{"meta":{"query_hash":"5ba9182f63f0","filters":{"venue":"Journal of Evidence-Based Medicine"},"cohort_total":78,"direct_labels_cover":73,"predictions_cover":78,"exported":78,"export_cap":100000,"truncated":false,"label_status":"direct model label, unvalidated","prediction_status":"machine_predicted_unvalidated (Codex and Gemma teacher distillation)","score_status":"score_only:v0-immature-baseline","snapshot":{"source":"OpenAlex, pinned release, all 482 partitions","release":"2026-06-24","frame_built":"2026-07-12"},"permalink":"https://metacan.xera.ac/q/5ba9182f63f0","api":"https://metacan.xera.ac/api/v1/cohort?venue=Journal+of+Evidence-Based+Medicine"},"results":[{"id":"W1483618109","doi":"10.1111/jebm.12022","title":"Visualization of evidence‐based medicine domain knowledge: production and citation of cochrane systematic reviews","year":2013,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"China Medical Board","keywords":"Citation; Knowledge translation; Systematic review; Cochrane Library; Psychological intervention; Production (economics); MEDLINE; Bibliometrics; Medicine; Developing country; Business; Political science; Alternative medicine; Computer science; Library science; Knowledge management; Economic growth; Nursing; Pathology; Economics","score_opus":0.842957060811485,"score_gpt":0.5825738731360253,"score_spread":0.26038318767545976,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1483618109","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.16316828,0.48662338,0.2692674,0.06798482,0.0028673229,0.009745757,0.0000025945822,0.000011460553,0.00032899817],"genre_scores_gemma":[0.98448795,0.008023638,0.0053435354,0.00045462672,0.0007536495,0.00007988853,0.0000029650887,0.000020663974,0.0008330784],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.94851714,0.019810535,0.022867136,0.0006303273,0.0079351,0.0002397445],"domain_scores_gemma":[0.89528346,0.03883656,0.044315193,0.0026883592,0.018417709,0.0004587246],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.24522083,0.00042188086,0.00805038,0.0019801538,0.000092950984,0.00008040948,0.0010313413,0.00012665949,0.0034963603],"category_scores_gemma":[0.5259732,0.00015062978,0.00085439114,0.0035202778,0.00054286694,0.0013830456,0.000035043533,0.00022318194,0.000079903395],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00074196566,0.0009508242,0.05608591,0.33120292,0.0012784749,0.000027577576,0.014033392,0.0016018267,0.21547139,0.002231777,0.3322066,0.044167336],"study_design_scores_gemma":[0.0028789924,0.0056508183,0.020996535,0.9264463,0.005308474,0.00017835115,0.0060002315,0.019107135,0.0038722663,0.005995139,0.0030185909,0.00054717675],"about_ca_topic_score_codex":0.00004119201,"about_ca_topic_score_gemma":0.000014188989,"teacher_disagreement_score":0.8213197,"about_ca_system_score_codex":0.00010743547,"about_ca_system_score_gemma":0.00034676565,"threshold_uncertainty_score":0.9974146},"labels":[{"model":"gemma","categories":["metaresearch","bibliometrics"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["bibliometrics","scholarly_communication"],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W1505158704","doi":"10.1111/j.1756-5391.2011.01118.x","title":"International comparative analyses of healthcare risk management","year":2011,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Patient Safety and Medication Errors","field":"Health Professions","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Government (linguistics); Health care; Risk management; China; Public administration; Business; Administration (probate law); Descriptive statistics; Political science; Medicine; Public relations; Law; Finance","score_opus":0.8040728549478809,"score_gpt":0.6454915615042782,"score_spread":0.15858129344360272,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1505158704","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000005735646,0.98489845,0.0039245114,0.0036332398,0.0034215897,0.000980178,0.00004519734,0.0000115546,0.0030795587],"genre_scores_gemma":[0.0007269705,0.99541897,0.0015295024,0.0005299909,0.0011648145,0.000053609376,0.00005300035,0.000021999216,0.00050112495],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99232054,0.002245185,0.0037410466,0.000251325,0.001134645,0.0003072378],"domain_scores_gemma":[0.9827917,0.002955895,0.012017331,0.0004397855,0.0014268791,0.00036840086],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0039383303,0.0003533026,0.00280104,0.0012600853,0.0001750137,4.7303118e-7,0.00082842493,0.0002934652,0.0015392066],"category_scores_gemma":[0.0019247348,0.00018950186,0.00054572197,0.00064387004,0.00028270096,0.00020481103,0.00007976305,0.0017045161,0.00006043726],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007637594,0.0004017035,0.0018883841,0.078211315,0.003215978,0.000090057896,0.004772067,0.00004860903,0.000001056194,0.0012091715,0.034597352,0.87480056],"study_design_scores_gemma":[0.0008275001,0.0005242347,0.0003229859,0.23502062,0.002463945,0.000006832372,0.0010104359,0.000006652948,7.368748e-7,0.000067623034,0.7596399,0.00010853437],"about_ca_topic_score_codex":0.00019953294,"about_ca_topic_score_gemma":0.000015614749,"teacher_disagreement_score":0.874692,"about_ca_system_score_codex":0.00046944802,"about_ca_system_score_gemma":0.0016693987,"threshold_uncertainty_score":0.9993735},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W1509438354","doi":"10.1111/j.1756-5391.2012.01174.x","title":"Assessing the recommendations for the use of diagnostic imaging in clinical practice guidelines","year":2012,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Radiation Dose and Imaging","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Manitoba; Children's Hospital Research Institute of Manitoba; Manitoba Health","funders":"","keywords":"Clinical Practice; Appropriate Use Criteria; Medicine; Medical physics; Medical imaging; Diagnostic accuracy; Diagnostic test; Evidence-based medicine; Quality of evidence; Evidence-based practice; Radiology; Alternative medicine; Pediatrics; Physical therapy; Randomized controlled trial; Pathology; Internal medicine","score_opus":0.6575977203867903,"score_gpt":0.5880777676982109,"score_spread":0.06951995268857947,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1509438354","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.02234512,0.024197794,0.051669706,0.8995479,0.0017032534,0.0004937591,0.0000012142491,0.0000060061147,0.000035283538],"genre_scores_gemma":[0.8981182,0.008297198,0.03613645,0.052578535,0.0047782464,0.000015868854,0.000003317791,0.000021850092,0.000050311075],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99696153,0.00037053516,0.0018647504,0.00008894761,0.00050673366,0.0002074748],"domain_scores_gemma":[0.9194107,0.07682622,0.0015561682,0.00029775614,0.0017648601,0.00014426306],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.012469264,0.00010817129,0.00038899246,0.00020066404,0.00009943194,0.000036838643,0.00016706181,0.000034202392,0.000060400842],"category_scores_gemma":[0.33479673,0.00004579954,0.00016208357,0.00035078125,0.00019965786,0.0020524312,0.000017240965,0.00043272163,0.0000012274888],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00030272282,0.00029902466,0.3017039,0.000077062876,0.0001084617,0.000021031636,0.0006045619,0.0005528312,0.00054405635,0.000065385415,0.22803059,0.46769038],"study_design_scores_gemma":[0.0041281255,0.00053502223,0.3048803,0.011503477,0.0017845352,0.00045761743,0.005071886,0.017957829,0.00038327315,0.000050187526,0.65313137,0.00011641714],"about_ca_topic_score_codex":0.00007245103,"about_ca_topic_score_gemma":0.0000021498604,"teacher_disagreement_score":0.87577313,"about_ca_system_score_codex":0.00007576768,"about_ca_system_score_gemma":0.00049140776,"threshold_uncertainty_score":0.6708066},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"evaluation","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W1518695103","doi":"10.1111/jebm.12005","title":"Should measuring haemoglobin among chronic kidney disease patients be a performance measure?","year":2012,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Erythropoietin and Anemia Treatment","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Kidney disease; Renal function; Observational study; Disease; Intensive care medicine; Internal medicine","score_opus":0.15440220857958156,"score_gpt":0.32853897968639495,"score_spread":0.1741367711068134,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1518695103","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.94355756,0.015043487,0.00019093978,0.038884334,0.0012929651,0.0005491033,0.000004838747,0.000040447696,0.0004363002],"genre_scores_gemma":[0.9908466,0.0016788003,0.00044619702,0.0037747654,0.0028342335,0.000015042325,0.000014972974,0.000043126347,0.0003463092],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99600595,0.00014998986,0.00096576184,0.0002286841,0.0020223362,0.00062725035],"domain_scores_gemma":[0.9955958,0.00025876064,0.0006842926,0.00042857399,0.00080287695,0.0022296975],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0019118731,0.0003641509,0.0007848332,0.0003478742,0.00014590315,0.000013924782,0.00021278062,0.000115406394,0.00064941554],"category_scores_gemma":[0.003630409,0.0002079916,0.00028979516,0.000361289,0.0003609706,0.00073181104,0.00003439759,0.0006227649,0.000020630581],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001360428,0.0007565606,0.9640394,0.00087338657,0.00015565059,0.00014871682,0.00029809616,0.000037676906,0.0015355279,0.000005453757,0.022393147,0.008395976],"study_design_scores_gemma":[0.01445578,0.004949485,0.9133561,0.04311643,0.0020444677,0.00014259598,0.00015822292,0.00028986516,0.006799503,0.000006172384,0.01431199,0.00036937516],"about_ca_topic_score_codex":0.000032093136,"about_ca_topic_score_gemma":0.0000024532908,"teacher_disagreement_score":0.050683267,"about_ca_system_score_codex":0.0010544625,"about_ca_system_score_gemma":0.0012576496,"threshold_uncertainty_score":0.84816515},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1520569313","doi":"10.1111/jebm.12113","title":"Aortic arch atheroma and the risk of stroke: a meta‐analysis","year":2014,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Aortic Thrombus and Embolism","field":"Medicine","cited_by":23,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Atheroma; Aortic arch; Meta-analysis; Medicine; Stroke (engine); Internal medicine; Epidemiology; Observational study; Cardiology; Aorta","score_opus":0.23131658846694136,"score_gpt":0.420063952514773,"score_spread":0.18874736404783166,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1520569313","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00027522276,0.9922998,0.002030093,0.004435037,0.00020493078,0.0006539673,0.000012476227,0.0000072012936,0.00008130195],"genre_scores_gemma":[0.005130186,0.9922954,0.0006464375,0.00044307826,0.0010563108,0.00001916118,0.0000047023377,0.000043202665,0.0003615469],"study_design_codex":"design_other","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9935699,0.0013283796,0.0029443244,0.00032085687,0.0015047219,0.00033182657],"domain_scores_gemma":[0.98596525,0.0071924147,0.004880185,0.00085858704,0.0006860141,0.00041754253],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.009105579,0.00053855474,0.011472246,0.0010253541,0.00008148011,0.0000136066365,0.0004393175,0.0002652499,0.00053669856],"category_scores_gemma":[0.005744356,0.00019191351,0.0045001316,0.001044384,0.001201449,0.00008797084,0.000058132562,0.001311446,0.0000039920055],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011896398,0.00021951285,0.00075867557,0.033620927,0.44979513,0.00018252704,0.0005688673,0.000034304503,0.000006964959,0.00020119977,0.0017449888,0.51167727],"study_design_scores_gemma":[0.0026231501,0.0012807805,0.00029753588,0.02571732,0.84602964,0.00029086784,0.00007378198,0.00016264092,0.0000020305147,0.00008681073,0.12327763,0.00015781107],"about_ca_topic_score_codex":0.0003079776,"about_ca_topic_score_gemma":0.000013601255,"teacher_disagreement_score":0.51151943,"about_ca_system_score_codex":0.0000762309,"about_ca_system_score_gemma":0.00089977635,"threshold_uncertainty_score":0.7842198},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"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,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_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","score_opus":0.9254087096431709,"score_gpt":0.6418932184013989,"score_spread":0.283515491241772,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1544000882","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.38564122,0.48095945,0.071555436,0.040302485,0.009081956,0.012014627,0.000026140666,0.000008131528,0.00041056535],"genre_scores_gemma":[0.98966813,0.003777136,0.0035599447,0.00042142146,0.0016908749,0.00003660743,0.0000011957098,0.000017875831,0.00082682143],"study_design_codex":"observational","study_design_gemma":"systematic_review","domain_scores_codex":[0.6303498,0.23563015,0.10921378,0.00080768246,0.023329087,0.0006694524],"domain_scores_gemma":[0.38558534,0.3831993,0.2162828,0.0060567074,0.007872635,0.0010032163],"candidate_categories":["metaresearch","metaepi_broad","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.8492686,0.0005973679,0.037499912,0.0018333021,0.00012852259,0.00010640393,0.0027436735,0.00017772199,0.0063042007],"category_scores_gemma":[0.933216,0.0001485867,0.008555588,0.0025896144,0.00021962752,0.0010803626,0.00025539182,0.00037255188,0.000069134985],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018399614,0.0014401148,0.5241669,0.19458123,0.0052865897,0.000058488655,0.011027563,0.00016563926,0.06524648,0.000490934,0.14578661,0.04990954],"study_design_scores_gemma":[0.022962999,0.009835026,0.058029577,0.7365054,0.068451986,0.0005901997,0.009401921,0.004052696,0.009448092,0.011387677,0.06703042,0.0023039936],"about_ca_topic_score_codex":0.00001280025,"about_ca_topic_score_gemma":0.000006322286,"teacher_disagreement_score":0.6040269,"about_ca_system_score_codex":0.00010273201,"about_ca_system_score_gemma":0.0002551734,"threshold_uncertainty_score":0.9946042},"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":"W1556696529","doi":"10.1111/j.1756-5391.2011.01116.x","title":"A mathematical model for evaluating tuberculosis screening strategies","year":2011,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Tuberculosis Research and Epidemiology","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Tuberculosis; Transmission (telecommunications); Immigration; Disease; Basic reproduction number; Medicine; Population; Environmental health; Order (exchange); Demography; Computer science; Geography; Telecommunications; Internal medicine; Business; Pathology","score_opus":0.5856095892322939,"score_gpt":0.4915099256631166,"score_spread":0.09409966356917732,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1556696529","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.16298346,0.0042559323,0.78731227,0.042984232,0.0001479416,0.00082969747,0.0000035327498,0.000036680478,0.0014462684],"genre_scores_gemma":[0.8178634,0.000430113,0.17785658,0.0029256516,0.00068530097,0.000036037258,0.0000030662245,0.00003202499,0.0001678388],"study_design_codex":"bench_or_experimental","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9967068,0.00024134731,0.0013301533,0.00023503277,0.0009561683,0.00053052267],"domain_scores_gemma":[0.9944072,0.0028318856,0.0005513624,0.00031637403,0.0013066854,0.0005865286],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0074859587,0.00023097798,0.0010553433,0.00045257024,0.0000970082,0.000012400966,0.00024474954,0.0001505518,0.0005540883],"category_scores_gemma":[0.021014323,0.00012641426,0.0003600718,0.00025552465,0.00031669365,0.00044022303,0.000028847497,0.00054782955,0.000007879507],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.08332994,0.004265452,0.07328532,0.014892064,0.0059918156,0.0018381386,0.020703021,0.03218862,0.2967478,0.023788637,0.16297978,0.2799894],"study_design_scores_gemma":[0.0066354247,0.013328961,0.007366401,0.012778995,0.0013991828,0.0007079498,0.0020609547,0.9361802,0.0020022963,0.017109778,0.00018626441,0.00024363806],"about_ca_topic_score_codex":0.000038457816,"about_ca_topic_score_gemma":0.0000043176897,"teacher_disagreement_score":0.9039915,"about_ca_system_score_codex":0.000109059205,"about_ca_system_score_gemma":0.0008964315,"threshold_uncertainty_score":0.9872321},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"simulation_or_modeling","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"simulation_or_modeling","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1580273645","doi":"10.1111/jebm.12169","title":"Evidence‐informed health policy making in Canada: past, present, and future","year":2015,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Primary Care and Health Outcomes","field":"Health Professions","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"London Health Sciences Centre; Western University","funders":"Western University of Health Sciences","keywords":"Psychological intervention; Political science; Evidence-based policy; Health policy; Policy making; Field (mathematics); Public relations; Public administration; Health care; Medicine; Alternative medicine; Nursing; Law","score_opus":0.3298856697761844,"score_gpt":0.5201995464452814,"score_spread":0.19031387666909705,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1580273645","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.01639821,0.053802803,0.000101620695,0.92588615,0.0022390154,0.00061081117,0.0000042689658,0.000012794666,0.0009443165],"genre_scores_gemma":[0.62100005,0.02718788,0.0012850685,0.31326768,0.036654126,0.00003451965,0.0000044249878,0.000047072386,0.0005191675],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9947683,0.0009029408,0.0020277593,0.00019693103,0.0012779243,0.0008261188],"domain_scores_gemma":[0.9923279,0.0039073983,0.0016601343,0.00030651904,0.00076595857,0.0010320874],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005621934,0.00023506231,0.0010212755,0.00067598815,0.00023386179,0.0000071609074,0.00034673908,0.0001346909,0.00016968988],"category_scores_gemma":[0.0034869239,0.00014170627,0.000055984056,0.0007093539,0.000093828276,0.0006972712,0.00008263863,0.0012936008,0.0000044769954],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.0007759854,0.000016828944,0.28756538,0.0013964007,0.000019038334,0.0001609852,0.0035508308,0.000026645706,0.0000060915922,0.00012911567,0.65719426,0.049158424],"study_design_scores_gemma":[0.0043826676,0.0014285947,0.3343969,0.023210686,0.000043410168,0.000078771176,0.012379544,0.00007417398,0.000002620944,0.0003224204,0.62348014,0.00020003345],"about_ca_topic_score_codex":0.6435193,"about_ca_topic_score_gemma":0.59063315,"teacher_disagreement_score":0.6126185,"about_ca_system_score_codex":0.009044662,"about_ca_system_score_gemma":0.13345873,"threshold_uncertainty_score":0.99475944},"labels":[{"model":"gemma","categories":["scholarly_communication"],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":true,"about_ca_topic":true,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":true,"about_ca_topic":true,"confidence":"medium"}],"label_agreement":"split"},{"id":"W1601804926","doi":"10.1111/j.1756-5391.2011.01119.x","title":"International comparative analyses of incidents reporting systems for healthcare risk management","year":2011,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Patient Safety and Medication Errors","field":"Health Professions","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Government (linguistics); Health care; Confidentiality; Risk management; Business; Incident report; Patient safety; Medicine; Medical emergency; Political science; Computer security; Finance; Law","score_opus":0.8600866229067199,"score_gpt":0.6659249706519029,"score_spread":0.19416165225481696,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1601804926","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000029679532,0.97647166,0.013783365,0.0009263595,0.0057053906,0.0023754924,0.000059847178,0.000012577491,0.00063565036],"genre_scores_gemma":[0.0021900677,0.9940241,0.0013485646,0.00020484252,0.0014928778,0.00021920016,0.00008858542,0.000027809123,0.0004039266],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9880272,0.0016545886,0.008534928,0.00029294365,0.0011530509,0.00033729253],"domain_scores_gemma":[0.9451914,0.0036992482,0.048021186,0.00041537947,0.002356279,0.0003165464],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.010001965,0.00034786738,0.0033742578,0.0011283214,0.00024050924,0.0000010830264,0.00067968905,0.00029786653,0.00015734894],"category_scores_gemma":[0.009763259,0.00019581441,0.0006141691,0.0004833278,0.00016159199,0.00022262044,0.00007260931,0.0011556484,0.000010784568],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001611098,0.0005534626,0.010959607,0.36231902,0.006890894,0.00017173169,0.009404319,0.00044421147,0.0000023796522,0.0010761524,0.044241227,0.5623259],"study_design_scores_gemma":[0.00091430295,0.0006080505,0.00031203296,0.31492266,0.0026710168,0.000015494865,0.0020354828,0.00005059699,4.341334e-7,0.000035836045,0.67832047,0.000113594324],"about_ca_topic_score_codex":0.00034161514,"about_ca_topic_score_gemma":0.000010831414,"teacher_disagreement_score":0.6340793,"about_ca_system_score_codex":0.0006486793,"about_ca_system_score_gemma":0.0017305372,"threshold_uncertainty_score":0.99857795},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1603245870","doi":"10.1111/j.1756-5391.2011.01158.x","title":"Effects and patient compliance of sustained‐release versus immediate‐release glipizides in patients with type 2 diabetes mellitus: a systematic review and meta‐analysis","year":2011,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Diabetes Treatment and Management","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"National Natural Science Foundation of China; Pfizer","keywords":"Glipizide; Medicine; Postprandial; Internal medicine; Hypoglycemia; Odds ratio; Insulin; Diabetes mellitus; Type 2 diabetes; Meta-analysis; Cochrane Library; Randomized controlled trial; Type 2 Diabetes