{"meta":{"page":1,"per_page":50,"max_per_page":100,"total":60,"total_is_capped":false,"direct_labels_cover":60,"predictions_cover":60,"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":"cf12b98fc3a7","filters":{"category":"metaepi_narrow"}},"results":[{"id":"W1561379059","doi":"10.1371/journal.pone.0132634","title":"PEDro or Cochrane to Assess the Quality of Clinical Trials? A Meta-Epidemiological Study","year":2015,"lang":"en","type":"review","venue":"PLoS ONE","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":173,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"Fondation pour la Recherche Médicale; Physiotherapy Foundation of Canada; Alberta Innovates; Killam Trusts; National Institutes of Health; Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung; University of Alberta; Alberta Innovates - Health Solutions; Women and Children's Health Research Institute; Children's Health Research Institute; Bill and Melinda Gates Foundation; Canadian Institutes of Health Research; National Science Foundation","keywords":"Blinding; Medicine; Meta-analysis; Systematic review; Clinical trial; MEDLINE; Physical therapy; Internal medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.9978832001452151,"gpt":0.8031065913405215,"spread":0.1947766088046936,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","open_science","insufficient_payload"],"consensus_categories":["metaresearch","metaepi_broad","insufficient_payload"],"category_scores_codex":[0.9341789,0.001104338,0.1499178,0.0003863922,0.0001198605,0.0005167568,0.007663617,0.0005806173,0.01905786],"category_scores_gemma":[0.9231139,0.0002626773,0.02818557,0.002980128,0.0001806183,0.0001010436,0.0009175164,0.001075259,0.005094427],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000042313,"about_ca_system_score_gemma":0.0007532155,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002597151,"about_ca_topic_score_gemma":0.00009582097,"domain_scores_codex":[0.1056691,0.7451407,0.121547,0.003305061,0.02376516,0.0005729304],"domain_scores_gemma":[0.1135538,0.7561658,0.09788492,0.02273848,0.008476942,0.001180119],"domain_codex":"methods","domain_gemma":"methods","domain_candidate":"methods","domain_consensus":"methods","study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0001988807,0.01895623,0.001917213,0.01861689,0.1459693,0.00002723639,0.0003440042,0.000002567477,2.793002e-7,0.000930585,0.1064532,0.7065836],"study_design_scores_gemma":[0.0003326504,0.0007960787,0.0002471897,0.001921868,0.1447193,0.000003288578,0.0003832311,0.00003202651,6.239053e-8,0.0006348988,0.8505335,0.0003959252],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00049421,0.984337,0.0007223492,0.0004572561,0.0002733341,0.01202155,0.0002125917,0.00001025036,0.001471453],"genre_scores_gemma":[0.000144131,0.9763132,0.005798894,0.0003487881,0.0007326962,0.001820064,0.00002176389,0.00003707811,0.01478335],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.7440803,"threshold_uncertainty_score":0.9999825,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"evaluation","study_design":"observational","genre":"review","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":"W2018803201","doi":"10.1016/s0895-4356(99)00160-2","title":"Randomized trials versus observational studies in adolescent pregnancy prevention","year":2000,"lang":"en","type":"article","venue":"Journal of Clinical Epidemiology","topic":"Adolescent Sexual and Reproductive Health","field":"Health Professions","cited_by":94,"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":"Observational study; Randomized controlled trial; Medicine; Psychological intervention; CINAHL; PsycINFO; Pregnancy; MEDLINE; Family medicine; Psychiatry; Surgery; Internal medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.9380215926437274,"gpt":0.7360098352874899,"spread":0.2020117573562376,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.1770054,0.000191645,0.007367685,0.0001289413,0.0001721783,9.952893e-7,0.0002116023,0.0004799194,0.0008701438],"category_scores_gemma":[0.4180903,0.0001115708,0.00108259,0.0001747701,0.0004860715,0.0001510685,0.00004963954,0.002277213,0.0001145771],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001482299,"about_ca_system_score_gemma":0.0007974164,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001204573,"about_ca_topic_score_gemma":0.00002022483,"domain_scores_codex":[0.8795796,0.1023109,0.01685805,0.000401983,0.0002808241,0.0005686444],"domain_scores_gemma":[0.6621632,0.3215996,0.01422607,0.0004697287,0.001097191,0.0004442509],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"randomized_trial","study_design_gemma":"observational","study_design_scores_codex":[0.4769093,0.001945216,0.3827553,0.000774741,0.0008516642,0.00003201828,0.001006764,0.0004091212,0.000002331863,0.009457876,0.02192391,0.1039318],"study_design_scores_gemma":[0.3896682,0.001267016,0.5075793,0.005719026,0.0002546151,0.00001017788,0.0008609321,0.0001647545,6.55908e-7,0.07298058,0.02128543,0.0002093402],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9095352,0.02140174,0.001525789,0.04481123,0.01838167,0.002795219,0.000006712345,0.00002970165,0.00151268],"genre_scores_gemma":[0.7869185,0.1487844,0.01304187,0.01558077,0.02078501,0.0002140157,0.00002558165,0.00006459543,0.01458533],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2410849,"threshold_uncertainty_score":0.9893484,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W3110084249","doi":"10.1016/j.joca.2020.10.004","title":"Visual Analogue Scale has higher assay sensitivity than WOMAC pain in detecting between-group differences in treatment effects: a meta-epidemiological study","year":2020,"lang":"en","type":"review","venue":"Osteoarthritis and Cartilage","topic":"Osteoarthritis Treatment and Mechanisms","field":"Medicine","cited_by":85,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Impact; Dalhousie University; McMaster University; University of Toronto; St. Michael's Hospital","funders":"Arthritis Society","keywords":"WOMAC; Visual analogue scale; Medicine; Osteoarthritis; Meta-analysis; Randomized controlled trial; Placebo; Physical therapy; Internal medicine; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.08019099511676102,"gpt":0.321414104342709,"spread":0.2412231092259479,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.003201714,0.001304107,0.008256466,0.0003837108,0.0001810684,0.0001267445,0.0001003894,0.000657927,0.00004408307],"category_scores_gemma":[0.0004299374,0.0008770716,0.001139031,0.0006258423,0.0001662061,0.0001181884,0.000161213,0.0007736042,0.0000218449],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003119498,"about_ca_system_score_gemma":0.0001157848,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002059016,"about_ca_topic_score_gemma":0.007864297,"domain_scores_codex":[0.9889678,0.006916886,0.001321366,0.001506055,0.0004081302,0.0008797527],"domain_scores_gemma":[0.995715,0.003015121,0.0003726061,0.0004011873,0.00002983059,0.0004662345],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"nonrandomized_trial","study_design_scores_codex":[0.00007027166,0.000726229,0.02863054,0.00128948,0.0002192847,0.00653936,0.0002900121,2.04711e-8,0.0001253394,0.000005871539,0.00000298585,0.9621006],"study_design_scores_gemma":[0.07494086,0.3626202,0.04163206,0.08879831,0.08005293,0.001853316,0.004789604,0.00003422134,0.001202415,0.0003673017,0.3341056,0.009603172],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"empirical","genre_scores_codex":[0.3400274,0.6559978,0.000009322458,0.00005713804,0.0001258486,0.003526066,0.0000963381,0.00008752849,0.00007257099],"genre_scores_gemma":[0.6172987,0.3805401,0.00008721308,0.00005505942,0.0002719334,0.001380742,0.0002279221,0.0000735585,0.00006479561],"genre_candidate":"review","genre_consensus":null,"teacher_disagreement_score":0.9524974,"threshold_uncertainty_score":0.999971,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"observational","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":"W2287844929","doi":"10.1136/bmjopen-2015-008562","title":"What is the influence of randomisation sequence generation and allocation concealment on treatment effects of physical therapy trials? A meta-epidemiological study","year":2015,"lang":"en","type":"review","venue":"BMJ Open","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":81,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Alberta Health; University of Alberta","funders":"Canadian Institutes of Health Research; Alberta Innovates; University of Alberta; Physiotherapy Foundation of Canada","keywords":"Medicine; Epidemiology; Meta-analysis; Alternative medicine; Sequence (biology); Public health; MEDLINE; Physical therapy; Gerontology; Internal medicine; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.9578471903528902,"gpt":0.6968038048601565,"spread":0.2610433854927338,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.2597782,0.0006103638,0.0228565,0.0001473997,0.00008529533,0.0007136922,0.001684872,0.0001291728,0.0001144386],"category_scores_gemma":[0.04089486,0.0001620402,0.003546937,0.0005939548,0.0001101464,0.0004019874,0.0001864226,0.0001053055,0.00006227122],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00009638568,"about_ca_system_score_gemma":0.0003140209,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00008946395,"about_ca_topic_score_gemma":0.00001204875,"domain_scores_codex":[0.8752403,0.1083557,0.0116893,0.00113067,0.003440675,0.0001433316],"domain_scores_gemma":[0.9373533,0.0403589,0.01786323,0.00338965,0.0009448658,0.00009009578],"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":[0.0001262602,0.0004347092,0.00002290528,0.0007320002,0.00606437,0.000001217195,0.002651113,0.0002392935,0.00001898614,0.0004433747,0.0009003665,0.9883654],"study_design_scores_gemma":[0.01500105,0.01221674,0.0005663764,0.009909393,0.0917971,0.00002279456,0.002132072,0.005955318,0.0002177866,0.004592742,0.8561955,0.001393102],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.01744215,0.9030035,0.0001322808,0.0003005661,0.00014601,0.07889667,0.00002191834,0.000001786729,0.00005516427],"genre_scores_gemma":[0.01200136,0.9725461,0.0001832829,0.0001878162,0.0001232537,0.01447628,0.00002994931,0.00001249472,0.0004394458],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.9869723,"threshold_uncertainty_score":0.9916502,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2111376665","doi":"10.1177/0306624x12447771","title":"A Meta-Analysis of Moral Reconation Therapy","year":2012,"lang":"en","type":"review","venue":"International Journal of Offender Therapy and Comparative Criminology","topic":"Psychology of Moral and Emotional Judgment","field":"Neuroscience","cited_by":76,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Saskatchewan","funders":"","keywords":"Psychology; Medicine; Psychotherapist","retraction":null,"screen_n_in":null,"score":{"opus":0.9301869343671447,"gpt":0.4773109657371682,"spread":0.4528759686299765,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004255987,0.000356944,0.002895234,0.001160796,0.00004985245,0.0000279123,0.0007305106,0.0002108196,0.0008785339],"category_scores_gemma":[0.00002050247,0.0002267919,0.001956008,0.0002308444,0.0003480701,0.0002387958,0.00005559848,0.0004270689,0.000009801402],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005117994,"about_ca_system_score_gemma":0.00008674535,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000005479511,"about_ca_topic_score_gemma":0.000001321657,"domain_scores_codex":[0.9970163,0.000815378,0.001147601,0.000314353,0.0005065094,0.0001998567],"domain_scores_gemma":[0.9967557,0.0006872258,0.001925314,0.0001637413,0.0003837558,0.00008430172],"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":[0.0005264998,0.0009601979,0.00002140243,0.00007955974,0.2451028,0.00003822367,0.001119522,0.00002350269,0.0002597783,0.01910421,0.0002833051,0.7324811],"study_design_scores_gemma":[0.001344087,0.0007326173,0.0005744123,0.0001024657,0.09277081,0.0008899009,0.00008974843,0.00001908826,0.000917912,0.003676353,0.8984349,0.0004477268],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.004499001,0.9928672,0.00012986,0.0005068771,0.001099209,0.0001964001,0.00008434778,0.000006672552,0.000610405],"genre_scores_gemma":[0.04249522,0.9563383,0.0001507079,0.0006325035,0.00018698,0.0000190431,0.0000209884,0.00001053667,0.0001457138],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.8981516,"threshold_uncertainty_score":0.961933,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gpt","categories":["metaepi_narrow","metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3172845662","doi":"10.1186/s12916-021-02008-2","title":"Utilization of the evidence from studies with no events in meta-analyses of adverse events: an empirical investigation","year":2021,"lang":"en","type":"article","venue":"BMC Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":53,"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":"Medical Research Council; National Health and Medical Research Council; Qatar National Library","keywords":"Systematic review; Medicine; Meta-analysis; Adverse effect; MEDLINE; Internal medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.951712016036363,"gpt":0.6398872374389903,"spread":0.3118247785973727,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.03723853,0.0001921969,0.003425883,0.0001945337,0.00003790624,0.000004260979,0.0007599278,0.000043115,0.002704557],"category_scores_gemma":[0.1431453,0.0000610181,0.0006530058,0.002445743,0.000171923,0.0002729145,0.0001057921,0.00007761607,0.000034273],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002526644,"about_ca_system_score_gemma":0.0001896791,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001632446,"about_ca_topic_score_gemma":0.00139813,"domain_scores_codex":[0.9757277,0.01329116,0.005685157,0.0006231412,0.004574462,0.00009840299],"domain_scores_gemma":[0.9833187,0.006911695,0.004364193,0.002618571,0.002708128,0.000078731],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00001749747,0.00004992611,0.9927999,0.00009534943,0.001282055,0.000001121695,0.002172557,0.0007645584,0.001464318,0.00002571529,0.001187529,0.0001394199],"study_design_scores_gemma":[0.0004963183,0.0001315343,0.9685357,0.0010076,0.00902125,0.000002390137,0.007297488,0.005882036,0.002920698,0.004396257,0.0002022134,0.0001065316],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9843494,0.007303016,0.006665473,0.001046759,0.0001398708,0.0003977926,0.000006446047,0.000001536853,0.00008973976],"genre_scores_gemma":[0.9937397,0.0001107135,0.004849961,0.0002520579,0.0000423824,0.00002182352,0.000009947655,0.000005014155,0.0009684507],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1059068,"threshold_uncertainty_score":0.9982071,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2802073505","doi":"10.1183/13993003.00376-2018","title":"The effect of anti-acid therapy on survival in idiopathic pulmonary fibrosis: a methodological review of observational