{"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":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.1969918,0.56262505,0.09063973,0.021850685,0.0028049825,0.009472122,0.066775344,0.005508141,0.043332167],"genre_scores_gemma":[0.61807877,0.09863792,0.2576155,0.0009493515,0.0011885662,0.009006278,0.01268549,0.0006246274,0.0012135145],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.78986937,0.11838539,0.054269593,0.006664722,0.029514838,0.0012959933],"domain_scores_gemma":[0.26884204,0.6147547,0.07042522,0.01427464,0.030167127,0.0015362328],"candidate_categories":["metaresearch","bibliometrics"],"consensus_categories":[],"category_scores_codex":[0.14277764,0.0026441505,0.0060729356,0.15513745,0.0017452817,0.016756939,0.0028038998,0.0031258096,0.00835994],"category_scores_gemma":[0.63356954,0.001714444,0.008436172,0.12963445,0.0018659551,0.012238738,0.008451415,0.0016555515,0.000770418],"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":[0.0013355805,0.00014849196,0.06831377,0.21225347,0.009994734,0.0006717789,0.016948413,0.006973325,0.0019316003,0.023327017,0.02484769,0.6332541],"study_design_scores_gemma":[0.0019551152,0.0015592898,0.19471969,0.30325162,0.045266524,0.0041694944,0.024963941,0.05282988,0.008577334,0.14625639,0.21493791,0.0015129487],"about_ca_topic_score_codex":0.0040508215,"about_ca_topic_score_gemma":0.0039790147,"teacher_disagreement_score":0.8572224,"about_ca_system_score_codex":0.0059670843,"about_ca_system_score_gemma":0.01319337,"threshold_uncertainty_score":0.75508934},"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":"empirical","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.54703295,0.24112187,0.0055300784,0.005502528,0.00054897385,0.0006572648,0.0042809816,0.00007929581,0.19524603],"genre_scores_gemma":[0.96043736,0.034334235,0.0020980467,0.00028606522,0.00015677816,0.00022622004,0.0012185287,0.000014064088,0.0012286844],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.982081,0.011026379,0.0020440985,0.00079262146,0.003373174,0.00068266597],"domain_scores_gemma":[0.9538963,0.028307518,0.0077686002,0.0016970553,0.007861659,0.00046888017],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.013290055,0.00033713935,0.00046812723,0.014585497,0.00079924916,0.0023531155,0.00060498406,0.00043083975,0.0074764853],"category_scores_gemma":[0.051299218,0.000104882674,0.00074710615,0.021071697,0.0011295802,0.0022773207,0.0015069491,0.0003821121,0.00022810746],"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":[0.001311633,0.0004653421,0.26441902,0.018897893,0.003807588,0.00087355933,0.03423669,0.002364265,0.000601324,0.106364354,0.014674654,0.55198365],"study_design_scores_gemma":[0.00017266248,0.0013405501,0.7681578,0.014342222,0.0029000777,0.0015358949,0.08019087,0.0013034659,0.0014741138,0.013309129,0.11517348,0.00009975836],"about_ca_topic_score_codex":0.0049594725,"about_ca_topic_score_gemma":0.0043372046,"teacher_disagreement_score":0.014585497,"about_ca_system_score_codex":0.0035875277,"about_ca_system_score_gemma":0.002872756,"threshold_uncertainty_score":0.07028538},"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":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.047120776,0.15463823,0.07781922,0.5314178,0.02845033,0.055185575,0.013652365,0.0013091719,0.09040655],"genre_scores_gemma":[0.17070144,0.039823797,0.68651754,0.06515051,0.004117975,0.024843773,0.005610812,0.0003310483,0.0029030412],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.55611014,0.20364186,0.16495372,0.00498626,0.06706482,0.00324322],"domain_scores_gemma":[0.18111303,0.47764054,0.11058995,0.014322339,0.20487818,0.0114559755],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.30332038,0.0019361038,0.0041785897,0.020463191,0.00317358,0.010702836,0.010170023,0.018918514,0.0071654883],"category_scores_gemma":[0.7879986,0.0030248023,0.00862762,0.014437948,0.0034164202,0.0087761665,0.006631265,0.013834038,0.00328888],"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":[0.002112472,0.0008927308,0.058921542,0.13739519,0.015184653,0.0017266934,0.010456219,0.0049950187,0.0018777609,0.026724678,0.3022562,0.43745697],"study_design_scores_gemma":[0.0055597113,0.0014255545,0.04021833,0.4956194,0.02179432,0.0015061675,0.0071391743,0.0147699015,0.002363535,0.04995788,0.35834262,0.0013033885],"about_ca_topic_score_codex":0.012705057,"about_ca_topic_score_gemma":0.019874426,"teacher_disagreement_score":0.6966796,"about_ca_system_score_codex":0.011741436,"about_ca_system_score_gemma":0.030841755,"threshold_uncertainty_score":0.85913},"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":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.15310682,0.36315295,0.04719318,0.39235574,0.012031885,0.0041468535,0.004183537,0.00021885366,0.023610197],"genre_scores_gemma":[0.87280923,0.031662162,0.048008215,0.038357582,0.0034260987,0.003847437,0.0013522012,0.000040983647,0.00049610296],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.8112191,0.12993772,0.024256224,0.005380229,0.02704727,0.0021595208],"domain_scores_gemma":[0.47149542,0.32380426,0.10657877,0.010083202,0.08107333,0.0069649806],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.22454338,0.0007077347,0.002553045,0.004085634,0.0015380387,0.0057148663,0.0028826033,0.004325384,0.0012625795],"category_scores_gemma":[0.50247496,0.000549293,0.001973748,0.0062484583,0.0033811678,0.00748566,0.0020310024,0.0042382395,0.0003271742],"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":[0.0024323617,0.0013276146,0.35848197,0.045912117,0.0034384003,0.00014119895,0.003709301,0.005107698,0.00051071594,0.02653498,0.036804784,0.51559883],"study_design_scores_gemma":[0.0017718687,0.013982033,0.59785813,0.17861407,0.0060796947,0.00046552927,0.008817019,0.018414121,0.0042238305,0.05923225,0.109881304,0.00066026],"about_ca_topic_score_codex":0.010246943,"about_ca_topic_score_gemma":0.010431911,"teacher_disagreement_score":0.7754566,"about_ca_system_score_codex":0.008032629,"about_ca_system_score_gemma":0.020826455,"threshold_uncertainty_score":0.95627606},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.047594987,0.9475637,0.0019827974,0.0004318996,0.00036465397,0.00027066833,0.0011373344,0.00006923049,0.0005847646],"genre_scores_gemma":[0.6368829,0.3543518,0.0043143937,0.00074419775,0.00058548443,0.000847306,0.0015111583,0.000052959753,0.00070989295],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.99471647,0.0026474525,0.0010509975,0.00084381725,0.0004839769,0.00025733473],"domain_scores