Mellitus; Endocrinology","score_opus":0.08992171394064337,"score_gpt":0.32841836815324693,"score_spread":0.23849665421260358,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1603245870","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000950497,0.9944469,0.0000079912,0.00033128457,0.00009844838,0.00413415,0.000009991644,0.000008370157,0.0000123841555],"genre_scores_gemma":[0.0048941425,0.99403864,0.00029471132,0.00049653987,0.000038128597,0.00013630223,0.000033501292,0.00004437529,0.000023685638],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9955909,0.0005481972,0.0021946526,0.00042150944,0.00086142163,0.00038329005],"domain_scores_gemma":[0.9926136,0.0023790926,0.0031182035,0.0005236668,0.00090590975,0.0004595081],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.001286995,0.0007095203,0.010535303,0.0012435116,0.00003165996,0.000013337312,0.00021058045,0.00013278148,0.00007386451],"category_scores_gemma":[0.0024609985,0.0003328584,0.0007552041,0.0014680788,0.00028598352,0.00021932379,0.00006523581,0.0003876368,0.000002447438],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00060995156,0.0004715877,0.0005470939,0.91018826,0.0715746,0.00013511884,0.000060440583,0.0000010564696,1.1381498e-7,0.0000088669085,0.00013269448,0.016270224],"study_design_scores_gemma":[0.0029216532,0.0063087116,0.00021841093,0.40480345,0.58451444,0.0000024404835,0.000032305878,0.000004109467,0.0000017372363,0.0000026727498,0.0010007331,0.00018933529],"about_ca_topic_score_codex":0.00002424417,"about_ca_topic_score_gemma":0.0000045274587,"teacher_disagreement_score":0.5129398,"about_ca_system_score_codex":0.00020182681,"about_ca_system_score_gemma":0.0003037367,"threshold_uncertainty_score":0.9999123},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1625449191","doi":"10.1111/j.1756-5391.2011.01127.x","title":"Visualization studies on evidence‐based medicine domain knowledge (series 2): structural diagrams of author networks","year":2011,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"China Medical Board","keywords":"Zhàng; China; Visualization; Citation; Cluster analysis; Graph drawing; Computer science; Domain (mathematical analysis); Library science; Data science; Information retrieval; Geography; Data mining; Artificial intelligence; Mathematics","score_opus":0.9085942578151824,"score_gpt":0.6040802977165222,"score_spread":0.3045139600986603,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1625449191","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.41678205,0.27888423,0.235272,0.05330828,0.011890634,0.0027905842,0.000007784271,0.00003287692,0.0010315503],"genre_scores_gemma":[0.9911826,0.0014852963,0.003692849,0.0010416247,0.001627153,0.000016787626,0.0000025855218,0.000026011481,0.0009250678],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9627207,0.012035186,0.016222617,0.0007479131,0.007862512,0.00041107624],"domain_scores_gemma":[0.93909436,0.028417561,0.020545464,0.002320327,0.00909028,0.0005319876],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.15449385,0.00061888737,0.0065466203,0.0015931365,0.00021500493,0.00008201125,0.0020909049,0.00018040203,0.0055659716],"category_scores_gemma":[0.2378163,0.00022957209,0.001342356,0.0032798485,0.0011493532,0.0011326063,0.000083176965,0.00046448005,0.000051824278],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.007976664,0.0009861822,0.22533655,0.004489833,0.0050740233,0.00063049525,0.049717266,0.016271925,0.0062286714,0.03304093,0.52974576,0.12050169],"study_design_scores_gemma":[0.016926404,0.07174381,0.23323873,0.30279425,0.018608226,0.0005903781,0.07009902,0.13864766,0.01043741,0.089099266,0.04435157,0.003463283],"about_ca_topic_score_codex":0.00002353862,"about_ca_topic_score_gemma":0.000052151347,"teacher_disagreement_score":0.57440054,"about_ca_system_score_codex":0.00017011445,"about_ca_system_score_gemma":0.00043702533,"threshold_uncertainty_score":0.9953431},"labels":[{"model":"gemma","categories":["bibliometrics"],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["bibliometrics"],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1709885097","doi":"10.1111/jebm.12141","title":"The methodological quality assessment tools for preclinical and clinical studies, systematic review and meta‐analysis, and clinical practice guideline: a systematic review","year":2015,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Animal testing and alternatives","field":"Veterinary","cited_by":2088,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Systematic review; Critical appraisal; Checklist; Medicine; Guideline; Randomized controlled trial; Psychological intervention; Clinical study design; MEDLINE; Medical physics; Clinical trial; Psychology; Alternative medicine; Nursing; Pathology","score_opus":0.9641987842621917,"score_gpt":0.7638045496857743,"score_spread":0.20039423457641747,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1709885097","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[6.1593215e-7,0.97002155,0.0032552353,0.017919566,0.0003179767,0.008416981,0.000030488067,0.00002468812,0.000012925252],"genre_scores_gemma":[0.0000026426117,0.9538238,0.037422236,0.0074514854,0.0006864537,0.00047403635,0.000009973553,0.00004277325,0.000086573535],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.92361534,0.053296633,0.020008624,0.0009977978,0.0016636817,0.00041793825],"domain_scores_gemma":[0.49991372,0.4677504,0.025329707,0.0013561097,0.0048336997,0.0008163518],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.3204975,0.0009497236,0.034438703,0.00019410557,0.00025217005,0.00015140121,0.00091662176,0.00040684047,0.000015606216],"category_scores_gemma":[0.70256656,0.0003525215,0.0038809828,0.0006425223,0.001180856,0.00051672664,0.00036069524,0.0020886546,0.0000021026196],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015114858,0.00012443263,0.000005977323,0.88212377,0.10785063,0.00011483513,0.000009434597,3.1533474e-8,1.0053365e-8,0.00006160371,0.0029307543,0.0066273506],"study_design_scores_gemma":[0.00034860728,0.00393047,0.000003962003,0.4914607,0.48662508,0.00082912855,0.00011949999,0.000015177279,1.0844924e-9,0.000058287835,0.016410694,0.00019838312],"about_ca_topic_score_codex":0.000011131066,"about_ca_topic_score_gemma":0.0000034125349,"teacher_disagreement_score":0.4237016,"about_ca_system_score_codex":0.00017082883,"about_ca_system_score_gemma":0.0011111263,"threshold_uncertainty_score":0.99989265},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2021893832","doi":"10.1111/j.1756-5391.2009.01030.x","title":"An Educational Approach to Improving Healthcare Safety and Quality*","year":2009,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Patient Safety and Medication Errors","field":"Health Professions","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Health care; Quality (philosophy); Patient safety; Business; Nursing; Risk analysis (engineering); Computer science; Medicine; Political science; Philosophy","score_opus":0.3683131285364435,"score_gpt":0.5357444901409603,"score_spread":0.16743136160451677,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2021893832","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08192934,0.0030770951,0.07983702,0.8314282,0.0014634406,0.0009231718,0.000008382497,0.00003042239,0.0013029546],"genre_scores_gemma":[0.95513177,0.0003154428,0.0066153356,0.035186097,0.0024089552,0.000010049878,0.000015650627,0.000009667544,0.00030704206],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9963816,0.0008728595,0.0014112437,0.00021579636,0.00079170056,0.0003268336],"domain_scores_gemma":[0.9954048,0.0013860093,0.0011838243,0.0002866533,0.00082996825,0.0009087931],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005372237,0.00014552998,0.00047087861,0.0003419194,0.0004020797,0.0000012182322,0.00024763696,0.00012909298,0.00012818536],"category_scores_gemma":[0.005692789,0.00009198132,0.000047853202,0.00036034582,0.000093704606,0.00041550476,0.000013188215,0.0007099689,0.000008872678],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.015624218,0.0023575046,0.36639422,0.0057478417,0.00015084137,0.000024596227,0.08178734,0.003178914,0.029538816,0.06651256,0.029748164,0.398935],"study_design_scores_gemma":[0.0034776505,0.004036202,0.9457317,0.00788311,0.000090471076,0.00002268708,0.007980895,0.0004596932,0.000043161566,0.00090683775,0.029120132,0.00024747784],"about_ca_topic_score_codex":0.00014197831,"about_ca_topic_score_gemma":0.000016689492,"teacher_disagreement_score":0.87320244,"about_ca_system_score_codex":0.00027368686,"about_ca_system_score_gemma":0.0018787817,"threshold_uncertainty_score":0.6815211},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2028332028","doi":"10.1111/jebm.12101","title":"Production and citation of cochrane systematic reviews: a bibliometrics analysis","year":2014,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Citation; Bibliometrics; Systematic review; Cochrane Library; Psychological intervention; Knowledge translation; Production (economics); Medicine; MEDLINE; Business; Political science; Meta-analysis; Library science; Computer science; Economics; Nursing; Knowledge management; Pathology","score_opus":0.8622232322885919,"score_gpt":0.5786419193064429,"score_spread":0.28358131298214895,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2028332028","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.21536368,0.18661815,0.5736138,0.020319225,0.0013453573,0.0024270318,0.0000021399328,0.0000046495707,0.00030595285],"genre_scores_gemma":[0.9873437,0.0058032623,0.0056288172,0.0003244275,0.00035875986,0.000012247806,9.3580013e-7,0.000006796563,0.00052101235],"study_design_codex":"observational","study_design_gemma":"systematic_review","domain_scores_codex":[0.9533605,0.017425068,0.019110652,0.00048748942,0.009451026,0.00016528454],"domain_scores_gemma":[0.88723487,0.0490125,0.047401432,0.0031161965,0.0128177665,0.00041724066],"candidate_categories":["metaresearch","bibliometrics","insufficient_payload"],"consensus_categories":["metaresearch","bibliometrics"],"category_scores_codex":[0.407464,0.0002458996,0.008243145,0.018508721,0.00006494156,0.00015286429,0.0008644093,0.00006911812,0.0009799162],"category_scores_gemma":[0.7342431,0.00008815822,0.0015418634,0.040661726,0.00015976245,0.0006073672,0.00002471358,0.00017897304,0.000040908657],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00052128994,0.000785338,0.39234075,0.21555634,0.010662745,0.00004642158,0.0066227317,0.010740627,0.01871514,0.0016978261,0.22918886,0.11312194],"study_design_scores_gemma":[0.0044271844,0.009323125,0.23589638,0.44162133,0.11135337,0.00059854134,0.007862932,0.15408729,0.0016410734,0.011138619,0.020561524,0.0014886314],"about_ca_topic_score_codex":0.00000906137,"about_ca_topic_score_gemma":0.000008356973,"teacher_disagreement_score":0.77198005,"about_ca_system_score_codex":0.000043305048,"about_ca_system_score_gemma":0.00008820736,"threshold_uncertainty_score":0.9999333},"labels":[],"label_agreement":null},{"id":"W2028766252","doi":"10.1111/jebm.12074","title":"Editorial for issue 4 of 2013, Journal of Evidence Based Medicine","year":2013,"lang":"en","type":"editorial","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Systematic review; Cochrane collaboration; Medicine; Health care; China; Alternative medicine; MEDLINE; Family medicine; Public relations; Library science; Medical education; Political science; Law; Computer science","score_opus":0.8002989571297778,"score_gpt":0.5755311978888985,"score_spread":0.22476775924087933,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2028766252","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00008129318,0.05653611,0.034571793,0.041421827,0.8659048,0.0013819295,0.00004497079,0.000003239839,0.000054003678],"genre_scores_gemma":[0.0006762965,0.007775657,0.0071140616,0.0006472155,0.9801059,0.000029657758,0.000014368333,0.00007758902,0.0035592965],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.89012593,0.014171874,0.046092097,0.0012324781,0.047581982,0.00079565647],"domain_scores_gemma":[0.51964575,0.25664502,0.12538938,0.0052838763,0.09144241,0.0015935502],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","open_science","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch","metaepi_narrow","research_integrity"],"category_scores_codex":[0.3817669,0.0013358752,0.020027343,0.004218463,0.00018664735,0.0003435825,0.008713016,0.0013942029,0.020416422],"category_scores_gemma":[0.7878143,0.00048779798,0.0058421455,0.0027559241,0.0011150987,0.0020753595,0.00014323513,0.00269357,0.00016905606],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012933606,0.0001385973,0.0001594164,0.0017965643,0.0007771544,0.000039864542,0.00025092,0.00034572618,0.00094456447,0.000007806545,0.98791504,0.006331004],"study_design_scores_gemma":[0.0042245314,0.008244799,0.000042798412,0.05383369,0.005913921,0.00003544609,0.0004302402,0.00082567416,0.00015756839,0.0008885492,0.92500806,0.00039472998],"about_ca_topic_score_codex":0.00019804205,"about_ca_topic_score_gemma":0.000020367674,"teacher_disagreement_score":0.40604743,"about_ca_system_score_codex":0.00044064352,"about_ca_system_score_gemma":0.0068996954,"threshold_uncertainty_score":0.99993926},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2049110830","doi":"10.1111/j.1756-5391.2009.01006.x","title":"Meta‐analysis of randomized trials for health care interventions: one for all?","year":2009,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"National Natural Science Foundation of China","keywords":"Randomized controlled trial; Meta-analysis; Psychological intervention; Medicine; Psychology; Nursing","score_opus":0.9820464684378005,"score_gpt":0.719992384647849,"score_spread":0.26205408378995154,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2049110830","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[1.13670424e-7,0.7192618,0.26279855,0.010104834,0.00043579226,0.0072151064,0.00016688164,0.0000017867412,0.000015140643],"genre_scores_gemma":[0.000011414157,0.96260685,0.033649325,0.0012782946,0.00079625315,0.00059876835,0.000099727375,0.00002975015,0.0009296482],"study_design_codex":"design_other","study_design_gemma":"meta_analysis","domain_scores_codex":[0.7764243,0.09685356,0.11382454,0.001256445,0.0111153135,0.0005258562],"domain_scores_gemma":[0.5080314,0.30070832,0.17418838,0.0033379009,0.013187457,0.0005464837],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","insufficient_payload"],"consensus_categories":["metaresearch","metaepi_broad"],"category_scores_codex":[0.72090745,0.00092039973,0.15800548,0.00713946,0.0001184443,0.00024526223,0.0032670805,0.00030899426,0.0051611923],"category_scores_gemma":[0.6310614,0.00029857343,0.17433842,0.004538936,0.00022337568,0.00032382138,0.00004303902,0.00043552736,0.000012853081],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002352204,0.00012596803,3.0173433e-7,0.037118502,0.43341592,0.0000021443395,0.0002040211,0.00008146564,1.790099e-7,0.00019387815,0.04864854,0.47785687],"study_design_scores_gemma":[0.009224164,0.00084544194,2.434127e-7,0.033094298,0.55324346,0.000004658536,0.0001160126,0.00011111546,2.7325945e-7,0.0005538053,0.4026764,0.0001301632],"about_ca_topic_score_codex":0.000017047647,"about_ca_topic_score_gemma":0.0000320321,"teacher_disagreement_score":0.4777267,"about_ca_system_score_codex":0.00025904897,"about_ca_system_score_gemma":0.002057784,"threshold_uncertainty_score":0.99994665},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2068393204","doi":"10.1111/j.1756-5391.2010.01105.x","title":"Effects of restrictions on reimbursement for pharmaceutical drugs","year":2010,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Healthcare Policy and Management","field":"Economics, Econometrics and Finance","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria","funders":"","keywords":"Reimbursement; Incentive; Health care; Pharmaceutical care; Systematic review; Business; Actuarial science; Pharmaceutical industry; MEDLINE; Medicine; Public economics; Public relations; Family medicine; Economics; Political science; Pharmacology; Pharmacy; Law; Economic growth","score_opus":0.20302714102185632,"score_gpt":0.39963071016154106,"score_spread":0.19660356913968474,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2068393204","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.38736528,0.010202776,0.04651414,0.53821385,0.010830667,0.002170397,0.000035495108,0.00003226519,0.0046351408],"genre_scores_gemma":[0.99230576,0.0012475583,0.0013264187,0.0037412923,0.0011490867,0.000021929143,7.389591e-7,0.00001157823,0.000195631],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.9985267,0.000029237019,0.0009728856,0.00014053842,0.0001247582,0.00020589175],"domain_scores_gemma":[0.99720645,0.0014676889,0.0007864961,0.00018321184,0.00015184649,0.00020433635],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0023160807,0.00010292376,0.0004325976,0.00049584376,0.00005535114,0.000006629301,0.00019361058,0.000065547334,0.000112080175],"category_scores_gemma":[0.0034545988,0.000076662334,0.0001392713,0.00021571398,0.00009200288,0.00011877063,0.000012233436,0.00034351877,0.000011495854],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019993177,0.0014312393,0.0054413085,0.003784307,0.0004297416,0.00006260132,0.001220644,0.00040913344,0.009148408,0.7997077,0.094993085,0.08137249],"study_design_scores_gemma":[0.008650443,0.00736721,0.016699111,0.0034560757,0.0002293982,0.000011358729,0.00013734188,0.0016050754,0.017834386,0.02202714,0.92168975,0.00029270357],"about_ca_topic_score_codex":0.00019126276,"about_ca_topic_score_gemma":0.0000069329244,"teacher_disagreement_score":0.8266967,"about_ca_system_score_codex":0.00008173755,"about_ca_system_score_gemma":0.00009711298,"threshold_uncertainty_score":0.4135727},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2125660710","doi":"10.1111/jebm.12137","title":"Health equity in humanitarian emergencies: a role for evidence aid","year":2015,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Food Security and Health in Diverse Populations","field":"Health Professions","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Cochrane; Bruyère","funders":"Canadian Institutes of Health Research","keywords":"Health equity; Equity (law); Health care; Disadvantaged; Psychological intervention; Business; Political science; Public relations; Economic growth; Public economics; Medicine; Economics; Nursing","score_opus":0.7822702770008315,"score_gpt":0.6122804426939947,"score_spread":0.16998983430683678,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2125660710","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.14973056,0.087967865,0.010601576,0.73492104,0.0102924565,0.004847204,0.000046017485,0.00009072354,0.0015025551],"genre_scores_gemma":[0.97724605,0.002121886,0.0039044903,0.013552954,0.0027867032,0.00008035741,0.0000074983914,0.000025764151,0.00027426993],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9946711,0.00097629207,0.002337757,0.00023482721,0.0010332275,0.00074679096],"domain_scores_gemma":[0.99229795,0.0032167635,0.0017328459,0.00033729317,0.0015540351,0.00086109136],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.015419218,0.00019930975,0.0008669317,0.0005969064,0.0005561877,0.000007383803,0.00049876305,0.0001855401,0.00040166613],"category_scores_gemma":[0.020613201,0.0001397517,0.00012382453,0.00059392647,0.00014563586,0.0007799367,0.00009052247,0.0010856445,0.00003619674],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0070747402,0.0008007665,0.2531239,0.007815451,0.00009971912,0.0000868419,0.10237145,0.0020520124,0.00032082584,0.013792544,0.5966194,0.015842319],"study_design_scores_gemma":[0.02737833,0.037600115,0.12689653,0.19444472,0.00043343523,0.000052277745,0.11700098,0.007854672,0.00004896615,0.06353323,0.42365912,0.0010976215],"about_ca_topic_score_codex":0.0035622637,"about_ca_topic_score_gemma":0.0059864125,"teacher_disagreement_score":0.8275155,"about_ca_system_score_codex":0.0015428268,"about_ca_system_score_gemma":0.007489957,"threshold_uncertainty_score":0.99813664},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2375639168","doi":"","title":"Principles for International Registration of Protocol Information and Results from Human Trials of Health-related Interventions-The Ottawa Statement,Part One","year":2006,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Pharmaceutical studies and practices","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Statement (logic); Protocol (science); Psychological intervention; Trial registration; Computer science; Clinical trial; Medicine; Alternative medicine; Political science; Law; Pathology; Nursing","score_opus":0.5417398428296815,"score_gpt":0.5303620964796135,"score_spread":0.011377746350067941,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2375639168","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.1265496,0.0072833914,0.021852689,0.7801019,0.0014235348,0.060380384,0.00041873084,0.00002640459,0.0019633893],"genre_scores_gemma":[0.9875852,0.00096046354,0.0068288436,0.0009842082,0.0013754528,0.0016626315,0.00027499278,0.000012934491,0.00031528602],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9920835,0.00035205987,0.006769392,0.00007924944,0.00061878696,0.000097008524],"domain_scores_gemma":[0.98504776,0.002747204,0.01092347,0.00011566577,0.0011026948,0.00006323726],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.021854997,0.00009234158,0.0006477699,0.0001569119,0.00007934452,0.000012435389,0.00010123438,0.000041741565,0.000121125144],"category_scores_gemma":[0.024575857,0.000049121398,0.00016045534,0.0001236733,0.00017535182,0.0004688003,0.000016420407,0.00017188993,2.2964763e-7],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.11150448,0.0054344716,0.034335922,0.019042315,0.0059749717,0.000026076308,0.0046880348,0.0018381771,0.05104076,0.01204512,0.46514282,0.28892684],"study_design_scores_gemma":[0.04758866,0.012338519,0.06539263,0.040858634,0.0015024585,0.000023360208,0.0011072896,0.0012444215,0.0126375565,0.002436894,0.8147058,0.00016378096],"about_ca_topic_score_codex":0.00074641994,"about_ca_topic_score_gemma":0.00007025313,"teacher_disagreement_score":0.8610356,"about_ca_system_score_codex":0.00008213826,"about_ca_system_score_gemma":0.00020569292,"threshold_uncertainty_score":0.98364055},"labels":[{"model":"gemma","categories":["metaresearch","open_science"],"domain":"reporting","study_design":"not_applicable","genre":"methods","about_ca_system":true,"about_ca_topic":true,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"reporting","study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2610165645","doi":"10.1111/jebm.12256","title":"Prehospital notification for major trauma patients requiring emergency hospital transport: A systematic review","year":2017,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Trauma and Emergency Care Studies","field":"Medicine","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Surgical Specialties (Canada); University of Saskatchewan","funders":"Department of Industry and Science, Australian Government; Australia-India Strategic Research Fund; Department of Industry, Innovation and Science, Australian Government; Department of Science and Technology, Ministry of Science and