{"meta":{"page":1,"per_page":50,"max_per_page":100,"total":78,"total_is_capped":false,"direct_labels_cover":73,"predictions_cover":78,"direct_label_status":"direct model label, unvalidated","prediction_status":"machine_predicted_unvalidated (Codex and Gemma teacher distillation)","score_status":"score_only:v0-immature-baseline (scores rank; they never assert a category)","snapshot":{"source":"OpenAlex, pinned release, all 482 partitions","release":"2026-06-24","frame_built":"2026-07-12"},"query_hash":"5ba9182f63f0","filters":{"venue":"Journal of Evidence-Based Medicine"}},"results":[{"id":"W1709885097","doi":"10.1111/jebm.12141","title":"The methodological quality assessment tools for preclinical and clinical studies, systematic review and meta‐analysis, and clinical practice guideline: a systematic review","year":2015,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Animal testing and alternatives","field":"Veterinary","cited_by":2088,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Systematic review; Critical appraisal; Checklist; Medicine; Guideline; Randomized controlled trial; Psychological intervention; Clinical study design; MEDLINE; Medical physics; Clinical trial; Psychology; Alternative medicine; Nursing; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.9641987842621917,"gpt":0.7638045496857743,"spread":0.2003942345764175,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.3204975,0.0009497236,0.0344387,0.0001941056,0.0002521701,0.0001514012,0.0009166218,0.0004068405,0.00001560622],"category_scores_gemma":[0.7025666,0.0003525215,0.003880983,0.0006425223,0.001180856,0.0005167266,0.0003606952,0.002088655,0.00000210262],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001708288,"about_ca_system_score_gemma":0.001111126,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001113107,"about_ca_topic_score_gemma":0.000003412535,"domain_scores_codex":[0.9236153,0.05329663,0.02000862,0.0009977978,0.001663682,0.0004179383],"domain_scores_gemma":[0.4999137,0.4677504,0.02532971,0.00135611,0.0048337,0.0008163518],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","study_design_scores_codex":[0.0001511486,0.0001244326,0.000005977323,0.8821238,0.1078506,0.0001148351,0.000009434597,3.153347e-8,1.005337e-8,0.00006160371,0.002930754,0.006627351],"study_design_scores_gemma":[0.0003486073,0.00393047,0.000003962003,0.4914607,0.4866251,0.0008291285,0.0001195,0.00001517728,1.084492e-9,0.00005828783,0.01641069,0.0001983831],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[6.159322e-7,0.9700215,0.003255235,0.01791957,0.0003179767,0.008416981,0.00003048807,0.00002468812,0.00001292525],"genre_scores_gemma":[0.000002642612,0.9538238,0.03742224,0.007451485,0.0006864537,0.0004740364,0.000009973553,0.00004277325,0.00008657353],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.4237016,"threshold_uncertainty_score":0.9998927,"prediction_status":"machine_predicted_unvalidated"},"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":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W3117672471","doi":"10.1111/j.1756-5391.2010.01086.x","title":"Revised STandards for Reporting Interventions in Clinical Trials of Acupuncture (STRICTA): Extending the CONSORT Statement","year":2010,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Acupuncture Treatment Research Studies","field":"Medicine","cited_by":443,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Ottawa Hospital; University of Ottawa","funders":"National Institute for Health and Care Research","keywords":"Checklist; Consolidated Standards of Reporting Trials; Psychological intervention; Clinical trial; Systematic review; Psychology; Alternative medicine; Medicine; Medical education; Family medicine; MEDLINE; Political science; Psychiatry; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.7114688404628366,"gpt":0.651952867294202,"spread":0.05951597316863466,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.1056098,0.0003542869,0.003242849,0.0005350381,0.0001066618,0.00001596093,0.0002540189,0.0001371899,0.0004097356],"category_scores_gemma":[0.4326503,0.0001013159,0.001649832,0.0004749163,0.000614477,0.0001711455,0.00005155136,0.001194182,3.683242e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002954807,"about_ca_system_score_gemma":0.001851366,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0000266534,"about_ca_topic_score_gemma":0.00004559631,"domain_scores_codex":[0.9865924,0.001012693,0.009918773,0.0002703257,0.001836471,0.0003693755],"domain_scores_gemma":[0.9503417,0.03041558,0.0142824,0.0006838232,0.003882133,0.000394379],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.02202939,0.001990434,0.5567662,0.007474781,0.003419645,0.001702796,0.001205238,0.00007713186,0.05294873,0.0004010397,0.1706523,0.1813323],"study_design_scores_gemma":[0.07786334,0.04724757,0.4647627,0.1809922,0.01859732,0.001416047,0.006655577,0.0006801456,0.02171734,0.001919807,0.1773734,0.0007745918],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.6649062,0.0440814,0.03240976,0.2456187,0.003761576,0.00882486,0.00009931976,0.00003719709,0.0002610052],"genre_scores_gemma":[0.9865871,0.005298365,0.005568483,0.0006123708,0.001592753,0.00007249289,0.00001078321,0.00002374531,0.000233878],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3270405,"threshold_uncertainty_score":0.9209629,"prediction_status":"machine_predicted_unvalidated"},"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":"W3161745299","doi":"10.1111/jebm.12427","title":"The ROBINS‐I and the NOS had similar reliability but differed in applicability: A random sampling observational studies of systematic reviews/meta‐analysis","year":2021,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":126,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Observational study; Meta-analysis; Statistics; Statistic; Reliability (semiconductor); Medicine; Systematic review; MEDLINE; Mathematics; Internal medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.96789418365725,"gpt":0.6353889867762997,"spread":0.3325051968809503,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch","metaepi_broad"],"category_scores_codex":[0.571514,0.000988608,0.06604463,0.001133563,0.0003074123,0.000362098,0.00355192,0.0002371106,0.0005029948],"category_scores_gemma":[0.6919101,0.0002375432,0.02075466,0.006824356,0.001346402,0.000290021,0.0002713922,0.001072404,0.00001123255],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002264566,"about_ca_system_score_gemma":0.0008928844,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002862961,"about_ca_topic_score_gemma":0.00007094997,"domain_scores_codex":[0.7990568,0.1244367,0.06276263,0.001497731,0.01177972,0.0004664159],"domain_scores_gemma":[0.255356,0.6435508,0.08066415,0.009041565,0.01101726,0.0003702688],"domain_codex":null,"domain_gemma":"methods","domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","study_design_scores_codex":[0.0004116682,0.0003153878,0.00119094,0.6632034,0.2497101,0.00004689022,0.001380171,0.001275674,0.000002300726,0.0009667292,0.006015617,0.07548108],"study_design_scores_gemma":[0.001656408,0.0001446151,0.0001092899,0.2155318,0.661736,0.0000567301,0.001231082,0.001206674,2.804332e-7,0.002022839,0.1159815,0.0003228727],"study_design_candidate":"meta_analysis","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00006394702,0.9758888,0.007758718,0.009399815,0.0002793829,0.006583275,0.00001380654,0.000001610364,0.0000106236],"genre_scores_gemma":[0.0004474575,0.9963246,0.0014849,0.0003901882,0.000234181,0.0007467406,0.000004852131,0.00001887611,0.0003482062],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.5437008,"threshold_uncertainty_score":0.9890522,"prediction_status":"machine_predicted_unvalidated"},"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":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W3095207500","doi":"10.1111/jebm.12412","title":"Association between homocysteine and obesity: A meta‐analysis","year":2020,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Folate and B Vitamins Research","field":"Medicine","cited_by":64,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"Yunnan Provincial Science and Technology Department; National Natural Science Foundation of China","keywords":"Homocysteine; Medicine; Meta-analysis; Internal medicine; Obesity; Cochrane Library; Insulin resistance; Confidence interval; Strictly standardized mean difference; Endocrinology","retraction":null,"screen_n_in":null,"score":{"opus":0.4488428679630203,"gpt":0.4667018673714734,"spread":0.01785899940845309,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.005026523,0.0005333817,0.009962489,0.002051295,0.00007854329,0.00004316536,0.0003330246,0.0004774626,0.0006176743],"category_scores_gemma":[0.006333503,0.0002662966,0.002976988,0.002931031,0.0001400678,0.0002406754,0.00007493433,0.001749593,0.00002678088],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000602469,"about_ca_system_score_gemma":0.0015715,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00006190048,"about_ca_topic_score_gemma":0.000008013489,"domain_scores_codex":[0.9939369,0.000789605,0.002084671,0.0004138455,0.00237228,0.0004026747],"domain_scores_gemma":[0.9906617,0.004497972,0.00249151,0.0004132773,0.001090506,0.0008449747],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","study_design_scores_codex":[0.0001595894,0.0001271598,0.002986464,0.03290238,0.5446914,0.002074521,0.00009257484,0.00000116573,0.000008053829,0.000003619346,0.007945421,0.4090076],"study_design_scores_gemma":[0.000690377,0.00171252,0.0002157377,0.01479762,0.524861,0.0001147039,0.00001254823,0.00001064174,0.000002934567,0.000004995404,0.457445,0.0001319112],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00004330122,0.9775575,0.0002672539,0.02131098,0.0001024319,0.000615218,0.00002211815,0.00001676453,0.00006440054],"genre_scores_gemma":[0.0003445629,0.9949504,0.0003719493,0.000489132,0.002356578,0.00001395259,0.00004592152,0.00004625332,0.001381248],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.4494996,"threshold_uncertainty_score":0.9999789,"prediction_status":"machine_predicted_unvalidated"},"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":"W4241679221","doi":"10.1111/jebm.12104","title":"Hip protectors for preventing hip fractures in older people","year":2014,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Bone health and osteoporosis research","field":"Medicine","cited_by":50,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University","funders":"","keywords":"Hip fracture; Older people; Medicine; Physical medicine and rehabilitation; Gerontology; Osteoporosis; Internal medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.1074336176220591,"gpt":0.4192378048710034,"spread":0.3118041872489443,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.006309673,0.0001945377,0.0008477428,0.0006644318,0.00008582108,0.00001377148,0.000195054,0.0001563975,0.0003528028],"category_scores_gemma":[0.01700872,0.0001122298,0.0001833484,0.0004969383,0.0001018516,0.0002124491,0.00001860438,0.0008168106,0.000004368659],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001884103,"about_ca_system_score_gemma":0.001019761,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001516291,"about_ca_topic_score_gemma":0.00007144232,"domain_scores_codex":[0.9968007,0.0002632838,0.00114206,0.0002245819,0.00103419,0.000535209],"domain_scores_gemma":[0.9953334,0.002475312,0.0006441696,0.0002767151,0.000717089,0.0005533398],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.01403021,0.001342505,0.5128237,0.01455903,0.0001515486,0.0002214657,0.003737014,0.0002888922,0.06713805,0.00006996591,0.05866656,0.3269711],"study_design_scores_gemma":[0.02402977,0.01383561,0.8805878,0.03599247,0.0003229738,0.0002641458,0.0008658898,0.003351524,0.01968716,0.0006101166,0.02012557,0.0003270271],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8862837,0.006103132,0.006454184,0.09841275,0.0005320223,0.002066588,0.000001036311,0.00001993927,0.0001266697],"genre_scores_gemma":[0.9915979,0.0002352269,0.002288247,0.003702631,0.001801536,0.00006252754,0.000003403154,0.00003035563,0.0002781898],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3677641,"threshold_uncertainty_score":0.9912714,"prediction_status":"machine_predicted_unvalidated"},"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":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W3100141195","doi":"10.1111/jebm.12418","title":"Analytics with artificial intelligence to advance the treatment of acute respiratory distress syndrome","year":2020,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Respiratory Support and Mechanisms","field":"Medicine","cited_by":47,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Health Sciences Centre; University of Alberta Hospital; University of Alberta","funders":"National Natural Science Foundation of China","keywords":"ARDS; Analytics; Medicine; Population; Artificial intelligence; Big data; Identification (biology); Computer science; Intensive care medicine; Machine learning; Data science; Data mining; Internal medicine; Lung","retraction":null,"screen_n_in":null,"score":{"opus":0.2641202297817875,"gpt":0.4285162414375897,"spread":0.1643960116558022,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.001002953,0.0006554322,0.004181384,0.0005933775,0.00007374866,0.00001745453,0.0005709057,0.0002038187,0.0001810335],"category_scores_gemma":[0.000660597,0.0002638245,0.0007913758,0.001363359,0.0003757396,0.0001386025,0.0000283067,0.0006253653,0.0000167995],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004744037,"about_ca_system_score_gemma":0.002649095,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000005583643,"about_ca_topic_score_gemma":0.000004760488,"domain_scores_codex":[0.9949851,0.000251152,0.002471025,0.00042257,0.001497185,0.0003729659],"domain_scores_gemma":[0.994226,0.0009662953,0.00243308,0.00084069,0.0007591883,0.0007747319],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.001258689,0.0003009985,0.00004282465,0.00482628,0.001309169,0.007409476,0.0001698819,0.00003317673,0.00006076211,0.0000775716,0.0003812678,0.9841299],"study_design_scores_gemma":[0.0003931123,0.03760044,0.00001072089,0.09303724,0.01326275,0.001261835,0.0002163175,0.000011614,0.0004535214,0.00003528078,0.8534769,0.0002403038],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.0001959294,0.9802514,0.01310181,0.004356873,0.000542504,0.001471598,0.00004353584,0.00001417006,0.00002217846],"genre_scores_gemma":[0.000633518,0.995491,0.001327173,0.001053214,0.001250978,0.00004474549,0.0000122902,0.00007194059,0.0001150946],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.9838896,"threshold_uncertainty_score":0.9999814,"prediction_status":"machine_predicted_unvalidated"},"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":"low"}],"label_agreement":"split"},{"id":"W2947332497","doi":"10.1111/jebm.12346","title":"A