studies","year":2018,"lang":"en","type":"review","venue":"European Respiratory Journal","topic":"Interstitial Lung Diseases and Idiopathic Pulmonary Fibrosis","field":"Medicine","cited_by":50,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McGill University; Jewish General Hospital","funders":"","keywords":"Medicine; Observational study; Pulmonary fibrosis; Idiopathic pulmonary fibrosis; Intensive care medicine; Internal medicine; Fibrosis; Lung","retraction":null,"screen_n_in":null,"score":{"opus":0.2886020444561501,"gpt":0.4418224539358594,"spread":0.1532204094797093,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.01655984,0.0005982317,0.003737854,0.0002527365,0.0002052364,0.00002358278,0.0006151007,0.0001580423,0.00004274313],"category_scores_gemma":[0.003589423,0.0002979955,0.001759242,0.0004939828,0.0005351189,0.00007560585,0.0001670747,0.001239795,0.00002977589],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001551596,"about_ca_system_score_gemma":0.0004540628,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":4.856948e-7,"about_ca_topic_score_gemma":1.225161e-7,"domain_scores_codex":[0.9822217,0.01361535,0.002453263,0.0004500392,0.0008870597,0.0003726119],"domain_scores_gemma":[0.9951721,0.001634465,0.001902992,0.0006512205,0.0004630026,0.0001761895],"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":[0.005455421,0.0003663014,0.0007200238,0.2255891,0.001417437,0.001847278,0.0000513023,7.565509e-7,0.0002616068,0.0001160183,0.003426924,0.7607478],"study_design_scores_gemma":[0.0005292386,0.01197743,0.001455397,0.4056505,0.0007482783,0.0003594836,0.00001255881,5.33522e-7,0.00002879275,0.000007635885,0.5789949,0.0002352533],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.001163295,0.9961259,0.000004197668,0.00008129908,0.0007622684,0.0009602903,0.00005086165,0.0000155628,0.0008363171],"genre_scores_gemma":[0.0005326872,0.9975029,0.00004789231,0.0003539831,0.001398185,0.00003082334,0.00001400371,0.00008287852,0.00003663695],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.7605125,"threshold_uncertainty_score":0.9999472,"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_narrow","metaepi_broad","metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2096833901","doi":"10.1371/journal.pone.0088008","title":"Quantifying Bias in Randomized Controlled Trials in Child Health: A Meta-Epidemiological Study","year":2014,"lang":"en","type":"review","venue":"PLoS ONE","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":43,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"Canadian Institutes of Health Research","keywords":"Epidemiology; Randomized controlled trial; Meta-analysis; Medicine; MEDLINE; Environmental health; Biology; Internal medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.9817470593674227,"gpt":0.6519590330899844,"spread":0.3297880262774383,"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_narrow","metaepi_broad","insufficient_payload"],"category_scores_codex":[0.9492888,0.001662827,0.4463966,0.003379816,0.0001339287,0.0008348197,0.003374443,0.0005250056,0.007294501],"category_scores_gemma":[0.9237,0.0005016659,0.0609404,0.003533772,0.000098009,0.0001198615,0.000298566,0.001320991,0.002405271],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001366804,"about_ca_system_score_gemma":0.0003052994,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001237182,"about_ca_topic_score_gemma":0.0005874893,"domain_scores_codex":[0.04541874,0.8011624,0.1395652,0.00308304,0.009936431,0.0008341788],"domain_scores_gemma":[0.08946429,0.7860366,0.1148095,0.008275912,0.0009261728,0.0004875707],"domain_codex":"methods","domain_gemma":"methods","domain_candidate":"methods","domain_consensus":"methods","study_design_codex":"design_other","study_design_gemma":"randomized_trial","study_design_scores_codex":[0.01791234,0.03241106,0.001923733,0.05848724,0.4068525,0.0001533917,0.001592524,0.00005472443,1.170478e-7,0.004295456,0.007902191,0.4684147],"study_design_scores_gemma":[0.3408185,0.0005750929,0.00005245396,0.05057988,0.3075927,0.00001591912,0.0003522638,0.003479317,3.592372e-8,0.002379395,0.2924202,0.001734246],"study_design_candidate":"meta_analysis","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.0001497156,0.9530569,0.0003556894,0.0008172194,0.0001473438,0.04447864,0.00002128772,0.0000115455,0.0009617034],"genre_scores_gemma":[0.0007163549,0.9849271,0.001863931,0.0003703247,0.0002098683,0.0107133,0.00002520355,0.00004051879,0.001133401],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.4666805,"threshold_uncertainty_score":0.9997435,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2102685275","doi":"10.1186/2046-4053-2-24","title":"A protocol for a systematic review on the impact of unpublished studies and studies published in the gray literature in meta-analyses","year":2013,"lang":"en","type":"review","venue":"Systematic Reviews","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":41,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University","funders":"European Commission; McMaster University","keywords":"Medicine; Grey literature; Meta-analysis; Systematic review; MEDLINE; Cochrane Library; Publication bias; Gray (unit); Protocol (science); Inclusion and exclusion criteria; Gerontology; Alternative medicine; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.9180444448269799,"gpt":0.6734652623171064,"spread":0.2445791825098735,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","scholarly_communication","open_science"],"consensus_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"category_scores_codex":[0.5709826,0.003320971,0.1390343,0.002275784,0.0002317146,0.004017989,0.008564133,0.0004152449,0.0006192197],"category_scores_gemma":[0.6502359,0.0006125668,0.03255719,0.01166025,0.0002821168,0.000995461,0.0004855628,0.001062481,0.0004301972],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003485228,"about_ca_system_score_gemma":0.0003241393,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001876829,"about_ca_topic_score_gemma":0.00005221459,"domain_scores_codex":[0.5639909,0.3256722,0.09641302,0.003443433,0.009424353,0.001056055],"domain_scores_gemma":[0.5283254,0.2607465,0.1674074,0.03237944,0.0107757,0.0003655773],"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":[0.000001691074,0.00008234962,0.00000136138,0.8880777,0.01672149,0.000004649245,0.001427192,7.394016e-7,1.829819e-8,0.0007809234,0.09211556,0.000786303],"study_design_scores_gemma":[0.0001905577,0.000137278,5.838609e-7,0.8876513,0.05262838,0.00008192773,0.001496682,0.00003054093,1.908065e-8,0.002265433,0.05502931,0.0004880366],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"review","genre_gemma":"protocol","genre_scores_codex":[1.395982e-7,0.5444773,0.00001847993,0.0003146271,0.00005147674,0.4549735,0.00003212164,0.000003616765,0.0001286807],"genre_scores_gemma":[0.000001487291,0.4987681,0.0001676879,0.0002172592,0.00003579773,0.4999435,0.000009013341,0.00003716607,0.0008200838],"genre_candidate":"review","genre_consensus":null,"teacher_disagreement_score":0.1064771,"threshold_uncertainty_score":0.9996326,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2749388036","doi":"10.1177/0022034517725049","title":"Impact of Selection Bias on Treatment Effect Size Estimates in Randomized Trials of Oral Health Interventions: A Meta-epidemiological Study","year":2017,"lang":"en","type":"review","venue":"Journal of Dental Research","topic":"Statistical Methods in Clinical Trials","field":"Mathematics","cited_by":41,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Institute of Health Economics; University of Alberta","funders":"","keywords":"Selection bias; Meta-analysis; Epidemiology; Randomized controlled trial; Psychological intervention; Medicine; Sample size determination; Oral health; Effect modification; Environmental health; Selection (genetic algorithm); Statistics; Dentistry; Confidence interval; Internal medicine; Computer science; Mathematics; Psychiatry; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.9859906141623999,"gpt":0.8396295782838669,"spread":0.146361035878533,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch","metaepi_broad"],"category_scores_codex":[0.3966984,0.0006344393,0.04049053,0.001241036,0.00009446629,0.00006680262,0.0007997012,0.0004029538,0.0005341789],"category_scores_gemma":[0.9190332,0.000255489,0.01776442,0.000551915,0.0004685759,0.00007477305,0.0001882026,0.001716402,0.00000618795],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001341495,"about_ca_system_score_gemma":0.001015337,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0004981416,"about_ca_topic_score_gemma":0.00006759151,"domain_scores_codex":[0.7887486,0.1932061,0.01505588,0.0004762811,0.001915487,0.0005976062],"domain_scores_gemma":[0.240725,0.7440867,0.01379035,0.0005592316,0.0005505166,0.0002882045],"domain_codex":null,"domain_gemma":"methods","domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"randomized_trial","study_design_scores_codex":[0.05611753,0.009903557,0.000677562,0.01477324,0.03156062,0.00009783376,0.00004973455,0.000004180853,0.000001256548,0.0002918303,0.0003792176,0.8861434],"study_design_scores_gemma":[0.2951093,0.2226424,0.001380991,0.1331023,0.06186797,0.0003510771,0.0001101267,0.00007055094,0.00004966435,0.2839568,0.0006320681,0.0007267734],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.007672248,0.9780841,0.0008882897,0.00003158007,0.0003973281,0.01262134,0.0002589594,0.000008245147,0.00003784236],"genre_scores_gemma":[0.002802078,0.9369317,0.05957658,0.000001208208,0.000160267,0.0004215132,0.000002213889,0.0000480612,0.00005631909],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.8854166,"threshold_uncertainty_score":0.9999897,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"meta_analysis","genre":"review","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":"high"}],"label_agreement":"agree"},{"id":"W2756094298","doi":"10.1136/ebmed-2017-110798","title":"What does expert opinion in guidelines mean? a meta-epidemiological study","year":2017,"lang":"en","type":"review","venue":"Evidence-Based Medicine","topic":"Clinical practice guidelines implementation","field":"Medicine","cited_by":38,"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; Observational study; CLARITY; Medicine; Expert opinion; Evidence-based medicine; Evidence-based practice; Family medicine; MEDLINE; Alternative medicine; Pathology; Intensive care medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.934234501038434,"gpt":0.7009597189922305,"spread":0.2332747820462034,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.02875996,0.001530234,0.01195205,0.0009434737,0.0002009736,0.0001269515,0.001087627,0.0007750569,0.002173872],"category_scores_gemma":[0.2157052,0.0007154682,0.001842877,0.0006673709,0.0005355329,0.001400383,0.0003349974,0.001880563,0.000150315],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0005990202,"about_ca_system_score_gemma":0.0019079,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002254777,"about_ca_topic_score_gemma":0.0005972738,"domain_scores_codex":[0.9823679,0.003986571,0.008747175,0.002097477,0.001950445,0.0008504786],"domain_scores_gemma":[0.964281,0.02669299,0.004089397,0.003162913,0.001136432,0.0006372398],"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":[0.0004012448,0.000711383,0.0003713684,0.00715709,0.00162842,0.0003425265,0.0002399477,0.000003738416,0.000001242444,0.000008524416,0.02925732,0.9598772],"study_design_scores_gemma":[0.00312249,0.002506773,0.0001375172,0.1395384,0.01100705,0.00007603854,0.00158214,0.0000770148,5.303465e-7,0.00006510001,0.8413488,0.000538161],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00001497996,0.9121943,0.0001927012,0.07438689,0.005917788,0.007102703,0.00001099489,0.0001412333,0.00003841961],"genre_scores_gemma":[0.00006778321,0.9787335,0.002031612,0.00871586,0.007611544,0.001902507,0.0003813958,0.0001273962,0.0004284065],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.959339,"threshold_uncertainty_score":0.9997447,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2901826044","doi":"10.1093/ije/dyy239","title":"How often can meta-analyses of individual-level data individualize treatment? A meta-epidemiologic study","year":2018,"lang":"en","type":"review","venue":"International Journal of Epidemiology","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":32,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Ottawa Hospital","funders":"Canadian Institutes of Health Research; Nederlandse Organisatie voor Wetenschappelijk Onderzoek; Laura and John Arnold Foundation","keywords":"Subgroup analysis; Meta-analysis; Medicine; Cochrane Library; MEDLINE; Randomized controlled trial; Internal medicine; Biology","retraction":null,"screen_n_in":null,"score":{"opus":0.9907924205974797,"gpt":0.7172303309820024,"spread":0.2735620896154773,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","open_science","insufficient_payload"],"consensus_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"category_scores_codex":[0.470414,0.001815958,0.07460891,0.003700684,0.0001158933,0.0005911972,0.02699345,0.0007299705,0.008299517],"category_scores_gemma":[0.5215519,0.0006480249,0.03508837,0.001619026,0.0005851132,0.0006768075,0.002405402,0.000927817,0.0002608826],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002679116,"about_ca_system_score_gemma":0.001089187,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002061561,"about_ca_topic_score_gemma":0.0002072823,"domain_scores_codex":[0.7422636,0.1848127,0.05985018,0.003104958,0.009160831,0.000807739],"domain_scores_gemma":[0.6393232,0.1803047,0.1580404,0.01248094,0.009204438,0.0006462953],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"meta_analysis","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00001857029,0.0006218968,0.001219303,0.0002433568,0.6077243,0.0001186731,0.0002001873,0.00002277863,1.568061e-7,0.0004393044,0.1330898,0.2563016],"study_design_scores_gemma":[0.0002906785,0.0005725845,0.0001824553,0.0001921865,0.2953204,0.0004728513,0.0002028498,0.00002210259,2.886335e-7,0.00577676,0.6966591,0.0003077263],"study_design_candidate":"meta_analysis","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00004119512,0.9816108,0.00591651,0.003324144,0.001642613,0.001435252,0.005821424,0.000004738504,0.0002033454],"genre_scores_gemma":[0.0004568326,0.9651061,0.0269336,0.001126954,0.001342982,0.0001612979,0.0008198588,0.00006403263,0.003988348],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.5635693,"threshold_uncertainty_score":0.9995971,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3097681315","doi":"10.1186/s40814-020-00712-z","title":"More pilot trials could plan to use qualitative data: a meta-epidemiological study","year":2020,"lang":"en","type":"article","venue":"Pilot and Feasibility Studies","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":29,"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; Population Health Research