Technology, India; Australian Government","keywords":"Medicine; Observational study; Psychological intervention; Emergency medicine; Emergency medical services; Odds ratio; Confidence interval; Major trauma; Emergency department; Medical emergency; Internal medicine; Nursing","score_opus":0.33115534002554575,"score_gpt":0.45649160879190104,"score_spread":0.12533626876635529,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2610165645","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000043091244,0.98377323,0.0003246907,0.0032286635,0.004355981,0.008013673,0.000053196916,0.000031409083,0.00017606597],"genre_scores_gemma":[0.00084446184,0.99527067,0.00032127995,0.00013908169,0.0020517162,0.00057551777,0.000133148,0.00011972215,0.0005443992],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9913237,0.00023089159,0.0056835315,0.000582809,0.0016346889,0.00054436474],"domain_scores_gemma":[0.98776364,0.00061084033,0.007278502,0.0011621903,0.002716301,0.00046854297],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.003071833,0.00090654136,0.0083202785,0.0006023482,0.0002726833,0.000013359857,0.0008118389,0.00033607826,0.0002105005],"category_scores_gemma":[0.011954678,0.0004920484,0.0028913366,0.0003941375,0.00021924081,0.00047053766,0.00002451635,0.00077174715,0.000018674187],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000093499904,0.00026796985,0.0001656331,0.85665303,0.0010809293,0.00006334141,0.00018321312,1.02399596e-7,4.1013845e-7,0.0000065270697,0.0071159666,0.13436936],"study_design_scores_gemma":[0.0015681792,0.004384223,0.00043356058,0.8383707,0.040403645,0.00006610431,0.00004998094,0.0000010414044,0.0000010947783,0.000013373655,0.114300884,0.00040726038],"about_ca_topic_score_codex":0.000011745163,"about_ca_topic_score_gemma":0.0000047120957,"teacher_disagreement_score":0.1339621,"about_ca_system_score_codex":0.00042987335,"about_ca_system_score_gemma":0.0008204688,"threshold_uncertainty_score":0.9997531},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2897220320","doi":"10.1111/jebm.12339","title":"Flexible piecewise linear model for investigating dose‐response relationship in meta‐analysis: Methodology, examples, and comparison","year":2019,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Radiation Dose and Imaging","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; St. Joseph’s Healthcare Hamilton; Impact","funders":"","keywords":"Piecewise linear function; Piecewise; Linear model; Simple (philosophy); Mathematics; Segmented regression; Applied mathematics; Function (biology); Spline (mechanical); Mathematical optimization; Algorithm; Computer science; Statistics; Linear regression; Mathematical analysis; Polynomial regression","score_opus":0.6721250343346913,"score_gpt":0.4885286508437766,"score_spread":0.18359638349091467,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2897220320","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6920411,0.033345968,0.22113474,0.052498505,0.00014916917,0.00076849223,0.0000042558604,0.000024551817,0.000033188247],"genre_scores_gemma":[0.8981242,0.0001884293,0.098179065,0.0029115223,0.00018939137,0.000015362726,0.000007558268,0.000018884126,0.00036558544],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9969239,0.00072982017,0.0013222892,0.00025885078,0.00051150116,0.00025360106],"domain_scores_gemma":[0.987394,0.010707951,0.0009180376,0.0002929699,0.00038508192,0.00030197744],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.011294648,0.00020435722,0.0016361242,0.0014190931,0.00006233425,0.000016076343,0.00013527602,0.00011317853,0.00009243782],"category_scores_gemma":[0.014904818,0.00012970496,0.00035499572,0.0009793262,0.00016182392,0.00045957955,0.000017703489,0.00053481304,0.0000025291756],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.007839732,0.00020686055,0.6634203,0.0007044991,0.012299351,0.000036973193,0.00398067,0.27909464,0.027694345,0.0003772384,0.0017615196,0.0025838816],"study_design_scores_gemma":[0.006474178,0.0015577658,0.10152536,0.001447813,0.049690314,0.000058339763,0.0008793,0.83460706,0.002157235,0.0010101432,0.00039988448,0.00019262185],"about_ca_topic_score_codex":0.00006374613,"about_ca_topic_score_gemma":0.000014861756,"teacher_disagreement_score":0.56189495,"about_ca_system_score_codex":0.0001108157,"about_ca_system_score_gemma":0.0006473627,"threshold_uncertainty_score":0.99339306},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"simulation_or_modeling","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"simulation_or_modeling","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2902864468","doi":"10.1111/jebm.12332","title":"Traffic accidents, maxillofacial injuries and risk factors: A systematic review of observational studies","year":2018,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Facial Trauma and Fracture Management","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Observational study; Medicine; Web of science; Cochrane Library; Poison control; Demography; Injury prevention; Meta-analysis; Environmental health; Internal medicine","score_opus":0.40014623957982814,"score_gpt":0.4597323569938587,"score_spread":0.05958611741403058,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2902864468","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00026196105,0.9949056,0.000055601686,0.0016012916,0.0008127943,0.0023110488,0.00002157456,0.000014351797,0.00001581469],"genre_scores_gemma":[0.00009920229,0.9977387,0.00031242656,0.0007602896,0.00086408877,0.000038536047,0.000026880729,0.000041471543,0.00011837409],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9929094,0.0005853227,0.004120125,0.00033600398,0.0017820538,0.0002671328],"domain_scores_gemma":[0.9892764,0.0021052607,0.0061511304,0.0005028855,0.0016857979,0.00027851274],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00480774,0.0006578878,0.008358857,0.0007703543,0.00009376825,0.000016968004,0.00037739615,0.00022838674,0.00014225903],"category_scores_gemma":[0.026569478,0.00030706837,0.00095522177,0.0006814549,0.0006304494,0.0002769701,0.000071213595,0.0007230208,0.0000059710464],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000921819,0.000084658765,0.0002498482,0.95522004,0.0027607845,0.00006990052,0.00024036613,5.3786334e-7,1.9392733e-7,0.000011096791,0.0038317356,0.037438635],"study_design_scores_gemma":[0.0006130859,0.0020083596,0.00024195419,0.82693785,0.04177759,0.00010512111,0.00023399833,0.0000016796313,0.0000010930773,0.000014364567,0.12787324,0.0001916563],"about_ca_topic_score_codex":0.000011833385,"about_ca_topic_score_gemma":0.0000026691027,"teacher_disagreement_score":0.1282822,"about_ca_system_score_codex":0.0002597031,"about_ca_system_score_gemma":0.00077486783,"threshold_uncertainty_score":0.99993813},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2902907490","doi":"10.1111/jebm.12330","title":"Quality of clinical practice guidelines about red blood cell transfusion","year":2018,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Blood transfusion and management","field":"Medicine","cited_by":30,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Instituto de Salud Carlos III; Canadian Medical Association; National Institute for Health and Care Excellence; Agency for Healthcare Research and Quality; Universitat Autònoma de Barcelona; World Health Organization","keywords":"Medicine; Guideline; Scope (computer science); CLARITY; Intensive care medicine; Computer science; Pathology","score_opus":0.58470875204432,"score_gpt":0.577849232896576,"score_spread":0.006859519147744075,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2902907490","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00006763927,0.96709234,0.0016790415,0.025030725,0.002557545,0.0015341102,0.000009274614,0.000027768512,0.0020015736],"genre_scores_gemma":[0.000037363196,0.97057235,0.018062305,0.0048972345,0.0055179857,0.000013692707,0.000018291596,0.0000774668,0.00080329605],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9826748,0.0018446189,0.011509027,0.0006215531,0.0029456757,0.00040432232],"domain_scores_gemma":[0.97711223,0.005826502,0.008367586,0.0011387037,0.0067824232,0.00077253446],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.026558816,0.00069875363,0.00741563,0.00086856243,0.000075382544,0.0000138208825,0.00067217177,0.0007099748,0.00089031504],"category_scores_gemma":[0.039592072,0.0003722601,0.0024944614,0.0009374417,0.0005468834,0.00030810153,0.00005597997,0.0017797089,0.000019470137],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002494689,0.0021885037,0.000077535406,0.07011123,0.0014620671,0.0006826237,0.00009027894,0.0000010118575,0.00009122872,0.000026546088,0.08782565,0.83494866],"study_design_scores_gemma":[0.0069159726,0.005916966,0.000105586405,0.20920613,0.02175607,0.00037241494,0.000119021504,0.0000034096079,0.000044432996,0.00000944073,0.75531137,0.00023919743],"about_ca_topic_score_codex":0.00011763995,"about_ca_topic_score_gemma":0.000010791487,"teacher_disagreement_score":0.83470947,"about_ca_system_score_codex":0.000084346495,"about_ca_system_score_gemma":0.0031693475,"threshold_uncertainty_score":0.9998729},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"evaluation","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2915214430","doi":"10.1111/j.1756-5391.2011.01110.x","title":"Editorial for issue 1 of 2011, Journal of Evidence Based Medicine","year":2011,"lang":"en","type":"editorial","venue":"Journal of Evidence-Based Medicine","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"MEDLINE; Cochrane Library; Health care; Alternative medicine; Evidence-based medicine; National library; Political science; Medicine; Public relations; Library science; Law; Computer science","score_opus":0.5158854550997702,"score_gpt":0.5784606755840381,"score_spread":0.06257522048426789,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2915214430","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00014931985,0.017856171,0.004288956,0.05776702,0.918416,0.0013806112,0.000047567006,0.000010365244,0.00008395425],"genre_scores_gemma":[0.0007158784,0.016776433,0.0024502196,0.0012159089,0.977537,0.000057304205,0.00002560363,0.000068885674,0.0011528094],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.97932345,0.0029141575,0.0071812775,0.0005068643,0.008708588,0.0013656896],"domain_scores_gemma":[0.8830569,0.07129046,0.014319974,0.00087083405,0.028446613,0.0020152284],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch","research_integrity"],"category_scores_codex":[0.05969933,0.000629635,0.003214807,0.0024121497,0.0005775855,0.000015878877,0.0025034097,0.0016402014,0.0056278445],"category_scores_gemma":[0.2706736,0.0003263138,0.0005324241,0.0010069666,0.0014020548,0.0013934168,0.00008717977,0.0055378918,0.000042367006],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.007509106,0.00019473278,0.00044959763,0.006805011,0.000095712894,0.000022103459,0.0016284391,0.00001800197,0.0007869069,0.000008484478,0.9803497,0.0021321701],"study_design_scores_gemma":[0.0058723805,0.017680142,0.00021295894,0.10181246,0.00057822425,0.0000065423105,0.0014159634,0.00006437192,0.00015203537,0.00041289095,0.87154704,0.00024498606],"about_ca_topic_score_codex":0.0016888093,"about_ca_topic_score_gemma":0.00007888076,"teacher_disagreement_score":0.21097428,"about_ca_system_score_codex":0.0012877851,"about_ca_system_score_gemma":0.055210993,"threshold_uncertainty_score":0.9999189},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2947332497","doi":"10.1111/jebm.12346","title":"A Chinese practice guideline of the assisted reproductive technology strategies for women with advanced age","year":2019,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Reproductive Health and Technologies","field":"Medicine","cited_by":46,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Hamilton Health Sciences; Impact","funders":"","keywords":"Guideline; Fertility; Medicine; Grading (engineering); Family medicine; Assisted reproductive technology; Delphi method; Infertility; Reproductive medicine; Population; Medical education; Pregnancy; Environmental health; Engineering","score_opus":0.06425853403512855,"score_gpt":0.4064706327631867,"score_spread":0.34221209872805813,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2947332497","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.77135766,0.005497706,0.0018997514,0.2185426,0.00049985776,0.0016253472,0.0000024300102,0.0000502389,0.00052440766],"genre_scores_gemma":[0.97274476,0.0003864689,0.024376223,0.0016580558,0.00045314938,0.000055111428,0.000001221444,0.000026000076,0.00029898022],"study_design_codex":"bench_or_experimental","study_design_gemma":"observational","domain_scores_codex":[0.9972121,0.0001043343,0.0011849699,0.00039174757,0.0007591099,0.00034774977],"domain_scores_gemma":[0.99214005,0.000993566,0.002125824,0.001085831,0.0035371033,0.00011764036],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0030438446,0.0002400073,0.0010296173,0.0005230912,0.00006283757,0.0000068695335,0.00036316284,0.00014802747,0.000039225197],"category_scores_gemma":[0.031236146,0.00009914461,0.00010350076,0.0013928778,0.00069959107,0.00050555426,0.000043057582,0.00065509713,0.0000011090717],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.09906268,0.0017597836,0.0762459,0.0053567835,0.0014499446,0.0005686959,0.0027813334,0.0015794829,0.66419244,0.0028859277,0.009614644,0.13450241],"study_design_scores_gemma":[0.059254874,0.18235797,0.23745994,0.04262249,0.002537601,0.008997114,0.113350585,0.00024923348,0.08323428,0.033563845,0.23519619,0.0011758773],"about_ca_topic_score_codex":0.000016191852,"about_ca_topic_score_gemma":0.0000037003015,"teacher_disagreement_score":0.5809581,"about_ca_system_score_codex":0.000245997,"about_ca_system_score_gemma":0.0017315492,"threshold_uncertainty_score":0.9769242},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2947494553","doi":"10.1111/jebm.12343","title":"Evaluating people's ability to assess treatment claims: Validating a test in Mandarin from Claim Evaluation Tools database","year":2019,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Psychometric Methodologies and Testing","field":"Decision Sciences","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Hamilton Health Sciences; Impact","funders":"Ministry of Science and Technology of the People's Republic of China","keywords":"Mandarin Chinese; Test (biology); Database; Computer science; Psychology; Linguistics","score_opus":0.9102972500589558,"score_gpt":0.6288915867040393,"score_spread":0.28140566335491646,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2947494553","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9718843,0.0011023126,0.016203426,0.007995904,0.0014744992,0.0010015434,0.000021611246,0.000015062252,0.00030133678],"genre_scores_gemma":[0.9336103,0.0000689669,0.06496891,0.00060853857,0.0006462821,0.000027849776,0.000009195249,0.000014816835,0.000045128378],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.98650205,0.0034891495,0.0030779722,0.0008008554,0.005674278,0.000455705],"domain_scores_gemma":[0.70325613,0.2884846,0.0030136716,0.0015788499,0.0031751385,0.00049162295],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.091621354,0.00031180336,0.0012788471,0.0014332952,0.00012188428,0.00023885896,0.0010638079,0.00011235818,0.0031343112],"category_scores_gemma":[0.6151812,0.00017170301,0.0001939765,0.0036337494,0.00007639664,0.0012480809,0.0001352852,0.0004733135,0.00005403132],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00035026902,0.00016238035,0.6041794,0.000023045282,0.000016046455,0.000022239368,0.00064054807,0.008866451,0.028929505,0.0000061478227,0.00040382496,0.35640016],"study_design_scores_gemma":[0.0068546142,0.006841865,0.8889723,0.005872721,0.00024542608,0.000033298642,0.0045773047,0.079286836,0.002605677,0.003968144,0.00040395293,0.0003378676],"about_ca_topic_score_codex":0.0007511951,"about_ca_topic_score_gemma":0.00020597293,"teacher_disagreement_score":0.52355987,"about_ca_system_score_codex":0.0009480509,"about_ca_system_score_gemma":0.001313654,"threshold_uncertainty_score":0.997777},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2949493839","doi":"10.1111/jebm.12347","title":"An evaluation of WHO emergency guidelines for Zika virus disease","year":2019,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval; Hôpital de l'Enfant-Jésus","funders":"UNICEF; Wellcome Trust; Wellcome; World Health Organization","keywords":"Guideline; Zika virus; Context (archaeology); Medicine; CLARITY; Medical emergency; Presentation (obstetrics); Critical appraisal; Checklist; Family medicine; Alternative medicine; Psychology; Pathology","score_opus":0.9321876143821779,"score_gpt":0.7733480368508203,"score_spread":0.15883957753135758,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2949493839","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.81666744,0.0016814829,0.010725914,0.16502236,0.0033092108,0.0024168105,0.00004645017,0.000013157543,0.000117208976],"genre_scores_gemma":[0.967694,0.00085500017,0.003362454,0.02459475,0.0029527987,0.00010637724,0.000006979105,0.000025570285,0.00040211368],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.9925827,0.0015727069,0.0029871135,0.00020411758,0.0022469885,0.0004063682],"domain_scores_gemma":[0.9814065,0.0030537103,0.002754426,0.00043979604,0.011696248,0.0006492889],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.029298432,0.00014370299,0.0005481645,0.0004489851,0.00019608435,0.0000028057257,0.0004127674,0.000072596165,0.0053246836],"category_scores_gemma":[0.052550934,0.000088544606,0.0001069188,0.0005067644,0.00008120087,0.00074355275,0.000018117635,0.00026337555,0.000031258278],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0028698866,0.00018696803,0.37816232,0.004095847,0.00007217257,0.0000049438268,0.008454384,0.004314592,0.068684794,0.0006652089,0.49381062,0.03867825],"study_design_scores_gemma":[0.029202957,0.011485478,0.65207356,0.03522173,0.0017135309,0.000007039308,0.011863697,0.09607743,0.0029338063,0.008248857,0.15043485,0.000737059],"about_ca_topic_score_codex":0.00018969923,"about_ca_topic_score_gemma":0.00006982395,"teacher_disagreement_score":0.34337577,"about_ca_system_score_codex":0.0002933702,"about_ca_system_score_gemma":0.005612399,"threshold_uncertainty_score":0.9995415},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2999302299","doi":"10.1111/jebm.12369","title":"Estimating patient‐reported outcomes for glaucoma management: A cross‐sectional study","year":2020,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Ophthalmology and Visual Impairment Studies","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St Joseph's Health Care; Lawson Health Research Institute; Western University","funders":"","keywords":"Medicine; Logistic regression; Glaucoma; Quality of life (healthcare); Cross-sectional study; Medical diagnosis; Ophthalmology","score_opus":0.23462447982855247,"score_gpt":0.4574047976704842,"score_spread":0.22278031784193172,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2999302299","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9751958,0.0004884139,0.0034262785,0.018736752,0.0009365803,0.0011011742,0.0000017844965,0.000031657666,0.00008151204],"genre_scores_gemma":[0.98906666,0.000016034248,0.005715323,0.004138537,0.00087412575,0.000035415083,0.000003362543,0.00002018447,0.00013037217],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9971753,0.000095338306,0.0013729144,0.00025816928,0.0008359462,0.00026236946],"domain_scores_gemma":[0.99697804,0.0009343645,0.0009326192,0.0001695596,0.0006900979,0.0002953083],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0013176039,0.0002315459,0.000790814,0.00022907983,0.00019694022,0.000016254724,0.00013827853,0.00006678259,0.00017414973],"category_scores_gemma":[0.0049940064,0.00013470842,0.00024046097,0.00030013014,0.00017837714,0.00024033923,0.000046804653,0.0003137566,0.0000048429047],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.003442626,0.00088535616,0.98894936,0.000503304,0.00095248077,0.0010006306,0.00074735325,0.00006419464,0.00015272123,0.000008858165,0.0021538155,0.0011393083],"study_design_scores_gemma":[0.010205807,0.031046903,0.9535327,0.0013245189,0.0010082035,0.00031609033,0.0010458527,0.0010693616,0.00005635831,0.00007142266,0.00020580184,0.00011699872],"about_ca_topic_score_codex":0.000008620412,"about_ca_topic_score_gemma":3.8649753e-7,"teacher_disagreement_score":0.03541667,"about_ca_system_score_codex":0.00008770225,"about_ca_system_score_gemma":0.0001270903,"threshold_uncertainty_score":0.5978653},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W3027694758","doi":"10.1111/jebm.12384","title":"Clinical research methods for treatment, diagnosis, prognosis, etiology, screening, and prevention: A narrative review","year":2020,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":41,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Juravinski Cancer Centre; McMaster University; Impact","funders":"","keywords":"Narrative review; Perspective (graphical); Etiology; Narrative; Systematic review; Medicine; Quality (philosophy); Psychology; MEDLINE; Psychotherapist; Psychiatry; Computer science; Political science","score_opus":0.9833962916424582,"score_gpt":0.8007399255518473,"score_spread":0.18265636609061098,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3027694758","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[5.560547e-7,0.91262263,0.04626661,0.033895228,0.00059617264,0.006544364,0.000014243701,0.0000035496619,0.00005666851],"genre_scores_gemma":[7.788315e-7,0.8844382,0.11025729,0.00085490226,0.0020916893,0.0011468002,0.000016114704,0.000034791185,0.0011594071],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.7735262,0.17654833,0.040374655,0.0021136792,0.006812871,0.00062427006],"domain_scores_gemma":[0.6605853,0.27914178,0.04496417,0.0032137402,0.010800372,0.0012946698],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.54152566,0.0009062331,0.030351061,0.0013826201,0.00030678418,0.00029573182,0.0028040179,0.00048723994,0.003300256],"category_scores_gemma":[0.5882525,0.0003136201,0.00961421,0.0035462684,0.0008709453,0.00040659323,0.00017994216,0.0011912155,0.000074538504],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000032526525,0.00008084111,0.00024221816,0.014584968,0.0011498558,0.000026369284,0.00008927618,1.2984263e-7,5.3345342e-8,0.000033390643,0.23821732,0.74554306],"study_design_scores_gemma":[0.0005768561,0.0050553065,0.000021506577,0.25527126,0.01016753,0.00014105883,0.00017169905,0.000035008634,1.9481675e-7,0.0006396557,0.72774094,0.00017900206],"about_ca_topic_score_codex":0.0000050788667,"about_ca_topic_score_gemma":0.0000059295985,"teacher_disagreement_score":0.745364,"about_ca_system_score_codex":0.00020148174,"about_ca_system_score_gemma":0.0025528118,"threshold_uncertainty_score":0.9999316},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3034590128","doi":"10.1111/jebm.12393","title":"The