma":[0.006086809,0.0016200861,0.00160091,0.9726662,0.014730388,0.00008563542,0.000024588158,0.000031763746,0.0031535844],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9935529,0.004275716,0.0006033744,0.00042348122,0.00093242,0.00021219312],"domain_scores_gemma":[0.96502995,0.029137924,0.0017381712,0.00096113863,0.0018709215,0.0012619144],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.009283495,0.0007538986,0.0009017023,0.00074467197,0.0019813727,0.0022671388,0.0010910289,0.018801171,0.010294027],"category_scores_gemma":[0.09994937,0.000534392,0.00082173775,0.0004782061,0.0050675827,0.0084072435,0.0019245197,0.030267986,0.0061458987],"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":[0.00008078515,0.000024719184,0.00042146223,0.00011643307,0.000014894664,0.0016043172,0.0006536948,0.000078949895,0.00017439792,0.016999668,0.96189994,0.01793068],"study_design_scores_gemma":[0.0001868019,0.000110789566,0.0009674099,0.0010774931,0.00002833089,0.009595878,0.001060339,0.0007929437,0.00027888108,0.090967774,0.8948592,0.000074096904],"about_ca_topic_score_codex":0.002305,"about_ca_topic_score_gemma":0.002476564,"teacher_disagreement_score":0.018801171,"about_ca_system_score_codex":0.002981444,"about_ca_system_score_gemma":0.0017723789,"threshold_uncertainty_score":0.049096406},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","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":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2884004244","doi":"10.1136/bmjebm-2018-110992","title":"Uncertainty surrounds anticoagulation risks and benefits in patients with chronic kidney disease with atrial fibrillation","year":2018,"lang":"en","type":"letter","venue":"BMJ evidence-based medicine","topic":"Atrial Fibrillation Management and Outcomes","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":"Ottawa Hospital; Institute for Clinical Evaluative Sciences; University of Ottawa","funders":"","keywords":"Medicine; Atrial fibrillation; Stroke (engine); Kidney disease; Observational study; Population; Internal medicine; Intensive care medicine; Cardiology; Disease; Randomized controlled trial; Heart failure","score_opus":0.104221354622322,"score_gpt":0.3482517148296416,"score_spread":0.2440303602073196,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2884004244","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00020866025,0.0037580037,0.00004302077,0.98132324,0.013835628,0.0000074184964,0.00012213858,0.0000113336355,0.0006905497],"genre_scores_gemma":[0.004041793,0.0056053814,0.00015295022,0.9178718,0.07118679,0.000042806474,0.00009402084,0.000018431372,0.0009860613],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.98763704,0.0054839025,0.0022291471,0.0008222235,0.003229918,0.0005977067],"domain_scores_gemma":[0.9071908,0.06966701,0.004340102,0.0011927024,0.013443104,0.004166307],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.011594492,0.00072308583,0.0030637824,0.001819639,0.0032412151,0.0036468238,0.002982259,0.038386863,0.010095075],"category_scores_gemma":[0.13927701,0.0008390796,0.0019204571,0.0020212836,0.0033904454,0.00639674,0.0020039522,0.03985406,0.008311592],"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":[0.00010773858,0.000020277488,0.00082502945,0.0004414978,0.00004976048,0.0005946821,0.00025547546,0.000039612514,0.000045962937,0.0013059826,0.9848125,0.011501518],"study_design_scores_gemma":[0.0008573208,0.0002210782,0.005145434,0.009939819,0.00030361922,0.005221149,0.001966105,0.00061565003,0.00029513674,0.025666483,0.94951093,0.00025727437],"about_ca_topic_score_codex":0.01253595,"about_ca_topic_score_gemma":0.014357712,"teacher_disagreement_score":0.038386863,"about_ca_system_score_codex":0.0052436525,"about_ca_system_score_gemma":0.006379129,"threshold_uncertainty_score":0.06131828},"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":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2885383088","doi":"10.1136/bmjebm-2018-111009","title":"In critically ill children, fluid overload is consistently associated with worse outcomes","year":2018,"lang":"en","type":"letter","venue":"BMJ evidence-based medicine","topic":"Sepsis Diagnosis and Treatment","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":"Sinai Health System; Toronto General Hospital; University Health Network","funders":"","keywords":"Critically ill; Intensive care medicine; Medicine; Critical illness","score_opus":0.12533883239647847,"score_gpt":0.38304514254165956,"score_spread":0.2577063101451811,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2885383088","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000590483,0.09651747,0.00021431732,0.84047633,0.05838242,0.000046105182,0.0014945418,0.000043194865,0.0022350969],"genre_scores_gemma":[0.015932258,0.0958192,0.0005371439,0.72512126,0.1580166,0.00019692234,0.0007083133,0.0000505413,0.0036177742],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9940422,0.0024401785,0.0010760048,0.00054660684,0.001646907,0.000248167],"domain_scores_gemma":[0.94636035,0.03719208,0.0047858,0.0005796673,0.010079867,0.0010021665],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0061487826,0.0010062874,0.0028553037,0.0023536582,0.00094037515,0.0018909087,0.002372667,0.011699447,0.013395068],"category_scores_gemma":[0.08911793,0.00063100917,0.0023897667,0.002636354,0.00090553105,0.0023712374,0.0008256835,0.009180706,0.00492107],"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":[0.00042004354,0.000015316531,0.0015705733,0.009979854,0.00067444437,0.00032134508,0.00011468473,0.00005467763,0.00008494667,0.00079015427,0.96268404,0.023289949],"study_design_scores_gemma":[0.0037132313,0.00046447536,0.017856106,0.11672832,0.009668237,0.0034308678,0.0009010222,0.0009167169,0.0007877334,0.013838285,0.8313521,0.00034279554],"about_ca_topic_score_codex":0.011794627,"about_ca_topic_score_gemma":0.013717777,"teacher_disagreement_score":0.013395068,"about_ca_system_score_codex":0.002299329,"about_ca_system_score_gemma":0.0039012455,"threshold_uncertainty_score":0.04481095},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","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":"split"},{"id":"W2887808077","doi":"10.1136/bmjebm-2018-110975","title":"Meta-analysis suggests that metformin may reduce pre-eclampsia compared with insulin use during pregnancy","year":2018,"lang":"en","type":"letter","venue":"BMJ evidence-based medicine","topic":"Pregnancy and preeclampsia studies","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 Toronto; Mount Sinai Hospital","funders":"","keywords":"Metformin; Medicine; Polycystic ovary; Eclampsia; Gestational diabetes; Pregnancy; Obstetrics; Preeclampsia; Type 2 diabetes; Diabetes mellitus; Meta-analysis; Insulin; Insulin resistance; Internal medicine; Endocrinology; Gestation; Biology","score_opus":0.2691340093160504,"score_gpt":0.3723894350958936,"score_spread":0.1032554257798432,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2887808077","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.002153231,0.2847563,0.0006837426,0.64866805,0.05462699,0.00013929556,0.002358048,0.00008797141,0.006526346],"genre_scores_gemma":[0.048580356,0.20530742,0.0010507983,0.6070077,0.13001712,0.00035882727,0.0010307517,0.00009836041,0.006548749],"study_design_codex":"not_applicable","study_design_gemma":"meta_analysis","domain_scores_codex":[0.993054,0.0034527942,0.0012358748,0.00059020094,0.0014397103,0.00022737407],"domain_scores_gemma":[0.9310405,0.057145048,0.0035277999,0.00063951395,0.0067791236,0.0008681193],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00948863,0.0011741907,0.0046266997,0.002214565,0.0008443568,0.002030303,0.0028291875,0.011615053,0.010333603],"category_scores_gemma":[0.1090777,0.0006748655,0.0049146432,0.0026370364,0.0008101684,0.0018657065,0.00079051655,0.009396478,0.002290053],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004665626,0.000056728753,0.004474516,0.03826969,0.015619846,0.0012447343,0.00021820208,0.000611497,0.00028723865,0.0030496619,0.87141573,0.06008645],"study_design_scores_gemma":[0.02072539,0.0016244628,0.025062919,0.13154207,0.12767191,0.0035771395,0.00062542077,0.0032107348,0.00135599,0.027135035,0.65696865,0.0005002521],"about_ca_topic_score_codex":0.011210502,"about_ca_topic_score_gemma":0.014939449,"teacher_disagreement_score":0.011615053,"about_ca_system_score_codex":0.002668558,"about_ca_system_score_gemma":0.0040233205,"threshold_uncertainty_score":0.05018127},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2887866346","doi":"10.1136/bmjebm-2018-111040","title":"More than 2 billion pairs of eyeballs: Why aren’t you sharing medical knowledge on Wikipedia?","year":2018,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Wikis in Education and Collaboration","field":"Social Sciences","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":"Queen's University","funders":"","keywords":"Reading (process); Knowledge translation; Internet privacy; World Wide Web; Public relations; Knowledge sharing; Encyclopedia; Resource (disambiguation); Information Dissemination; Computer science; Medical education; Political science; Library science; Knowledge management; Medicine; Law","score_opus":0.11614870997935471,"score_gpt":0.45045097746982166,"score_spread":0.33430226749046693,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2887866346","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.09620333,0.03273565,0.007737419,0.7757362,0.0089267045,0.00009599153,0.0022293502,0.0008511896,0.07548412],"genre_scores_gemma":[0.7074915,0.031508945,0.008191612,0.18023525,0.009839619,0.00021090753,0.0027031328,0.0011404739,0.058678586],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9942002,0.0030224645,0.0003011314,0.0006593342,0.0013841295,0.0004326397],"domain_scores_gemma":[0.97650105,0.011557912,0.002390532,0.0014558363,0.0032148773,0.004879835],"candidate_categories":["scholarly_communication","open_science"],"consensus_categories":[],"category_scores_codex":[0.006098276,0.0004545269,0.000451853,0.0020646576,0.007071405,0.00762608,0.0010710934,0.004033742,0.023021696],"category_scores_gemma":[0.036286265,0.0005023506,0.00055910175,0.0020737674,0.00427073,0.016181724,0.005278031,0.0037760914,0.005868815],"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":[0.00030513207,0.00009196724,0.022331217,0.00078472873,0.00016738876,0.000996945,0.022168346,0.00018119763,0.00057520164,0.023735506,0.7142459,0.2144164],"study_design_scores_gemma":[0.000040155865,0.000099486286,0.013827849,0.0015993562,0.00009735989,0.0016425903,0.03511271,0.000689816,0.00054563116,0.03589777,0.9103121,0.00013517714],"about_ca_topic_score_codex":0.010310659,"about_ca_topic_score_gemma":0.014107617,"teacher_disagreement_score":0.9989289,"about_ca_system_score_codex":0.002186321,"about_ca_system_score_gemma":0.0017832698,"threshold_uncertainty_score":0.07701528},"labels":[{"model":"gemma","categories":["scholarly_communication","open_science"],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["scholarly_communication"],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2899136888","doi":"10.1136/bmjebm-2018-111089","title":"Understanding Risk for Better Stroke Prevention","year":2018,"lang":"en","type":"letter","venue":"BMJ evidence-based medicine","topic":"Acute Ischemic Stroke Management","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":"Institute for Clinical Evaluative Sciences; University of Toronto; St. Michael's Hospital","funders":"","keywords":"Stroke (engine); Medicine; Physical medicine and rehabilitation; Engineering","score_opus":0.24907372318812274,"score_gpt":0.3807631036811311,"score_spread":0.13168938049300835,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2899136888","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000033472494,0.0027567632,0.00003254309,0.96999544,0.026035625,0.000007497926,0.00008136598,0.000017047934,0.0010402144],"genre_scores_gemma":[0.0007647603,0.0035176508,0.00010550884,0.9087889,0.08407494,0.00003651529,0.00007608414,0.000020073787,0.0026155452],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99294996,0.0026140928,0.0009619379,0.00064957904,0.002276528,0.00054798234],"domain_scores_gemma":[0.96686226,0.01861421,0.001288756,0.0007415031,0.009219508,0.0032737732],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0068171555,0.0016514496,0.002538839,0.0022604966,0.0033996692,0.0046707187,0.0037019765,0.043481253,0.031472437],"category_scores_gemma":[0.06637035,0.00095078076,0.0023349614,0.0015350977,0.004470737,0.009986634,0.0029060165,0.0556946,0.030059583],"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":[0.000015992298,0.00000866764,0.000056408593,0.00007146829,0.000006001895,0.00006858673,0.000037464764,0.000011687356,0.0000125628585,0.0009062946,0.99470156,0.0041033197],"study_design_scores_gemma":[0.00025641025,0.00006042213,0.0006608163,0.0024315305,0.000054146974,0.00059258466,0.0003808001,0.00016325447,0.00008240106,0.01357918,0.98166627,0.00007222671],"about_ca_topic_score_codex":0.012818993,"about_ca_topic_score_gemma":0.012730119,"teacher_disagreement_score":0.043481253,"about_ca_system_score_codex":0.006601144,"about_ca_system_score_gemma":0.008738963,"threshold_uncertainty_score":0.10528582},"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":"low"}],"label_agreement":"agree"},{"id":"W2901928126","doi":"10.1136/bmjebm-2018-111019","title":"Evidence-based practice education for healthcare professions: an expert view","year":2018,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":317,"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; Health care; Stakeholder; Medical education; Professional development; Stakeholder engagement; Best practice; Sociology; Pedagogy; Medicine; Psychology; Political science; Nursing; Public relations","score_opus":0.5722276379167234,"score_gpt":0.6578693978875588,"score_spread":0.08564175997083545,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2901928126","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0019069238,0.17729487,0.0044341153,0.8011119,0.006609945,0.00007525217,0.000042681037,0.00003252897,0.008491819],"genre_scores_gemma":[0.090184994,0.3708879,0.03738271,0.4784451,0.016488109,0.00043673336,0.0001741822,0.00011644473,0.005883782],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.8855371,0.056718376,0.021564446,0.004572665,0.028314816,0.0032925573],"domain_scores_gemma":[0.48284,0.38511276,0.011568191,0.008651355,0.08655125,0.025276395],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.16198976,0.00080353743,0.002838315,0.007954631,0.0038322376,0.016607828,0.0062784944,0.02541879,0.003977958],"category_scores_gemma":[0.22973968,0.0010393843,0.0014482733,0.0064860536,0.0136416145,0.018667834,0.011946776,0.027011506,0.0016147924],"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":[0.0001662476,0.0004859289,0.0019698674,0.029981578,0.00026088752,0.0018820085,0.04018292,0.0012318876,0.0019687198,0.08726583,0.25518048,0.5794237],"study_design_scores_gemma":[0.00006330119,0.00020810441,0.0022697027,0.134632,0.00016489092,0.0023617183,0.029763848,0.0011173008,0.0005239793,0.083672196,0.7450286,0.00019427233],"about_ca_topic_score_codex":0.0034303097,"about_ca_topic_score_gemma":0.004875663,"teacher_disagreement_score":0.83801025,"about_ca_system_score_codex":0.009604085,"about_ca_system_score_gemma":0.03320775,"threshold_uncertainty_score":0.856694},"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":"W2905106223","doi":"10.1136/bmjebm-2018-111121","title":"Drug discovery today: no molecules required","year":2018,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Chemical and Physical Studies","field":"Biochemistry, Genetics and Molecular Biology","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":"Drug discovery; Drug; Computer science; Medicine; Computational biology; Data science; Pharmacology; Bioinformatics; Biology","score_opus":0.03348223847339627,"score_gpt":0.318468529405376,"score_spread":0.28498629093197975,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2905106223","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0044541475,0.36140302,0.021404902,0.5080785,0.03846473,0.00016507771,0.00042792573,0.0007337711,0.06486786],"genre_scores_gemma":[0.09385333,0.42891073,0.020376548,0.33726552,0.04819556,0.00045769758,0.0008261594,0.00052445865,0.06958996],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99200773,0.0025258993,0.0004435155,0.000906136,0.0035253083,0.0005913926],"domain_scores_gemma":[0.9897031,0.0047067124,0.0006465887,0.0011253046,0.0019964236,0.0018218837],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.013456966,0.001213807,0.002231905,0.0011326553,0.0018477787,0.007759953,0.0028301226,0.0069985474,0.03149103],"category_scores_gemma":[0.014336376,0.00080032914,0.0007753352,0.0008252422,0.008868873,0.025542427,0.0053460747,0.015067221,0.013167098],"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":[0.00042507437,0.00033064868,0.000497856,0.0030910121,0.00013417464,0.00044649106,0.0006843288,0.0006314947,0.004342067,0.3974575,0.30534473,0.2866146],"study_design_scores_gemma":[0.00009680361,0.000319563,0.00014351384,0.0009206083,0.000028746628,0.0003139148,0.00031138444,0.00027667978,0.00091598596,0.117876716,0.87875485,0.00004125067],"about_ca_topic_score_codex":0.0004927983,"about_ca_topic_score_gemma":0.0008366327,"teacher_disagreement_score":0.03149103,"about_ca_system_score_codex":0.0028969848,"about_ca_system_score_gemma":0.0027209711,"threshold_uncertainty_score":0.10534805},"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":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2911416111","doi":"10.1136/bmjebm-2018-111069","title":"Causal inference for clinicians","year":2019,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Advanced Causal Inference Techniques","field":"Mathematics","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":"Jewish General Hospital","funders":"","keywords":"Inference; Causal inference; Psychology; Computer science; Artificial intelligence; Econometrics; Mathematics","score_opus":0.38310932339326553,"score_gpt":0.5368590741435255,"score_spread":0.15374975075025998,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2911416111","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0010245813,0.010517914,0.82330376,0.13148777,0.003511692,0.0005379261,0.00087280513,0.00060868025,0.028134916],"genre_scores_gemma":[0.10478613,0.016911425,0.82452714,0.029835826,0.009032702,0.0035353398,0.0011517666,0.000462189,0.009757541],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.92437714,0.061859425,0.0033987125,0.0044709714,0.005227085,0.00066661095],"domain_scores_gemma":[0.673441,0.28606725,0.008927881,0.01679979,0.012900952,0.0018630952],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.095994756,0.0018417476,0.0028760103,0.0059007835,0.0028388319,0.009466123,0.004762341,0.0089463685,0.031455826],"category_scores_gemma":[0.33025423,0.0015890806,0.0030731936,0.0043149274,0.012683518,0.012403453,0.006053475,0.016934175,0.0065989494],"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":[0.00008380202,0.000044910397,0.0007514376,0.0010613069,0.00016977623,0.00020297205,0.0010728516,0.003387652,0.00009821145,0.8656096,0.04249931,0.085018195],"study_design_scores_gemma":[0.000036898877,0.000011539558,0.000072514595,0.00049528794,0.000024368652,0.00008613234,0.00012014419,0.0034867888,0.00006548674,0.9647131,0.030870315,0.000017343258],"about_ca_topic_score_codex":0.0042017153,"about_ca_topic_score_gemma":0.002708673,"teacher_disagreement_score":0.095994756,"about_ca_system_score_codex":0.0052030734,"about_ca_system_score_gemma":0.011103861,"threshold_uncertainty_score":0.5076749},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2911968494","doi":"10.1136/bmjebm-2018-111057","title":"Use of modified intention-to-treat analysis in studies of direct oral anticoagulants and risk of selection bias: a systematic review","year":2019,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Atrial Fibrillation Management and Outcomes","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":"Centre Hospitalier de l’Université de Montréal","funders":"Fonds de Recherche du Québec - Santé","keywords":"Selection (genetic algorithm); Risk analysis (engineering); Selection bias; Medicine; Intensive care medicine; Psychology; Actuarial science; Computer science; Business; Artificial intelligence; Pathology","score_opus":0.5686784527537067,"score_gpt":0.5045638600117248,"score_spread":0.06411459274198184,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2911968494","genre_codex":"review","genre_gemma":"review","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0058078594,0.8283394,0.045791768,0.003719161,0.004755591,0.10579068,0.0033360445,0.0005808411,0.0018786954],"genre_scores_gemma":[0.15070623,0.21847288,0.15787537,0.008169284,0.0018476213,0.4591471,0.002182031,0.00038465252,0.0012148943],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.37677386,0.3963726,0.15953498,0.020718534,0.045105726,0.0014943678],"domain_scores_gemma":[0.3592418,0.51686525,0.081887424,0.020096533,0.020613734,0.0012952499],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.33654216,0.004568919,0.035400894,0.02557627,0.0022527387,0.00951536,0.007784355,0.010070801,0.0061234515],"category_scores_gemma":[0.5951705,0.00403466,0.036942676,0.019918203,0.007163558,0.010913661,0.0045663808,0.0046970947,0.0009691797],"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":[0.0009495462,0.000045350975,0.0018898385,0.8853156,0.07637611,0.00012962325,0.00071381713,0.0003528645,0.00014050826,0.0016847361,0.0019249099,0.03047708],"study_design_scores_gemma":[0.0077197775,0.0012123806,0.0041318196,0.7441987,0.20964512,0.00045483292,0.0007709154,0.003594737,0.000925353,0.011015952,0.015921347,0.00040907625],"about_ca_topic_score_codex":0.0032180732,"about_ca_topic_score_gemma":0.0062249578,"teacher_disagreement_score":0.96459913,"about_ca_system_score_codex":0.013065534,"about_ca_system_score_gemma":0.01951845,"threshold_uncertainty_score":0.8181616},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_narrow","metaepi_broad","metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2925846695","doi":"10.1136/bmjebm-2019-111174","title":"Assessing assumptions for statistical analyses in randomised clinical trials","year":2019,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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":"McMaster University; Impact","funders":"Novo Nordisk Fonden","keywords":"Statistical power; Statistical hypothesis testing; Clinical trial; Statistical analysis; Computer science; Statistical model; Econometrics; Medical physics; Management science; Data science; Medicine; Statistics; Machine learning; Mathematics; Engineering","score_opus":0.9827549661353883,"score_gpt":0.784218247525338,"score_spread":0.1985367186100503,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2925846695","genre_codex":"methods","genre_gemma":"methods","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004192232,0.011218509,0.94769186,0.015460548,0.0042282282,0.010965699,0.0011585417,0.00059905153,0.0044852966],"genre_scores_gemma":[0.096852556,0.0034928985,0.82260597,0.009046494,0.0020133003,0.06392762,0.0009970332,0.00043424455,0.00063002633],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.06053643,0.7837995,0.0916197,0.013930265,0.048149377,0.0019646648],"domain_scores_gemma":[0.020037211,0.90023804,0.031229852,0.034318577,0.013547658,0.0006287035],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.84948564,0.005363189,0.014069123,0.013285584,0.003933985,0.017276812,0.013983264,0.022882616,0.005406741],"category_scores_gemma":[0.9426373,0.0056980434,0.018713705,0.013893276,0.030021513,0.020107383,0.012464552,0.034401227,0.0024316981],"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":[0.009226309,0.00057431974,0.021239704,0.078515805,0.030215198,0.0034489993,0.020200463,0.039822575,0.0020628655,0.5360821,0.0257454,0.2328663],"study_design_scores_gemma":[0.002407596,0.0019607367,0.0047382563,0.02573226,0.0035993047,0.0013154132,0.0014935929,0.061990563,0.0029144515,0.85862356,0.034637127,0.00058719696],"about_ca_topic_score_codex":0.0024788985,"about_ca_topic_score_gemma":0.0012631393,"teacher_disagreement_score":0.15051436,"about_ca_system_score_codex":0.014736601,"about_ca_system_score_gemma":0.024045764,"threshold_uncertainty_score":0.18561107},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2935997213","doi":"10.1136/bmjebm-2019-111193","title":"Guidelines do not self-implement: time for a research paradigm shift from massive creation to effective implementation in evidence-based medicine research in China","year":2019,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Clinical practice guidelines implementation","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":"Ottawa Hospital; University of Ottawa","funders":"European Society of Cardiology","keywords":"Guideline; Context (archaeology); Quality (philosophy); China; Paradigm shift; Engineering ethics; Chen; Clinical Practice; Quality management; Computer science; Psychology; Medical education; Management science; Medicine; Political science; Nursing; Engineering; Epistemology; Operations management","score_opus":0.5299034806332391,"score_gpt":0.6399346695960708,"score_spread":0.1100311889628317,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2935997213","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.03023689,0.034472488,0.003458596,0.92459977,0.0033406732,0.0001727442,0.00013957561,0.0000651354,0.0035141986],"genre_scores_gemma":[0.64751273,0.03764548,0.02620235,0.27884817,0.0043464378,0.0010868576,0.00051428116,0.00017304455,0.003670591],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.90566444,0.056965504,0.008218203,0.005630551,0.015826099,0.007695265],"domain_scores_gemma":[0.72261506,0.17376056,0.019423245,0.017892662,0.033329964,0.032978527],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.1552188,0.00056038407,0.00182872,0.0030296752,0.0069095977,0.016828857,0.0047442317,0.0103001455,0.0068080663],"category_scores_gemma":[0.19221489,0.0009793446,0.0015890441,0.006679174,0.017183093,0.030877583,0.011531112,0.012780515,0.0006767361],"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":[0.000413015,0.0004021173,0.039742783,0.011337508,0.00039777183,0.0015767706,0.10936074,0.000773027,0.0014116251,0.2322283,0.1134707,0.48888558],"study_design_scores_gemma":[0.00046999345,0.00149883,0.112566784,0.022929454,0.00039264627,0.0014228228,0.09206832,0.003138312,0.0023871376,0.23840308,0.52408695,0.0006357313],"about_ca_topic_score_codex":0.037525166,"about_ca_topic_score_gemma":0.04372904,"teacher_disagreement_score":0.8447812,"about_ca_system_score_codex":0.024914142,"about_ca_system_score_gemma":0.120214425,"threshold_uncertainty_score":0.82088524},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2942870102","doi":"10.1136/bmjebm-2019-111186","title":"Training curriculum to help patient representatives participate meaningfully in the development of clinical practice guidelines","year":2019,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Mental Health and Patient Involvement","field":"Health Professions","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 Calgary","funders":"","keywords":"Guideline; Curriculum; Medical education; Grading (engineering); Likert scale; Medicine; Curriculum development; Face validity; Psychology; Scale (ratio); Readability; Pedagogy; Engineering; Computer science","score_opus":0.787443666098379,"score_gpt":0.6554227771389366,"score_spread":0.13202088895944242,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2942870102","genre_codex":"empirical","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4541788,0.0034919232,0.24811767,0.08225363,0.008625566,0.065392286,0.0018292113,0.008590483,0.12752041],"genre_scores_gemma":[0.2728816,0.0043227635,0.6172676,0.016205061,0.001945843,0.039243337,0.0021924549,0.00032060983,0.045620773],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9960311,0.0017333201,0.00036819687,0.00029048376,0.0008901196,0.0006867195],"domain_scores_gemma":[0.9889135,0.002700467,0.0011147748,0.0006815151,0.0021444836,0.004445313],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.007754942,0.0008886444,0.0006262216,0.0012490988,0.0011263363,0.0013028124,0.0017017445,0.0018053881,0.015701666],"category_scores_gemma":[0.017752,0.0005529844,0.0008573932,0.000788787,0.0005294666,0.0014168972,0.0041660327,0.002912383,0.0066794395],"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":[0.00033942965,0.020114254,0.008025156,0.0018556982,0.000033997567,0.00059550285,0.007370417,0.0022150318,0.011275189,0.0031671408,0.14052309,0.8044851],"study_design_scores_gemma":[0.0019525909,0.014022622,0.11687566,0.004040262,0.0001626706,0.0028499567,0.009430522,0.006422998,0.012307206,0.011749533,0.81988996,0.00029600642],"about_ca_topic_score_codex":0.000871215,"about_ca_topic_score_gemma":0.0021593075,"teacher_disagreement_score":0.015701666,"about_ca_system_score_codex":0.0016911847,"about_ca_system_score_gemma":0.009892088,"threshold_uncertainty_score":0.052527368},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"nonrandomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2951259638","doi":"10.1136/bmjebm-2019-111212","title":"Evidence from large Danish cohort does not support an association between the MMR vaccine and autism: facts in a post-truth world","year":2019,"lang":"en","type":"letter","venue":"BMJ evidence-based medicine","topic":"Vaccine Coverage and Hesitancy","field":"Social Sciences","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 Alberta","funders":"","keywords":"Danish; Autism; Association (psychology); MMR vaccine; Cohort; Cohort study; Medicine; Psychology; Virology; Psychiatry; Vaccination; Philosophy; Rubella; Internal medicine; Linguistics; Psychotherapist","score_opus":0.05852519767125543,"score_gpt":0.34461354752757245,"score_spread":0.28608834985631704,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2951259638","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.13874741,0.048606135,0.0010440452,0.7381701,0.013495979,0.00007687899,0.010797597,0.00007419401,0.048987683],"genre_scores_gemma":[0.5639256,0.035301186,0.0010310736,0.3723915,0.014025905,0.000062020925,0.0039665466,0.0000735426,0.009222572],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9963301,0.0012670466,0.000699346,0.00042862387,0.0010590923,0.00021584032],"domain_scores_gemma":[0.94022655,0.045068286,0.005824988,0.0015612844,0.0063854805,0.00093335664],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0071021696,0.00026812652,0.0007271043,0.0015936701,0.0011510849,0.0014526256,0.0009347533,0.0052833585,0.0055573382],"category_scores_gemma":[0.038661424,0.00035618639,0.00086928,0.0018559572,0.0012250055,0.0011717761,0.0006716637,0.004103965,0.0021254576],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0027010995,0.00010032773,0.45734245,0.0026402615,0.0025249198,0.008121689,0.001894734,0.00018216553,0.0012983917,0.003690784,0.45452031,0.06498296],"study_design_scores_gemma":[0.00091683306,0.00052362424,0.58634895,0.00945673,0.0039542215,0.0118387565,0.004977477,0.00097242167,0.001151069,0.0043874476,0.37522212,0.0002503317],"about_ca_topic_score_codex":0.035210934,"about_ca_topic_score_gemma":0.03872903,"teacher_disagreement_score":0.035210934,"about_ca_system_score_codex":0.0018581448,"about_ca_system_score_gemma":0.0013399726,"threshold_uncertainty_score":0.07001197},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2966910029","doi":"10.1136/bmjebm-2019-111236","title":"Treating postoperative pain? Avoid tramadol, long-acting opioid analgesics and long-term use","year":2019,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Opioid Use Disorder Treatment","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":"National Institute for Health and Care Research","keywords":"Tramadol; Medicine; Opioid; Anesthesia; Term (time); Pain ladder; Analgesic; Internal medicine","score_opus":0.058941608157173385,"score_gpt":0.3453892618272062,"score_spread":0.28644765367003283,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2966910029","genre_codex":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.49689186,0.11349705,0.0007450999,0.34870595,0.0045857877,0.00009800792,0.008672379,0.000024289404,0.026779551],"genre_scores_gemma":[0.9029263,0.04787383,0.0011940862,0.035520468,0.0034805064,0.00008914693,0.0024901968,0.000023154786,0.006402293],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.99952376,0.00013214312,0.000077077966,0.00005657739,0.00014008928,0.0000702849],"domain_scores_gemma":[0.9971902,0.00080018985,0.0011015367,0.00010373032,0.00043373022,0.00037068976],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0011432072,0.000072798226,0.00019802214,0.0003866063,0.00032895317,0.00067297247,0.00032647405,0.0009893422,0.012329676],"category_scores_gemma":[0.008350001,0.00009641836,0.0005061132,0.00086449005,0.0002186086,0.0006631302,0.00034569702,0.00093627744,0.0009820444],"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":[0.0008428446,0.00023777036,0.793047,0.0016445167,0.00055436743,0.00023949835,0.0004662769,0.000055828903,0.00027465692,0.0014608249,0.077848226,0.1233281],"study_design_scores_gemma":[0.00012602519,0.00022615954,0.9662542,0.0022091055,0.00035002845,0.00054863526,0.000665467,0.00008704181,0.000093334595,0.0011821953,0.028232783,0.000025135138],"about_ca_topic_score_codex":0.013272476,"about_ca_topic_score_gemma":0.02754212,"teacher_disagreement_score":0.013272476,"about_ca_system_score_codex":0.00046661095,"about_ca_system_score_gemma":0.00094111287,"threshold_uncertainty_score":0.04124683},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","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":"low"}],"label_agreement":"agree"},{"id":"W2969601771","doi":"10.1136/bmjebm-2019-111225","title":"Challenges