_gemma":[0.9900613,0.006938071,0.0014123588,0.0006062401,0.0007409164,0.00024106175],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.009144917,0.002409408,0.010413197,0.004174411,0.00073566084,0.0029346964,0.0018407474,0.0020859991,0.0033925297],"category_scores_gemma":[0.015499551,0.0016283294,0.04395048,0.0055458182,0.0005090956,0.0011120145,0.0012078551,0.0022315152,0.00032313075],"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":[0.0050799013,0.000058395133,0.024234636,0.042049684,0.92107606,0.00022726861,0.00006273995,0.0009026394,0.00029854217,0.000104972605,0.00046080988,0.005444248],"study_design_scores_gemma":[0.00072335266,0.00020869443,0.012134362,0.0025713702,0.9827466,0.00012725545,0.000028350316,0.0005567749,0.000091002454,0.00018723562,0.00060400815,0.000021017415],"about_ca_topic_score_codex":0.004845065,"about_ca_topic_score_gemma":0.0070035732,"teacher_disagreement_score":0.010413197,"about_ca_system_score_codex":0.0011076059,"about_ca_system_score_gemma":0.0013793967,"threshold_uncertainty_score":0.048363507},"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":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015923152,0.9094245,0.020600228,0.027582252,0.003281126,0.014769997,0.0036806562,0.00018611533,0.004551968],"genre_scores_gemma":[0.5286565,0.28856152,0.07050012,0.026325062,0.0032435143,0.07867394,0.0027117513,0.00033166012,0.0009960318],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.20404193,0.51803386,0.21192892,0.018774731,0.04504545,0.002175119],"domain_scores_gemma":[0.11560245,0.74767303,0.09019625,0.02031721,0.025179023,0.0010321158],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.5837778,0.0032627408,0.01802771,0.018407097,0.0042729974,0.014767172,0.006378443,0.006267817,0.006144984],"category_scores_gemma":[0.82465434,0.0033447582,0.014541748,0.025135128,0.008381182,0.019385848,0.008650056,0.004436077,0.0007784676],"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":[0.0027009032,0.00011666533,0.025359116,0.7685531,0.070980445,0.00033589697,0.0067268577,0.0009698062,0.00049588847,0.007162373,0.00543368,0.11116525],"study_design_scores_gemma":[0.005131751,0.0014183922,0.042080276,0.7079582,0.17876682,0.0007740536,0.003589564,0.0035249633,0.0020935666,0.025523793,0.028619893,0.0005187931],"about_ca_topic_score_codex":0.009148925,"about_ca_topic_score_gemma":0.015064589,"teacher_disagreement_score":0.9819723,"about_ca_system_score_codex":0.01910528,"about_ca_system_score_gemma":0.035429824,"threshold_uncertainty_score":0.5132761},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.037055016,0.0011209325,0.9239963,0.0029096217,0.0002617027,0.00020060847,0.0013871744,0.000342924,0.032725677],"genre_scores_gemma":[0.8178307,0.0023518323,0.14274196,0.0006183694,0.00032091016,0.0013056978,0.0011186006,0.00013501258,0.03357689],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9992256,0.00034030617,0.000048190897,0.0001133804,0.00016408961,0.000108350956],"domain_scores_gemma":[0.9970794,0.0019796675,0.00035269096,0.00009321743,0.00038174557,0.000113312555],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0016474036,0.0010758938,0.0010306936,0.000999744,0.00059628615,0.0015695848,0.0019973523,0.002333549,0.0063849287],"category_scores_gemma":[0.0084169805,0.000428549,0.001025481,0.00097267923,0.00097135577,0.0016801993,0.00092841976,0.0018035817,0.0008527198],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000022080661,0.000027543452,0.000608236,0.000053014643,0.000020395542,0.00011720182,0.000049406783,0.9156486,0.0004472923,0.07901099,0.0011084393,0.0028868665],"study_design_scores_gemma":[0.000013450774,0.000016010845,0.00010003897,0.000008045908,0.0000088946845,0.000020538304,0.000009996138,0.9869949,0.00005837208,0.0116065275,0.0011560484,0.0000071737118],"about_ca_topic_score_codex":0.019360447,"about_ca_topic_score_gemma":0.007498811,"teacher_disagreement_score":0.019360447,"about_ca_system_score_codex":0.002710397,"about_ca_system_score_gemma":0.0022654298,"threshold_uncertainty_score":0.03849554},"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":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0098837465,0.0990679,0.0027498598,0.8605,0.001170657,0.00010563061,0.00039785312,0.000074816795,0.026049621],"genre_scores_gemma":[0.5899032,0.22861814,0.05094728,0.12084478,0.0011066666,0.00019849312,0.00069142756,0.00009858895,0.0075915684],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9613332,0.014590838,0.0020963799,0.0016193761,0.012723199,0.0076369527],"domain_scores_gemma":[0.8537695,0.055909254,0.0062014083,0.0028125183,0.05136474,0.029942576],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0607685,0.00078702293,0.0013595135,0.006310034,0.016356783,0.026159447,0.0045589283,0.0071828407,0.004953438],"category_scores_gemma":[0.08345711,0.0007493283,0.0011929866,0.012837858,0.017223436,0.009035436,0.008606905,0.012792072,0.0003252516],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.00023979219,0.00029298675,0.02463149,0.008129464,0.00036458825,0.0009110826,0.019213647,0.0040731085,0.0008829008,0.37393782,0.13302708,0.4342961],"study_design_scores_gemma":[0.00018462657,0.00018388013,0.06002499,0.025380274,0.0004047402,0.00029248866,0.05528554,0.0041760937,0.0015265561,0.16667166,0.6853044,0.0005648226],"about_ca_topic_score_codex":0.98124456,"about_ca_topic_score_gemma":0.9912519,"teacher_disagreement_score":0.9392315,"about_ca_system_score_codex":0.22085322,"about_ca_system_score_gemma":0.69809276,"threshold_uncertainty_score":0.9036998},"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":"empirical","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93037313,0.02363166,0.0055591464,0.001822263,0.0002271577,0.000812854,0.004706915,0.00015182656,0.032715138],"genre_scores_gemma":[0.99195486,0.0028195695,0.0020085569,0.00009853585,0.00005380511,0.00021884475,0.0022771321,0.000026765143,0.0005419904],"study_design_codex":"observational","study_design_gemma":"systematic_review","domain_scores_codex":[0.93901175,0.036556937,0.009151825,0.0022103505,0.011212301,0.0018568942],"domain_scores_gemma":[0.73938096,0.10618225,0.06576283,0.011607126,0.07472255,0.0023443094],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.05410062,0.00037606494,0.0004634859,0.014022642,0.0006501042,0.0033851939,0.0011729621,0.00043942596,0.003291069],"category_scores_gemma":[0.15071113,0.0002507943,0.0012142366,0.021536008,0.0011393613,0.003202666,0.001754678,0.00058228587,0.00037140897],"