Chinese practice guideline of the assisted reproductive technology strategies for women with advanced age","year":2019,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Reproductive Health and Technologies","field":"Medicine","cited_by":46,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Hamilton Health Sciences; Impact","funders":"","keywords":"Guideline; Fertility; Medicine; Grading (engineering); Family medicine; Assisted reproductive technology; Delphi method; Infertility; Reproductive medicine; Population; Medical education; Pregnancy; Environmental health; Engineering","retraction":null,"screen_n_in":null,"score":{"opus":0.06425853403512855,"gpt":0.4064706327631867,"spread":0.3422120987280581,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.003043845,0.0002400073,0.001029617,0.0005230912,0.00006283757,0.000006869534,0.0003631628,0.0001480275,0.0000392252],"category_scores_gemma":[0.03123615,0.00009914461,0.0001035008,0.001392878,0.0006995911,0.0005055543,0.00004305758,0.0006550971,0.000001109072],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000245997,"about_ca_system_score_gemma":0.001731549,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001619185,"about_ca_topic_score_gemma":0.000003700301,"domain_scores_codex":[0.9972121,0.0001043343,0.00118497,0.0003917476,0.0007591099,0.0003477498],"domain_scores_gemma":[0.9921401,0.000993566,0.002125824,0.001085831,0.003537103,0.0001176404],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"bench_or_experimental","study_design_gemma":"observational","study_design_scores_codex":[0.09906268,0.001759784,0.0762459,0.005356784,0.001449945,0.0005686959,0.002781333,0.001579483,0.6641924,0.002885928,0.009614644,0.1345024],"study_design_scores_gemma":[0.05925487,0.182358,0.2374599,0.04262249,0.002537601,0.008997114,0.1133506,0.0002492335,0.08323428,0.03356384,0.2351962,0.001175877],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7713577,0.005497706,0.001899751,0.2185426,0.0004998578,0.001625347,0.00000243001,0.0000502389,0.0005244077],"genre_scores_gemma":[0.9727448,0.0003864689,0.02437622,0.001658056,0.0004531494,0.00005511143,0.000001221444,0.00002600008,0.0002989802],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5809581,"threshold_uncertainty_score":0.9769242,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4386592036","doi":"10.1111/jebm.12547","title":"Artificial intelligence and digital health in improving primary health care service delivery in LMICs: A systematic review","year":2023,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Mobile Health and mHealth Applications","field":"Health Professions","cited_by":44,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Psychological intervention; Digital health; Medicine; Health care; Service delivery framework; mHealth; Nursing; Service (business); Business; Political science","retraction":null,"screen_n_in":null,"score":{"opus":0.33261185357091,"gpt":0.518243514726389,"spread":0.185631661155479,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.01505237,0.0006162039,0.007700081,0.001666078,0.0003544439,0.00001860999,0.0006829533,0.0004013172,0.00002521266],"category_scores_gemma":[0.007018471,0.0003980882,0.0002619323,0.002801334,0.0001052617,0.000378321,0.0001497012,0.003374334,0.00006415672],"about_ca_system_candidate":true,"about_ca_system_consensus":true,"about_ca_system_score_codex":0.004439063,"about_ca_system_score_gemma":0.03068569,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001449938,"about_ca_topic_score_gemma":0.002210553,"domain_scores_codex":[0.9828646,0.003066616,0.01133181,0.0006337381,0.0009107849,0.001192437],"domain_scores_gemma":[0.9794487,0.008599869,0.009219076,0.0007409269,0.0008187451,0.001172673],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","study_design_scores_codex":[0.00002669994,0.00003166445,0.00001296947,0.6204133,0.00001183325,0.00002124663,0.0004721627,0.000001648038,1.525336e-8,0.00002999591,0.0008812996,0.3780972],"study_design_scores_gemma":[0.0002566552,0.0006300093,0.00001326955,0.9399476,0.000296667,0.00005307494,0.001678308,0.00002778026,6.072592e-9,0.00008338531,0.05680018,0.0002130468],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00000342798,0.9365954,0.0003416832,0.04723164,0.0006131344,0.01510361,0.00006112722,0.00003665104,0.00001335564],"genre_scores_gemma":[0.00001371096,0.963921,0.0002562313,0.03236749,0.0007376351,0.002505659,0.00009234017,0.00008565705,0.00002025517],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.3778841,"threshold_uncertainty_score":0.9998471,"prediction_status":"machine_predicted_unvalidated"},"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":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W3027694758","doi":"10.1111/jebm.12384","title":"Clinical research methods for treatment, diagnosis, prognosis, etiology, screening, and prevention: A narrative review","year":2020,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":41,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Juravinski Cancer Centre; McMaster University; Impact","funders":"","keywords":"Narrative review; Perspective (graphical); Etiology; Narrative; Systematic review; Medicine; Quality (philosophy); Psychology; MEDLINE; Psychotherapist; Psychiatry; Computer science; Political science","retraction":null,"screen_n_in":null,"score":{"opus":0.9833962916424582,"gpt":0.8007399255518473,"spread":0.182656366090611,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.5415257,0.0009062331,0.03035106,0.00138262,0.0003067842,0.0002957318,0.002804018,0.0004872399,0.003300256],"category_scores_gemma":[0.5882525,0.0003136201,0.00961421,0.003546268,0.0008709453,0.0004065932,0.0001799422,0.001191215,0.0000745385],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002014817,"about_ca_system_score_gemma":0.002552812,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000005078867,"about_ca_topic_score_gemma":0.000005929599,"domain_scores_codex":[0.7735262,0.1765483,0.04037466,0.002113679,0.006812871,0.0006242701],"domain_scores_gemma":[0.6605853,0.2791418,0.04496417,0.00321374,0.01080037,0.00129467],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00003252652,0.00008084111,0.0002422182,0.01458497,0.001149856,0.00002636928,0.00008927618,1.298426e-7,5.334534e-8,0.00003339064,0.2382173,0.7455431],"study_design_scores_gemma":[0.0005768561,0.005055306,0.00002150658,0.2552713,0.01016753,0.0001410588,0.0001716991,0.00003500863,1.948167e-7,0.0006396557,0.7277409,0.0001790021],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[5.560547e-7,0.9126226,0.04626661,0.03389523,0.0005961726,0.006544364,0.0000142437,0.000003549662,0.00005666851],"genre_scores_gemma":[7.788315e-7,0.8844382,0.1102573,0.0008549023,0.002091689,0.0011468,0.0000161147,0.00003479118,0.001159407],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.745364,"threshold_uncertainty_score":0.9999316,"prediction_status":"machine_predicted_unvalidated"},"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":["metaresearch"],"domain":"methods","study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4289132654","doi":"10.1111/jebm.12483","title":"Clinical manifestations of COVID‐19: An overview of 102 systematic reviews with evidence mapping","year":2022,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"COVID-19 Clinical Research Studies","field":"Medicine","cited_by":37,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Medicine; Cochrane Library; Meta-analysis; MEDLINE; Systematic review; Coronavirus disease 2019 (COVID-19); Internal medicine; Web of science; Disease; Infectious disease (medical specialty)","retraction":null,"screen_n_in":null,"score":{"opus":0.8614589479472008,"gpt":0.6529562840453669,"spread":0.2085026639018339,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.05561811,0.0008010268,0.01497536,0.001588351,0.0001608514,0.00001520555,0.00142041,0.0003039819,0.00214188],"category_scores_gemma":[0.6480944,0.000392589,0.002350847,0.002540203,0.001411448,0.0004637675,0.0002669457,0.002668417,0.00001009309],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001082275,"about_ca_system_score_gemma":0.01752802,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001076245,"about_ca_topic_score_gemma":0.00001309598,"domain_scores_codex":[0.9728775,0.008794836,0.0125896,0.0007593715,0.004435524,0.0005431523],"domain_scores_gemma":[0.6896216,0.2745279,0.02602543,0.00314933,0.003543643,0.003132129],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","study_design_scores_codex":[0.0004925497,0.0004918118,0.0004206474,0.8712119,0.001019656,0.001298672,0.0002215555,0.00002171122,0.000002855775,0.00002827857,0.001727691,0.1230627],"study_design_scores_gemma":[0.00121813,0.009020588,0.00009419899,0.6124322,0.009729881,0.0003810406,0.0002133976,0.00002929021,1.877599e-7,0.00002424059,0.3666558,0.0002009927],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00001343821,0.9853114,0.002007911,0.00679436,0.0004148094,0.005394042,0.00002009891,0.00001834548,0.00002563037],"genre_scores_gemma":[0.0001109771,0.9935884,0.002501956,0.00237979,0.0008690371,0.0002625099,0.00001672648,0.00007598893,0.0001946066],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.5924762,"threshold_uncertainty_score":0.9998526,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaepi_broad","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":"medium"}],"label_agreement":"split"},{"id":"W3198694028","doi":"10.1111/jebm.12443","title":"Progress and challenges of network meta‐analysis","year":2021,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":35,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Guideline; Management science; Medicine; Engineering; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.9673007878232126,"gpt":0.6345168843397608,"spread":0.3327839034834518,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","insufficient_payload"],"consensus_categories":["metaresearch","metaepi_broad"],"category_scores_codex":[0.2713143,0.0008134356,0.05347907,0.002456108,0.00007841006,0.0002212067,0.002637185,0.0003098416,0.008317138],"category_scores_gemma":[0.07922479,0.0002617976,0.02103387,0.006530784,0.0004234728,0.0003243489,0.0001590056,0.0006636665,0.00002678723],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004927113,"about_ca_system_score_gemma":0.0008917294,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000004255513,"about_ca_topic_score_gemma":0.0000130982,"domain_scores_codex":[0.9288904,0.03005401,0.02847452,0.001055624,0.01116363,0.0003618065],"domain_scores_gemma":[0.9019688,0.04077169,0.0477191,0.003494592,0.005569496,0.0004763516],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.000007811445,0.00005545213,0.0001002352,0.01436507,0.1888975,0.0001356716,0.00008651608,0.00008869106,3.686598e-8,0.0003627341,0.01242783,0.7834725],"study_design_scores_gemma":[0.00008592424,0.0002033297,0.00001909063,0.02080626,0.4347483,0.00008051155,0.0001045304,0.00004599213,7.864684e-8,0.0001549965,0.543622,0.0001290175],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.000001234075,0.989364,0.001845792,0.007621468,0.0004056024,0.0006194967,0.000006690653,0.000001474693,0.000134238],"genre_scores_gemma":[0.00006217131,0.9937922,0.004473493,0.0001478306,0.0009584122,0.00002757447,0.000003807778,0.00002300137,0.0005114764],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.7833434,"threshold_uncertainty_score":0.9999834,"prediction_status":"machine_predicted_unvalidated"},"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":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4293254834","doi":"10.1111/jebm.12485","title":"Cognitive behavior therapy for insomnia in cancer patients: a systematic review and network meta‐analysis","year":2022,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Sleep and related disorders","field":"Psychology","cited_by":34,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"National Key Research and Development Program of China","keywords":"Insomnia; Meta-analysis; Sleep onset latency; Sleep onset; Randomized controlled trial; Medicine; Confidence interval; Internal medicine; Physical therapy; Psychiatry","retraction":null,"screen_n_in":null,"score":{"opus":0.3100531130145915,"gpt":0.4771257613616753,"spread":0.1670726483470839,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004051115,0.0005339799,0.009356597,0.001069943,0.00008672081,0.00001251576,0.0004012866,0.0002696604,0.004198421],"category_scores_gemma":[0.001537688,0.000263715,0.00218581,0.002289815,0.0001409327,0.0001107558,0.0000252118,0.0009713853,0.000002651898],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001869969,"about_ca_system_score_gemma":0.0003452333,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001144013,"about_ca_topic_score_gemma":0.00002849029,"domain_scores_codex":[0.9937127,0.002037499,0.002848737,0.0004085741,0.000628844,0.0003636115],"domain_scores_gemma":[0.988955,0.006130591,0.003985976,0.0003174214,0.0004264942,0.0001845846],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","study_design_scores_codex":[0.0002648899,0.0009543605,0.0002455421,0.5269572,0.3375486,0.0001346458,0.0003997946,0.00001152015,6.201376e-9,0.000009290988,0.004129094,0.1293451],"study_design_scores_gemma":[0.00183862,0.001122583,0.00001738694,0.3116993,0.6455768,0.00001828895,0.0001053673,8.814227e-7,3.435833e-9,0.000004873102,0.03942271,0.0001931708],"study_design_candidate":"meta_analysis","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.000005516148,0.9925113,0.00006778934,0.0007185836,0.0007295745,0.005892475,0.00004509808,0.000007042972,0.00002267548],"genre_scores_gemma":[0.0000270484,0.9925327,0.00002152989,0.002886582,0.0002115029,0.004118309,0.00005005519,0.00005286677,0.00009940008],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.3080282,"threshold_uncertainty_score":0.9999815,"prediction_status":"machine_predicted_unvalidated"},"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":"W4361267876","doi":"10.1111/jebm.12526","title":"Acupuncture for treatment of knee osteoarthritis: A clinical practice guideline","year":2023,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Acupuncture Treatment Research Studies","field":"Medicine","cited_by":34,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"McMaster University","funders":"National Science Fund for Distinguished Young Scholars","keywords":"WOMAC; Acupuncture; Medicine; Physical therapy; Osteoarthritis; Certainty; Evidence-based medicine; Alternative