association between allergic rhinitis and risk of rheumatoid arthritis: A systematic review and meta‐analysis","year":2020,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Allergic Rhinitis and Sensitization","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Meta-analysis; Odds ratio; Rheumatoid arthritis; Confidence interval; Relative risk; Internal medicine; Cohort study; Case-control study; MEDLINE","score_opus":0.11339027055341291,"score_gpt":0.349170584944384,"score_spread":0.2357803143909711,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3034590128","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000007262512,0.98875296,0.00031679895,0.009242392,0.00006870511,0.0015413975,0.000036339523,0.000009769053,0.000024385776],"genre_scores_gemma":[0.00027704798,0.9986362,0.0001632789,0.00053748826,0.00022752257,0.0000323804,0.000028881213,0.000034804812,0.00006240775],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.99231195,0.0019596536,0.0038174107,0.0002961087,0.0013896084,0.0002252447],"domain_scores_gemma":[0.98084205,0.00844021,0.008841483,0.00038527386,0.0011323105,0.00035867118],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.007244861,0.0004719831,0.012730525,0.00044967842,0.0001442311,0.000037103247,0.00019977168,0.0002536593,0.00006958865],"category_scores_gemma":[0.015974035,0.000210473,0.001686708,0.0010831796,0.00019076822,0.00016066854,0.000039952552,0.00084538,0.0000023883626],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000034305333,0.000025672889,0.00019212023,0.6845174,0.28812033,0.00018811905,0.00016099206,0.0000011864489,1.2216687e-7,0.000021803617,0.0013869185,0.02535102],"study_design_scores_gemma":[0.0004865525,0.0005157874,0.00004480779,0.3881654,0.58033574,0.00032368198,0.00005012473,0.000014813433,1.6341184e-7,0.000017599334,0.02992408,0.00012123779],"about_ca_topic_score_codex":0.00004173272,"about_ca_topic_score_gemma":0.000017404096,"teacher_disagreement_score":0.296352,"about_ca_system_score_codex":0.00016392188,"about_ca_system_score_gemma":0.0005568623,"threshold_uncertainty_score":0.9923148},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W3095207500","doi":"10.1111/jebm.12412","title":"Association between homocysteine and obesity: A meta‐analysis","year":2020,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Folate and B Vitamins Research","field":"Medicine","cited_by":64,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"Yunnan Provincial Science and Technology Department; National Natural Science Foundation of China","keywords":"Homocysteine; Medicine; Meta-analysis; Internal medicine; Obesity; Cochrane Library; Insulin resistance; Confidence interval; Strictly standardized mean difference; Endocrinology","score_opus":0.4488428679630203,"score_gpt":0.4667018673714734,"score_spread":0.01785899940845309,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3095207500","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00004330122,0.97755754,0.00026725387,0.021310981,0.00010243195,0.00061521796,0.00002211815,0.000016764532,0.00006440054],"genre_scores_gemma":[0.00034456293,0.9949504,0.0003719493,0.000489132,0.002356578,0.000013952588,0.000045921515,0.000046253324,0.0013812481],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9939369,0.00078960496,0.0020846713,0.00041384553,0.00237228,0.00040267472],"domain_scores_gemma":[0.99066174,0.004497972,0.0024915102,0.00041327727,0.0010905057,0.00084497465],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0050265226,0.00053338165,0.009962489,0.0020512945,0.00007854329,0.000043165357,0.00033302463,0.0004774626,0.00061767426],"category_scores_gemma":[0.0063335034,0.00026629656,0.0029769882,0.0029310307,0.00014006783,0.00024067538,0.00007493433,0.0017495933,0.000026780885],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015958937,0.0001271598,0.0029864637,0.03290238,0.54469144,0.0020745213,0.00009257484,0.0000011657297,0.000008053829,0.0000036193462,0.007945421,0.4090076],"study_design_scores_gemma":[0.000690377,0.0017125205,0.00021573767,0.014797617,0.524861,0.00011470392,0.000012548234,0.000010641737,0.0000029345665,0.000004995404,0.45744503,0.0001319112],"about_ca_topic_score_codex":0.00006190048,"about_ca_topic_score_gemma":0.000008013489,"teacher_disagreement_score":0.4494996,"about_ca_system_score_codex":0.000602469,"about_ca_system_score_gemma":0.0015715001,"threshold_uncertainty_score":0.9999789},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W3100141195","doi":"10.1111/jebm.12418","title":"Analytics with artificial intelligence to advance the treatment of acute respiratory distress syndrome","year":2020,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Respiratory Support and Mechanisms","field":"Medicine","cited_by":47,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; University of Alberta Hospital; University of Alberta","funders":"National Natural Science Foundation of China","keywords":"ARDS; Analytics; Medicine; Population; Artificial intelligence; Big data; Identification (biology); Computer science; Intensive care medicine; Machine learning; Data science; Data mining; Internal medicine; Lung","score_opus":0.2641202297817875,"score_gpt":0.4285162414375897,"score_spread":0.1643960116558022,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3100141195","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00019592943,0.98025143,0.01310181,0.0043568728,0.000542504,0.0014715978,0.00004353584,0.000014170063,0.000022178461],"genre_scores_gemma":[0.00063351804,0.995491,0.0013271725,0.0010532138,0.0012509775,0.000044745495,0.000012290203,0.000071940594,0.00011509464],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9949851,0.00025115197,0.0024710246,0.00042257004,0.0014971851,0.0003729659],"domain_scores_gemma":[0.99422604,0.00096629525,0.0024330795,0.00084069,0.00075918826,0.00077473186],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0010029528,0.00065543223,0.0041813836,0.00059337745,0.000073748655,0.000017454531,0.00057090574,0.00020381872,0.00018103352],"category_scores_gemma":[0.000660597,0.0002638245,0.00079137576,0.0013633586,0.00037573962,0.00013860245,0.000028306698,0.0006253653,0.0000167995],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001258689,0.00030099845,0.00004282465,0.00482628,0.0013091688,0.0074094757,0.00016988188,0.00003317673,0.00006076211,0.000077571596,0.0003812678,0.9841299],"study_design_scores_gemma":[0.00039311234,0.037600443,0.000010720893,0.09303724,0.01326275,0.001261835,0.00021631748,0.000011614002,0.00045352144,0.000035280784,0.8534769,0.0002403038],"about_ca_topic_score_codex":0.0000055836426,"about_ca_topic_score_gemma":0.0000047604876,"teacher_disagreement_score":0.9838896,"about_ca_system_score_codex":0.00047440373,"about_ca_system_score_gemma":0.002649095,"threshold_uncertainty_score":0.9999814},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W3117672471","doi":"10.1111/j.1756-5391.2010.01086.x","title":"Revised STandards for Reporting Interventions in Clinical Trials of Acupuncture (STRICTA): Extending the CONSORT Statement","year":2010,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Acupuncture Treatment Research Studies","field":"Medicine","cited_by":443,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital; University of Ottawa","funders":"National Institute for Health and Care Research","keywords":"Checklist; Consolidated Standards of Reporting Trials; Psychological intervention; Clinical trial; Systematic review; Psychology; Alternative medicine; Medicine; Medical education; Family medicine; MEDLINE; Political science; Psychiatry; Pathology","score_opus":0.7114688404628366,"score_gpt":0.651952867294202,"score_spread":0.05951597316863466,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3117672471","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6649062,0.044081405,0.032409757,0.24561869,0.0037615756,0.00882486,0.00009931976,0.00003719709,0.00026100522],"genre_scores_gemma":[0.9865871,0.005298365,0.005568483,0.0006123708,0.0015927528,0.00007249289,0.000010783211,0.00002374531,0.00023387803],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.98659235,0.0010126932,0.009918773,0.00027032575,0.0018364711,0.0003693755],"domain_scores_gemma":[0.9503417,0.030415576,0.014282397,0.0006838232,0.0038821332,0.00039437902],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.105609804,0.00035428687,0.003242849,0.0005350381,0.000106661784,0.000015960934,0.00025401893,0.00013718988,0.00040973563],"category_scores_gemma":[0.4326503,0.00010131595,0.0016498322,0.00047491628,0.00061447703,0.00017114545,0.000051551357,0.0011941815,3.683242e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.022029387,0.0019904338,0.5567662,0.007474781,0.0034196451,0.0017027963,0.0012052383,0.00007713186,0.05294873,0.00040103967,0.17065229,0.18133229],"study_design_scores_gemma":[0.07786334,0.04724757,0.46476266,0.18099225,0.018597322,0.0014160465,0.006655577,0.0006801456,0.021717336,0.0019198068,0.17737336,0.0007745918],"about_ca_topic_score_codex":0.000026653404,"about_ca_topic_score_gemma":0.000045596305,"teacher_disagreement_score":0.3270405,"about_ca_system_score_codex":0.00029548074,"about_ca_system_score_gemma":0.0018513663,"threshold_uncertainty_score":0.9209629},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"reporting","study_design":"design_other","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3128250231","doi":"10.1111/jebm.12419","title":"Consistency of recommendations and methodological quality of guidelines for the diagnosis and treatment of COVID‐19","year":2021,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Clinical practice guidelines implementation","field":"Medicine","cited_by":29,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Hamilton Health Sciences; Impact","funders":"","keywords":"Guideline; Interim; Medicine; CLARITY; Consistency (knowledge bases); Coronavirus disease 2019 (COVID-19); MEDLINE; Checklist; Quality (philosophy); Concordance; Data collection; Medical physics; Family medicine; Psychology; Computer science; Pathology; Statistics; Political science; Internal medicine","score_opus":0.9688802759637457,"score_gpt":0.738142788820996,"score_spread":0.23073748714274966,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3128250231","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000154535,0.874626,0.007990737,0.115577355,0.00027219797,0.0012754801,0.00009782749,0.0000025551476,0.0000033491895],"genre_scores_gemma":[0.00014541995,0.95438457,0.04354593,0.0012607148,0.00050428935,0.0000691037,0.000051605704,0.000016197417,0.0000221696],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9901848,0.0019824805,0.006800352,0.0002649811,0.00062785496,0.00013955771],"domain_scores_gemma":[0.83551466,0.15235585,0.007837146,0.00042157248,0.0035891812,0.00028160893],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.016763262,0.0002912842,0.0043727127,0.00029065806,0.00005901894,0.000005259652,0.00014014773,0.0001935119,0.00028137796],"category_scores_gemma":[0.3261049,0.00013047231,0.00072601734,0.00036405632,0.0006493598,0.000102775564,0.00004131565,0.00025296432,7.131941e-8],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005400826,0.00033224438,0.0013173786,0.0135724,0.00094138866,0.000011098391,0.00011430749,0.000003031281,0.000031059753,0.00010277778,0.007657442,0.9753768],"study_design_scores_gemma":[0.005940962,0.009488078,0.0005350578,0.07912947,0.022307876,0.00032553013,0.0020224028,0.000031574444,0.000057993097,0.00022663733,0.87977886,0.0001555482],"about_ca_topic_score_codex":0.000370635,"about_ca_topic_score_gemma":0.000049794373,"teacher_disagreement_score":0.9752212,"about_ca_system_score_codex":0.00015393866,"about_ca_system_score_gemma":0.0025923403,"threshold_uncertainty_score":0.67957157},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"evaluation","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W3161745299","doi":"10.1111/jebm.12427","title":"The ROBINS‐I and the NOS had similar reliability but differed in applicability: A random sampling observational studies of systematic reviews/meta‐analysis","year":2021,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":126,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Observational study; Meta-analysis; Statistics; Statistic; Reliability (semiconductor); Medicine; Systematic review; MEDLINE; Mathematics; Internal medicine","score_opus":0.96789418365725,"score_gpt":0.6353889867762997,"score_spread":0.33250519688095026,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3161745299","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00006394702,0.97588885,0.0077587184,0.0093998145,0.00027938286,0.006583275,0.000013806544,0.0000016103637,0.0000106235975],"genre_scores_gemma":[0.0004474575,0.9963246,0.0014849005,0.00039018822,0.00023418105,0.0007467406,0.0000048521315,0.00001887611,0.0003482062],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.79905677,0.12443673,0.06276263,0.0014977312,0.011779723,0.00046641592],"domain_scores_gemma":[0.255356,0.64355075,0.08066415,0.009041565,0.011017264,0.00037026877],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch","metaepi_broad"],"category_scores_codex":[0.57151395,0.000988608,0.06604463,0.0011335632,0.00030741227,0.000362098,0.00355192,0.00023711057,0.0005029948],"category_scores_gemma":[0.6919101,0.00023754321,0.020754656,0.006824356,0.0013464021,0.00029002098,0.00027139217,0.0010724037,0.00001123255],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00041166818,0.00031538776,0.0011909404,0.6632034,0.24971013,0.00004689022,0.0013801714,0.001275674,0.0000023007262,0.00096672925,0.006015617,0.07548108],"study_design_scores_gemma":[0.0016564076,0.00014461514,0.00010928993,0.21553178,0.66173595,0.0000567301,0.0012310822,0.0012066741,2.8043323e-7,0.0020228392,0.11598145,0.00032287266],"about_ca_topic_score_codex":0.000028629614,"about_ca_topic_score_gemma":0.000070949965,"teacher_disagreement_score":0.54370075,"about_ca_system_score_codex":0.00022645656,"about_ca_system_score_gemma":0.00089288445,"threshold_uncertainty_score":0.98905224},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"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,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Joseph’s Healthcare Hamilton; McMaster University; 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","score_opus":0.4126879839645758,"score_gpt":0.4530451666427994,"score_spread":0.040357182678223635,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3192941963","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06696076,0.92504597,0.00002864152,0.004064195,0.00031459786,0.003576755,0.0000030604488,0.0000033912756,0.0000026280004],"genre_scores_gemma":[0.06112304,0.9376012,0.0004967728,0.00041745626,0.00015300167,0.00010538545,0.0000024145488,0.000024855202,0.00007587598],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9936243,0.0022019346,0.0028932043,0.00031655785,0.0007198243,0.00024411987],"domain_scores_gemma":[0.9640715,0.032305624,0.002402266,0.0003535662,0.0005691212,0.00029790783],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005626046,0.00049764523,0.009934409,0.00056723866,0.000046442678,0.0000132371615,0.00014251452,0.00012817461,0.0000127400135],"category_scores_gemma":[0.019548103,0.00020372853,0.00059209426,0.00033205366,0.00039301027,0.00011607902,0.00007535798,0.0005239042,4.7287023e-7],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.2870779,0.0010979978,0.0012137223,0.39266878,0.0031533996,0.0033663078,0.004840796,0.00007961935,0.00029116674,0.00008149346,0.00020475202,0.30592406],"study_design_scores_gemma":[0.08447317,0.009027326,0.00006670343,0.898142,0.0055692657,0.0006854057,0.00025827446,0.0001972598,0.00024560103,0.000012937215,0.0011538452,0.00016821148],"about_ca_topic_score_codex":0.00019053646,"about_ca_topic_score_gemma":0.000022222697,"teacher_disagreement_score":0.5054732,"about_ca_system_score_codex":0.00029581325,"about_ca_system_score_gemma":0.0006914561,"threshold_uncertainty_score":0.98871064},"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":"W3198694028","doi":"10.1111/jebm.12443","title":"Progress and challenges of network meta‐analysis","year":2021,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":35,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Guideline; Management science; Medicine; Engineering; Pathology","score_opus":0.9673007878232126,"score_gpt":0.6345168843397608,"score_spread":0.33278390348345177,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3198694028","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000012340748,0.989364,0.001845792,0.007621468,0.0004056024,0.00061949674,0.0000066906528,0.0000014746931,0.00013423803],"genre_scores_gemma":[0.00006217131,0.99379224,0.0044734925,0.00014783064,0.0009584122,0.000027574468,0.0000038077776,0.000023001367,0.0005114764],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9288904,0.030054014,0.028474525,0.0010556241,0.011163632,0.00036180654],"domain_scores_gemma":[0.9019688,0.04077169,0.0477191,0.0034945922,0.0055694957,0.0004763516],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","insufficient_payload"],"consensus_categories":["metaresearch","metaepi_broad"],"category_scores_codex":[0.27131426,0.0008134356,0.053479068,0.0024561083,0.00007841006,0.00022120669,0.002637185,0.00030984162,0.008317138],"category_scores_gemma":[0.07922479,0.00026179763,0.021033874,0.006530784,0.00042347275,0.00032434886,0.00015900556,0.00066366646,0.000026787233],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000007811445,0.00005545213,0.00010023519,0.014365071,0.1888975,0.0001356716,0.00008651608,0.000088691064,3.686598e-8,0.00036273408,0.0124278255,0.7834725],"study_design_scores_gemma":[0.000085924235,0.00020332968,0.000019090632,0.020806262,0.43474832,0.00008051155,0.00010453036,0.00004599213,7.864684e-8,0.00015499648,0.54362196,0.00012901753],"about_ca_topic_score_codex":0.000004255513,"about_ca_topic_score_gemma":0.000013098204,"teacher_disagreement_score":0.78334343,"about_ca_system_score_codex":0.00004927113,"about_ca_system_score_gemma":0.00089172943,"threshold_uncertainty_score":0.99998343},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4200188781","doi":"10.1111/jebm.12461","title":"Aerosol generated by dental procedures: A scoping review","year":2021,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Dental Research and COVID-19","field":"Dentistry","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"Coordenação de Aperfeiçoamento de Pessoal de Nível Superior","keywords":"Aerosol; Protocol (science); Coronavirus disease 2019 (COVID-19); Environmental science; Pandemic; Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); Contamination; Environmental health; Meteorology; Geography; Medicine; Infectious disease (medical specialty); Biology; Pathology; Alternative medicine","score_opus":0.287049416494756,"score_gpt":0.501716132993358,"score_spread":0.21466671649860203,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4200188781","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000009575206,0.99554044,0.00046686435,0.00089703425,0.0012590137,0.0016936833,0.00003600178,0.00002703642,0.00007034339],"genre_scores_gemma":[0.000008561849,0.9922359,0.00011307477,0.004188915,0.0014770769,0.000108464985,0.000148607,0.00009445583,0.0016249415],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9907139,0.0013466977,0.0033575355,0.0006219537,0.003247772,0.00071212166],"domain_scores_gemma":[0.99013865,0.0038779397,0.0030233252,0.00069721555,0.0011035424,0.0011593049],"candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0038150963,0.0007588997,0.0048848465,0.0005170326,0.0001581659,0.00014347049,0.0013405475,0.0003684441,0.0051270346],"category_scores_gemma":[0.0459726,0.00044623518,0.0013433523,0.0016425668,0.00029102503,0.0004996584,0.00016644449,0.0016901749,0.00021833133],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000057351903,0.00012386861,0.00000930402,0.39774913,0.0002958775,0.0048430013,0.0000063326656,5.856044e-7,0.00005565963,0.0000034150917,0.449829,0.14702648],"study_design_scores_gemma":[0.00089154695,0.0005315172,2.3252242e-7,0.749415,0.001753631,0.003513854,0.000023484936,0.0000012733589,0.000032268636,0.0000031293976,0.24353254,0.0003015531],"about_ca_topic_score_codex":0.00004588189,"about_ca_topic_score_gemma":0.00011460883,"teacher_disagreement_score":0.35166585,"about_ca_system_score_codex":0.0008971628,"about_ca_system_score_gemma":0.01614568,"threshold_uncertainty_score":0.99979895},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4223485503","doi":"10.1111/jebm.12466","title":"An extension of the RIGHT statement for introductions and interpretations of clinical practice guidelines: RIGHT for INT","year":2022,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Clinical practice guidelines implementation","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Checklist; Guideline; Brainstorming; Medical education; Delphi method; Delphi; Psychology; Medicine; Family medicine; Computer science; Pathology","score_opus":0.7615709211221467,"score_gpt":0.6777831758827233,"score_spread":0.08378774523942345,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4223485503","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00006318912,0.8647852,0.044732966,0.08196307,0.0042166966,0.004092988,0.00012621735,0.000008591059,0.000011065505],"genre_scores_gemma":[0.00012254508,0.92412055,0.06882352,0.0030407174,0.0035341878,0.00013699723,0.00011761284,0.00005213744,0.000051766525],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9844867,0.0013957624,0.012092365,0.00047121593,0.0013307295,0.00022325375],"domain_scores_gemma":[0.9260974,0.047411278,0.016255155,0.0010127538,0.008900287,0.00032315904],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.028264815,0.0003611811,0.0032207912,0.0005143572,0.0001568516,0.000012162149,0.00042854075,0.00020043395,0.00040023497],"category_scores_gemma":[0.33311707,0.00017947816,0.0010937017,0.00056948164,0.00044414052,0.00058773026,0.00008915206,0.0010285705,3.471208e-7],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0026720045,0.0009156977,0.00002751376,0.012611338,0.0009127808,0.000010708441,0.00015948566,0.00007162705,0.00004599353,0.00018794586,0.095695585,0.8866893],"study_design_scores_gemma":[0.0035856832,0.009408448,0.000036734484,0.03907415,0.017663896,0.0003625385,0.0009622022,0.0006859202,0.000014124471,0.00025767184,0.9278081,0.00014052691],"about_ca_topic_score_codex":0.00005267343,"about_ca_topic_score_gemma":0.000016327422,"teacher_disagreement_score":0.8865488,"about_ca_system_score_codex":0.00025897718,"about_ca_system_score_gemma":0.0041591562,"threshold_uncertainty_score":0.97960776},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"reporting","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4241679221","doi":"10.1111/jebm.12104","title":"Hip