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":[0.0016041446,0.00022352807,0.84207875,0.0020174116,0.0010138636,0.00013164456,0.011376917,0.000991149,0.00040576304,0.006704695,0.003976279,0.12947588],"study_design_scores_gemma":[0.00005974776,0.0005234461,0.97441703,0.0008086297,0.00040902672,0.00018227226,0.012629231,0.0010170218,0.0005960989,0.0004304933,0.008880682,0.000046408666],"about_ca_topic_score_codex":0.008421892,"about_ca_topic_score_gemma":0.0057923244,"teacher_disagreement_score":0.05410062,"about_ca_system_score_codex":0.004649131,"about_ca_system_score_gemma":0.0029613797,"threshold_uncertainty_score":0.2861148},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0065330444,0.9924198,0.00031068662,0.00010845223,0.00011022968,0.00021257541,0.00015007141,0.000015025392,0.00014003436],"genre_scores_gemma":[0.16464779,0.8313723,0.0017918899,0.0005539358,0.00020618696,0.0008436908,0.00033518265,0.000021713258,0.00022729783],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9914384,0.004279159,0.002498064,0.0006789762,0.000926657,0.00017887088],"domain_scores_gemma":[0.985631,0.010586211,0.0024422344,0.00034327238,0.0008069729,0.00019031134],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.010881953,0.002438958,0.018902201,0.0045520626,0.00046294215,0.0026218886,0.0017170911,0.002084224,0.0021215545],"category_scores_gemma":[0.02383207,0.0014937794,0.030804455,0.005826704,0.00066304044,0.0015734658,0.0010915165,0.0016099893,0.00017751056],"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":[0.0041539245,0.00006159337,0.002525451,0.37658224,0.5981347,0.000115887284,0.00009442285,0.00040794702,0.00021490645,0.00008601612,0.00033772486,0.01728526],"study_design_scores_gemma":[0.0013190847,0.0005047939,0.0032654512,0.020431574,0.972758,0.00009849858,0.000050802144,0.00021505413,0.00014989608,0.00009969559,0.0010828828,0.000024242598],"about_ca_topic_score_codex":0.0047560576,"about_ca_topic_score_gemma":0.009545323,"teacher_disagreement_score":0.018902201,"about_ca_system_score_codex":0.0019461188,"about_ca_system_score_gemma":0.0023155217,"threshold_uncertainty_score":0.057549894},"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":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6756814,0.0036439463,0.23412034,0.0050073382,0.00022014468,0.00045087215,0.004780907,0.0019013771,0.074193686],"genre_scores_gemma":[0.9329338,0.0012602534,0.061065625,0.00006383539,0.00004946539,0.00023221679,0.00115578,0.00008489711,0.003154049],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9991817,0.00043809044,0.00005334327,0.000093634815,0.00017952279,0.00005368372],"domain_scores_gemma":[0.9879529,0.009265362,0.0012293119,0.00033438762,0.0009954227,0.00022259781],"candidate_categories":["metaresearch","bibliometrics"],"consensus_categories":[],"category_scores_codex":[0.0015794729,0.00040217608,0.0001872903,0.008236869,0.00080129446,0.0022748485,0.00040928702,0.0004900441,0.00827344],"category_scores_gemma":[0.011396889,0.00017082211,0.0004939288,0.006540817,0.0006016779,0.0029876016,0.0010472379,0.00038176693,0.0004042589],"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":[0.00051071454,0.00019210605,0.17922783,0.0026659574,0.00035703843,0.001783181,0.06226995,0.056149516,0.016655607,0.26135626,0.028742947,0.3900889],"study_design_scores_gemma":[0.00012938915,0.00030258243,0.17968442,0.0022708427,0.000493496,0.0020438978,0.04668961,0.31337708,0.021570938,0.23476054,0.19850783,0.00016943691],"about_ca_topic_score_codex":0.0044742934,"about_ca_topic_score_gemma":0.003441599,"teacher_disagreement_score":0.99842054,"about_ca_system_score_codex":0.0010891655,"about_ca_system_score_gemma":0.0011097145,"threshold_uncertainty_score":0.027677476},"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":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0041255085,0.57271504,0.14987038,0.022830987,0.0077410927,0.2205976,0.015909947,0.001815848,0.004393671],"genre_scores_gemma":[0.022989688,0.12360175,0.38881603,0.003954916,0.00077292824,0.4530863,0.005471343,0.00042258567,0.00088450586],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.39885274,0.2556985,0.28056124,0.009250412,0.053229816,0.0024072598],"domain_scores_gemma":[0.35398498,0.37138912,0.12476554,0.027902575,0.118053734,0.0039040905],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.40250728,0.0051605464,0.022766903,0.039733667,0.0045094974,0.011550509,0.008707971,0.008673324,0.0063858875],"category_scores_gemma":[0.6324881,0.006277239,0.034863282,0.032064326,0.006245182,0.009400433,0.010698547,0.009583072,0.0021066794],"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":[0.00068015815,0.00010223473,0.0035752782,0.8108572,0.021000205,0.00015244236,0.0019304429,0.00096080994,0.00058392505,0.004427847,0.031129496,0.1246],"study_design_scores_gemma":[0.0027774295,0.0004315407,0.00705586,0.8477969,0.05146795,0.0006140049,0.0012507263,0.0029872588,0.0012822909,0.013136068,0.070712574,0.0004873023],"about_ca_topic_score_codex":0.005552594,"about_ca_topic_score_gemma":0.007959792,"teacher_disagreement_score":0.5974927,"about_ca_system_score_codex":0.018736683,"about_ca_system_score_gemma":0.07133729,"threshold_uncertainty_score":0.7368149},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.006139718,0.008151572,0.040712032,0.7655734,0.0055215633,0.0005238284,0.000099275305,0.00060590537,0.17267278],"genre_scores_gemma":[0.3631783,0.034500994,0.19281197,0.29919025,0.010225861,0.0020776542,0.00028827056,0.00029709836,0.09742956],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9916898,0.0041619665,0.0003049345,0.00043255364,0.0026027851,0.000807907],"domain_scores_gemma":[0.9886595,0.003006073,0.0011221085,0.0005820435,0.0022059784,0.004424325],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.010071029,0.0010834773,0.0004883524,0.0025082522,0.00482941,0.008116916,0.0024349773,0.006736906,0.015407174],"category_scores_gemma":[0.014456838,0.00033895377,0.00067894714,0.0014022205,0.0058018635,0.0052135857,0.010751003,0.007949331,0.0034417508],"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":[0.000055615124,0.0021855547,0.0034010075,0.0016261284,0.0000827401,0.0002707641,0.0062320805,0.0009783759,0.00093396596,0.2089675,0.2802271,0.49503917],"study_design_scores_gemma":[0.00014794797,0.00072714005,0.00797909,0.0041014906,0.00006813279,0.0004661696,0.00885495,0.001772711,0.0010590253,0.2034084,0.7713393,0.00007563328],"about_ca_topic_score_codex":0.0042931116,"about_ca_topic_score_gemma":0.0059