medicine; Guideline","retraction":null,"screen_n_in":null,"score":{"opus":0.3030239936294578,"gpt":0.5266039029249645,"spread":0.2235799092955068,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004256031,0.0004763977,0.002320918,0.0006140404,0.00009029067,0.000009130054,0.0001708386,0.0001806262,0.0001653503],"category_scores_gemma":[0.04483596,0.0001677052,0.0008856084,0.0007774775,0.0004737581,0.0002859501,0.00003433379,0.000252852,0.00001504014],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003240393,"about_ca_system_score_gemma":0.001193876,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002927832,"about_ca_topic_score_gemma":0.00000898826,"domain_scores_codex":[0.9950757,0.0002745421,0.002320945,0.0003184043,0.001551913,0.0004585028],"domain_scores_gemma":[0.9819646,0.01224821,0.001337013,0.0004830267,0.003496819,0.0004703354],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0316327,0.001758782,0.02285057,0.001032236,0.002791604,0.004389369,0.001010138,0.0001179641,0.006804964,0.00006278563,0.4975209,0.430028],"study_design_scores_gemma":[0.04114838,0.1667183,0.01199859,0.009858039,0.007730466,0.001136989,0.002585238,0.0002917338,0.003215843,0.000273788,0.7547688,0.0002737942],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.09233853,0.1177073,0.008788078,0.7692014,0.0033556,0.007299146,0.00007547451,0.000188163,0.001046359],"genre_scores_gemma":[0.7113721,0.08909582,0.1460601,0.02032494,0.02449252,0.0005443579,0.0002664598,0.0002735863,0.007570125],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.7488765,"threshold_uncertainty_score":0.9632098,"prediction_status":"machine_predicted_unvalidated"},"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":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2902907490","doi":"10.1111/jebm.12330","title":"Quality of clinical practice guidelines about red blood cell transfusion","year":2018,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Blood transfusion and management","field":"Medicine","cited_by":30,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"","funders":"Instituto de Salud Carlos III; Canadian Medical Association; National Institute for Health and Care Excellence; Agency for Healthcare Research and Quality; Universitat Autònoma de Barcelona; World Health Organization","keywords":"Medicine; Guideline; Scope (computer science); CLARITY; Intensive care medicine; Computer science; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.58470875204432,"gpt":0.577849232896576,"spread":0.006859519147744075,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.02655882,0.0006987536,0.00741563,0.0008685624,0.00007538254,0.00001382088,0.0006721718,0.0007099748,0.000890315],"category_scores_gemma":[0.03959207,0.0003722601,0.002494461,0.0009374417,0.0005468834,0.0003081015,0.00005597997,0.001779709,0.00001947014],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00008434649,"about_ca_system_score_gemma":0.003169348,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00011764,"about_ca_topic_score_gemma":0.00001079149,"domain_scores_codex":[0.9826748,0.001844619,0.01150903,0.0006215531,0.002945676,0.0004043223],"domain_scores_gemma":[0.9771122,0.005826502,0.008367586,0.001138704,0.006782423,0.0007725345],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.002494689,0.002188504,0.00007753541,0.07011123,0.001462067,0.0006826237,0.00009027894,0.000001011858,0.00009122872,0.00002654609,0.08782565,0.8349487],"study_design_scores_gemma":[0.006915973,0.005916966,0.0001055864,0.2092061,0.02175607,0.0003724149,0.0001190215,0.000003409608,0.000044433,0.00000944073,0.7553114,0.0002391974],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00006763927,0.9670923,0.001679042,0.02503072,0.002557545,0.00153411,0.000009274614,0.00002776851,0.002001574],"genre_scores_gemma":[0.0000373632,0.9705724,0.0180623,0.004897234,0.005517986,0.00001369271,0.0000182916,0.0000774668,0.000803296],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.8347095,"threshold_uncertainty_score":0.9998729,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"evaluation","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W3128250231","doi":"10.1111/jebm.12419","title":"Consistency of recommendations and methodological quality of guidelines for the diagnosis and treatment of COVID‐19","year":2021,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Clinical practice guidelines implementation","field":"Medicine","cited_by":29,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Hamilton Health Sciences; Impact","funders":"","keywords":"Guideline; Interim; Medicine; CLARITY; Consistency (knowledge bases); Coronavirus disease 2019 (COVID-19); MEDLINE; Checklist; Quality (philosophy); Concordance; Data collection; Medical physics; Family medicine; Psychology; Computer science; Pathology; Statistics; Political science; Internal medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.9688802759637457,"gpt":0.738142788820996,"spread":0.2307374871427497,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01676326,0.0002912842,0.004372713,0.0002906581,0.00005901894,0.000005259652,0.0001401477,0.0001935119,0.000281378],"category_scores_gemma":[0.3261049,0.0001304723,0.0007260173,0.0003640563,0.0006493598,0.0001027756,0.00004131565,0.0002529643,7.131941e-8],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001539387,"about_ca_system_score_gemma":0.00259234,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000370635,"about_ca_topic_score_gemma":0.00004979437,"domain_scores_codex":[0.9901848,0.001982481,0.006800352,0.0002649811,0.000627855,0.0001395577],"domain_scores_gemma":[0.8355147,0.1523558,0.007837146,0.0004215725,0.003589181,0.0002816089],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0005400826,0.0003322444,0.001317379,0.0135724,0.0009413887,0.00001109839,0.0001143075,0.000003031281,0.00003105975,0.0001027778,0.007657442,0.9753768],"study_design_scores_gemma":[0.005940962,0.009488078,0.0005350578,0.07912947,0.02230788,0.0003255301,0.002022403,0.00003157444,0.0000579931,0.0002266373,0.8797789,0.0001555482],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.000154535,0.874626,0.007990737,0.1155774,0.000272198,0.00127548,0.00009782749,0.000002555148,0.000003349189],"genre_scores_gemma":[0.00014542,0.9543846,0.04354593,0.001260715,0.0005042893,0.0000691037,0.0000516057,0.00001619742,0.0000221696],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.9752212,"threshold_uncertainty_score":0.6795716,"prediction_status":"machine_predicted_unvalidated"},"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":"W1520569313","doi":"10.1111/jebm.12113","title":"Aortic arch atheroma and the risk of stroke: a meta‐analysis","year":2014,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Aortic Thrombus and Embolism","field":"Medicine","cited_by":23,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Atheroma; Aortic arch; Meta-analysis; Medicine; Stroke (engine); Internal medicine; Epidemiology; Observational study; Cardiology; Aorta","retraction":null,"screen_n_in":null,"score":{"opus":0.2313165884669414,"gpt":0.420063952514773,"spread":0.1887473640478317,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.009105579,0.0005385547,0.01147225,0.001025354,0.00008148011,0.00001360664,0.0004393175,0.0002652499,0.0005366986],"category_scores_gemma":[0.005744356,0.0001919135,0.004500132,0.001044384,0.001201449,0.00008797084,0.00005813256,0.001311446,0.000003992006],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0000762309,"about_ca_system_score_gemma":0.0008997764,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003079776,"about_ca_topic_score_gemma":0.00001360125,"domain_scores_codex":[0.9935699,0.00132838,0.002944324,0.0003208569,0.001504722,0.0003318266],"domain_scores_gemma":[0.9859653,0.007192415,0.004880185,0.000858587,0.0006860141,0.0004175425],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"meta_analysis","study_design_scores_codex":[0.00118964,0.0002195128,0.0007586756,0.03362093,0.4497951,0.000182527,0.0005688673,0.0000343045,0.000006964959,0.0002011998,0.001744989,0.5116773],"study_design_scores_gemma":[0.00262315,0.001280781,0.0002975359,0.02571732,0.8460296,0.0002908678,0.00007378198,0.0001626409,0.000002030515,0.00008681073,0.1232776,0.0001578111],"study_design_candidate":"meta_analysis","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.0002752228,0.9922998,0.002030093,0.004435037,0.0002049308,0.0006539673,0.00001247623,0.000007201294,0.00008130195],"genre_scores_gemma":[0.005130186,0.9922954,0.0006464375,0.0004430783,0.001056311,0.00001916118,0.000004702338,0.00004320266,0.0003615469],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.5115194,"threshold_uncertainty_score":0.7842198,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"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":"W4283066010","doi":"10.1111/jebm.12475","title":"More work is needed on cost‐utility analyses of robotic‐assisted surgery","year":2022,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Prostate Cancer Diagnosis and Treatment","field":"Medicine","cited_by":22,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"National Social Science Fund of China","keywords":"Medicine; Robotic surgery; Surgery; General surgery; Prostatectomy; Internal medicine; Prostate cancer","retraction":null,"screen_n_in":null,"score":{"opus":0.637030306476812,"gpt":0.5034306589085736,"spread":0.1335996475682384,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002814516,0.0007427706,0.007166942,0.001482398,0.0000968364,0.00001496903,0.0003657621,0.000257938,0.003341211],"category_scores_gemma":[0.004088506,0.0003961264,0.002375509,0.002148326,0.0003831946,0.0001304947,0.0000560171,0.001237365,0.000007200717],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001310945,"about_ca_system_score_gemma":0.003726989,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001283391,"about_ca_topic_score_gemma":0.00000144097,"domain_scores_codex":[0.9928271,0.0006364463,0.003354742,0.0005135384,0.002218336,0.0004498181],"domain_scores_gemma":[0.9842956,0.008938399,0.004483669,0.0009623682,0.0006954138,0.0006245612],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.001308892,0.0009796885,0.001943196,0.01868332,0.002239827,0.001486947,0.00009534764,0.0000257935,0.0000029678,7.986228e-7,0.07807767,0.8951555],"study_design_scores_gemma":[0.001746783,0.003021423,0.002127234,0.3083016,0.01690086,0.0002611777,0.0001712507,0.000005931703,0.000107765,0.000004538933,0.6670761,0.0002753999],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.0002273375,0.9865069,0.00004504084,0.01044662,0.001427149,0.001210275,0.00003995693,0.00001687051,0.0000798378],"genre_scores_gemma":[0.001453688,0.9944413,0.0001454154,0.002475061,0.001050246,0.0001109365,0.0001070574,0.0000723116,0.0001439565],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.8948802,"threshold_uncertainty_score":0.9998491,"prediction_status":"machine_predicted_unvalidated"},"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":"low"}],"label_agreement":"split"},{"id":"W4244511973","doi":"10.1111/j.1756-5391.2010.01095.x","title":"Medical versus surgical management for gastro‐oesophageal reflux disease (GORD) in adults","year":2010,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Gastroesophageal reflux and treatments","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Medicine; Heartburn; Regurgitation (circulation); Reflux; Disease; Swallowing; Proton-pump inhibitor; Gastroenterology; Gastric acid; Internal medicine; Gastro-; Stomach; Surgery","retraction":null,"screen_n_in":null,"score":{"opus":0.06964112816479469,"gpt":0.3810043665589661,"spread":0.3113632383941715,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002109364,0.0003418382,0.0008828667,0.000593439,0.00006688487,0.00001279023,0.0003392172,0.0001679217,0.0009747218],"category_scores_gemma":[0.003718979,0.0001998974,0.0003449967,0.0003767792,0.0003116838,0.0002354356,0.0000444689,0.0009039094,0.0000169132],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001953742,"about_ca_system_score_gemma":0.0006787586,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001021799,"about_ca_topic_score_gemma":0.00001362367,"domain_scores_codex":[0.9957098,0.0001040844,0.001209156,0.0003755525,0.002022736,0.0005787081],"domain_scores_gemma":[0.9957693,0.001254995,0.0004555816,0.0004337528,0.000440251,0.001646114],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.1070384,0.003304284,0.6482876,0.002405111,0.0006845295,0.1472816,0.0001949623,0.00003888199,0.0001982745,0.0009159929,0.007285573,0.08236471],"study_design_scores_gemma":[0.153821,0.0195456,0.7955596,0.01973934,0.001958321,0.00110397,0.0001871318,0.001337209,0.0003502903,0.0003461149,0.005763542,0.0002879547],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9738908,0.001598871,0.0004736278,0.02013758,0.002377043,0.0009989005,0.00000906794,0.00002966382,0.0004844778],"genre_scores_gemma":[0.9925215,0.0004241998,0.003670636,0.0007417689,0.002190512,0.00004882101,0.00002501925,0.00003656358,0.000340957],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.147272,"threshold_uncertainty_score":0.9999385,"prediction_status":"machine_predicted_unvalidated"},"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":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2610165645","doi":"10.1111/jebm.12256","title":"Prehospital notification for major trauma patients requiring emergency hospital transport: A systematic review","year":2017,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Trauma and Emergency Care Studies","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Surgical Specialties (Canada); University of Saskatchewan","funders":"Department of Industry and Science, Australian Government; Australia-India Strategic Research Fund; Department of Industry, Innovation and Science, Australian Government; Department of Science and Technology, Ministry of Science and Technology, India; Australian Government","keywords":"Medicine; Observational study; Psychological intervention; Emergency medicine; Emergency medical services; Odds ratio; Confidence interval; Major trauma; Emergency department; Medical emergency; Internal medicine; Nursing","retraction":null,"screen_n_in":null,"score":{"opus":0.3311553400255458,"gpt":0.456491608791901,"spread":0.1253362687663553,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.003071833,0.0009065414,0.008320278,0.0006023482,0.0002726833,0.00001335986,0.0008118389,0.0003360783,0.0002105005],"category_scores_gemma":[0.01195468,0.0004920484,0.002891337,0.0003941375,0.0002192408,0.0004705377,0.00002451635,0.0007717471,0.00001867419],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004298734,"about_ca_system_score_gemma":0.0008204688,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001174516,"about_ca_topic_score_gemma":0.000004712096,"domain_scores_codex":[0.9913237,0.0002308916,0.005683532,0.000582809,0.001634689,0.0005443647],"domain_scores_gemma":[0.9877636,0.0006108403,0.007278502,0.00116219,0.002716301,0.000468543],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","study_design_scores_codex":[0.0000934999,0.0002679699,0.0001656331,0.856653,0.001080929,0.00006334141,0.0001832131,1.023996e-7,4.101385e-7,0.00000652707,0.007115967,0.1343694],"study_design_scores_gemma":[0.001568179,0.004384223,0.0004335606,0.8383707,0.04040365,0.00006610431,0.00004998094,0.000001041404,0.000001094778,0.00001337365,0.1143009,0.0004072604],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00004309124,0.9837732,0.0003246907,0.003228663,0.004355981,0.008013673,0.00005319692,0.00003140908,0.000176066],"genre_scores_gemma":[0.0008444618,0.9952707,0.0003212799,0.0001390817,0.002051716,0.0005755178,0.000133148,0.0001197221,0.0005443992],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.1339621,"threshold_uncertainty_score":0.9997531,"prediction_status":"machine_predicted_unvalidated"},"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":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W1601804926","doi":"10.1111/j.1756-5391.2011.01119.x","title":"International