Institute; Laurentian University; McMaster University; Hamilton Health Sciences; Toronto General Hospital; University Health Network; Impact; Health Sciences Centre","funders":"","keywords":"Qualitative research; Medicine; Qualitative property; Psychological intervention; Clinical trial; Research design; Randomized controlled trial; Confidence interval; Data collection; Family medicine; Odds ratio; Clinical study design; Physical therapy; Nursing; Internal medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.9954596694666988,"gpt":0.7332717440502268,"spread":0.2621879254164721,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.3544163,0.0006434463,0.01532445,0.0001405645,0.0003638921,0.0006633941,0.001971982,0.00004699366,0.0007359965],"category_scores_gemma":[0.6835256,0.0002393008,0.001241031,0.001152077,0.0002908739,0.0004962566,0.001721309,0.0002532118,0.0004620591],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002637569,"about_ca_system_score_gemma":0.00003529192,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001075637,"about_ca_topic_score_gemma":0.0003327067,"domain_scores_codex":[0.8851428,0.08963238,0.01611051,0.00395832,0.004640639,0.0005153386],"domain_scores_gemma":[0.7321855,0.2402699,0.009515725,0.01275678,0.003853143,0.001418994],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"qualitative","study_design_scores_codex":[0.001693128,0.001878224,0.3874937,0.000259242,0.03111058,0.00002772935,0.1222907,0.00003163363,0.00006568444,0.001089037,0.4522654,0.001794981],"study_design_scores_gemma":[0.003379995,0.01234343,0.4426449,0.00004455004,0.02155379,0.000005762323,0.4517403,0.001269549,0.000006812019,0.01780252,0.04787775,0.001330691],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9435537,0.01132828,0.004095923,0.03081939,0.000270165,0.008770748,0.0009606556,0.00003912206,0.0001620177],"genre_scores_gemma":[0.9916359,0.0001294315,0.003873181,0.003302464,0.0001325067,0.0002676658,0.00001746659,0.00001086931,0.0006305315],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.4043877,"threshold_uncertainty_score":0.9992941,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_narrow","metaepi_broad","metaresearch"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2889636695","doi":"10.1002/lom3.10273","title":"Methodological review and meta‐analysis of dilution assays for estimates of virus‐ and grazer‐mediated phytoplankton mortality","year":2018,"lang":"en","type":"article","venue":"Limnology and Oceanography Methods","topic":"Microbial Community Ecology and Physiology","field":"Environmental Science","cited_by":22,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Toronto","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Dilution; Virus; Mortality rate; Biology; Meta-analysis; Demography; Virology; Medicine; Internal medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.187349636260773,"gpt":0.4322050671732794,"spread":0.2448554309125064,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003211567,0.0001628883,0.001235987,0.0001375325,0.0001629626,0.000002502991,0.0001335419,0.0002536047,0.0005505969],"category_scores_gemma":[0.0005293211,0.0001218994,0.0002348028,0.0005332879,0.002516757,0.00006545908,0.0002025424,0.0001534269,4.554425e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000002786895,"about_ca_system_score_gemma":0.000003577892,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0000848128,"about_ca_topic_score_gemma":0.0002864313,"domain_scores_codex":[0.9970608,0.002049801,0.0003510225,0.0003065167,0.00003854114,0.0001932677],"domain_scores_gemma":[0.996415,0.002995816,0.0002563407,0.0002462984,0.00003072669,0.00005585507],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0006625596,0.0005676963,0.6547018,0.001852018,0.09037374,0.000001696626,0.001172669,0.00001593214,0.2361429,0.00656048,0.0008656117,0.007082851],"study_design_scores_gemma":[0.0003811535,0.001178199,0.8000473,0.00001075379,0.1440639,0.00001108788,0.000040369,0.000444212,0.01827098,0.03410996,0.001158871,0.000283179],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9419317,0.022467,0.03428797,0.0003351273,0.00005705145,0.0004648515,0.0001837716,0.00002723176,0.000245276],"genre_scores_gemma":[0.7052842,0.01178743,0.281606,0.001162812,0.000009486039,0.00003520779,0.00009859611,0.000008511189,0.00000777199],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.247318,"threshold_uncertainty_score":0.9273095,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2164404559","doi":"10.1136/bmjopen-2014-005297","title":"Interpreting trial results following use of different intention-to-treat approaches for preventing attrition bias: a meta-epidemiological study protocol","year":2014,"lang":"en","type":"article","venue":"BMJ Open","topic":"Advanced Causal Inference Techniques","field":"Mathematics","cited_by":21,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University","funders":"Parker Institute for Cancer Immunotherapy; Oak Foundation","keywords":"Medicine; Protocol (science); Attrition; Epidemiology; Selection bias; Publication bias; Alternative medicine; Meta-analysis; Environmental health; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.8468178363736624,"gpt":0.5865171235954936,"spread":0.2603007127781688,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.008661238,0.0003238663,0.001413947,0.0001134771,0.0001121315,0.0001432549,0.0005623759,0.000115618,0.00001171537],"category_scores_gemma":[0.04870853,0.0002068602,0.0005423679,0.0001438328,0.00002969379,0.0004133851,0.0008269837,0.0001441723,0.000001322192],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00006847153,"about_ca_system_score_gemma":0.00002005343,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00005373886,"about_ca_topic_score_gemma":0.00007925493,"domain_scores_codex":[0.9957054,0.001476373,0.001665447,0.0005614203,0.0002812899,0.0003100508],"domain_scores_gemma":[0.9922353,0.005903223,0.0009811192,0.0006703017,0.000130256,0.00007979857],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"randomized_trial","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.5567439,0.0423352,0.03909899,0.01047575,0.01190302,0.00003555328,0.01893378,0.0004348588,0.01094541,0.1707581,0.01814938,0.1201861],"study_design_scores_gemma":[0.1456308,0.03836998,0.001761612,0.007716423,0.005774295,0.00001216587,0.004884338,0.01824298,0.03085453,0.7425047,0.001392648,0.002855515],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"protocol","genre_gemma":"protocol","genre_scores_codex":[0.1302695,3.398663e-7,0.2869096,0.000136371,0.00005070685,0.5819992,0.00001915477,0.000161063,0.0004540131],"genre_scores_gemma":[0.3364595,3.565352e-8,0.1515738,0.00002903171,0.00003961213,0.5116646,0.00001178815,0.00002645822,0.0001952771],"genre_candidate":"protocol","genre_consensus":"protocol","teacher_disagreement_score":0.5717466,"threshold_uncertainty_score":0.9593046,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"protocol","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":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3048211970","doi":"10.1016/j.jclinepi.2020.08.002","title":"Abbreviated and comprehensive literature searches led to identical or very similar effect estimates: a meta-epidemiological study","year":2020,"lang":"en","type":"review","venue":"Journal of Clinical Epidemiology","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":20,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Institute of Health Economics; University of Alberta","funders":"","keywords":"Interquartile range; Meta-analysis; Confidence interval; Medicine; Odds ratio; MEDLINE; Epidemiology; Statistics; Systematic review; Internal medicine; Mathematics; Biology","retraction":null,"screen_n_in":null,"score":{"opus":0.9657825327774562,"gpt":0.7373159784020075,"spread":0.2284665543754487,"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","metaepi_narrow","metaepi_broad","research_integrity","insufficient_payload"],"category_scores_codex":[0.6571743,0.002284627,0.1811862,0.001158811,0.0001774346,0.0005645743,0.004992217,0.001843481,0.00283884],"category_scores_gemma":[0.9543487,0.0007077146,0.04348714,0.003222625,0.0005618383,0.000273481,0.001370078,0.005560776,0.00151281],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0000732152,"about_ca_system_score_gemma":0.0006131437,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000003907091,"about_ca_topic_score_gemma":0.000005131004,"domain_scores_codex":[0.3516843,0.5268719,0.1129122,0.00334431,0.004151956,0.001035326],"domain_scores_gemma":[0.1183691,0.81235,0.05981969,0.003720739,0.003190396,0.002550007],"domain_codex":"methods","domain_gemma":"methods","domain_candidate":"methods","domain_consensus":"methods","study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0007511529,0.0006272795,0.01135225,0.00647701,0.05140508,0.002639292,0.0001586679,0.00007999734,9.162844e-8,0.000334513,0.1296183,0.7965564],"study_design_scores_gemma":[0.0007525465,0.006232766,0.002774846,0.004219295,0.03344564,0.001366121,0.00006581525,0.0005991593,1.27554e-8,0.002857459,0.9471851,0.0005012195],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.0005916429,0.9838045,0.006010921,0.00456506,0.001246107,0.003678452,0.00005672118,0.00001132849,0.00003533173],"genre_scores_gemma":[0.0001447802,0.973561,0.02020552,0.004390755,0.001276568,0.0001158403,0.00001495736,0.00005458124,0.0002360217],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.8175668,"threshold_uncertainty_score":0.9995374,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4293483968","doi":"10.1136/bmjopen-2021-060237","title":"Factors Associated with the Magnitude Of acUpuncture treatment effectS (FAMOUS): a meta-epidemiological study of acupuncture randomised controlled trials","year":2022,"lang":"en","type":"article","venue":"BMJ Open","topic":"Acupuncture Treatment Research Studies","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa; Western University; York University; Hamilton Health Sciences; University of Toronto; McMaster University; Impact","funders":"China Academy of Chinese Medical Sciences; National Natural Science Foundation of China","keywords":"Medicine; Acupuncture; Meta-analysis; Meta-regression; Randomized controlled trial; MEDLINE; Sample size determination; Internal medicine; Clinical trial; Physical therapy; Alternative medicine; Pathology; Statistics","retraction":null,"screen_n_in":null,"score":{"opus":0.3686658038902662,"gpt":0.4846991083205923,"spread":0.1160333044303261,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","insufficient_payload"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.01170421,0.001366369,0.01657531,0.0001710512,0.0004394227,0.00003766343,0.0007169736,0.0001374269,0.0009792444],"category_scores_gemma":[0.02733517,0.000249672,0.002774457,0.0005514708,0.0002352378,0.00007089529,0.0008077302,0.0004723915,0.000001285591],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004119521,"about_ca_system_score_gemma":0.0003795646,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003942172,"about_ca_topic_score_gemma":0.0003618652,"domain_scores_codex":[0.9842198,0.01148234,0.001516573,0.0006424852,0.001608296,0.000530538],"domain_scores_gemma":[0.957311,0.0392055,0.001886229,0.001022574,0.0003990599,0.0001756793],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","study_design_scores_codex":[0.4226965,0.02002073,0.119564,0.0003316597,0.3962075,0.0006841345,0.01187811,0.001083106,0.003323519,0.00005883659,0.02341879,0.0007331725],"study_design_scores_gemma":[0.6197907,0.05193467,0.1870033,0.0001243404,0.1262649,0.00002180396,0.01086579,0.00004293053,0.002619791,0.00009239372,0.0007147355,0.0005246529],"study_design_candidate":"randomized_trial","study_design_consensus":"randomized_trial","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8683627,0.001105414,0.000009793547,0.003042048,0.0000761755,0.1263892,0.0001458558,0.00004716672,0.0008216298],"genre_scores_gemma":[0.9567671,0.00002382015,0.00006031606,0.0003626813,0.00005345043,0.03999095,0.0001555593,0.00004127572,0.002544864],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2699426,"threshold_uncertainty_score":0.9999955,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_narrow","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":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"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":"W3142062327","doi":"10.9778/cmajo.20200140","title":"Completeness of reporting for COVID-19 case reports, January to April 2020: a meta-epidemiologic study","year":2021,"lang":"en","type":"article","venue":"CMAJ Open","topic":"COVID-19 and healthcare impacts","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":true,"about_ca":false},"ca_institutions":"Queen's University; University of Toronto; St. Michael's Hospital","funders":"Amgen","keywords":"Coronavirus disease 2019 (COVID-19); Pandemic; Completeness (order theory); 2019-20 coronavirus outbreak; Medicine; Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); Coronavirus; Meta-analysis; MEDLINE; Disease; Infectious disease (medical specialty); Outbreak; Virology; Pathology; Political science; Mathematics","retraction":null,"screen_n_in":null,"score":{"opus":0.5234770417953649,"gpt":0.546135499860114,"spread":0.02265845806474909,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.008556453,0.0002502609,0.002133385,0.0001013245,0.0002458479,0.00004592417,0.0001635605,0.0001106151,0.0003760047],"category_scores_gemma":[0.06764556,0.0001913483,0.0003668159,0.0005115985,0.00004195537,0.0001009928,0.0005971259,0.0002041386,0.000006268917],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002655605,"about_ca_system_score_gemma":0.002947439,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.02494087,"about_ca_topic_score_gemma":0.007419407,"domain_scores_codex":[0.9953443,0.0003747258,0.002700856,0.000830357,0.0002868661,0.0004629129],"domain_scores_gemma":[0.9937241,0.001620377,0.001808339,0.001308138,0.0006271129,0.0009119546],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"not_applicable","study_design_scores_codex":[0.001244282,0.001492048,0.5867663,0.003579455,0.002352097,0.3191847,0.005315554,0.0002452922,0.002188545,0.0002148224,0.07266314,0.004753764],"study_design_scores_gemma":[0.01001814,0.007630518,0.3885523,0.000616833,0.009858015,0.1473893,0.02971311,0.0002699586,0.002776606,0.003814382,0.3978684,0.001492408],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9432207,0.001410708,0.003768477,0.04131962,0.0003825825,0.008769542,0.00008916399,0.00008161094,0.0009576473],"genre_scores_gemma":[0.962405,0.00001723882,0.006517008,0.02918504,0.0001694959,0.0006227753,0.00009046778,0.0000303216,0.0009626684],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3252053,"threshold_uncertainty_score":0.9815521,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4413408738","doi":"10.1177/01410768251366880","title":"Quantifying the impact of immortal time bias: empirical