819915,"teacher_disagreement_score":0.015407174,"about_ca_system_score_codex":0.0055303397,"about_ca_system_score_gemma":0.028705714,"threshold_uncertainty_score":0.05326128},"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":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.13073376,0.6988926,0.034193408,0.0125973355,0.002063158,0.019586476,0.07283429,0.001457168,0.027641844],"genre_scores_gemma":[0.53378844,0.25402218,0.1431528,0.0010276453,0.002085076,0.03847721,0.02518695,0.0005626016,0.0016971404],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.54152125,0.18162398,0.15404932,0.011633272,0.10897249,0.0021997348],"domain_scores_gemma":[0.20652771,0.63526577,0.0912378,0.014381857,0.0505149,0.0020719296],"candidate_categories":["metaresearch","bibliometrics"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.21858351,0.003114114,0.014450372,0.32925162,0.0036031543,0.015723512,0.0037038026,0.003193752,0.005210003],"category_scores_gemma":[0.7191654,0.0019523463,0.014832025,0.38180503,0.0035410158,0.011538462,0.0096030235,0.0018813753,0.0007693446],"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":[0.0013000667,0.00023501081,0.14147258,0.40992954,0.032569576,0.00084094657,0.0068288585,0.0054160776,0.001104067,0.01291043,0.021215914,0.36617693],"study_design_scores_gemma":[0.001985714,0.0018250492,0.36731339,0.23376447,0.11069733,0.0052399533,0.013633277,0.030187562,0.004120558,0.058763325,0.17101379,0.0014556175],"about_ca_topic_score_codex":0.0055219308,"about_ca_topic_score_gemma":0.0048767226,"teacher_disagreement_score":0.7814165,"about_ca_system_score_codex":0.013324483,"about_ca_system_score_gemma":0.027729314,"threshold_uncertainty_score":0.96362567},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000011948619,0.0028026728,0.00006464421,0.0121199135,0.9843813,0.000034159286,0.000062029336,0.000052918902,0.00047040975],"genre_scores_gemma":[0.00020487537,0.005517155,0.000121070756,0.0150625855,0.9747218,0.000075025426,0.00011208261,0.000071625815,0.0041136215],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9820093,0.002878812,0.002521183,0.0023087694,0.008933683,0.0013482135],"domain_scores_gemma":[0.9314374,0.01945311,0.005001766,0.0020062185,0.0349218,0.0071797916],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.018110873,0.006392626,0.010906514,0.00911949,0.005121207,0.017916616,0.00975084,0.018752737,0.049131613],"category_scores_gemma":[0.09288609,0.0025925362,0.007787449,0.0039891,0.0046231877,0.009648431,0.003092197,0.030126855,0.032429397],"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":[0.00006224142,0.000017396169,0.000027174156,0.0007185419,0.000032456,0.000043084274,0.000009405675,0.000017280227,0.000034460943,0.00018580475,0.9927049,0.0061472543],"study_design_scores_gemma":[0.0002098506,0.00007854657,0.00061486766,0.0032686642,0.00016230663,0.0003080209,0.000052913918,0.00020433926,0.00017252722,0.001306912,0.99356407,0.000056963254],"about_ca_topic_score_codex":0.0033253566,"about_ca_topic_score_gemma":0.005275771,"teacher_disagreement_score":0.049131613,"about_ca_system_score_codex":0.007926234,"about_ca_system_score_gemma":0.00900342,"threshold_uncertainty_score":0.16436166},"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":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00053244166,0.93965244,0.02832219,0.025003573,0.0051043234,0.0003113278,0.00023411452,0.00024479037,0.0005948313],"genre_scores_gemma":[0.03987218,0.8340078,0.09378235,0.01916493,0.00950727,0.0023615742,0.00032032825,0.0003755224,0.000608009],"study_design_codex":"meta_analysis","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.72308266,0.23657039,0.022803558,0.005049174,0.011685335,0.0008088314],"domain_scores_gemma":[0.6275851,0.3305184,0.016283335,0.01707132,0.007438306,0.0011035909],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.20148939,0.0061765243,0.038735855,0.0139235025,0.001498174,0.012446804,0.008192836,0.008221423,0.0031554354],"category_scores_gemma":[0.4508076,0.003937359,0.022462545,0.016974175,0.003351887,0.013804184,0.0044001797,0.015019799,0.001173507],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019767277,0.00011789552,0.0038320073,0.21429384,0.41177753,0.00053828757,0.0009861551,0.0053408644,0.0005905377,0.033346873,0.054838743,0.2723605],"study_design_scores_gemma":[0.003688393,0.0008609186,0.0032631333,0.19572935,0.3881942,0.0009337055,0.00081194186,0.009289515,0.0016532475,0.29663488,0.09820433,0.0007364737],"about_ca_topic_score_codex":0.0027837101,"about_ca_topic_score_gemma":0.0045883283,"teacher_disagreement_score":0.7985106,"about_ca_system_score_codex":0.006146799,"about_ca_system_score_gemma":0.010114182,"threshold_uncertainty_score":0.98470575},"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":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0048591155,0.9801657,0.00046243146,0.0073923552,0.0007009448,0.00018541669,0.00092425814,0.000019834915,0.005289889],"genre_scores_gemma":[0.20682628,0.76938707,0.0023050054,0.016773261,0.001971318,0.0007713945,0.00086678285,0.00006745875,0.0010314487],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.8821295,0.06150716,0.024887273,0.0049762274,0.02406428,0.0024355827],"domain_scores_gemma":[0.50140744,0.42807385,0.054275446,0.00440596,0.010066863,0.0017703298],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.07520523,0.0010769694,0.0049041887,0.005715394,0.0006800435,0.004251729,0.0024927396,0.004044144,0.016066132],"category_scores_gemma":[0.2881312,0.001171597,0.011704305,0.0061458694,0.002313018,0.0052349716,0.0033624843,0.006233675,0.00056444947],"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":[0.012088323,0.0002627193,0.014242736,0.52249354,0.08927928,0.0005173332,0.000696114,0.0024894068,0.00049886876,0.020297732,0.02137724,0.31575668],"study_design_scores_gemma":[0.005245076,0.0021763956,0.059577618,0.5134453,0.18575166,0.0012649556,0.0010481471,0.00093277585,0.001397934,0.01653861,0.21239206,0.0002295343],"about_ca_topic_score_codex":0.008958074,"about_ca_topic_score_gemma":0.012326477,"teacher_disagreement_score":0.07520523,"about_ca_system_score_codex":0.008273143,"about_ca_system_score_gemma":0.006951948,"threshold_uncertainty_score":0.39772803},"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":"commentary","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005461935,0.12417921,0.0038016788,0.85839653,0.0037799715,0.00014072981,0.000049573777,0.000051398994,0.009054743],"genre_scores_gemma":[0.1911886,0