in interpreting results from ‘multiple regression’ when there is interaction between covariates","year":2019,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Advanced Causal Inference Techniques","field":"Mathematics","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; McGill University; Jewish General Hospital","funders":"","keywords":"Covariate; Observational study; Outcome (game theory); Popularity; Meta-regression; Regression; Population; Econometrics; Regression analysis; Statistics; Regression toward the mean; Pace; Baseline (sea); Medicine; Psychology; Meta-analysis; Mathematics; Geography; Social psychology; Internal medicine","score_opus":0.4058680440234082,"score_gpt":0.4765542601293872,"score_spread":0.07068621610597897,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2969601771","genre_codex":"methods","genre_gemma":"methods","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0028281035,0.01695735,0.8986076,0.0663288,0.007105791,0.00063074636,0.0007471026,0.0010425302,0.0057520634],"genre_scores_gemma":[0.105742395,0.011529177,0.8250647,0.040016696,0.01016809,0.0025541082,0.00047361944,0.001975047,0.0024761935],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.36043748,0.54414386,0.03732034,0.01606362,0.040510494,0.001524228],"domain_scores_gemma":[0.11505887,0.81109583,0.02683652,0.032560404,0.013710051,0.0007382973],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.49097246,0.0030907656,0.0063886987,0.006976268,0.003003038,0.013150924,0.00822909,0.00825945,0.006006943],"category_scores_gemma":[0.841028,0.0028542439,0.008171708,0.0089480765,0.017116914,0.012329758,0.008648192,0.022466136,0.0034884585],"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":[0.0011958467,0.00019106221,0.02293351,0.015064785,0.010301799,0.0032803358,0.011724391,0.011844326,0.0018476401,0.34766552,0.09628245,0.4776684],"study_design_scores_gemma":[0.00030415264,0.00025254753,0.0059679537,0.005970404,0.0014396855,0.0024725795,0.0014419926,0.014666115,0.0017805602,0.8989168,0.066479206,0.00030799126],"about_ca_topic_score_codex":0.006414877,"about_ca_topic_score_gemma":0.0048768823,"teacher_disagreement_score":0.50902754,"about_ca_system_score_codex":0.0032212562,"about_ca_system_score_gemma":0.007521761,"threshold_uncertainty_score":0.6277216},"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":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2981608944","doi":"10.1136/bmjebm-2019-111273","title":"Challenges facing early-career and mid-career researchers: potential solutions to safeguard the future of evidence-based medicine","year":2019,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","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":"Hospital for Sick Children","funders":"National Institute for Health and Care Research","keywords":"Globe; Attendance; Career development; Career Pathways; General partnership; Public relations; Medical education; Psychology; Management; Political science; Medicine","score_opus":0.6353307993045753,"score_gpt":0.534381934536827,"score_spread":0.10094886476774823,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2981608944","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"incentives","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"incentives","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0013992966,0.0119113205,0.001657766,0.97727156,0.004598591,0.000060277634,0.000032595835,0.000062098006,0.0030063991],"genre_scores_gemma":[0.09182617,0.07008363,0.037089664,0.77202916,0.015694197,0.0009925077,0.0003436324,0.0003797481,0.011561351],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.7946424,0.13872819,0.011577987,0.0074003455,0.032308836,0.0153422095],"domain_scores_gemma":[0.5624126,0.28458166,0.02401196,0.01290429,0.044741046,0.07134843],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.16172606,0.0016203098,0.0022886058,0.0041429647,0.024219219,0.0413555,0.008286605,0.045468282,0.023282532],"category_scores_gemma":[0.24353342,0.0022148232,0.0029477838,0.0044326517,0.03133301,0.060338806,0.040263224,0.062191777,0.008484862],"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":[0.000273256,0.00047146998,0.0036820197,0.0072255987,0.00018141883,0.0018961675,0.06429774,0.00076785043,0.0010428608,0.12972745,0.59333766,0.19709654],"study_design_scores_gemma":[0.00010417092,0.000321132,0.0023661104,0.01677374,0.000057966856,0.0022428166,0.12369236,0.00080192136,0.00061427185,0.18992576,0.6627011,0.0003986712],"about_ca_topic_score_codex":0.0068107178,"about_ca_topic_score_gemma":0.012551448,"teacher_disagreement_score":0.83827394,"about_ca_system_score_codex":0.020347679,"about_ca_system_score_gemma":0.10995812,"threshold_uncertainty_score":0.85529935},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"incentives","study_design":"theoretical_or_conceptual","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"incentives","study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2990740094","doi":"10.1136/bmjebm-2019-111272","title":"Building capacity in evidence-based medicine in low-income and middle-income countries: problems and potential solutions","year":2019,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","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","funders":"National Institute for Health and Care Research","keywords":"Low and middle income countries; Income distribution; Middle income; Development economics; Economics; Demographic economics; Developing country; Economic growth; Mathematics; Inequality","score_opus":0.2712587169033339,"score_gpt":0.46077319020027663,"score_spread":0.1895144732969427,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2990740094","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00051701674,0.010102383,0.0005750309,0.98399156,0.0032866,0.0000202149,0.000014848258,0.000017104458,0.001475161],"genre_scores_gemma":[0.10944789,0.036519162,0.006144774,0.825151,0.019922493,0.00040749664,0.0000986569,0.0001416529,0.002166874],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.79315484,0.15100701,0.011396805,0.008822833,0.022395704,0.013222732],"domain_scores_gemma":[0.40804195,0.49366283,0.020378385,0.009427462,0.03886984,0.029619584],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.13246115,0.0014724882,0.0018810055,0.0043426543,0.019078882,0.027863907,0.008348583,0.03380034,0.0093130125],"category_scores_gemma":[0.29346046,0.0014349376,0.0020330332,0.0040301094,0.059017673,0.04183073,0.030235177,0.043447133,0.0017964058],"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":[0.00009242182,0.000099830366,0.0016706224,0.0073969406,0.00016494443,0.00163591,0.061869275,0.00088256784,0.00034817818,0.277007,0.57859004,0.0702424],"study_design_scores_gemma":[0.00008949813,0.00008724607,0.0013289534,0.028898878,0.00005528373,0.0006791483,0.07075049,0.0008126,0.0003422425,0.25565457,0.6410605,0.00024059489],"about_ca_topic_score_codex":0.020638937,"about_ca_topic_score_gemma":0.026704157,"teacher_disagreement_score":0.86753887,"about_ca_system_score_codex":0.03202544,"about_ca_system_score_gemma":0.12734948,"threshold_uncertainty_score":0.7005299},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"incentives","study_design":"theoretical_or_conceptual","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2993142824","doi":"10.1136/bmjebm-2019-111300","title":"The need for clinical judgement in the application of evidence-based medicine","year":2019,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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":"Robarts Clinical Trials; Western University","funders":"","keywords":"Judgement; Evidence-based medicine; Medicine; Intensive care medicine; MEDLINE; Alternative medicine; Psychotherapist; Psychology; Epistemology; Pathology","score_opus":0.9715138239990201,"score_gpt":0.7161295285443602,"score_spread":0.2553842954546599,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2993142824","genre_codex":"commentary","genre_gemma":"review","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0024687985,0.16422175,0.05487976,0.7172309,0.04826686,0.0029237366,0.0004113387,0.00038780356,0.009209027],"genre_scores_gemma":[0.18201195,0.076109454,0.28059825,0.39608234,0.050789185,0.011335941,0.0006336754,0.00062995765,0.0018092829],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.103400126,0.6724174,0.10820466,0.018234609,0.09465396,0.003089265],"domain_scores_gemma":[0.03314018,0.8522196,0.02608744,0.031826224,0.05154103,0.0051855845],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.75352436,0.003114172,0.021116506,0.023095518,0.008999371,0.03335258,0.018609792,0.042055685,0.0061771097],"category_scores_gemma":[0.9039241,0.005367352,0.008315869,0.012953792,0.060131803,0.03824427,0.022902086,0.052372765,0.0033996447],"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":[0.002497998,0.00040333503,0.0042613796,0.109489635,0.0094145695,0.0015991997,0.033706,0.0026847199,0.00094179506,0.20586187,0.295913,0.33322656],"study_design_scores_gemma":[0.0025375267,0.00047673917,0.0020902755,0.15319084,0.0021545547,0.0011958294,0.0063405596,0.0047380934,0.0005844908,0.6944494,0.13130471,0.0009369488],"about_ca_topic_score_codex":0.0063719032,"about_ca_topic_score_gemma":0.0066342554,"teacher_disagreement_score":0.24647564,"about_ca_system_score_codex":0.022163313,"about_ca_system_score_gemma":0.064863324,"threshold_uncertainty_score":0.3039484},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W3003086189","doi":"10.1136/bmjebm-2019-111268","title":"Assessment of assumptions of statistical analysis methods in randomised clinical trials: the what and how","year":2020,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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":"McMaster University; Impact","funders":"Novo Nordisk Fonden","keywords":"Statistical analysis; Clinical trial; Statistics; Medical physics; Medicine; Econometrics; Mathematics; Internal medicine","score_opus":0.9498731355574201,"score_gpt":0.7462174734380085,"score_spread":0.20365566211941166,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3003086189","genre_codex":"methods","genre_gemma":"methods","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.003341599,0.026534887,0.8254543,0.115404785,0.010679191,0.010517211,0.0010384283,0.0009276357,0.006101951],"genre_scores_gemma":[0.06334601,0.006671546,0.8661659,0.026150018,0.004106612,0.03177232,0.000527811,0.0006746474,0.000585066],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.032673262,0.84262574,0.080868,0.007779454,0.034844883,0.001208712],"domain_scores_gemma":[0.014167875,0.8884415,0.03144087,0.040197745,0.024886629,0.0008653263],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.9058305,0.005288937,0.013278829,0.01360856,0.006070852,0.029782407,0.015280397,0.022062121,0.0041856356],"category_scores_gemma":[0.9604863,0.0076146885,0.01707912,0.012568569,0.054401867,0.039124873,0.015874395,0.040071286,0.0033708713],"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":[0.0047874507,0.0004649649,0.012484965,0.07195243,0.01594942,0.0010646746,0.042538736,0.017336382,0.0015773642,0.39616135,0.054418348,0.38126397],"study_design_scores_gemma":[0.0021803868,0.0012260192,0.00340906,0.08189339,0.0027838373,0.00087740755,0.00316777,0.02417178,0.0032634484,0.8077349,0.067992,0.0012999561],"about_ca_topic_score_codex":0.0040468606,"about_ca_topic_score_gemma":0.0026954892,"teacher_disagreement_score":0.0941695,"about_ca_system_score_codex":0.019840768,"about_ca_system_score_gemma":0.042911623,"threshold_uncertainty_score":0.14395553},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"design_other","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3009190329","doi":"10.1136/bmjebm-2019-111312","title":"Ibuprofen for acute pain relief in the emergency department","year":2020,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Anesthesia and Pain Management","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 British Columbia","funders":"","keywords":"Ibuprofen; Medicine; Emergency department; Analgesic; Antipyretic; Acute pain; Adverse effect; Anesthesia; Medical prescription; Analgesic agents; Acetaminophen; Emergency medicine; Intensive care medicine; Internal medicine; Pharmacology","score_opus":0.10219635562718302,"score_gpt":0.37857842951306253,"score_spread":0.27638207388587954,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3009190329","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0015226996,0.9296248,0.00037625496,0.009798364,0.0028105597,0.000107236134,0.0004809326,0.000057753692,0.05522134],"genre_scores_gemma":[0.02906843,0.92794806,0.0019210931,0.015892003,0.005711717,0.00012975499,0.00095241773,0.000039793977,0.018336723],"study_design_codex":"design_other","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9988902,0.0003862733,0.00021071936,0.00007033721,0.00036368053,0.00007876906],"domain_scores_gemma":[0.9976082,0.0010745014,0.000651213,0.00007517181,0.00039367983,0.00019724252],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0011842219,0.00030777827,0.0007243103,0.0021551051,0.00037372223,0.0013215408,0.00044036546,0.0014894188,0.049062133],"category_scores_gemma":[0.005671021,0.00016806024,0.001133156,0.0015206113,0.00040326777,0.0009753364,0.0006710054,0.0017815739,0.016824072],"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":[0.0009823615,0.0002727956,0.0006093819,0.030040199,0.0004185932,0.0004986398,0.00010108532,0.000060218958,0.0011792178,0.003343329,0.14267932,0.81981486],"study_design_scores_gemma":[0.0008016217,0.00042978075,0.0046474193,0.037361864,0.00031707837,0.0022694534,0.00011762485,0.00004448623,0.000760105,0.0025868784,0.95063496,0.000028779721],"about_ca_topic_score_codex":0.00070651504,"about_ca_topic_score_gemma":0.002159997,"teacher_disagreement_score":0.049062133,"about_ca_system_score_codex":0.0005195242,"about_ca_system_score_gemma":0.0009810294,"threshold_uncertainty_score":0.1641292},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W3011157861","doi":"10.1136/bmjebm-2019-111331","title":"Adjuvant-containing control arms in pivotal quadrivalent human papillomavirus vaccine trials: restoration of previously unpublished methodology","year":2020,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Cervical Cancer and HPV Research","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":"U.S. Food and Drug Administration; National Institute for Health and Care Research; Laura and John Arnold Foundation; European Respiratory Society; American Association of Colleges of Pharmacy","keywords":"Clinical trial; Medicine; Blinding; Family medicine; Informed consent; Placebo; Alternative medicine; Internal medicine; Pathology","score_opus":0.5405509864717235,"score_gpt":0.5166450541185875,"score_spread":0.023905932353135984,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3011157861","genre_codex":"methods","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.057043824,0.05281457,0.6750666,0.029014098,0.01505289,0.12927125,0.021831019,0.0015505877,0.018355154],"genre_scores_gemma":[0.17658828,0.00624786,0.5938154,0.012016861,0.0037846123,0.1960841,0.007787171,0.0010174185,0.002658372],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.059374038,0.6800893,0.18979937,0.018314598,0.050819885,0.0016027077],"domain_scores_gemma":[0.041855313,0.5239198,0.14018793,0.24360722,0.048840944,0.0015887412],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.8006513,0.003158124,0.0066258293,0.028342767,0.003647125,0.016119942,0.009494014,0.0059537296,0.0052141766],"category_scores_gemma":[0.9055918,0.0037527278,0.009191978,0.028081542,0.014337198,0.013423758,0.016744735,0.00798778,0.0022661812],"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":[0.01111284,0.00078238006,0.048827264,0.119112454,0.012022664,0.0008064749,0.046479587,0.0032460652,0.0031069908,0.10005361,0.03498355,0.6194661],"study_design_scores_gemma":[0.01072576,0.005550502,0.052033406,0.14385034,0.017695291,0.001016863,0.0062343306,0.010048673,0.013524662,0.1887992,0.54914045,0.0013805453],"about_ca_topic_score_codex":0.00166671,"about_ca_topic_score_gemma":0.0025264348,"teacher_disagreement_score":0.19934869,"about_ca_system_score_codex":0.0064255297,"about_ca_system_score_gemma":0.02846826,"threshold_uncertainty_score":0.2458325},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3027820216","doi":"10.1136/bmjebm-2020-111371","title":"Problem with patient decision aids","year":2020,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Patient-Provider Communication in Healthcare","field":"Health Professions","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":"Université Laval","funders":"","keywords":"Decision aids; Certainty; Presentation (obstetrics); Health care; Decision support system; Medicine; Decision quality; Decision analysis; Value (mathematics); Quality (philosophy); Medical emergency; Patient satisfaction; Nursing; Computer science; Alternative medicine; Artificial intelligence; Surgery","score_opus":0.38599478060132236,"score_gpt":0.47030998431235405,"score_spread":0.08431520371103168,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3027820216","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0066297194,0.02392703,0.1254089,0.7396234,0.011249915,0.0013117515,0.0009173514,0.0014327139,0.089499295],"genre_scores_gemma":[0.19670601,0.023518696,0.2433227,0.49158478,0.011512895,0.0044873627,0.0010139795,0.001265872,0.026587797],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.6600566,0.2660839,0.03004417,0.0075498885,0.033088878,0.0031765199],"domain_scores_gemma":[0.32507345,0.59602606,0.01412093,0.020820785,0.037902087,0.0060566207],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.18205465,0.0020134605,0.0017796592,0.0041551823,0.0055978876,0.013451522,0.0076462864,0.016554225,0.029399361],"category_scores_gemma":[0.523292,0.0017638819,0.004127203,0.0035835141,0.014553643,0.028029663,0.01204998,0.028991152,0.013670987],"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":[0.0008012339,0.00032929663,0.0055648196,0.004379717,0.000342493,0.0028936788,0.019181086,0.0017528518,0.00039112457,0.19959992,0.369339,0.39542478],"study_design_scores_gemma":[0.00032218144,0.00025260856,0.0006615887,0.008562765,0.00020886875,0.009438505,0.007792506,0.002649122,0.0009458554,0.16607133,0.8028195,0.00027520422],"about_ca_topic_score_codex":0.0025505207,"about_ca_topic_score_gemma":0.002363794,"teacher_disagreement_score":0.18205465,"about_ca_system_score_codex":0.007225909,"about_ca_system_score_gemma":0.014243654,"threshold_uncertainty_score":0.96280855},"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":"low"}],"label_agreement":"agree"},{"id":"W3028006966","doi":"10.1136/bmjebm-2020-111386","title":"Thrombolysis with alteplase 3–4.5 hours after acute ischaemic stroke: trial reanalysis adjusted for baseline imbalances","year":2020,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Acute Ischemic Stroke Management","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; St. Joseph’s Healthcare Hamilton; Impact","funders":"","keywords":"Medicine; Thrombolysis; Stroke (engine); Modified Rankin Scale; Ischaemic stroke; Clinical trial; Placebo; Internal medicine; Physical therapy; Randomized controlled trial; Tissue plasminogen activator; Ischemic stroke; Ischemia; Myocardial infarction","score_opus":0.06341808429054929,"score_gpt":0.33722726806590375,"score_spread":0.27380918377535446,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3028006966","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6330467,0.11599363,0.10162653,0.003918781,0.008741502,0.046283826,0.07557968,0.0046907323,0.010118622],"genre_scores_gemma":[0.88113713,0.0051739584,0.046963573,0.0011327297,0.00080835505,0.034388714,0.02343018,0.0006387094,0.0063266917],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.97952646,0.01098802,0.0034639402,0.0029243573,0.0022742252,0.0008230519],"domain_scores_gemma":[0.9848585,0.0054899696,0.0039569573,0.0036319476,0.0018661213,0.0001964536],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.027835507,0.0021324635,0.0053245677,0.0016737754,0.0004210851,0.002261167,0.0018819404,0.00092295813,0.0068542967],"category_scores_gemma":[0.047312755,0.0008693616,0.016488075,0.0024384789,0.00047074945,0.0014145559,0.0012922051,0.0017872968,0.0010208432],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.35749197,0.000587022,0.042040534,0.026016211,0.3868874,0.000637534,0.0005266301,0.0051609655,0.004378553,0.0018079499,0.032741472,0.1417238],"study_design_scores_gemma":[0.23907965,0.020335272,0.16427149,0.0037563352,0.48481497,0.00086826994,0.00029797634,0.02085793,0.004906149,0.0084965145,0.05171667,0.0005988267],"about_ca_topic_score_codex":0.0025438876,"about_ca_topic_score_gemma":0.003438388,"teacher_disagreement_score":0.9721645,"about_ca_system_score_codex":0.001022134,"about_ca_system_score_gemma":0.002519591,"threshold_uncertainty_score":0.14720994},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3032356396","doi":"10.1136/bmjebm-2019-111326","title":"Evaluating the quality of systematic reviews and meta-analyses published on direct oral anticoagulants in the past 5 years","year":2020,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Venous Thromboembolism Diagnosis and Management","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 Ottawa; Public Health Ontario; University of Toronto; McMaster University","funders":"","keywords":"Systematic review; Checklist; Medicine; Data extraction; MEDLINE; Guideline; Meta-analysis; Quality (philosophy); Quality Score; Medical physics; Family medicine; Internal medicine; Pathology; Psychology","score_opus":0.7738296929297955,"score_gpt":0.5664963431092327,"score_spread":0.20733334982056284,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3032356396","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008509415,0.9672453,0.0075882305,0.003426699,0.0014208481,0.0047288444,0.0060615125,0.00013516947,0.00088399404],"genre_scores_gemma":[0.24458165,0.6457055,0.053174436,0.0057323882,0.0020777169,0.039430175,0.008373919,0.00024365366,0.0006806983],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.5075968,0.16642147,0.2642822,0.016138038,0.043651488,0.0019100404],"domain_scores_gemma":[0.24008578,0.53163373,0.1621177,0.014466413,0.049425747,0.0022706906],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.39873445,0.0036945553,0.025044065,0.049204286,0.0019883958,0.010292356,0.0056075924,0.004806061,0.003785017],"category_scores_gemma":[0.6656754,0.0039019294,0.03645596,0.03774399,0.0041229012,0.008385087,0.006643271,0.0033518362,0.00045980923],"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":[0.0006587032,0.000027362807,0.011100797,0.79267585,0.16717428,0.00022011044,0.0011515872,0.00060010253,0.00023069071,0.0006547856,0.0020384153,0.023467287],"study_design_scores_gemma":[0.0011740915,0.00061341276,0.024171915,0.4322419,0.51690197,0.0006629032,0.00093973556,0.0008564483,0.0006659415,0.002731532,0.018806554,0.0002336968],"about_ca_topic_score_codex":0.009241579,"about_ca_topic_score_gemma":0.022583868,"teacher_disagreement_score":0.9749559,"about_ca_system_score_codex":0.010880295,"about_ca_system_score_gemma":0.029267384,"threshold_uncertainty_score":0.7414675},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W3039340826","doi":"10.1136/bmjebm-2020-111412","title":"Dietary Recommendations for Familial Hypercholesterolaemia: an Evidence-Free Zone","year":2020,"lang":"en","type":"letter","venue":"BMJ evidence-based medicine","topic":"Diet and metabolism studies","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":"Canadian VIGOUR Centre","funders":"","keywords":"Medicine; Insulin resistance; Saturated fat; Internal medicine; Cholesterol; Obesity; Endocrinology; Coronary heart disease; Low carbohydrate; Cardiology; Weight loss","score_opus":0.24919084720613657,"score_gpt":0.4015866636004412,"score_spread":0.15239581639430463,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3039340826","genre_codex":"review","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.002924484,0.97291857,0.0009763775,0.016611775,0.0017610716,0.0010563583,0.00056872354,0.00008078047,0.003101859],"genre_scores_gemma":[0.02451812,0.9495558,0.008949392,0.012505564,0.00083595054,0.0015526466,0.000974211,0.00002429116,0.0010840353],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99323624,0.0023855984,0.0020024236,0.00033313007,0.0018823133,0.00016023804],"domain_scores_gemma":[0.9853837,0.0076235905,0.0032002132,0.00028332006,0.002967149,0.00054195867],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.011334249,0.0009213243,0.0030275974,0.0041749105,0.00073942216,0.0026095333,0.0031458894,0.004258664,0.0060215434],"category_scores_gemma":[0.030530283,0.0007384556,0.0038365198,0.0028506983,0.0006612133,0.0025165263,0.0016620252,0.0037959414,0.0015232177],"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":[0.0013335196,0.0005063147,0.0014810923,0.19247244,0.0033585879,0.0001314027,0.0002530526,0.00025298938,0.0003950368,0.0013705085,0.028345523,0.7700995],"study_design_scores_gemma":[0.0030863509,0.00090444117,0.00984116,0.77264637,0.011713679,0.0002582115,0.00033511175,0.00034214972,0.00034494078,0.0017594799,0.19868532,0.000082702776],"about_ca_topic_score_codex":0.0053273393,"about_ca_topic_score_gemma":0.011577516,"teacher_disagreement_score":0.011334249,"about_ca_system_score_codex":0.0022059577,"about_ca_system_score_gemma":0.0045268415,"threshold_uncertainty_score":0.059942007},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"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":"W3045810088","doi":"10.1136/bmjebm-2020-111339","title":"Partnering with patients in the production of evidence","year":2020,"lang":"en","type":"letter","venue":"BMJ evidence-based medicine","topic":"Mental Health and Patient Involvement","field":"Health Professions","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":"Institute for Clinical Evaluative Sciences; SickKids Foundation; Hospital for Sick Children; University of Toronto","funders":"","keywords":"General partnership; Manifesto; Evidence-based medicine; Healthcare delivery; Evidence-based practice; Health care; Production (economics); Trustworthiness; Public relations; Medicine; Knowledge management; Engineering ethics; Psychology; Business; Political science; Alternative medicine; Computer science; Engineering; Social psychology","score_opus":0.6388868732396912,"score_gpt":0.5055099970302084,"score_spread":0.1333768762094828,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3045810088","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.010566393,0.016963424,0.04423185,0.88696516,0.006328301,0.0016663449,0.00019265352,0.0006301498,0.03245571],"genre_scores_gemma":[0.29761526,0.04479487,0.16075033,0.4575497,0.015938655,0.0062586754,0.0008697895,0.0011150291,0.015107796],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.393326,0.53275144,0.022697702,0.011578007,0.02950568,0.010141204],"domain_scores_gemma":[0.37890115,0.3862918,0.0381199,0.039107632,0.050539013,0.1070405],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.33091575,0.0016665688,0.0031819844,0.005734693,0.015823886,0.03941189,0.0080612255,0.018328132,0.02716767],"category_scores_gemma":[0.5051556,0.0024606215,0.004300668,0.0038375587,0.015887452,0.053313468,0.07385454,0.041715343,0.012745913],"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":[0.00049742503,0.0011506377,0.012687581,0.0053198333,0.0007371667,0.0053554173,0.17353813,0.001182427,0.0018917731,0.07835127,0.3139654,0.40532288],"study_design_scores_gemma":[0.00035662335,0.00073419954,0.0019559825,0.016839866,0.00025707606,0.004901131,0.10197552,0.0016957574,0.0016235517,0.17933816,0.68981975,0.0005024018],"about_ca_topic_score_codex":0.0015170253,"about_ca_topic_score_gemma":0.0026058992,"teacher_disagreement_score":0.66908425,"about_ca_system_score_codex":0.01006107,"about_ca_system_score_gemma":0.0646259,"threshold_uncertainty_score":0.8251},"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":"W3045872399","doi":"10.1136/bmjebm-2020-111399","title":"Development of five online modules for teaching evidence-informed healthcare: the West coast Interprofessional Clinical Knowledge Evidence Disseminator (WICKED) Project","year":2020,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","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":"University of British Columbia","funders":"University of British Columbia","keywords":"Psychological intervention; Interactivity; Evidence-based practice; Critical appraisal; Medical education; Health care; Psychology; Evidence-based medicine; Nursing; Knowledge management; Medicine; Computer science; Alternative medicine; Multimedia","score_opus":0.6310685343411546,"score_gpt":0.6791755430567533,"score_spread":0.0481070087155987,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3045872399","genre_codex":"empirical","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.48722413,0.0019940222,0.25450227,0.024986561,0.0033246523,0.13539246,0.010011133,0.009713862,0.072850846],"genre_scores_gemma":[0.101428166,0.001366125,0.81530696,0.00226718,0.0002350446,0.05576887,0.0059051793,0.000705951,0.017016454],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99166507,0.005022138,0.00046529245,0.00046869362,0.0016022497,0.00077646977],"domain_scores_gemma":[0.98274106,0.0035445709,0.00086205883,0.0011747581,0.002494326,0.009183279],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.021570755,0.0011158213,0.0008706295,0.0015884255,0.0013770597,0.002702019,0.0037312293,0.0009920752,0.018931502],"category_scores_gemma":[0.025626462,0.0007934062,0.0011327735,0.00079430983,0.001199401,0.0019935889,0.010798907,0.0030792612,0.005704909],"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":[0.0007984232,0.019004507,0.007089349,0.0012785612,0.00009604456,0.00063381135,0.005599033,0.0017527633,0.0022129954,0.0025757519,0.06649348,0.89246523],"study_design_scores_gemma":[0.0071975254,0.029968726,0.081591874,0.007090775,0.00044482757,0.0019295212,0.0153644765,0.016067341,0.01812102,0.031341463,0.7903964,0.0004860431],"about_ca_topic_score_codex":0.0016747453,"about_ca_topic_score_gemma":0.004408137,"teacher_disagreement_score":0.021570755,"about_ca_system_score_codex":0.0028133756,"about_ca_system_score_gemma":0.02049287,"threshold_uncertainty_score":0.1140784},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W3048641009","doi":"10.1136/bmjebm-2020-111517","title":"Thrombolysis with alteplase 3–4.5 hours after acute ischaemic stroke: trial reanalysis adjusted for baseline imbalances","year":2020,"lang":"en","type":"letter","venue":"BMJ evidence-based medicine","topic":"Acute Ischemic Stroke Management","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":"Thrombolysis; Medicine; Ischaemic stroke; Stroke (engine); Baseline (sea); Acute stroke; Clinical trial; Internal medicine; Cardiology; Emergency medicine; Tissue plasminogen activator; Myocardial infarction; Ischemia","score_opus":0.05651902862013785,"score_gpt":0.3302200308703915,"score_spread":0.2737010022502536,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3048641009","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.012718393,0.020754194,0.0042408267,0.8497366,0.088650696,0.001153051,0.01224468,0.0004605712,0.010040928],"genre_scores_gemma":[0.10606116,0.009142596,0.014844567,0.74328226,0.09799863,0.0052653416,0.006973286,0.00080625515,0.015625827],"study_design_codex":"not_applicable","study_design_gemma":"meta_analysis","domain_scores_codex":[0.95204914,0.030292606,0.008438078,0.001587524,0.0066191168,0.0010134648],"domain_scores_gemma":[0.8282306,0.109123014,0.01658846,0.011390972,0.03212504,0.0025418538],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.04640432,0.0008591138,0.0026708345,0.0030868864,0.0011356971,0.0030856996,0.0016652246,0.006628036,0.004473314],"category_scores_gemma":[0.2613973,0.0007600316,0.003649522,0.004550734,0.0010930992,0.0021538404,0.0010198263,0.007438438,0.0065173963],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.005426001,0.0000786755,0.0153972255,0.0020315084,0.0023368353,0.00069923507,0.00025730854,0.00036586614,0.0003345924,0.001537842,0.8584273,0.11310761],"study_design_scores_gemma":[0.016112447,0.0013148624,0.08290674,0.013323766,0.013440232,0.002993829,0.0005041767,0.010983739,0.0027050602,0.037147027,0.8177976,0.0007704739],"about_ca_topic_score_codex":0.0072164154,"about_ca_topic_score_gemma":0.013525874,"teacher_disagreement_score":0.9535957,"about_ca_system_score_codex":0.0028289305,"about_ca_system_score_gemma":0.0044575687,"threshold_uncertainty_score":0.24541241},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W3048654576","doi":"10.1136/bmjebm-2020-111405","title":"Preventing hospital readmission through better medication continuity after hospital discharge","year":2020,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Pharmaceutical Practices and Patient