evidence from meta-analyses","year":2025,"lang":"en","type":"article","venue":"Journal of the Royal Society of Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":15,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Ottawa Hospital; Mental Health Research Canada; University of Ottawa","funders":"","keywords":"Interquartile range; Meta-analysis; Subgroup analysis; MEDLINE; Study heterogeneity; Medicine; Systematic review; Confidence interval; Publication bias; Random effects model; Statistics; Surgery; Internal medicine; Mathematics","retraction":null,"screen_n_in":null,"score":{"opus":0.9106758573085312,"gpt":0.6310800049927433,"spread":0.2795958523157879,"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":[0.1215139,0.0002840542,0.00609241,0.0001352999,0.0001515658,0.00009103699,0.003480765,0.00009586285,0.009851587],"category_scores_gemma":[0.06652995,0.00006824869,0.01577712,0.001696631,0.0005048004,0.000135135,0.0002416697,0.0004420233,0.00002807653],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005860204,"about_ca_system_score_gemma":0.0002689352,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0004944771,"about_ca_topic_score_gemma":0.00000961485,"domain_scores_codex":[0.9743255,0.008088527,0.01046474,0.0003303785,0.006604689,0.0001860955],"domain_scores_gemma":[0.9595501,0.02142123,0.01455113,0.002294593,0.00207315,0.0001097375],"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":[0.00004144627,0.000080212,0.07830391,0.00004869026,0.02827418,0.000001667253,0.002334488,0.009292421,0.002978736,0.000009579023,0.8768764,0.001758263],"study_design_scores_gemma":[0.001712779,0.0007801814,0.5605192,0.002389088,0.09706553,0.00003622042,0.009610614,0.3127747,0.001822175,0.006573475,0.006365812,0.0003502509],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8890441,0.05732874,0.011997,0.04011418,0.0004926226,0.0003969331,0.0000175616,0.000001798357,0.0006070748],"genre_scores_gemma":[0.9944589,0.0001417324,0.001619903,0.0005725486,0.0002184955,0.000001452576,4.108071e-7,0.000005744671,0.002980804],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.8705106,"threshold_uncertainty_score":0.9940796,"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":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2102903042","doi":"10.1186/2046-4053-2-88","title":"How should we evaluate the risk of bias of physical therapy trials?: a psychometric and meta-epidemiological approach towards developing guidelines for the design, conduct, and reporting of RCTs in Physical Therapy (PT) area: a study protocol","year":2013,"lang":"en","type":"article","venue":"Systematic Reviews","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":15,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"Alberta Innovates; Killam Trusts; Fondation pour la Recherche Médicale; University of Alberta; Canadian Institutes of Health Research; Physiotherapy Foundation of Canada; Alberta Innovates - Health Solutions; Women and Children's Health Research Institute; Children's Health Research Institute","keywords":"Medicine; Epidemiology; Randomized controlled trial; Meta-analysis; Physical therapy; Research design; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.9844898175982594,"gpt":0.6756415085022591,"spread":0.3088483090960003,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.6258188,0.0008596817,0.03121051,0.0004195696,0.0001442283,0.0003723549,0.001596969,0.0001061338,0.00004568441],"category_scores_gemma":[0.6230486,0.0001740346,0.004965774,0.002838066,0.0002445298,0.000231042,0.0001492538,0.0002420324,0.000003528145],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001932814,"about_ca_system_score_gemma":0.0001172961,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00006179273,"about_ca_topic_score_gemma":0.00000492366,"domain_scores_codex":[0.8063946,0.146532,0.04128274,0.001299398,0.004121628,0.0003695877],"domain_scores_gemma":[0.706642,0.1666053,0.1130716,0.00686688,0.006641236,0.0001729587],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0008441836,0.006052303,0.03355064,0.2129912,0.05720051,0.000002898533,0.05232253,0.0009858508,0.003334693,0.003805155,0.02544422,0.6034658],"study_design_scores_gemma":[0.01927337,0.01028265,0.03895077,0.01608657,0.05781135,0.0001692507,0.06447995,0.651675,0.00198802,0.1161909,0.01973835,0.003353886],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"protocol","genre_gemma":"protocol","genre_scores_codex":[0.1045594,0.03544134,0.09523479,0.002529718,0.00006434312,0.7621102,0.00001040934,0.000006338203,0.00004353897],"genre_scores_gemma":[0.442138,0.002224769,0.03057184,0.0001487101,0.0001124735,0.5246442,8.874256e-7,0.00003140551,0.0001277589],"genre_candidate":"protocol","genre_consensus":"protocol","teacher_disagreement_score":0.6506891,"threshold_uncertainty_score":0.9831722,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4404923291","doi":"10.1186/s12916-024-03778-1","title":"Integration of non-randomized studies with randomized controlled trials in meta-analyses of clinical studies: a meta-epidemiological study on effect estimation of interventions","year":2024,"lang":"en","type":"review","venue":"BMC Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","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; St. Joseph’s Healthcare Hamilton; Impact","funders":"West China Hospital, Sichuan University; Sichuan Provincial Administration of Traditional Chinese Medicine; National Science Fund for Distinguished Young Scholars; Sichuan University; National Natural Science Foundation of China","keywords":"Medicine; Randomized controlled trial; Meta-analysis; Systematic review; MEDLINE; Concordance; Psychological intervention; Internal medicine; Psychiatry","retraction":null,"screen_n_in":null,"score":{"opus":0.9831690436353597,"gpt":0.7817473869103255,"spread":0.2014216567250342,"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_narrow","metaepi_broad"],"category_scores_codex":[0.9116199,0.001717479,0.4426759,0.003587058,0.0000430413,0.00004907613,0.001492483,0.0003429127,0.001220269],"category_scores_gemma":[0.963791,0.0003585033,0.1321228,0.003518847,0.00181706,0.0001021724,0.0002142423,0.0007100983,0.00004587861],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005602309,"about_ca_system_score_gemma":0.0002612046,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00005569394,"about_ca_topic_score_gemma":0.0001122422,"domain_scores_codex":[0.1177663,0.6848536,0.1862911,0.002453255,0.00831945,0.0003163455],"domain_scores_gemma":[0.03587907,0.8400344,0.1163552,0.003933676,0.003666256,0.0001314959],"domain_codex":"methods","domain_gemma":"methods","domain_candidate":"methods","domain_consensus":"methods","study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","study_design_scores_codex":[0.1218288,0.0007427845,0.00001173623,0.03124613,0.818803,0.0000078472,0.001075773,0.0003226587,3.948126e-7,0.0004972322,0.002194961,0.02326868],"study_design_scores_gemma":[0.2826625,0.001542609,0.00000281825,0.02165341,0.6887743,0.000002129683,0.001825359,0.00209797,5.237453e-7,0.001122058,0.0001575611,0.0001587795],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.0004492085,0.9335089,0.0307881,0.0001416531,0.0005605518,0.03436271,0.00003249281,0.000006254421,0.0001500725],"genre_scores_gemma":[0.01309245,0.9673918,0.005331669,0.00002449246,0.0001404666,0.01293486,0.00003765713,0.0000326177,0.001013922],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.3105531,"threshold_uncertainty_score":0.9998867,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","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":"high"}],"label_agreement":"agree"},{"id":"W4389687393","doi":"10.23736/s1973-9087.23.08310-7","title":"The influence of bias in randomized controlled trials on rehabilitation intervention effect estimates: what we have learned from meta-epidemiological studies","year":2023,"lang":"en","type":"article","venue":"European Journal of Physical and Rehabilitation Medicine","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":13,"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":"Ministero della Salute","keywords":"Blinding; Rehabilitation; Randomized controlled trial; Medicine; Meta-analysis; Publication bias; Physical therapy; Physical medicine and rehabilitation; Epidemiology; Intervention (counseling); Reporting bias; Selection bias; MEDLINE; Clinical psychology; Psychiatry; Surgery; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.4888137266518677,"gpt":0.5026068912095041,"spread":0.01379316455763635,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.2763181,0.0002239862,0.007261251,0.0004572154,0.00009966029,0.00003498687,0.0001663806,0.00003942669,0.00001835788],"category_scores_gemma":[0.5861516,0.0001093087,0.001187505,0.000232294,0.0007810596,0.0003889727,0.00003423317,0.0002679585,0.00004368811],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00009108092,"about_ca_system_score_gemma":0.00001808662,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00005688978,"about_ca_topic_score_gemma":0.000007919981,"domain_scores_codex":[0.9593482,0.03001408,0.00987702,0.0003309058,0.0002399418,0.0001898553],"domain_scores_gemma":[0.664882,0.3284986,0.006132776,0.0001865473,0.0002207918,0.00007924246],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.2578515,0.002146413,0.03015062,0.006569513,0.0284825,0.00003708972,0.2988199,0.1206051,0.0009844317,0.1599248,0.02009472,0.0743334],"study_design_scores_gemma":[0.2749509,0.01428898,0.1279911,0.008159671,0.001444002,0.000002638199,0.03740158,0.01963064,0.00001177259,0.514631,0.001100597,0.0003871261],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8686016,0.01330634,0.0001751897,0.1163685,0.0003220176,0.001165815,0.000008188365,0.00001208093,0.00004028303],"genre_scores_gemma":[0.9946722,0.0039491,0.0004513761,0.0004596183,0.0003250103,0.00008655893,0.00000588585,0.00001443086,0.00003585493],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3547062,"threshold_uncertainty_score":0.7451828,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"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":"agree"},{"id":"W4392947957","doi":"10.1371/journal.pone.0298526","title":"Do statistical heterogeneity methods impact the results of meta- analyses? A meta epidemiological study","year":2024,"lang":"en","type":"article","venue":"PLoS ONE","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":10,"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":"Meta-analysis; Random effects model; Statistics; Confidence interval; Study heterogeneity; Pooling; Medicine; Restricted maximum likelihood; Interquartile range; Mathematics; Sample size determination; Maximum likelihood; Internal medicine; Computer science; Artificial intelligence","retraction":null,"screen_n_in":null,"score":{"opus":0.9732907961212317,"gpt":0.7002165803757118,"spread":0.2730742157455199,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.3911566,0.0004932888,0.01248794,0.0003362551,0.0001175624,0.0007191339,0.002120336,0.00009560939,0.01507089],"category_scores_gemma":[0.2280201,0.0001281643,0.007486404,0.00235438,0.0001702354,0.0001497275,0.0003386188,0.0003621902,0.0008332206],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002954422,"about_ca_system_score_gemma":0.00006461768,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001049215,"about_ca_topic_score_gemma":0.0000514527,"domain_scores_codex":[0.8029982,0.1635684,0.01959282,0.002622323,0.01061759,0.000600726],"domain_scores_gemma":[0.8826216,0.1057133,0.003455351,0.006709646,0.001201373,0.0002987238],"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":[0.0002176372,0.00871711,0.0147417,0.0003257605,0.9072847,0.0001120907,0.001835107,0.0008168663,0.004244561,0.005723354,0.04495695,0.01102419],"study_design_scores_gemma":[0.0006927595,0.002383409,0.05280069,0.00007469857,0.7296615,0.00002339247,0.001178412,0.1406931,0.002940054,0.06205536,0.006654484,0.0008421561],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7365089,0.05355693,0.1978735,0.002861503,0.0001657592,0.004153626,0.001489591,0.00004486915,0.003345331],"genre_scores_gemma":[0.8979655,0.00005287331,0.1006577,0.0001015087,0.0000600315,0.0001687712,0.000009448489,0.00001256468,0.000971652],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1776232,"threshold_uncertainty_score":0.9999447,"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":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2767159407","doi":"10.1136/bmjopen-2017-018856","title":"Association between statistical significance and time to publication among systematic reviews: a study protocol for a meta-epidemiological investigation","year":2017,"lang":"en","type":"article","venue":"BMJ Open","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":10,"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":"Japan Society for the Promotion of Science","keywords":"Medicine; Publication bias; Protocol (science); Family medicine; Clinical trial; Systematic review; Meta-analysis; Epidemiology; Publication; Peer review; Psychological intervention; MEDLINE; Medical education; Alternative medicine; Pathology; Psychiatry","retraction":null,"screen_n_in":null,"score":{"opus":0.9313157082465333,"gpt":0.6675016834766426,"spread":0.2638140247698907,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","scholarly_communication","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.4374733,0.0003279986,0.007554243,0.0001437651,0.0005588166,0.006006367,0.003107501,0.0001105086,0.0028724],"category_scores_gemma":[0.6221032,0.0001401375,0.0006667792,0.0003478492,0.00005453124,0.0009712196,0.000613615,0.0001046884,0.002754623],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00008978706,"about_ca_system_score_gemma":0.00008086727,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00006146828,"about_ca_topic_score_gemma":0.00005321489,"domain_scores_codex":[0.9462198,0.03615031,0.0126338,0.001480057,0.003195172,0.000320874],"domain_scores_gemma":[0.9526178,0.02080832,0.01953631,0.004754412,0.00192688,0.0003562405],"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":[0.00003896914,0.0001203851,0.4307758,0.004538728,0.002199604,9.023231e-7,0.0005080431,0.000004691786,0.00001025913,0.001469859,0.5579295,0.00240334],"study_design_scores_gemma":[0.001533049,0.0007942432,0.8873501,0.001677333,0.008051565,0.000001916697,0.0002730306,0.008953924,0.00000679766,0.02316717,0.06747135,0.0007195033],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"protocol","genre_gemma":"protocol","genre_scores_codex":[0.001056606,0.000007639369,0.02433292,0.004663386,0.00002228043,0.9677949,0.00005783667,0.000007191483,0.002057258],"genre_scores_gemma":[0.005251097,1.54582e-7,0.01581944,0.0003433247,0.00008147814,0.9580355,0.00001432568,0.00001053269,0.0204441],"genre_candidate":"protocol","genre_consensus":"protocol","teacher_disagreement_score":0.4904581,"threshold_uncertainty_score":0.9980391,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"evaluation","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_narrow","metaepi_broad","metaresearch"],"domain":"evaluation","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1998831987","doi":"10.1136/bmjopen-2014-005491","title":"Assessing