comparative analyses of incidents reporting systems for healthcare risk management","year":2011,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Patient Safety and Medication Errors","field":"Health Professions","cited_by":17,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Government (linguistics); Health care; Confidentiality; Risk management; Business; Incident report; Patient safety; Medicine; Medical emergency; Political science; Computer security; Finance; Law","retraction":null,"screen_n_in":null,"score":{"opus":0.8600866229067199,"gpt":0.6659249706519029,"spread":0.194161652254817,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01000196,0.0003478674,0.003374258,0.001128321,0.0002405092,0.000001083026,0.0006796891,0.0002978665,0.0001573489],"category_scores_gemma":[0.009763259,0.0001958144,0.0006141691,0.0004833278,0.000161592,0.0002226204,0.00007260931,0.001155648,0.00001078457],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0006486793,"about_ca_system_score_gemma":0.001730537,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003416151,"about_ca_topic_score_gemma":0.00001083141,"domain_scores_codex":[0.9880272,0.001654589,0.008534928,0.0002929436,0.001153051,0.0003372925],"domain_scores_gemma":[0.9451914,0.003699248,0.04802119,0.0004153795,0.002356279,0.0003165464],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.001611098,0.0005534626,0.01095961,0.362319,0.006890894,0.0001717317,0.009404319,0.0004442115,0.000002379652,0.001076152,0.04424123,0.5623259],"study_design_scores_gemma":[0.0009143029,0.0006080505,0.000312033,0.3149227,0.002671017,0.00001549486,0.002035483,0.00005059699,4.341334e-7,0.00003583604,0.6783205,0.0001135943],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00002967953,0.9764717,0.01378336,0.0009263595,0.005705391,0.002375492,0.00005984718,0.00001257749,0.0006356504],"genre_scores_gemma":[0.002190068,0.9940241,0.001348565,0.0002048425,0.001492878,0.0002192002,0.00008858542,0.00002780912,0.0004039266],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.6340793,"threshold_uncertainty_score":0.998578,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"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":"W2947494553","doi":"10.1111/jebm.12343","title":"Evaluating people's ability to assess treatment claims: Validating a test in Mandarin from Claim Evaluation Tools database","year":2019,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":17,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Hamilton Health Sciences; Impact","funders":"Ministry of Science and Technology of the People's Republic of China","keywords":"Mandarin Chinese; Test (biology); Database; Computer science; Psychology; Linguistics","retraction":null,"screen_n_in":null,"score":{"opus":0.9102972500589558,"gpt":0.6288915867040393,"spread":0.2814056633549165,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.09162135,0.0003118034,0.001278847,0.001433295,0.0001218843,0.000238859,0.001063808,0.0001123582,0.003134311],"category_scores_gemma":[0.6151812,0.000171703,0.0001939765,0.003633749,0.00007639664,0.001248081,0.0001352852,0.0004733135,0.00005403132],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0009480509,"about_ca_system_score_gemma":0.001313654,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0007511951,"about_ca_topic_score_gemma":0.0002059729,"domain_scores_codex":[0.9865021,0.00348915,0.003077972,0.0008008554,0.005674278,0.000455705],"domain_scores_gemma":[0.7032561,0.2884846,0.003013672,0.00157885,0.003175138,0.000491623],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.000350269,0.0001623804,0.6041794,0.00002304528,0.00001604646,0.00002223937,0.0006405481,0.008866451,0.02892951,0.000006147823,0.000403825,0.3564002],"study_design_scores_gemma":[0.006854614,0.006841865,0.8889723,0.005872721,0.0002454261,0.00003329864,0.004577305,0.07928684,0.002605677,0.003968144,0.0004039529,0.0003378676],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9718843,0.001102313,0.01620343,0.007995904,0.001474499,0.001001543,0.00002161125,0.00001506225,0.0003013368],"genre_scores_gemma":[0.9336103,0.0000689669,0.06496891,0.0006085386,0.0006462821,0.00002784978,0.000009195249,0.00001481683,0.00004512838],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5235599,"threshold_uncertainty_score":0.997777,"prediction_status":"machine_predicted_unvalidated"},"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":"low"}],"label_agreement":"agree"},{"id":"W4200188781","doi":"10.1111/jebm.12461","title":"Aerosol generated by dental procedures: A scoping review","year":2021,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Dental Research and COVID-19","field":"Dentistry","cited_by":17,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"Coordenação de Aperfeiçoamento de Pessoal de Nível Superior","keywords":"Aerosol; Protocol (science); Coronavirus disease 2019 (COVID-19); Environmental science; Pandemic; Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); Contamination; Environmental health; Meteorology; Geography; Medicine; Infectious disease (medical specialty); Biology; Pathology; Alternative medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.287049416494756,"gpt":0.501716132993358,"spread":0.214666716498602,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.003815096,0.0007588997,0.004884847,0.0005170326,0.0001581659,0.0001434705,0.001340547,0.0003684441,0.005127035],"category_scores_gemma":[0.0459726,0.0004462352,0.001343352,0.001642567,0.000291025,0.0004996584,0.0001664445,0.001690175,0.0002183313],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0008971628,"about_ca_system_score_gemma":0.01614568,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00004588189,"about_ca_topic_score_gemma":0.0001146088,"domain_scores_codex":[0.9907139,0.001346698,0.003357535,0.0006219537,0.003247772,0.0007121217],"domain_scores_gemma":[0.9901386,0.00387794,0.003023325,0.0006972156,0.001103542,0.001159305],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","study_design_scores_codex":[0.0000573519,0.0001238686,0.00000930402,0.3977491,0.0002958775,0.004843001,0.000006332666,5.856044e-7,0.00005565963,0.000003415092,0.449829,0.1470265],"study_design_scores_gemma":[0.000891547,0.0005315172,2.325224e-7,0.749415,0.001753631,0.003513854,0.00002348494,0.000001273359,0.00003226864,0.000003129398,0.2435325,0.0003015531],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.000009575206,0.9955404,0.0004668644,0.0008970342,0.001259014,0.001693683,0.00003600178,0.00002703642,0.00007034339],"genre_scores_gemma":[0.000008561849,0.9922359,0.0001130748,0.004188915,0.001477077,0.000108465,0.000148607,0.00009445583,0.001624941],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.3516659,"threshold_uncertainty_score":0.999799,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"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":"W2021893832","doi":"10.1111/j.1756-5391.2009.01030.x","title":"An Educational Approach to Improving Healthcare Safety and Quality*","year":2009,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Patient Safety and Medication Errors","field":"Health Professions","cited_by":17,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Health care; Quality (philosophy); Patient safety; Business; Nursing; Risk analysis (engineering); Computer science; Medicine; Political science; Philosophy","retraction":null,"screen_n_in":null,"score":{"opus":0.3683131285364435,"gpt":0.5357444901409603,"spread":0.1674313616045168,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005372237,0.00014553,0.0004708786,0.0003419194,0.0004020797,0.000001218232,0.000247637,0.000129093,0.0001281854],"category_scores_gemma":[0.005692789,0.00009198132,0.0000478532,0.0003603458,0.00009370461,0.0004155048,0.00001318822,0.0007099689,0.000008872678],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002736869,"about_ca_system_score_gemma":0.001878782,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001419783,"about_ca_topic_score_gemma":0.00001668949,"domain_scores_codex":[0.9963816,0.0008728595,0.001411244,0.0002157964,0.0007917006,0.0003268336],"domain_scores_gemma":[0.9954048,0.001386009,0.001183824,0.0002866533,0.0008299683,0.0009087931],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.01562422,0.002357505,0.3663942,0.005747842,0.0001508414,0.00002459623,0.08178734,0.003178914,0.02953882,0.06651256,0.02974816,0.398935],"study_design_scores_gemma":[0.00347765,0.004036202,0.9457317,0.00788311,0.00009047108,0.00002268708,0.007980895,0.0004596932,0.00004316157,0.0009068378,0.02912013,0.0002474778],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.08192934,0.003077095,0.07983702,0.8314282,0.001463441,0.0009231718,0.000008382497,0.00003042239,0.001302955],"genre_scores_gemma":[0.9551318,0.0003154428,0.006615336,0.0351861,0.002408955,0.00001004988,0.00001565063,0.000009667544,0.0003070421],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8732024,"threshold_uncertainty_score":0.6815211,"prediction_status":"machine_predicted_unvalidated"},"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":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W1544000882","doi":"10.1111/j.1756-5391.2012.01173.x","title":"Systematic reviews of TCM trials: how does inclusion of Chinese trials affect outcome?","year":2012,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":17,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"","keywords":"Inclusion (mineral); English language; Clinical trial; Chinese language; Systematic review; MEDLINE; Randomized controlled trial; China; Alternative medicine; Medicine; Computer science; Medical education; Psychology; Pathology; Linguistics; Social psychology; Political science","retraction":null,"screen_n_in":null,"score":{"opus":0.9254087096431709,"gpt":0.6418932184013989,"spread":0.283515491241772,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_broad","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.8492686,0.0005973679,0.03749991,0.001833302,0.0001285226,0.0001064039,0.002743674,0.000177722,0.006304201],"category_scores_gemma":[0.933216,0.0001485867,0.008555588,0.002589614,0.0002196275,0.001080363,0.0002553918,0.0003725519,0.00006913499],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000102732,"about_ca_system_score_gemma":0.0002551734,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001280025,"about_ca_topic_score_gemma":0.000006322286,"domain_scores_codex":[0.6303498,0.2356302,0.1092138,0.0008076825,0.02332909,0.0006694524],"domain_scores_gemma":[0.3855853,0.3831993,0.2162828,0.006056707,0.007872635,0.001003216],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"systematic_review","study_design_scores_codex":[0.001839961,0.001440115,0.5241669,0.1945812,0.00528659,0.00005848865,0.01102756,0.0001656393,0.06524648,0.000490934,0.1457866,0.04990954],"study_design_scores_gemma":[0.022963,0.009835026,0.05802958,0.7365054,0.06845199,0.0005901997,0.009401921,0.004052696,0.009448092,0.01138768,0.06703042,0.002303994],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"empirical","genre_scores_codex":[0.3856412,0.4809594,0.07155544,0.04030249,0.009081956,0.01201463,0.00002614067,0.000008131528,0.0004105653],"genre_scores_gemma":[0.9896681,0.003777136,0.003559945,0.0004214215,0.001690875,0.00003660743,0.00000119571,0.00001787583,0.0008268214],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.6040269,"threshold_uncertainty_score":0.9946042,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"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":"medium"}],"label_agreement":"split"},{"id":"W2897220320","doi":"10.1111/jebm.12339","title":"Flexible piecewise linear model for investigating dose‐response relationship in meta‐analysis: Methodology, examples, and comparison","year":2019,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Radiation Dose and Imaging","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; St. Joseph’s Healthcare Hamilton; Impact","funders":"","keywords":"Piecewise linear function; Piecewise; Linear model; Simple (philosophy); Mathematics; Segmented regression; Applied mathematics; Function (biology); Spline (mechanical); Mathematical optimization; Algorithm; Computer science; Statistics; Linear regression; Mathematical analysis; Polynomial regression","retraction":null,"screen_n_in":null,"score":{"opus":0.6721250343346913,"gpt":0.4885286508437766,"spread":0.1835963834909147,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01129465,0.0002043572,0.001636124,0.001419093,0.00006233425,0.00001607634,0.000135276,0.0001131785,0.00009243782],"category_scores_gemma":[0.01490482,0.000129705,0.0003549957,0.0009793262,0.0001618239,0.0004595795,0.00001770349,0.000534813,0.000002529176],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001108157,"about_ca_system_score_gemma":0.0006473627,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00006374613,"about_ca_topic_score_gemma":0.00001486176,"domain_scores_codex":[0.9969239,0.0007298202,0.001322289,0.0002588508,0.0005115012,0.0002536011],"domain_scores_gemma":[0.987394,0.01070795,0.0009180376,0.0002929699,0.0003850819,0.0003019774],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.007839732,0.0002068606,0.6634203,0.0007044991,0.01229935,0.00003697319,0.00398067,0.2790946,0.02769434,0.0003772384,0.00176152,0.002583882],"study_design_scores_gemma":[0.006474178,0.001557766,0.1015254,0.001447813,0.04969031,0.00005833976,0.0008793,0.8346071,0.002157235,0.001010143,0.0003998845,0.0001926219],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.6920411,0.03334597,0.2211347,0.0524985,0.0001491692,0.0007684922,0.00000425586,0.00002455182,0.00003318825],"genre_scores_gemma":[0.8981242,0.0001884293,0.09817906,0.002911522,0.0001893914,0.00001536273,0.000007558268,0.00001888413,0.0003655854],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.561895,"threshold_uncertainty_score":0.9933931,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"simulation_or_modeling","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"simulation_or_modeling","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W4324307220","doi":"10.1111/jebm.12522","title":"STandards