Outcomes","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; University of British Columbia Hospital","funders":"","keywords":"Medicine; Discontinuation; Transitional care; Psychological intervention; Confusion; Hospital discharge; Health care; Hospital readmission; Hospital care; Medication Reconciliation; Patient discharge; Emergency medicine; Intensive care medicine; Medical emergency; MEDLINE; Family medicine; Nursing; Psychiatry; Pharmacy; Psychology","score_opus":0.1439962306500625,"score_gpt":0.4227465836670851,"score_spread":0.2787503530170226,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3048654576","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.70594656,0.061614968,0.007366669,0.12920523,0.0026256985,0.0012014814,0.0019475964,0.00081793417,0.08927388],"genre_scores_gemma":[0.9422018,0.019321581,0.014753475,0.011444034,0.0030396737,0.0003964128,0.0015164235,0.0000885852,0.007238048],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9974868,0.0009425528,0.00039175065,0.00016416659,0.00063113234,0.0003834177],"domain_scores_gemma":[0.9892753,0.0023880978,0.004613663,0.000510646,0.0012191639,0.0019931174],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003937628,0.00023752972,0.00082672195,0.0010395879,0.000622606,0.002185592,0.0012030458,0.0016807849,0.012847094],"category_scores_gemma":[0.03512779,0.00027558295,0.0013003882,0.00081950246,0.00022527433,0.0018106798,0.0018618478,0.0017053483,0.0016770501],"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":[0.0012516973,0.0077445935,0.13874431,0.0033604247,0.0011244683,0.0003461663,0.0013807797,0.0008919444,0.0005440873,0.0023783688,0.044338096,0.79789513],"study_design_scores_gemma":[0.002038089,0.003982845,0.91289884,0.009837132,0.0011195863,0.00085472467,0.001393856,0.0015004004,0.0006857218,0.0032450648,0.06234941,0.00009442911],"about_ca_topic_score_codex":0.0056884144,"about_ca_topic_score_gemma":0.009139976,"teacher_disagreement_score":0.012847094,"about_ca_system_score_codex":0.0011538129,"about_ca_system_score_gemma":0.002753373,"threshold_uncertainty_score":0.04297781},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W3082235742","doi":"10.1136/bmjebm-2020-111362","title":"Conventional compared to network meta-analysis to evaluate antibiotic prophylaxis in patients with cancer and haematopoietic stem cell transplantation recipients","year":2020,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Neutropenia and Cancer Infections","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":"SickKids Foundation; Hospital for Sick Children; Pediatric Oncology Group; Toronto General Hospital; University of Toronto","funders":"National Institute for Health and Care Research; Pediatric Oncology Group of Ontario","keywords":"Meta-analysis; Medicine; Internal medicine; Placebo; Transplantation; Trimethoprim; Relative risk; Hematopoietic stem cell transplantation; Antibiotics; Oncology; Surgery; Confidence interval; Biology; Microbiology; Pathology","score_opus":0.14006914249045047,"score_gpt":0.3592804178242128,"score_spread":0.21921127533376236,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3082235742","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.054601964,0.9054106,0.028787319,0.001773389,0.0015856319,0.0014691335,0.0030447077,0.00027874406,0.0030486274],"genre_scores_gemma":[0.86798155,0.10656599,0.016603742,0.002133218,0.00083908136,0.0022998012,0.0020926474,0.000113476126,0.0013704402],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.969652,0.024139911,0.0023097214,0.0022696054,0.0012914359,0.00033731523],"domain_scores_gemma":[0.96607935,0.029233174,0.0020435723,0.001433323,0.00095284503,0.00025767589],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.02956597,0.003514493,0.010740607,0.0042545144,0.0004183388,0.0028719802,0.0020979194,0.0021569622,0.0035864573],"category_scores_gemma":[0.056219593,0.0012142388,0.031991262,0.0036708764,0.0005752027,0.0023726772,0.0017430246,0.0023769587,0.0003313557],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.007952663,0.0000393059,0.006301546,0.03959427,0.92974246,0.00012489758,0.00003589025,0.006462781,0.00026757445,0.000668244,0.0005914053,0.00821886],"study_design_scores_gemma":[0.0031379098,0.0006302089,0.0038586347,0.0025180725,0.9809155,0.00012939083,0.000024228377,0.004504544,0.00020733246,0.0022024661,0.0018410764,0.0000305935],"about_ca_topic_score_codex":0.004535854,"about_ca_topic_score_gemma":0.0057271663,"teacher_disagreement_score":0.02956597,"about_ca_system_score_codex":0.0021101094,"about_ca_system_score_gemma":0.0012082827,"threshold_uncertainty_score":0.1563617},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"meta_analysis","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3090996960","doi":"10.1136/bmjebm-2020-111499","title":"Quality of early evidence on the pathogenesis, diagnosis, prognosis and treatment of COVID-19","year":2020,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"COVID-19 Clinical Research Studies","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":true,"ca_institutions":"University of Ottawa; McMaster University","funders":"","keywords":"Checklist; MEDLINE; Critical appraisal; Medicine; Rigour; Systematic review; Consolidated Standards of Reporting Trials; Declaration; Family medicine; Publication bias; Clinical study design; Data extraction; Clinical trial; Alternative medicine; Psychology; Meta-analysis; Pathology; Political science","score_opus":0.5759781474648025,"score_gpt":0.5831047699438358,"score_spread":0.007126622479033284,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3090996960","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0016156193,0.98852766,0.0006295204,0.0052855653,0.0011814793,0.0003868308,0.001304896,0.0000148295585,0.0010536055],"genre_scores_gemma":[0.056280125,0.92323244,0.0058774026,0.0073303166,0.0016605865,0.0023219853,0.0027438847,0.00006568297,0.0004876236],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.8327247,0.05702997,0.07975136,0.005266262,0.023084277,0.0021434578],"domain_scores_gemma":[0.33409613,0.507836,0.08725915,0.013140859,0.05482434,0.0028435497],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.14162813,0.0015020488,0.010531313,0.025126794,0.0013338063,0.012187891,0.0035337328,0.0053005423,0.010072108],"category_scores_gemma":[0.4954859,0.002120092,0.01059217,0.019831993,0.0027985133,0.007833429,0.004405054,0.0036860558,0.0012992907],"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":[0.00068425917,0.000042430314,0.004677441,0.8922037,0.008430389,0.00016989956,0.00062253064,0.00013370079,0.000266281,0.001033577,0.005765162,0.08597068],"study_design_scores_gemma":[0.00025447083,0.00021401979,0.008807097,0.9299307,0.019041076,0.0002805671,0.0004925877,0.000086952765,0.0002568475,0.0013836777,0.03920227,0.000049756607],"about_ca_topic_score_codex":0.0072557917,"about_ca_topic_score_gemma":0.0141042555,"teacher_disagreement_score":0.85837185,"about_ca_system_score_codex":0.009691318,"about_ca_system_score_gemma":0.023830103,"threshold_uncertainty_score":0.7490101},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W3094203153","doi":"10.1136/bmjebm-2020-111452","title":"Optimising the process for conducting scoping reviews","year":2020,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","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":"Izaak Walton Killam Health Centre; BC Children's Hospital; University of Toronto; Dalhousie University; York University; University of British Columbia","funders":"Canadian Institutes of Health Research","keywords":"Process (computing); Systematic review; Resource (disambiguation); Knowledge management; Process management; Best practice; Key (lock); Computer science; Management science; Data science; Engineering; MEDLINE; Political science","score_opus":0.895103126102174,"score_gpt":0.7197301643337385,"score_spread":0.17537296176843553,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3094203153","genre_codex":"methods","genre_gemma":"methods","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0062505063,0.0139287375,0.66766816,0.11127601,0.01151051,0.1630308,0.0019610208,0.009800607,0.014573674],"genre_scores_gemma":[0.008316619,0.0041361414,0.90249765,0.0049083172,0.00093642884,0.0758557,0.0005979041,0.0007746096,0.0019766877],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.1258001,0.6962239,0.11128879,0.01176149,0.050574787,0.0043509547],"domain_scores_gemma":[0.068693176,0.5828554,0.046897877,0.07584948,0.21591786,0.009786132],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.82123923,0.0058590528,0.0074828085,0.027808784,0.010625598,0.032847375,0.011758534,0.019298667,0.023356589],"category_scores_gemma":[0.8806435,0.01015258,0.011412302,0.025384791,0.012165252,0.038679518,0.027176954,0.026919348,0.034459062],"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":[0.0013726739,0.0006001614,0.0027419622,0.047488596,0.0008933267,0.0013117647,0.07888341,0.0029531121,0.004483044,0.028103862,0.12300587,0.70816225],"study_design_scores_gemma":[0.0036074698,0.0012030159,0.0054341457,0.1814388,0.0016855724,0.0013935027,0.027296862,0.007669007,0.0069986507,0.12761028,0.633782,0.001880687],"about_ca_topic_score_codex":0.008166284,"about_ca_topic_score_gemma":0.017948475,"teacher_disagreement_score":0.17876077,"about_ca_system_score_codex":0.022359844,"about_ca_system_score_gemma":0.14455946,"threshold_uncertainty_score":0.2204439},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3098080120","doi":"10.1136/bmjebm-2020-111503","title":"Ten years later: a review of the US 2009 institute of medicine report on conflicts of interest and solutions for further reform","year":2020,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Pharmaceutical industry and healthcare","field":"Pharmacology, Toxicology and Pharmaceutics","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":"University of Toronto; St. Michael's Hospital","funders":"","keywords":"Conflict of interest; Guideline; Alternative medicine; Political science; Health care; Medical research; Public relations; Medical education; Engineering ethics; Medicine; Engineering; Law","score_opus":0.7738178979150093,"score_gpt":0.6174324461110864,"score_spread":0.15638545180392294,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3098080120","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00009244071,0.9857875,0.000073169234,0.010409229,0.0028499183,0.000015529324,0.00009592084,0.000004384083,0.0006718979],"genre_scores_gemma":[0.0017244144,0.9800166,0.00038237817,0.014766053,0.0019234812,0.000042391614,0.00017424059,0.000008059223,0.00096241635],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9940719,0.0020349708,0.0014435278,0.00037018186,0.0018729827,0.00020657782],"domain_scores_gemma":[0.98021406,0.010120374,0.0031408495,0.00036257008,0.0056683733,0.00049381645],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.009856354,0.00081975554,0.001357236,0.008476085,0.0005471519,0.0024276553,0.00083594216,0.0027951342,0.005748082],"category_scores_gemma":[0.03670816,0.0005141713,0.0022340138,0.008702375,0.0008279723,0.0036842104,0.001334983,0.0034481713,0.0014616256],"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":[0.00013187322,0.000037585516,0.00054989324,0.079928294,0.0006300008,0.00025328,0.000606795,0.00014190185,0.00043732117,0.00906571,0.36464205,0.5435753],"study_design_scores_gemma":[0.000023032082,0.000034554247,0.0018239366,0.07359527,0.00047170714,0.0003212517,0.00021370628,0.000021988475,0.00012456786,0.0009431794,0.92239904,0.000027775157],"about_ca_topic_score_codex":0.007869109,"about_ca_topic_score_gemma":0.02253497,"teacher_disagreement_score":0.99014366,"about_ca_system_score_codex":0.0042418772,"about_ca_system_score_gemma":0.012855811,"threshold_uncertainty_score":0.05212605},"labels":[{"model":"gemma","categories":["metaresearch","research_integrity"],"domain":"incentives","study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","research_integrity"],"domain":"methods","study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3101218852","doi":"10.1136/bmjebm-2020-111577","title":"Blinding practices during acute point-of-care ultrasound research: the BLIND-US meta-research study","year":2020,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Ultrasound in Clinical Applications","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; University of Ottawa","funders":"","keywords":"Blinding; Medicine; Subgroup analysis; Point of care ultrasound; Medical physics; Meta-analysis; Research design; Physical therapy; Clinical trial; Radiology; Ultrasound; Internal medicine; Statistics","score_opus":0.717634259424709,"score_gpt":0.5996602644583762,"score_spread":0.11797399496633287,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3101218852","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06414367,0.84741634,0.055780858,0.0033413148,0.005381966,0.014730499,0.003854012,0.00060227327,0.004749095],"genre_scores_gemma":[0.84279233,0.07288862,0.048696224,0.005265628,0.002091256,0.023996383,0.0022624421,0.0004630744,0.0015440615],"study_design_codex":"meta_analysis","study_design_gemma":"observational","domain_scores_codex":[0.54963964,0.30815774,0.09644975,0.023479417,0.01951068,0.0027627426],"domain_scores_gemma":[0.37569925,0.47953957,0.086072765,0.042055767,0.0152867315,0.0013458602],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.28872517,0.003808098,0.014337589,0.0059784427,0.0024067918,0.009802326,0.0036125823,0.0060292627,0.0052354606],"category_scores_gemma":[0.4913675,0.002751881,0.032995287,0.0074984985,0.0052015083,0.008109644,0.004805532,0.00340511,0.000715706],"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":[0.024047794,0.00017474336,0.05506813,0.27720243,0.5916769,0.00023403835,0.0014722798,0.0015847719,0.00071282435,0.0033657777,0.0026433954,0.041817024],"study_design_scores_gemma":[0.014839847,0.0028693757,0.03024489,0.08677756,0.82724494,0.00076748943,0.00064233557,0.002784639,0.0025998903,0.016271826,0.014527032,0.00043024975],"about_ca_topic_score_codex":0.0022655635,"about_ca_topic_score_gemma":0.002868507,"teacher_disagreement_score":0.71127486,"about_ca_system_score_codex":0.0035876895,"about_ca_system_score_gemma":0.006456157,"threshold_uncertainty_score":0.8771285},"labels":[],"label_agreement":null},{"id":"W3103117573","doi":"10.1136/bmjebm-2020-111479","title":"Ten papers for teachers of evidence-based medicine and health care: Sicily workshop 2019","year":2020,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","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":"National Institute for Health and Care Research","keywords":"Health care; Medical education; Library science; Psychology; Medicine; Political science; Computer science","score_opus":0.7018070483281779,"score_gpt":0.6353810479535997,"score_spread":0.06642600037457813,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3103117573","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0059440155,0.5529749,0.002302995,0.18265109,0.22706413,0.008154951,0.001676569,0.0002662295,0.018965134],"genre_scores_gemma":[0.04336166,0.6010505,0.018245915,0.13684054,0.07671671,0.021665443,0.0064517925,0.00043740546,0.09523008],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9795248,0.008632412,0.0032521465,0.0009048596,0.0058239917,0.0018619425],"domain_scores_gemma":[0.96669066,0.010425811,0.0032798494,0.00070941256,0.010216155,0.008678217],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.043442152,0.0015364467,0.0022828912,0.008364135,0.0023477827,0.0070053777,0.0035904145,0.007953762,0.02434136],"category_scores_gemma":[0.049856503,0.0012358968,0.004291251,0.004078804,0.001314113,0.004381694,0.012974252,0.00661009,0.006764136],"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":[0.0011073744,0.00028723918,0.00048504004,0.043614082,0.00026418973,0.0013475121,0.0025294975,0.00015336306,0.0013115092,0.0027733317,0.69260377,0.25352305],"study_design_scores_gemma":[0.000324754,0.00022497593,0.0019878107,0.029903835,0.00018470858,0.0003433681,0.00092360587,0.00003904235,0.0003083373,0.00096554647,0.9647459,0.000048112754],"about_ca_topic_score_codex":0.0034304876,"about_ca_topic_score_gemma":0.009740744,"teacher_disagreement_score":0.95655787,"about_ca_system_score_codex":0.01100473,"about_ca_system_score_gemma":0.024179347,"threshold_uncertainty_score":0.22974676},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3112825017","doi":"10.1136/bmjebm-2020-111493","title":"Framework for the synthesis of non-randomised studies and randomised controlled trials: a guidance on conducting a systematic review and meta-analysis for healthcare decision making","year":2020,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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":"Canadian Agency for Drugs and Technologies in Health","funders":"Sanofi Pasteur; Sanofi","keywords":"Health care; Randomized controlled trial; Evidence-based medicine; Systematic review; Delphi method; Management science; Medicine; Quality (philosophy); Transparency (behavior); Best practice; MEDLINE; Computer science; Alternative medicine; Engineering; Artificial intelligence; Political science","score_opus":0.9173359485862664,"score_gpt":0.6565785159814549,"score_spread":0.26075743260481155,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3112825017","genre_codex":"methods","genre_gemma":"review","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00063134235,0.050043862,0.7646204,0.06598317,0.007022474,0.09253572,0.00824932,0.002464286,0.008449397],"genre_scores_gemma":[0.003419625,0.007242658,0.91147065,0.003941924,0.00047628465,0.071812436,0.00096376386,0.00014884135,0.00052380294],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.1827789,0.683921,0.091878384,0.010808221,0.027825963,0.0027875681],"domain_scores_gemma":[0.13010004,0.7301701,0.03300595,0.029964458,0.07171988,0.005039571],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.64560235,0.010276116,0.022018937,0.041781154,0.004983778,0.022129172,0.020057468,0.030002289,0.018700033],"category_scores_gemma":[0.73635215,0.010247436,0.037047468,0.026455032,0.016649067,0.013710276,0.017566223,0.03226669,0.015288032],"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":[0.0014226192,0.00025088328,0.0008376849,0.31226465,0.008417912,0.0010546226,0.015074011,0.015839024,0.0010042507,0.26261172,0.10257549,0.27864715],"study_design_scores_gemma":[0.003034781,0.00045104593,0.0008360644,0.3493623,0.006249523,0.0006366745,0.0017431969,0.012779174,0.0016427311,0.35227254,0.2701228,0.0008691237],"about_ca_topic_score_codex":0.012410114,"about_ca_topic_score_gemma":0.010499183,"teacher_disagreement_score":0.35439765,"about_ca_system_score_codex":0.024735074,"about_ca_system_score_gemma":0.11888003,"threshold_uncertainty_score":0.43703544},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W3117477626","doi":"10.1136/bmjebm-2020-111393","title":"Successes, lessons and opportunities: 15-year follow-up of an integrated evidence-based medicine curriculum","year":2020,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","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":"University of British Columbia; College of Family Physicians of Canada; University of Alberta","funders":"University of Alberta","keywords":"Curriculum; Medical education; Evidence-based medicine; Critical appraisal; Curriculum development; Evidence-based practice; Medicine; Family medicine; Psychology; Alternative medicine; Pedagogy; Pathology","score_opus":0.5774288423614425,"score_gpt":0.5805625078589207,"score_spread":0.0031336654974781775,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3117477626","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.40214413,0.5725552,0.0013404194,0.010618931,0.0013659843,0.0030413945,0.0016461398,0.0001052352,0.0071825045],"genre_scores_gemma":[0.7580757,0.22838902,0.005225967,0.0037136073,0.0002603417,0.0013270496,0.0010210156,0.000032944896,0.0019543346],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9915627,0.0031723483,0.0012602421,0.00041450237,0.0029633052,0.0006267382],"domain_scores_gemma":[0.97836703,0.008042345,0.0037379356,0.0010949204,0.006552484,0.002205296],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.021567535,0.0004339431,0.002049055,0.0020824266,0.00058185065,0.002073269,0.0014956988,0.0013427461,0.0011868514],"category_scores_gemma":[0.06597543,0.0004022729,0.0016943581,0.0023953277,0.00044265983,0.002355808,0.0031051831,0.001734492,0.00033503154],"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":[0.0010973222,0.0013456866,0.009257972,0.029557487,0.0012637059,0.00035759155,0.0046720398,0.00031829032,0.00053561904,0.0006066402,0.0037257397,0.94726187],"study_design_scores_gemma":[0.0041473215,0.027095823,0.43680885,0.18565345,0.013354565,0.0023397019,0.019170756,0.00161155,0.004727157,0.0024901505,0.30216312,0.00043751774],"about_ca_topic_score_codex":0.014807178,"about_ca_topic_score_gemma":0.023602437,"teacher_disagreement_score":0.021567535,"about_ca_system_score_codex":0.0051963367,"about_ca_system_score_gemma":0.012486677,"threshold_uncertainty_score":0.114061415},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"qualitative","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"qualitative","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W3121563112","doi":"10.1136/bmjebm-2020-111604","title":"Barriers to reporting guideline adherence in point-of-care ultrasound research: a cross-sectional survey of authors and journal editors","year":2021,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Ultrasound in Clinical Applications","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":"Ottawa Hospital; University of Ottawa","funders":"","keywords":"Likert scale; Guideline; Medicine; Family medicine; Cross-sectional study; Alternative medicine; Survey research; MEDLINE; Primary care; Medical education; Medical physics; Psychology; Pathology; Applied psychology","score_opus":0.37060179582142727,"score_gpt":0.5493210041737503,"score_spread":0.17871920835232302,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3121563112","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99205834,0.002000788,0.0009331699,0.0023239914,0.00006993573,0.00026511203,0.000744352,0.000030625386,0.0015737005],"genre_scores_gemma":[0.9946465,0.0011724487,0.001974999,0.0011294637,0.0001000667,0.00038143128,0.00032251998,0.000012806435,0.00025961714],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9415052,0.026516126,0.016418941,0.0028955145,0.011422276,0.0012419679],"domain_scores_gemma":[0.62149125,0.18095177,0.14485374,0.0089084245,0.037395213,0.006399573],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.054235134,0.00022095894,0.0009957104,0.003079646,0.000977018,0.0032264506,0.0008050834,0.001593341,0.0018178091],"category_scores_gemma":[0.21601485,0.00082913606,0.00085979025,0.0032903359,0.0010872398,0.0037012221,0.0018968199,0.001333742,0.0006809488],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011502774,0.00017533347,0.98402345,0.00044931492,0.0001369894,0.00010628156,0.007859662,0.00003886296,0.00023560286,0.000045766305,0.0010855881,0.0057282243],"study_design_scores_gemma":[0.00008288265,0.0011044313,0.97085303,0.00068753335,0.00015037006,0.0014779933,0.020688292,0.0004381186,0.00037206922,0.00016399924,0.0039114235,0.00006992764],"about_ca_topic_score_codex":0.0014761419,"about_ca_topic_score_gemma":0.0017255389,"teacher_disagreement_score":0.94576484,"about_ca_system_score_codex":0.00078344735,"about_ca_system_score_gemma":0.0021255426,"threshold_uncertainty_score":0.2868262},"labels":[],"label_agreement":null},{"id":"W3136262608","doi":"10.1136/bmjebm-2020-111542","title":"Understanding of research results, evidence summaries and their applicability—not critical appraisal—are core skills of medical curriculum","year":2021,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","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":"McMaster University; Impact","funders":"Terveyden Tutkimuksen Toimikunta; Sigrid Juséliuksen Säätiö; Helsingin ja Uudenmaan Sairaanhoitopiiri","keywords":"Critical appraisal; Rigour; Evidence-based medicine; Observational study; Systematic review; Medical education; Psychology; Medicine; Critical reading; Evidence-based practice; Psychological intervention; Medical literature; Curriculum; Reading (process); MEDLINE; Alternative medicine; Nursing; Pedagogy; Epistemology; Pathology","score_opus":0.8221500894503884,"score_gpt":0.6922985889787482,"score_spread":0.12985150047164018,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3136262608","genre_codex":"commentary","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0025229983,0.13594458,0.29296002,0.5151906,0.017014442,0.006317669,0.00096191256,0.0013096508,0.027778208],"genre_scores_gemma":[0.054584496,0.17005725,0.6712,0.0667123,0.020280382,0.009421028,0.0008916821,0.00080852315,0.0060443226],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.62210935,0.2686529,0.046471875,0.007486236,0.05343013,0.0018496258],"domain_scores_gemma":[0.20514283,0.66851753,0.03590088,0.03743497,0.04769779,0.0053060185],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.3134953,0.0022127419,0.007019938,0.013344471,0.0039161253,0.026389372,0.0055004116,0.012310789,0.01010947],"category_scores_gemma":[0.58677745,0.0024667825,0.0039153555,0.008505732,0.023838323,0.026517795,0.0089762835,0.02525024,0.009969677],"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":[0.0002627643,0.00045682845,0.0013321443,0.047197297,0.0011116992,0.00042583267,0.020134708,0.0022581008,0.0021574565,0.21395624,0.22583202,0.48487487],"study_design_scores_gemma":[0.00026691385,0.0004685485,0.0020261854,0.06129942,0.00037132512,0.0012280004,0.0043698503,0.001876775,0.0010422774,0.4693773,0.45730457,0.00036874585],"about_ca_topic_score_codex":0.0025547782,"about_ca_topic_score_gemma":0.002743289,"teacher_disagreement_score":0.6865047,"about_ca_system_score_codex":0.010456808,"about_ca_system_score_gemma":0.04457637,"threshold_uncertainty_score":0.84658253},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3150860698","doi":"10.1136/bmjebm-2020-111652","title":"Completeness of reporting for systematic reviews of point-of-care ultrasound: a meta-research study","year":2021,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Ultrasound in Clinical Applications","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":"Vancouver General Hospital; McMaster University; Queen's University; University of British Columbia; University of Ottawa","funders":"","keywords":"Checklist; Systematic review; Medicine; Cochrane Library; MEDLINE; Meta-analysis; Subgroup analysis; Publication bias; Medical physics; Pathology; Psychology","score_opus":0.6497725731400538,"score_gpt":0.5765663789344181,"score_spread":0.07320619420563568,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3150860698","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.023504859,0.7564083,0.08841067,0.008265713,0.0049583516,0.09413768,0.019726979,0.0009611974,0.0036262309],"genre_scores_gemma":[0.28255975,0.18830144,0.16649592,0.005311754,0.0016120143,0.34409866,0.010091334,0.00057915325,0.0009499988],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.37317908,0.31995586,0.22584175,0.022999216,0.05540752,0.0026166234],"domain_scores_gemma":[0.26425704,0.5013101,0.13407764,0.044844896,0.053665143,0.0018451684],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.45039555,0.00350956,0.023455994,0.025605742,0.0032051147,0.010570443,0.0062267426,0.0060479804,0.0068419497],"category_scores_gemma":[0.7356257,0.00535228,0.058942754,0.028037457,0.00504277,0.012171741,0.008410189,0.004880577,0.0010601751],"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":[0.0018761071,0.00009384803,0.015286001,0.72510475,0.2118805,0.00021512393,0.0013611586,0.0007775647,0.00034954908,0.002160719,0.00629114,0.0346035],"study_design_scores_gemma":[0.0050755613,0.00092526677,0.026632821,0.4514366,0.48395568,0.00086108287,0.0008544173,0.0025279627,0.0011074031,0.0077184024,0.0183586,0.0005461756],"about_ca_topic_score_codex":0.0035991543,"about_ca_topic_score_gemma":0.0064719925,"teacher_disagreement_score":0.5496044,"about_ca_system_score_codex":0.012768164,"about_ca_system_score_gemma":0.025728041,"threshold_uncertainty_score":0.6777601},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.003281369,0.05895252,0.157535,0.66818756,0.071858235,0.00020520904,0.00077475957,0.00095996313,0.038245387],"genre_scores_gemma":[0.11836228,0.09304679,0.24386868,0.39064327,0.10998973,0.00087886263,0.0010100362,0.0015587693,0.04064158],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9953571,0.0026899443,0.0006513448,0.00035211464,0.0008128776,0.00013658215],"domain_scores_gemma":[0.9740865,0.02133372,0.0010417197,0.0006260268,0.0024535104,0.00045853545],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.006368312,0.0012749194,0.00084494305,0.0015113386,0.0012546434,0.0030374497,0.001726109,0.0046221786,0.009035505],"category_scores_gemma":[0.0603965,0.0004207761,0.001024891,0.0010916226,0.0046457956,0.00904748,0.0019967612,0.013510066,0.0054558064],"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":[0.00011324971,0.000055540033,0.001617868,0.0007953953,0.000027423344,0.0028614493,0.002311223,0.0010909751,0.000517045,0.16581118,0.7074218,0.11737691],"study_design_scores_gemma":[0.000041721007,0.00010913577,0.00065279193,0.0013538694,0.000038034425,0.00837246,0.00075544073,0.0030995526,0.0005881891,0.19710189,0.787816,0.00007092077],"about_ca_topic_score_codex":0.0018821231,"about_ca_topic_score_gemma":0.0019600787,"teacher_disagreement_score":0.009035505,"about_ca_system_score_codex":0.0015734589,"about_ca_system_score_gemma":0.0016162575,"threshold_uncertainty_score":0.033679247},"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":"W3165462729","doi":"10.1136/bmjebm-2021-111675","title":"Quality of systematic reviews supporting the 2017 ACC/AHA and 2018 ESC/ESH guidelines for the management of hypertension","year":2021,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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":"Checklist; Systematic review; Medicine; Protocol (science); MEDLINE; Family medicine; Canadian Cardiovascular Society; Alternative medicine; Internal medicine; Psychology; Pathology; Myocardial infarction","score_opus":0.9622276600800553,"score_gpt":0.6563818915020886,"score_spread":0.30584576857796675,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3165462729","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.011824369,0.8873807,0.024568891,0.029922834,0.0060005626,0.01781815,0.017625537,0.00044885653,0.0044100727],"genre_scores_gemma":[0.2910725,0.44218385,0.14225137,0.02019385,0.003945223,0.08512388,0.013850829,0.00048583033,0.00089274143],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.33857787,0.28128275,0.29183444,0.01752518,0.067999646,0.0027800724],"domain_scores_gemma":[0.13844149,0.58689684,0.15117308,0.022346847,0.09869954,0.002442244],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.4788518,0.0027024406,0.0143151665,0.026334012,0.0033781712,0.012822029,0.006014213,0.0068886206,0.0062474376],"category_scores_gemma":[0.8013855,0.0038346925,0.026013128,0.021812802,0.0046776226,0.009246024,0.007260464,0.0054112864,0.00082063285],"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":[0.00086975173,0.000055924298,0.011613929,0.84665674,0.06559455,0.00013080049,0.0015584067,0.00062844204,0.00039887923,0.0026125908,0.013951907,0.055928066],"study_design_scores_gemma":[0.0013540232,0.00031390486,0.018818326,0.8239193,0.11609717,0.00026707814,0.00095748715,0.0013161928,0.0007153601,0.00510485,0.030882107,0.00025424498],"about_ca_topic_score_codex":0.013370885,"about_ca_topic_score_gemma":0.024182227,"teacher_disagreement_score":0.5211482,"about_ca_system_score_codex":0.012760593,"about_ca_system_score_gemma":0.04597981,"threshold_uncertainty_score":0.64266855},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"evaluation","study_design":"systematic_review","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3201255490","doi":"10.1136/bmjebm-2021-111772","title":"Appraising qualitative health research—towards a differentiated approach","year":2021,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Primary Care and Health Outcomes","field":"Health Professions","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":"Nipissing University","funders":"","keywords":"Critical appraisal; Qualitative research; Management science; Health care; Research design; Neglect; Qualitative property; Engineering ethics; Psychology; Data science; Computer science; Knowledge management; Sociology; Medicine; Social science; Engineering; Political science; Nursing; Alternative