bias in osteoarthritis trials included in Cochrane reviews: protocol for a meta-epidemiological study","year":2014,"lang":"en","type":"article","venue":"BMJ Open","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":9,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa; Ottawa Hospital; Cochrane","funders":"National Health and Medical Research Council; University of Ottawa; Medical Research Council; Parker Institute for Cancer Immunotherapy; Canadian Institutes of Health Research; Oak Foundation","keywords":"Medicine; Epidemiology; Protocol (science); Meta-analysis; Osteoarthritis; Alternative medicine; Cochrane collaboration; Systematic review; MEDLINE; Physical therapy; Family medicine; Internal medicine; Pathology; Cochrane Library","retraction":null,"screen_n_in":null,"score":{"opus":0.9811986829709822,"gpt":0.7555210505783807,"spread":0.2256776323926015,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_broad","scholarly_communication","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.8568103,0.0004690507,0.02350057,0.0004497837,0.0001274764,0.002633702,0.00364518,0.0001376692,0.007564977],"category_scores_gemma":[0.718495,0.0001793267,0.003000118,0.001705036,0.00004249064,0.0007665585,0.0008315811,0.0002059979,0.0006942448],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0000432807,"about_ca_system_score_gemma":0.0001216488,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001121491,"about_ca_topic_score_gemma":0.001101248,"domain_scores_codex":[0.5092608,0.4163206,0.06432213,0.003283454,0.0059475,0.0008655379],"domain_scores_gemma":[0.865464,0.1036536,0.02260533,0.007075701,0.0009339904,0.0002673365],"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":[0.001011915,0.003013588,0.2468209,0.001722025,0.001440612,0.00006476354,0.002528449,0.0002979147,0.0001116331,0.003830029,0.1849426,0.5542156],"study_design_scores_gemma":[0.01616513,0.001791629,0.05224356,0.002409727,0.001628391,0.00002024707,0.002396625,0.006749896,0.00002750835,0.03204161,0.8835153,0.001010373],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"protocol","genre_gemma":"protocol","genre_scores_codex":[0.00481998,0.0002389941,0.007371284,0.001461725,0.00009130582,0.979772,0.000007672538,0.000004567497,0.006232471],"genre_scores_gemma":[0.01991165,0.000002657076,0.01762311,0.0007686071,0.0001006615,0.9602454,0.000004173326,0.00001297857,0.001330811],"genre_candidate":"protocol","genre_consensus":"protocol","teacher_disagreement_score":0.6985727,"threshold_uncertainty_score":0.9984016,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"design_other","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4385287363","doi":"10.1136/bmjopen-2023-073232","title":"Evaluating the impact of including non-randomised studies of interventions in meta-analysis of randomised controlled trials: a protocol for a meta-epidemiological study","year":2023,"lang":"en","type":"article","venue":"BMJ Open","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":9,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"St. Joseph’s Healthcare Hamilton; McMaster University; Impact","funders":"National Science Fund for Distinguished Young Scholars; Canadian Institutes of Health Research; National Natural Science Foundation of China","keywords":"Medicine; Meta-analysis; Randomized controlled trial; Concordance; Systematic review; Psychological intervention; MEDLINE; Publication bias; Population; Epidemiology; Meta-regression; Internal medicine; Environmental health; Psychiatry","retraction":null,"screen_n_in":null,"score":{"opus":0.9906337942003213,"gpt":0.8021268322848917,"spread":0.1885069619154296,"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":[0.8671608,0.0004754017,0.06762571,0.001491658,0.0001493974,0.0002238514,0.003193431,0.00007207707,0.00594781],"category_scores_gemma":[0.7753061,0.0001263651,0.07732247,0.006045366,0.0001257156,0.0001920618,0.0009487888,0.0001179377,0.00001206044],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00003479394,"about_ca_system_score_gemma":0.0002033592,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000246719,"about_ca_topic_score_gemma":0.0002805329,"domain_scores_codex":[0.4995381,0.3829458,0.105855,0.002235148,0.008728497,0.0006974823],"domain_scores_gemma":[0.3763623,0.53663,0.07500934,0.006060208,0.005812945,0.0001252717],"domain_codex":null,"domain_gemma":"methods","domain_candidate":"methods","domain_consensus":null,"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","study_design_scores_codex":[0.0313667,0.0006140629,0.002712646,0.0002599083,0.9270189,8.975695e-7,0.002865176,0.03174194,0.0001805298,0.00005785241,0.002826178,0.0003551942],"study_design_scores_gemma":[0.1532123,0.001157699,0.003649245,0.00007701236,0.6824954,2.578858e-7,0.006997047,0.1488452,0.00002096849,0.003393176,0.000007399385,0.0001442645],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","genre_codex":"protocol","genre_gemma":"protocol","genre_scores_codex":[0.07209261,0.0003353595,0.003472593,0.0003831836,0.00002675472,0.9232632,0.0001001453,0.000003026766,0.0003231299],"genre_scores_gemma":[0.3352576,0.00000133444,0.001591951,0.00001161208,0.000009405894,0.6624733,0.000003217794,0.000006750723,0.0006448408],"genre_candidate":"protocol","genre_consensus":"protocol","teacher_disagreement_score":0.263165,"threshold_uncertainty_score":0.9949609,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2001201455","doi":"10.1136/bmjopen-2013-004527","title":"Methodological characteristics and treatment effect sizes in oral health randomised controlled trials: Is there a relationship? Protocol for a meta-epidemiological study","year":2014,"lang":"en","type":"article","venue":"BMJ Open","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":9,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"Canadian Institutes of Health Research; Alberta Innovates; Killam Trusts; University of Alberta; Alberta Innovates - Health Solutions; Women and Children's Health Research Institute; Children's Health Research Institute; Fondation pour la Recherche Médicale; Alpha Omega Foundation","keywords":"Medicine; Sample size determination; Protocol (science); Randomized controlled trial; Meta-analysis; Epidemiology; Intervention (counseling); Research design; MEDLINE; Clinical trial; Placebo; Alternative medicine; Physical therapy; Statistics; Psychiatry; Internal medicine; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.9742725078857929,"gpt":0.7329682702713417,"spread":0.2413042376144512,"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":[0.800721,0.0005041721,0.03224141,0.0001784782,0.0002122437,0.0009455589,0.001072096,0.0001329654,0.003254321],"category_scores_gemma":[0.722715,0.0001400415,0.003930789,0.0003939734,0.00004889585,0.0001338169,0.0001999074,0.0001242647,0.00008213591],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004842288,"about_ca_system_score_gemma":0.0001087614,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00007471153,"about_ca_topic_score_gemma":0.0001269565,"domain_scores_codex":[0.4678264,0.507021,0.02136775,0.001574983,0.001797806,0.0004120804],"domain_scores_gemma":[0.3664968,0.6131425,0.01656716,0.003016924,0.0004626427,0.0003139584],"domain_codex":"methods","domain_gemma":"methods","domain_candidate":"methods","domain_consensus":"methods","study_design_codex":"observational","study_design_gemma":"randomized_trial","study_design_scores_codex":[0.1087847,0.00234135,0.6614413,0.000331321,0.01312758,0.00001180646,0.002403493,0.00004079699,0.000008195908,0.003527425,0.02640446,0.1815776],"study_design_scores_gemma":[0.4737465,0.01634541,0.3769048,0.0001917549,0.009576299,0.00001331841,0.0009250059,0.03257449,0.000005140173,0.02754301,0.06140215,0.0007721654],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"protocol","genre_gemma":"protocol","genre_scores_codex":[0.02170576,0.00006744426,0.01966978,0.003174828,0.00004415247,0.9547039,0.00002221753,0.000006136696,0.0006058107],"genre_scores_gemma":[0.07789339,0.000001657778,0.02048736,0.0005873015,0.00005982403,0.8994638,0.000004192081,0.00000897566,0.00149344],"genre_candidate":"protocol","genre_consensus":"protocol","teacher_disagreement_score":0.3649618,"threshold_uncertainty_score":0.9976568,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"protocol","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":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4403591944","doi":"10.1016/j.atherosclerosis.2024.118624","title":"Publication bias in pharmacogenetics of statin-associated muscle symptoms: A meta-epidemiological study","year":2024,"lang":"en","type":"review","venue":"Atherosclerosis","topic":"Lipoproteins and Cardiovascular Health","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"Association Nationale de la Recherche et de la Technologie","keywords":"Pharmacogenetics; Medicine; Statin; Epidemiology; Meta-analysis; Hydroxymethylglutaryl-CoA Reductase Inhibitors; Publication bias; Internal medicine; Pharmacology; Genetics; Genotype; Biology","retraction":null,"screen_n_in":null,"score":{"opus":0.3292323609253614,"gpt":0.4284963298699932,"spread":0.0992639689446318,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.004932845,0.0006147937,0.005419479,0.0003655099,0.00003891536,0.00003431655,0.0002634045,0.0004387505,0.0001715101],"category_scores_gemma":[0.001125721,0.0003674162,0.002637188,0.001459265,0.00006915321,0.00004255593,0.0001491897,0.0007452077,0.00007515602],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003293108,"about_ca_system_score_gemma":0.0004064929,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0005224644,"about_ca_topic_score_gemma":0.00004527688,"domain_scores_codex":[0.9945064,0.001632477,0.001937684,0.0007807981,0.000698766,0.0004438571],"domain_scores_gemma":[0.9976245,0.0003505246,0.0007243029,0.000898396,0.0002150462,0.0001872843],"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":[0.00000835024,0.002000741,0.001540624,0.01232045,0.01520752,0.00001288395,0.0003744961,0.000001232512,0.000007282215,0.0000318169,0.0008631481,0.9676315],"study_design_scores_gemma":[0.002634809,0.001919471,0.02248095,0.01360255,0.1249716,0.0000167447,0.0002052603,0.00005645525,0.00000636995,0.00004176799,0.8331925,0.0008714318],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.01311238,0.9821911,0.00001110144,0.0000875447,0.0001480537,0.004044984,0.000182017,0.0001034575,0.0001193298],"genre_scores_gemma":[0.01223786,0.985398,0.0001320261,0.0001519966,0.0001176144,0.00146733,0.0001981546,0.0001119739,0.0001850276],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.96676,"threshold_uncertainty_score":0.9998778,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"systematic_review","genre":"review","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":"high"}],"label_agreement":"split"},{"id":"W4409203502","doi":"10.1016/j.drugalcdep.2025.112665","title":"Covariates in studies examining longitudinal relationships between substance use and mental health problems among youth: A meta-epidemiologic review","year":2025,"lang":"en","type":"review","venue":"Drug and Alcohol Dependence","topic":"Substance Abuse Treatment and Outcomes","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Centre Hospitalier Universitaire Sainte-Justine; McMaster University; Université de Montréal; St. Joseph’s Healthcare Hamilton","funders":"Canadian Institutes of Health Research","keywords":"Mental health; Substance use; Meta-analysis; Psychology; Longitudinal study; Covariate; Clinical psychology; Psychiatry; Medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.517409130212774,"gpt":0.4333312985156157,"spread":0.08407783169715832,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.003045928,0.0007154206,0.005472331,0.0003040149,0.0002713489,0.00004262281,0.0001437927,0.0001971891,0.000009022882],"category_scores_gemma":[0.001191908,0.000456415,0.0003281529,0.0006635414,0.000218541,0.000351637,0.0001048,0.0008356447,0.000004310232],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002095957,"about_ca_system_score_gemma":0.000180311,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001340635,"about_ca_topic_score_gemma":0.001252184,"domain_scores_codex":[0.9958615,0.0009961858,0.001391504,0.0009719021,0.0002919152,0.0004869819],"domain_scores_gemma":[0.9949621,0.003633301,0.000723925,0.0004171234,0.00007677677,0.0001867851],"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":[0.000009862072,0.00004700393,0.8996386,0.05444982,0.002913631,0.00006838176,0.002600343,3.885508e-7,1.905805e-8,0.00009023936,0.0003182233,0.03986352],"study_design_scores_gemma":[0.002644672,0.0002221059,0.1891973,0.7589585,0.03388461,0.0001549071,0.003059888,0.000004483285,5.21929e-7,0.0002857238,0.01019035,0.001396899],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.007791575,0.9889402,0.000007712839,0.0003453615,0.00008705921,0.002659769,0.00007341758,0.00006863113,0.00002622085],"genre_scores_gemma":[0.01004304,0.9878708,0.0005537714,0.0001576341,0.00004041882,0.0003439849,0.0003540693,0.00002908105,0.0006072695],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.7104412,"threshold_uncertainty_score":0.9997888,"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":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3203021509","doi":"10.1016/j.ocarto.2021.100215","title":"The state of trial registrations in the field of Orthopaedics in years 2015–2020. A meta-epidemiological study","year":2021,"lang":"en","type":"review","venue":"Osteoarthritis and Cartilage Open","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":6,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Western University; Hand and Upper Limb Clinic; St. Joseph's Hospital; University of Toronto; Queen's University; Canadian Institute for Military and Veteran Health Research","funders":"","keywords":"Medicine; Randomized controlled trial; Epidemiology; Population; Meta-analysis; Orthopedic surgery; Trial registration; MEDLINE; Prospective cohort study; Surgery; Internal medicine; Environmental