.3581189,0.096536696,0.3315118,0.017458329,0.0019102057,0.00021497789,0.00015940727,0.0029010295],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.7638999,0.18850139,0.017545495,0.004969473,0.019558456,0.0055252686],"domain_scores_gemma":[0.22802965,0.6922306,0.018218182,0.018217437,0.026772242,0.016531868],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.3600235,0.0016619151,0.00474925,0.01521617,0.0061051296,0.038969535,0.007373513,0.028577665,0.010707589],"category_scores_gemma":[0.43306172,0.0017180167,0.0034077552,0.01079046,0.042729475,0.049141392,0.043189995,0.03464449,0.0010179597],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018884768,0.00024687257,0.0027161601,0.019368151,0.0005518298,0.00033609537,0.004505904,0.0010124858,0.00015891578,0.54679435,0.08497509,0.33914542],"study_design_scores_gemma":[0.0002887614,0.00032183196,0.002217109,0.1213364,0.0003275333,0.0004164297,0.00990273,0.00068240403,0.00024084089,0.5055397,0.35849696,0.00022927784],"about_ca_topic_score_codex":0.0063118455,"about_ca_topic_score_gemma":0.008774093,"teacher_disagreement_score":0.3600235,"about_ca_system_score_codex":0.020057524,"about_ca_system_score_gemma":0.083854266,"threshold_uncertainty_score":0.78920496},"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":"methods","genre_gemma":"other","domain_codex":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"other","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.003929837,0.02323085,0.50143385,0.063971184,0.024519388,0.292123,0.017971395,0.0060104607,0.066810094],"genre_scores_gemma":[0.0144366035,0.006979391,0.62374073,0.020631582,0.0026453466,0.31076202,0.007012089,0.0012032844,0.012588953],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.25275534,0.42653775,0.22923453,0.015678132,0.062119447,0.013674822],"domain_scores_gemma":[0.16317223,0.4062825,0.08125337,0.15839419,0.18047199,0.01042564],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.64706874,0.005531512,0.018701008,0.02455129,0.017742457,0.031920854,0.029204229,0.05752732,0.012844706],"category_scores_gemma":[0.66222936,0.018211206,0.026847422,0.031890877,0.029396372,0.015181091,0.011334592,0.07725605,0.02657185],"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":[0.009580787,0.0018776937,0.003621982,0.046579603,0.0041507166,0.0012745092,0.02249246,0.0027066318,0.0040617795,0.2789238,0.45191854,0.17281152],"study_design_scores_gemma":[0.01309203,0.0018534639,0.016915318,0.057176583,0.0058790776,0.0021478792,0.0032379036,0.004222302,0.008142403,0.09832635,0.78675115,0.0022556053],"about_ca_topic_score_codex":0.03259796,"about_ca_topic_score_gemma":0.0487968,"teacher_disagreement_score":0.35293126,"about_ca_system_score_codex":0.027322466,"about_ca_system_score_gemma":0.17600673,"threshold_uncertainty_score":0.43522716},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0022351511,0.99620605,0.00016286848,0.00027654573,0.00011524615,0.00049943116,0.000303052,0.000009110692,0.00019259407],"genre_scores_gemma":[0.037780274,0.958998,0.0008189768,0.000622691,0.000119195414,0.0012534646,0.00027510108,0.000006069197,0.00012624353],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.990072,0.0041618817,0.003215645,0.00075856876,0.0015152823,0.0002765393],"domain_scores_gemma":[0.962078,0.027754135,0.007013236,0.00046431358,0.0022887797,0.00040158664],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.010085773,0.0014166286,0.008105713,0.00665778,0.00058907317,0.0026834286,0.0018050907,0.0021119888,0.0036068351],"category_scores_gemma":[0.0551216,0.0012116099,0.008407899,0.006968447,0.0008572723,0.002404781,0.0013106301,0.0013834401,0.00021364786],"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":[0.00026750134,0.000020981397,0.0008389246,0.96930546,0.010891688,0.00007095607,0.00016911143,0.00007427089,0.00008221688,0.000103383536,0.0007476037,0.017427897],"study_design_scores_gemma":[0.0005162092,0.00037866068,0.00576825,0.8806465,0.10317069,0.00031521884,0.00032950766,0.00011897306,0.00017623117,0.00018577119,0.008354866,0.00003920678],"about_ca_topic_score_codex":0.008699069,"about_ca_topic_score_gemma":0.021238297,"teacher_disagreement_score":0.010085773,"about_ca_system_score_codex":0.00364461,"about_ca_system_score_gemma":0.013640132,"threshold_uncertainty_score":0.053339303},"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":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0044916645,0.015452376,0.97279245,0.0012908714,0.00041149746,0.0018395671,0.0016277073,0.0010269195,0.0010670333],"genre_scores_gemma":[0.1699499,0.011227622,0.79523313,0.0016005517,0.0002903808,0.015922248,0.0018206416,0.0005234272,0.003432032],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9208911,0.0701295,0.0024100253,0.0034485264,0.0026599588,0.00046095776],"domain_scores_gemma":[0.9137672,0.07609789,0.003092418,0.0047526066,0.0021217393,0.00016808894],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.08123306,0.00249038,0.0044373656,0.003960289,0.0005783334,0.001991793,0.004152166,0.0029236278,0.01113078],"category_scores_gemma":[0.1430289,0.0010794994,0.012291114,0.0053104213,0.0010401683,0.0028600811,0.0016143634,0.004046369,0.0013166305],"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":[0.005859751,0.0004206343,0.015930146,0.039652277,0.04830032,0.0016657126,0.0011857261,0.27541932,0.0027118681,0.1517268,0.02955569,0.42757162],"study_design_scores_gemma":[0.0027843665,0.002628888,0.007037178,0.005756286,0.030588007,0.0014019847,0.00033079725,0.63489765,0.0028105206,0.25699914,0.05440002,0.00036516634],"about_ca_topic_score_codex":0.004802062,"about_ca_topic_score_gemma":0.0030682567,"teacher_disagreement_score":0.9187669,"about_ca_system_score_codex":0.001798637,"about_ca_system_score_gemma":0.0022381851,"threshold_uncertainty_score":0.42960662},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0042072684,0.9944622,0.00019913467,0.00022320004,0.0000969381,0.00030834522,0.00036316807,0.000004819185,0.00013501679],"genre_scores_gemma":[0.05861397,0.93750626,0.0014132726,0.0006230214,0.00018319713,0.0010999012,0.00045969486,0.0000058317733,0.0000948416],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9803388,0.0056278026,0.009830977,0.0012383267,0.0025996412,0.00036441715],"domain_scores_gemma":[0.9019985,0.0685286,0.021903317,0.0014637813,0.005533085,0.00057274307],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.016763926,0.0017084487,0.008548334,0.016315391,0.00090229744,0.003244498,0.0017196005,0.0024226902,0.0031128