medicine; Pathology","score_opus":0.745284703788777,"score_gpt":0.6760284951873954,"score_spread":0.0692562086013816,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3201255490","genre_codex":"methods","genre_gemma":"methods","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.002369726,0.05095193,0.7939872,0.12086829,0.008825747,0.008182213,0.0003104147,0.00034947894,0.014154971],"genre_scores_gemma":[0.03742551,0.02504895,0.90472203,0.016028168,0.0026614366,0.012084888,0.00033244616,0.00029285328,0.0014037504],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.153636,0.7749116,0.032380834,0.0073289583,0.030096078,0.0016464936],"domain_scores_gemma":[0.11749118,0.7738834,0.013232616,0.0273261,0.06433995,0.003726771],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.67463523,0.0032613734,0.0069401497,0.018563088,0.006206528,0.038114887,0.017577073,0.014637462,0.0055696],"category_scores_gemma":[0.67755634,0.0036351678,0.0037400583,0.01901748,0.05454705,0.028228741,0.032965183,0.024074852,0.0025381746],"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":[0.00023344657,0.000093167946,0.00087183673,0.032965355,0.0005277825,0.00051451195,0.15376128,0.0022874745,0.0011465662,0.6069375,0.018889586,0.18177141],"study_design_scores_gemma":[0.00014650026,0.0001414651,0.00037764633,0.05124124,0.00017417573,0.00034347002,0.030620344,0.0031666937,0.00049887504,0.7735794,0.1395323,0.00017783679],"about_ca_topic_score_codex":0.003956042,"about_ca_topic_score_gemma":0.003586546,"teacher_disagreement_score":0.32536477,"about_ca_system_score_codex":0.024256486,"about_ca_system_score_gemma":0.054121397,"threshold_uncertainty_score":0.40123272},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4200239034","doi":"10.1136/bmjebm-2021-111850","title":"Users' Guides to the Medical Literature series on social media (part 2): how to appraise studies using data from platforms","year":2021,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Social Media in Health Education","field":"Social Sciences","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; Impact; Université de Montréal","funders":"","keywords":"Series (stratigraphy); Social media; Data science; Computer science; World Wide Web; Information retrieval; Geology","score_opus":0.64309403782733,"score_gpt":0.5438279267303098,"score_spread":0.09926611109702022,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4200239034","genre_codex":"dataset","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00458054,0.086759515,0.1541855,0.17769597,0.027207145,0.09728623,0.27305073,0.034841903,0.14439246],"genre_scores_gemma":[0.007880724,0.06411894,0.6028565,0.05954301,0.008343103,0.10478596,0.048965305,0.01659699,0.08690946],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.92024285,0.02951111,0.037590668,0.0012870305,0.010380091,0.0009882598],"domain_scores_gemma":[0.34016055,0.5128958,0.030411826,0.01305911,0.09830622,0.0051664463],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.08205412,0.0039065448,0.0077911974,0.047005944,0.002430219,0.00858243,0.0048798868,0.0087290425,0.20503259],"category_scores_gemma":[0.4333638,0.0048466283,0.0047199614,0.033668056,0.0037676536,0.010298007,0.010802493,0.009469652,0.107299596],"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":[0.0003790371,0.0001340981,0.00036345114,0.028626017,0.00013749604,0.000119073404,0.0025153349,0.000255412,0.0008345715,0.0017834184,0.8292948,0.13555737],"study_design_scores_gemma":[0.0005570835,0.00011858742,0.001481033,0.02448634,0.00021661154,0.00023607712,0.001486198,0.00043757027,0.00069531234,0.005986457,0.96410733,0.00019149146],"about_ca_topic_score_codex":0.008431567,"about_ca_topic_score_gemma":0.022115985,"teacher_disagreement_score":0.91794586,"about_ca_system_score_codex":0.0029146648,"about_ca_system_score_gemma":0.022277761,"threshold_uncertainty_score":0.68590236},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"evaluation","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W4200294250","doi":"10.1136/bmjebm-2021-111817","title":"Users' Guides to the Medical Literature series on social media (part 1): how to interpret healthcare information available on platforms","year":2021,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Social Media in Health Education","field":"Social Sciences","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; Impact; Université de Montréal","funders":"","keywords":"Social media; Health care; Medical information; Data science; Internet privacy; Computer science; Public relations; World Wide Web; Knowledge management; Political science","score_opus":0.20942782448776642,"score_gpt":0.434394690256741,"score_spread":0.22496686576897457,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4200294250","genre_codex":"other","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.007710845,0.034987256,0.13198066,0.15390739,0.022389328,0.025772981,0.25424662,0.037535552,0.33146942],"genre_scores_gemma":[0.01871004,0.05482719,0.36381844,0.055204265,0.007178114,0.03169107,0.089900434,0.026032588,0.35263798],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9823523,0.006631898,0.0056814887,0.0005131133,0.0043657846,0.00045532125],"domain_scores_gemma":[0.7315167,0.19418056,0.007817599,0.0069729146,0.055754453,0.0037576507],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.030293338,0.0020157509,0.00247321,0.02795436,0.002461667,0.005758062,0.002323894,0.0044637793,0.17583527],"category_scores_gemma":[0.17435734,0.0017560207,0.0014835233,0.019767523,0.0021296649,0.007719893,0.0069090417,0.004647937,0.118850365],"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":[0.00009895867,0.000060760205,0.00031408298,0.0038109876,0.000012965087,0.000075926146,0.0033352065,0.000083987616,0.00069554645,0.0015907967,0.8659652,0.12395564],"study_design_scores_gemma":[0.000027892958,0.000024555264,0.00050145097,0.002078303,0.00001563755,0.000063575535,0.0009904954,0.000093266484,0.00037974492,0.0011075194,0.99468094,0.000036595717],"about_ca_topic_score_codex":0.0130034415,"about_ca_topic_score_gemma":0.030607654,"teacher_disagreement_score":0.96970665,"about_ca_system_score_codex":0.0019931763,"about_ca_system_score_gemma":0.015978176,"threshold_uncertainty_score":0.58822757},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","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":"low"}],"label_agreement":"split"},{"id":"W4220907181","doi":"10.1136/bmjebm-2022-111953","title":"Advocating for evidence-informed decisions to make healthcare fit for each person","year":2022,"lang":"en","type":"editorial","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","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; St. Michael's Hospital; University of British Columbia","funders":"","keywords":"Health care; Psychology; Public relations; Internet privacy; Social psychology; Political science; Computer science; Law","score_opus":0.555224960111113,"score_gpt":0.6254457142454214,"score_spread":0.0702207541343084,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4220907181","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000015933314,0.0069637215,0.00023121467,0.053179972,0.9383877,0.0000620702,0.00006244918,0.000043960037,0.001053052],"genre_scores_gemma":[0.00036172415,0.0067755957,0.0005494682,0.034382917,0.9541946,0.00011246976,0.000044814326,0.000064902575,0.0035134319],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.94025296,0.020049172,0.008679644,0.0031849944,0.02602111,0.0018120942],"domain_scores_gemma":[0.6566604,0.22711043,0.01328734,0.0070317467,0.07679841,0.019111698],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.06854481,0.0062075667,0.010695826,0.010080537,0.007893515,0.024450213,0.007376086,0.050294194,0.025701342],"category_scores_gemma":[0.2623177,0.0037644817,0.0066119656,0.005204285,0.008781238,0.013556932,0.004958896,0.0553566,0.018852677],"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":[0.000061165425,0.000015947982,0.000025160527,0.0016119907,0.00011536826,0.00005939442,0.000053566815,0.000026735035,0.00004096568,0.00062032783,0.9911055,0.0062639397],"study_design_scores_gemma":[0.0005111422,0.00006320543,0.0002639438,0.008847323,0.00059422455,0.0002590834,0.0002879942,0.0003450613,0.00016608306,0.010279449,0.97829175,0.00009087279],"about_ca_topic_score_codex":0.0036931336,"about_ca_topic_score_gemma":0.010105072,"teacher_disagreement_score":0.06854481,"about_ca_system_score_codex":0.0067545734,"about_ca_system_score_gemma":0.018082466,"threshold_uncertainty_score":0.36250395},"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":"high"}],"label_agreement":"agree"},{"id":"W4223598199","doi":"10.1136/bmjebm-2022-111978","title":"In memory of Richard Saitz (1963–2022): former editor of BMJ Evidence-Based Medicine","year":2022,"lang":"en","type":"letter","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","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":"Psychoanalysis; Medicine; History; Classics; Psychology; Cognitive science","score_opus":0.2879360818303817,"score_gpt":0.5279963544396121,"score_spread":0.2400602726092304,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4223598199","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00017087252,0.0025102159,0.00007457752,0.84049845,0.15526807,0.000023164812,0.00008363516,0.000025236093,0.0013458198],"genre_scores_gemma":[0.0015482533,0.0015031968,0.00015266248,0.85911655,0.12840727,0.000052733638,0.000041197745,0.000033400134,0.009144803],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9924468,0.001709969,0.001370536,0.00097120967,0.0026192956,0.00088224787],"domain_scores_gemma":[0.97667456,0.010724322,0.0031431098,0.0005471882,0.0038723843,0.0050384155],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.007986325,0.002117076,0.002815313,0.0022115242,0.007032007,0.009416019,0.002493297,0.0831894,0.009781584],"category_scores_gemma":[0.08171262,0.0016918345,0.0026516833,0.0018088751,0.0037470474,0.00674224,0.0038283532,0.06796973,0.01227727],"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":[0.000103195525,0.00002881597,0.00038464056,0.00013006665,0.000039247723,0.0013050594,0.00012358688,0.00003402982,0.00009444393,0.0005678304,0.9939857,0.0032034589],"study_design_scores_gemma":[0.00016529103,0.00009313613,0.0015794926,0.00096407946,0.00010991159,0.001243307,0.00052197627,0.00035415802,0.00020087612,0.0028009808,0.9918727,0.000094117866],"about_ca_topic_score_codex":0.0071625314,"about_ca_topic_score_gemma":0.013665118,"teacher_disagreement_score":0.0831894,"about_ca_system_score_codex":0.005548396,"about_ca_system_score_gemma":0.007350192,"threshold_uncertainty_score":0.04223621},"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":"high"}],"label_agreement":"agree"},{"id":"W4224057184","doi":"10.1136/bmjebm-2021-111866","title":"Which actionable statements qualify as good practice statements In Covid-19 guidelines? A systematic appraisal","year":2022,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Clinical practice guidelines implementation","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 Laval; Ottawa Hospital; Centre hospitalier universitaire de Québec; McMaster University; Impact; Cochrane; Bruyère; Hôpital de l'Enfant-Jésus; University of Ottawa","funders":"Canadian Institutes of Health Research; World Health Organization","keywords":"Guideline; Global Positioning System; Critical appraisal; Systematic review; Grading (engineering); MEDLINE; Quality (philosophy); Medicine; Computer science; Alternative medicine; Engineering; Political science","score_opus":0.7104620991194237,"score_gpt":0.7006955760346055,"score_spread":0.00976652308481818,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4224057184","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.010911501,0.8992161,0.00940219,0.018236648,0.003577796,0.04711847,0.007641957,0.00018058022,0.0037147196],"genre_scores_gemma":[0.2175987,0.54377556,0.08421792,0.014515331,0.0015389569,0.12900923,0.0083288625,0.00016334708,0.0008521441],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.52131987,0.21852641,0.19113505,0.012624606,0.053118777,0.003275301],"domain_scores_gemma":[0.34845912,0.3778788,0.15242204,0.019512312,0.09747023,0.00425765],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.27122638,0.002413488,0.014103552,0.025892936,0.0026432464,0.009988977,0.005206937,0.0062497677,0.0050812154],"category_scores_gemma":[0.66026837,0.0031383156,0.012140082,0.02399252,0.0056433454,0.01174076,0.0055642226,0.0041339584,0.0007745329],"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":[0.00042080885,0.000054357166,0.0049624913,0.902535,0.008052177,0.00009997069,0.0023834885,0.00019116631,0.00014824732,0.0020533998,0.0071891877,0.07190962],"study_design_scores_gemma":[0.00062017934,0.00017244923,0.004940994,0.9646682,0.013209526,0.000119453354,0.0013252421,0.00023457434,0.00017424871,0.001657474,0.012788871,0.00008873997],"about_ca_topic_score_codex":0.011996188,"about_ca_topic_score_gemma":0.025184447,"teacher_disagreement_score":0.7287736,"about_ca_system_score_codex":0.017861884,"about_ca_system_score_gemma":0.06761216,"threshold_uncertainty_score":0.8987076},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"evaluation","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W4224095215","doi":"10.1136/bmjebm-2022-111962","title":"Good or best practice statements: proposal for the operationalisation and implementation of GRADE guidance","year":2022,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":164,"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 Laval; Ottawa Hospital; Centre hospitalier universitaire de Québec; McMaster University; Impact; Cochrane; Bruyère; Hôpital de l'Enfant-Jésus; University of Ottawa","funders":"Canadian Institutes of Health Research; World Health Organization","keywords":"Grading (engineering); Guideline; Global Positioning System; Computer science; Fidelity; Process management; Best practice; Medical education; Knowledge management; Medicine; Business; Engineering; Political science","score_opus":0.5414977190713219,"score_gpt":0.5727772921882497,"score_spread":0.03127957311692775,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4224095215","genre_codex":"commentary","genre_gemma":"methods","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0033237806,0.010356175,0.3728243,0.44642776,0.025893617,0.08214523,0.0066616787,0.0040240902,0.048343364],"genre_scores_gemma":[0.012456869,0.0035057992,0.91573423,0.022171076,0.0015829733,0.036872264,0.0033986773,0.0006165721,0.003661576],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.28384727,0.39332634,0.20637436,0.016524093,0.088509455,0.011418484],"domain_scores_gemma":[0.16831572,0.30955753,0.059816193,0.061756115,0.37977198,0.02078249],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.59937763,0.004795094,0.007826364,0.040962946,0.008503218,0.042242527,0.028643738,0.03601551,0.012868568],"category_scores_gemma":[0.7815488,0.0075935777,0.018442022,0.033884283,0.020527726,0.03542134,0.032227173,0.04324681,0.01654191],"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":[0.0005101424,0.0005500319,0.0028970218,0.021370627,0.00064367487,0.00065452,0.014374566,0.0029673183,0.00075388193,0.29933771,0.36519328,0.2907472],"study_design_scores_gemma":[0.001051722,0.0006995463,0.0035527297,0.060048223,0.0009700473,0.00073945103,0.0077507114,0.0051982473,0.0014580555,0.20042005,0.7171434,0.0009678406],"about_ca_topic_score_codex":0.024822364,"about_ca_topic_score_gemma":0.022440018,"teacher_disagreement_score":0.40062237,"about_ca_system_score_codex":0.034343634,"about_ca_system_score_gemma":0.18167605,"threshold_uncertainty_score":0.49403876},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W4283587010","doi":"10.1136/bmjebm-2022-111928","title":"Theory and practice of Bayesian and frequentist frameworks for network meta-analysis","year":2022,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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":"St. Joseph’s Healthcare Hamilton; McMaster University; Ted Rogers Centre for Heart Research; University Health Network; Impact","funders":"","keywords":"Frequentist inference; Bayesian probability; Computer science; Bayesian network; Bayesian statistics; Frequentist probability; Machine learning; Bayesian inference; Artificial intelligence; Data science","score_opus":0.7536134177763099,"score_gpt":0.5727628547311456,"score_spread":0.18085056304516423,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4283587010","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00038189045,0.004620253,0.9883957,0.0034593376,0.0002372906,0.00026496584,0.00014751057,0.00016357732,0.0023295667],"genre_scores_gemma":[0.040538065,0.0076047517,0.9435964,0.002176965,0.0010662088,0.0039733653,0.00017581352,0.00017855185,0.00068983674],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.69468594,0.26734605,0.0097828135,0.010533305,0.016634416,0.0010174748],"domain_scores_gemma":[0.44655433,0.51311576,0.011819017,0.01792074,0.009688114,0.0009020799],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.26744863,0.0044340156,0.00835998,0.01484073,0.0026334669,0.011795032,0.010424491,0.00783312,0.006612868],"category_scores_gemma":[0.42623168,0.0035498193,0.008692788,0.011632901,0.015016849,0.010514863,0.0072329384,0.013336263,0.001851403],"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":[0.00006467801,0.000039536542,0.00058503397,0.0024779341,0.0013585996,0.00011607489,0.000711232,0.032053534,0.00008643475,0.9151263,0.0037183815,0.04366236],"study_design_scores_gemma":[0.000079389596,0.00003535756,0.00012606315,0.0007989316,0.00021472304,0.00007175138,0.00005164659,0.03138826,0.00011259007,0.95934546,0.007721314,0.000054609478],"about_ca_topic_score_codex":0.007943441,"about_ca_topic_score_gemma":0.0052587595,"teacher_disagreement_score":0.73255134,"about_ca_system_score_codex":0.009950807,"about_ca_system_score_gemma":0.009664898,"threshold_uncertainty_score":0.90336627},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4285794520","doi":"10.1136/bmjebm-2022-111952","title":"Adapt or die: how the pandemic made the shift from EBM to EBM+ more urgent","year":2022,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Healthcare cost, quality, practices","field":"Health Professions","cited_by":123,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"CARE Canada; Association of Professional Engineers and Geoscientists of Manitoba; Public Health Ontario; University of Toronto","funders":"National Institute for Health and Care Research; Canadian Institutes of Health Research; NIHR Oxford Biomedical Research Centre; National Medical Research Council; National Health and Medical Research Council; Wellcome Trust","keywords":"Context (archaeology); Checklist; Evidence-based medicine; Psychological intervention; Quality (philosophy); Yesterday; Evidence-based practice; Health care; Scientific evidence; Intervention (counseling); Psychology; Engineering ethics; MEDLINE; Management science; Medicine; Political science; Engineering; Alternative medicine; Nursing; Epistemology; History","score_opus":0.7945608969354264,"score_gpt":0.5892763489863074,"score_spread":0.205284547949119,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4285794520","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0020873495,0.031779625,0.0022376815,0.95139223,0.006175212,0.000028167005,0.00011023529,0.00006603532,0.0061235027],"genre_scores_gemma":[0.17421833,0.101840965,0.01353888,0.6875112,0.010890503,0.00027923152,0.00034107227,0.00049984903,0.010879972],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.96479005,0.024216073,0.0015244606,0.0024421152,0.003965328,0.003061888],"domain_scores_gemma":[0.9515327,0.030650418,0.0026582514,0.002525493,0.0072966064,0.005336456],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.05034799,0.0010562022,0.0015903731,0.0025562248,0.0073094456,0.018583812,0.0037148902,0.018454554,0.009761367],"category_scores_gemma":[0.09026349,0.0006719765,0.002015602,0.0023959295,0.03300995,0.035593152,0.010559329,0.026880305,0.0028036013],"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":[0.00037654355,0.00013092764,0.0041593825,0.002794574,0.0002567532,0.0016407394,0.027079418,0.00128482,0.00060466817,0.41498858,0.35163638,0.19504723],"study_design_scores_gemma":[0.00005158923,0.00013429373,0.0020626436,0.005399647,0.000090829395,0.0009192132,0.026268264,0.0006709329,0.00034163514,0.42670286,0.5371958,0.00016236548],"about_ca_topic_score_codex":0.019125113,"about_ca_topic_score_gemma":0.020930981,"teacher_disagreement_score":0.949652,"about_ca_system_score_codex":0.014066285,"about_ca_system_score_gemma":0.020381683,"threshold_uncertainty_score":0.26626885},"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":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4291021075","doi":"10.1136/bmjebm-2022-111944","title":"Knowledge user survey and Delphi process to inform development of a new risk of bias tool to assess systematic reviews with network meta-analysis (RoB NMA tool)","year":2022,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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":"McGill University; Simon Fraser University; Ottawa Hospital; University of Ottawa; University of Toronto; University of British Columbia; Cochrane; St. Michael's Hospital","funders":"Canadian Institutes of Health Research; Medical Research Council; National Institute for Health and Care Research; World Health Organization","keywords":"Delphi method; Knowledge translation; Systematic review; Delphi; Psychology; Process (computing); Quality (philosophy); Knowledge management; Computer science; MEDLINE; Political science; Artificial intelligence","score_opus":0.9019285677618601,"score_gpt":0.5686405846106061,"score_spread":0.3332879831512541,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4291021075","genre_codex":"methods","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.040570676,0.0018307087,0.5909587,0.012233036,0.0011372621,0.33250502,0.004078472,0.0018904058,0.014795748],"genre_scores_gemma":[0.04202044,0.0006484098,0.48962122,0.002584311,0.00017146267,0.46336734,0.00064190704,0.0001741242,0.00077064335],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.37576243,0.5240004,0.06530117,0.007994113,0.022199545,0.0047422717],"domain_scores_gemma":[0.38545132,0.44642136,0.028037887,0.029758006,0.105906196,0.00442525],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.50579834,0.0026764048,0.0034221576,0.017686516,0.0037448637,0.005822336,0.0033582537,0.0038223423,0.020021435],"category_scores_gemma":[0.5936334,0.0025768445,0.0064881924,0.010708165,0.004321433,0.009954034,0.015255704,0.0053486517,0.0046697566],"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":[0.002637726,0.0010930079,0.016029514,0.041753594,0.001264084,0.0006936703,0.13775495,0.006955925,0.005021791,0.032765944,0.078918315,0.6751115],"study_design_scores_gemma":[0.0073453225,0.0062070624,0.0347178,0.06960042,0.0025515407,0.0012757102,0.08640241,0.0681191,0.012520284,0.17136599,0.53781015,0.0020842433],"about_ca_topic_score_codex":0.002295161,"about_ca_topic_score_gemma":0.003601621,"teacher_disagreement_score":0.49420166,"about_ca_system_score_codex":0.008511299,"about_ca_system_score_gemma":0.029563658,"threshold_uncertainty_score":0.60943866},"labels":[],"label_agreement":null},{"id":"W4297996183","doi":"10.1136/bmjebm-2022-112002","title":"Proof of concept: all-virtual guideline development workshops using GRADE during the COVID-19 pandemic","year":2022,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","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":"Grading (engineering); Guideline; Medical education; Pandemic; Psychological intervention; Coronavirus disease 2019 (COVID-19); Psychology; Medicine; Nursing; Engineering","score_opus":0.548356336968319,"score_gpt":0.5763784541983684,"score_spread":0.028022117230049437,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4297996183","genre_codex":"protocol","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.05243178,0.007396342,0.21320747,0.100661404,0.07917354,0.45804918,0.020247772,0.006347364,0.06248521],"genre_scores_gemma":[0.111296564,0.0043274877,0.37095997,0.040440097,0.0071310285,0.44222984,0.0038346506,0.0011856778,0.018594658],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.93163776,0.048492055,0.004524457,0.00299662,0.009439121,0.002909999],"domain_scores_gemma":[0.8029503,0.09560983,0.015448533,0.020477349,0.046593208,0.0189208],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.08880705,0.0020462594,0.0018197733,0.0010340185,0.002580951,0.005696768,0.0043366216,0.007067589,0.100826144],"category_scores_gemma":[0.26981992,0.0011032973,0.0035120407,0.00060593645,0.002099756,0.004510876,0.006829325,0.0049126544,0.0181132],"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":[0.018473856,0.009094156,0.002657599,0.044336915,0.0012924743,0.0019981936,0.005967986,0.0013682896,0.005469805,0.012362114,0.36653516,0.5304434],"study_design_scores_gemma":[0.05385996,0.03202429,0.007910415,0.041045558,0.0020120821,0.002076101,0.00443195,0.002263488,0.015734361,0.016329061,0.82185674,0.0004559915],"about_ca_topic_score_codex":0.0009920591,"about_ca_topic_score_gemma":0.0023680034,"teacher_disagreement_score":0.91119295,"about_ca_system_score_codex":0.002634108,"about_ca_system_score_gemma":0.0192301,"threshold_uncertainty_score":0.4696622},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4306181334","doi":"10.1136/bmjebm-2022-112065","title":"The Pandora’s Box of Evidence Synthesis and the case for a living Evidence Synthesis Taxonomy","year":2022,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Healthcare cost, quality, practices","field":"Health Professions","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":"Public Health Ontario; St. Michael's Hospital; University of Toronto; McMaster University; Queen's University; Impact","funders":"National Health and Medical Research Council; University of Adelaide","keywords":"Context (archaeology); Pandemic; Nursing; Scope (computer science); Public relations; Psychology; Medical education; Business; Medicine; Coronavirus disease 2019 (COVID-19); Political science; Geography; Computer science","score_opus":0.7695973013225446,"score_gpt":0.5618726478439796,"score_spread":0.20772465347856495,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4306181334","genre_codex":"commentary","genre_gemma":"methods","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00018222148,0.1476762,0.030308781,0.79411286,0.02326062,0.00033703665,0.00017653665,0.0001899446,0.0037559213],"genre_scores_gemma":[0.019610098,0.18879984,0.2491108,0.4797497,0.05428101,0.004795475,0.00039349313,0.00063670793,0.0026228512],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.30258477,0.50416183,0.08699825,0.016909083,0.08527999,0.004066009],"domain_scores_gemma":[0.13873294,0.7709654,0.019473342,0.025866045,0.037650198,0.007312109],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.6188834,0.003965951,0.01877986,0.02939516,0.010973298,0.054510098,0.013666191,0.044339567,0.009256523],"category_scores_gemma":[0.6970709,0.005326293,0.007948692,0.020453794,0.09123459,0.08047324,0.02652624,0.0852803,0.0048735486],"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":[0.00041554685,0.00008091886,0.00031108095,0.019571295,0.0006407805,0.00017550781,0.003177686,0.0006364357,0.00015191225,0.74016786,0.15329364,0.08137737],"study_design_scores_gemma":[0.0003555666,0.00012835517,0.00019234979,0.040777225,0.00022923521,0.0002333598,0.00112328,0.0010510915,0.00009463369,0.7792349,0.17638768,0.00019231548],"about_ca_topic_score_codex":0.0077738836,"about_ca_topic_score_gemma":0.0070399963,"teacher_disagreement_score":0.38111663,"about_ca_system_score_codex":0.026432471,"about_ca_system_score_gemma":0.07332203,"threshold_uncertainty_score":0.4699847},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W4309247470","doi":"10.1136/bmjebm-2022-112064","title":"What is the place in therapy for nirmatrelvir/ritonavir?","year":2022,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Healthcare cost, quality, practices","field":"Health Professions","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; McGill University","funders":"","keywords":"Ritonavir; Medicine; Antiretroviral therapy; Virology; Human immunodeficiency virus (HIV); Viral load","score_opus":0.7470211808498358,"score_gpt":0.6191185209135557,"score_spread":0.12790265993628014,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4309247470","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.03773919,0.04291615,0.0012942399,0.87124544,0.011593434,0.00006918214,0.0013848312,0.00006604738,0.03369155],"genre_scores_gemma":[0.8001122,0.03800649,0.006140994,0.12712161,0.017327508,0.00015395391,0.0011514665,0.00034354563,0.009642266],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.98293465,0.007854437,0.0017226205,0.000806866,0.003776292,0.0029052107],"domain_scores_gemma":[0.9518842,0.012416655,0.009648002,0.000865567,0.009465819,0.015719732],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01002806,0.0003436987,0.0017016854,0.0012543519,0.0029597778,0.008979482,0.0028872136,0.0062435335,0.024870228],"category_scores_gemma":[0.053934716,0.00048558446,0.002400728,0.0026469876,0.0034386641,0.00928057,0.0020963696,0.010330058,0.0036809843],"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":[0.003538247,0.0015951603,0.11036362,0.005662701,0.0011321347,0.0034094763,0.004033249,0.00094126083,0.0014929499,0.030265925,0.2279381,0.6096272],"study_design_scores_gemma":[0.0015120393,0.0038519164,0.19527395,0.05431861,0.0019874095,0.007259302,0.05329088,0.002253479,0.0029407619,0.05521555,0.6211722,0.00092392584],"about_ca_topic_score_codex":0.021131642,"about_ca_topic_score_gemma":0.055345632,"teacher_disagreement_score":0.024870228,"about_ca_system_score_codex":0.00740245,"about_ca_system_score_gemma":0.020307805,"threshold_uncertainty_score":0.08319914},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4312061410","doi":"10.1136/bmjebm-2022-112024","title":"Retrieval barriers in individual participant data reviews with network meta-analysis","year":2022,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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":"Public Health Ontario; University of Toronto; St. Michael's Hospital","funders":"","keywords":"Randomized controlled trial; MEDLINE; Medicine; Psychological intervention; Systematic review; Unavailability; Meta-analysis; Data sharing; Computer science; Alternative medicine; Nursing; Political science; Surgery; Engineering","score_opus":0.9577443300487736,"score_gpt":0.6060650932160673,"score_spread":0.3516792368327063,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4312061410","genre_codex":"methods","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.046885844,0.22591,0.46209097,0.08166687,0.008494404,0.10855504,0.033800177,0.00396855,0.028628195],"genre_scores_gemma":[0.21781033,0.028473882,0.5037732,0.021200359,0.0023968266,0.21278913,0.00923655,0.0020057962,0.0023139082],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.07896572,0.6723237,0.19937898,0.01579482,0.03220421,0.0013325249],"domain_scores_gemma":[0.01885861,0.8743971,0.047165457,0.041897606,0.016967464,0.00071384816],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.74176466,0.005169207,0.014375246,0.028320404,0.0049338914,0.017239619,0.0116452975,0.009095522,0.0151247755],"category_scores_gemma":[0.9281144,0.0075647193,0.017102143,0.027606564,0.00761626,0.023280863,0.023232315,0.0065316493,0.0034869104],"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":[0.0076556657,0.00034020812,0.033160735,0.50806546,0.057526115,0.0019929046,0.040639102,0.00620626,0.0016620547,0.04127849,0.037860174,0.26361287],"study_design_scores_gemma":[0.009367499,0.00205807,0.020733759,0.47577673,0.053369343,0.0029559454,0.010848748,0.016243543,0.005616111,0.14861841,0.25220892,0.0022030163],"about_ca_topic_score_codex":0.006315651,"about_ca_topic_score_gemma":0.0071468516,"teacher_disagreement_score":0.25823534,"about_ca_system_score_codex":0.01401682,"about_ca_system_score_gemma":0.043476693,"threshold_uncertainty_score":0.3184502},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4315475599","doi":"10.1136/bmjebm-2022-112117","title":"Expanding the measurement of overdiagnosis in the context of disease precursors and risk factors","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Colorectal Cancer Screening and Detection","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":"Public Health Agency of Canada; Université de Montréal; Université Laval; University of Alberta; Public Health Ontario; University of Toronto; McGill University","funders":"","keywords":"Overdiagnosis; Context (archaeology); Disease; Medicine; Intensive care medicine; Risk analysis (engineering); Internal medicine; History","score_opus":0.11546235201575854,"score_gpt":0.35582039208908084,"score_spread":0.2403580400733223,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4315475599","genre_codex":"empirical","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8460732,0.035383735,0.046189103,0.012695297,0.0026286093,0.0011947836,0.00805977,0.00020460866,0.047570877],"genre_scores_gemma":[0.978101,0.0022053954,0.01548387,0.0016771021,0.00052817137,0.00039592918,0.00080744905,0.000026696445,0.0007742954],"study_design_codex":"observational","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9549891,0.026735235,0.005535162,0.0031692095,0.008251033,0.0013202251],"domain_scores_gemma":[0.82000065,0.12552536,0.028168092,0.009305498,0.014825821,0.0021745802],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.031256557,0.0008427503,0.0015782635,0.005048501,0.0010829386,0.0045205927,0.00167677,0.0026182875,0.0020641978],"category_scores_gemma":[0.1889427,0.0007392643,0.0013565588,0.004884031,0.0013812167,0.003141807,0.0039481604,0.0029696177,0.00047576724],"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":[0.000393292,0.0002450197,0.9437943,0.0015727608,0.0006584106,0.0001442947,0.001736394,0.0004899256,0.0003935939,0.0021391949,0.002436204,0.045996666],"study_design_scores_gemma":[0.00006887527,0.00056909444,0.9687958,0.0015749861,0.0011990076,0.001238471,0.0017273786,0.0042452808,0.001449811,0.0072402977,0.011791232,0.00009971529],"about_ca_topic_score_codex":0.009523551,"about_ca_topic_score_gemma":0.014115917,"teacher_disagreement_score":0.96874344,"about_ca_system_score_codex":0.0011812133,"about_ca_system_score_gemma":0.0027827255,"threshold_uncertainty_score":0.16530246},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4318973081","doi":"10.1136/bmjebm-2022-112053","title":"Different