for Reporting Interventions in Clinical Trials Of Tuina/Massage (STRICTOTM): Extending the CONSORT statement","year":2023,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Complementary and Alternative Medicine Studies","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Massage; Checklist; Consolidated Standards of Reporting Trials; Psychological intervention; Medicine; Physical therapy; Chiropractic; Clinical trial; Guideline; Alternative medicine; Intervention (counseling); Nursing; Psychology","retraction":null,"screen_n_in":null,"score":{"opus":0.8701649821119687,"gpt":0.6756244896524661,"spread":0.1945404924595026,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.1058027,0.0001867203,0.001959073,0.0006971246,0.0001032333,0.000007097475,0.0002000264,0.0000249628,0.0003195237],"category_scores_gemma":[0.1903829,0.00008841097,0.0008482827,0.0007330365,0.0003930544,0.0001613051,0.00006215653,0.000477487,6.007493e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002192297,"about_ca_system_score_gemma":0.0007398161,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00004346398,"about_ca_topic_score_gemma":0.0000263768,"domain_scores_codex":[0.9870164,0.000952011,0.01016235,0.0002202197,0.001340848,0.0003081693],"domain_scores_gemma":[0.9700735,0.01860011,0.009399343,0.0003030834,0.001438342,0.0001856184],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.006945222,0.0007257848,0.702274,0.003615496,0.0017864,0.00150005,0.001659334,0.0002249946,0.004908761,0.0005022513,0.07317646,0.2026813],"study_design_scores_gemma":[0.04416841,0.03961199,0.6488619,0.1492054,0.004932395,0.0002810457,0.02701158,0.001678213,0.003746588,0.002013633,0.07810141,0.0003874299],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8841287,0.01264972,0.0256242,0.07257911,0.002687684,0.002099489,0.0000730363,0.00002538017,0.0001326635],"genre_scores_gemma":[0.9915122,0.005442155,0.0007630856,0.0003845193,0.001597515,0.0000378274,0.00001399972,0.00001618389,0.000232508],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2022938,"threshold_uncertainty_score":0.9207643,"prediction_status":"machine_predicted_unvalidated"},"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":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4378194782","doi":"10.1111/jebm.12534","title":"Introducing value‐based healthcare perspectives into hospital performance assessment: A scoping review","year":2023,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Primary Care and Health Outcomes","field":"Health Professions","cited_by":16,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Health care; Protocol (science); Reliability (semiconductor); Process management; Process (computing); Computer science; Knowledge management; Medicine; Management science; Engineering; Alternative medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.3583688597916175,"gpt":0.5924324040321615,"spread":0.2340635442405439,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.01476888,0.000937605,0.00745658,0.001541721,0.0010519,0.00002016933,0.001195283,0.0006219888,0.0007401888],"category_scores_gemma":[0.01178003,0.0005553975,0.001087051,0.002060508,0.0003062681,0.0006893663,0.0001969076,0.005565892,0.0001422685],"about_ca_system_candidate":true,"about_ca_system_consensus":true,"about_ca_system_score_codex":0.00633465,"about_ca_system_score_gemma":0.05814563,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003124581,"about_ca_topic_score_gemma":0.00004353775,"domain_scores_codex":[0.9853883,0.004032043,0.006307272,0.0008344621,0.002223371,0.001214518],"domain_scores_gemma":[0.9764819,0.01113824,0.00795655,0.00114895,0.002436315,0.0008380334],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","study_design_scores_codex":[0.00004106071,0.00005157745,0.0003030574,0.7311671,0.0001439424,0.0001217475,0.0006778196,0.000005124614,1.48962e-7,0.00005338641,0.03675129,0.2306838],"study_design_scores_gemma":[0.0008426306,0.001753866,0.00008058601,0.7552313,0.001629315,0.00001854681,0.000620702,0.00001843876,8.180096e-8,0.00002184378,0.239413,0.0003696894],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.000006187142,0.8605086,0.0003466515,0.1304226,0.004176557,0.004299847,0.00001053341,0.0001017745,0.0001273464],"genre_scores_gemma":[0.0000134997,0.9678656,0.002386338,0.02170988,0.007230479,0.0003633654,0.00004744199,0.0001667091,0.0002167281],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.2303141,"threshold_uncertainty_score":0.9996898,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2902864468","doi":"10.1111/jebm.12332","title":"Traffic accidents, maxillofacial injuries and risk factors: A systematic review of observational studies","year":2018,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Facial Trauma and Fracture Management","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Observational study; Medicine; Web of science; Cochrane Library; Poison control; Demography; Injury prevention; Meta-analysis; Environmental health; Internal medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.4001462395798281,"gpt":0.4597323569938587,"spread":0.05958611741403058,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00480774,0.0006578878,0.008358857,0.0007703543,0.00009376825,0.000016968,0.0003773962,0.0002283867,0.000142259],"category_scores_gemma":[0.02656948,0.0003070684,0.0009552218,0.0006814549,0.0006304494,0.0002769701,0.0000712136,0.0007230208,0.000005971046],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002597031,"about_ca_system_score_gemma":0.0007748678,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001183338,"about_ca_topic_score_gemma":0.000002669103,"domain_scores_codex":[0.9929094,0.0005853227,0.004120125,0.000336004,0.001782054,0.0002671328],"domain_scores_gemma":[0.9892764,0.002105261,0.00615113,0.0005028855,0.001685798,0.0002785127],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","study_design_scores_codex":[0.0000921819,0.00008465876,0.0002498482,0.95522,0.002760784,0.00006990052,0.0002403661,5.378633e-7,1.939273e-7,0.00001109679,0.003831736,0.03743863],"study_design_scores_gemma":[0.0006130859,0.00200836,0.0002419542,0.8269379,0.04177759,0.0001051211,0.0002339983,0.000001679631,0.000001093077,0.00001436457,0.1278732,0.0001916563],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.000261961,0.9949056,0.00005560169,0.001601292,0.0008127943,0.002311049,0.00002157456,0.0000143518,0.00001581469],"genre_scores_gemma":[0.00009920229,0.9977387,0.0003124266,0.0007602896,0.0008640888,0.00003853605,0.00002688073,0.00004147154,0.0001183741],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.1282822,"threshold_uncertainty_score":0.9999381,"prediction_status":"machine_predicted_unvalidated"},"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_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4393097347","doi":"10.1111/jebm.12593","title":"Polypharmacy and risk of fractures in older adults: A systematic review","year":2024,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Pharmaceutical Practices and Patient Outcomes","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Université de Montréal; McGill University Health Centre; Université Laval; Centre hospitalier de l'Université Laval; Université du Québec à Rimouski","funders":"Fonds de Recherche du Québec - Santé; Ministère de l'Éducation et de l'Enseignement supérieur","keywords":"Polypharmacy; Medicine; CINAHL; Cochrane Library; Confounding; MEDLINE; Gerontology; Meta-analysis; Intensive care medicine; Internal medicine; Psychiatry; Psychological intervention","retraction":null,"screen_n_in":null,"score":{"opus":0.2701375324967158,"gpt":0.5235085103135608,"spread":0.253370977816845,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005121082,0.0005335605,0.006932918,0.0009560863,0.00002119555,0.00001241627,0.0002793118,0.0001967087,0.000334975],"category_scores_gemma":[0.01665256,0.0002295497,0.000840874,0.0008478569,0.0001791078,0.0002642343,0.00004923222,0.002150884,0.00001208106],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001725937,"about_ca_system_score_gemma":0.0008290332,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00004838152,"about_ca_topic_score_gemma":0.00000222373,"domain_scores_codex":[0.9932704,0.00106681,0.003895476,0.000301545,0.001200008,0.0002657849],"domain_scores_gemma":[0.9884247,0.006081967,0.004291372,0.000398611,0.0003743998,0.0004289178],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","study_design_scores_codex":[0.0002461703,0.0001279534,0.0002627854,0.9224305,0.00078151,0.0005943441,0.00009957359,1.702577e-7,4.168144e-7,0.000002259932,0.002092503,0.07336184],"study_design_scores_gemma":[0.001719754,0.0008183377,0.00004295996,0.8690354,0.06843274,0.0008790703,0.00004790552,0.000038355,0.000001383378,0.000006750188,0.05881701,0.0001603535],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00006503492,0.9922522,0.00001724559,0.004068589,0.0005479314,0.002989468,0.00001272418,0.000009158664,0.00003760054],"genre_scores_gemma":[0.0005069182,0.9965045,0.0001067528,0.002235838,0.0004842944,0.00003746007,0.000006181659,0.00004754349,0.00007054597],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.07320149,"threshold_uncertainty_score":0.9916306,"prediction_status":"machine_predicted_unvalidated"},"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":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1603245870","doi":"10.1111/j.1756-5391.2011.01158.x","title":"Effects and patient compliance of sustained‐release versus immediate‐release glipizides in patients with type 2 diabetes mellitus: a systematic review and meta‐analysis","year":2011,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Diabetes Treatment and Management","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University","funders":"National Natural Science Foundation of China; Pfizer","keywords":"Glipizide; Medicine; Postprandial; Internal medicine; Hypoglycemia; Odds ratio; Insulin; Diabetes mellitus; Type 2 diabetes; Meta-analysis; Cochrane Library; Randomized controlled trial; Type 2 Diabetes Mellitus; Endocrinology","retraction":null,"screen_n_in":null,"score":{"opus":0.08992171394064337,"gpt":0.3284183681532469,"spread":0.2384966542126036,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.001286995,0.0007095203,0.0105353,0.001243512,0.00003165996,0.00001333731,0.0002105804,0.0001327815,0.00007386451],"category_scores_gemma":[0.002460998,0.0003328584,0.0007552041,0.001468079,0.0002859835,0.0002193238,0.00006523581,0.0003876368,0.000002447438],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002018268,"about_ca_system_score_gemma":0.0003037367,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002424417,"about_ca_topic_score_gemma":0.000004527459,"domain_scores_codex":[0.9955909,0.0005481972,0.002194653,0.0004215094,0.0008614216,0.0003832901],"domain_scores_gemma":[0.9926136,0.002379093,0.003118203,0.0005236668,0.0009059098,0.0004595081],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","study_design_scores_codex":[0.0006099516,0.0004715877,0.0005470939,0.9101883,0.0715746,0.0001351188,0.00006044058,0.00000105647,1.13815e-7,0.000008866909,0.0001326945,0.01627022],"study_design_scores_gemma":[0.002921653,0.006308712,0.0002184109,0.4048035,0.5845144,0.000002440484,0.00003230588,0.000004109467,0.000001737236,0.00000267275,0.001000733,0.0001893353],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.000950497,0.9944469,0.0000079912,0.0003312846,0.00009844838,0.00413415,0.000009991644,0.000008370157,0.00001238416],"genre_scores_gemma":[0.004894143,0.9940386,0.0002947113,0.0004965399,0.0000381286,0.0001363022,0.00003350129,0.00004437529,0.00002368564],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.5129398,"threshold_uncertainty_score":0.9999123,"prediction_status":"machine_predicted_unvalidated"},"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":"W3034590128","doi":"10.1111/jebm.12393","title":"The association between allergic rhinitis and risk of rheumatoid arthritis: A systematic review and meta‐analysis","year":2020,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Allergic Rhinitis and Sensitization","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Medicine; Meta-analysis; Odds ratio; Rheumatoid arthritis; Confidence interval; Relative risk; Internal medicine; Cohort study; Case-control study; MEDLINE","retraction":null,"screen_n_in":null,"score":{"opus":0.1133902705534129,"gpt":0.349170584944384,"spread":0.2357803143909711,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.007244861,0.0004719831,0.01273052,0.0004496784,0.0001442311,0.00003710325,0.0001997717,0.0002536593,0.00006958865],"category_scores_gemma":[0.01597404,0.000210473,0.001686708,0.00108318,0.0001907682,0.0001606685,0.00003995255,0.00084538,0.000002388363],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001639219,"about_ca_system_score_gemma":0.0005568623,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00004173272,"about_ca_topic_score_gemma":0.0000174041,"domain_scores_codex":[0.992312,0.001959654,0.003817411,0.0002961087,0.001389608,0.0002252447],"domain_scores_gemma":[0.9808421,0.00844021,0.008841483,0.0003852739,0.00113231,0.0003586712],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","study_design_scores_codex":[0.00003430533,0.00002567289,0.0001921202,0.6845174,0.2881203,0.0001881191,0.0001609921,0.000001186449,1.221669e-7,0.00002180362,0.001386918,0.02535102],"study_design_scores_gemma":[0.0004865525,0.0005157874,0.00004480779,0.3881654,0.5803357,0.000323682,0.00005012473,0.00001481343,1.634118e-7,0.00001759933,0.02992408,0.0001212378],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.000007262512,0.988753,0.0003167989,0.009242392,0.00006870511,0.001541398,0.00003633952,0.000009769053,0.00002438578],"genre_scores_gemma":[0.000277048,0.9986362,0.0001632789,0.0005374883,0.0002275226,0.0000323804,0.00002888121,0.00003480481,0.00006240775],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.296352,"threshold_uncertainty_score":0.9923148,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"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":"W4283766318","doi":"10.1111/jebm.12476","title":"Malignancy risk stratification for solitary pulmonary nodule: A clinical practice guideline","year":2022,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Lung Cancer Diagnosis and Treatment","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University","funders":"","keywords":"Malignancy; Medicine; Risk stratification; Nodule (geology); Guideline; Radiology; Family history; Clinical Practice; Internal medicine; Pathology; Family medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.1658120201995652,"gpt":0.4598695820264167,"spread":0.2940575618268515,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.007647634,0.0002138577,0.0007816017,0.0002489618,0.0002299013,0.00001039071,0.0002176473,0.00007169138,0.0006339847],"category_scores_gemma":[0.009915962,0.0001333288,0.0003847055,0.0003535343,0.0001286638,0.0003388511,0.00003458578,0.0006484851,0.000004191987],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0006222349,"about_ca_system_score_gemma":0.001482737,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001267301,"about_ca_topic_score_gemma":0.000004173188,"domain_scores_codex":[0.9954594,0.0004870353,0.00221285,0.0003378085,0.001235579,0.0002673464],"domain_scores_gemma":[0.9916292,0.003975516,0.002039774,0.0004748063,0.00152513,0.0003555164],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.01938955,0.006040151,0.08874732,0.0006525872,0.001355184,0.002616756,0.0003906287,0.001447694,0.001324464,0.000428606,0.6544957,0.2231113],"study_design_scores_gemma":[0.02290509,0.05022039,0.05966842,0.007535517,0.01099303,0.00357782,0.003405381,0.006847684,0.0008033344,0.001588972,0.8319703,0.0004841093],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.05398038,0.1024108,0.02517471,0.8080401,0.005517277,0.003449727,0.0001279504,0.00007063317,0.001228453],"genre_scores_gemma":[0.9166702,0.01157039,0.04448341,0.02115976,0.005377551,0.0002428873,0.0001019498,0.00005554549,0.0003383401],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8626898,"threshold_uncertainty_score":0.9984239,"prediction_status":"machine_predicted_unvalidated"},"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":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W4296785985","doi":"10.1111/jebm.12493","title":"Endovascular thrombectomy and intravenous alteplase in patients with acute ischemic stroke due to large vessel occlusion: A clinical practice guideline","year":2022,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Acute Ischemic Stroke Management","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Dalhousie University; Vancouver General Hospital; University of British Columbia; Impact; University of Calgary; Hamilton General Hospital; Heart and Stroke Foundation; McMaster University","funders":"","keywords":"Medicine; Guideline; Stroke (engine); Randomized controlled trial; Intensive care medicine; Occlusion; Emergency medicine; Internal medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.03341171726631387,"gpt":0.3447126050320186,"spread":0.3113008877657047,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.006082252,0.0003333013,0.001262685,0.0006304877,0.0001317085,0.00001449988,0.0003476884,0.00008360417,0.000461541],"category_scores_gemma":[0.005915944,0.0002141999,0.0001560156,0.000737032,0.0001484766,0.0003596501,0.0003550283,0.001342625,0.000004344302],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004480999,"about_ca_system_score_gemma":0.0007178653,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00005535864,"about_ca_topic_score_gemma":0.00000535857,"domain_scores_codex":[0.994423,0.0003753658,0.001939511,0.0004786803,0.002300263,0.0004831393],"domain_scores_gemma":[0.995603,0.001209472,0.001092014,0.0005693753,0.000932842,0.0005932766],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.04098236,0.004901446,0.2186917,0.0003773626,0.002689354,0.03549157,0.00139799,0.0004501067,0.004239686,0.00001886281,0.6716197,0.01913988],"study_design_scores_gemma":[0.06355131,0.03757317,0.1242373,0.004815313,0.005100757,0.005196927,0.002911408,0.0007487343,0.0007880823,0.000005244634,0.7544932,0.0005784816],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9270618,0.002853578,0.007849785,0.05993273,0.0005506633,0.001261806,0.00002476922,0.00002272839,0.0004421109],"genre_scores_gemma":[0.9244877,0.0007632265,0.0317786,0.04163779,0.001001397,0.00003503966,0.00004299672,0.00006649572,0.0001867634],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.09445439,"threshold_uncertainty_score":0.8734818,"prediction_status":"machine_predicted_unvalidated"},"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":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1483618109","doi":"10.1111/jebm.12022","title":"Visualization of evidence‐based medicine domain knowledge: production and citation of cochrane systematic reviews","year":2013,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":12,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"China Medical Board","keywords":"Citation; Knowledge translation; Systematic review; Cochrane Library; Psychological intervention; Production (economics); MEDLINE; Bibliometrics; Medicine; Developing country; Business; Political science; Alternative medicine; Computer science; Library science; Knowledge management; Economic growth; Nursing; Pathology; Economics","retraction":null,"screen_n_in":null,"score":{"opus":0.842957060811485,"gpt":0.5825738731360253,"spread":0.2603831876754598,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.2452208,0.0004218809,0.00805038,0.001980154,0.00009295098,0.00008040948,0.001031341,0.0001266595,0.00349636],"category_scores_gemma":[0.5259732,0.0001506298,0.0008543911,0.003520278,0.0005428669,0.001383046,0.00003504353,0.0002231819,0.0000799034],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001074355,"about_ca_system_score_gemma":0.0003467656,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00004119201,"about_ca_topic_score_gemma":0.00001418899,"domain_scores_codex":[0.9485171,0.01981054,0.02286714,0.0006303273,0.0079351,0.0002397445],"domain_scores_gemma":[0.8952835,0.03883656,0.04431519,0.002688359,0.01841771,0.0004587246],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","study_design_scores_codex":[0.0007419657,0.0009508242,0.05608591,0.3312029,0.001278475,0.00002757758,0.01403339,0.001601827,0.2154714,0.002231777,0.3322066,0.04416734],"study_design_scores_gemma":[0.002878992,0.005650818,0.02099654,0.9264463,0.005308474,0.0001783512,0.006000231,0.01910714,0.003872266,0.005995139,0.003018591,0.0005471768],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"empirical","genre_scores_codex":[0.1631683,0.4866234,0.2692674,0.06798482,0.002867323,0.009745757,0.000002594582,0.00001146055,0.0003289982],"genre_scores_gemma":[0.984488,0.008023638,0.005343535,0.0004546267,0.0007536495,0.00007988853,0.000002965089,0.00002066397,0.0008330784],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8213197,"threshold_uncertainty_score":0.9974146,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","bibliometrics"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["bibliometrics","scholarly_communication"],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4386863815","doi":"10.1111/jebm.12550","title":"Methodological quality, risk of bias, and reporting quality: A confusion persists","year":2023,"lang":"en","type":"letter","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":12,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Confusion; Quality (philosophy); Risk analysis (engineering); Psychology; Business; Epistemology; Philosophy","retraction":null,"screen_n_in":null,"score":{"opus":0.9914672909457757,"gpt":0.7089936643344512,"spread":0.2824736266113245,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_broad","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.763588,0.0006473223,0.0175817,0.001892828,0.0001534059,0.0002542044,0.002131007,0.0007311305,0.003872971],"category_scores_gemma":[0.9174014,0.0002329406,0.004755065,0.002048478,0.0006132431,0.0002833962,0.0002356016,0.002607133,0.00006591737],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00009564062,"about_ca_system_score_gemma":0.0006205698,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002426899,"about_ca_topic_score_gemma":0.00002042229,"domain_scores_codex":[0.7402718,0.1506106,0.08131086,0.001628228,0.02559857,0.0005798982],"domain_scores_gemma":[0.3733329,0.3340454,0.2770647,0.004545217,0.01051836,0.0004933693],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00009261722,0.00002325681,0.05641784,0.001365983,0.0006840665,0.0006589346,0.0004653234,0.00003891691,0.0003606619,0.00001837639,0.9242777,0.01559638],"study_design_scores_gemma":[0.002145964,0.002766012,0.1317421,0.03456401,0.01001611,0.0008469386,0.004849093,0.0009592641,0.000122268,0.01320897,0.7977671,0.001012131],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.06023142,0.01710489,0.02553499,0.8941204,0.00190613,0.0007693921,0.00003034282,0.00000900796,0.000293425],"genre_scores_gemma":[0.1089832,0.02043672,0.0828644,0.7254857,0.02878032,0.00005204776,0.00007331696,0.0002353987,0.03308889],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.3669389,"threshold_uncertainty_score":0.9996939,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2125660710","doi":"10.1111/jebm.12137","title":"Health equity in humanitarian emergencies: a role for evidence aid","year":2015,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Food Security and Health in Diverse Populations","field":"Health Professions","cited_by":8,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Cochrane; Bruyère","funders":"Canadian Institutes of Health Research","keywords":"Health equity; Equity (law); Health care; Disadvantaged; Psychological intervention; Business; Political science; Public relations; Economic growth; Public economics; Medicine; Economics; Nursing","retraction":null,"screen_n_in":null,"score":{"opus":0.7822702770008315,"gpt":0.6122804426939947,"spread":0.1699898343068368,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01541922,0.0001993098,0.0008669317,0.0005969064,0.0005561877,0.000007383803,0.0004987631,0.0001855401,0.0004016661],"category_scores_gemma":[0.0206132,0.0001397517,0.0001238245,0.0005939265,0.0001456359,0.0007799367,0.00009052247,0.001085644,0.00003619674],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001542827,"about_ca_system_score_gemma":0.007489957,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003562264,"about_ca_topic_score_gemma":0.005986413,"domain_scores_codex":[0.9946711,0.0009762921,0.002337757,0.0002348272,0.001033228,0.000746791],"domain_scores_gemma":[0.9922979,0.003216763,0.001732846,0.0003372932,0.001554035,0.0008610914],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00707474,0.0008007665,0.2531239,0.007815451,0.00009971912,0.0000868419,0.1023714,0.002052012,0.0003208258,0.01379254,0.5966194,0.01584232],"study_design_scores_gemma":[0.02737833,0.03760011,0.1268965,0.1944447,0.0004334352,0.00005227774,0.117001,0.007854672,0.00004896615,0.06353323,0.4236591,0.001097622],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.1497306,0.08796787,0.01060158,0.734921,0.01029246,0.004847204,0.00004601749,0.00009072354,0.001502555],"genre_scores_gemma":[0.977246,0.002121886,0.00390449,0.01355295,0.002786703,0.00008035741,0.000007498391,0.00002576415,0.0002742699],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8275155,"threshold_uncertainty_score":0.9981366,"prediction_status":"machine_predicted_unvalidated"},"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":"W4408981663","doi":"10.1111/jebm.70019","title":"Synthesizing Public Health Preparedness Mechanisms for High‐Impact Infectious Disease Threats: A Jurisdictional Scan","year":2025,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Viral Infections and Outbreaks Research","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"Public Health Agency","keywords":"Preparedness; Public health; International Health Regulations; Health care; Business; Emergency management; Global health; Infectious disease (medical specialty); Workforce; Pandemic; Public relations; Medicine; Medical emergency; Political science; Disease; Coronavirus disease 2019 (COVID-19); Nursing","retraction":null,"screen_n_in":null,"score":{"opus":0.3352340853519318,"gpt":0.5054519080436269,"spread":0.1702178226916951,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.005829643,0.0007457731,0.004367379,0.003095082,0.0004116395,0.00009972617,0.0004079223,0.0002872231,0.0004817571],"category_scores_gemma":[0.008398798,0.0004095205,0.001860298,0.001770053,0.0002770195,0.0004763897,0.00007383413,0.001291217,0.000007692736],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.003741561,"about_ca_system_score_gemma":0.03174959,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002100981,"about_ca_topic_score_gemma":0.00002051081,"domain_scores_codex":[0.9936343,0.0007533493,0.002482496,0.0005931482,0.001670371,0.0008662852],"domain_scores_gemma":[0.9890626,0.004235223,0.002181084,0.0006555718,0.002033971,0.001831551],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"systematic_review","study_design_scores_codex":[0.0006494818,0.000567561,0.0000712579,0.05456408,0.001236933,0.0001810281,0.00006004652,0.00002106239,0.000004538817,0.0006164978,0.01090441,0.9311231],"study_design_scores_gemma":[0.004434856,0.01086986,0.0001891921,0.5603871,0.006955971,0.001272709,0.0001035846,0.0001547974,0.000004815567,0.002154182,0.4129038,0.0005691005],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00002659032,0.944921,0.03867387,0.01201494,0.001425168,0.002666459,0.0001284759,0.00006884279,0.00007464699],"genre_scores_gemma":[0.002806574,0.9915301,0.0007180597,0.0009524503,0.00278069,0.000314535,0.0001345471,0.00009141716,0.0006716418],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.930554,"threshold_uncertainty_score":0.9999539,"prediction_status":"machine_predicted_unvalidated"},"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":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W1580273645","doi":"10.1111/jebm.12169","title":"Evidence‐informed health policy making in Canada: past, present, and future","year":2015,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Primary Care and Health Outcomes","field":"Health Professions","cited_by":7,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"London Health Sciences Centre; Western University","funders":"Western University of Health Sciences","keywords":"Psychological intervention; Political science; Evidence-based policy; Health policy; Policy making; Field (mathematics); Public relations; Public administration; Health care; Medicine; Alternative