health","retraction":null,"screen_n_in":null,"score":{"opus":0.6498741816734732,"gpt":0.5627773797430697,"spread":0.08709680193040348,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.1530054,0.000330561,0.01057177,0.0001267655,0.00008622296,0.0007553137,0.002692266,0.0001272312,0.000775272],"category_scores_gemma":[0.0380061,0.0001157483,0.002647586,0.001549477,0.0001037704,0.0001227763,0.000526346,0.0003843982,0.00001679874],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000008438367,"about_ca_system_score_gemma":0.0002270775,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001105895,"about_ca_topic_score_gemma":0.008588145,"domain_scores_codex":[0.9359346,0.04649558,0.01392765,0.0009824472,0.002388206,0.0002715149],"domain_scores_gemma":[0.9712659,0.02087075,0.004793054,0.002750374,0.0002627044,0.00005720518],"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":[0.00006242393,0.0001601842,0.0001311093,0.0002076678,0.0002079248,0.0001027602,0.0006674199,0.000003100953,1.251571e-7,0.0003875298,0.01059599,0.9874738],"study_design_scores_gemma":[0.001794501,0.0006660359,0.00006903293,0.0008566565,0.001795533,0.00001816728,0.001910372,0.000003294516,1.30798e-7,0.0007640437,0.9919803,0.000141992],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.005001375,0.9876148,0.000007963356,0.0003249201,0.0001260628,0.005372757,0.00007533456,5.199424e-7,0.001476315],"genre_scores_gemma":[0.005577389,0.9918593,0.00008490477,0.00008787668,0.00003720619,0.0009162332,0.0000151309,0.000009072951,0.001412853],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.9873317,"threshold_uncertainty_score":0.9700972,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3039344762","doi":"10.1136/annrheumdis-2020-217237","title":"Population characteristics as important contextual factors in rheumatological trials: an exploratory meta-epidemiological study from an OMERACT Working Group","year":2020,"lang":"en","type":"review","venue":"Annals of the Rheumatic Diseases","topic":"Rheumatoid Arthritis Research and Therapies","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Ottawa Hospital; Bruyère; University of Ottawa","funders":"Odense Universitetshospital; Oak Foundation","keywords":"Medicine; Epidemiology; Population; Meta-analysis; Physical therapy; Family medicine; Internal medicine; Gerontology; Environmental health","retraction":null,"screen_n_in":null,"score":{"opus":0.4342632187980843,"gpt":0.4729137158865182,"spread":0.03865049708843388,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.004540863,0.0009866487,0.01192476,0.0002875377,0.0001736736,0.0001314891,0.0008617882,0.0003383947,0.0005076461],"category_scores_gemma":[0.02544659,0.0004829425,0.002515723,0.0004818514,0.0002589373,0.0003759673,0.0002744062,0.0008191894,0.00001876616],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005939816,"about_ca_system_score_gemma":0.0003426992,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000682298,"about_ca_topic_score_gemma":0.0001385897,"domain_scores_codex":[0.9810923,0.01189446,0.004445974,0.0009115578,0.001046662,0.0006090309],"domain_scores_gemma":[0.9886943,0.006441811,0.0027706,0.001175261,0.0001278637,0.0007901883],"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":[0.0008434941,0.005061229,0.08493754,0.00225309,0.008458403,0.0001733446,0.001381918,9.308982e-7,5.054006e-7,0.0003321359,0.0001594656,0.8963979],"study_design_scores_gemma":[0.008444112,0.01235018,0.718294,0.1104298,0.01631878,0.0001501361,0.01657665,0.0004328763,0.000004717983,0.01153205,0.1013298,0.004136932],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.3189391,0.6776414,0.000003931536,0.0001280766,0.0002201118,0.002507542,0.0004811339,0.00007483437,0.000003875802],"genre_scores_gemma":[0.3943312,0.6040716,0.00001718313,0.00009477978,0.00008273454,0.0002925438,0.001056393,0.00005169489,0.000001849645],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.892261,"threshold_uncertainty_score":0.9997622,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4379135156","doi":"10.1016/j.mayocpiqo.2023.04.010","title":"Inadequate Reporting of Cointerventions, Other Methodological Factors, and Treatment Estimates in Cardiovascular Trials: A Meta-Epidemiological Study","year":2023,"lang":"en","type":"article","venue":"Mayo Clinic Proceedings Innovations Quality & Outcomes","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":6,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Toronto; St. Michael's Hospital","funders":"Arts and Humanities Research Council; University of California, San Francisco; University of Bern; University of Toronto; Inselspital, Universitätsspital Bern; Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung; National Science Foundation","keywords":"Blinding; Medicine; Meta-analysis; Publication bias; Psychological intervention; Clinical trial; Odds ratio; MEDLINE; Confidence interval; Relative risk; Physical therapy; Internal medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.9823887722089711,"gpt":0.7142536049889211,"spread":0.26813516722005,"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":[0.6955549,0.0005639153,0.0198233,0.00117917,0.0001811776,0.0003156724,0.0008049853,0.0001982999,0.001366177],"category_scores_gemma":[0.8059064,0.0002232017,0.007222652,0.004904819,0.0001766378,0.000258474,0.0003129454,0.0002728556,0.00008845781],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00006975272,"about_ca_system_score_gemma":0.00005852457,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0004386644,"about_ca_topic_score_gemma":0.00007669386,"domain_scores_codex":[0.8958231,0.02709021,0.07028135,0.002263051,0.003983977,0.0005582916],"domain_scores_gemma":[0.5940554,0.2770793,0.1153204,0.006680341,0.006463333,0.0004012677],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00001227515,0.0003310172,0.982532,0.00005680834,0.006286945,0.00000321052,0.001122568,0.00001482173,0.00001647646,0.004423933,0.0005756465,0.004624349],"study_design_scores_gemma":[0.0006537297,0.0002295922,0.9603747,0.00002654955,0.002711083,0.000002912242,0.01022752,0.0003075387,0.00002384038,0.02345035,0.001783013,0.0002092186],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9917697,0.0004844495,0.003265051,0.001845255,0.000137541,0.002064429,0.00004403573,0.00004615335,0.000343367],"genre_scores_gemma":[0.9857632,0.00002874539,0.01228019,0.0001144216,0.00002735836,0.0004598708,0.00001473407,0.00001485449,0.001296614],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3017678,"threshold_uncertainty_score":0.9995467,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4390765507","doi":"10.1007/s11655-023-3617-0","title":"Factors and Their Impact on Treatment Effect of Acupuncture in Different Outcomes: A Meta-Regression of Acupuncture Randomized Controlled Trials","year":2024,"lang":"en","type":"review","venue":"Chinese Journal of Integrative Medicine","topic":"Acupuncture Treatment Research Studies","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Acupuncture; Medicine; Randomized controlled trial; Meta-analysis; Confidence interval; Strictly standardized mean difference; Physical therapy; Meta-regression; MEDLINE; Cochrane Library; Internal medicine; Alternative medicine; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.1108893715685012,"gpt":0.4947666850293225,"spread":0.3838773134608213,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaepi_narrow","metaepi_broad"],"category_scores_codex":[0.007940953,0.007900817,0.1722339,0.003021921,0.00004558393,0.00001951305,0.0003201257,0.0005794021,0.0002971027],"category_scores_gemma":[0.07548755,0.0005474354,0.03919699,0.0007791932,0.0008686429,0.00007796018,0.0001034966,0.001940069,8.17016e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001114789,"about_ca_system_score_gemma":0.0004977192,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00007031114,"about_ca_topic_score_gemma":0.00003039558,"domain_scores_codex":[0.9844813,0.007043555,0.006021263,0.0005782424,0.001438304,0.0004373271],"domain_scores_gemma":[0.8864808,0.1058908,0.005991019,0.0005667273,0.0006677807,0.0004029373],"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":[0.3032271,0.001460792,0.00293966,0.03038731,0.5701377,0.0006612977,0.009363643,0.000002933831,0.0001756745,0.000017282,0.001888146,0.07973842],"study_design_scores_gemma":[0.3964468,0.03242037,0.0006317933,0.1449391,0.4213583,0.00031211,0.0007458886,0.00001644967,0.0003570257,0.0002665216,0.00199248,0.0005132428],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.03867168,0.9481384,0.0000408235,0.0009111289,0.0005644429,0.01135064,0.0002197268,0.00001381882,0.00008936683],"genre_scores_gemma":[0.1193719,0.8791373,0.00001186861,0.00001905568,0.0004316201,0.0004244846,0.0001196652,0.00008502077,0.0003991417],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.1487795,"threshold_uncertainty_score":0.9996977,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_narrow","metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4377010956","doi":"10.1093/fampra/cmad052","title":"The quality of paediatric asthma guidelines: evidence underpinning diagnostic test recommendations from a meta-epidemiological study","year":2023,"lang":"en","type":"article","venue":"Family Practice","topic":"Asthma and respiratory diseases","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Hospital for Sick Children; University of Toronto","funders":"U.S. Department of Health and Human Services; National Institutes of Health; Canadian Institutes of Health Research; National Institute for Health and Care Research; Canadian Thoracic Society; National Institute for Health and Care Excellence","keywords":"Medicine; Asthma; Spirometry; Epidemiology; Family medicine; Test (biology); Diagnostic test; Pediatrics; MEDLINE; Primary care; Intensive care medicine; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.4121088092352826,"gpt":0.4964183921697496,"spread":0.08430958293446694,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.006676428,0.0002191436,0.0006789506,0.00008920669,0.0003387451,0.00003735376,0.0002461128,0.00008635356,0.00007176868],"category_scores_gemma":[0.4643736,0.0001207503,0.0003049862,0.0009306925,0.0001190429,0.0003205961,0.0001916572,0.0003530454,0.0002144946],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004499277,"about_ca_system_score_gemma":0.0002726684,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0006498115,"about_ca_topic_score_gemma":0.00003388775,"domain_scores_codex":[0.995338,0.002155445,0.001214824,0.000426024,0.0005779564,0.0002877281],"domain_scores_gemma":[0.7961112,0.2013999,0.0007823939,0.0007436235,0.0007741653,0.0001887263],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.001037098,0.003770191,0.6687768,0.0001479162,0.003523981,0.0003987883,0.001098318,0.0005120308,0.001209914,0.000317323,0.2340651,0.08514264],"study_design_scores_gemma":[0.0007430311,0.0008585611,0.8412705,0.0001090866,0.002777966,0.000004450318,0.0243563,0.000318533,0.000007932683,0.0002111304,0.1291636,0.0001788539],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8591828,0.04594523,0.001749276,0.08799274,0.0008295893,0.002418165,0.0002082976,0.0004799361,0.00119394],"genre_scores_gemma":[0.9859086,0.005633902,0.001884032,0.005564691,0.0004492987,0.0002514406,0.0000420838,0.00002415144,0.0002417656],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.4576972,"threshold_uncertainty_score":0.5401382,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"evaluation","study_design":"systematic_review","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"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":"W3134334065","doi":"10.1016/j.jebdp.2021.101544","title":"Influence of Sponsorship Bias on Treatment Effect Size Estimates in Randomized Trials of Oral Health Interventions: A Meta-epidemiological Study","year":2021,"lang":"en","type":"review","venue":"Journal of Evidence Based Dental Practice","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":4,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Toronto; St. Michael's Hospital; University of Alberta","funders":"Alberta Heritage Foundation for Medical Research; Alberta Innovates; Alberta Innovates - Health Solutions","keywords":"Meta-analysis; Publication bias; Randomized controlled trial; Medicine; Epidemiology; Clinical trial; Confidence interval; Sample size determination; Psychological intervention; Random effects model; Selection bias; Reporting bias; Relative risk; MEDLINE; Internal medicine; Statistics; Psychiatry; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.9725301236095575,"gpt":0.7303173162000859,"spread":0.2422128074094716,"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":[0.8537911,0.001168001,0.09789299,0.001829168,0.00007672135,0.0004104681,0.002381393,0.0002639194,0.002113875],"category_scores_gemma":[0.9787639,0.0003649835,0.04794583,0.003092394,0.0002159213,0.001403072,0.0001765974,0.0009305573,0.00009273147],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0005174481,"about_ca_system_score_gemma":0.00153589,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002299755,"about_ca_topic_score_gemma":0.00005319107,"domain_scores_codex":[0.2197435,0.6939877,0.07645977,0.001285676,0.008109398,0.0004139022],"domain_scores_gemma":[0.05071898,0.8060229,0.1380452,0.002466612,0.002446646,0.0002995839],"domain_codex":"methods","domain_gemma":"methods","domain_candidate":"methods","domain_consensus":"methods","study_design_codex":"design_other","study_design_gemma":"meta_analysis","study_design_scores_codex":[0.09401018,0.02173981,0.00626162,0.04900627,0.1171144,0.002266701,0.0006337338,0.01344856,0.000006148468,0.000427569,0.002387829,0.6926972],"study_design_scores_gemma":[0.0683477,0.03079771,0.001801256,0.3784227,0.4406719,0.0018678,0.001962197,0.0004912569,0.00002761783,0.0006568917,0.07363982,0.001313065],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.01271835,0.9787424,0.0006284066,0.0006047137,0.0002418726,0.007012586,0.00004090865,0.000001992175,0.000008741875],"genre_scores_gemma":[0.02299077,0.9669019,0.009312577,0.0002155323,0.0000513947,0.0003773709,0.000004171643,0.00002495205,0.0001213636],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.6913841,"threshold_uncertainty_score":0.9998802,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4318973081","doi":"10.1136/bmjebm-2022-112053","title":"Different meta-analysis methods can change judgements about imprecision of effect estimates: a meta-epidemiological study","year":2023,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":4,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Meta-analysis; Statistics; Random effects model; Mathematics; Medicine; Restricted maximum likelihood; Maximum likelihood; Econometrics; Internal medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.9820049695252123,"gpt":0.7362051535903333,"spread":0.245799815934879,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","open_science","insufficient_payload"],"consensus_categories":["metaresearch","metaepi_narrow","metaepi_broad","insufficient_payload"],"category_scores_codex":[0.6144974,0.003642757,0.1512528,0.006392405,0.0002824329,0.0003842319,0.008078281,0.0007782417,0.03555686],"category_scores_gemma":[0.536687,0.001118133,0.07427885,0.0168604,0.0004931893,0.0002986594,0.001113968,0.001148717,0.0009962902],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002750828,"about_ca_system_score_gemma":0.0003508826,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0006465082,"about_ca_topic_score_gemma":0.0002507571,"domain_scores_codex":[0.6700647,0.2484504,0.05377088,0.006069376,0.02036029,0.001284344],"domain_scores_gemma":[0.6266546,0.3138774,0.04061433,0.01524002,0.002658684,0.0009549544],"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":[0.00003525776,0.0003082722,0.0009351403,0.01305675,0.610729,0.00008756603,0.0001970974,0.00006026426,5.269549e-7,0.0000292776,0.01122006,0.3633407],"study_design_scores_gemma":[0.0003803852,0.001572929,0.0006170247,0.005560864,0.8966656,0.000006883195,0.0001104045,0.001438823,0.000001173906,0.000353586,0.09257792,0.0007144275],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00002647937,0.941686,0.03555174,0.001298298,0.0008834205,0.02025403,0.0001937537,0.00005173471,0.00005459706],"genre_scores_gemma":[0.0002460588,0.964655,0.01114032,0.0007026578,0.000682681,0.01978834,0.00023435,0.0001440458,0.002406525],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.3626263,"threshold_uncertainty_score":0.9997815,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3158629300","doi":"10.1186/s12893-021-01278-0","title":"Bias estimation in study design: a meta-epidemiological analysis of transcatheter versus surgical aortic valve replacement","year":2021,"lang":"en","type":"article","venue":"BMC Surgery","topic":"Advanced Causal Inference Techniques","field":"Mathematics","cited_by":3,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"St. Michael's Hospital; Women's College Hospital; University of Toronto; Institute for Work & Health; Western University","funders":"University of Toronto; Canadian Institutes of Health Research; University of Ottawa","keywords":"Medicine; Aortic valve replacement; Propensity score matching; Stenosis; Randomized controlled trial; Aortic valve stenosis; Surgery; Aortic valve; Internal medicine; Epidemiology; Cardiology; Meta-analysis","retraction":null,"screen_n_in":null,"score":{"opus":0.6141875518604735,"gpt":0.4766740758921051,"spread":0.1375134759683684,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005078647,0.0002435561,0.00171646,0.0004179604,0.00002600526,0.00001314942,0.0001053388,0.0001190696,0.0005053596],"category_scores_gemma":[0.007827071,0.0001831226,0.001549542,0.001320547,0.00005491301,0.0001097858,0.00004870683,0.0001566728,0.000003274589],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00009463671,"about_ca_system_score_gemma":0.00009923479,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002624987,"about_ca_topic_score_gemma":0.0001118249,"domain_scores_codex":[0.9958096,0.001870553,0.001233905,0.0004402717,0.0003620416,0.0002836028],"domain_scores_gemma":[0.973603,0.02527038,0.000347872,0.000589661,0.000125099,0.00006397584],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.003865922,0.01102683,0.8866314,0.0004165863,0.04319983,0.0002907133,0.003070212,0.0328311,0.0003920413,0.01203181,0.0009442208,0.0052993],"study_design_scores_gemma":[0.007409279,0.003181731,0.3450291,0.0004555208,0.1404983,0.00004649019,0.005217322,0.2148935,0.03061477,0.2484448,0.0002103746,0.003998787],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7083968,0.0001630727,0.2907833,0.00003388581,0.00005162584,0.0003895335,0.000005605732,0.000108183,0.00006800531],"genre_scores_gemma":[0.9282903,0.00003065562,0.07133145,0.00001545347,0.000007628333,0.0002622665,0.00001981469,0.00001559447,0.00002681401],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5416024,"threshold_uncertainty_score":0.93703,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2911968494","doi":"10.1136/bmjebm-2018-111057","title":"Use of modified intention-to-treat analysis in studies of direct oral anticoagulants and risk of selection bias: a systematic review","year":2019,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Atrial Fibrillation Management and Outcomes","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Centre Hospitalier de l’Université de Montréal","funders":"Fonds de Recherche du Québec - Santé","keywords":"Selection (genetic algorithm); Risk analysis (engineering); Selection bias; Medicine; Intensive care medicine; Psychology; Actuarial science; Computer science; Business; Artificial intelligence; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.5686784527537067,"gpt":0.5045638600117248,"spread":0.06411459274198184,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.004419893,0.0004326321,0.0116443,0.002261446,0.00001971659,0.000004655714,0.0001247467,0.0001436813,0.00003339507],"category_scores_gemma":[0.02744275,0.0002542412,0.0008559121,0.003703985,0.0001633839,0.0001261673,0.00006659931,0.0001594014,0.000003317134],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001298437,"about_ca_system_score_gemma":0.0001980014,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0008178712,"about_ca_topic_score_gemma":0.0001059575,"domain_scores_codex":[0.993974,0.001284781,0.003299204,0.0004456423,0.0008087247,0.0001876828],"domain_scores_gemma":[0.9904341,0.004906978,0.003218818,0.0006552861,0.0006900176,0.00009481291],"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":[0.0001501129,0.00002065706,0.05646099,0.9288368,0.00451411,0.000005588136,0.0000613212,0.000055253,0.000004993511,0.00001095807,0.0001694744,0.009709801],"study_design_scores_gemma":[0.0006563519,0.0005455387,0.002612767,0.8563666,0.1388536,0.000004195192,0.00006500849,0.0003246689,0.000008871146,0.000002305232,0.0003968511,0.0001632725],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.001771104,0.9899677,0.0002542753,0.0002827995,0.000103989,0.007567951,0.00002343657,0.00001712889,0.00001158511],"genre_scores_gemma":[0.00520394,0.993831,0.000313943,0.00005580629,0.00006044392,0.00005923559,0.00005837406,0.00002461369,0.0003926161],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.1343395,"threshold_uncertainty_score":0.999991,"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_narrow","metaepi_broad","metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4387222396","doi":"10.1007/s40279-023-01941-x","title":"Certainty of Evidence Assessment in Systematic Reviews Published by High-Impact Sports Science Journals: A Meta-epidemiological Study","year":2023,"lang":"en","type":"article","venue":"Sports Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":3,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact; University of Saskatchewan","funders":"","keywords":"Certainty; Systematic review; Grading (engineering); Meta-analysis; Publication bias; Quality of evidence; Medicine; Citation; Evidence-based medicine; Sports medicine; MEDLINE; Bibliometrics; Psychology; Alternative medicine; Computer science; Library science; Political science; Pathology; Mathematics; Engineering","retraction":null,"screen_n_in":null,"score":{"opus":0.9041380651400333,"gpt":0.6472378102149855,"spread":0.2569002549250478,"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":[0.7694244,0.0007368456,0.02229149,0.002555462,0.0001805604,0.000370945,0.003802448,0.0001502026,0.0252052],"category_scores_gemma":[0.4696264,0.0002205636,0.002551777,0.01498771,0.0005052719,0.0009255991,0.0003634101,0.0007274688,0.0001641498],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002140847,"about_ca_system_score_gemma":0.0003820141,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002958689,"about_ca_topic_score_gemma":0.00003734431,"domain_scores_codex":[0.8210157,0.05519967,0.07288496,0.004594807,0.04454222,0.001762692],"domain_scores_gemma":[0.9244934,0.02825088,0.03292613,0.009689182,0.003638601,0.00100181],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00002012417,0.0004272578,0.7546117,0.003539614,0.0008684395,0.0003288568,0.001598331,0.0004219797,0.0002632537,0.0003196708,0.2330334,0.004567423],"study_design_scores_gemma":[0.0005561693,0.0005275941,0.9660062,0.0116107,0.004371765,0.00005875682,0.006218756,0.003270346,0.000004259856,0.005136278,0.00183479,0.0004044038],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9399861,0.03163027,0.0018552,0.01115628,0.001274195,0.01255468,0.00001269712,0.00003766326,0.001492888],"genre_scores_gemma":[0.9953136,0.001364428,0.0004707185,0.0004039913,0.00008522618,0.0003657997,0.000006772526,0.00001423468,0.001975269],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.299798,"threshold_uncertainty_score":0.9922236,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4400747098","doi":"10.1111/dom.15771","title":"Bariatric surgery and all‐cause mortality: A methodological review of studies using a non‐surgical comparator","year":2024,"lang":"en","type":"review","venue":"Diabetes Obesity and Metabolism","topic":"Bariatric Surgery and Outcomes","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McGill University; Jewish General Hospital","funders":"Atara Biotherapeutics; Kowa Company; Amarin Corporation; Novo Nordisk; Merck; Novartis; Pfizer; Seqirus; AstraZeneca; National Institute of Diabetes and Digestive and Kidney Diseases; Boehringer Ingelheim","keywords":"Medicine; Surgical procedures; Surgery; General surgery; Intensive care medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.3688343445838422,"gpt":0.4663856746911559,"spread":0.0975513301073137,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.004993422,0.0007756712,0.01034572,0.0005332898,0.00009353479,0.00003601067,0.0001191032,0.0005100039,0.00003484881],"category_scores_gemma":[0.001667182,0.0004811863,0.00145869,0.001029769,0.0003141072,0.0001091209,0.0003811949,0.000747418,0.00001012785],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00003488439,"about_ca_system_score_gemma":0.0003337537,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001261602,"about_ca_topic_score_gemma":4.100053e-7,"domain_scores_codex":[0.9952779,0.00130019,0.001649799,0.0008183719,0.0004302893,0.0005234714],"domain_scores_gemma":[0.9942502,0.004130312,0.0005954749,0.0004972758,0.0001688618,0.0003579117],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00001444873,0.0002462951,0.00512268,0.8550358,0.01558073,0.0004589658,0.00008014446,2.582239e-8,0.000001833153,0.0006928751,0.002320412,0.1204457],"study_design_scores_gemma":[0.0003123978,0.0000290902,0.002399815,0.1104897,0.1253115,0.0003140708,0.00001824972,0.000008289128,0.000005136537,0.0001587119,0.7603327,0.0006203233],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.02285461,0.9749678,0.000003612865,0.0000850945,0.000922512,0.001006586,0.00007413575,0.00005298241,0.00003261117],"genre_scores_gemma":[0.0001365432,0.9979402,0.0007005308,0.0004135983,0.0005814693,0.00006282477,0.00005732454,0.00004856275,0.00005889306],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.7580123,"threshold_uncertainty_score":0.999764,"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":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1974429147","doi":"10.1136/annrheumdis-2014-eular.2210","title":"THU0416 Attrition Bias in Rheumatoid Arthritis Randomised Trials with Different Modified Intention-To-Treat Approaches: A Meta-Epidemiological Study","year":2014,"lang":"en","type":"article","venue":"Annals of the Rheumatic Diseases","topic":"Advanced Causal Inference Techniques","field":"Mathematics","cited_by":2,"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":"Medicine; Randomized controlled trial; Population; Rheumatoid arthritis; Meta-analysis; Internal medicine; Intention-to-treat analysis; Sample size determination; Physical therapy; Imputation (statistics); Study heterogeneity; Missing data; Statistics","retraction":null,"screen_n_in":null,"score":{"opus":0.6572351615281036,"gpt":0.4474743812337799,"spread":0.2097607802943237,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005061229,0.0005110107,0.003192602,0.0002206265,0.0001021804,0.00005184078,0.0005588681,0.000094773,0.00009092163],"category_scores_gemma":[0.03363305,0.0002422111,0.0008568122,0.0003531135,0.0001875214,0.0001815049,0.0001820125,0.0001852763,0.000004593654],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00003229902,"about_ca_system_score_gemma":0.00003143047,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00009714511,"about_ca_topic_score_gemma":0.0001006856,"domain_scores_codex":[0.9923261,0.004237101,0.00197975,0.0004704779,0.0005766143,0.0004099044],"domain_scores_gemma":[0.9866925,0.01083375,0.001112337,0.0009907261,0.0001768296,0.0001938524],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.04920784,0.06672943,0.09167966,0.006591593,0.02046179,0.00004156403,0.01211201,0.006114079,0.0004670206,0.4710855,0.01622865,0.2592809],"study_design_scores_gemma":[0.01584935,0.001538428,0.02320855,0.003472989,0.001248603,0.000008746874,0.001630544,0.002030072,0.002389935,0.9478989,0.000008056888,0.000715867],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9457786,0.0005577875,0.04773262,0.0009078414,0.00004002063,0.004547081,0.0000582392,0.0001901363,0.0001876633],"genre_scores_gemma":[0.9948238,0.000497969,0.002404711,0.0001434373,0.00001032301,0.002030643,0.00001331194,0.00003755099,0.00003829405],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.4768134,"threshold_uncertainty_score":0.987708,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4411977700","doi":"10.1186/s12916-025-04202-y","title":"Sex and gender considerations in randomized controlled trials in critical care nephrology: a meta-epidemiologic study","year":2025,"lang":"en","type":"review","venue":"BMC Medicine","topic":"Sex and Gender in Healthcare","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Alberta Health Services; University of Alberta","funders":"University of Alberta; King Chulalongkorn Memorial Hospital","keywords":"Medicine; Randomized controlled trial; Nephrology; Meta-analysis; Internal medicine; MEDLINE; Intensive care medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.6544556132467321,"gpt":0.5824175902738793,"spread":0.07203802297285278,"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":[0.05869656,0.0008731958,0.05746891,0.001592072,0.00008804581,0.00001374656,0.0001239484,0.0008359354,0.0006278006],"category_scores_gemma":[0.3637586,0.00042656,0.00205808,0.0005282473,0.0005294894,0.00002788085,0.00009170176,0.001284539,0.000004513462],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001678545,"about_ca_system_score_gemma":0.001980674,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003833111,"about_ca_topic_score_gemma":0.0008106688,"domain_scores_codex":[0.9465253,0.04347046,0.007713037,0.001261745,0.0004475199,0.0005819335],"domain_scores_gemma":[0.7879894,0.2100276,0.0007281433,0.0007527269,0.0002433999,0.000258676],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"randomized_trial","study_design_scores_codex":[0.1202965,0.002582846,0.01154896,0.5606242,0.0461942,0.00473503,0.03104202,0.00001115141,6.376209e-7,0.01676723,0.01033576,0.1958615],"study_design_scores_gemma":[0.7844065,0.001118963,0.0002762914,0.01471446,0.15384,0.000375365,0.01251586,0.00005355151,3.665616e-8,0.006810991,0.02529,0.0005979377],"study_design_candidate":"randomized_trial","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00004974037,0.9721523,0.0000907993,0.002546267,0.0005282955,0.02351429,0.00002849539,0.00005275002,0.001037061],"genre_scores_gemma":[0.004955468,0.9807958,0.0005864268,0.003728187,0.0005738863,0.00893557,0.0001481619,0.00003460273,0.0002418832],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.6641101,"threshold_uncertainty_score":0.9998186,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaepi_broad","metaepi_narrow"],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4405223579","doi":"10.1093/jnci/djae318","title":"Justification, margin