meta-analysis methods can change judgements about imprecision of effect estimates: a meta-epidemiological study","year":2023,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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; Impact","funders":"","keywords":"Meta-analysis; Statistics; Random effects model; Mathematics; Medicine; Restricted maximum likelihood; Maximum likelihood; Econometrics; Internal medicine","score_opus":0.9820049695252123,"score_gpt":0.7362051535903333,"score_spread":0.24579981593487898,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4318973081","genre_codex":"review","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.07259309,0.7599143,0.13563804,0.014048181,0.004762994,0.0060245413,0.0013071193,0.00047067902,0.0052410904],"genre_scores_gemma":[0.67572343,0.09723771,0.20566167,0.008544596,0.0018030117,0.009344206,0.00074471725,0.0003418459,0.00059885555],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.23611733,0.66761416,0.055996377,0.0096991295,0.029705776,0.0008672691],"domain_scores_gemma":[0.07915136,0.85974246,0.031956133,0.019031867,0.009654375,0.00046374576],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.52488333,0.0047944183,0.010054894,0.013212399,0.0017836511,0.013354538,0.0053790645,0.010224596,0.0024265368],"category_scores_gemma":[0.8112393,0.003440132,0.0466065,0.0103593785,0.0044246293,0.009911896,0.0052897,0.00863645,0.0003976717],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.018402154,0.00035817298,0.043099377,0.09420867,0.67202526,0.00079606264,0.0035138705,0.011045581,0.0012114411,0.009483041,0.0038987657,0.14195758],"study_design_scores_gemma":[0.016322209,0.005534968,0.03397552,0.096296035,0.7514195,0.0019067223,0.0017108571,0.023487408,0.0048072264,0.050429583,0.012790962,0.0013189776],"about_ca_topic_score_codex":0.0016326802,"about_ca_topic_score_gemma":0.002557819,"teacher_disagreement_score":0.9899451,"about_ca_system_score_codex":0.0050377194,"about_ca_system_score_gemma":0.003658023,"threshold_uncertainty_score":0.5859035},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4322758031","doi":"10.1136/bmjebm-2022-112098","title":"Decision Analysis in SHared decision making for Thromboprophylaxis during Pregnancy (DASH-TOP): a sequential explanatory mixed-methods pilot study","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Patient-Provider Communication in Healthcare","field":"Health Professions","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":"Mount Sinai Hospital; McMaster University; Impact","funders":"","keywords":"Decision aids; Intervention (counseling); Decision analysis; Medicine; Psychology; Nursing; Alternative medicine","score_opus":0.4146435220118736,"score_gpt":0.5623066713106735,"score_spread":0.14766314929879987,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4322758031","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9493163,0.00032494657,0.01574191,0.00052328984,0.00008887859,0.03170879,0.00030276654,0.00006478455,0.0019283255],"genre_scores_gemma":[0.8423684,0.00040647943,0.08043208,0.00045281282,0.00007529885,0.07516066,0.00018782973,0.000020981646,0.0008954786],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.95683914,0.037208002,0.0017409967,0.0013303848,0.0015493077,0.0013321851],"domain_scores_gemma":[0.8827217,0.10240949,0.005589871,0.0036960114,0.003183227,0.0023996243],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.04760683,0.0014315459,0.0012935571,0.0014828235,0.0019153266,0.0017113195,0.0019880629,0.0014521298,0.00750388],"category_scores_gemma":[0.06381873,0.0011481522,0.0016620591,0.0013449755,0.0017048995,0.0019242292,0.0031285433,0.0018589073,0.00057587575],"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":[0.0548129,0.11260321,0.039972544,0.011655538,0.0008926051,0.0011653918,0.13258937,0.0089074755,0.0034381272,0.00707823,0.0020376001,0.6248471],"study_design_scores_gemma":[0.111917935,0.5622155,0.08684803,0.006022925,0.002397778,0.001162012,0.106665224,0.052550502,0.012134931,0.02776992,0.02952949,0.0007857252],"about_ca_topic_score_codex":0.0015893665,"about_ca_topic_score_gemma":0.0022665737,"teacher_disagreement_score":0.04760683,"about_ca_system_score_codex":0.0031208806,"about_ca_system_score_gemma":0.0072076176,"threshold_uncertainty_score":0.251772},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"qualitative","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4323542543","doi":"10.1136/bmjebm-2022-112100","title":"Proposed triggers for retiring a living systematic review","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Statistical Methods in Clinical Trials","field":"Mathematics","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; Impact","funders":"","keywords":"Construct (python library); Computer science; Certainty; Actuarial science; Risk analysis (engineering); Business","score_opus":0.8168270424800806,"score_gpt":0.6492158980082676,"score_spread":0.16761114447181302,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4323542543","genre_codex":"commentary","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.010439377,0.021239351,0.03929584,0.87637264,0.020160614,0.0054031587,0.0021031054,0.0016788243,0.023307057],"genre_scores_gemma":[0.1716627,0.016837139,0.24892472,0.5127135,0.017097319,0.016001048,0.0027987994,0.0011668424,0.012797941],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.5168786,0.23788501,0.12574703,0.017590677,0.08685717,0.015041443],"domain_scores_gemma":[0.15480644,0.55006427,0.10423706,0.046042014,0.12352794,0.021322386],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.41919708,0.0017574605,0.0031519963,0.010642812,0.005892214,0.014284036,0.007963019,0.021718549,0.022086233],"category_scores_gemma":[0.7526866,0.0021120461,0.008795298,0.008895095,0.0066202395,0.01629147,0.013923759,0.015802113,0.0059628407],"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":[0.0018860637,0.00027101295,0.01933347,0.03612659,0.0022007104,0.0046476657,0.008782658,0.0026548977,0.0015796532,0.18672638,0.44823098,0.2875599],"study_design_scores_gemma":[0.0013610576,0.0008895722,0.012912495,0.04807502,0.0019729699,0.0036101623,0.005075566,0.006567005,0.002467504,0.23836938,0.6777214,0.0009778643],"about_ca_topic_score_codex":0.0032397304,"about_ca_topic_score_gemma":0.00958802,"teacher_disagreement_score":0.5808029,"about_ca_system_score_codex":0.013014653,"about_ca_system_score_gemma":0.054954488,"threshold_uncertainty_score":0.7162334},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4328055071","doi":"10.1136/bmjebm-2022-112005","title":"US Food and Drug Administration regulatory reviewer disagreements and postmarket safety actions among new therapeutics","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Pharmaceutical Economics and Policy","field":"Economics, Econometrics and Finance","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":"Dalhousie University","funders":"Canadian Institutes of Health Research","keywords":"Medicine; Food and drug administration; Drug; Pharmacology; Patient safety; Health care","score_opus":0.18732288818352935,"score_gpt":0.37053351688652847,"score_spread":0.18321062870299912,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4328055071","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.808353,0.13352297,0.011651366,0.014230839,0.0013369883,0.0013768857,0.0072363988,0.00019097897,0.022100562],"genre_scores_gemma":[0.9894985,0.0043990896,0.0022211531,0.0011555176,0.0005033646,0.0003144631,0.0012719934,0.000038407157,0.00059755536],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.7277871,0.15911782,0.049391866,0.015271131,0.045849785,0.0025823114],"domain_scores_gemma":[0.14906706,0.44996038,0.34250492,0.012685331,0.042975925,0.002806307],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.16083992,0.0005123436,0.0014244438,0.008500561,0.0011838276,0.0022530626,0.0021145782,0.002139912,0.003284586],"category_scores_gemma":[0.5066848,0.0006997475,0.0022337937,0.006557297,0.0014369383,0.003440688,0.0032359632,0.002504866,0.0005648838],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001335207,0.000087038454,0.9702059,0.0017840927,0.0013716036,0.00023230916,0.0014694205,0.00028846838,0.00012600471,0.00041751296,0.004629537,0.018052865],"study_design_scores_gemma":[0.00027563068,0.00046696397,0.975966,0.0032260055,0.0014489513,0.002077961,0.0016571912,0.0021462487,0.0007380313,0.0010987839,0.010766064,0.00013214843],"about_ca_topic_score_codex":0.004240031,"about_ca_topic_score_gemma":0.0031772433,"teacher_disagreement_score":0.8391601,"about_ca_system_score_codex":0.0024928,"about_ca_system_score_gemma":0.0028413935,"threshold_uncertainty_score":0.85061294},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4366386724","doi":"10.1136/bmjebm-2022-112070","title":"Rapid Reviews Methods Series: Involving patient and public partners, healthcare providers and policymakers as knowledge users","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Mental Health and Patient Involvement","field":"Health Professions","cited_by":48,"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 Agency for Drugs and Technologies in Health; Public Health Ontario; University of Toronto; St. Michael's Hospital; Cochrane; Public Health Agency of Canada; University of Ottawa","funders":"","keywords":"Timeline; Health care; Commission; Knowledge translation; Psychology; Knowledge management; Public relations; Management science; Computer science; Business; Political science; Engineering; Mathematics","score_opus":0.6031864510411815,"score_gpt":0.5799916421850461,"score_spread":0.02319480885613545,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4366386724","genre_codex":"protocol","genre_gemma":"methods","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0016755345,0.09414656,0.24410754,0.058253936,0.05661181,0.4857649,0.01074039,0.005084731,0.043614633],"genre_scores_gemma":[0.0037769158,0.04209061,0.53759366,0.01132641,0.0062067714,0.38403553,0.0019837096,0.0011567859,0.011829592],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.35790735,0.4793361,0.09183775,0.013301279,0.05522333,0.0023941903],"domain_scores_gemma":[0.19175161,0.62072456,0.044300206,0.057470024,0.079921484,0.005832165],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.46273714,0.0049830712,0.007117579,0.021038506,0.0046691564,0.01237796,0.009378505,0.0101018995,0.050671306],"category_scores_gemma":[0.6694083,0.004558043,0.007261568,0.02029228,0.007897388,0.016216885,0.011950829,0.01026262,0.036632273],"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":[0.00075096457,0.00020506668,0.00055573875,0.13711917,0.001127734,0.00038545206,0.007439855,0.0006607997,0.0019828991,0.032037307,0.37913212,0.43860295],"study_design_scores_gemma":[0.00066113507,0.0002849408,0.0006687911,0.06076902,0.0004594886,0.00044078272,0.0010973718,0.0005232224,0.0009283532,0.023994803,0.909966,0.00020614671],"about_ca_topic_score_codex":0.0023751468,"about_ca_topic_score_gemma":0.006569678,"teacher_disagreement_score":0.53726286,"about_ca_system_score_codex":0.012142096,"about_ca_system_score_gemma":0.05913802,"threshold_uncertainty_score":0.6625408},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"design_other","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4366386735","doi":"10.1136/bmjebm-2022-112185","title":"Rapid reviews methods series: Guidance on team considerations, study selection, data extraction and risk of bias assessment","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":78,"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":"Data extraction; Correctness; Computer science; Systematic review; Completeness (order theory); Selection (genetic algorithm); Data mining; Information retrieval; MEDLINE; Artificial intelligence; Mathematics; Algorithm","score_opus":0.9188994641379558,"score_gpt":0.6765426507253051,"score_spread":0.2423568134126507,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4366386735","genre_codex":"protocol","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00085790985,0.017875228,0.30240455,0.03245893,0.017069595,0.51822764,0.06789096,0.016372614,0.026842585],"genre_scores_gemma":[0.0011688486,0.009677381,0.5090877,0.0027150405,0.001741885,0.4573298,0.006488589,0.0020335044,0.00975727],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.5360598,0.31632477,0.10645171,0.006594282,0.03226621,0.0023032862],"domain_scores_gemma":[0.24392892,0.5496194,0.051641434,0.04512235,0.105037965,0.0046500145],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.3852699,0.0059487144,0.011508372,0.022639886,0.0028355832,0.009906277,0.009126025,0.010461362,0.16644643],"category_scores_gemma":[0.65114707,0.010998541,0.014908863,0.020515319,0.0043372638,0.008315594,0.008311663,0.014030779,0.13228066],"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":[0.0011112499,0.00016968431,0.00032747956,0.09639452,0.00093570736,0.00021742562,0.0016996927,0.000894719,0.00095920195,0.010791734,0.69835037,0.18814814],"study_design_scores_gemma":[0.0034760055,0.00029884433,0.0014597103,0.06357348,0.0007918326,0.00041143992,0.00045122227,0.0017884012,0.0013996137,0.025724933,0.90021116,0.00041332707],"about_ca_topic_score_codex":0.004873833,"about_ca_topic_score_gemma":0.010417989,"teacher_disagreement_score":0.6147301,"about_ca_system_score_codex":0.00747936,"about_ca_system_score_gemma":0.06292624,"threshold_uncertainty_score":0.75807166},"labels":[],"label_agreement":null},{"id":"W4366386740","doi":"10.1136/bmjebm-2022-112136","title":"Implementing hierarchical network meta-analysis incorporating exchangeable dose effects compared to standard hierarchical network meta-analysis","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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; University of Calgary; St. Michael's Hospital","funders":"Medical Research Council","keywords":"Meta-analysis; Network analysis; Computer science; Multilevel model; Medicine; Internal medicine; Engineering; Machine learning","score_opus":0.8066750011977708,"score_gpt":0.5551599690728524,"score_spread":0.25151503212491844,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4366386740","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.022357354,0.011405422,0.95081824,0.002640032,0.0014041162,0.003667185,0.002247545,0.0024182529,0.0030418765],"genre_scores_gemma":[0.25971928,0.0021146082,0.7288379,0.0009819353,0.00031450845,0.005023648,0.0011227375,0.00060262566,0.001282708],"study_design_codex":"meta_analysis","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.86749005,0.12043659,0.0045422325,0.0042492156,0.0027372222,0.000544755],"domain_scores_gemma":[0.82897115,0.15347062,0.0034209667,0.0107836565,0.0027245444,0.0006290171],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.08658935,0.0032618712,0.006312381,0.0040016463,0.0008876738,0.0039677294,0.0038551546,0.0024474103,0.012410646],"category_scores_gemma":[0.2457097,0.0024098286,0.016765568,0.0034862708,0.0007608533,0.0056558843,0.0047592027,0.0045578764,0.00078983966],"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":[0.02112908,0.0008158832,0.010798827,0.025088953,0.33058596,0.0006691461,0.0010666096,0.26253885,0.0018138494,0.03744081,0.0112438705,0.29680815],"study_design_scores_gemma":[0.016320856,0.0025159607,0.004297286,0.0023985861,0.1651163,0.00024642126,0.00018509534,0.66082686,0.0016354986,0.13565241,0.010497493,0.0003073376],"about_ca_topic_score_codex":0.00905247,"about_ca_topic_score_gemma":0.011741817,"teacher_disagreement_score":0.91341066,"about_ca_system_score_codex":0.0021798508,"about_ca_system_score_gemma":0.0044570165,"threshold_uncertainty_score":0.45793372},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"simulation_or_modeling","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4366386744","doi":"10.1136/bmjebm-2022-112111","title":"Rapid reviews methods series: Guidance on assessing the certainty of evidence","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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":"McMaster University; Impact","funders":"Public Health Agency","keywords":"Grading (engineering); Systematic review; Computer science; Rating system; Certainty; Management science; Medicine; Actuarial science; MEDLINE; Engineering; Mathematics; Business; Political science","score_opus":0.9546345879208633,"score_gpt":0.6949051985894442,"score_spread":0.2597293893314191,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4366386744","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0011019373,0.088481516,0.40947878,0.0478053,0.027798299,0.27247205,0.06777643,0.010140937,0.07494468],"genre_scores_gemma":[0.0029646952,0.040429503,0.6968956,0.0048406622,0.0022450958,0.23071359,0.009766254,0.0021402615,0.010004425],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.42795932,0.3825945,0.11792526,0.0060842503,0.06307795,0.002358705],"domain_scores_gemma":[0.24152313,0.543166,0.0509262,0.040211935,0.11998892,0.0041838447],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.38939396,0.0064374343,0.012766715,0.027645554,0.0025336456,0.012951415,0.01166051,0.013225298,0.11768156],"category_scores_gemma":[0.7202661,0.0090915905,0.021350417,0.022634288,0.0039684908,0.01031031,0.011936654,0.016028095,0.085586496],"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":[0.0009608174,0.00012012022,0.00039386319,0.11768626,0.0018255521,0.00018422377,0.0013563549,0.0009837855,0.0005025557,0.021263583,0.52641255,0.32831043],"study_design_scores_gemma":[0.0022583678,0.0002511105,0.001530731,0.10595478,0.0015430044,0.00042856944,0.00035921478,0.0013455001,0.0010603053,0.029610936,0.8552622,0.0003953025],"about_ca_topic_score_codex":0.007355266,"about_ca_topic_score_gemma":0.008736609,"teacher_disagreement_score":0.6106061,"about_ca_system_score_codex":0.009554849,"about_ca_system_score_gemma":0.057185877,"threshold_uncertainty_score":0.75298595},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W4367675492","doi":"10.1136/bmjebm-2023-112343","title":"Correspondence on ‘The limits of shared decision making’ by Elwyn<i>et al</i>: We call for a more nuanced approach","year":2023,"lang":"en","type":"letter","venue":"BMJ evidence-based medicine","topic":"Patient-Provider Communication in Healthcare","field":"Health Professions","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":"Carleton University; University of Toronto","funders":"","keywords":"Computer science; Psychology; Sociology","score_opus":0.5148313715036523,"score_gpt":0.52626582144505,"score_spread":0.01143444994139764,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4367675492","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000043733602,0.00030838486,0.000046340585,0.9872313,0.011710921,0.00000511134,0.000010011031,0.0000067357378,0.00063745567],"genre_scores_gemma":[0.00070037745,0.00021048546,0.000102477505,0.97970027,0.018016217,0.000023135688,0.0000064258966,0.000018507651,0.00122221],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.93057936,0.030417603,0.0077343024,0.005301477,0.019416368,0.006550846],"domain_scores_gemma":[0.73615575,0.19531994,0.013930163,0.0064600497,0.027274763,0.020859422],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.045617238,0.0014912033,0.003230491,0.002088956,0.01535763,0.021367364,0.0060546347,0.1459257,0.017406441],"category_scores_gemma":[0.2476516,0.0025816814,0.003342454,0.0027967412,0.021215416,0.020390306,0.008766197,0.14542325,0.010850377],"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":[0.000026007527,0.000012161127,0.000107683845,0.0000730777,0.000013887754,0.0003004071,0.00047461502,0.000024470939,0.00004214805,0.003030414,0.99352807,0.0023670301],"study_design_scores_gemma":[0.00015632856,0.00003702415,0.0005567514,0.0011949114,0.00003634538,0.0008530701,0.0020297265,0.00030129444,0.00014346761,0.018942403,0.97558683,0.00016183502],"about_ca_topic_score_codex":0.015563866,"about_ca_topic_score_gemma":0.025476316,"teacher_disagreement_score":0.1459257,"about_ca_system_score_codex":0.013533797,"about_ca_system_score_gemma":0.021453833,"threshold_uncertainty_score":0.24124992},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"other","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":"low"}],"label_agreement":"agree"},{"id":"W4376639745","doi":"10.1136/bmjebm-2022-112215","title":"Catalogue of bias: novelty bias","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Computational Drug Discovery Methods","field":"Computer Science","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 Toronto; St. Michael's Hospital","funders":"","keywords":"Novelty; Computer science; Psychology; Social psychology","score_opus":0.3599031462954102,"score_gpt":0.4300701984081084,"score_spread":0.07016705211269819,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4376639745","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.020499812,0.08399442,0.5389443,0.059593484,0.03756754,0.0934605,0.04358701,0.004227433,0.118125536],"genre_scores_gemma":[0.3479334,0.024703637,0.3414507,0.069927305,0.01371895,0.16531134,0.011977314,0.0018365222,0.023140887],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.35876435,0.330472,0.16704181,0.03784878,0.102421165,0.0034519173],"domain_scores_gemma":[0.1772877,0.53183746,0.087781586,0.13933443,0.061327703,0.0024311284],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.3506263,0.0031173192,0.010167757,0.0142475655,0.0043192953,0.011591269,0.006975581,0.013296491,0.043148562],"category_scores_gemma":[0.7201375,0.00238134,0.008403885,0.012923769,0.009958203,0.012458643,0.01088416,0.009462006,0.008669836],"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":[0.004108486,0.00029693492,0.026095647,0.120936304,0.011828769,0.002123167,0.01122524,0.0019189814,0.001771249,0.20051838,0.16871147,0.45046538],"study_design_scores_gemma":[0.002332451,0.000558004,0.015415686,0.11620298,0.0067124423,0.006634113,0.0022702971,0.008927713,0.0048185624,0.37322754,0.46197167,0.0009285909],"about_ca_topic_score_codex":0.002332044,"about_ca_topic_score_gemma":0.0019629297,"teacher_disagreement_score":0.6493737,"about_ca_system_score_codex":0.0079795765,"about_ca_system_score_gemma":0.011470716,"threshold_uncertainty_score":0.8007934},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4378806110","doi":"10.1136/bmjebm-2022-112200","title":"Plain language summaries: enhancing patient-centred care and improving accessibility of health research","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Health Literacy and Information Accessibility","field":"Health Professions","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":"Ontario Stroke Network; University Health Network; University of British Columbia","funders":"","keywords":"Plain language; Health care; Medicine; Computer science; Nursing; Political science","score_opus":0.264962570399348,"score_gpt":0.5450406684158028,"score_spread":0.2800780980164548,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4378806110","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.30494156,0.073936746,0.09416919,0.33962178,0.008811609,0.0046969145,0.009851387,0.008189313,0.15578142],"genre_scores_gemma":[0.61969906,0.061211605,0.24283749,0.040313095,0.00859252,0.0027209842,0.0045002205,0.0009844977,0.019140553],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.97594255,0.016503336,0.003943908,0.00058029726,0.0027178382,0.00031209373],"domain_scores_gemma":[0.8163881,0.1363764,0.02446415,0.005856057,0.01051154,0.0064037116],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.02136286,0.00046334983,0.0011016504,0.0024017377,0.0007962184,0.005294188,0.0011034596,0.001981834,0.022817198],"category_scores_gemma":[0.20306848,0.0006100397,0.0010019109,0.0017491749,0.00052718335,0.006600228,0.0039477516,0.0018747787,0.004209535],"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":[0.0019345201,0.0013770934,0.016554443,0.01058473,0.00029563697,0.0006382447,0.007008804,0.00044327977,0.0016990632,0.0033927173,0.15901418,0.7970573],"study_design_scores_gemma":[0.0063049584,0.0076732505,0.20214352,0.030256825,0.0033412839,0.008204596,0.016403066,0.0042324727,0.010412769,0.05639978,0.65365857,0.00096890796],"about_ca_topic_score_codex":0.0008482824,"about_ca_topic_score_gemma":0.0024242639,"teacher_disagreement_score":0.97863716,"about_ca_system_score_codex":0.00079716166,"about_ca_system_score_gemma":0.0038545288,"threshold_uncertainty_score":0.112978935},"labels":[{"model":"gemma","categories":["scholarly_communication"],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["scholarly_communication","open_science"],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4384663396","doi":"10.1136/bmjebm-2022-111921","title":"Reducing the residue of retractions in evidence synthesis: ways to minimise inappropriate citation and use of retracted data","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Academic integrity and plagiarism","field":"Social Sciences","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 Regina","funders":"Alfred P. Sloan Foundation","keywords":"Citation; Systematic review; MEDLINE; Computer science; Medical literature; Data science; Medicine; Political science; Library science; Law","score_opus":0.35376102033369716,"score_gpt":0.42625773168766457,"score_spread":0.0724967113539674,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4384663396","genre_codex":"methods","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"reproducibility","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reproducibility","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.016126912,0.034930427,0.5921119,0.28099477,0.026113845,0.025872981,0.0016428118,0.005749232,0.016457161],"genre_scores_gemma":[0.09770248,0.012550441,0.8108369,0.032694466,0.011183909,0.024320882,0.0012222149,0.002552387,0.006936308],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.08465452,0.6081745,0.21870518,0.014639262,0.07086193,0.0029646272],"domain_scores_gemma":[0.010268321,0.76090086,0.08196793,0.07945542,0.06485247,0.0025550013],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.8411662,0.005878992,0.013374125,0.047464512,0.011588958,0.039131165,0.01748821,0.029331936,0.016474787],"category_scores_gemma":[0.9667117,0.010106288,0.013660851,0.045974668,0.022361616,0.049743045,0.032424826,0.02941531,0.013261273],"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":[0.0027841204,0.0005057997,0.009439609,0.035853572,0.0034738916,0.0012884554,0.040269356,0.0034943847,0.0025324936,0.031094272,0.08325893,0.786005],"study_design_scores_gemma":[0.004970047,0.0021562604,0.020364178,0.120363235,0.007630837,0.003457134,0.01875216,0.028997509,0.012194821,0.3198967,0.45870796,0.0025091313],"about_ca_topic_score_codex":0.0040390384,"about_ca_topic_score_gemma":0.005126503,"teacher_disagreement_score":0.9706681,"about_ca_system_score_codex":0.015374174,"about_ca_system_score_gemma":0.07406521,"threshold_uncertainty_score":0.1958704},"labels":[{"model":"gemma","categories":["metaresearch","research_integrity"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","research_integrity"],"domain":"methods","study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4385063979","doi":"10.1136/bmjebm-2023-112270","title":"High-value care education can learn from the evidence-based medicine movement: moving beyond competencies and curricula to culture","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Healthcare cost, quality, practices","field":"Health Professions","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":"Health Sciences Centre; Women in Science and Engineering Newfoundland and Labrador; University of Toronto; Sunnybrook Health Science Centre","funders":"","keywords":"Curriculum; Movement (music); Value (mathematics); Medical education; Medicine; Pedagogy; Psychology; Computer science; Art","score_opus":0.46851158345077826,"score_gpt":0.5371929324341386,"score_spread":0.06868134898336031,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4385063979","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.037803777,0.008912083,0.015382339,0.83853203,0.004627374,0.00019474875,0.00008001846,0.00018352951,0.09428412],"genre_scores_gemma":[0.66375726,0.02775396,0.07589371,0.20619957,0.0039188103,0.0007399984,0.00025360592,0.0002315099,0.02125158],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9845697,0.007464261,0.0006929687,0.00060444576,0.004956311,0.0017122949],"domain_scores_gemma":[0.88134867,0.058403328,0.006564293,0.0060458337,0.010596546,0.037041407],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.034462843,0.00046395644,0.0006413962,0.0021579496,0.0035700612,0.015521343,0.0020516124,0.0063505825,0.010883463],"category_scores_gemma":[0.085344546,0.0005628703,0.0008078915,0.0011686232,0.008610894,0.0133949425,0.016836617,0.01224041,0.002789811],"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":[0.00007772915,0.0019901283,0.022196976,0.0024419865,0.00013151535,0.00034711015,0.031600982,0.0006634178,0.0011975187,0.12907317,0.11363404,0.6966455],"study_design_scores_gemma":[0.00020285234,0.0006653505,0.05180903,0.020865006,0.00013060698,0.0007608258,0.05934783,0.001119644,0.0025668177,0.45964876,0.40258744,0.00029578555],"about_ca_topic_score_codex":0.0050686984,"about_ca_topic_score_gemma":0.016671253,"teacher_disagreement_score":0.96553713,"about_ca_system_score_codex":0.0047624153,"about_ca_system_score_gemma":0.043199975,"threshold_uncertainty_score":0.18225908},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"theoretical_or_conceptual","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":"medium"}],"label_agreement":"split"},{"id":"W4385064669","doi":"10.1136/bmjebm-2023-112344","title":"Ten resources for understanding bias in health research: EBM live workshop 2022","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","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; University of Alberta; University of Ottawa; University of Toronto","funders":"","keywords":"Psychology","score_opus":0.8979152418329284,"score_gpt":0.6624939927835675,"score_spread":0.2354212490493609,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4385064669","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0028416887,0.019850893,0.021906352,0.91538155,0.018665766,0.002246707,0.0013681541,0.00052460365,0.017214248],"genre_scores_gemma":[0.08637768,0.045403507,0.26358792,0.48251572,0.023091357,0.02080308,0.005869715,0.0020591975,0.07029187],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.86890215,0.093767546,0.008175667,0.0028716046,0.019855112,0.0064280173],"domain_scores_gemma":[0.6491837,0.26450846,0.0067202006,0.011843919,0.03741998,0.0303238],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.28578338,0.0019383702,0.0023536158,0.0055617513,0.0059650172,0.0154725285,0.008766931,0.0277942,0.039093513],"category_scores_gemma":[0.29876035,0.0024368428,0.0039178333,0.0021423886,0.0070483005,0.015595496,0.04716118,0.030968336,0.006655082],"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":[0.00059496827,0.000713396,0.0011485942,0.0062744017,0.00031658358,0.0006750827,0.01214706,0.0008484194,0.0011953529,0.024622781,0.56761456,0.38384885],"study_design_scores_gemma":[0.0003560941,0.00015667295,0.0016556937,0.018496094,0.00017564301,0.0002432632,0.008618353,0.00079517456,0.0010904623,0.053765338,0.9144862,0.00016099137],"about_ca_topic_score_codex":0.007607536,"about_ca_topic_score_gemma":0.015467205,"teacher_disagreement_score":0.7142166,"about_ca_system_score_codex":0.012666935,"about_ca_system_score_gemma":0.07775483,"threshold_uncertainty_score":0.88075626},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"other","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":"split"},{"id":"W4385064816","doi":"10.1136/bmjebm-2023-112266","title":"Choosing wisely 10 years later: reflection and looking ahead","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Healthcare cost, quality, practices","field":"Health Professions","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 Toronto","funders":"","keywords":"Reflection (computer programming); Look-ahead; Computer science; Algorithm; Programming