014],"category_scores_gemma":[0.0688396,0.0015633606,0.008524862,0.016925296,0.0012188769,0.0027920618,0.0015001813,0.0011384077,0.00023579206],"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":[0.00022310358,0.000027749315,0.0069718305,0.9581485,0.019587394,0.00017810112,0.0002974256,0.00006776002,0.00014564498,0.00010942574,0.00044609673,0.013797138],"study_design_scores_gemma":[0.00031668536,0.00040188374,0.024449598,0.7961835,0.16843265,0.0010481314,0.0009602747,0.00012381127,0.00019904459,0.00031850068,0.0075062797,0.000059662892],"about_ca_topic_score_codex":0.005617793,"about_ca_topic_score_gemma":0.014399854,"teacher_disagreement_score":0.016763926,"about_ca_system_score_codex":0.0025394727,"about_ca_system_score_gemma":0.010279061,"threshold_uncertainty_score":0.0886572},"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":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.025333693,0.3729993,0.11410384,0.33952105,0.024176518,0.032064304,0.016360603,0.00178045,0.07366028],"genre_scores_gemma":[0.22123723,0.14170311,0.5320304,0.04960326,0.00320531,0.034691263,0.013984469,0.00048223173,0.0030627458],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.6208403,0.18067628,0.1245374,0.008480007,0.062068097,0.0033979078],"domain_scores_gemma":[0.3140572,0.34212103,0.081178635,0.018609045,0.23643838,0.007595793],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.24809413,0.0014314586,0.00368434,0.017925955,0.002930544,0.007367384,0.008078688,0.0060740295,0.0041958177],"category_scores_gemma":[0.55639124,0.0015370145,0.0057533714,0.015474926,0.0042697666,0.0042315535,0.0072385212,0.0069415076,0.001299115],"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":[0.0004883709,0.00037906898,0.028974019,0.1476045,0.0027796554,0.00032747112,0.01007636,0.002521656,0.0012025438,0.02180385,0.19401737,0.58982515],"study_design_scores_gemma":[0.0012769906,0.0006046991,0.036635462,0.4146184,0.00530075,0.00080009137,0.0066539017,0.0035943012,0.002227659,0.034745622,0.49310827,0.0004339439],"about_ca_topic_score_codex":0.031419113,"about_ca_topic_score_gemma":0.023912942,"teacher_disagreement_score":0.75190586,"about_ca_system_score_codex":0.015930833,"about_ca_system_score_gemma":0.07604656,"threshold_uncertainty_score":0.9272338},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000018651996,0.0021907843,0.00005913002,0.01748232,0.979779,0.00001897829,0.00003695082,0.000036683203,0.00037752197],"genre_scores_gemma":[0.00025578536,0.0037969893,0.00009841205,0.017776314,0.97505707,0.00003708919,0.000053302138,0.000045460136,0.002879504],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9871777,0.0019442943,0.0019079816,0.0016870113,0.006416499,0.0008664231],"domain_scores_gemma":[0.9384785,0.020442586,0.004529105,0.0015587802,0.028007437,0.006983677],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.012841127,0.0046691806,0.0057613472,0.005465437,0.0044018095,0.013062497,0.0054057105,0.015388871,0.023800371],"category_scores_gemma":[0.06981748,0.0015596005,0.004579632,0.0024971326,0.0031548026,0.0070361537,0.0020114484,0.024973126,0.016207373],"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":[0.00004835951,0.0000139149115,0.00003132208,0.00042744345,0.000020300773,0.000057454905,0.000011640067,0.00001090094,0.000035908066,0.00013643217,0.9948132,0.0043931967],"study_design_scores_gemma":[0.00012209342,0.00009222749,0.00057796313,0.0020192615,0.00010362436,0.0004355208,0.00007498679,0.00015043332,0.00019249598,0.00092517503,0.9952667,0.000039636336],"about_ca_topic_score_codex":0.0015550427,"about_ca_topic_score_gemma":0.0029906784,"teacher_disagreement_score":0.023800371,"about_ca_system_score_codex":0.004766275,"about_ca_system_score_gemma":0.00566943,"threshold_uncertainty_score":0.07962018},"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":"review","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008206058,0.36210918,0.07258599,0.28455237,0.025759554,0.044385012,0.013766873,0.0012793024,0.18735565],"genre_scores_gemma":[0.057051737,0.33004707,0.39361733,0.096177146,0.003250827,0.048288047,0.02482602,0.00024577277,0.04649606],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.98978335,0.0032382,0.0028594625,0.00046686165,0.0032263263,0.00042572824],"domain_scores_gemma":[0.9867929,0.0031372055,0.0012224833,0.0003339815,0.0077571603,0.0007562969],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.012258971,0.0014606849,0.0018897986,0.007254741,0.001989027,0.002456045,0.00424745,0.0062852777,0.0077975695],"category_scores_gemma":[0.026626393,0.00072199886,0.0035509262,0.005361933,0.0013411912,0.0026156774,0.002434462,0.004615641,0.0035261922],"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":[0.0000762751,0.00021554106,0.0018833277,0.027444068,0.00024761588,0.0011175221,0.0023958012,0.00085291255,0.001290109,0.009456249,0.6070069,0.34801364],"study_design_scores_gemma":[0.00046267427,0.00014802747,0.007623453,0.062702246,0.0009937198,0.0018197291,0.0009904315,0.0012115127,0.000674055,0.007870164,0.9153438,0.00016011232],"about_ca_topic_score_codex":0.041379318,"about_ca_topic_score_gemma":0.057707597,"teacher_disagreement_score":0.041379318,"about_ca_system_score_codex":0.006050197,"about_ca_system_score_gemma":0.046677873,"threshold_uncertainty_score":0.08227694},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.975603,0.00023705531,0.009253485,0.00049961125,0.0000868783,0.002665418,0.003297337,0.00029316734,0.008064087],"genre_scores_gemma":[0.963907,0.00015355313,0.027224489,0.0002978352,0.000054259643,0.0033745775,0.0037522705,0.000045797646,0.0011902376],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9769702,0.010372449,0.0053381515,0.0010029186,0.0055808,0.0007354501],"domain_scores_gemma":[0.9167365,0.046856012,0.013053379,0.00625748,0.01586096,0.001235653],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.029676732,0.00048297585,0.00077190937,0.003477973,0.0007392097,0.001691982,0.001359349,0.0008911599,0.003708544],"category_scores_gemma":[0.08816238,0.00032305773,0.0016053903,0.0020020858,0.00074529456,0.0014719723,0.0021683688,0.00080105703,0.0010772096],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014759636,0.0021515992,0.79297006,0.0005232818,0.00035812327,0.00023412188,0.0033304254,0.0017180984,0.0020907172,0.0010852902,0.006099553,0.18796279],"study_design_scores_gemma":[0.0002702068,0.002000195,0.97041243,0.0003074496,0.0001499276,0.0004069198,0.0016006354,0.0099844225,0.0041163242,0.0009749842,0.009670