protectors for preventing hip fractures in older people","year":2014,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Bone health and osteoporosis research","field":"Medicine","cited_by":50,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Hip fracture; Older people; Medicine; Physical medicine and rehabilitation; Gerontology; Osteoporosis; Internal medicine","score_opus":0.10743361762205914,"score_gpt":0.4192378048710034,"score_spread":0.3118041872489443,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4241679221","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8862837,0.0061031324,0.006454184,0.09841275,0.0005320223,0.0020665885,0.0000010363111,0.000019939274,0.00012666966],"genre_scores_gemma":[0.9915979,0.00023522691,0.0022882468,0.0037026308,0.001801536,0.000062527535,0.0000034031536,0.000030355633,0.00027818984],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99680066,0.00026328384,0.0011420597,0.00022458186,0.0010341902,0.000535209],"domain_scores_gemma":[0.9953334,0.0024753117,0.0006441696,0.0002767151,0.00071708905,0.0005533398],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.006309673,0.00019453766,0.00084774283,0.00066443183,0.00008582108,0.000013771476,0.00019505397,0.00015639752,0.00035280277],"category_scores_gemma":[0.01700872,0.000112229834,0.00018334841,0.0004969383,0.000101851554,0.00021244906,0.00001860438,0.00081681064,0.000004368659],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.014030209,0.0013425051,0.5128237,0.014559034,0.00015154862,0.00022146567,0.0037370143,0.00028889225,0.067138046,0.00006996591,0.05866656,0.32697105],"study_design_scores_gemma":[0.024029767,0.013835612,0.88058776,0.03599247,0.0003229738,0.00026414584,0.00086588983,0.0033515235,0.019687157,0.00061011664,0.02012557,0.00032702705],"about_ca_topic_score_codex":0.00015162914,"about_ca_topic_score_gemma":0.00007144232,"teacher_disagreement_score":0.36776406,"about_ca_system_score_codex":0.00018841027,"about_ca_system_score_gemma":0.0010197612,"threshold_uncertainty_score":0.99127144},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4244511973","doi":"10.1111/j.1756-5391.2010.01095.x","title":"Medical versus surgical management for gastro‐oesophageal reflux disease (GORD) in adults","year":2010,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Gastroesophageal reflux and treatments","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Heartburn; Regurgitation (circulation); Reflux; Disease; Swallowing; Proton-pump inhibitor; Gastroenterology; Gastric acid; Internal medicine; Gastro-; Stomach; Surgery","score_opus":0.06964112816479469,"score_gpt":0.38100436655896613,"score_spread":0.31136323839417146,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4244511973","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9738908,0.001598871,0.00047362776,0.020137576,0.0023770428,0.0009989005,0.00000906794,0.000029663823,0.0004844778],"genre_scores_gemma":[0.9925215,0.00042419982,0.0036706356,0.00074176886,0.0021905124,0.000048821006,0.000025019252,0.000036563582,0.00034095696],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9957098,0.00010408438,0.0012091562,0.00037555248,0.0020227355,0.0005787081],"domain_scores_gemma":[0.9957693,0.0012549945,0.00045558158,0.0004337528,0.000440251,0.0016461143],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0021093644,0.00034183817,0.0008828667,0.000593439,0.000066884866,0.000012790233,0.0003392172,0.00016792171,0.0009747218],"category_scores_gemma":[0.0037189792,0.00019989743,0.0003449967,0.0003767792,0.00031168383,0.0002354356,0.000044468896,0.00090390944,0.000016913196],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.107038446,0.0033042843,0.6482876,0.0024051105,0.0006845295,0.14728165,0.00019496225,0.000038881986,0.0001982745,0.00091599295,0.007285573,0.08236471],"study_design_scores_gemma":[0.15382096,0.0195456,0.7955596,0.019739337,0.0019583213,0.0011039701,0.00018713178,0.0013372086,0.0003502903,0.00034611492,0.005763542,0.0002879547],"about_ca_topic_score_codex":0.00001021799,"about_ca_topic_score_gemma":0.0000136236695,"teacher_disagreement_score":0.14727198,"about_ca_system_score_codex":0.00019537422,"about_ca_system_score_gemma":0.0006787586,"threshold_uncertainty_score":0.99993855},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W4283066010","doi":"10.1111/jebm.12475","title":"More work is needed on cost‐utility analyses of robotic‐assisted surgery","year":2022,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Prostate Cancer Diagnosis and Treatment","field":"Medicine","cited_by":22,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"National Social Science Fund of China","keywords":"Medicine; Robotic surgery; Surgery; General surgery; Prostatectomy; Internal medicine; Prostate cancer","score_opus":0.637030306476812,"score_gpt":0.5034306589085736,"score_spread":0.1335996475682384,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4283066010","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00022733751,0.98650694,0.000045040844,0.010446622,0.0014271486,0.0012102755,0.000039956925,0.00001687051,0.000079837795],"genre_scores_gemma":[0.0014536877,0.99444133,0.00014541538,0.0024750615,0.001050246,0.000110936475,0.00010705744,0.000072311595,0.00014395651],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9928271,0.0006364463,0.0033547424,0.0005135384,0.0022183363,0.00044981812],"domain_scores_gemma":[0.9842956,0.008938399,0.0044836686,0.0009623682,0.0006954138,0.00062456116],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0028145157,0.00074277056,0.007166942,0.0014823984,0.0000968364,0.000014969028,0.0003657621,0.000257938,0.003341211],"category_scores_gemma":[0.0040885056,0.00039612644,0.0023755094,0.0021483263,0.0003831946,0.00013049472,0.000056017096,0.0012373653,0.000007200717],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013088924,0.0009796885,0.0019431957,0.018683316,0.002239827,0.0014869472,0.00009534764,0.000025793495,0.0000029677997,7.986228e-7,0.078077674,0.89515555],"study_design_scores_gemma":[0.0017467826,0.0030214225,0.0021272337,0.3083016,0.016900863,0.00026117772,0.00017125068,0.000005931703,0.00010776504,0.000004538933,0.66707605,0.00027539989],"about_ca_topic_score_codex":0.00012833915,"about_ca_topic_score_gemma":0.0000014409704,"teacher_disagreement_score":0.8948802,"about_ca_system_score_codex":0.0013109448,"about_ca_system_score_gemma":0.0037269895,"threshold_uncertainty_score":0.9998491},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4283766318","doi":"10.1111/jebm.12476","title":"Malignancy risk stratification for solitary pulmonary nodule: A clinical practice guideline","year":2022,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Lung Cancer Diagnosis and Treatment","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Malignancy; Medicine; Risk stratification; Nodule (geology); Guideline; Radiology; Family history; Clinical Practice; Internal medicine; Pathology; Family medicine","score_opus":0.16581202019956523,"score_gpt":0.4598695820264167,"score_spread":0.2940575618268515,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4283766318","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.053980384,0.10241079,0.025174709,0.8080401,0.0055172774,0.003449727,0.00012795038,0.00007063317,0.0012284528],"genre_scores_gemma":[0.9166702,0.0115703875,0.04448341,0.021159755,0.0053775506,0.00024288731,0.000101949816,0.00005554549,0.00033834015],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9954594,0.00048703532,0.0022128497,0.00033780854,0.0012355789,0.00026734645],"domain_scores_gemma":[0.99162924,0.003975516,0.0020397743,0.0004748063,0.0015251304,0.00035551641],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.007647634,0.00021385768,0.0007816017,0.00024896182,0.00022990128,0.000010390708,0.00021764732,0.00007169138,0.00063398475],"category_scores_gemma":[0.009915962,0.00013332878,0.00038470552,0.00035353427,0.00012866383,0.00033885112,0.00003458578,0.0006484851,0.0000041919866],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.019389553,0.006040151,0.088747315,0.0006525872,0.001355184,0.0026167561,0.00039062867,0.001447694,0.0013244642,0.000428606,0.6544957,0.22311133],"study_design_scores_gemma":[0.02290509,0.050220385,0.05966842,0.007535517,0.010993032,0.0035778196,0.003405381,0.006847684,0.0008033344,0.0015889722,0.8319703,0.00048410927],"about_ca_topic_score_codex":0.00012673014,"about_ca_topic_score_gemma":0.000004173188,"teacher_disagreement_score":0.8626898,"about_ca_system_score_codex":0.00062223495,"about_ca_system_score_gemma":0.0014827368,"threshold_uncertainty_score":0.99842393},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W4284698039","doi":"10.1111/jebm.12481","title":"Factors associated with concurrent malignancy risk among patients with incidental solitary pulmonary nodule: A systematic review taskforce for developing rapid recommendations","year":2022,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Lung Cancer Diagnosis and Treatment","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"Sichuan Province Science and Technology Support Program","keywords":"Medicine; Malignancy; Lung cancer; Odds ratio; Nodule (geology); Confidence interval; Checklist; Family history; Internal medicine; Radiology","score_opus":0.2008800434411215,"score_gpt":0.38444564758022526,"score_spread":0.18356560413910375,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4284698039","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00029327365,0.98877144,0.00027122462,0.0012890804,0.0006799923,0.008361622,0.00029405742,0.000027488546,0.00001185143],"genre_scores_gemma":[0.0014230564,0.99412805,0.00017854312,0.00066789833,0.00024423306,0.0013456523,0.0018404362,0.00012333867,0.000048795646],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9931556,0.0009397011,0.003097023,0.000542045,0.0017771786,0.0004884681],"domain_scores_gemma":[0.9852898,0.005116026,0.007542063,0.0005409409,0.0010964801,0.00041466602],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0025105104,0.00090199884,0.006299243,0.0006080246,0.00030021532,0.000023382949,0.00040254302,0.00016216525,0.00037433713],"category_scores_gemma":[0.004993395,0.00040686954,0.0008843143,0.0009760133,0.00017982791,0.00036439011,0.000056838242,0.00094943406,0.0000013380113],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00016681387,0.000863522,0.019357512,0.91112417,0.0056180484,0.00024363647,0.00012020284,0.0000025530223,1.8556033e-8,0.0000045122597,0.0058315913,0.056667417],"study_design_scores_gemma":[0.0033609534,0.0054518892,0.0023146952,0.9095308,0.05259623,0.000081644175,0.0001284818,0.0000058296896,5.6617085e-7,0.0000019622323,0.026108986,0.00041794148],"about_ca_topic_score_codex":0.000045763954,"about_ca_topic_score_gemma":0.000021289503,"teacher_disagreement_score":0.056249473,"about_ca_system_score_codex":0.004315818,"about_ca_system_score_gemma":0.0027315607,"threshold_uncertainty_score":0.9998383},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W4289132654","doi":"10.1111/jebm.12483","title":"Clinical manifestations of COVID‐19: An overview of 102 systematic reviews with evidence mapping","year":2022,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"COVID-19 Clinical Research Studies","field":"Medicine","cited_by":37,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Medicine; Cochrane Library; Meta-analysis; MEDLINE; Systematic review; Coronavirus disease 2019 (COVID-19); Internal medicine; Web of science; Disease; Infectious disease (medical specialty)","score_opus":0.8614589479472008,"score_gpt":0.6529562840453669,"score_spread":0.20850266390183392,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4289132654","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000013438209,0.9853114,0.0020079107,0.0067943595,0.00041480944,0.005394042,0.00002009891,0.000018345483,0.000025630374],"genre_scores_gemma":[0.00011097715,0.9935884,0.0025019557,0.0023797904,0.00086903706,0.00026250992,0.000016726484,0.00007598893,0.0001946066],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9728775,0.008794836,0.012589605,0.0007593715,0.0044355243,0.00054315233],"domain_scores_gemma":[0.68962157,0.27452788,0.026025428,0.0031493297,0.0035436426,0.003132129],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.055618107,0.0008010268,0.014975359,0.0015883511,0.00016085137,0.000015205552,0.0014204097,0.00030398185,0.00214188],"category_scores_gemma":[0.64809436,0.00039258902,0.0023508468,0.0025402033,0.0014114482,0.00046376753,0.0002669457,0.0026684166,0.000010093093],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00049254973,0.0004918118,0.00042064735,0.8712119,0.0010196564,0.0012986717,0.00022155553,0.000021711217,0.000002855775,0.000028278573,0.0017276913,0.12306269],"study_design_scores_gemma":[0.0012181305,0.009020588,0.00009419899,0.61243224,0.009729881,0.0003810406,0.00021339764,0.000029290208,1.8775988e-7,0.000024240591,0.3666558,0.00020099274],"about_ca_topic_score_codex":0.00010762454,"about_ca_topic_score_gemma":0.00001309598,"teacher_disagreement_score":0.59247625,"about_ca_system_score_codex":0.0010822753,"about_ca_system_score_gemma":0.017528016,"threshold_uncertainty_score":0.9998526},"labels":[{"model":"gemma","categories":["metaepi_broad","metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4289520494","doi":"10.1111/jebm.12484","title":"The reporting quality of N‐of‐1 trials and protocols still needs improvement","year":2022,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Hamilton Health Sciences; Impact","funders":"National Social Science Fund of China","keywords":"Checklist; Consolidated Standards of Reporting Trials; Protocol (science); Medicine; Clinical trial; Sample size determination; Quality (philosophy); Informed consent; Family medicine; Medical physics; Alternative medicine; Psychology; Statistics; Internal medicine; Pathology; Mathematics","score_opus":0.9645233545769438,"score_gpt":0.6807940684084278,"score_spread":0.283729286168516,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4289520494","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8911448,0.019914193,0.017322192,0.05000352,0.0012040859,0.01993064,0.000014783356,0.0000028931277,0.00046288743],"genre_scores_gemma":[0.9968017,0.00011591485,0.0011881716,0.00039700078,0.00026335564,0.000436106,2.9164212e-7,0.0000062074146,0.00079130696],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.8052113,0.050178647,0.11795465,0.0006702179,0.02557076,0.0004143899],"domain_scores_gemma":[0.67628413,0.085943855,0.22884579,0.0029164862,0.0057228315,0.00028690888],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.73450273,0.00017477859,0.006036967,0.0005223672,0.00016981097,0.0000881622,0.0012723027,0.000029473853,0.0023834857],"category_scores_gemma":[0.6565475,0.000054186534,0.0014393649,0.0014299888,0.00021590438,0.00018817597,0.00013510845,0.00029212382,0.0000013440358],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002656489,0.00046740661,0.24189848,0.0019974702,0.0017965998,0.00006676971,0.0055585057,0.0016729891,0.13180606,0.0023477706,0.09302743,0.516704],"study_design_scores_gemma":[0.016805582,0.032202262,0.16114154,0.02007869,0.0069662617,0.000498977,0.08661062,0.0074067893,0.021636449,0.04223525,0.60330385,0.0011137449],"about_ca_topic_score_codex":0.00004940763,"about_ca_topic_score_gemma":0.0000052566115,"teacher_disagreement_score":0.5155903,"about_ca_system_score_codex":0.000057732785,"about_ca_system_score_gemma":0.00053047226,"threshold_uncertainty_score":0.9985285},"labels":[],"label_agreement":null},{"id":"W4289889759","doi":"10.1111/jebm.12487","title":"A conversation on evidence‐based medicine in the COVID‐era, patient revolution, and academic career with Dr. Victor Montori","year":2022,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Healthcare cost, quality, practices","field":"Health Professions","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Population Health Research Institute; McMaster University; Impact; Nova Scotia Health Authority; Dalhousie University","funders":"","keywords":"Context (archaeology); Privilege (computing); Pandemic; Conversation; Medicine; Coronavirus disease 2019 (COVID-19); Medical education; Family medicine; Psychology; Disease; History; Law; Political science; Internal medicine; Infectious disease (medical specialty)","score_opus":0.7555222829583214,"score_gpt":0.5494773220065372,"score_spread":0.20604496095178415,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4289889759","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2951033,0.01564811,0.00065498735,0.68567955,0.0013195991,0.0014980865,0.000007685692,0.000022693765,0.00006594826],"genre_scores_gemma":[0.8835327,0.0022797654,0.00015197019,0.11229787,0.0014566217,0.00018697752,0.000007779092,0.000027191492,0.00005913751],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9823798,0.010624882,0.0022472686,0.0004104335,0.0037700864,0.000567532],"domain_scores_gemma":[0.96974754,0.024882328,0.0029941131,0.00054343307,0.0013237016,0.0005088601],"candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.027118636,0.00031618035,0.00082123704,0.00073633087,0.00085486774,0.000008131817,0.0005593939,0.00020103346,0.000934613],"category_scores_gemma":[0.03393469,0.00016462726,0.00006199851,0.0011533445,0.00045393367,0.0008251571,0.00006664308,0.00454155,0.000006337553],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.019241255,0.00037374048,0.6217441,0.0029644258,0.00012404835,0.0007661589,0.105901085,0.0018436266,0.002056252,0.0007208722,0.23564778,0.008616686],"study_design_scores_gemma":[0.023596533,0.052641142,0.34674233,0.09337781,0.0011375978,0.0003142193,0.18824849,0.003098948,0.00009636595,0.0010934442,0.28871247,0.0009406411],"about_ca_topic_score_codex":0.0024076109,"about_ca_topic_score_gemma":0.0003944338,"teacher_disagreement_score":0.5884294,"about_ca_system_score_codex":0.0020994986,"about_ca_system_score_gemma":0.0043046013,"threshold_uncertainty_score":0.99997866},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4293254834","doi":"10.1111/jebm.12485","title":"Cognitive behavior therapy for insomnia in cancer patients: a systematic review and network meta‐analysis","year":2022,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Sleep and related disorders","field":"Psychology","cited_by":34,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"National Key Research and Development Program of China","keywords":"Insomnia; Meta-analysis; Sleep onset latency; Sleep onset; Randomized controlled trial; Medicine; Confidence interval; Internal medicine; Physical therapy; Psychiatry","score_opus":0.3100531130145915,"score_gpt":0.47712576136167534,"score_spread":0.16707264834708385,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4293254834","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000005516148,0.9925113,0.00006778934,0.0007185836,0.0007295745,0.005892475,0.000045098077,0.0000070429724,0.000022675482],"genre_scores_gemma":[0.000027048402,0.99253273,0.000021529893,0.0028865817,0.00021150293,0.004118309,0.000050055187,0.000052866773,0.00009940008],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9937127,0.0020374993,0.0028487374,0.0004085741,0.00062884396,0.00036361153],"domain_scores_gemma":[0.98895496,0.0061305906,0.003985976,0.00031742142,0.00042649423,0.00018458457],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0040511154,0.0005339799,0.0093565965,0.0010699433,0.000086720815,0.000012515764,0.00040128655,0.00026966038,0.0041984213],"category_scores_gemma":[0.0015376877,0.00026371496,0.0021858097,0.002289815,0.0001409327,0.00011075585,0.000025211799,0.0009713853,0.0000026518983],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00026488985,0.0009543605,0.0002455421,0.5269572,0.3375486,0.0001346458,0.00039979463,0.000011520147,6.2013763e-9,0.000009290988,0.004129094,0.12934506],"study_design_scores_gemma":[0.00183862,0.0011225832,0.000017386943,0.31169927,0.64557683,0.00001828895,0.00010536729,8.814227e-7,3.4358327e-9,0.000004873102,0.039422713,0.00019317084],"about_ca_topic_score_codex":0.0001144013,"about_ca_topic_score_gemma":0.000028490293,"teacher_disagreement_score":0.30802822,"about_ca_system_score_codex":0.00018699694,"about_ca_system_score_gemma":0.0003452333,"threshold_uncertainty_score":0.9999815},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4296785985","doi":"10.1111/jebm.12493","title":"Endovascular thrombectomy and intravenous alteplase in patients with acute ischemic stroke due to large vessel occlusion: A clinical practice guideline","year":2022,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Acute Ischemic Stroke Management","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University; Vancouver General Hospital; University of British Columbia; Impact; University of Calgary; Hamilton General Hospital; Heart and Stroke Foundation; McMaster University","funders":"","keywords":"Medicine; Guideline; Stroke (engine); Randomized controlled trial; Intensive care medicine; Occlusion; Emergency medicine; Internal medicine","score_opus":0.03341171726631387,"score_gpt":0.3447126050320186,"score_spread":0.31130088776570475,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4296785985","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9270618,0.0028535784,0.007849785,0.059932735,0.00055066333,0.0012618059,0.000024769224,0.000022728385,0.00044211093],"genre_scores_gemma":[0.9244877,0.0007632265,0.031778596,0.041637786,0.0010013968,0.00003503966,0.000042996722,0.00006649572,0.00018676344],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99442303,0.00037536584,0.0019395113,0.00047868027,0.0023002631,0.00048313933],"domain_scores_gemma":[0.995603,0.001209472,0.0010920144,0.0005693753,0.000932842,0.00059327664],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0060822517,0.0003333013,0.0012626854,0.00063048775,0.00013170848,0.00001449988,0.00034768836,0.00008360417,0.00046154103],"category_scores_gemma":[0.005915944,0.00021419988,0.00015601562,0.00073703204,0.0001484766,0.00035965015,0.00035502826,0.0013426247,0.000004344302],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.040982362,0.0049014464,0.2186917,0.0003773626,0.0026893541,0.03549157,0.00139799,0.00045010672,0.004239686,0.000018862815,0.67161965,0.01913988],"study_design_scores_gemma":[0.063551314,0.037573166,0.12423732,0.0048153126,0.005100757,0.005196927,0.0029114075,0.00074873434,0.0007880823,0.000005244634,0.75449324,0.0005784816],"about_ca_topic_score_codex":0.000055358643,"about_ca_topic_score_gemma":0.0000053585704,"teacher_disagreement_score":0.09445439,"about_ca_system_score_codex":0.00044809986,"about_ca_system_score_gemma":0.0007178653,"threshold_uncertainty_score":0.8734818},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4324307220","doi":"10.1111/jebm.12522","title":"STandards