language","score_opus":0.7360109360885645,"score_gpt":0.6133176081544891,"score_spread":0.12269332793407539,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4385064816","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0023634958,0.003186728,0.00091182854,0.97652644,0.014699945,0.00002188158,0.000037937716,0.0000329646,0.0022187429],"genre_scores_gemma":[0.09075049,0.011066305,0.008226695,0.86565304,0.015784418,0.000227167,0.00015691116,0.00023080048,0.007904151],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.93732876,0.037658587,0.00409323,0.0031450577,0.010023968,0.0077504334],"domain_scores_gemma":[0.7793556,0.09131614,0.009517556,0.0067193448,0.052528046,0.060563337],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.071943715,0.0011000711,0.0020314897,0.0017366525,0.013862671,0.019679891,0.0053459276,0.037229892,0.012936172],"category_scores_gemma":[0.23386595,0.001069075,0.002790766,0.001441897,0.014432238,0.035126638,0.01449101,0.11009045,0.0063477457],"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":[0.00018244857,0.0007943974,0.0065981206,0.0006505723,0.00022035891,0.0019969328,0.055114772,0.00025240312,0.000592736,0.021544095,0.84315234,0.06890082],"study_design_scores_gemma":[0.00013630597,0.0003000029,0.0047076675,0.004279346,0.00016743393,0.0023082262,0.23661765,0.0005849092,0.00076138275,0.07919934,0.6703675,0.00057026616],"about_ca_topic_score_codex":0.009612484,"about_ca_topic_score_gemma":0.022857862,"teacher_disagreement_score":0.071943715,"about_ca_system_score_codex":0.013221876,"about_ca_system_score_gemma":0.040320236,"threshold_uncertainty_score":0.38047928},"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":"medium"}],"label_agreement":"agree"},{"id":"W4385270131","doi":"10.1136/bmjebm-2023-112482","title":"Exploring advanced methods for network meta-analysis","year":2023,"lang":"en","type":"editorial","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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":"Institute for Work & Health; St. Michael's Hospital","funders":"","keywords":"Network analysis; Computer science; Engineering","score_opus":0.9687966746492905,"score_gpt":0.6861621719315322,"score_spread":0.28263450271775836,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4385270131","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0002347539,0.0949126,0.04082438,0.059183266,0.80011356,0.000678939,0.0012144477,0.0010813528,0.0017567546],"genre_scores_gemma":[0.007398041,0.08385216,0.08202378,0.05635626,0.75721383,0.0029621979,0.00076994504,0.0015243747,0.007899409],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.90098953,0.07152584,0.008926016,0.0026808376,0.015331257,0.00054653635],"domain_scores_gemma":[0.4431702,0.5048618,0.0063416427,0.0076260525,0.035614543,0.002385676],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.106610686,0.0059439004,0.011533291,0.00963712,0.0019361215,0.01508748,0.008928819,0.015145395,0.016272003],"category_scores_gemma":[0.45522738,0.0032514278,0.011547449,0.008108723,0.0038702402,0.0070793102,0.0040159957,0.02575446,0.005869213],"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":[0.00098986,0.000054408487,0.00015453795,0.026093943,0.009585236,0.00023885415,0.00018838797,0.0013464863,0.00020389071,0.008024836,0.8968332,0.056286477],"study_design_scores_gemma":[0.008480045,0.00041083206,0.0011919924,0.040165916,0.033786632,0.0005863646,0.00022200613,0.028030152,0.0008352428,0.2039321,0.68173015,0.000628514],"about_ca_topic_score_codex":0.0035735688,"about_ca_topic_score_gemma":0.007127847,"teacher_disagreement_score":0.89338934,"about_ca_system_score_codex":0.004011615,"about_ca_system_score_gemma":0.008274493,"threshold_uncertainty_score":0.563818},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","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":"split"},{"id":"W4385863795","doi":"10.1136/bmjebm-2023-112271","title":"Reducing low value care: opportunities and challenges for Choosing Wisely campaigns","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Healthcare cost, quality, practices","field":"Health Professions","cited_by":11,"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; Ottawa Hospital; University of Ottawa","funders":"Canadian Institutes of Health Research","keywords":"Value (mathematics); Business; Computer science","score_opus":0.8963123753308139,"score_gpt":0.6013933185560272,"score_spread":0.2949190567747867,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4385863795","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0034379254,0.0041457727,0.00086997426,0.98274773,0.0015126049,0.000018642824,0.000018354178,0.000023742063,0.0072252243],"genre_scores_gemma":[0.33300555,0.027878938,0.018961329,0.5994463,0.0107991835,0.0005690588,0.00011489708,0.0002711623,0.008953538],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.93344945,0.043694347,0.0018211423,0.0018097275,0.013581588,0.0056437105],"domain_scores_gemma":[0.7636189,0.15871435,0.009781028,0.0038145978,0.020640941,0.043430082],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.06466514,0.0007229008,0.0015510558,0.0022389211,0.012354763,0.024381965,0.00369695,0.018573994,0.023342779],"category_scores_gemma":[0.16431911,0.0008001263,0.0011606098,0.0024033699,0.015091675,0.019639207,0.0118205715,0.024103124,0.0027180742],"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":[0.00021876907,0.0010476379,0.011071149,0.0019607502,0.00024223636,0.00047728227,0.020511877,0.0003539096,0.0006173228,0.086849876,0.5176759,0.35897338],"study_design_scores_gemma":[0.0003844486,0.00045708969,0.014384422,0.013376498,0.0002095347,0.00049943704,0.09556967,0.0014792478,0.0006032493,0.36997214,0.50269926,0.0003650771],"about_ca_topic_score_codex":0.00960146,"about_ca_topic_score_gemma":0.026092583,"teacher_disagreement_score":0.06466514,"about_ca_system_score_codex":0.008806071,"about_ca_system_score_gemma":0.042930435,"threshold_uncertainty_score":0.341986},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"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":"W4386767565","doi":"10.1136/bmjebm-2023-112582","title":"Enhanced access but increased testing with virtual care: insights from a multinational survey","year":2023,"lang":"en","type":"letter","venue":"BMJ evidence-based medicine","topic":"Telemedicine and Telehealth Implementation","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; Canadian Institute for Health Information; University of Manitoba; Beef Farmers of Ontario","funders":"","keywords":"Multinational corporation; Business; Survey research; Computer science; Business administration","score_opus":0.18317465042704004,"score_gpt":0.41496418873076385,"score_spread":0.2317895383037238,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4386767565","genre_codex":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7323291,0.00051817077,0.00040609136,0.2554598,0.0002756715,0.000036680594,0.0006647084,0.0000138100095,0.010296078],"genre_scores_gemma":[0.9172192,0.000641999,0.00047258803,0.07896411,0.0003033063,0.00006357355,0.00015981533,0.000022692308,0.0021527254],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9938279,0.0033126594,0.00037320663,0.00037222452,0.0010355142,0.0010785614],"domain_scores_gemma":[0.95568043,0.02595706,0.008186275,0.0019118144,0.0032814906,0.0049829376],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.009078154,0.00014267232,0.00027520067,0.00031580942,0.0014213881,0.0021304789,0.0008182977,0.0034844324,0.0061664577],"category_scores_gemma":[0.035599396,0.0002610086,0.00029265656,0.001345242,0.0014193092,0.0021782552,0.0012752181,0.004065515,0.00043377705],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00034060044,0.00061193766,0.90886545,0.00017895433,0.00009031471,0.0018586633,0.010838774,0.00009452957,0.00058977737,0.0015541158,0.041450202,0.03352671],"study_design_scores_gemma":[0.000071743365,0.00034933147,0.9435996,0.00039332054,0.000071356204,0.0011584737,0.03243962,0.00046084562,0.00019970855,0.0010929193,0.020123953,0.000039119368],"about_ca_topic_score_codex":0.07246515,"about_ca_topic_score_gemma":0.08767375,"teacher_disagreement_score":0.07246515,"about_ca_system_score_codex":0.0016017961,"about_ca_system_score_gemma":0.0037370638,"threshold_uncertainty_score":0.14408672},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4387504423","doi":"10.1136/bmjebm-2023-112352","title":"Making sustainable healthcare decisions: three turns towards sustainable guidelines","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Agriculture Sustainability and Environmental Impact","field":"Environmental Science","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":"McMaster University; Impact","funders":"FP7 People: Marie-Curie Actions; European Commission","keywords":"Health care; Business; Environmental planning; Environmental resource management; Geography; Environmental science; Economics; Economic growth","score_opus":0.19240987544358096,"score_gpt":0.416253864773184,"score_spread":0.22384398932960303,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4387504423","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00051552354,0.0050231474,0.007652566,0.9736183,0.0026328687,0.000060649487,0.00007138375,0.00007505221,0.010350352],"genre_scores_gemma":[0.17669244,0.074771605,0.18808936,0.52412575,0.020432454,0.0012951331,0.0008315344,0.00090705615,0.012854651],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.85228974,0.10485651,0.010676377,0.0023855139,0.024493735,0.005298112],"domain_scores_gemma":[0.65867305,0.27258563,0.007974589,0.008621665,0.032820657,0.019324515],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.17766924,0.0012147021,0.0028472603,0.0052669547,0.0094446065,0.03464827,0.006512788,0.036550328,0.0152306305],"category_scores_gemma":[0.27456307,0.0017868349,0.0023263013,0.004050935,0.042164605,0.030519953,0.020081365,0.044678774,0.0058898875],"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":[0.00008974771,0.00013771928,0.0007048562,0.0016335383,0.00023181218,0.0006318989,0.012692483,0.0014350241,0.00032569177,0.37498626,0.49404743,0.11308356],"study_design_scores_gemma":[0.000048340018,0.00004134816,0.00043358025,0.0043763835,0.00004932612,0.00011567131,0.008534756,0.00047131846,0.000211028,0.7162989,0.26928636,0.0001329951],"about_ca_topic_score_codex":0.008704674,"about_ca_topic_score_gemma":0.017602542,"teacher_disagreement_score":0.17766924,"about_ca_system_score_codex":0.011944339,"about_ca_system_score_gemma":0.059917513,"threshold_uncertainty_score":0.93961596},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4387558756","doi":"10.1136/bmjebm-2022-112125","title":"Can a replication revolution resolve the duplication crisis in systematic reviews?","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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":"Université de Montréal; Cochrane; Ottawa Hospital; Bruyère; University of Ottawa","funders":"","keywords":"Replication (statistics); Gene duplication; Political science; Biology; Genetics; Virology","score_opus":0.831327983766012,"score_gpt":0.5804957268733995,"score_spread":0.25083225689261257,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4387558756","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"reproducibility","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reproducibility","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.001562606,0.05504706,0.037357483,0.86418974,0.035901252,0.0005757169,0.00053749944,0.0009146488,0.00391399],"genre_scores_gemma":[0.11826889,0.052858524,0.19793919,0.52597374,0.091453105,0.005577127,0.0015643444,0.002011637,0.0043534692],"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.3341709,0.5007023,0.05442533,0.021461887,0.08410823,0.0051313485],"domain_scores_gemma":[0.03472396,0.7973827,0.024431946,0.082434684,0.055743415,0.0052833003],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.7418719,0.0050481916,0.015418737,0.016846376,0.007202324,0.024983738,0.01619554,0.039216418,0.017151996],"category_scores_gemma":[0.9253106,0.005797591,0.011334708,0.013363978,0.032027394,0.0751094,0.02230235,0.034568444,0.0072404062],"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":[0.0038320213,0.0005927726,0.0064913314,0.028563816,0.0073436513,0.00094162766,0.011532427,0.0027626322,0.0011518922,0.17426994,0.38862732,0.3738905],"study_design_scores_gemma":[0.0034346173,0.0007834784,0.0031883286,0.029786738,0.0031693375,0.00073221873,0.0026616694,0.006372314,0.0010040288,0.73023844,0.21796589,0.00066304597],"about_ca_topic_score_codex":0.009428742,"about_ca_topic_score_gemma":0.013155083,"teacher_disagreement_score":0.2581281,"about_ca_system_score_codex":0.012436587,"about_ca_system_score_gemma":0.06916459,"threshold_uncertainty_score":0.31831795},"labels":[],"label_agreement":null},{"id":"W4388006999","doi":"10.1136/bmjebm-2023-112475","title":"How to use the regulatory data from Health Canada for secondary analyses on new drugs, biologics and vaccines","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Clinical practice guidelines implementation","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":true,"ca_institutions":"Dalhousie University","funders":"Dalhousie University","keywords":"Medicine; Public health; Internet privacy; Business; Computer science","score_opus":0.7468377814715961,"score_gpt":0.5827913881962639,"score_spread":0.16404639327533221,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4388006999","genre_codex":"commentary","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0041891797,0.01403573,0.107490554,0.3253722,0.0131196715,0.0942071,0.28931007,0.012661047,0.13961443],"genre_scores_gemma":[0.026214005,0.013798913,0.7113831,0.07368306,0.003054955,0.09133817,0.04629453,0.0052906754,0.028942572],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.63187444,0.18025571,0.083856896,0.007515632,0.09033855,0.006158773],"domain_scores_gemma":[0.19900413,0.40354058,0.04280556,0.07751775,0.26811874,0.009013189],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.31042376,0.0023337689,0.0067219925,0.029423425,0.004439494,0.01978562,0.0068896315,0.009567846,0.06469373],"category_scores_gemma":[0.73580444,0.005446836,0.008098961,0.028549466,0.0054984814,0.011891642,0.0077721034,0.0076710926,0.04787107],"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":[0.00056085055,0.00012489373,0.0025112503,0.008774149,0.00047199466,0.00016456771,0.0014763928,0.0008049454,0.0004221741,0.005486634,0.83598745,0.14321466],"study_design_scores_gemma":[0.0032273687,0.0001226986,0.011424112,0.02795721,0.0007559967,0.00016473909,0.0014733404,0.002548979,0.0022100299,0.019805923,0.929512,0.0007974498],"about_ca_topic_score_codex":0.52047825,"about_ca_topic_score_gemma":0.6435565,"teacher_disagreement_score":0.9813545,"about_ca_system_score_codex":0.018645542,"about_ca_system_score_gemma":0.20340805,"threshold_uncertainty_score":0.9646915},"labels":[{"model":"gemma","categories":["metaresearch","open_science"],"domain":"reproducibility","study_design":"not_applicable","genre":"methods","about_ca_system":true,"about_ca_topic":true,"confidence":"high"},{"model":"gpt","categories":["metaresearch","open_science"],"domain":"methods","study_design":"design_other","genre":"methods","about_ca_system":false,"about_ca_topic":true,"confidence":"high"}],"label_agreement":"split"},{"id":"W4389207515","doi":"10.1136/bmjebm-2023-112397","title":"Integrating a climate lens into the design of education programmes for health professionals","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Innovations in Medical Education","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; McMaster University; University of Waterloo","funders":"","keywords":"Lens (geology); Health professionals; Through-the-lens metering; Medical education; Sociology; Engineering ethics; Public relations; Engineering; Political science; Medicine; Health care","score_opus":0.16808932245712369,"score_gpt":0.5137538397847501,"score_spread":0.3456645173276264,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4389207515","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.039166674,0.014986135,0.5128949,0.13770176,0.0049400125,0.0027267633,0.00041965485,0.00091577333,0.2862482],"genre_scores_gemma":[0.5624514,0.0074396883,0.39479962,0.012994802,0.0008811314,0.0037330983,0.0001387087,0.0003528938,0.017208664],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.9827252,0.013908608,0.00036148948,0.00046563658,0.0013421478,0.0011968549],"domain_scores_gemma":[0.9801332,0.013891098,0.0009840109,0.0006015513,0.0016242159,0.0027659792],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.016311258,0.0008920205,0.00044437635,0.0021632346,0.0042956257,0.008313875,0.0024076893,0.0035476764,0.022630919],"category_scores_gemma":[0.023648435,0.0006076854,0.00079009816,0.0012295002,0.005608305,0.006328178,0.006589491,0.004013595,0.0016823323],"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":[0.0002868521,0.0016965256,0.005917027,0.0033330068,0.00010367215,0.0004146681,0.023386935,0.015485877,0.0023521173,0.6063013,0.032743927,0.3079781],"study_design_scores_gemma":[0.00029022104,0.0009439263,0.006277698,0.005608872,0.00010624758,0.0002625896,0.025494842,0.008163355,0.002502017,0.500158,0.4500704,0.00012185597],"about_ca_topic_score_codex":0.004424171,"about_ca_topic_score_gemma":0.010561968,"teacher_disagreement_score":0.022630919,"about_ca_system_score_codex":0.009587118,"about_ca_system_score_gemma":0.015340976,"threshold_uncertainty_score":0.08626318},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4390977446","doi":"10.1136/bmjebm-2023-112608","title":"Preventing overdiagnosis and overuse: proposed guidance for guideline panels","year":2024,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Healthcare cost, quality, practices","field":"Health Professions","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":"Overdiagnosis; Guideline; Medicine; Intensive care medicine; Medical physics; Internal medicine; Pathology","score_opus":0.7344073371348866,"score_gpt":0.6267851970459801,"score_spread":0.10762214008890647,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4390977446","genre_codex":"commentary","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.002048753,0.008776927,0.038381718,0.9202328,0.0067524514,0.004104223,0.0010521205,0.00086006714,0.017791081],"genre_scores_gemma":[0.036841795,0.0116758,0.66699505,0.255567,0.008382246,0.010087477,0.0018138057,0.0003596005,0.008277183],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.8108451,0.11156862,0.034669273,0.0056051053,0.030656775,0.006655196],"domain_scores_gemma":[0.5152479,0.28258005,0.023277333,0.012004522,0.14437032,0.022519818],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.15433931,0.003296592,0.00711014,0.008803296,0.007505365,0.018125474,0.013423682,0.06523141,0.008786165],"category_scores_gemma":[0.4554431,0.0030671062,0.012829623,0.006598518,0.007690274,0.0116294585,0.009224976,0.033885706,0.007633539],"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":[0.00029439272,0.0009492321,0.008775497,0.004201945,0.00067351025,0.0010866411,0.0041943355,0.0030292242,0.0007731553,0.038421683,0.77159196,0.1660085],"study_design_scores_gemma":[0.003925166,0.0011100725,0.018351106,0.061046105,0.0033888686,0.0023757631,0.0054714093,0.018172946,0.002334214,0.24554977,0.6367864,0.0014882075],"about_ca_topic_score_codex":0.028614646,"about_ca_topic_score_gemma":0.06854515,"teacher_disagreement_score":0.8456607,"about_ca_system_score_codex":0.010405604,"about_ca_system_score_gemma":0.08662247,"threshold_uncertainty_score":0.81623405},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","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":"low"}],"label_agreement":"split"},{"id":"W4391022170","doi":"10.1136/bmjebm-2023-112784","title":"Development of the Reporting Infographics and Visual Abstracts of Comparative studies (RIVA-C) checklist and guide","year":2024,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Delphi Technique in Research","field":"Social Sciences","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 British Columbia","funders":"National Health and Medical Research Council; Medical Research Council","keywords":"Infographic; Checklist; Psychological intervention; Delphi method; Reading (process); Medical education; Psychology; Computer science; Medicine","score_opus":0.48597114545384495,"score_gpt":0.5992018600550433,"score_spread":0.11323071460119832,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4391022170","genre_codex":"protocol","genre_gemma":"methods","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.007750813,0.004403376,0.27359605,0.02783726,0.007740705,0.6328737,0.02653548,0.0050508063,0.014211896],"genre_scores_gemma":[0.0053827227,0.0020478223,0.4211148,0.0030636692,0.0006011705,0.55651474,0.0075478936,0.0005406055,0.0031865279],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.58116686,0.26796347,0.10191294,0.005846127,0.03934786,0.0037627693],"domain_scores_gemma":[0.2702932,0.3453095,0.054822702,0.04814731,0.27105796,0.010369378],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.41096503,0.004348631,0.0062641115,0.03566065,0.00489721,0.0103855645,0.011291598,0.0066908514,0.02717777],"category_scores_gemma":[0.6209948,0.006160642,0.008727854,0.01747458,0.004915765,0.014870877,0.012303393,0.014247825,0.012525119],"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":[0.0015639235,0.00080778555,0.0036881643,0.057596065,0.0004737289,0.0008457224,0.022432957,0.003008037,0.0024852077,0.03197763,0.53818953,0.33693132],"study_design_scores_gemma":[0.002362273,0.001007241,0.00898482,0.0664459,0.0005180045,0.0008740319,0.015334098,0.005643704,0.004147615,0.032034248,0.8616701,0.0009779488],"about_ca_topic_score_codex":0.0062856413,"about_ca_topic_score_gemma":0.0073871757,"teacher_disagreement_score":0.589035,"about_ca_system_score_codex":0.01569257,"about_ca_system_score_gemma":0.073083796,"threshold_uncertainty_score":0.726385},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"reporting","study_design":"design_other","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4391035776","doi":"10.1136/bmjebm-2023-112530","title":"Rapid review methods series: Guidance on the use of supportive software","year":2024,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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":"Systematic review; Software; Computer science; Automation; Process (computing); Software engineering; Engineering; MEDLINE","score_opus":0.9106387882086024,"score_gpt":0.6251835614097659,"score_spread":0.2854552267988365,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4391035776","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0012620847,0.017736692,0.5710121,0.054070454,0.02686156,0.17689438,0.05144454,0.049207453,0.051510766],"genre_scores_gemma":[0.0017118803,0.010039933,0.8086444,0.005681696,0.0029989744,0.14139114,0.007214283,0.006648644,0.015668979],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.6289047,0.25303888,0.08217783,0.004790421,0.028322948,0.0027652718],"domain_scores_gemma":[0.20754585,0.6113456,0.04483411,0.036697004,0.093661025,0.0059164697],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.31896532,0.0053876867,0.007863686,0.024188964,0.002181345,0.009698868,0.009637056,0.012173456,0.18187971],"category_scores_gemma":[0.6687196,0.0124971,0.013358341,0.020962441,0.004324776,0.012223357,0.011623984,0.015186016,0.17637044],"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":[0.0009911391,0.00016695257,0.00031336525,0.046675816,0.00048380514,0.0003718539,0.0015911883,0.0010813829,0.0010416957,0.009974387,0.6872794,0.250029],"study_design_scores_gemma":[0.002537374,0.00025836882,0.0013120276,0.04541227,0.00043499714,0.0006759281,0.00039796424,0.0021487873,0.0016140501,0.028870033,0.9159444,0.0003938208],"about_ca_topic_score_codex":0.0029599245,"about_ca_topic_score_gemma":0.0055719037,"teacher_disagreement_score":0.6810347,"about_ca_system_score_codex":0.004981896,"about_ca_system_score_gemma":0.043416865,"threshold_uncertainty_score":0.839837},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4392545360","doi":"10.1136/bmjebm-2024-112868","title":"Overly complex methods may impair pragmatic use of core evidence-based medicine principles","year":2024,"lang":"en","type":"editorial","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","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; Impact","funders":"","keywords":"Core (optical fiber); Psychology; Epistemology; Medicine; Computer science; Philosophy","score_opus":0.6553516264834451,"score_gpt":0.6381847810439233,"score_spread":0.01716684543952185,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4392545360","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00005437132,0.008300344,0.0007000474,0.053698257,0.9353348,0.00014325303,0.0001055296,0.00009441902,0.001568924],"genre_scores_gemma":[0.0018942979,0.012250789,0.002680482,0.07754212,0.898292,0.0004464542,0.00008027899,0.0001945556,0.0066190236],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.88730234,0.039703127,0.020204173,0.0035588264,0.047405384,0.0018261817],"domain_scores_gemma":[0.5321652,0.37237814,0.013958446,0.008558633,0.06482141,0.008118149],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.11780408,0.0049262885,0.0069921715,0.00907513,0.0060329875,0.016425857,0.007926932,0.029899001,0.013249777],"category_scores_gemma":[0.4173596,0.0034051589,0.0065571666,0.0049560685,0.007671713,0.008237422,0.004284075,0.035723925,0.008281406],"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":[0.00014268767,0.000025340425,0.000045463243,0.0037992736,0.0003384493,0.00007398,0.00014162359,0.000071465016,0.00007480269,0.0017067099,0.9757544,0.017825875],"study_design_scores_gemma":[0.0010014377,0.00009241033,0.00065889786,0.013203418,0.0011829594,0.00026186433,0.00029073446,0.0008706431,0.00039315593,0.0165928,0.96529126,0.00016033404],"about_ca_topic_score_codex":0.004091731,"about_ca_topic_score_gemma":0.013874409,"teacher_disagreement_score":0.88219595,"about_ca_system_score_codex":0.0064430633,"about_ca_system_score_gemma":0.014980577,"threshold_uncertainty_score":0.6230149},"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":"W4392583344","doi":"10.1136/bmjebm-2023-112695","title":"CARE-radiology statement explanation and elaboration: reporting guideline for radiological case reports","year":2024,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Delphi Technique in Research","field":"Social Sciences","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; Impact","funders":"Fundamental Research Funds for the Central Universities; Lanzhou University","keywords":"Guideline; Radiological weapon; Checklist; Medicine; Delphi method; Multidisciplinary approach; Documentation; Medical physics; Radiology; Psychology; Pathology; Computer science","score_opus":0.3723114543237489,"score_gpt":0.5789761313316922,"score_spread":0.20666467700794333,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4392583344","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008472835,0.0033303653,0.4390923,0.07376936,0.0056389594,0.42567426,0.013005895,0.005896487,0.025119588],"genre_scores_gemma":[0.009615068,0.0022949923,0.7071806,0.00740627,0.00074664765,0.2616494,0.0068408516,0.0005023863,0.0037638806],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.5033019,0.31817433,0.14394684,0.0039423285,0.026864007,0.0037705451],"domain_scores_gemma":[0.33656758,0.29281938,0.07301204,0.0424469,0.24967119,0.005482914],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.31732264,0.0030477857,0.003174282,0.019889176,0.004036072,0.00616891,0.007582948,0.008868375,0.014304796],"category_scores_gemma":[0.5090412,0.0036259468,0.0056458553,0.009197391,0.004829341,0.008247955,0.008990976,0.008089723,0.015028809],"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":[0.00053476344,0.00086284505,0.0043524397,0.026361683,0.000207773,0.0017171404,0.029939888,0.0025907606,0.0071332026,0.020764168,0.6382973,0.26723802],"study_design_scores_gemma":[0.0011035847,0.0007029183,0.010878438,0.05680329,0.00021332173,0.0032910088,0.0148479985,0.0066100196,0.007193824,0.023783784,0.8737934,0.00077840616],"about_ca_topic_score_codex":0.0048685493,"about_ca_topic_score_gemma":0.0050360016,"teacher_disagreement_score":0.6826774,"about_ca_system_score_codex":0.00864225,"about_ca_system_score_gemma":0.037591312,"threshold_uncertainty_score":0.84186274},"labels":[],"label_agreement":null},{"id":"W4392797608","doi":"10.1136/bmjebm-2023-112722","title":"Rapid reviews methods series: assessing the appropriateness of conducting a rapid review","year":2024,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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":"Ottawa Public Health; Public Health Agency of Canada; University of Ottawa","funders":"","keywords":"Scope (computer science); Context (archaeology); Risk analysis (engineering); Resource (disambiguation); Process (computing); Management science; Computer science; Systematic review; Operations research; Medicine; MEDLINE; Political science; Engineering","score_opus":0.9588174281060408,"score_gpt":0.7032138568070663,"score_spread":0.2556035712989745,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4392797608","genre_codex":"protocol","genre_gemma":"methods","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0020935212,0.036517665,0.24107094,0.060184862,0.053351697,0.5265575,0.02907767,0.013445762,0.03770035],"genre_scores_gemma":[0.0038562268,0.017458431,0.51089686,0.0059191887,0.004390974,0.44403183,0.0041264393,0.0032810173,0.00603894],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.17133193,0.5977991,0.15116079,0.008190682,0.06921409,0.0023034746],"domain_scores_gemma":[0.0505859,0.6996047,0.07038724,0.054038085,0.12029237,0.005091785],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.6970117,0.0062058778,0.010780219,0.03354989,0.0044262465,0.020797234,0.008674599,0.010941311,0.102747396],"category_scores_gemma":[0.9129382,0.010493086,0.014925919,0.03223076,0.0075929593,0.020227814,0.014473935,0.013659122,0.07721359],"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":[0.0031235723,0.00025594915,0.0014374019,0.18832593,0.0037735263,0.00026694118,0.002222329,0.0012733728,0.0012252799,0.017717285,0.3975109,0.3828676],"study_design_scores_gemma":[0.0059152907,0.0013421281,0.004078202,0.16561942,0.0032346584,0.0007528592,0.0009667766,0.0031241358,0.002195171,0.03792263,0.773932,0.0009167545],"about_ca_topic_score_codex":0.0026474143,"about_ca_topic_score_gemma":0.0053232852,"teacher_disagreement_score":0.3029883,"about_ca_system_score_codex":0.014290099,"about_ca_system_score_gemma":0.08802902,"threshold_uncertainty_score":0.37363857},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4394722187","doi":"10.1136/bmjebm-2023-112767","title":"Trivalent and quadrivalent seasonal influenza vaccine in adults aged 60 and older: a systematic review and network meta-analysis","year":2024,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Influenza Virus Research Studies","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Public Health Ontario; University of Ottawa; McMaster University; University of Toronto; Public Health Agency of Canada; Ottawa Hospital; Institute for Work & Health; St. Michael's Hospital","funders":"Canadian Institutes of Health Research; Canada Excellence Research Chairs, Government of Canada; Public Health Agency; Health Research Board","keywords":"Medicine; Meta-analysis; Confidence interval; Odds ratio; Influenza vaccine; Placebo; Randomized controlled trial; Systematic review; Internal medicine; MEDLINE; Pediatrics; Vaccination; Immunology; Alternative medicine","score_opus":0.3326713281150513,"score_gpt":0.4941296525993264,"score_spread":0.1614583244842751,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4394722187","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004519348,0.99219704,0.0010279741,0.00027011888,0.00020087359,0.0009352775,0.0005126054,0.00003229071,0.0003044632],"genre_scores_gemma":[0.16800252,0.8157245,0.007061314,0.0011258133,0.00034154,0.006304761,0.00095642585,0.000034048546,0.00044906235],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9767254,0.012366303,0.006447552,0.0017363284,0.0022890633,0.00043530832],"domain_scores_gemma":[0.97220516,0.01962065,0.0050820163,0.00068884547,0.0021178008,0.0002855037],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.025529204,0.0033567937,0.022626057,0.011087835,0.0006863261,0.0037784958,0.00240381,0.002702344,0.0036574793],"category_scores_gemma":[0.058385935,0.0018488141,0.04300221,0.0082961125,0.0007129449,0.0030352224,0.002041618,0.0017942188,0.00031275535],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013835825,0.000031805812,0.0025871203,0.39884305,0.58567065,0.000082414175,0.00008202479,0.00066011184,0.00011195038,0.00014757579,0.00038384696,0.010015902],"study_design_scores_gemma":[0.00073470053,0.00018600805,0.0021029606,0.03792342,0.9570494,0.000082461105,0.000041855663,0.00035047074,0.00006802201,0.00022075672,0.0012161733,0.000023825223],"about_ca_topic_score_codex":0.0074173976,"about_ca_topic_score_gemma":0.015594198,"teacher_disagreement_score":0.025529204,"about_ca_system_score_codex":0.0044827675,"about_ca_system_score_gemma":0.0053857123,"threshold_uncertainty_score":0.13501298},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":true,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4396967666","doi":"10.1136/bmjebm-2023-112622","title":"Direct-to-consumer advertising: a modifiable driver of overdiagnosis and overtreatment","year":2024,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Pharmaceutical industry and healthcare","field":"Pharmacology, Toxicology and Pharmaceutics","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":"York University","funders":"","keywords":"Overdiagnosis; Advertising; Medicine; Business; Internal