medicine; Nursing; Law","retraction":null,"screen_n_in":null,"score":{"opus":0.3298856697761844,"gpt":0.5201995464452814,"spread":0.190313876669097,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005621934,0.0002350623,0.001021276,0.0006759882,0.0002338618,0.000007160907,0.0003467391,0.0001346909,0.0001696899],"category_scores_gemma":[0.003486924,0.0001417063,0.00005598406,0.0007093539,0.00009382828,0.0006972712,0.00008263863,0.001293601,0.000004476995],"about_ca_system_candidate":true,"about_ca_system_consensus":true,"about_ca_system_score_codex":0.009044662,"about_ca_system_score_gemma":0.1334587,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.6435193,"about_ca_topic_score_gemma":0.5906332,"domain_scores_codex":[0.9947683,0.0009029408,0.002027759,0.000196931,0.001277924,0.0008261188],"domain_scores_gemma":[0.9923279,0.003907398,0.001660134,0.000306519,0.0007659586,0.001032087],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0007759854,0.00001682894,0.2875654,0.001396401,0.00001903833,0.0001609852,0.003550831,0.00002664571,0.000006091592,0.0001291157,0.6571943,0.04915842],"study_design_scores_gemma":[0.004382668,0.001428595,0.3343969,0.02321069,0.00004341017,0.00007877118,0.01237954,0.00007417398,0.000002620944,0.0003224204,0.6234801,0.0002000334],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.01639821,0.0538028,0.0001016207,0.9258862,0.002239015,0.0006108112,0.000004268966,0.00001279467,0.0009443165],"genre_scores_gemma":[0.6210001,0.02718788,0.001285068,0.3132677,0.03665413,0.00003451965,0.000004424988,0.00004707239,0.0005191675],"genre_candidate":"commentary","genre_consensus":null,"teacher_disagreement_score":0.6126185,"threshold_uncertainty_score":0.9947594,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["scholarly_communication"],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":true,"about_ca_topic":true,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":true,"about_ca_topic":true,"confidence":"medium"}],"label_agreement":"split"},{"id":"W1505158704","doi":"10.1111/j.1756-5391.2011.01118.x","title":"International comparative analyses of healthcare risk management","year":2011,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Patient Safety and Medication Errors","field":"Health Professions","cited_by":6,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Government (linguistics); Health care; Risk management; China; Public administration; Business; Administration (probate law); Descriptive statistics; Political science; Medicine; Public relations; Law; Finance","retraction":null,"screen_n_in":null,"score":{"opus":0.8040728549478809,"gpt":0.6454915615042782,"spread":0.1585812934436027,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00393833,0.0003533026,0.00280104,0.001260085,0.0001750137,4.730312e-7,0.0008284249,0.0002934652,0.001539207],"category_scores_gemma":[0.001924735,0.0001895019,0.000545722,0.00064387,0.000282701,0.000204811,0.00007976305,0.001704516,0.00006043726],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000469448,"about_ca_system_score_gemma":0.001669399,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001995329,"about_ca_topic_score_gemma":0.00001561475,"domain_scores_codex":[0.9923205,0.002245185,0.003741047,0.000251325,0.001134645,0.0003072378],"domain_scores_gemma":[0.9827917,0.002955895,0.01201733,0.0004397855,0.001426879,0.0003684009],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0007637594,0.0004017035,0.001888384,0.07821131,0.003215978,0.0000900579,0.004772067,0.00004860903,0.000001056194,0.001209171,0.03459735,0.8748006],"study_design_scores_gemma":[0.0008275001,0.0005242347,0.0003229859,0.2350206,0.002463945,0.000006832372,0.001010436,0.000006652948,7.368748e-7,0.00006762303,0.7596399,0.0001085344],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.000005735646,0.9848984,0.003924511,0.00363324,0.00342159,0.000980178,0.00004519734,0.0000115546,0.003079559],"genre_scores_gemma":[0.0007269705,0.995419,0.001529502,0.0005299909,0.001164814,0.00005360938,0.00005300035,0.00002199922,0.0005011249],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.874692,"threshold_uncertainty_score":0.9993735,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","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":"W4393934783","doi":"10.1111/jebm.12603","title":"Fluids in the treatment of diabetic ketoacidosis in children: A systematic review","year":2024,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Diabetes and associated disorders","field":"Biochemistry, Genetics and Molecular Biology","cited_by":6,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Diabetic ketoacidosis; Meta-analysis; Population; Randomized controlled trial; Intensive care medicine; Pediatrics; Surgery; Internal medicine; Insulin","retraction":null,"screen_n_in":null,"score":{"opus":0.05882681303615756,"gpt":0.3550558840351998,"spread":0.2962290709990422,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002328774,0.0003691749,0.002920227,0.0003373712,0.00001252877,0.00001120564,0.0004782054,0.0001956268,0.00001619496],"category_scores_gemma":[0.001353218,0.0001537554,0.0008138131,0.0006845251,0.00009296938,0.000005599627,0.00001667307,0.0002552622,0.00000365376],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001329729,"about_ca_system_score_gemma":0.0004981328,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0000197927,"about_ca_topic_score_gemma":0.000008292672,"domain_scores_codex":[0.9963798,0.0007673332,0.001937783,0.0002278661,0.000455762,0.0002314371],"domain_scores_gemma":[0.9979107,0.0003395985,0.001151386,0.0004213575,0.0001104863,0.00006643],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","study_design_scores_codex":[0.00001730626,0.0005339286,0.0003988426,0.9337019,0.001012865,0.00006046732,0.00009781485,0.000003588654,0.00003890747,0.000003914611,0.003639008,0.06049146],"study_design_scores_gemma":[0.0005959114,0.002390526,0.00002074906,0.9619826,0.007887154,0.00004085948,0.0000493872,0.000001798062,0.00001267058,0.000007874529,0.02685848,0.0001519853],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.0004035206,0.9969935,0.00000314394,0.001022854,0.0001402654,0.001404547,0.00001071184,0.000001390422,0.00002010243],"genre_scores_gemma":[0.001028508,0.9980485,0.00001216807,0.0004577248,0.0002328765,0.0001147715,0.00003475998,0.00002618116,0.00004452333],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.06033947,"threshold_uncertainty_score":0.6269965,"prediction_status":"machine_predicted_unvalidated"},"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":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W4284698039","doi":"10.1111/jebm.12481","title":"Factors associated with concurrent malignancy risk among patients with incidental solitary pulmonary nodule: A systematic review taskforce for developing rapid recommendations","year":2022,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Lung Cancer Diagnosis and Treatment","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University","funders":"Sichuan Province Science and Technology Support Program","keywords":"Medicine; Malignancy; Lung cancer; Odds ratio; Nodule (geology); Confidence interval; Checklist; Family history; Internal medicine; Radiology","retraction":null,"screen_n_in":null,"score":{"opus":0.2008800434411215,"gpt":0.3844456475802253,"spread":0.1835656041391037,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00251051,0.0009019988,0.006299243,0.0006080246,0.0003002153,0.00002338295,0.000402543,0.0001621652,0.0003743371],"category_scores_gemma":[0.004993395,0.0004068695,0.0008843143,0.0009760133,0.0001798279,0.0003643901,0.00005683824,0.0009494341,0.000001338011],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.004315818,"about_ca_system_score_gemma":0.002731561,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00004576395,"about_ca_topic_score_gemma":0.0000212895,"domain_scores_codex":[0.9931556,0.0009397011,0.003097023,0.000542045,0.001777179,0.0004884681],"domain_scores_gemma":[0.9852898,0.005116026,0.007542063,0.0005409409,0.00109648,0.000414666],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","study_design_scores_codex":[0.0001668139,0.000863522,0.01935751,0.9111242,0.005618048,0.0002436365,0.0001202028,0.000002553022,1.855603e-8,0.00000451226,0.005831591,0.05666742],"study_design_scores_gemma":[0.003360953,0.005451889,0.002314695,0.9095308,0.05259623,0.00008164418,0.0001284818,0.00000582969,5.661709e-7,0.000001962232,0.02610899,0.0004179415],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.0002932737,0.9887714,0.0002712246,0.00128908,0.0006799923,0.008361622,0.0002940574,0.00002748855,0.00001185143],"genre_scores_gemma":[0.001423056,0.994128,0.0001785431,0.0006678983,0.0002442331,0.001345652,0.001840436,0.0001233387,0.00004879565],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.05624947,"threshold_uncertainty_score":0.9998383,"prediction_status":"machine_predicted_unvalidated"},"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":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2949493839","doi":"10.1111/jebm.12347","title":"An evaluation of WHO emergency guidelines for Zika virus disease","year":2019,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":6,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Université Laval; Hôpital de l'Enfant-Jésus","funders":"UNICEF; Wellcome Trust; Wellcome; World Health Organization","keywords":"Guideline; Zika virus; Context (archaeology); Medicine; CLARITY; Medical emergency; Presentation (obstetrics); Critical appraisal; Checklist; Family medicine; Alternative medicine; Psychology; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.9321876143821779,"gpt":0.7733480368508203,"spread":0.1588395775313576,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.02929843,0.000143703,0.0005481645,0.0004489851,0.0001960844,0.000002805726,0.0004127674,0.00007259617,0.005324684],"category_scores_gemma":[0.05255093,0.00008854461,0.0001069188,0.0005067644,0.00008120087,0.0007435528,0.00001811763,0.0002633756,0.00003125828],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002933702,"about_ca_system_score_gemma":0.005612399,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001896992,"about_ca_topic_score_gemma":0.00006982395,"domain_scores_codex":[0.9925827,0.001572707,0.002987114,0.0002041176,0.002246988,0.0004063682],"domain_scores_gemma":[0.9814065,0.00305371,0.002754426,0.000439796,0.01169625,0.0006492889],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"observational","study_design_scores_codex":[0.002869887,0.000186968,0.3781623,0.004095847,0.00007217257,0.000004943827,0.008454384,0.004314592,0.06868479,0.0006652089,0.4938106,0.03867825],"study_design_scores_gemma":[0.02920296,0.01148548,0.6520736,0.03522173,0.001713531,0.000007039308,0.0118637,0.09607743,0.002933806,0.008248857,0.1504349,0.000737059],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8166674,0.001681483,0.01072591,0.1650224,0.003309211,0.00241681,0.00004645017,0.00001315754,0.000117209],"genre_scores_gemma":[0.967694,0.0008550002,0.003362454,0.02459475,0.002952799,0.0001063772,0.000006979105,0.00002557029,0.0004021137],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3433758,"threshold_uncertainty_score":0.9995415,"prediction_status":"machine_predicted_unvalidated"},"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":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4392938758","doi":"10.1111/jebm.12600","title":"Handling time‐varying treatments in observational studies: A scoping review and recommendations","year":2024,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Advanced Causal Inference Techniques","field":"Mathematics","cited_by":5,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; St. Joseph’s Healthcare Hamilton; Impact","funders":"National Natural Science Foundation of China","keywords":"Observational study; Protocol (science); Workflow; Observational methods in psychology; Computer science; Set (abstract data type); Process (computing); Medical physics; Management science; Selection (genetic algorithm); Medicine; Alternative medicine; Machine learning; Engineering","retraction":null,"screen_n_in":null,"score":{"opus":0.833002005037274,"gpt":0.6314645032262648,"spread":0.2015375018110092,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.004029371,0.0004655198,0.003699056,0.000739722,0.00006453998,0.00002250334,0.0002567206,0.0001456656,0.0001291875],"category_scores_gemma":[0.017628,0.0002571627,0.0002677214,0.0008047404,0.0001323782,0.0004847056,0.00007844762,0.0008094297,0.000007316356],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0007047621,"about_ca_system_score_gemma":0.001044717,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000002431085,"about_ca_topic_score_gemma":0.000003384562,"domain_scores_codex":[0.9958295,0.0004109809,0.002715466,0.000302729,0.0005074122,0.0002339689],"domain_scores_gemma":[0.9873541,0.00973238,0.002044111,0.0002736691,0.0004705953,0.0001251366],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","study_design_scores_codex":[0.00001255324,0.00008026924,0.000005939592,0.5247491,0.0006466462,0.0002085265,0.0001921479,0.000001301303,0.000003970945,0.0003858242,0.01209285,0.4616209],"study_design_scores_gemma":[0.000258884,0.0003458553,1.900384e-7,0.9668669,0.006658873,0.0001529174,0.00002379945,0.00001082249,0.000001948589,0.008415849,0.0170657,0.0001982589],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.000001153051,0.9912341,0.001148525,0.005989903,0.0002020823,0.001349834,0.00000756657,0.00004060375,0.00002622278],"genre_scores_gemma":[5.276739e-7,0.9681724,0.03063214,0.0005326254,0.0003106679,0.000134495,0.00001513119,0.00004885899,0.0001531649],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.4614226,"threshold_uncertainty_score":0.9999881,"prediction_status":"machine_predicted_unvalidated"},"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"},{"id":"W1556696529","doi":"10.1111/j.1756-5391.2011.01116.x","title":"A mathematical model for evaluating tuberculosis screening strategies","year":2011,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Tuberculosis Research and Epidemiology","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Tuberculosis; Transmission (telecommunications); Immigration; Disease; Basic reproduction number; Medicine; Population; Environmental health; Order (exchange); Demography; Computer science; Geography; Telecommunications; Internal medicine; Business; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.5856095892322939,"gpt":0.4915099256631166,"spread":0.09409966356917732,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.007485959,0.000230978,0.001055343,0.0004525702,0.0000970082,0.00001240097,0.0002447495,0.0001505518,0.0005540883],"category_scores_gemma":[0.02101432,0.0001264143,0.0003600718,0.0002555246,0.0003166937,0.000440223,0.0000288475,0.0005478295,0.000007879507],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001090592,"about_ca_system_score_gemma":0.0008964315,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003845782,"about_ca_topic_score_gemma":0.00000431769,"domain_scores_codex":[0.9967068,0.0002413473,0.001330153,0.0002350328,0.0009561683,0.0005305227],"domain_scores_gemma":[0.9944072,0.002831886,0.0005513624,0.000316374,0.001306685,0.0005865286],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"bench_or_experimental","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.08332994,0.004265452,0.07328532,0.01489206,0.005991816,0.001838139,0.02070302,0.03218862,0.2967478,0.02378864,0.1629798,0.2799894],"study_design_scores_gemma":[0.006635425,0.01332896,0.007366401,0.01277899,0.001399183,0.0007079498,0.002060955,0.9361802,0.002002296,0.01710978,0.0001862644,0.0002436381],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1629835,0.004255932,0.7873123,0.04298423,0.0001479416,0.0008296975,0.00000353275,0.00003668048,0.001446268],"genre_scores_gemma":[0.8178634,0.000430113,0.1778566,0.002925652,0.000685301,0.00003603726,0.000003066225,0.00003202499,0.0001678388],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.9039915,"threshold_uncertainty_score":0.9872321,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"simulation_or_modeling","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"simulation_or_modeling","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W3192941963","doi":"10.1111/jebm.12440","title":"Comparing