for Reporting Interventions in Clinical Trials Of Tuina/Massage (STRICTOTM): Extending the CONSORT statement","year":2023,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Complementary and Alternative Medicine Studies","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Massage; Checklist; Consolidated Standards of Reporting Trials; Psychological intervention; Medicine; Physical therapy; Chiropractic; Clinical trial; Guideline; Alternative medicine; Intervention (counseling); Nursing; Psychology","score_opus":0.8701649821119687,"score_gpt":0.6756244896524661,"score_spread":0.19454049245950256,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4324307220","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.88412875,0.012649722,0.025624195,0.07257911,0.0026876838,0.0020994889,0.000073036295,0.000025380168,0.00013266345],"genre_scores_gemma":[0.9915122,0.005442155,0.00076308555,0.00038451934,0.0015975148,0.0000378274,0.000013999719,0.00001618389,0.00023250803],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9870164,0.000952011,0.0101623535,0.00022021974,0.0013408483,0.00030816928],"domain_scores_gemma":[0.9700735,0.018600114,0.009399343,0.00030308342,0.001438342,0.00018561837],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.10580272,0.00018672028,0.0019590731,0.0006971246,0.00010323325,0.000007097475,0.00020002644,0.000024962805,0.0003195237],"category_scores_gemma":[0.1903829,0.00008841097,0.0008482827,0.00073303655,0.00039305445,0.00016130507,0.000062156534,0.00047748696,6.0074933e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.006945222,0.0007257848,0.70227396,0.0036154957,0.0017864002,0.0015000504,0.0016593339,0.00022499464,0.0049087615,0.0005022513,0.07317646,0.20268126],"study_design_scores_gemma":[0.04416841,0.03961199,0.6488619,0.14920543,0.004932395,0.00028104574,0.027011579,0.0016782125,0.0037465882,0.0020136328,0.07810141,0.00038742987],"about_ca_topic_score_codex":0.000043463977,"about_ca_topic_score_gemma":0.000026376803,"teacher_disagreement_score":0.20229383,"about_ca_system_score_codex":0.00021922965,"about_ca_system_score_gemma":0.0007398161,"threshold_uncertainty_score":0.92076427},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"reporting","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4361267876","doi":"10.1111/jebm.12526","title":"Acupuncture for treatment of knee osteoarthritis: A clinical practice guideline","year":2023,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Acupuncture Treatment Research Studies","field":"Medicine","cited_by":34,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"McMaster University","funders":"National Science Fund for Distinguished Young Scholars","keywords":"WOMAC; Acupuncture; Medicine; Physical therapy; Osteoarthritis; Certainty; Evidence-based medicine; Alternative medicine; Guideline","score_opus":0.30302399362945776,"score_gpt":0.5266039029249645,"score_spread":0.22357990929550675,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4361267876","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.09233853,0.11770728,0.008788078,0.7692014,0.0033555997,0.0072991457,0.000075474505,0.00018816305,0.0010463587],"genre_scores_gemma":[0.71137214,0.089095816,0.14606006,0.02032494,0.02449252,0.0005443579,0.0002664598,0.00027358634,0.0075701247],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9950757,0.0002745421,0.0023209446,0.00031840435,0.0015519134,0.0004585028],"domain_scores_gemma":[0.9819646,0.012248205,0.0013370133,0.00048302667,0.0034968194,0.0004703354],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0042560305,0.00047639772,0.0023209183,0.00061404036,0.00009029067,0.0000091300535,0.00017083855,0.00018062619,0.00016535025],"category_scores_gemma":[0.04483596,0.00016770525,0.0008856084,0.00077747746,0.0004737581,0.00028595014,0.000034333785,0.000252852,0.000015040136],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.031632703,0.0017587823,0.022850571,0.0010322364,0.0027916045,0.004389369,0.001010138,0.00011796406,0.0068049636,0.00006278563,0.49752086,0.43002802],"study_design_scores_gemma":[0.04114838,0.1667183,0.011998588,0.009858039,0.0077304663,0.0011369885,0.0025852383,0.0002917338,0.0032158426,0.00027378803,0.75476885,0.00027379417],"about_ca_topic_score_codex":0.000029278315,"about_ca_topic_score_gemma":0.00000898826,"teacher_disagreement_score":0.74887645,"about_ca_system_score_codex":0.0003240393,"about_ca_system_score_gemma":0.0011938759,"threshold_uncertainty_score":0.9632098},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4362658427","doi":"10.1111/jebm.12521","title":"Reporting quality and risk of bias of Cochrane individual participant data meta‐analyses: A cross‐sectional study","year":2023,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Cochrane Library; Medicine; Systematic review; Meta-analysis; MEDLINE; Cochrane collaboration; Internal medicine","score_opus":0.9918231896452125,"score_gpt":0.7345700359268312,"score_spread":0.2572531537183813,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4362658427","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.978624,0.016696515,0.0024427013,0.0014040134,0.0003070498,0.00043875232,0.000061540544,0.0000031433424,0.000022309181],"genre_scores_gemma":[0.99743146,0.00068292243,0.0013972265,0.00008078928,0.00023056795,0.000006732869,0.0000070854176,0.000008607911,0.00015458874],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.8922699,0.027056772,0.059434514,0.0010640033,0.019811006,0.00036381645],"domain_scores_gemma":[0.7731077,0.078892104,0.13424988,0.0054584555,0.007813076,0.00047876226],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.6567453,0.0002963584,0.008389996,0.0012507397,0.0001322773,0.00018342557,0.0021134538,0.00007465283,0.0022218144],"category_scores_gemma":[0.7106444,0.000109841516,0.0016073175,0.0031469413,0.00040958318,0.0006542404,0.00031116485,0.0003766452,0.00001646092],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001244399,0.0001732227,0.9802789,0.00027691975,0.011723573,0.00006256014,0.00095981004,0.0015591049,0.0005761174,0.000010690473,0.0031087762,0.0011458852],"study_design_scores_gemma":[0.0009234232,0.00084470713,0.9540205,0.0007846198,0.035316445,0.000044692766,0.0040286714,0.0030907206,0.00013504285,0.0005469119,0.00016127017,0.00010296969],"about_ca_topic_score_codex":0.00048682187,"about_ca_topic_score_gemma":0.000116353076,"teacher_disagreement_score":0.11916216,"about_ca_system_score_codex":0.000013951065,"about_ca_system_score_gemma":0.0005004782,"threshold_uncertainty_score":0.9986903},"labels":[],"label_agreement":null},{"id":"W4378194782","doi":"10.1111/jebm.12534","title":"Introducing value‐based healthcare perspectives into hospital performance assessment: A scoping review","year":2023,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Primary Care and Health Outcomes","field":"Health Professions","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Health care; Protocol (science); Reliability (semiconductor); Process management; Process (computing); Computer science; Knowledge management; Medicine; Management science; Engineering; Alternative medicine","score_opus":0.35836885979161753,"score_gpt":0.5924324040321615,"score_spread":0.23406354424054393,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4378194782","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000006187142,0.86050856,0.00034665148,0.13042256,0.0041765566,0.004299847,0.000010533413,0.00010177451,0.00012734639],"genre_scores_gemma":[0.000013499702,0.9678656,0.0023863383,0.021709876,0.007230479,0.00036336543,0.000047441987,0.00016670911,0.00021672808],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.98538834,0.004032043,0.006307272,0.00083446206,0.0022233706,0.0012145182],"domain_scores_gemma":[0.9764819,0.0111382445,0.0079565495,0.0011489502,0.0024363154,0.0008380334],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.014768878,0.000937605,0.00745658,0.0015417206,0.0010518996,0.000020169331,0.0011952828,0.0006219888,0.00074018876],"category_scores_gemma":[0.011780032,0.00055539754,0.001087051,0.0020605084,0.00030626808,0.0006893663,0.00019690755,0.005565892,0.00014226852],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.00004106071,0.00005157745,0.00030305743,0.7311671,0.00014394244,0.00012174754,0.0006778196,0.000005124614,1.48962e-7,0.000053386415,0.036751285,0.23068379],"study_design_scores_gemma":[0.00084263063,0.001753866,0.00008058601,0.7552313,0.0016293146,0.000018546812,0.00062070205,0.000018438763,8.1800955e-8,0.000021843776,0.239413,0.00036968943],"about_ca_topic_score_codex":0.00031245808,"about_ca_topic_score_gemma":0.000043537748,"teacher_disagreement_score":0.23031409,"about_ca_system_score_codex":0.0063346503,"about_ca_system_score_gemma":0.058145627,"threshold_uncertainty_score":0.99968976},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4380870584","doi":"10.1111/jebm.12538","title":"Estimation of death risk factors associated with the coronavirus pandemic in the Middle East and North Africa","year":2023,"lang":"en","type":"letter","venue":"Journal of Evidence-Based Medicine","topic":"COVID-19 Clinical Research Studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Prince Edward Island","funders":"","keywords":"Case fatality rate; Pandemic; Socioeconomic status; Demography; Middle East; Geography; Mortality rate; Environmental health; Coronavirus disease 2019 (COVID-19); Medicine; Population; Disease; Infectious disease (medical specialty)","score_opus":0.5200755105736113,"score_gpt":0.4380864895685413,"score_spread":0.08198902100507005,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4380870584","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.37992054,0.0037682254,0.00026286172,0.6150442,0.00015138483,0.0007825742,0.00004036912,0.000016274736,0.000013597377],"genre_scores_gemma":[0.9350825,0.0055615176,0.000027281285,0.057893872,0.0011437922,0.000021856042,0.000032883334,0.000041555715,0.00019474933],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9943687,0.00096452015,0.0012089178,0.0002755662,0.0027602317,0.00042210703],"domain_scores_gemma":[0.93528163,0.06145096,0.0019567995,0.00042851662,0.00074245327,0.00013965539],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.006142185,0.00034727858,0.0013509174,0.0004117399,0.00012727531,0.000016996803,0.00047541776,0.00028560605,0.000030061186],"category_scores_gemma":[0.12173236,0.00012404534,0.00020157799,0.000960048,0.0008904746,0.0001148581,0.00006458668,0.004116081,0.0000016918626],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019773562,0.0002530453,0.7476311,0.0020843076,0.0014400339,0.0020552343,0.01036377,0.00244016,0.000026949869,0.0000013487052,0.22390762,0.007819126],"study_design_scores_gemma":[0.005519852,0.008509179,0.92780733,0.03383777,0.0032422538,0.00002228194,0.0018470155,0.0024142137,0.000003687983,0.0001140584,0.016429003,0.00025335056],"about_ca_topic_score_codex":0.00045304658,"about_ca_topic_score_gemma":0.00023907145,"teacher_disagreement_score":0.5571503,"about_ca_system_score_codex":0.00036613538,"about_ca_system_score_gemma":0.0011029324,"threshold_uncertainty_score":0.99818146},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4386592036","doi":"10.1111/jebm.12547","title":"Artificial intelligence and digital health in improving primary health care service delivery in LMICs: A systematic review","year":2023,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Mobile Health and mHealth Applications","field":"Health Professions","cited_by":44,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Psychological intervention; Digital health; Medicine; Health care; Service delivery framework; mHealth; Nursing; Service (business); Business; Political science","score_opus":0.33261185357091,"score_gpt":0.518243514726389,"score_spread":0.18563166115547897,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4386592036","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000034279797,0.9365954,0.00034168316,0.04723164,0.0006131344,0.01510361,0.00006112722,0.00003665104,0.000013355635],"genre_scores_gemma":[0.0000137109555,0.963921,0.00025623132,0.032367487,0.00073763507,0.0025056587,0.000092340175,0.00008565705,0.000020255167],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9828646,0.003066616,0.011331806,0.00063373806,0.0009107849,0.0011924373],"domain_scores_gemma":[0.97944874,0.008599869,0.0092190765,0.00074092695,0.00081874506,0.0011726726],"candidate_categories":["metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.015052369,0.0006162039,0.007700081,0.0016660779,0.00035444385,0.000018609991,0.00068295334,0.00040131723,0.000025212656],"category_scores_gemma":[0.007018471,0.00039808822,0.00026193226,0.0028013343,0.00010526166,0.00037832104,0.00014970115,0.0033743344,0.000064156724],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.000026699938,0.00003166445,0.000012969473,0.6204133,0.000011833249,0.000021246633,0.0004721627,0.0000016480377,1.5253361e-8,0.000029995907,0.00088129955,0.37809718],"study_design_scores_gemma":[0.0002566552,0.0006300093,0.000013269548,0.9399476,0.00029666696,0.000053074942,0.0016783084,0.00002778026,6.072592e-9,0.00008338531,0.056800175,0.00021304676],"about_ca_topic_score_codex":0.0014499377,"about_ca_topic_score_gemma":0.0022105526,"teacher_disagreement_score":0.37788412,"about_ca_system_score_codex":0.004439063,"about_ca_system_score_gemma":0.030685687,"threshold_uncertainty_score":0.9998471},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4386863815","doi":"10.1111/jebm.12550","title":"Methodological quality, risk of bias, and reporting quality: A confusion persists","year":2023,"lang":"en","type":"letter","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Confusion; Quality (philosophy); Risk analysis (engineering); Psychology; Business; Epistemology; Philosophy","score_opus":0.9914672909457757,"score_gpt":0.7089936643344512,"score_spread":0.28247362661132447,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4386863815","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.060231417,0.01710489,0.025534995,0.8941204,0.0019061303,0.0007693921,0.000030342817,0.00000900796,0.00029342496],"genre_scores_gemma":[0.10898319,0.020436725,0.082864396,0.72548574,0.028780315,0.000052047755,0.00007331696,0.00023539866,0.033088893],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.7402718,0.15061064,0.08131086,0.0016282285,0.025598569,0.0005798982],"domain_scores_gemma":[0.37333295,0.33404538,0.27706474,0.0045452174,0.010518361,0.0004933693],"candidate_categories":["metaresearch","metaepi_broad","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.763588,0.00064732233,0.0175817,0.001892828,0.00015340585,0.00025420438,0.002131007,0.00073113054,0.0038729706],"category_scores_gemma":[0.9174014,0.00023294058,0.0047550653,0.002048478,0.0006132431,0.0002833962,0.00023560162,0.002607133,0.000065917375],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009261722,0.000023256813,0.056417838,0.0013659829,0.0006840665,0.0006589346,0.00046532336,0.000038916914,0.0003606619,0.000018376386,0.92427766,0.015596382],"study_design_scores_gemma":[0.0021459637,0.002766012,0.1317421,0.034564015,0.01001611,0.0008469386,0.0048490926,0.00095926406,0.00012226801,0.013208971,0.7977671,0.0010121305],"about_ca_topic_score_codex":0.00024268993,"about_ca_topic_score_gemma":0.000020422294,"teacher_disagreement_score":0.36693886,"about_ca_system_score_codex":0.00009564062,"about_ca_system_score_gemma":0.0006205698,"threshold_uncertainty_score":0.9996939},"labels":[],"label_agreement":null},{"id":"W4389899943","doi":"10.1111/jebm.12577","title":"Reporting, handling, and interpretation of time‐varying drug treatments in observational studies using routinely collected healthcare data","year":2023,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Advanced Causal Inference Techniques","field":"Mathematics","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"National Natural Science Foundation of China","keywords":"Observational study; Discontinuation; Medicine; Logistic regression; Protocol (science); Pharmacoepidemiology; Meta-analysis; Internal medicine; Alternative medicine; Pharmacology; Pathology; Medical prescription","score_opus":0.8265753926692315,"score_gpt":0.5867571796628459,"score_spread":0.23981821300638562,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4389899943","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9845934,0.0030345093,0.007225073,0.0046365308,0.0001223043,0.00032186988,0.000008493677,0.000048509723,0.000009329142],"genre_scores_gemma":[0.9517133,0.0013013463,0.04658158,0.00015537547,0.00011707258,0.000003964426,0.000020136453,0.000019770541,0.00008745844],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99647254,0.00019610881,0.0024171637,0.00019177228,0.00055501336,0.00016741607],"domain_scores_gemma":[0.9908666,0.003566196,0.004308049,0.00028330545,0.00090464985,0.00007117],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.004663147,0.00015176751,0.00081067503,0.0004986415,0.00006006685,0.000007993114,0.00019649402,0.000049396207,0.000011133138],"category_scores_gemma":[0.04846497,0.00010176848,0.000038935465,0.00076092145,0.00013459552,0.0005696526,0.00009086275,0.00021596039,2.3722356e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0041998234,0.00077012484,0.73316103,0.009010569,0.001600936,0.0018119342,0.035136454,0.008638682,0.14444155,0.0011363965,0.02684068,0.033251822],"study_design_scores_gemma":[0.008592253,0.004225258,0.060237315,0.20474118,0.0015552604,0.00040617428,0.0070794076,0.52590567,0.026016502,0.16033934,0.00006128583,0.00084037706],"about_ca_topic_score_codex":0.00010933447,"about_ca_topic_score_gemma":0.000039883715,"teacher_disagreement_score":0.67292374,"about_ca_system_score_codex":0.00024232508,"about_ca_system_score_gemma":0.0004187647,"threshold_uncertainty_score":0.9595502},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"systematic_review","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"reporting","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4391230770","doi":"10.1111/jebm.12582","title":"Consensus for a primary care clinical decision‐making tool for assessing, diagnosing, and managing low back pain in Alberta, Canada","year":2024,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Musculoskeletal pain and rehabilitation","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Alberta Children's Hospital; Alberta Health Services; Mount Royal University; Alberta Health; Alberta Bone and Joint Health Institute; Canada Energy Regulator; University of Calgary","funders":"","keywords":"Delphi method; Medicine; Health care; Low back pain; Clinical pathway; Clinical decision making; Delphi; MEDLINE; Physical therapy; Nursing; Family medicine; Alternative medicine","score_opus":0.0514674864526475,"score_gpt":0.3939665490086469,"score_spread":0.34249906255599943,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4391230770","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.832811,0.03779178,0.04526382,0.07906012,0.002649312,0.0022749864,0.000007570228,0.000016407565,0.00012502557],"genre_scores_gemma":[0.9704993,0.00049453526,0.022699863,0.004958966,0.0011817014,0.000030569587,0.000010707179,0.000033650587,0.00009074666],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9972657,0.00026203343,0.001384991,0.00029697307,0.000524391,0.0002659224],"domain_scores_gemma":[0.9588869,0.039937545,0.00032487648,0.00016869628,0.00048678694,0.00019517877],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0063833026,0.00019171045,0.0007500939,0.0003350523,0.000064664266,0.000040114563,0.00008771906,0.00011638283,0.000030962834],"category_scores_gemma":[0.025831096,0.0001174546,0.00022146506,0.0002445099,0.00018234589,0.00015529894,0.000014112434,0.00037548316,3.5487457e-7],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019807166,0.00011349963,0.12621759,0.030078681,0.00013336336,0.00029071246,0.0005452362,0.00019249001,0.00039674153,0.0000638293,0.043580476,0.7964067],"study_design_scores_gemma":[0.015389311,0.009377709,0.3959445,0.4920572,0.00080118695,0.00018017528,0.0032121337,0.036356468,0.000054240987,0.0013502263,0.044709466,0.00056736806],"about_ca_topic_score_codex":0.004899637,"about_ca_topic_score_gemma":0.020428993,"teacher_disagreement_score":0.7958393,"about_ca_system_score_codex":0.00053555064,"about_ca_system_score_gemma":0.0027888427,"threshold_uncertainty_score":0.99744564},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"qualitative","genre":"empirical","about_ca_system":true,"about_ca_topic":true,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":true,"confidence":"high"}],"label_agreement":"split"},{"id":"W4392938758","doi":"10.1111/jebm.12600","title":"Handling time‐varying treatments in observational studies: A scoping review and recommendations","year":2024,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Advanced Causal Inference Techniques","field":"Mathematics","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; St. Joseph’s Healthcare Hamilton; Impact","funders":"National Natural Science Foundation of China","keywords":"Observational study; Protocol (science); Workflow; Observational methods in psychology; Computer science; Set (abstract data type); Process (computing); Medical physics; Management science; Selection (genetic algorithm); Medicine; Alternative medicine; Machine