medicine","score_opus":0.48704821847550556,"score_gpt":0.5796537362761751,"score_spread":0.09260551780066956,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4396967666","genre_codex":"empirical","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.86637926,0.036728755,0.0020633913,0.03814285,0.0012102405,0.00012445769,0.021658508,0.00015837823,0.03353427],"genre_scores_gemma":[0.99050856,0.0032961152,0.0006492416,0.0017348364,0.0005043224,0.000019305255,0.0014195091,0.000034935907,0.0018332509],"study_design_codex":"observational","study_design_gemma":"not_applicable","domain_scores_codex":[0.99635786,0.0010675313,0.00049943896,0.0006435083,0.00096857187,0.00046303836],"domain_scores_gemma":[0.9616903,0.01669305,0.014052962,0.002839074,0.0030895488,0.0016351296],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0022586056,0.00031158986,0.0005543185,0.0018529906,0.00065713015,0.002506703,0.0008830567,0.0021240332,0.012471533],"category_scores_gemma":[0.03682034,0.00052349817,0.001457555,0.004297288,0.0005441798,0.0015750516,0.0009768914,0.0036177007,0.001149962],"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":[0.0002070266,0.00010305091,0.9791341,0.00024982681,0.00035075305,0.0002905627,0.00027268624,0.000057633584,0.00010915749,0.00074906205,0.0047252327,0.013750884],"study_design_scores_gemma":[0.000022098595,0.000061854924,0.9891973,0.0005212936,0.00056902325,0.0012927728,0.00052646425,0.00055769226,0.00018158618,0.0012307505,0.0058072847,0.000031815114],"about_ca_topic_score_codex":0.041783407,"about_ca_topic_score_gemma":0.03230875,"teacher_disagreement_score":0.041783407,"about_ca_system_score_codex":0.0010587677,"about_ca_system_score_gemma":0.0023682872,"threshold_uncertainty_score":0.08308035},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4398251068","doi":"10.1136/bmjebm-2023-112681","title":"Guidelines for the development and validation of patient-reported outcome measures: a scoping review","year":2024,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Cancer survivorship and care","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 Alberta","funders":"","keywords":"Prom; PsycINFO; CLARITY; Guideline; MEDLINE; Scopus; Medicine; Data extraction; Computer science; Medical physics; Pathology","score_opus":0.5391998939064311,"score_gpt":0.5249969430455457,"score_spread":0.01420295086088541,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398251068","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0025852635,0.6102853,0.11622368,0.06146443,0.013800623,0.14830214,0.014164509,0.0016678732,0.03150621],"genre_scores_gemma":[0.011653942,0.40394408,0.35924768,0.009652215,0.0014016734,0.20025161,0.009518115,0.0005019932,0.0038286764],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.51028883,0.20432684,0.22981124,0.008237061,0.043512538,0.0038235139],"domain_scores_gemma":[0.35985783,0.37359217,0.064513184,0.021232584,0.17746617,0.003338098],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.33305204,0.0045285304,0.010016422,0.064278394,0.0057056504,0.014230684,0.012238145,0.015259796,0.010014979],"category_scores_gemma":[0.58664393,0.0057906304,0.017949307,0.044719193,0.007462299,0.014723349,0.01215823,0.010094822,0.00810999],"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":[0.00027278447,0.00015695968,0.0015818251,0.5541681,0.0015361037,0.0005128641,0.005991751,0.0012600758,0.00090285734,0.0154673625,0.10746527,0.31068408],"study_design_scores_gemma":[0.00018917928,0.00007438611,0.0011933658,0.8137029,0.0016928467,0.00029756333,0.0014951497,0.00044669735,0.0005058622,0.0083971545,0.17189373,0.00011117447],"about_ca_topic_score_codex":0.0123869805,"about_ca_topic_score_gemma":0.01676684,"teacher_disagreement_score":0.66694796,"about_ca_system_score_codex":0.015850144,"about_ca_system_score_gemma":0.093279466,"threshold_uncertainty_score":0.8224656},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4400491807","doi":"10.1136/bmjebm-2023-112719","title":"Cancer screening attendance rates in transgender and gender-diverse patients: a systematic review and meta-analysis","year":2024,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"LGBTQ Health, Identity, and Policy","field":"Psychology","cited_by":9,"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":"Academy of Medical Sciences; Wellcome Trust","keywords":"Transgender; Meta-analysis; Attendance; Cancer; Medicine; Clinical psychology; Psychology; Demography; Gerontology; Sociology; Internal medicine; Political science; Psychoanalysis","score_opus":0.5172235772782489,"score_gpt":0.5521889312201794,"score_spread":0.034965353941930566,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4400491807","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.005854981,0.992243,0.0003994449,0.00017145202,0.000107403626,0.0003632053,0.00062065944,0.000019137795,0.00022073193],"genre_scores_gemma":[0.1535053,0.84028405,0.0022780655,0.00080025336,0.00017680365,0.0017312478,0.00094783754,0.0000232747,0.00025322352],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9909859,0.0037961004,0.0028866539,0.00095476373,0.0010417823,0.00033483157],"domain_scores_gemma":[0.9787649,0.015736822,0.0032917503,0.0005329768,0.0014417188,0.00023196662],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0144234765,0.002395656,0.018347025,0.007708128,0.0006372875,0.0030988322,0.00237847,0.002201642,0.003174244],"category_scores_gemma":[0.037516102,0.0014635184,0.039174147,0.0075259055,0.00065555354,0.0018444406,0.0015807624,0.0014397249,0.00024209039],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012887699,0.000030359372,0.0056300797,0.5361016,0.44458857,0.00012081383,0.0001247866,0.0003886537,0.00016015921,0.0001105119,0.00052153855,0.010934068],"study_design_scores_gemma":[0.0003864045,0.00021170243,0.005051791,0.06497013,0.9273582,0.00013158724,0.00009360292,0.00014525375,0.000109443434,0.00012435734,0.0013929433,0.000024571416],"about_ca_topic_score_codex":0.009318724,"about_ca_topic_score_gemma":0.019769447,"teacher_disagreement_score":0.018347025,"about_ca_system_score_codex":0.0029778755,"about_ca_system_score_gemma":0.0046477034,"threshold_uncertainty_score":0.07627952},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4400582003","doi":"10.1136/bmjebm-2024-112886","title":"Pharmacological interventions for preventing upper gastrointestinal bleeding in people admitted to intensive care units: a network meta-analysis","year":2024,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Nosocomial Infections in ICU","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":"Psychological intervention; Intensive care; Medicine; Upper gastrointestinal bleeding; Intensive care medicine; Gastrointestinal bleeding; Meta-analysis; Intensive care unit; Internal medicine; Nursing","score_opus":0.4932109188940237,"score_gpt":0.5329212801273934,"score_spread":0.039710361233369684,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4400582003","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.013187915,0.9820712,0.0017039378,0.00035939954,0.00035377542,0.000993034,0.0007593269,0.00006913028,0.0005022603],"genre_scores_gemma":[0.43238297,0.5497104,0.007946056,0.0015316606,0.00063787977,0.0051517305,0.0017041806,0.00007801001,0.0008571291],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9789723,0.012357476,0.00454123,0.0018009237,0.0019042982,0.00042369246],"domain_scores_gemma":[0.97423935,0.018762277,0.004096493,0.0009800317,0.001599252,0.00032256756],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.025175188,0.0034330527,0.023256995,0.009009727,0.0007327236,0.004201408,0.0026227979,0.002644178,0.004847088],"category_scores_gemma":[0.047564298,0.0016825453,0.052775774,0.007087675,0.0007784349,0.002481546,0.0023213595,0.0020235567,0.00041613317],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0020953005,0.000024047398,0.0023839793,0.113351,0.87812144,0.000063921616,0.0000386059,0.00033295696,0.00008901888,0.000060006263,0.00017169332,0.0032679397],"study_design_scores_gemma":[0.0007605856,0.00015582044,0.0016876864,0.0071242983,0.98944783,0.000042170595,0.000018920351,0.00019537161,0.000045103083,0.000095401214,0.00041444015,0.000012435493],"about_ca_topic_score_codex":0.005210883,"about_ca_topic_score_gemma":0.009090087,"teacher_disagreement_score":0.025175188,"about_ca_system_score_codex":0.0030371621,"about_ca_system_score_gemma":0.002969331,"threshold_uncertainty_score":0.13314068},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4400879434","doi":"10.1136/bmjebm-2024-112899","title":"Rapid review method series: interim guidance for the reporting of rapid reviews","year":2024,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":55,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Public Health Ontario; University of Toronto; St. Michael's Hospital; University of Alberta; Public Health Agency of Canada; Bruyère; University of Ottawa","funders":"Canadian Institutes of Health Research; Canada Research Chairs","keywords":"Interim; Systematic review; Set (abstract data type); Computer science; Psychology; MEDLINE; Political science","score_opus":0.9004752747875066,"score_gpt":0.647176774034603,"score_spread":0.2532985007529035,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4400879434","genre_codex":"protocol","genre_gemma":"methods","domain_codex":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0009555778,0.010187738,0.28165498,0.036491785,0.041259788,0.5135577,0.07358022,0.017071487,0.025240725],"genre_scores_gemma":[0.001581265,0.0053698346,0.4098487,0.003246478,0.00235,0.55856436,0.0082581155,0.0021589268,0.008622335],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.39367947,0.4122575,0.131495,0.017075947,0.04121351,0.004278587],"domain_scores_gemma":[0.17103568,0.48697615,0.07554027,0.07095124,0.1883744,0.0071222675],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.49539238,0.0075541013,0.013396782,0.02681083,0.004181833,0.020328598,0.009711759,0.0155551955,0.17102012],"category_scores_gemma":[0.73480266,0.016792841,0.018985087,0.02951867,0.00954439,0.01646572,0.010006449,0.028243711,0.20622179],"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":[0.0018499558,0.00021254142,0.00055438903,0.08652606,0.0010892436,0.00021081865,0.002330285,0.0009049435,0.0013207913,0.018029608,0.71979,0.16718136],"study_design_scores_gemma":[0.0050261016,0.0004183246,0.0018754347,0.078027554,0.0009308354,0.00033906338,0.0008776308,0.0019063415,0.0017943691,0.02818638,0.8800321,0.0005857222],"about_ca_topic_score_codex":0.0027008282,"about_ca_topic_score_gemma":0.004662409,"teacher_disagreement_score":0.5046076,"about_ca_system_score_codex":0.01074788,"about_ca_system_score_gemma":0.0779357,"threshold_uncertainty_score":0.62227106},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","research_integrity"],"domain":"reporting","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4401198502","doi":"10.1136/bmjebm-2023-112815","title":"Successes, shortcomings and learning opportunities for evidence-based medicine from the COVID-19 pandemic","year":2024,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","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; Impact","funders":"","keywords":"Coronavirus disease 2019 (COVID-19); Pandemic; 2019-20 coronavirus outbreak; Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); Virology; Computer science; Medicine; Infectious disease (medical specialty); Internal medicine; Disease","score_opus":0.7098921007921705,"score_gpt":0.591445131736033,"score_spread":0.1184469690561375,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4401198502","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.018033842,0.015923928,0.0008054415,0.9492478,0.0025080137,0.000028929055,0.00010846146,0.00004956908,0.013294022],"genre_scores_gemma":[0.7396491,0.06924477,0.014801399,0.16005631,0.008536381,0.0002271582,0.00055745576,0.00030324102,0.006624135],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.93996114,0.030482475,0.0054426612,0.001967128,0.015357603,0.0067889383],"domain_scores_gemma":[0.74259394,0.12343227,0.014517151,0.007690542,0.04051727,0.07124879],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.14449418,0.00044749095,0.00088040583,0.0030896626,0.005338256,0.019624807,0.0025132513,0.006871883,0.007945436],"category_scores_gemma":[0.25516912,0.00063421007,0.0009850909,0.0025148045,0.008999898,0.0190775,0.02095994,0.021091718,0.0016106656],"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":[0.0006398632,0.00064212753,0.0398256,0.0038501443,0.0002199031,0.001584934,0.027788518,0.0012689161,0.00073841424,0.10423669,0.21890315,0.60030174],"study_design_scores_gemma":[0.0002551632,0.00085944147,0.081682526,0.022243278,0.00017343412,0.0020291363,0.102838226,0.0025750524,0.0016174066,0.14802776,0.63719195,0.0005066074],"about_ca_topic_score_codex":0.00771773,"about_ca_topic_score_gemma":0.020441826,"teacher_disagreement_score":0.8555058,"about_ca_system_score_codex":0.010598782,"about_ca_system_score_gemma":0.042183146,"threshold_uncertainty_score":0.76416737},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","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":"low"}],"label_agreement":"split"},{"id":"W4401991382","doi":"10.1136/bmjebm-2024-113031","title":"Meta-analysis of continuous outcomes: a user’s guide for analysis and interpretation","year":2024,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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; Impact","funders":"","keywords":"Interpretation (philosophy); Computer science; Meta-analysis; Psychology; Medicine; Internal medicine; Programming language","score_opus":0.8238197139119845,"score_gpt":0.6024751740025874,"score_spread":0.22134453990939706,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4401991382","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0013696089,0.01392551,0.7352453,0.0027916422,0.003778576,0.011411738,0.11925082,0.105490446,0.006736458],"genre_scores_gemma":[0.0024900138,0.002359343,0.9452515,0.0013556348,0.00043484257,0.020056186,0.008540245,0.015554449,0.0039578425],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9520978,0.028462537,0.011982912,0.0028236583,0.0040420624,0.00059106515],"domain_scores_gemma":[0.68013614,0.2770126,0.008461008,0.013649633,0.01946017,0.0012804556],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.056469552,0.011044704,0.01574902,0.015709186,0.0014776031,0.0093822535,0.010267827,0.007049889,0.17244652],"category_scores_gemma":[0.23945463,0.011998795,0.015329881,0.020201333,0.002496939,0.005721744,0.0052538426,0.013608092,0.09144697],"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":[0.0019795282,0.0004407503,0.0017186102,0.04942711,0.007888954,0.00084716873,0.000981169,0.0054700198,0.0032044782,0.014203667,0.7755232,0.13831536],"study_design_scores_gemma":[0.009958379,0.0007652193,0.008693672,0.029002726,0.0114813,0.003728684,0.00072225987,0.067964055,0.014601593,0.10456584,0.74612033,0.0023959535],"about_ca_topic_score_codex":0.0045833616,"about_ca_topic_score_gemma":0.009500087,"teacher_disagreement_score":0.94353044,"about_ca_system_score_codex":0.0018699818,"about_ca_system_score_gemma":0.011035511,"threshold_uncertainty_score":0.57689106},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4406820295","doi":"10.1136/bmjebm-2024-113134","title":"Proposed framework for unifying disease definitions in guideline development","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Clinical practice guidelines implementation","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":"Guideline; Disease; Computer science; Management science; Medicine; Engineering; Pathology","score_opus":0.5406540875678616,"score_gpt":0.5688595876034911,"score_spread":0.028205500035629494,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406820295","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0042069894,0.0018741387,0.8906015,0.074352674,0.0012524148,0.0056091105,0.0012278577,0.0012240125,0.019651292],"genre_scores_gemma":[0.01691097,0.00032636174,0.97837037,0.0013498019,0.00015145795,0.001493945,0.0005497544,0.00005050763,0.00079683104],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.85461694,0.07922063,0.0320071,0.009582731,0.020017194,0.0045554205],"domain_scores_gemma":[0.78788054,0.09263197,0.012891703,0.016891306,0.08096562,0.008738862],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.14887883,0.0027679869,0.0030493364,0.018419929,0.008756995,0.027518427,0.009916932,0.012548612,0.00813491],"category_scores_gemma":[0.18634157,0.0030462458,0.0069435555,0.012820905,0.011976142,0.017983701,0.014842815,0.012976994,0.0043949024],"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":[0.00013508777,0.0003926061,0.0058261883,0.0017843028,0.00044725926,0.0005841006,0.008737291,0.0048367223,0.0010693158,0.8414135,0.022563886,0.11220979],"study_design_scores_gemma":[0.00031897152,0.0003349483,0.00393583,0.006207954,0.00092631055,0.0009518018,0.0056676283,0.03505483,0.0025489444,0.82383263,0.11986633,0.00035375313],"about_ca_topic_score_codex":0.029131249,"about_ca_topic_score_gemma":0.03737574,"teacher_disagreement_score":0.8511212,"about_ca_system_score_codex":0.013581397,"about_ca_system_score_gemma":0.059223466,"threshold_uncertainty_score":0.7873559},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4407132017","doi":"10.1136/bmjebm-2023-112389","title":"Rapid reviews methods series: guidance on rapid scoping, mapping and evidence and gap map (‘Big Picture Reviews’)","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Delphi Technique in Research","field":"Social Sciences","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":"St. Michael's Hospital; Ottawa Hospital","funders":"","keywords":"Series (stratigraphy); Data science; Computer science; Information retrieval; Biology","score_opus":0.4793826492030341,"score_gpt":0.5678402709174526,"score_spread":0.08845762171441851,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4407132017","genre_codex":"protocol","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00095090765,0.080996856,0.22587153,0.077740595,0.026854146,0.243366,0.11876016,0.0229191,0.2025407],"genre_scores_gemma":[0.0024288157,0.09599863,0.5181314,0.015086427,0.0038338208,0.2661666,0.032126512,0.006426341,0.059801467],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.76838696,0.13437216,0.04956115,0.0064121764,0.037941374,0.0033262633],"domain_scores_gemma":[0.48274922,0.34033132,0.040182203,0.025158003,0.105415896,0.0061632893],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.21553165,0.0052657733,0.0074261157,0.022230232,0.0025237352,0.013678589,0.01145873,0.017556943,0.20712584],"category_scores_gemma":[0.44783148,0.008633054,0.012148183,0.029629365,0.0031737324,0.012134382,0.012656998,0.0127474805,0.21730188],"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":[0.00018939492,0.00007655984,0.00013272569,0.045400985,0.00017114027,0.00012144623,0.0010930975,0.00050243246,0.00038362129,0.008859184,0.78556883,0.15750058],"study_design_scores_gemma":[0.0004318945,0.00007103515,0.00051718106,0.03907127,0.00013430853,0.00014629876,0.00018984285,0.0002642951,0.00034582234,0.006909426,0.9517938,0.00012490705],"about_ca_topic_score_codex":0.009372389,"about_ca_topic_score_gemma":0.015145409,"teacher_disagreement_score":0.78446835,"about_ca_system_score_codex":0.009178703,"about_ca_system_score_gemma":0.061067805,"threshold_uncertainty_score":0.96738917},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4407402958","doi":"10.1136/bmjebm-2024-112919","title":"What makes a ‘good’ decision with artificial intelligence? A grounded theory study in paediatric care","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Artificial Intelligence in Healthcare and Education","field":"Medicine","cited_by":8,"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; York University; SickKids Foundation; University of Toronto; Canadian Institute for Advanced Research; Vector Institute","funders":"Canadian Institutes of Health Research; Hospital for Sick Children","keywords":"Grounded theory; Nonprobability sampling; Snowball sampling; Psychology; Psychological intervention; Qualitative research; Health care; Medical education; Applied psychology; Nursing; Medicine","score_opus":0.2084493904989476,"score_gpt":0.4812033996992079,"score_spread":0.2727540092002603,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4407402958","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95851237,0.00263165,0.014095111,0.013790655,0.00018760243,0.0019406521,0.00014619398,0.000021795879,0.008673896],"genre_scores_gemma":[0.9854165,0.0014867433,0.009723378,0.0015879124,0.000024348465,0.0010582015,0.00005400753,0.000019303185,0.0006297399],"study_design_codex":"qualitative","study_design_gemma":"qualitative","domain_scores_codex":[0.93224025,0.0598116,0.0012831627,0.0013280094,0.002653473,0.0026835892],"domain_scores_gemma":[0.88249207,0.10448895,0.0040148194,0.0018881398,0.0029974221,0.0041186325],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.08369147,0.0006694353,0.0009785813,0.0043973294,0.010422926,0.008579546,0.0037158006,0.0030238763,0.002560251],"category_scores_gemma":[0.08051379,0.0010866397,0.0007342512,0.0037226635,0.022805022,0.010361295,0.008511202,0.006749003,0.00030035537],"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":[0.00003457801,0.00010515735,0.0021621052,0.0003338275,0.000005724445,0.00036747725,0.98598725,0.00007267717,0.0001054142,0.0059327055,0.00029315255,0.0045998613],"study_design_scores_gemma":[0.000023129682,0.00011512516,0.0012685085,0.0007708012,0.000006523641,0.000198846,0.98954165,0.00022444979,0.000108571505,0.003219659,0.004506742,0.000016035923],"about_ca_topic_score_codex":0.006127112,"about_ca_topic_score_gemma":0.00938037,"teacher_disagreement_score":0.08369147,"about_ca_system_score_codex":0.015045887,"about_ca_system_score_gemma":0.017278172,"threshold_uncertainty_score":0.44260812},"labels":[{"model":"gemma","categories":["sts"],"domain":null,"study_design":"qualitative","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"qualitative","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4407402962","doi":"10.1136/bmjebm-2023-112588","title":"Expanded disease definitions in Alzheimer’s disease and the new era of disease-modifying drugs","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Dementia and Cognitive Impairment Research","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; University of Alberta","funders":"","keywords":"Disease; Medicine; Pathology","score_opus":0.09279324019988813,"score_gpt":0.3975662631856471,"score_spread":0.3047730229857589,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4407402962","genre_codex":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00911033,0.7483684,0.013606967,0.17782891,0.031820465,0.0001809205,0.0017371821,0.00011081893,0.01723605],"genre_scores_gemma":[0.15392187,0.55786407,0.08094284,0.11837641,0.075669356,0.0014641295,0.0058006905,0.00019866928,0.005761947],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.95424527,0.024162982,0.011928366,0.0016573662,0.0071937046,0.00081224623],"domain_scores_gemma":[0.89990807,0.057531048,0.016015308,0.003872891,0.018642578,0.0040301583],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.045021202,0.0015228442,0.004175087,0.0092933765,0.0012703926,0.007065066,0.003125444,0.0037575888,0.0034123072],"category_scores_gemma":[0.07468522,0.00057014637,0.003975078,0.007741594,0.0042863335,0.010678992,0.0049171178,0.010437514,0.00057356275],"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":[0.0016892889,0.00051160896,0.01704092,0.025503196,0.0012761264,0.00062167284,0.0029818104,0.0021741264,0.0018084026,0.32331002,0.21749547,0.40558738],"study_design_scores_gemma":[0.00069490145,0.00068098033,0.037961975,0.026687471,0.0009796678,0.002978516,0.0019820996,0.0019261756,0.00039759566,0.19135651,0.7340449,0.00030925786],"about_ca_topic_score_codex":0.0061281663,"about_ca_topic_score_gemma":0.009943749,"teacher_disagreement_score":0.045021202,"about_ca_system_score_codex":0.005197567,"about_ca_system_score_gemma":0.00692297,"threshold_uncertainty_score":0.23809773},"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":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4408245171","doi":"10.1136/bmjebm-2024-113077","title":"Rating certainty when the target threshold is the null and the point estimate is close to the null","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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":"McMaster University; Impact","funders":"Einstein Stiftung Berlin; National Natural Science Foundation of China","keywords":"Certainty; Null hypothesis; Null (SQL); Harm; Confidence interval; Point estimation; Point (geometry); Range (aeronautics); Grading (engineering); Econometrics; Mathematics; Statistics; Computer science; Psychology; Social psychology; Data mining; Engineering","score_opus":0.5544545072542626,"score_gpt":0.533779388139586,"score_spread":0.020675119114676566,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4408245171","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08527002,0.07233574,0.592009,0.114768185,0.018273143,0.0069656083,0.004682026,0.002210617,0.10348565],"genre_scores_gemma":[0.6347712,0.01749398,0.3129394,0.019895421,0.0048625562,0.0049159206,0.0014666127,0.00053860975,0.0031162994],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.6835962,0.18475533,0.06761726,0.009912582,0.051911004,0.0022076266],"domain_scores_gemma":[0.27708504,0.57537633,0.048929315,0.030795425,0.06345524,0.004358645],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.22673933,0.00137505,0.0039627426,0.008177809,0.0016802953,0.009179367,0.0025578253,0.007506573,0.007860707],"category_scores_gemma":[0.72581035,0.0012305616,0.00438716,0.004398793,0.0046921773,0.010323806,0.007836872,0.0099554015,0.0022892645],"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":[0.005624315,0.0003396986,0.038683347,0.03934272,0.0056852964,0.0011353131,0.015602721,0.0069341287,0.0064334036,0.08204316,0.10117554,0.69700044],"study_design_scores_gemma":[0.0031519358,0.003732901,0.047863178,0.060731523,0.008627771,0.006470548,0.010815299,0.033648074,0.012339677,0.5521806,0.2577937,0.0026448343],"about_ca_topic_score_codex":0.0021651036,"about_ca_topic_score_gemma":0.0020473045,"teacher_disagreement_score":0.77326065,"about_ca_system_score_codex":0.0036233515,"about_ca_system_score_gemma":0.0036338416,"threshold_uncertainty_score":0.9535681},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4408383217","doi":"10.1136/bmjebm-2024-113527","title":"Improving peer review of systematic reviews and related review types by involving librarians and information specialists as methodological peer reviewers: a randomised controlled trial","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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 Toronto; Ottawa Hospital","funders":"","keywords":"Peer review; Psychological intervention; Systematic review; Medicine; Randomized controlled trial; Family medicine; Inclusion (mineral); Intervention (counseling); MEDLINE; Medical education; Psychology; Nursing; Internal medicine; Social psychology","score_opus":0.6790106756671164,"score_gpt":0.5482039948117332,"score_spread":0.13080668085538316,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4408383217","genre_codex":"protocol","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.089567944,0.2799331,0.03340385,0.015393976,0.019755326,0.5507604,0.00208864,0.0018597806,0.0072369883],"genre_scores_gemma":[0.27920756,0.08355214,0.11967115,0.0076953904,0.007378998,0.4994931,0.0009051376,0.0002775562,0.0018189097],"study_design_codex":"systematic_review","study_design_gemma":"randomized_trial","domain_scores_codex":[0.42325917,0.44023123,0.07632563,0.019112488,0.037084185,0.0039871866],"domain_scores_gemma":[0.501711,0.29676977,0.11927483,0.03349531,0.037719652,0.011029415],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.29428184,0.004513172,0.01833039,0.007883723,0.0043315925,0.011235902,0.005166547,0.014678328,0.022301642],"category_scores_gemma":[0.51881397,0.004619185,0.014457797,0.008335113,0.008545554,0.015976317,0.007106895,0.0076001813,0.003707177],"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":[0.24746215,0.0057290085,0.0028235968,0.4931655,0.038778536,0.00018402262,0.0018504554,0.0010886942,0.0015651473,0.0021733535,0.008804013,0.19637556],"study_design_scores_gemma":[0.82407224,0.028088734,0.004295249,0.090791814,0.03311172,0.00025903303,0.00059566234,0.0017773317,0.0010523315,0.0031439024,0.012414481,0.0003975586],"about_ca_topic_score_codex":0.0016644084,"about_ca_topic_score_gemma":0.002599104,"teacher_disagreement_score":0.70571816,"about_ca_system_score_codex":0.007529724,"about_ca_system_score_gemma":0.027396139,"threshold_uncertainty_score":0.87027615},"labels":[],"label_agreement":null},{"id":"W4408488905","doi":"10.1136/bmjebm-2024-113050","title":"Identifying actionable statements in Chinese health guidelines: a cross-sectional study","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Clinical practice guidelines implementation","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; Cochrane","funders":"National Natural Science Foundation of China","keywords":"Guideline; Grading (engineering); Specialty; Family medicine; Medicine; Global Positioning System; Medical education; Computer science; Engineering; Pathology","score_opus":0.5934378988714254,"score_gpt":0.6822762201283746,"score_spread":0.08883832125694924,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4408488905","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.993934,0.0021867137,0.0005299433,0.00049117167,0.000020947504,0.00068495114,0.0011811763,0.0000066466578,0.0009643073],"genre_scores_gemma":[0.99495846,0.0010437211,0.0015653817,0.0004304354,0.000015770294,0.00086300686,0.0009685605,0.00000405129,0.0001506033],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.974981,0.00807084,0.008892061,0.0023768265,0.004679651,0.0009996311],"domain_scores_gemma":[0.9211523,0.029761726,0.03347328,0.0026835934,0.011484748,0.0014443555],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.027706442,0.0003444397,0.0010309912,0.005345935,0.0011744575,0.0015017578,0.0010569819,0.0011310133,0.0018649598],"category_scores_gemma":[0.078157105,0.000641149,0.0009024574,0.009545201,0.0011171071,0.0038703764,0.0015989647,0.0009833833,0.0002378495],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007894427,0.00015909917,0.9796912,0.0015349386,0.00014012842,0.00017384185,0.008159491,0.000039928145,0.000089943096,0.000275128,0.0007026,0.008954725],"study_design_scores_gemma":[0.000059486636,0.0004313943,0.9763307,0.001451943,0.0003248695,0.0005148486,0.016571112,0.00047464098,0.00016168943,0.0002901433,0.0033407893,0.000048283953],"about_ca_topic_score_codex":0.018038439,"about_ca_topic_score_gemma":0.020987881,"teacher_disagreement_score":0.97229356,"about_ca_system_score_codex":0.003047361,"about_ca_system_score_gemma":0.006887635,"threshold_uncertainty_score":0.14652741},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4408555504","doi":"10.1136/bmjebm-2024-112974","title":"Analgesic effects of non-surgical and non-interventional treatments for low back pain: a systematic review and meta-analysis of placebo-controlled randomised trials","year":2025,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Musculoskeletal pain and rehabilitation","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":"Memorial University of Newfoundland","funders":"","keywords":"Medicine; Placebo; Meta-analysis; Analgesic; Randomized controlled trial; Low back pain; Chronic pain; Physical therapy; Confidence interval; Strictly standardized mean difference; Relative risk; Clinical trial; Anesthesia; Internal medicine; Alternative medicine","score_opus":0.09458008446055835,"score_gpt":0.4391426482895198,"score_spread":0.34456256382896144,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4408555504","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0025207622,0.9948757,0.0006178336,0.00014311707,0.00019537953,0.001109129,0.00031685628,0.000023952822,0.00019724571],"genre_scores_gemma":[0.09009305,0.8966431,0.0045188013,0.0009386988,0.00030769664,0.0063514113,0.000707685,0.00003663788,0.00040289486],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.97331816,0.013593396,0.007782676,0.0019084968,0.0029529242,0.0004442865],"domain_scores_gemma":[0.9610647,0.02774537,0.0076783895,0.0009110977,0.0022546905,0.0003458125],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.026838355,0.003813662,0.030488312,0.009597851,0.0008987487,0.004489925,0.0029694284,0.0036186327,0.0051099416],"category_scores_gemma":[0.062642865,0.0021735292,0.04263743,0.008244152,0.0013594151,0.0029228453,0.0018890036,0.002311114,0.00040653397],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002597355,0.00003602945,0.00081573933,0.6809755,0.3046469,0.00006862214,0.00006863621,0.0003081367,0.00014621209,0.00011996482,0.00035279902,0.009864128],"study_design_scores_gemma":[0.002626324,0.000501526,0.001777252,0.09380666,0.8985794,0.00010694498,0.0000626169,0.00021733355,0.00018040331,0.00027443733,0.0018334214,0.000033674973],"about_ca_topic_score_codex":0.0053036204,"about_ca_topic_score_gemma":0.010380067,"teacher_disagreement_score":0.030488312,"about_ca_system_score_codex":0.0053630746,"about_ca_system_score_gemma":0.0060879034,"threshold_uncertainty_score":0.14193642},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W4408660425","doi":"10.1136/bmjebm-2024-113006","title":"The path of Chile towards the institutionalisation of evidence-based health policy","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Primary Care and