and combining evidence of treatment effects in randomized and nonrandomized studies on the use of misoprostol to prevent postpartum hemorrhage","year":2021,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Maternal and fetal healthcare","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"St. Joseph’s Healthcare Hamilton; McMaster University; Impact","funders":"","keywords":"Misoprostol; Medicine; Randomized controlled trial; Confidence interval; Odds ratio; Meta-analysis; Placebo; MEDLINE; Relative risk; Obstetrics; Pregnancy; Internal medicine; Abortion; Alternative medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.4126879839645758,"gpt":0.4530451666427994,"spread":0.04035718267822364,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005626046,0.0004976452,0.009934409,0.0005672387,0.00004644268,0.00001323716,0.0001425145,0.0001281746,0.00001274001],"category_scores_gemma":[0.0195481,0.0002037285,0.0005920943,0.0003320537,0.0003930103,0.000116079,0.00007535798,0.0005239042,4.728702e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002958133,"about_ca_system_score_gemma":0.0006914561,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001905365,"about_ca_topic_score_gemma":0.0000222227,"domain_scores_codex":[0.9936243,0.002201935,0.002893204,0.0003165579,0.0007198243,0.0002441199],"domain_scores_gemma":[0.9640715,0.03230562,0.002402266,0.0003535662,0.0005691212,0.0002979078],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","study_design_scores_codex":[0.2870779,0.001097998,0.001213722,0.3926688,0.0031534,0.003366308,0.004840796,0.00007961935,0.0002911667,0.00008149346,0.000204752,0.3059241],"study_design_scores_gemma":[0.08447317,0.009027326,0.00006670343,0.898142,0.005569266,0.0006854057,0.0002582745,0.0001972598,0.000245601,0.00001293721,0.001153845,0.0001682115],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.06696076,0.925046,0.00002864152,0.004064195,0.0003145979,0.003576755,0.000003060449,0.000003391276,0.000002628],"genre_scores_gemma":[0.06112304,0.9376012,0.0004967728,0.0004174563,0.0001530017,0.0001053855,0.000002414549,0.0000248552,0.00007587598],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.5054732,"threshold_uncertainty_score":0.9887106,"prediction_status":"machine_predicted_unvalidated"},"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":["metaepi_narrow","metaepi_broad","metaresearch"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4408070566","doi":"10.1111/jebm.70006","title":"Inadequate Reporting of Harm From Randomized Clinical Trials in Top Medical Publications","year":2025,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":5,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"National Natural Science Foundation of China","keywords":"Consolidated Standards of Reporting Trials; Harm; Medicine; Randomized controlled trial; Family medicine; Alternative medicine; Descriptive statistics; Medical journal; Psychology; Internal medicine; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.9714145874832354,"gpt":0.7366030185621091,"spread":0.2348115689211263,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_broad","insufficient_payload"],"consensus_categories":["metaresearch","metaepi_broad"],"category_scores_codex":[0.9342591,0.0007257995,0.09094124,0.004388515,0.00006212686,0.0002502177,0.004969014,0.000787591,0.019159],"category_scores_gemma":[0.9934424,0.0002432261,0.02096175,0.005642704,0.000617126,0.0004200699,0.000199665,0.002083716,0.00007989409],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000164963,"about_ca_system_score_gemma":0.009903991,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00009455562,"about_ca_topic_score_gemma":0.00002653564,"domain_scores_codex":[0.1478194,0.4069549,0.4142672,0.002068131,0.02826967,0.0006206638],"domain_scores_gemma":[0.009737177,0.6556275,0.3229223,0.004247249,0.006682217,0.0007835479],"domain_codex":"reporting","domain_gemma":"methods","domain_candidate":"methods","domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0007209166,0.0001251629,0.00049426,0.004380804,0.001292302,0.000105175,0.00006450417,0.000004173253,1.935776e-7,0.0001790017,0.05719391,0.9354396],"study_design_scores_gemma":[0.01683889,0.0001411491,0.00006436316,0.2667521,0.01381037,0.00005243263,0.0001278556,0.0005890349,4.924236e-7,0.002170511,0.6992477,0.0002051808],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.0000291338,0.943339,0.02982759,0.02182151,0.002126747,0.002322776,0.00001414844,0.000002979312,0.0005160479],"genre_scores_gemma":[0.0001520793,0.9923363,0.003333644,0.0006896951,0.001725241,0.00009074089,0.00001974599,0.00001801696,0.001634515],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.9352344,"threshold_uncertainty_score":0.9957089,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4399901519","doi":"10.1111/jebm.12622","title":"Impact of hormone replacement therapy on all‐cause and cancer‐specific mortality in colorectal cancer: A systematic review and dose‒response meta‐analysis of observational studies","year":2024,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Menopause: Health Impacts and Treatments","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"Natural Science Foundation of Sichuan Province","keywords":"Medicine; Meta-analysis; Hazard ratio; Internal medicine; Colorectal cancer; Hormone replacement therapy (female-to-male); Cohort study; Cochrane Library; Oncology; Cohort; Cancer; Confidence interval; Testosterone (patch)","retraction":null,"screen_n_in":null,"score":{"opus":0.6676574156498694,"gpt":0.5740848911037709,"spread":0.09357252454609843,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.007440194,0.0006506704,0.0134851,0.001484742,0.00003363629,0.00001019702,0.0001513674,0.0001467541,0.0001156954],"category_scores_gemma":[0.002303995,0.00027541,0.000554661,0.001672739,0.000247592,0.0001278023,0.00003120674,0.0005271497,3.510324e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001201229,"about_ca_system_score_gemma":0.002430171,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0005691063,"about_ca_topic_score_gemma":0.00006124006,"domain_scores_codex":[0.9927619,0.001295022,0.003925103,0.0004223191,0.001301139,0.0002945184],"domain_scores_gemma":[0.9901195,0.0040391,0.004206913,0.0004439285,0.0007912802,0.000399331],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","study_design_scores_codex":[0.005167545,0.0003377451,0.0008447156,0.624846,0.3651311,0.0002760632,0.0003084311,0.00001277716,0.00006827954,0.000004496674,0.0009528654,0.002050039],"study_design_scores_gemma":[0.002267655,0.01353411,0.005858697,0.4557185,0.5219544,0.00006362884,0.00002179731,0.000007819287,0.000002096246,0.000003421944,0.0004447473,0.0001230919],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.03935488,0.9539961,8.918393e-7,0.003072308,0.0001048998,0.003279169,0.0001860616,0.000004805212,9.278385e-7],"genre_scores_gemma":[0.009898029,0.9890897,0.00003448738,0.0004666071,0.00008994188,0.0003388748,0.00002047504,0.00003415769,0.00002774688],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.1691275,"threshold_uncertainty_score":0.9999698,"prediction_status":"machine_predicted_unvalidated"},"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":"W4394722496","doi":"10.1111/jebm.12606","title":"The detection of circulating tumor cells indicates poor therapeutic efficacy and prognosis in patients with nonsmall cell lung cancer: A systematic review and meta‐analysis","year":2024,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Cancer Cells and Metastasis","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Medicine; Meta-analysis; Internal medicine; Oncology; Circulating tumor cell; Cochrane Library; Hazard ratio; Lung cancer; Cancer; Confidence interval; Metastasis","retraction":null,"screen_n_in":null,"score":{"opus":0.08999203785124837,"gpt":0.3642577766812765,"spread":0.2742657388300281,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003113172,0.0005027922,0.007242504,0.0009375692,0.0000660255,0.00003277044,0.0001904245,0.00009814542,0.00002857472],"category_scores_gemma":[0.0004464527,0.0001917486,0.0008494268,0.001918339,0.0002099024,0.00009620111,0.00003444694,0.0007189314,3.767407e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002266209,"about_ca_system_score_gemma":0.0006796464,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001065398,"about_ca_topic_score_gemma":0.00003940744,"domain_scores_codex":[0.9955382,0.0005710069,0.002352894,0.0003638283,0.000911461,0.0002626268],"domain_scores_gemma":[0.9937072,0.001895011,0.003424704,0.0002860831,0.0004659657,0.0002210545],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","study_design_scores_codex":[0.0001125465,0.0001044321,0.0002257874,0.9397087,0.04307275,0.00004975554,0.00008543859,0.000003317229,0.00001913857,3.511511e-7,0.00001832073,0.01659947],"study_design_scores_gemma":[0.0007267434,0.0007766583,0.00006194378,0.3923241,0.6053117,0.00004245918,0.00002608333,0.00001908331,0.00002910969,4.39389e-7,0.0005650291,0.0001166887],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.000302249,0.9947087,0.00004488189,0.001041619,0.0000857599,0.003791542,0.00001644342,0.000006136994,0.000002687849],"genre_scores_gemma":[0.03068006,0.9685931,0.00007443246,0.0002826677,0.00007502355,0.0002224359,0.000003744237,0.00004582034,0.00002275673],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.5622389,"threshold_uncertainty_score":0.7819281,"prediction_status":"machine_predicted_unvalidated"},"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":"W1625449191","doi":"10.1111/j.1756-5391.2011.01127.x","title":"Visualization studies on evidence‐based medicine domain knowledge (series 2): structural diagrams of author networks","year":2011,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":4,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"China Medical Board","keywords":"Zhàng; China; Visualization; Citation; Cluster analysis; Graph drawing; Computer science; Domain (mathematical analysis); Library science; Data science; Information retrieval; Geography; Data mining; Artificial intelligence; Mathematics","retraction":null,"screen_n_in":null,"score":{"opus":0.9085942578151824,"gpt":0.6040802977165222,"spread":0.3045139600986603,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.1544939,0.0006188874,0.00654662,0.001593136,0.0002150049,0.00008201125,0.002090905,0.000180402,0.005565972],"category_scores_gemma":[0.2378163,0.0002295721,0.001342356,0.003279848,0.001149353,0.001132606,0.00008317696,0.0004644801,0.00005182428],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001701145,"about_ca_system_score_gemma":0.0004370253,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002353862,"about_ca_topic_score_gemma":0.00005215135,"domain_scores_codex":[0.9627207,0.01203519,0.01622262,0.0007479131,0.007862512,0.0004110762],"domain_scores_gemma":[0.9390944,0.02841756,0.02054546,0.002320327,0.00909028,0.0005319876],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","study_design_scores_codex":[0.007976664,0.0009861822,0.2253366,0.004489833,0.005074023,0.0006304953,0.04971727,0.01627192,0.006228671,0.03304093,0.5297458,0.1205017],"study_design_scores_gemma":[0.0169264,0.07174381,0.2332387,0.3027942,0.01860823,0.0005903781,0.07009902,0.1386477,0.01043741,0.08909927,0.04435157,0.003463283],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.4167821,0.2788842,0.235272,0.05330828,0.01189063,0.002790584,0.000007784271,0.00003287692,0.00103155],"genre_scores_gemma":[0.9911826,0.001485296,0.003692849,0.001041625,0.001627153,0.00001678763,0.000002585522,0.00002601148,0.0009250678],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5744005,"threshold_uncertainty_score":0.9953431,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["bibliometrics"],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["bibliometrics"],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2028332028","doi":"10.1111/jebm.12101","title":"Production and citation of cochrane systematic reviews: a bibliometrics analysis","year":2014,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":4,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Citation; Bibliometrics; Systematic review; Cochrane Library; Psychological intervention; Knowledge translation; Production (economics); Medicine; MEDLINE; Business; Political science; Meta-analysis; Library science; Computer science; Economics; Nursing; Knowledge management; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.8622232322885919,"gpt":0.5786419193064429,"spread":0.2835813129821489,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","bibliometrics","insufficient_payload"],"consensus_categories":["metaresearch","bibliometrics"],"category_scores_codex":[0.407464,0.0002458996,0.008243145,0.01850872,0.00006494156,0.0001528643,0.0008644093,0.00006911812,0.0009799162],"category_scores_gemma":[0.7342431,0.00008815822,0.001541863,0.04066173,0.0001597625,0.0006073672,0.00002471358,0.000178973,0.00004090866],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004330505,"about_ca_system_score_gemma":0.00008820736,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00000906137,"about_ca_topic_score_gemma":0.000008356973,"domain_scores_codex":[0.9533605,0.01742507,0.01911065,0.0004874894,0.009451026,0.0001652845],"domain_scores_gemma":[0.8872349,0.0490125,0.04740143,0.003116196,0.01281777,0.0004172407],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"systematic_review","study_design_scores_codex":[0.0005212899,0.000785338,0.3923407,0.2155563,0.01066274,0.00004642158,0.006622732,0.01074063,0.01871514,0.001697826,0.2291889,0.1131219],"study_design_scores_gemma":[0.004427184,0.009323125,0.2358964,0.4416213,0.1113534,0.0005985413,0.007862932,0.1540873,0.001641073,0.01113862,0.02056152,0.001488631],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.2153637,0.1866181,0.5736138,0.02031923,0.001345357,0.002427032,0.000002139933,0.000004649571,0.0003059529],"genre_scores_gemma":[0.9873437,0.005803262,0.005628817,0.0003244275,0.0003587599,0.00001224781,9.358001e-7,0.000006796563,0.0005210123],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.77198,"threshold_uncertainty_score":0.9999333,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null}]}