values, and analysis populations for oncologic noninferiority and equivalence trials: a meta-epidemiological study","year":2024,"lang":"en","type":"article","venue":"JNCI Journal of the National Cancer Institute","topic":"Statistical Methods in Clinical Trials","field":"Mathematics","cited_by":2,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"London Health Sciences Centre; Western University","funders":"National Cancer Institute; University of Texas MD Anderson Cancer Center; National Institutes of Health; Andrew Sabin Family Foundation","keywords":"Margin (machine learning); Equivalence (formal languages); Clinical endpoint; Meta-analysis; Population; Medicine; Statistics; Clinical trial; Mathematics; Computer science; Internal medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.9378516184211011,"gpt":0.7058903326026658,"spread":0.2319612858184353,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.02820563,0.0001884506,0.001511802,0.0002118806,0.0002371892,0.0001092648,0.0002874849,0.0001236639,0.00009834002],"category_scores_gemma":[0.2323558,0.00009534643,0.0006980029,0.000645889,0.0002725653,0.0002125958,0.0001014332,0.0003731234,4.896104e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00022347,"about_ca_system_score_gemma":0.0003412136,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001962841,"about_ca_topic_score_gemma":0.0000568006,"domain_scores_codex":[0.9950238,0.00197544,0.001888106,0.00029956,0.0006642019,0.0001488787],"domain_scores_gemma":[0.948314,0.04958966,0.001008962,0.0001508355,0.0008272994,0.0001092522],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.001181417,0.001257664,0.0240052,0.0007528696,0.03747094,0.00002476162,0.0009520547,0.00583053,0.0003494145,0.859831,0.03505692,0.03328726],"study_design_scores_gemma":[0.000667743,0.0001662181,0.03853773,0.00005660756,0.02011378,0.0000188042,0.00003925055,0.004059095,0.00001452906,0.934044,0.002162891,0.0001193319],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.2069209,0.008233471,0.7380403,0.02862268,0.01182417,0.003979298,0.002065611,0.00008925046,0.0002243911],"genre_scores_gemma":[0.5067679,0.0003575229,0.4906334,0.0003395065,0.001399407,0.0002090844,0.000002128166,0.00001283254,0.0002782289],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.299847,"threshold_uncertainty_score":0.9775565,"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":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4323536717","doi":"10.1101/2023.03.05.23286821","title":"The impact of blinding on trial results: A systematic review and meta-analysis","year":2023,"lang":"en","type":"review","venue":"medRxiv","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Impact; Dalhousie University; McMaster University","funders":"","keywords":"Blinding; Meta-analysis; Publication bias; Medicine; Randomized controlled trial; Systematic review; MEDLINE; Odds ratio; Strictly standardized mean difference; Clinical trial; Internal medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.9072607679519703,"gpt":0.6315100895610637,"spread":0.2757506783909066,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","scholarly_communication","insufficient_payload"],"consensus_categories":["metaresearch","metaepi_broad","insufficient_payload"],"category_scores_codex":[0.4079105,0.001146153,0.07879013,0.001825148,0.0002756293,0.001193437,0.004781353,0.0002532954,0.001390986],"category_scores_gemma":[0.2663695,0.0002978107,0.06866115,0.01289315,0.000122824,0.0001002677,0.0003215356,0.0004808673,0.002618606],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0000754468,"about_ca_system_score_gemma":0.0002191475,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002011534,"about_ca_topic_score_gemma":0.00002141339,"domain_scores_codex":[0.8716412,0.07037715,0.04652056,0.002213722,0.008734155,0.0005132481],"domain_scores_gemma":[0.8595928,0.0776829,0.04714656,0.01405164,0.001235917,0.0002901946],"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":[0.00003719983,0.00003217062,0.000001289444,0.4581632,0.5294082,0.000006615125,0.00003436982,0.000004677062,2.700259e-9,0.0001823258,0.01025878,0.001871119],"study_design_scores_gemma":[0.0003861679,0.0001024364,9.821096e-7,0.02169204,0.9199297,0.000006893657,0.0000210852,0.00003822897,7.390223e-9,0.0001176615,0.05745814,0.0002466843],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.000003112523,0.9888146,0.00004385553,0.0001539282,0.0001095638,0.009787385,0.0004060131,0.000009392403,0.0006721775],"genre_scores_gemma":[0.00001846364,0.9828207,0.00003528037,0.00003537085,0.00005168383,0.001158367,0.00002921737,0.00004330849,0.01580757],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.4364712,"threshold_uncertainty_score":0.9999474,"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":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4413340902","doi":"10.3390/curroncol32080462","title":"High Outcome-Reporting Bias in Randomized-Controlled Trials of Acupuncture for Cancer Chemotherapy-Induced Nausea and Vomiting: A Systematic Review and Meta-Epidemiological Study","year":2025,"lang":"en","type":"review","venue":"Current Oncology","topic":"Acupuncture Treatment Research Studies","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":true,"about_ca":false},"ca_institutions":"","funders":"Universitätsspital Zürich","keywords":"Medicine; Nausea; Acupuncture; Meta-analysis; Vomiting; Clinical trial; Reporting bias; Randomized controlled trial; MEDLINE; Publication bias; Metric (unit); Outcome (game theory); Prospective cohort study; Systematic review; Epidemiology; Physical therapy; Alternative medicine; Intensive care medicine; Internal medicine; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.6576657626141557,"gpt":0.6228960202366118,"spread":0.03476974237754382,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch","metaepi_narrow"],"category_scores_codex":[0.06208055,0.001687631,0.1515891,0.0006184895,0.00008192757,0.00001741294,0.0002021927,0.0005484416,0.00005662751],"category_scores_gemma":[0.3746964,0.0004753836,0.007257084,0.0004019207,0.0002233979,0.00003234429,0.000232789,0.0009114305,4.367278e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004915235,"about_ca_system_score_gemma":0.0008301812,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00004651705,"about_ca_topic_score_gemma":0.00006860693,"domain_scores_codex":[0.9688949,0.01425573,0.01481225,0.001041331,0.0004266395,0.0005691505],"domain_scores_gemma":[0.856505,0.1229086,0.01911776,0.0006766421,0.0005895777,0.0002023774],"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":[0.006668246,0.0007099137,0.000463897,0.9137176,0.05041975,0.00003227592,0.0001040798,4.862188e-8,1.179092e-7,0.00003411032,0.0003141628,0.02753576],"study_design_scores_gemma":[0.1698683,0.0007957713,0.0000251758,0.2189875,0.6033857,0.00004504359,0.00007103305,0.00000328004,3.182262e-7,0.00005737913,0.006460439,0.0003001169],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.0002669762,0.8937379,0.00001799942,0.001615342,0.0002935999,0.1039424,0.00007365516,0.00003174922,0.0000204827],"genre_scores_gemma":[0.0002923983,0.9332625,0.0001947315,0.0002563405,0.0001398489,0.06552716,0.00006227545,0.00003575877,0.0002290006],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.6947302,"threshold_uncertainty_score":0.9997698,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"reporting","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2536368941","doi":"10.1093/ejo/cjw062","title":"Historical controls in orthodontics: need for larger meta-epidemiological studies","year":2016,"lang":"en","type":"letter","venue":"European Journal of Orthodontics","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"University of Alberta; Alberta Innovates; Killam Trusts; Alberta Innovates - Health Solutions; Women and Children's Health Research Institute; Children's Health Research Institute","keywords":"Orthodontics; Epidemiology; Medicine; Dentistry; Pathology","retraction":null,"screen_n_in":null,"score":{"opus":0.9015952426937508,"gpt":0.5451367677650194,"spread":0.3564584749287314,"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":[0.4444004,0.001137525,0.02259487,0.001346695,0.0001421854,0.0005284761,0.004144919,0.0004043831,0.001360409],"category_scores_gemma":[0.343845,0.0004119498,0.01563861,0.0006949145,0.0001714627,0.0002931282,0.0003149257,0.002216594,0.0005298979],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003692301,"about_ca_system_score_gemma":0.0001605824,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":8.277536e-7,"about_ca_topic_score_gemma":0.000003466247,"domain_scores_codex":[0.8065398,0.1397509,0.04328827,0.001838952,0.007316481,0.001265593],"domain_scores_gemma":[0.891587,0.06587945,0.03267287,0.003339142,0.006224659,0.0002968845],"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":[0.00004665121,0.00003425745,0.009838042,0.000112258,0.005647989,0.001290283,0.0001176411,0.00002439961,0.0000083439,0.0001343659,0.9786949,0.004050933],"study_design_scores_gemma":[0.0008734048,0.0003227107,0.0007974132,0.0001882694,0.006786525,0.0001597348,0.0000727794,0.00004244061,8.095986e-7,0.001256878,0.9890705,0.0004285316],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.0005886607,0.04051342,0.2715336,0.6789598,0.004457394,0.001363821,0.00007515552,0.000009713173,0.002498424],"genre_scores_gemma":[0.002722776,0.003503647,0.2620895,0.5234507,0.02852981,0.00007795727,0.00002037855,0.0004027121,0.1792025],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.1767041,"threshold_uncertainty_score":0.9998332,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W7117489679","doi":"10.1016/j.semarthrit.2025.152907","title":"Determining the best discriminatory physical functioning outcome measurement instrument for psoriatic arthritis trials: A meta-epidemiological study","year":2025,"lang":"en","type":"article","venue":"Seminars in Arthritis and Rheumatism","topic":"Spondyloarthritis Studies and Treatments","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Krembil Foundation; University Health Network; Ottawa Hospital; University of Ottawa","funders":"Parker Institute for Cancer Immunotherapy; Oak Foundation","keywords":"Psoriatic arthritis; Outcome (game theory); Patient-reported outcome; Arthritis; MEDLINE; Psychometrics","retraction":null,"screen_n_in":null,"score":{"opus":0.1628121631337334,"gpt":0.3739478527659262,"spread":0.2111356896321928,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.003802227,0.0004274634,0.002218178,0.0001555284,0.0005681948,0.00007179128,0.0001025861,0.00010127,0.00002519328],"category_scores_gemma":[0.003319144,0.0002542397,0.0004950442,0.0002017219,0.0001921758,0.0001188237,0.0001492721,0.0002937488,0.000004334134],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001606742,"about_ca_system_score_gemma":0.00005793797,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001177848,"about_ca_topic_score_gemma":0.0004732812,"domain_scores_codex":[0.9967291,0.0005267838,0.001254311,0.0005940383,0.0004286473,0.0004671466],"domain_scores_gemma":[0.9977535,0.001411615,0.0002604277,0.0003430334,0.0001282435,0.000103185],"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":[0.0001484851,0.00124727,0.1420935,0.00006722141,0.001308233,0.00003596766,0.001148389,0.000003054829,0.000009829358,0.001411141,0.0003313863,0.8521955],"study_design_scores_gemma":[0.04708352,0.007687189,0.8866857,0.02199315,0.005190059,0.0001195455,0.01958729,0.0007037446,0.00007752029,0.007999592,0.002059714,0.0008129459],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8743727,0.1197384,0.0002269681,0.001372304,0.0006948928,0.003119537,0.00003323477,0.00004139225,0.0004004875],"genre_scores_gemma":[0.9589418,0.03742121,0.0008575907,0.0003027263,0.00007121373,0.002255791,0.00001282862,0.00002148278,0.0001153682],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.8513826,"threshold_uncertainty_score":0.999991,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"observational","genre":"empirical","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":"W2319079108","doi":"10.1016/j.joca.2016.01.101","title":"Assessing bias in osteoarthritis trials included in cochrane reviews: A meta-epidemiological study","year":2016,"lang":"en","type":"article","venue":"Osteoarthritis and Cartilage","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Ottawa Hospital; Bruyère; University of Ottawa; Cochrane","funders":"","keywords":"Blinding; Meta-analysis; Medicine; Systematic review; Publication bias; Reporting bias; MEDLINE; Study heterogeneity; Physical therapy; Sample size determination; Randomized controlled trial; Clinical trial; Psychological intervention; Epidemiology; Random effects model; Cochrane Library; Internal medicine; Statistics; Psychiatry","retraction":null,"screen_n_in":null,"score":{"opus":0.8062375650609488,"gpt":0.5538883201150719,"spread":0.2523492449458768,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","scholarly_communication","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.3947822,0.000703865,0.01701028,0.0007231162,0.0001631228,0.001207926,0.001021651,0.0002276163,0.008281958],"category_scores_gemma":[0.209901,0.0002750319,0.002800036,0.001719583,0.000182512,0.0009532439,0.0003584167,0.0002856595,0.0007834708],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005343648,"about_ca_system_score_gemma":0.00005699377,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003213224,"about_ca_topic_score_gemma":0.003131215,"domain_scores_codex":[0.8619792,0.1045355,0.02617534,0.002550225,0.003880122,0.000879526],"domain_scores_gemma":[0.967254,0.02421107,0.005109867,0.002731353,0.0003802839,0.0003134035],"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":[0.00007194213,0.0004230778,0.1870836,0.00007761864,0.0001026569,0.0005484025,0.00132526,0.000003645303,0.01066364,0.0005589061,0.004702922,0.7944384],"study_design_scores_gemma":[0.05442749,0.009386268,0.324402,0.009667398,0.008267507,0.0007979345,0.0278609,0.0001291691,0.003651456,0.05225936,0.5027995,0.006351022],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9298599,0.06259112,0.000260249,0.001330947,0.0002358082,0.003096169,0.00003916762,0.00001117681,0.002575439],"genre_scores_gemma":[0.9947252,0.002134489,0.0005589177,0.000423915,0.0001038828,0.0007242805,0.000004344964,0.00002018483,0.00130481],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7880874,"threshold_uncertainty_score":0.9999945,"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":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"}]}