learning; Engineering","score_opus":0.833002005037274,"score_gpt":0.6314645032262648,"score_spread":0.20153750181100916,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4392938758","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000011530515,0.9912341,0.0011485246,0.005989903,0.00020208232,0.0013498343,0.00000756657,0.000040603754,0.00002622278],"genre_scores_gemma":[5.276739e-7,0.9681724,0.030632144,0.00053262536,0.0003106679,0.00013449503,0.0000151311915,0.000048858994,0.00015316495],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99582946,0.00041098087,0.0027154656,0.00030272902,0.0005074122,0.00023396892],"domain_scores_gemma":[0.9873541,0.00973238,0.0020441106,0.00027366905,0.00047059535,0.0001251366],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0040293713,0.00046551984,0.0036990563,0.00073972205,0.00006453998,0.000022503344,0.00025672055,0.00014566565,0.00012918748],"category_scores_gemma":[0.017628001,0.00025716273,0.00026772136,0.00080474035,0.00013237819,0.00048470555,0.00007844762,0.00080942974,0.0000073163555],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000012553236,0.00008026924,0.0000059395916,0.5247491,0.0006466462,0.00020852653,0.0001921479,0.0000013013026,0.0000039709453,0.00038582415,0.01209285,0.46162087],"study_design_scores_gemma":[0.00025888399,0.0003458553,1.9003845e-7,0.9668669,0.0066588726,0.00015291736,0.00002379945,0.000010822493,0.0000019485892,0.008415849,0.017065698,0.00019825892],"about_ca_topic_score_codex":0.0000024310846,"about_ca_topic_score_gemma":0.0000033845622,"teacher_disagreement_score":0.46142262,"about_ca_system_score_codex":0.0007047621,"about_ca_system_score_gemma":0.001044717,"threshold_uncertainty_score":0.9999881},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W4393097347","doi":"10.1111/jebm.12593","title":"Polypharmacy and risk of fractures in older adults: A systematic review","year":2024,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Pharmaceutical Practices and Patient Outcomes","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; McGill University Health Centre; Université Laval; Centre hospitalier de l'Université Laval; Université du Québec à Rimouski","funders":"Fonds de Recherche du Québec - Santé; Ministère de l'Éducation et de l'Enseignement supérieur","keywords":"Polypharmacy; Medicine; CINAHL; Cochrane Library; Confounding; MEDLINE; Gerontology; Meta-analysis; Intensive care medicine; Internal medicine; Psychiatry; Psychological intervention","score_opus":0.27013753249671585,"score_gpt":0.5235085103135608,"score_spread":0.253370977816845,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4393097347","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00006503492,0.99225223,0.000017245593,0.0040685893,0.00054793135,0.0029894684,0.000012724179,0.000009158664,0.000037600545],"genre_scores_gemma":[0.0005069182,0.9965045,0.00010675281,0.0022358377,0.00048429437,0.000037460068,0.0000061816586,0.000047543486,0.00007054597],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9932704,0.0010668095,0.0038954758,0.00030154502,0.0012000079,0.00026578485],"domain_scores_gemma":[0.9884247,0.006081967,0.004291372,0.00039861095,0.00037439977,0.0004289178],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005121082,0.00053356047,0.006932918,0.00095608627,0.000021195552,0.000012416274,0.00027931185,0.00019670874,0.00033497496],"category_scores_gemma":[0.016652558,0.00022954968,0.000840874,0.0008478569,0.00017910782,0.00026423432,0.000049232218,0.0021508837,0.000012081062],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002461703,0.00012795339,0.0002627854,0.92243046,0.00078150997,0.0005943441,0.00009957359,1.702577e-7,4.1681443e-7,0.0000022599315,0.0020925035,0.073361844],"study_design_scores_gemma":[0.0017197544,0.00081833766,0.000042959957,0.86903536,0.06843274,0.00087907026,0.00004790552,0.000038354996,0.0000013833783,0.0000067501883,0.058817014,0.00016035355],"about_ca_topic_score_codex":0.000048381524,"about_ca_topic_score_gemma":0.00000222373,"teacher_disagreement_score":0.07320149,"about_ca_system_score_codex":0.00017259369,"about_ca_system_score_gemma":0.0008290332,"threshold_uncertainty_score":0.9916306},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4393232494","doi":"10.1111/jebm.12585","title":"Critical appraisal of a statistical analysis: A novel visual guide","year":2024,"lang":"en","type":"editorial","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Advancing Health Outcomes; Vancouver Coastal Health Research Institute; University of British Columbia; International Collaboration On Repair Discoveries; Vancouver Coastal Health","funders":"","keywords":"Columbia university; Library science; Population; Geography; History; Sociology; Media studies; Demography","score_opus":0.675440415036593,"score_gpt":0.6314371220015348,"score_spread":0.04400329303505823,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4393232494","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000779424,0.017203908,0.44619527,0.0078456625,0.52772367,0.00030052953,0.00023568909,0.000004943419,0.00041237168],"genre_scores_gemma":[0.017508224,0.00070705934,0.031859823,0.0004708608,0.94432384,0.000024959121,0.00008088308,0.00009005068,0.004934312],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.91305906,0.009795039,0.036829863,0.0014694985,0.038283873,0.00056268455],"domain_scores_gemma":[0.3495921,0.5714715,0.03521705,0.0055073635,0.0365229,0.0016891038],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.27219966,0.00089006405,0.016877351,0.004991125,0.000080732265,0.000644629,0.003821325,0.00076691114,0.012978182],"category_scores_gemma":[0.879412,0.00034984795,0.006467629,0.0062167402,0.00092160533,0.00051265914,0.00019973707,0.0022022175,0.00037102908],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00019657922,0.0002011273,0.00024721376,0.0010324332,0.0036620633,0.00031296865,0.00016327635,0.000166315,0.00012568812,0.00042215895,0.99144226,0.0020279353],"study_design_scores_gemma":[0.0007706243,0.0018294024,0.00022935947,0.009096071,0.037956603,0.000041593597,0.00042578464,0.009344059,0.000011151691,0.0007488931,0.93920684,0.00033959394],"about_ca_topic_score_codex":0.00008991859,"about_ca_topic_score_gemma":0.000029987412,"teacher_disagreement_score":0.6072123,"about_ca_system_score_codex":0.00021292394,"about_ca_system_score_gemma":0.0040022,"threshold_uncertainty_score":0.99989533},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4393934783","doi":"10.1111/jebm.12603","title":"Fluids in the treatment of diabetic ketoacidosis in children: A systematic review","year":2024,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Diabetes and associated disorders","field":"Biochemistry, Genetics and Molecular Biology","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Diabetic ketoacidosis; Meta-analysis; Population; Randomized controlled trial; Intensive care medicine; Pediatrics; Surgery; Internal medicine; Insulin","score_opus":0.05882681303615756,"score_gpt":0.3550558840351998,"score_spread":0.2962290709990422,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4393934783","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0004035206,0.9969935,0.0000031439401,0.0010228541,0.0001402654,0.0014045466,0.000010711838,0.0000013904216,0.000020102427],"genre_scores_gemma":[0.0010285083,0.9980485,0.000012168067,0.0004577248,0.00023287654,0.00011477145,0.000034759978,0.00002618116,0.000044523327],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9963798,0.0007673332,0.0019377825,0.00022786613,0.00045576197,0.00023143708],"domain_scores_gemma":[0.99791074,0.00033959854,0.001151386,0.00042135754,0.00011048634,0.00006643],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0023287742,0.0003691749,0.0029202271,0.0003373712,0.000012528769,0.000011205644,0.00047820545,0.0001956268,0.000016194956],"category_scores_gemma":[0.0013532178,0.00015375543,0.00081381307,0.0006845251,0.00009296938,0.0000055996275,0.000016673068,0.00025526216,0.0000036537601],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000017306264,0.0005339286,0.00039884262,0.9337019,0.0010128648,0.000060467322,0.000097814846,0.0000035886537,0.000038907467,0.0000039146107,0.0036390084,0.060491458],"study_design_scores_gemma":[0.0005959114,0.0023905258,0.000020749061,0.9619826,0.007887154,0.00004085948,0.0000493872,0.0000017980618,0.000012670583,0.000007874529,0.026858477,0.00015198528],"about_ca_topic_score_codex":0.000019792704,"about_ca_topic_score_gemma":0.000008292672,"teacher_disagreement_score":0.060339473,"about_ca_system_score_codex":0.00013297288,"about_ca_system_score_gemma":0.0004981328,"threshold_uncertainty_score":0.62699646},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W4394722496","doi":"10.1111/jebm.12606","title":"The detection of circulating tumor cells indicates poor therapeutic efficacy and prognosis in patients with nonsmall cell lung cancer: A systematic review and meta‐analysis","year":2024,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Cancer Cells and Metastasis","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Meta-analysis; Internal medicine; Oncology; Circulating tumor cell; Cochrane Library; Hazard ratio; Lung cancer; Cancer; Confidence interval; Metastasis","score_opus":0.08999203785124837,"score_gpt":0.3642577766812765,"score_spread":0.27426573883002814,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4394722496","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000302249,0.99470866,0.000044881894,0.0010416186,0.0000857599,0.0037915423,0.000016443422,0.0000061369938,0.0000026878488],"genre_scores_gemma":[0.030680064,0.96859306,0.000074432464,0.0002826677,0.00007502355,0.00022243595,0.0000037442367,0.000045820343,0.000022756727],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9955382,0.00057100685,0.002352894,0.0003638283,0.00091146096,0.00026262677],"domain_scores_gemma":[0.9937072,0.0018950108,0.003424704,0.00028608312,0.00046596574,0.00022105446],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0031131718,0.0005027922,0.0072425036,0.00093756925,0.000066025495,0.000032770437,0.00019042446,0.00009814542,0.000028574723],"category_scores_gemma":[0.0004464527,0.00019174861,0.00084942684,0.0019183388,0.00020990237,0.00009620111,0.000034446937,0.00071893143,3.7674067e-7],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011254646,0.00010443214,0.00022578736,0.9397087,0.04307275,0.000049755537,0.00008543859,0.000003317229,0.000019138573,3.511511e-7,0.000018320728,0.016599467],"study_design_scores_gemma":[0.00072674337,0.0007766583,0.00006194378,0.39232406,0.6053117,0.000042459178,0.000026083335,0.00001908331,0.000029109695,4.3938897e-7,0.0005650291,0.00011668868],"about_ca_topic_score_codex":0.000106539774,"about_ca_topic_score_gemma":0.00003940744,"teacher_disagreement_score":0.56223893,"about_ca_system_score_codex":0.00022662092,"about_ca_system_score_gemma":0.0006796464,"threshold_uncertainty_score":0.7819281},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4399901519","doi":"10.1111/jebm.12622","title":"Impact of hormone replacement therapy on all‐cause and cancer‐specific mortality in colorectal cancer: A systematic review and dose‒response meta‐analysis of observational studies","year":2024,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Menopause: Health Impacts and Treatments","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"Natural Science Foundation of Sichuan Province","keywords":"Medicine; Meta-analysis; Hazard ratio; Internal medicine; Colorectal cancer; Hormone replacement therapy (female-to-male); Cohort study; Cochrane Library; Oncology; Cohort; Cancer; Confidence interval; Testosterone (patch)","score_opus":0.6676574156498694,"score_gpt":0.5740848911037709,"score_spread":0.09357252454609843,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4399901519","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.039354883,0.95399606,8.918393e-7,0.0030723081,0.00010489984,0.0032791693,0.00018606163,0.000004805212,9.2783847e-7],"genre_scores_gemma":[0.009898029,0.98908967,0.00003448738,0.00046660707,0.00008994188,0.00033887476,0.000020475038,0.000034157685,0.000027746883],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9927619,0.0012950222,0.003925103,0.00042231908,0.0013011389,0.00029451842],"domain_scores_gemma":[0.99011946,0.0040391004,0.0042069126,0.0004439285,0.0007912802,0.00039933104],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0074401945,0.00065067044,0.013485099,0.0014847423,0.00003363629,0.000010197016,0.0001513674,0.0001467541,0.00011569539],"category_scores_gemma":[0.0023039954,0.00027540998,0.00055466103,0.0016727393,0.00024759196,0.00012780227,0.000031206735,0.0005271497,3.5103244e-7],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0051675453,0.00033774512,0.0008447156,0.624846,0.36513105,0.00027606316,0.00030843107,0.000012777159,0.000068279536,0.0000044966737,0.0009528654,0.0020500394],"study_design_scores_gemma":[0.0022676548,0.013534109,0.0058586965,0.45571852,0.5219544,0.00006362884,0.000021797312,0.000007819287,0.000002096246,0.0000034219436,0.00044474733,0.00012309186],"about_ca_topic_score_codex":0.0005691063,"about_ca_topic_score_gemma":0.00006124006,"teacher_disagreement_score":0.16912746,"about_ca_system_score_codex":0.0012012286,"about_ca_system_score_gemma":0.0024301708,"threshold_uncertainty_score":0.9999698},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4402619708","doi":"10.1111/jebm.12638","title":"Evaluation of the reliability and usability of CARE‐Radiology: A descriptive‐analytic study","year":2024,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Radiology practices and education","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Usability; Reliability (semiconductor); Descriptive research; Psychology; Computer science; Statistics; Mathematics; Human–computer interaction; Physics","score_opus":0.24208867524564498,"score_gpt":0.4457038113049224,"score_spread":0.20361513605927745,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402619708","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95409423,0.025792655,0.00007576066,0.018429719,0.00089397404,0.00062058447,9.978004e-7,0.0000038684884,0.00008821822],"genre_scores_gemma":[0.9989372,0.00040471004,0.00018126745,0.00011248504,0.00033311427,0.0000061734463,5.5007604e-7,0.000006190619,0.000018327484],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9965223,0.0012440189,0.000924856,0.00019360305,0.0010079966,0.00010720711],"domain_scores_gemma":[0.99465907,0.00205793,0.0006427822,0.0003784425,0.0021646435,0.00009712819],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.016249908,0.000111413676,0.0006008795,0.00020886121,0.000039095234,0.0000044144967,0.00012554249,0.00007562596,0.000112069545],"category_scores_gemma":[0.022586407,0.00005192077,0.00013331656,0.00042921735,0.0006081076,0.00023222616,0.000017237659,0.0003942676,4.1208943e-7],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0021589454,0.0007916034,0.92906094,0.0019958268,0.0007697532,0.000021265416,0.017775461,0.0011143424,0.01291598,0.000042236792,0.0015605863,0.031793047],"study_design_scores_gemma":[0.0024289538,0.008466108,0.9516119,0.005057238,0.008893462,0.00027314376,0.015340649,0.0062235445,0.00088353106,0.0005273267,0.00023000424,0.00006416351],"about_ca_topic_score_codex":0.00015975053,"about_ca_topic_score_gemma":0.000019728492,"teacher_disagreement_score":0.04484296,"about_ca_system_score_codex":0.0003766537,"about_ca_system_score_gemma":0.0019120803,"threshold_uncertainty_score":0.9856468},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W4402777205","doi":"10.1111/jebm.12642","title":"Validation of the Generic Version of China Health Related Outcomes Measures (CHROME‐G) among individuals with type 2 diabetes in China","year":2024,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"National Natural Science Foundation of China","keywords":"Cronbach's alpha; Convergent validity; Rank correlation; Spearman's rank correlation coefficient; Medicine; Quality of life (healthcare); Population; Correlation; Demography; Type 2 diabetes; SF-36; Statistics; Gerontology; Health related quality of life; Mathematics; Diabetes mellitus; Clinical psychology; Internal consistency; Psychometrics; Internal medicine; Environmental health","score_opus":0.24583446782666812,"score_gpt":0.40354834018930813,"score_spread":0.15771387236264,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402777205","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9086809,0.019653227,0.00016407254,0.07022829,0.0008185035,0.00036219633,0.000020242174,0.000007683974,0.00006487489],"genre_scores_gemma":[0.9977428,0.0008269502,0.00040850832,0.00080146454,0.00013011185,0.000004391588,0.000005665382,0.000019347368,0.000060769882],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99412596,0.00064903754,0.004388183,0.00021805473,0.00039780425,0.00022096743],"domain_scores_gemma":[0.99331814,0.0014055015,0.004679768,0.00031706053,0.00015866033,0.00012086129],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.027311336,0.00016060339,0.0012985356,0.00086418516,0.0000608646,0.000016801481,0.0003195255,0.00009507384,0.00018859337],"category_scores_gemma":[0.007761447,0.00010064939,0.00013355525,0.00094394526,0.0001979406,0.0005585163,0.000023955075,0.00034347124,0.000010367345],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000031902287,0.00005782185,0.9888554,0.0009602101,0.00012723541,0.0000016244713,0.0018915224,0.0036669562,0.00007011714,0.0003320242,0.003748348,0.00025680452],"study_design_scores_gemma":[0.0010970073,0.0009608126,0.98408276,0.009461557,0.000044055978,0.0000049111295,0.0002908299,0.0018331183,0.00027236226,0.0012456547,0.0005969895,0.00010995487],"about_ca_topic_score_codex":0.0008792458,"about_ca_topic_score_gemma":0.000044830354,"teacher_disagreement_score":0.08906188,"about_ca_system_score_codex":0.00047089008,"about_ca_system_score_gemma":0.0007649148,"threshold_uncertainty_score":0.9465619},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W4402890012","doi":"10.1111/jebm.12651","title":"International Traditional Medicine Clinical Trial Registry: A meaningful initiative and its future development","year":2024,"lang":"en","type":"letter","venue":"Journal of Evidence-Based Medicine","topic":"Complementary and Alternative Medicine Studies","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"China Academy of Chinese Medical Sciences; World Health Organization","keywords":"Clinical trial; Alternative medicine; Medicine; Family medicine; Internal medicine; Pathology","score_opus":0.565957851305087,"score_gpt":0.46926807889027355,"score_spread":0.09668977241481347,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402890012","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.007807208,0.02333981,0.00026220438,0.94890505,0.017354272,0.00091574463,0.00004247493,0.00003742287,0.0013358179],"genre_scores_gemma":[0.026556399,0.01701762,0.0018114017,0.6420503,0.30897725,0.00007318837,0.0006113057,0.00012876223,0.0027737233],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9914783,0.00047011007,0.0035337321,0.00073369604,0.0033267243,0.00045739577],"domain_scores_gemma":[0.99083865,0.004445339,0.002296064,0.00032136322,0.0016536025,0.00044498878],"candidate_categories":["metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004346788,0.0007732354,0.0026032638,0.0012263026,0.00015423188,0.000027747059,0.0005545154,0.00029516025,0.0025259904],"category_scores_gemma":[0.0073765586,0.0004275079,0.00042422832,0.00042397794,0.00095699955,0.0003040286,0.0001333862,0.005090207,0.000017105238],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.013204784,0.00016427034,0.00054998085,0.0010521719,0.0026162844,0.01405617,0.0010232371,3.7861506e-7,0.000025000214,0.000083831335,0.95519036,0.012033539],"study_design_scores_gemma":[0.04801162,0.016324088,0.0018500992,0.04101427,0.0025420869,0.002506207,0.0010509696,0.00003182052,0.000033180735,0.0003572931,0.8859346,0.00034375722],"about_ca_topic_score_codex":0.00000752453,"about_ca_topic_score_gemma":0.000003798077,"teacher_disagreement_score":0.3068547,"about_ca_system_score_codex":0.00048041533,"about_ca_system_score_gemma":0.0020784272,"threshold_uncertainty_score":0.99981767},"labels":[{"model":"gemma","categories":["open_science"],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4407665061","doi":"10.1111/jebm.70000","title":"Prognostic Significance of Hypophosphatemia in Kidney Transplant Patients: A Systematic Review and Meta‐Analysis","year":2025,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Parathyroid Disorders and Treatments","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Hypophosphatemia; Medicine; Kidney transplant; Meta-analysis; Internal medicine; Intensive care medicine; Kidney; Kidney transplantation","score_opus":0.11543966429191882,"score_gpt":0.38191782849746897,"score_spread":0.26647816420555015,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4407665061","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000006123649,0.9938558,0.00010407325,0.0014552023,0.00006751463,0.0044099423,0.00007017622,0.0000052766973,0.000025920102],"genre_scores_gemma":[0.00063457055,0.99754673,0.00017017408,0.0010855617,0.00002474297,0.0002816773,0.000059451202,0.000022209097,0.0001749091],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.99401605,0.00071881444,0.0037085589,0.00037198304,0.0009194644,0.00026515595],"domain_scores_gemma":[0.9924701,0.0026863287,0.0033063996,0.0004956913,0.00063837646,0.00040313075],"candidate_categories":["metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.0020180705,0.00059946853,0.015038197,0.0013834869,0.000026642407,0.0000072205503,0.0002746824,0.00019154057,0.0000598122],"category_scores_gemma":[0.0063417526,0.00028107825,0.0019146387,0.0023462602,0.00015565426,0.00012978574,0.000016115735,0.00049864297,0.0000018024335],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008805876,0.0004134281,0.00066254614,0.9338128,0.06318245,0.00026726673,0.00003227745,0.0000010497953,2.3066269e-7,0.0000032792288,0.00087635155,0.0006602753],"study_design_scores_gemma":[0.0011283458,0.0006943998,0.000038170656,0.47197536,0.5252122,0.000028935272,0.000005807387,0.0000016323388,5.8183315e-7,0.000003894783,0.00080969353,0.00010098752],"about_ca_topic_score_codex":0.000043966345,"about_ca_topic_score_gemma":0.0000062160807,"teacher_disagreement_score":0.46202976,"about_ca_system_score_codex":0.00017893482,"about_ca_system_score_gemma":0.0018939192,"threshold_uncertainty_score":0.9999641},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4407753376","doi":"10.1111/jebm.70001","title":"Association Between Recurrent Aphthous Ulcers and Oral Biodiversity: A Systematic Review and Meta‐Analysis","year":2025,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Oral Health Pathology and Treatment","field":"Dentistry","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"National Natural Science Foundation of China","keywords":"Meta-analysis; Cochrane Library; Medicine; Internal medicine; Confidence interval; Saliva; Beta diversity; Subgroup