394,0.00010623655],"about_ca_topic_score_codex":0.0024010318,"about_ca_topic_score_gemma":0.0039724135,"teacher_disagreement_score":0.029676732,"about_ca_system_score_codex":0.0012205587,"about_ca_system_score_gemma":0.0023591511,"threshold_uncertainty_score":0.15694743},"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":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.28861022,0.12837343,0.08072323,0.19943029,0.0087020695,0.18105423,0.021973923,0.0019381546,0.0891945],"genre_scores_gemma":[0.46350065,0.03325878,0.4180102,0.014277374,0.00082337373,0.058583345,0.008460004,0.0003158893,0.0027703198],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.49526426,0.27855292,0.14614019,0.0055758953,0.06972683,0.004739949],"domain_scores_gemma":[0.23128143,0.2940514,0.10369073,0.022816736,0.33964768,0.008512042],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.35822225,0.0010769146,0.0022216528,0.018422348,0.003006346,0.0065309526,0.0051099802,0.0031025049,0.0014107602],"category_scores_gemma":[0.66750586,0.0014734209,0.003525053,0.014776644,0.0018253459,0.005579084,0.0069774142,0.003490695,0.0005249034],"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":[0.0005571263,0.0010643721,0.09610815,0.05920277,0.0015604286,0.0012121148,0.042516254,0.002389736,0.0013296122,0.0046410635,0.11940442,0.67001396],"study_design_scores_gemma":[0.001813884,0.003859293,0.20154867,0.29296726,0.0049698623,0.0017783433,0.045200974,0.007459406,0.006182219,0.0064950134,0.4268554,0.00086979667],"about_ca_topic_score_codex":0.01927835,"about_ca_topic_score_gemma":0.040567234,"teacher_disagreement_score":0.64177775,"about_ca_system_score_codex":0.016144857,"about_ca_system_score_gemma":0.074546985,"threshold_uncertainty_score":0.79142624},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9984529,0.0002529172,0.0005744364,0.000052825624,0.000009457709,0.000075016935,0.00036649444,0.000003854754,0.00021225902],"genre_scores_gemma":[0.9984694,0.00011993817,0.00068275427,0.000072587405,0.000011268733,0.000116968346,0.00046532872,0.0000018981327,0.00005974238],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99627125,0.0022898812,0.00038691342,0.0004241744,0.00048690123,0.00014082312],"domain_scores_gemma":[0.9873052,0.0057480433,0.00407526,0.0009969145,0.0011965964,0.00067797326],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.007871119,0.00031863744,0.0004296143,0.00080197235,0.00055188313,0.0007438439,0.00039816793,0.0007366349,0.0011475409],"category_scores_gemma":[0.019036371,0.00044283734,0.00065372366,0.0009897226,0.00036018412,0.0009349533,0.0005929382,0.00091644254,0.0002903806],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005127995,0.00011879748,0.9987185,0.000013818063,0.0000814197,0.000014542459,0.00007054403,0.00002973142,0.000034983095,0.000010370672,0.000051166127,0.00080482],"study_design_scores_gemma":[0.00003353064,0.0011129107,0.9973334,0.000023776853,0.00010810031,0.00022670889,0.00033808098,0.0004745087,0.00007150058,0.000033478842,0.00023601191,0.000007921352],"about_ca_topic_score_codex":0.0019564247,"about_ca_topic_score_gemma":0.0020087804,"teacher_disagreement_score":0.007871119,"about_ca_system_score_codex":0.00035065916,"about_ca_system_score_gemma":0.00059799216,"threshold_uncertainty_score":0.04162699},"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":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0001422633,0.9911589,0.0013366702,0.0043965313,0.001190212,0.0003153993,0.00018528658,0.000012796143,0.0012619306],"genre_scores_gemma":[0.0027934825,0.98718655,0.0045714714,0.0031205975,0.00096987357,0.0008087774,0.00017456386,0.000011312509,0.00036329427],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.974441,0.013145176,0.0071568387,0.0010698582,0.003790113,0.00039686513],"domain_scores_gemma":[0.8957005,0.08491781,0.009204397,0.0020416495,0.0075040646,0.00063149014],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.026962938,0.0017831849,0.003730455,0.0131987035,0.0013695374,0.006360547,0.0021267436,0.0031790729,0.0073897387],"category_scores_gemma":[0.12039863,0.0010333394,0.0051721185,0.012728142,0.0022979532,0.0067801434,0.003587884,0.0044847983,0.0016673323],"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":[0.0002223438,0.0000639342,0.00073591364,0.60472214,0.002212701,0.00035621534,0.0016978208,0.00039480373,0.0004054721,0.032371502,0.057974145,0.29884285],"study_design_scores_gemma":[0.0000693949,0.00009055799,0.0008996477,0.6952534,0.0037543443,0.0008621451,0.0008423254,0.00014271792,0.00018511782,0.0079784645,0.28986382,0.000058084926],"about_ca_topic_score_codex":0.0032090673,"about_ca_topic_score_gemma":0.0076189376,"teacher_disagreement_score":0.97303706,"about_ca_system_score_codex":0.005259535,"about_ca_system_score_gemma":0.020902032,"threshold_uncertainty_score":0.14259535},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0055987802,0.99238974,0.00057301275,0.00019356952,0.00018540613,0.0004182972,0.0004217171,0.00002750662,0.00019195731],"genre_scores_gemma":[0.19565316,0.7965654,0.0033563206,0.00092754007,0.00042150618,0.0018385728,0.0007854997,0.00003691317,0.0004150764],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9864806,0.0064623496,0.0037343367,0.0012637526,0.001652924,0.00040595786],"domain_scores_gemma":[0.97840357,0.015024285,0.0038368287,0.00063767034,0.001831361,0.0002663188],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.016240463,0.0035939747,0.02279907,0.00817861,0.0008222488,0.004229107,0.0025391574,0.002613125,0.003426657],"category_scores_gemma":[0.036360987,0.001965636,0.038902763,0.008649692,0.00091497146,0.0021792713,0.0015943508,0.002018351,0.00032812628],"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":[0.0010665591,0.000023610339,0.0037821515,0.33245045,0.65626585,0.00013683662,0.00006144947,0.00042388533,0.00019081827,0.0000818897,0.00039615404,0.0051202895],"study_design_scores_gemma":[0.0002997318,0.00012539324,0.0026367004,0.020095723,0.975605,0.00009028757,0.000033224507,0.00018773541,0.00008588385,0.00010183875,0.00072076917,0.000017709837],"about_ca_topic_score_codex":0.006168094,"about_ca_topic_score_gemma":0.013898797,"teacher_disagreement_score":0.02279907,"about_ca_system_score_codex":0.0028719285,"about_ca_system_score_gemma":0.003948783,"threshold_uncertainty_score":0.0858888},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.018703816,0.976478,0.0022011753,0.0004328844,0.0004762453,0.00037732278,0.0008576103,0.00007076365,0.00040215053],"genre