Health Outcomes","field":"Health Professions","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":"Institutionalisation; Path (computing); Political science; Path dependence; Economics; Computer science; Neoclassical economics; Law","score_opus":0.32637414548648297,"score_gpt":0.5374067296486567,"score_spread":0.2110325841621737,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4408660425","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.15796345,0.008322202,0.0045052273,0.6784693,0.0013978188,0.00047780058,0.0010010933,0.0002415915,0.14762157],"genre_scores_gemma":[0.88810354,0.004295349,0.009729575,0.056218076,0.00082388433,0.00063151243,0.0006264229,0.000100163445,0.03947148],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"qualitative","domain_scores_codex":[0.9941626,0.0021690598,0.00030013206,0.0003770136,0.0010364113,0.0019548198],"domain_scores_gemma":[0.98089224,0.0070949746,0.0018068097,0.0020045715,0.0034634368,0.0047380305],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01681073,0.00039805405,0.0006050181,0.0015660885,0.002819742,0.0077993735,0.0017640591,0.005364604,0.010272266],"category_scores_gemma":[0.023226889,0.00051294704,0.0009883223,0.0018655147,0.010793692,0.004219257,0.010842863,0.011522229,0.00054064405],"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":[0.00035830835,0.0004265723,0.018887477,0.0013922927,0.00014298184,0.0011621747,0.008071459,0.0013755473,0.0013855578,0.89355016,0.029428486,0.043818943],"study_design_scores_gemma":[0.0005294113,0.0003033593,0.10396705,0.002905241,0.0000720703,0.00049475714,0.011995605,0.001612442,0.0015886496,0.15061668,0.72571665,0.00019805269],"about_ca_topic_score_codex":0.09666842,"about_ca_topic_score_gemma":0.06321699,"teacher_disagreement_score":0.9831893,"about_ca_system_score_codex":0.0344503,"about_ca_system_score_gemma":0.08366572,"threshold_uncertainty_score":0.2499556},"labels":[{"model":"gemma","categories":["sts"],"domain":null,"study_design":"qualitative","genre":"empirical","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":"low"}],"label_agreement":"split"},{"id":"W4409047603","doi":"10.1136/bmjebm-2024-113322","title":"Leading with options or issues to support purposeful shared decision-making in clinical practice","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Patient-Provider Communication in Healthcare","field":"Health Professions","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 Ottawa; Ottawa Hospital","funders":"","keywords":"Clinical Practice; Psychology; Process management; Medicine; Nursing; Business","score_opus":0.4683492103640248,"score_gpt":0.6216772883195422,"score_spread":0.15332807795551734,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4409047603","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.048668433,0.0059577,0.27658918,0.40055236,0.0075761015,0.0020665312,0.00037114037,0.0013851736,0.2568335],"genre_scores_gemma":[0.58435106,0.004495251,0.35052243,0.038262215,0.0032206574,0.0017475004,0.0005800969,0.00056422857,0.01625655],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.81014186,0.15634884,0.009851717,0.0038284522,0.015270633,0.0045584687],"domain_scores_gemma":[0.704462,0.2260134,0.019072525,0.017019408,0.015385028,0.018047707],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.09382113,0.0013327497,0.0011174434,0.0039452533,0.009596001,0.024360193,0.0041336734,0.011635072,0.032640506],"category_scores_gemma":[0.31842688,0.0010203623,0.002547498,0.0021606158,0.011549545,0.023702215,0.020991595,0.014090536,0.0075812726],"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":[0.00066082395,0.0011660154,0.016692242,0.003403668,0.0005247442,0.0054717683,0.12183062,0.0027775452,0.0020652383,0.266564,0.13294406,0.4458993],"study_design_scores_gemma":[0.00023283067,0.0003016089,0.002184139,0.0077563543,0.00028123189,0.00262614,0.05270711,0.005751936,0.0019534838,0.64620364,0.2796907,0.00031084396],"about_ca_topic_score_codex":0.0008206762,"about_ca_topic_score_gemma":0.0019113526,"teacher_disagreement_score":0.09382113,"about_ca_system_score_codex":0.0041914727,"about_ca_system_score_gemma":0.02968006,"threshold_uncertainty_score":0.49617952},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4409283522","doi":"10.1136/bmjebm-2024-113066","title":"Novel AI applications in systematic review: GPT-4 assisted data extraction, analysis, review of bias","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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":"Hospital for Sick Children; University of Toronto","funders":"","keywords":"Data extraction; Concordance; Comparability; Computer science; Meta-analysis; Systematic review; Statistics; Data mining; Artificial intelligence; Medicine; MEDLINE; Mathematics; Internal medicine; Biology","score_opus":0.8422006587199937,"score_gpt":0.6309828565757105,"score_spread":0.2112178021442832,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4409283522","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.009452806,0.013181113,0.8608833,0.006442535,0.0014328143,0.049834695,0.015436315,0.036555655,0.0067808577],"genre_scores_gemma":[0.01817814,0.0019190065,0.9368825,0.0007859995,0.00012901914,0.03902134,0.0018823736,0.00067846687,0.0005231753],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.8007211,0.15152499,0.029278196,0.0069291373,0.010744833,0.00080175407],"domain_scores_gemma":[0.47805986,0.4081268,0.03319387,0.04305436,0.03613365,0.0014315535],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.2317545,0.004182814,0.0051250076,0.013046492,0.0015706157,0.0087899165,0.0042043733,0.0035161767,0.023596426],"category_scores_gemma":[0.521881,0.004200181,0.013335892,0.012320661,0.001769624,0.006358456,0.0091581885,0.0034719654,0.005608978],"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":[0.00627632,0.00029298846,0.009082122,0.19386835,0.020070095,0.0007298156,0.0070368885,0.022174908,0.005420822,0.020496363,0.062596835,0.65195453],"study_design_scores_gemma":[0.020151004,0.0035214461,0.017852128,0.09323405,0.04453305,0.0036549894,0.002313534,0.32976732,0.017800802,0.15984704,0.30581045,0.0015141933],"about_ca_topic_score_codex":0.0030451282,"about_ca_topic_score_gemma":0.007958881,"teacher_disagreement_score":0.7682455,"about_ca_system_score_codex":0.005343271,"about_ca_system_score_gemma":0.021317424,"threshold_uncertainty_score":0.94738346},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"simulation_or_modeling","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4409642815","doi":"10.1136/bmjebm-2024-113171","title":"Challenges in the selection and measurement of outcomes in psychiatric trials","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","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":"McMaster University; St. Joseph’s Healthcare Hamilton; Impact","funders":"Danmarks Frie Forskningsfond","keywords":"Selection (genetic algorithm); Psychiatry; Psychology; Medicine; Computer science; Artificial intelligence","score_opus":0.7992621237512433,"score_gpt":0.5470442325903778,"score_spread":0.25221789116086546,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4409642815","genre_codex":"commentary","genre_gemma":"methods","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.020767212,0.13934742,0.2360535,0.5680715,0.014378759,0.0054360665,0.0042840834,0.00046498625,0.0111965155],"genre_scores_gemma":[0.4430617,0.04514427,0.34240398,0.11301027,0.027111921,0.024472658,0.0027898906,0.0007054493,0.0012998172],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.08867963,0.7653288,0.09537622,0.009380509,0.039093614,0.002141162],"domain_scores_gemma":[0.018042421,0.9220448,0.01979847,0.0146326665,0.023956036,0.0015255166],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.82562965,0.0021213393,0.014003844,0.01313898,0.0042695478,0.023155209,0.011267093,0.011243406,0.0034643272],"category_scores_gemma":[0.9448464,0.0037142283,0.00667528,0.019504793,0.019864623,0.017221402,0.013494519,0.022587707,0.0017435555],"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":[0.0032095404,0.000791259,0.0792201,0.033096477,0.015036001,0.0005083747,0.008476363,0.012994462,0.0004774149,0.15660025,0.120254196,0.56933564],"study_design_scores_gemma":[0.0026992857,0.0011696275,0.04886525,0.03934029,0.0042566606,0.0010363472,0.0041211825,0.031892784,0.0010226903,0.78965205,0.075121306,0.00082263054],"about_ca_topic_score_codex":0.012551803,"about_ca_topic_score_gemma":0.014437365,"teacher_disagreement_score":0.17437035,"about_ca_system_score_codex":0.012870979,"about_ca_system_score_gemma":0.03057786,"threshold_uncertainty_score":0.21502972},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"design_other","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4410343249","doi":"10.1136/bmjebm-2025-113825","title":"Reporting guideline for the use of Generative Artificial intelligence tools in MEdical Research: the GAMER Statement","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Artificial Intelligence in Healthcare and Education","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":"Queen's University","funders":"Chinese Academy of Medical Sciences; Lanzhou University","keywords":"Declaration; Delphi method; Checklist; Transparency (behavior); Delphi; Computer science; Guideline; Artificial intelligence; Medical education; Psychology; Data science; Medicine; Computer security","score_opus":0.8920578848643578,"score_gpt":0.6593057290771102,"score_spread":0.23275215578724762,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4410343249","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0037222828,0.03021605,0.43875304,0.24798195,0.020725718,0.180228,0.022085654,0.0062613175,0.05002597],"genre_scores_gemma":[0.011420136,0.01859069,0.71099055,0.037713487,0.0014364532,0.19843927,0.011226819,0.00088226766,0.009300355],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.44349295,0.34196058,0.16213919,0.0042725527,0.043198522,0.004936214],"domain_scores_gemma":[0.39958072,0.3406198,0.061522182,0.03692564,0.15628316,0.0050685615],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.334614,0.0038785895,0.004899815,0.017937342,0.0038578291,0.009714328,0.012794102,0.019613108,0.009020527],"category_scores_gemma":[0.52557665,0.0041759317,0.013603811,0.009297135,0.006658055,0.0073250183,0.016656797,0.013575604,0.010682066],"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":[0.00061138975,0.0004243801,0.0025342007,0.07782282,0.0006190959,0.0009725026,0.010732465,0.0027068593,0.0029333136,0.04426884,0.5796674,0.27670676],"study_design_scores_gemma":[0.0011584506,0.0004699369,0.0040472173,0.19542964,0.0008119739,0.0019123489,0.003788555,0.0036717553,0.003136403,0.03693591,0.7481069,0.000530866],"about_ca_topic_score_codex":0.008416137,"about_ca_topic_score_gemma":0.0078035584,"teacher_disagreement_score":0.66538596,"about_ca_system_score_codex":0.01079771,"about_ca_system_score_gemma":0.057337176,"threshold_uncertainty_score":0.8205394},"labels":[],"label_agreement":null},{"id":"W4411037095","doi":"10.1136/bmjebm-2025-113804","title":"Top 15 Choosing Wisely international campaign recommendations to reduce low-value care","year":2025,"lang":"en","type":"letter","venue":"BMJ evidence-based medicine","topic":"Healthcare cost, quality, practices","field":"Health Professions","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":"Public Health Ontario; University of Toronto; St. Michael's Hospital","funders":"","keywords":"Value (mathematics); Business; Computer science","score_opus":0.6366719556149726,"score_gpt":0.6106770473141256,"score_spread":0.02599490830084694,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411037095","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005599976,0.0005017489,0.00010137303,0.97679543,0.011370497,0.00004241676,0.00018825779,0.000029515302,0.010410845],"genre_scores_gemma":[0.0044296873,0.0006692067,0.00061765435,0.9658075,0.014687993,0.00012842764,0.00012616387,0.000052044685,0.013481334],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9847681,0.004472,0.0024612811,0.0008821108,0.004436095,0.0029804634],"domain_scores_gemma":[0.92496085,0.03842507,0.0030287956,0.0010828979,0.013559294,0.01894304],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.016004693,0.00086750725,0.0014037417,0.0023733883,0.009124353,0.009777396,0.0016793564,0.061271675,0.054503895],"category_scores_gemma":[0.10131521,0.0013005326,0.001908999,0.0016871025,0.0028147493,0.0046187965,0.0031952702,0.03252187,0.021864235],"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":[0.000049881273,0.000058701513,0.0013251452,0.00008158429,0.000018932598,0.00022131484,0.00016479284,0.00003053712,0.00007587757,0.0012276798,0.9839416,0.012803872],"study_design_scores_gemma":[0.00051172404,0.00010204167,0.0064386926,0.0018585685,0.00011313051,0.0004544188,0.0019037059,0.0005672321,0.00028691144,0.009144898,0.9784788,0.00013984749],"about_ca_topic_score_codex":0.029154293,"about_ca_topic_score_gemma":0.0788245,"teacher_disagreement_score":0.061271675,"about_ca_system_score_codex":0.008842443,"about_ca_system_score_gemma":0.023406154,"threshold_uncertainty_score":0.18233365},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"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":"W4411952500","doi":"10.1136/bmjebm-2024-113617","title":"Understanding synthetic data: artificial datasets for real-world evidence","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Machine Learning in Healthcare","field":"Computer Science","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":"Computer science; Artificial intelligence","score_opus":0.5243495068028722,"score_gpt":0.4894084368243749,"score_spread":0.034941069978497274,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411952500","genre_codex":"methods","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.27947,0.0070222216,0.59066033,0.017918708,0.0016622882,0.0019290413,0.085285485,0.002503688,0.013548218],"genre_scores_gemma":[0.6049941,0.0018768616,0.34483418,0.0014954386,0.0003484787,0.0012454136,0.043811146,0.00019222783,0.00120217],"study_design_codex":"simulation_or_modeling","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9913707,0.0061075925,0.0007986077,0.00069713086,0.0008994123,0.00012651339],"domain_scores_gemma":[0.89150876,0.095056444,0.0028470505,0.006984324,0.0027961524,0.00080729683],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.011632906,0.00082998927,0.0007280022,0.003205444,0.0005665608,0.0032440217,0.0020033533,0.0023562338,0.005670403],"category_scores_gemma":[0.09952826,0.0004781263,0.0012332946,0.0027939794,0.0013259875,0.002661357,0.0017792586,0.0022345835,0.0007278434],"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":[0.0021375797,0.0015138459,0.048271976,0.0069508725,0.0013275413,0.001700505,0.0015685715,0.4821263,0.0026708522,0.11708705,0.08907125,0.24557376],"study_design_scores_gemma":[0.00036988797,0.00038479693,0.0053248433,0.0012950863,0.00023940825,0.0007808478,0.0010327653,0.717244,0.002411223,0.22646198,0.044339947,0.00011511081],"about_ca_topic_score_codex":0.0022991607,"about_ca_topic_score_gemma":0.0035110645,"teacher_disagreement_score":0.9883671,"about_ca_system_score_codex":0.001442333,"about_ca_system_score_gemma":0.0016913515,"threshold_uncertainty_score":0.06152141},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4412521978","doi":"10.1136/bmjebm-2025-113715","title":"Co-production and implementation of an evidence collation strategy for a novel point-of-care information resource: gpevidence.org","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Biomedical Text Mining and Ontologies","field":"Biochemistry, Genetics and Molecular Biology","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":"National Institute for Health and Care Research","keywords":"Collation; Production (economics); Resource (disambiguation); Point (geometry); Information resource; Computer science; Business; Knowledge management; Mathematics; Economics","score_opus":0.07713404106508719,"score_gpt":0.4283125794856482,"score_spread":0.351178538420561,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4412521978","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.016910229,0.003506443,0.6462377,0.11919771,0.008517009,0.0230829,0.05949875,0.046981808,0.0760675],"genre_scores_gemma":[0.047798418,0.001310749,0.87702405,0.007734573,0.002224097,0.006506118,0.036863085,0.00714143,0.013397494],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9024421,0.04754167,0.021043764,0.0069606407,0.020168077,0.0018438869],"domain_scores_gemma":[0.5006598,0.26858705,0.024115546,0.07906412,0.113717966,0.013855616],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.13448413,0.0020271249,0.002775159,0.025949366,0.0027677643,0.012031636,0.004761803,0.0048266947,0.06316456],"category_scores_gemma":[0.4032358,0.0018657481,0.0033754797,0.017724864,0.0015183946,0.011043658,0.017267222,0.005219928,0.03252321],"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":[0.0009270587,0.00047433545,0.008132983,0.005422674,0.00069533545,0.0020086428,0.0035576248,0.0013893388,0.0038366637,0.008943129,0.44485143,0.51976097],"study_design_scores_gemma":[0.0011929739,0.0004893355,0.009349131,0.008915476,0.0013775195,0.001990768,0.003005616,0.013516929,0.014572475,0.037755158,0.90728176,0.0005528371],"about_ca_topic_score_codex":0.0052128113,"about_ca_topic_score_gemma":0.0066269627,"teacher_disagreement_score":0.8655159,"about_ca_system_score_codex":0.004076769,"about_ca_system_score_gemma":0.034504537,"threshold_uncertainty_score":0.71122855},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4414833709","doi":"10.1136/bmjebm-2025-113704","title":"Addressing misleading medical information on social media: a scoping review of current interventions","year":2025,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Misinformation and Its Impacts","field":"Social Sciences","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":"McMaster University; Impact","funders":"National Health and Medical Research Council; Medical Research Council","keywords":"Psychological intervention; Social media; Public health; Public policy; Medical information; MEDLINE; Digital health","score_opus":0.632896534837736,"score_gpt":0.6151385818461708,"score_spread":0.01775795299156524,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4414833709","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0002776528,0.9978397,0.0001645305,0.0006438753,0.00016968884,0.00023841126,0.00009178844,0.000008948717,0.0005654723],"genre_scores_gemma":[0.0031856098,0.9948736,0.00058644963,0.00048777912,0.0001317233,0.0005373282,0.00009533395,0.0000064211035,0.00009573079],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9830959,0.0071357135,0.005515532,0.00088723865,0.002906286,0.00045936013],"domain_scores_gemma":[0.854225,0.1263299,0.010041614,0.0013502645,0.0073343157,0.00071894156],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.023535846,0.002089112,0.006958963,0.019949814,0.0013972663,0.005511367,0.0036189093,0.0042625433,0.011486889],"category_scores_gemma":[0.11425246,0.0015175627,0.0076806303,0.013409677,0.0021733078,0.005197769,0.0033625802,0.0030263527,0.0013514818],"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":[0.00020024009,0.00007533641,0.0003101716,0.7973414,0.0017831372,0.00006437023,0.00044391528,0.00012358134,0.000101903235,0.0007134331,0.003277465,0.19556506],"study_design_scores_gemma":[0.0000595153,0.00009916356,0.0006724787,0.97127223,0.0058243545,0.00010230783,0.00029967396,0.000057486064,0.00010747607,0.00045980173,0.021028934,0.000016721184],"about_ca_topic_score_codex":0.0051107537,"about_ca_topic_score_gemma":0.011778492,"teacher_disagreement_score":0.023535846,"about_ca_system_score_codex":0.0049918527,"about_ca_system_score_gemma":0.017527351,"threshold_uncertainty_score":0.12447095},"labels":[{"model":"gemma","categories":["scholarly_communication"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4415279421","doi":"10.1136/bmjebm-2024-113391","title":"Over 1000 terms have been used to describe evidence synthesis: a scoping review","year":2025,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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; Public Health Ontario; Impact; McGill University; McMaster University; Queen's University; St. Michael's Hospital","funders":"National Health and Medical Research Council","keywords":"Identification (biology); Scope (computer science); Context (archaeology); Key (lock); Perspective (graphical); Systematic review","score_opus":0.8775033471675885,"score_gpt":0.6368221377890774,"score_spread":0.24068120937851112,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4415279421","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004041405,0.8040775,0.040115666,0.019398248,0.008946808,0.057385847,0.02534818,0.001064716,0.039621536],"genre_scores_gemma":[0.015554937,0.7492616,0.08114323,0.006124332,0.0014620991,0.12872264,0.011421742,0.00062971545,0.0056797],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.7834496,0.07119558,0.11212074,0.0056794067,0.025099477,0.0024552569],"domain_scores_gemma":[0.739401,0.16617078,0.049085785,0.007843337,0.035483707,0.002015376],"candidate_categories":["metaresearch","bibliometrics"],"consensus_categories":[],"category_scores_codex":[0.12752621,0.006085936,0.014588946,0.116568,0.0061164037,0.020054346,0.0042735655,0.010392447,0.035347786],"category_scores_gemma":[0.31110913,0.0029106955,0.010985989,0.13876234,0.0058771083,0.021236228,0.012666381,0.0061735925,0.015418391],"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":[0.00055996625,0.000085999076,0.0015282544,0.5010483,0.0020728852,0.00041018202,0.004692758,0.00066447107,0.0010412743,0.01697347,0.06149321,0.40942916],"study_design_scores_gemma":[0.000149549,0.000109925495,0.0018983507,0.78694147,0.002528267,0.00038311406,0.0013895468,0.00029874605,0.00040058044,0.0073792045,0.1983999,0.000121335754],"about_ca_topic_score_codex":0.00451181,"about_ca_topic_score_gemma":0.0061413622,"teacher_disagreement_score":0.88343203,"about_ca_system_score_codex":0.015909968,"about_ca_system_score_gemma":0.050272513,"threshold_uncertainty_score":0.6744311},"labels":[{"model":"gemma","categories":["metaresearch","bibliometrics"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4415406762","doi":"10.1136/bmjebm-2025-113675","title":"Comparing traditional and AI-enhanced strategies for developing patient decision aids: a multiple case study","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Healthcare Decision-Making and Restraints","field":"Psychology","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 Laval","funders":"","keywords":"Context (archaeology); Key (lock); Data collection; Identification (biology)","score_opus":0.24972038070113398,"score_gpt":0.478647880144518,"score_spread":0.228927499443384,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4415406762","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9680127,0.0014457863,0.014135048,0.002744217,0.00008528809,0.006805365,0.00013370896,0.00003252774,0.006605378],"genre_scores_gemma":[0.92654556,0.0014882933,0.062349852,0.0007493085,0.000044440196,0.007794051,0.00006866707,0.000024997633,0.00093492406],"study_design_codex":"qualitative","study_design_gemma":"qualitative","domain_scores_codex":[0.8739438,0.11006522,0.005675076,0.0020095413,0.0059380946,0.002368242],"domain_scores_gemma":[0.81140167,0.16447884,0.009175586,0.0048167063,0.0072411117,0.0028860217],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.09077914,0.00094722275,0.0010251527,0.0036857636,0.0052087056,0.0041557364,0.0034316317,0.0029271662,0.0028317862],"category_scores_gemma":[0.1536218,0.0008442735,0.0013574486,0.0025577242,0.0045717764,0.00674498,0.0065802303,0.0028521775,0.00033021413],"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":[0.0026267513,0.012258609,0.06248174,0.008276203,0.000387942,0.01342634,0.61549586,0.0031122395,0.0024271896,0.01864081,0.0037394238,0.25712687],"study_design_scores_gemma":[0.0037687698,0.02535821,0.032002557,0.010629883,0.00072292745,0.016796526,0.7954669,0.013308698,0.014242816,0.014932713,0.072223686,0.0005464205],"about_ca_topic_score_codex":0.0024362796,"about_ca_topic_score_gemma":0.0058228434,"teacher_disagreement_score":0.09077914,"about_ca_system_score_codex":0.010732903,"about_ca_system_score_gemma":0.008323847,"threshold_uncertainty_score":0.4800917},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"qualitative","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4415524767","doi":"10.1136/bmjebm-2025-114275","title":"Correspondence on ‘Less is more for patients, practitioners, public and planet: a taxonomy for the harms of too much medicine’","year":2025,"lang":"en","type":"letter","venue":"BMJ evidence-based medicine","topic":"Pharmaceutical industry and healthcare","field":"Pharmacology, Toxicology and Pharmaceutics","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 Addiction and Mental Health","funders":"","keywords":"Taxonomy (biology); Focus (optics)","score_opus":0.6890357720182831,"score_gpt":0.5719885798623117,"score_spread":0.11704719215597148,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4415524767","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00020290323,0.00046521839,0.000054935834,0.9782284,0.016448256,0.000026267735,0.000090493406,0.000018040058,0.004465392],"genre_scores_gemma":[0.0012107439,0.00026579338,0.00013691594,0.9698561,0.018708752,0.00004623807,0.000026477745,0.000020532132,0.009728472],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9891623,0.0031233216,0.0015266283,0.0011459336,0.0034900887,0.0015516735],"domain_scores_gemma":[0.9705128,0.01852221,0.001614448,0.0005912111,0.0037502693,0.00500921],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.007969844,0.0013258354,0.0024091732,0.0015564515,0.008739595,0.008072023,0.0023097575,0.09325087,0.021887759],"category_scores_gemma":[0.051646307,0.0014958404,0.0022867739,0.0017168501,0.004633719,0.0068348898,0.002881207,0.06263643,0.015136186],"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":[0.000045218287,0.000025198122,0.00024214145,0.000050439452,0.00001107359,0.00054502225,0.00019489725,0.000021836002,0.00010539286,0.0024273256,0.9930608,0.0032707185],"study_design_scores_gemma":[0.00019711372,0.00011015304,0.002084639,0.0006892633,0.0000350471,0.001063407,0.0010550445,0.00037673826,0.00016873107,0.012021177,0.9820809,0.000117859345],"about_ca_topic_score_codex":0.014888869,"about_ca_topic_score_gemma":0.034093264,"teacher_disagreement_score":0.09325087,"about_ca_system_score_codex":0.008321328,"about_ca_system_score_gemma":0.010634581,"threshold_uncertainty_score":0.0732218},"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":"low"}],"label_agreement":"agree"},{"id":"W4416430973","doi":"10.1136/bmjebm-2025-113904","title":"A primer to planetary health in evidence-based medicine and clinical decision-making","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Climate Change and Health Impacts","field":"Environmental Science","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":"Impact; Queen's University; McMaster University; Western University","funders":"","keywords":"Primer (cosmetics); MEDLINE; Public health; Biomedicine","score_opus":0.23348970559746515,"score_gpt":0.5002367385224676,"score_spread":0.26674703292500246,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4416430973","genre_codex":"commentary","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00021414834,0.046725158,0.07631591,0.852615,0.012575614,0.00017558922,0.0000656626,0.0001670941,0.011145798],"genre_scores_gemma":[0.03630053,0.09442542,0.32941335,0.48625723,0.041315068,0.00254872,0.00021849542,0.00054483255,0.008976314],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.887443,0.09596694,0.006232256,0.0019108944,0.007142166,0.0013047532],"domain_scores_gemma":[0.53916705,0.4294368,0.004948474,0.0046978295,0.013891091,0.007858765],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.13244633,0.0019640622,0.0035323836,0.006845674,0.0047856825,0.020168269,0.0059381216,0.030672738,0.012976755],"category_scores_gemma":[0.21432796,0.0020642497,0.0032815181,0.0043426976,0.03465215,0.028627213,0.0132523775,0.066677734,0.0040364633],"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":[0.00013764222,0.00025900154,0.00039713096,0.003849019,0.00021832073,0.0005158171,0.010944824,0.0012902247,0.0002891262,0.66344666,0.18280661,0.13584557],"study_design_scores_gemma":[0.00010956404,0.000061132145,0.00017095885,0.009873537,0.000055955457,0.00033293248,0.002225212,0.0009784184,0.00014371859,0.6111584,0.37481174,0.00007842026],"about_ca_topic_score_codex":0.003462917,"about_ca_topic_score_gemma":0.0038009374,"teacher_disagreement_score":0.13244633,"about_ca_system_score_codex":0.0071207876,"about_ca_system_score_gemma":0.020294685,"threshold_uncertainty_score":0.7004515},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"methods","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":"low"}],"label_agreement":"agree"},{"id":"W4417300166","doi":"10.1136/bmjebm-2025-113997","title":"Reporting GUideline for Intervention DEscription in Rehabilitation (GUIDE-Rehab): a tool to open the ‘black box’ of rehabilitation complex interventions","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Cerebral Palsy and Movement Disorders","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; Ontario Tech University; University of Alberta","funders":"","keywords":"Rehabilitation; Psychological intervention; Delphi method; Intervention (counseling); Guideline; Transparency (behavior); Quality (philosophy); Rehabilitation counseling","score_opus":0.16649353114868318,"score_gpt":0.47692042697819914,"score_spread":0.310426895829516,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4417300166","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0034207446,0.021035042,0.41444463,0.118888296,0.018142248,0.3176941,0.05158798,0.0086648865,0.046121992],"genre_scores_gemma":[0.0065130526,0.011071973,0.66658944,0.01782789,0.000957684,0.26491833,0.02210348,0.0013323169,0.008685781],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.5318612,0.28225222,0.14315301,0.004442582,0.03528221,0.0030087058],"domain_scores_gemma":[0.5311801,0.26112106,0.045376077,0.029779049,0.128643,0.0039007233],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.2932789,0.002938864,0.0051969336,0.012440874,0.0031310932,0.0074464944,0.007662915,0.009830218,0.015207691],"category_scores_gemma":[0.4231031,0.004069697,0.0084757,0.0100201955,0.0041150223,0.0075179352,0.009196592,0.011432681,0.01548667],"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":[0.0003808375,0.00024687062,0.0011543646,0.033528943,0.00033305993,0.0003243953,0.0056472453,0.0010602217,0.001203398,0.019329792,0.6925323,0.24425851],"study_design_scores_gemma":[0.0010747823,0.0003002746,0.0027269148,0.05323368,0.00034434797,0.00046761427,0.0021539992,0.0011939202,0.0016113927,0.01602421,0.92053276,0.00033613676],"about_ca_topic_score_codex":0.0122162085,"about_ca_topic_score_gemma":0.014456953,"teacher_disagreement_score":0.70672107,"about_ca_system_score_codex":0.010366112,"about_ca_system_score_gemma":0.053664166,"threshold_uncertainty_score":0.87151295},"labels":[],"label_agreement":null},{"id":"W4417321183","doi":"10.1136/bmjebm-2025-113925","title":"Current knowledge on the adaptation of clinical practice guidelines: a scoping review","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Clinical practice guidelines implementation","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; Hamilton Health Sciences","funders":"Chinese University of Hong Kong","keywords":"Adaptation (eye); Quality (philosophy); Clinical Practice; Current (fluid); Context (archaeology)","score_opus":0.7839957976028011,"score_gpt":0.70014721165082,"score_spread":0.08384858595198119,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4417321183","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0012569696,0.97987854,0.0036383036,0.007669634,0.0012854886,0.0019321926,0.00037946063,0.00006307881,0.0038963621],"genre_scores_gemma":[0.01517849,0.96204597,0.013969006,0.002866247,0.0004757558,0.004489533,0.0006376092,0.0000476868,0.00028964208],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.79758316,0.09843397,0.066286005,0.006321622,0.029159924,0.0022154166],"domain_scores_gemma":[0.38168874,0.50707406,0.039508186,0.016156271,0.053285558,0.0022872337],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.19152817,0.0024523817,0.0068133147,0.04529819,0.0035371028,0.013688004,0.0062327473,0.008514044,0.0067973966],"category_scores_gemma":[0.5135931,0.0034156556,0.0071653253,0.050888542,0.0059334286,0.017348891,0.009299134,0.0064549916,0.0019062889],"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":[0.00011216086,0.00010604017,0.0014543597,0.47292224,0.0010856607,0.00018427115,0.0051099495,0.0005501258,0.00026626835,0.0055447565,0.013512079,0.49915212],"study_design_scores_gemma":[0.0000328684,0.00003958954,0.0010033391,0.9459003,0.0012764233,0.00013698968,0.0014431077,0.00018790133,0.00013468102,0.0018830228,0.047927696,0.000034215722],"about_ca_topic_score_codex":0.011741236,"about_ca_topic_score_gemma":0.018611861,"teacher_disagreement_score":0.8084718,"about_ca_system_score_codex":0.013932986,"about_ca_system_score_gemma":0.07546607,"threshold_uncertainty_score":0.9969897},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"}]}