analysis; Gastroenterology; Biology; Biodiversity; Ecology","score_opus":0.27340091588399423,"score_gpt":0.45475368682364314,"score_spread":0.1813527709396489,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4407753376","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000005999329,0.9928717,0.00006440285,0.0052022124,0.0002792009,0.0014322483,0.0001231801,0.000012049065,0.000008981298],"genre_scores_gemma":[0.000028177621,0.9982238,0.00013953503,0.00091702363,0.00009896666,0.000038298705,0.000037452657,0.000008425496,0.00050833786],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9939324,0.0025735009,0.002147993,0.00035425395,0.00072372425,0.00026811182],"domain_scores_gemma":[0.9891362,0.0053045936,0.0044855853,0.00033600852,0.00044722218,0.00029038568],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0064073424,0.00046363997,0.010528426,0.0009992006,0.00014306366,0.00002429937,0.00024561517,0.00037220606,0.00015537678],"category_scores_gemma":[0.0037247634,0.00024058996,0.0013932469,0.0011527587,0.000106518906,0.00016684085,0.000062070896,0.00074530137,0.00001745846],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000013202591,0.00003161166,0.00085980754,0.881423,0.107593305,0.00035700313,0.000015901367,1.02470864e-7,2.6856979e-9,0.0000016447793,0.0029061765,0.0067982725],"study_design_scores_gemma":[0.00029578252,0.00035992652,0.00009757572,0.304474,0.6886185,0.00009769136,0.000009855369,2.35365e-7,3.4182598e-8,0.000007506565,0.005924801,0.00011412073],"about_ca_topic_score_codex":0.000043172,"about_ca_topic_score_gemma":0.000029170204,"teacher_disagreement_score":0.5810252,"about_ca_system_score_codex":0.00054414466,"about_ca_system_score_gemma":0.00049094966,"threshold_uncertainty_score":0.98109734},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W4408070566","doi":"10.1111/jebm.70006","title":"Inadequate Reporting of Harm From Randomized Clinical Trials in Top Medical Publications","year":2025,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"National Natural Science Foundation of China","keywords":"Consolidated Standards of Reporting Trials; Harm; Medicine; Randomized controlled trial; Family medicine; Alternative medicine; Descriptive statistics; Medical journal; Psychology; Internal medicine; Pathology","score_opus":0.9714145874832354,"score_gpt":0.7366030185621091,"score_spread":0.2348115689211263,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4408070566","genre_codex":"review","genre_gemma":"review","domain_codex":"reporting","domain_gemma":"methods","model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000029133804,0.94333905,0.029827585,0.02182151,0.0021267466,0.0023227762,0.000014148441,0.0000029793116,0.0005160479],"genre_scores_gemma":[0.00015207926,0.99233633,0.0033336435,0.0006896951,0.0017252413,0.00009074089,0.000019745989,0.000018016963,0.0016345155],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.14781943,0.40695485,0.41426724,0.002068131,0.028269675,0.0006206638],"domain_scores_gemma":[0.009737177,0.6556275,0.32292235,0.004247249,0.006682217,0.00078354793],"candidate_categories":["metaresearch","metaepi_broad","insufficient_payload"],"consensus_categories":["metaresearch","metaepi_broad"],"category_scores_codex":[0.9342591,0.00072579953,0.09094124,0.004388515,0.00006212686,0.0002502177,0.0049690143,0.00078759104,0.019158997],"category_scores_gemma":[0.9934424,0.00024322608,0.020961752,0.0056427037,0.000617126,0.0004200699,0.00019966504,0.002083716,0.00007989409],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007209166,0.00012516289,0.00049426,0.0043808036,0.0012923023,0.00010517502,0.000064504165,0.0000041732533,1.9357763e-7,0.00017900167,0.057193905,0.9354396],"study_design_scores_gemma":[0.016838888,0.00014114914,0.000064363165,0.26675206,0.013810368,0.00005243263,0.00012785563,0.0005890349,4.924236e-7,0.0021705108,0.69924766,0.00020518084],"about_ca_topic_score_codex":0.00009455562,"about_ca_topic_score_gemma":0.000026535636,"teacher_disagreement_score":0.9352344,"about_ca_system_score_codex":0.00016496298,"about_ca_system_score_gemma":0.009903991,"threshold_uncertainty_score":0.99570894},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4408454641","doi":"10.1111/jebm.70011","title":"Treatment for Depression Among Adults: An Evidence and Gap Map of Systematic Reviews","year":2025,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Treatment of Major Depression","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"National Social Science Fund of China; National Office for Philosophy and Social Sciences; Queen's University; University College London; Lanzhou University","keywords":"Systematic review; Psychological intervention; Pharmacotherapy; Depression (economics); Randomized controlled trial; Psychiatry; Psychology; Clinical psychology; Intervention (counseling); Major depressive disorder; Meta-analysis; MEDLINE; Medicine; Mood; Internal medicine","score_opus":0.23142318271998852,"score_gpt":0.423286228035958,"score_spread":0.19186304531596948,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4408454641","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.52081805,0.44638345,0.007428428,0.016623313,0.0008810688,0.007768059,0.0000049689133,0.00003590026,0.000056752775],"genre_scores_gemma":[0.9747339,0.017653562,0.006077734,0.00039855775,0.00037287752,0.0001997627,0.000004834861,0.000023529587,0.00053524465],"study_design_codex":"observational","study_design_gemma":"systematic_review","domain_scores_codex":[0.99643093,0.0004343428,0.0018948009,0.0003208661,0.00068771356,0.00023136048],"domain_scores_gemma":[0.99284077,0.003711291,0.00173903,0.00063388597,0.0007383887,0.00033664264],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0023274177,0.00032620004,0.0019520142,0.0005341724,0.00007544347,0.000013130164,0.00020727151,0.00013371668,0.000045748697],"category_scores_gemma":[0.009478798,0.00015318858,0.00028750053,0.00024073334,0.00021215262,0.0005300375,0.000025581146,0.00017416052,0.0000015511715],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.022716003,0.0026737992,0.4852425,0.34208423,0.00095863437,0.00047125504,0.0018159987,0.00006395219,0.0516447,0.00007791447,0.011217816,0.08103322],"study_design_scores_gemma":[0.01266847,0.0147497235,0.03911548,0.91294324,0.004746814,0.00015575609,0.00050350314,0.0011382098,0.01307347,0.00010766788,0.0006383361,0.00015931249],"about_ca_topic_score_codex":0.00007095064,"about_ca_topic_score_gemma":0.0000143332745,"teacher_disagreement_score":0.570859,"about_ca_system_score_codex":0.00029706495,"about_ca_system_score_gemma":0.0003500656,"threshold_uncertainty_score":0.99886477},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W4408981663","doi":"10.1111/jebm.70019","title":"Synthesizing Public Health Preparedness Mechanisms for High‐Impact Infectious Disease Threats: A Jurisdictional Scan","year":2025,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Viral Infections and Outbreaks Research","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"Public Health Agency","keywords":"Preparedness; Public health; International Health Regulations; Health care; Business; Emergency management; Global health; Infectious disease (medical specialty); Workforce; Pandemic; Public relations; Medicine; Medical emergency; Political science; Disease; Coronavirus disease 2019 (COVID-19); Nursing","score_opus":0.33523408535193183,"score_gpt":0.5054519080436269,"score_spread":0.1702178226916951,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4408981663","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000026590322,0.944921,0.03867387,0.01201494,0.0014251681,0.0026664592,0.00012847589,0.00006884279,0.00007464699],"genre_scores_gemma":[0.0028065739,0.99153006,0.0007180597,0.0009524503,0.0027806896,0.000314535,0.00013454705,0.00009141716,0.0006716418],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.99363434,0.00075334933,0.0024824964,0.00059314823,0.0016703706,0.0008662852],"domain_scores_gemma":[0.9890626,0.0042352234,0.0021810837,0.00065557176,0.002033971,0.0018315511],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0058296425,0.00074577314,0.0043673795,0.0030950815,0.00041163948,0.000099726174,0.00040792229,0.00028722314,0.00048175713],"category_scores_gemma":[0.008398798,0.00040952052,0.0018602978,0.0017700532,0.00027701948,0.0004763897,0.00007383413,0.0012912167,0.000007692736],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00064948184,0.000567561,0.0000712579,0.054564085,0.001236933,0.00018102808,0.000060046517,0.000021062395,0.0000045388174,0.0006164978,0.010904405,0.9311231],"study_design_scores_gemma":[0.0044348557,0.010869862,0.00018919208,0.56038713,0.0069559715,0.0012727091,0.000103584585,0.0001547974,0.000004815567,0.002154182,0.4129038,0.0005691005],"about_ca_topic_score_codex":0.0002100981,"about_ca_topic_score_gemma":0.000020510812,"teacher_disagreement_score":0.930554,"about_ca_system_score_codex":0.0037415612,"about_ca_system_score_gemma":0.031749588,"threshold_uncertainty_score":0.99995387},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4410942467","doi":"10.1111/jebm.70034","title":"Knowledge and Awareness of Generative Artificial Intelligence Use in Medicine Among International Stakeholders: A Cross‐Sectional Study","year":2025,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Artificial Intelligence in Healthcare and Education","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Psychology; Medical education; Cross-sectional study; Descriptive statistics; Medicine","score_opus":0.6922511064727699,"score_gpt":0.547229176445701,"score_spread":0.14502193002706887,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4410942467","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97302455,0.0022236824,0.010078599,0.011641676,0.0023621593,0.00057290005,0.0000020761056,0.000008888148,0.00008547088],"genre_scores_gemma":[0.9971559,0.000610428,0.00040608854,0.00045107192,0.0011682095,0.000018235944,0.0000058694886,0.000011554341,0.00017263698],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99609476,0.00027771541,0.0023249933,0.00032088225,0.0007570481,0.00022460277],"domain_scores_gemma":[0.99383485,0.0025056563,0.0007600799,0.0002456012,0.002447968,0.00020581979],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0032516683,0.00021620767,0.00075293594,0.0014570383,0.00009463335,0.000027739696,0.0002209723,0.00013863019,0.00022938],"category_scores_gemma":[0.009534719,0.00014339255,0.00008412925,0.0009622808,0.00084176194,0.0005586288,0.000037882808,0.00056364574,0.0000014143671],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017097828,0.0007787799,0.96996135,0.00021221349,0.000110627894,0.000066138826,0.004110751,0.0004833765,0.0017709029,0.0003452299,0.0004233509,0.020027488],"study_design_scores_gemma":[0.00063105585,0.0038420216,0.9572262,0.0095345825,0.00021911817,0.000047028258,0.012821175,0.0038633114,0.0092205675,0.0022941446,0.0001622458,0.00013854986],"about_ca_topic_score_codex":0.0024280006,"about_ca_topic_score_gemma":0.0012687824,"teacher_disagreement_score":0.024131356,"about_ca_system_score_codex":0.00035679224,"about_ca_system_score_gemma":0.0013593622,"threshold_uncertainty_score":0.9988084},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4413626650","doi":"10.1111/jebm.70058","title":"A Systematic Survey of the Optimal Strategy for Dealing With Missing Binary Outcomes in Simulation Studies of Randomized Controlled Trials","year":2025,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Statistical Methods and Bayesian Inference","field":"Mathematics","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"National Natural Science Foundation of China","keywords":"Missing data; Randomized controlled trial; Statistics; Imputation (statistics); Meta-analysis; Statistical power; Computer science; MEDLINE; Descriptive statistics; Psychology; Econometrics; Medicine; Mathematics","score_opus":0.484340746084108,"score_gpt":0.5428199664858457,"score_spread":0.05847922040173775,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413626650","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.051813968,0.008204716,0.9350284,0.0018697805,0.0002054726,0.0028582523,0.000005937927,0.000002852945,0.000010647295],"genre_scores_gemma":[0.8900226,0.00012555905,0.10972213,0.0000468581,0.00002606055,0.000035309222,1.7392065e-7,0.0000066750795,0.0000146884495],"study_design_codex":"randomized_trial","study_design_gemma":"systematic_review","domain_scores_codex":[0.9894543,0.0053015095,0.0044774213,0.00009533994,0.0005511818,0.0001202007],"domain_scores_gemma":[0.6644115,0.3291367,0.004561411,0.00016619879,0.0016900089,0.000034180302],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.058369163,0.000165926,0.0072236424,0.00027689463,0.00003931742,0.000009608906,0.00017126909,0.000052797623,0.000010427396],"category_scores_gemma":[0.52858543,0.000057204867,0.00041572755,0.00027033634,0.0002872281,0.00007499356,0.000012153716,0.000139197,1.2370845e-8],"study_design_candidate":"randomized_trial","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.7217598,0.0007586341,0.014115815,0.119227245,0.011195973,0.00003707374,0.0020905028,0.061718285,0.00429195,0.06041318,0.000103945575,0.00428759],"study_design_scores_gemma":[0.35153303,0.0018088958,0.004635808,0.37184802,0.00823005,0.000004200889,0.0012735997,0.15114588,0.00093994057,0.10842882,9.819571e-8,0.00015164525],"about_ca_topic_score_codex":0.00005352014,"about_ca_topic_score_gemma":0.000016169699,"teacher_disagreement_score":0.83820856,"about_ca_system_score_codex":0.000050620536,"about_ca_system_score_gemma":0.00043752062,"threshold_uncertainty_score":0.9696071},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4414717095","doi":"10.1111/jebm.70080","title":"Intervention at Circulating Homocysteine Levels &gt;10 µmol/L for Primary Prevention of Cardiovascular Diseases: A Systematic Review and Dose–Response Meta‐Analysis","year":2025,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Folate and B Vitamins Research","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Primary prevention; Intervention (counseling); Homocysteine; Primary care; Disease; Epidemiology","score_opus":0.2504137809890123,"score_gpt":0.4511091426588932,"score_spread":0.2006953616698809,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4414717095","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000043955843,0.9894661,0.0035977808,0.00084764446,0.000094043484,0.0058501246,0.00007959447,0.000012351937,0.000008412569],"genre_scores_gemma":[0.0003917162,0.9958747,0.00040192253,0.00014609405,0.00019279966,0.00040759644,0.00012404907,0.000047143618,0.0024139986],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.98950845,0.003221544,0.0044507254,0.00054027524,0.0019239587,0.00035506987],"domain_scores_gemma":[0.98674154,0.006091859,0.003739389,0.0011008733,0.0019374333,0.00038890095],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaepi_broad"],"category_scores_codex":[0.015963003,0.00062757626,0.016205367,0.0023646588,0.00009826654,0.000024801368,0.0004004539,0.00027207672,0.0003156131],"category_scores_gemma":[0.01829141,0.00033795525,0.011932241,0.0021938158,0.0002030597,0.00028232374,0.00012678218,0.00053271814,0.0000025017043],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005571025,0.00015966635,0.0000035925973,0.7995005,0.17772603,0.00016836266,0.000014317926,0.0000099956815,0.0000133598105,0.0000010158254,0.00021792593,0.021628078],"study_design_scores_gemma":[0.0012407255,0.0015612188,0.000030893934,0.4820275,0.5098622,0.00024015697,0.000011115665,0.000019713467,0.0000022965398,0.0000036211854,0.0048716017,0.00012898644],"about_ca_topic_score_codex":0.000018992661,"about_ca_topic_score_gemma":0.0000026133555,"teacher_disagreement_score":0.33213615,"about_ca_system_score_codex":0.0006883593,"about_ca_system_score_gemma":0.001770908,"threshold_uncertainty_score":0.99990726},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W7117138160","doi":"10.1111/jebm.70098","title":"Challenges Facing Evidence‐Synthesis and Clinical Practice Guidelines, and Their Solutions: Restoring Balance Through Core GRADE","year":2025,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Clinical practice guidelines implementation","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Clinical Practice; Core (optical fiber); Balance (ability); Component (thermodynamics); Key (lock); Core stability","score_opus":0.809906837804852,"score_gpt":0.6039205569126929,"score_spread":0.2059862808921591,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7117138160","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.04325209,0.2544661,0.016553255,0.6835593,0.0013700282,0.00056369574,0.0000026367163,0.000039208797,0.00019371253],"genre_scores_gemma":[0.5244402,0.40197036,0.048537407,0.02228943,0.002638916,0.000016505308,0.0000019150013,0.000029708182,0.00007553806],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9936431,0.0006836756,0.0038932888,0.00050499977,0.00090634683,0.00036860484],"domain_scores_gemma":[0.9371377,0.056542356,0.0021557484,0.00055467576,0.0032675005,0.00034200642],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.019877387,0.00032440774,0.0011985952,0.00032100073,0.00020528263,0.000035499605,0.00021971365,0.00021029223,0.000037820064],"category_scores_gemma":[0.4581867,0.00019868667,0.00018737267,0.00044454573,0.0005031949,0.0017004962,0.00012274043,0.001046735,0.0000020715313],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0069343033,0.0006686541,0.056518868,0.0019937605,0.0012002389,0.0005103174,0.0012385194,0.00008020663,0.0074783685,0.001158693,0.070385575,0.8518325],"study_design_scores_gemma":[0.022515839,0.014015279,0.19955082,0.24624595,0.014990157,0.00521518,0.046914533,0.011550601,0.0022950273,0.0074282116,0.4280342,0.0012442182],"about_ca_topic_score_codex":0.00021974326,"about_ca_topic_score_gemma":0.000049151426,"teacher_disagreement_score":0.85058826,"about_ca_system_score_codex":0.0002284938,"about_ca_system_score_gemma":0.0011478185,"threshold_uncertainty_score":0.8102208},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W7117315753","doi":"10.1111/jebm.70099","title":"REporting Guidelines for aCupuncture‐Related AdverSe Event Case Reports (RECASE): Elaboration and Explanation","year":2025,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Acupuncture Treatment Research Studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Hamilton Health Sciences","funders":"Korea Institute of Oriental Medicine; Korea Health Industry Development Institute","keywords":"Guideline; Harm; Event (particle physics); Causality (physics); Adverse effect; Delphi method; MEDLINE; Delphi; Pharmacovigilance","score_opus":0.2498401227798475,"score_gpt":0.48979017272476094,"score_spread":0.23995004994491345,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7117315753","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.56582016,0.041566644,0.038025167,0.34880465,0.0019675465,0.0034629658,0.000006611402,0.0000812703,0.00026499305],"genre_scores_gemma":[0.98462385,0.001828964,0.0089996755,0.0025580472,0.0007255987,0.000055393524,0.000030216152,0.000019831954,0.0011583982],"study_design_codex":"not_applicable","study_design_gemma":"case_report","domain_scores_codex":[0.99502885,0.000107413754,0.0035113115,0.0003040807,0.00079411065,0.0002542118],"domain_scores_gemma":[0.9892659,0.0015983727,0.0037373956,0.00029944762,0.004868015,0.00023087126],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.006320622,0.0003477544,0.0011410222,0.00066840526,0.00018809823,0.0000130229855,0.00004303531,0.00013469481,0.000055953795],"category_scores_gemma":[0.101525925,0.00013807938,0.00025902,0.00055561453,0.00016246691,0.0003337972,0.000028152623,0.00031516445,3.7770937e-7],"study_design_candidate":"case_report","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008621381,0.0007127729,0.10434906,0.0051734275,0.003350311,0.25389728,0.002126747,0.0019126479,0.049188122,0.00026851037,0.45994627,0.11045347],"study_design_scores_gemma":[0.05490095,0.02976974,0.08233192,0.1364339,0.031211196,0.4306003,0.018366562,0.019343477,0.050618343,0.007029036,0.13766666,0.0017279319],"about_ca_topic_score_codex":0.000060426893,"about_ca_topic_score_gemma":0.00003274115,"teacher_disagreement_score":0.41880372,"about_ca_system_score_codex":0.00037877078,"about_ca_system_score_gemma":0.0009222589,"threshold_uncertainty_score":0.9060423},"labels":[],"label_agreement":null},{"id":"W7117582584","doi":"10.1111/jebm.70106","title":"The Necessity and Feasibility Assessment Tool of the Clinical Prediction Model for Individual Prognosis Before Its Startup: A Multi‐Sectoral Delphi Consensus Study","year":2025,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Artificial Intelligence in Healthcare and Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Provincial Laboratory of Public Health; University of Calgary","funders":"Army Medical University; General Hospital of People’s Liberation Army; Peking University Third Hospital; Peking University; Shandong University; Nanjing Medical University; Sichuan University; Beijing Children's Hospital, Capital Medical University; Fudan University; Capital Medical University; National Natural Science Foundation of China; West China Hospital, Sichuan University; Tianjin University","keywords":"Delphi method; Predictive modelling; Delphi; MEDLINE; Risk assessment","score_opus":0.6466986557540526,"score_gpt":0.5807877204602196,"score_spread":0.06591093529383296,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7117582584","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9281089,0.0012178172,0.0050872755,0.06161956,0.0012824798,0.0026591364,0.000014369893,0.000008052454,0.000002441116],"genre_scores_gemma":[0.9970123,0.00021615892,0.0013792551,0.0007841575,0.00048185707,0.000056535984,0.0000022696154,0.000007671229,0.000059775233],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99650085,0.00048165888,0.0018634485,0.00022676925,0.00071777677,0.00020948185],"domain_scores_gemma":[0.9930888,0.0034058685,0.0009863563,0.0003257532,0.0020413916,0.00015187425],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.008980425,0.00014836955,0.0005293577,0.000119631695,0.00032410343,0.000019927242,0.00021514876,0.00011913283,0.000006273788],"category_scores_gemma":[0.013241732,0.00006711084,0.00017346501,0.00031568776,0.0004578606,0.000116029245,0.00004421125,0.0005565588,9.643716e-8],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0015858433,0.0011282836,0.95703584,0.00031368167,0.00015824307,0.0000022472157,0.0017246595,0.00014702186,0.00013132041,0.000030808333,0.0021497863,0.035592236],"study_design_scores_gemma":[0.0016701298,0.0077038766,0.90245306,0.003501446,0.0011915979,0.000008768075,0.0063798195,0.075758725,0.0005274063,0.00066206115,0.00007984004,0.00006325997],"about_ca_topic_score_codex":0.00010664287,"about_ca_topic_score_gemma":0.00034490268,"teacher_disagreement_score":0.0756117,"about_ca_system_score_codex":0.0002010901,"about_ca_system_score_gemma":0.002909371,"threshold_uncertainty_score":0.99507016},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"qualitative","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"qualitative","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"}]}