_scores_gemma":[0.48621437,0.50142896,0.0058652386,0.0014764789,0.0008083226,0.0016024796,0.0016205683,0.000084615924,0.00089893356],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.99200857,0.004492511,0.0016130265,0.0009350776,0.0006834864,0.00026729263],"domain_scores_gemma":[0.9910313,0.0063548144,0.0011255881,0.0004394666,0.00086171157,0.00018709971],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01107393,0.0040341695,0.017719857,0.006758961,0.0007989796,0.0034867378,0.001997629,0.0025625485,0.00428841],"category_scores_gemma":[0.01977846,0.0021011613,0.052888997,0.00663948,0.00051457144,0.0014321416,0.0014275992,0.0029533529,0.00040753698],"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":[0.0028512266,0.00003831136,0.005119854,0.05128268,0.9354557,0.00015830593,0.00003681529,0.00079261255,0.00024869893,0.00008075682,0.00036659263,0.0035685701],"study_design_scores_gemma":[0.00044595494,0.0001580835,0.0033200188,0.0025982668,0.9922746,0.00007237327,0.000018169436,0.00036961798,0.000096959804,0.00013574715,0.0004948675,0.000015362499],"about_ca_topic_score_codex":0.005628656,"about_ca_topic_score_gemma":0.008808332,"teacher_disagreement_score":0.017719857,"about_ca_system_score_codex":0.0020655931,"about_ca_system_score_gemma":0.0017103308,"threshold_uncertainty_score":0.05856526},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.01165391,0.49404025,0.35324907,0.09025786,0.0057219965,0.0003533668,0.003062777,0.0027621505,0.03889859],"genre_scores_gemma":[0.24983372,0.48199055,0.23556034,0.012484692,0.008686869,0.00043838107,0.0053689424,0.00034859194,0.0052878475],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99777156,0.0009529108,0.00025576827,0.00025953897,0.0006562712,0.00010398925],"domain_scores_gemma":[0.99436474,0.0037853208,0.00042250904,0.00030266005,0.00094862434,0.00017612251],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0030911362,0.0010659741,0.0011932884,0.0030497985,0.00037759473,0.0037745237,0.0011315278,0.0013041047,0.003087271],"category_scores_gemma":[0.012399715,0.00024545606,0.00120579,0.002625778,0.0011283444,0.0028612346,0.0018839828,0.0032687555,0.0012892209],"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":[0.00018060826,0.0002222149,0.008779321,0.0066092843,0.00073142437,0.0005206061,0.0004134519,0.02630685,0.0023640208,0.12534496,0.083581924,0.74494535],"study_design_scores_gemma":[0.00006305522,0.0002499673,0.0052253255,0.0051407223,0.00037929334,0.00051910063,0.00057343615,0.07788081,0.0027805443,0.60197735,0.30504525,0.00016518458],"about_ca_topic_score_codex":0.0013889583,"about_ca_topic_score_gemma":0.00117177,"teacher_disagreement_score":0.0037745237,"about_ca_system_score_codex":0.0009989796,"about_ca_system_score_gemma":0.002015533,"threshold_uncertainty_score":0.016347706},"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":"protocol","genre_gemma":"methods","domain_codex":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0038847812,0.027243545,0.3934519,0.034397867,0.046608612,0.4471135,0.033378176,0.0069805505,0.0069410955],"genre_scores_gemma":[0.009782373,0.0060517993,0.27143863,0.005101988,0.0030943942,0.6936944,0.007942254,0.0011062898,0.0017878406],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.12880334,0.5628995,0.27101347,0.008500237,0.02583235,0.0029511587],"domain_scores_gemma":[0.09138894,0.60060126,0.088208176,0.060329523,0.15640128,0.0030708665],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.653025,0.007331602,0.019305667,0.03350151,0.006386263,0.01588966,0.013229444,0.016438551,0.013525429],"category_scores_gemma":[0.8143518,0.0095828185,0.033717472,0.032856934,0.013834825,0.010243757,0.011090559,0.035641097,0.010384986],"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":[0.008056335,0.00055389944,0.0056918683,0.26013514,0.012260337,0.0008321485,0.010001944,0.005651111,0.0025476874,0.07306034,0.43115348,0.19005573],"study_design_scores_gemma":[0.017369753,0.0018392547,0.01769298,0.220277,0.009759726,0.002053581,0.0032721912,0.02053359,0.006309534,0.13587672,0.5629269,0.0020887211],"about_ca_topic_score_codex":0.0027048923,"about_ca_topic_score_gemma":0.0030008957,"teacher_disagreement_score":0.34697503,"about_ca_system_score_codex":0.020631878,"about_ca_system_score_gemma":0.03844201,"threshold_uncertainty_score":0.42788202},"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":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.19618478,0.38838378,0.15321226,0.0848396,0.010588233,0.0719879,0.05225934,0.0010253047,0.04151892],"genre_scores_gemma":[0.61085224,0.05150061,0.25381854,0.011236169,0.0009535645,0.056792602,0.0135431485,0.00036519312,0.00093793956],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.24258639,0.37692204,0.27764302,0.016769791,0.08312444,0.0029543499],"domain_scores_gemma":[0.09314739,0.66005474,0.103546254,0.03494867,0.1052825,0.003020471],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.5380876,0.00120767,0.00567779,0.03043096,0.0025852136,0.0102296835,0.0057752747,0.0037269865,0.0026550537],"category_scores_gemma":[0.84586763,0.002454634,0.007981014,0.025494415,0.004971018,0.0072069108,0.0077247964,0.0044854195,0.00044313964],"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":[0.0014385296,0.00031551803,0.20383759,0.321218,0.02599466,0.00033147025,0.03518729,0.002110751,0.0009070836,0.014632879,0.04174353,0.35228258],"study_design_scores_gemma":[0.0018294835,0.000709103,0.16218758,0.62457603,0.026199533,0.00075225276,0.018059395,0.0054034446,0.002599816,0.025109682,0.13175143,0.00082218915],"about_ca_topic_score_codex":0.01623086,"about_ca_topic_score_gemma":0.02321705,"teacher_disagreement_score":0.4619124,"about_ca_system_score_codex":0.013360138,"about_ca_system_score_gemma":0.041186236,"threshold_uncertainty_score":0.56962025},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.16319174,0.47847056,0.16574179,0.004938424,0.0038499977,0.15976146,0.015265187,0.0010489572,0.007731964],"genre_scores_gemma":[0.58702016,0.031826947,0.19202523,0.0016849882,0.0007164659,0.1812679,0.004361073,0.0001960327,0.00090126984],"study_design_codex":"meta_analysis","study_design_gemma":"observational","domain_scores_codex":[0.46192506,0.3179867,0.163873,0.017069139,0.037507832,0.0016382982],"domain_scores_gemma":[0.48171604,0.3681137,0.083272904,0.031627897,0.03401189,0.0012575747],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.35247537,0.0031234452,0.0127578955,0.013021502,0.0015144539,0.0048307236,0.0031414903,0.002667741,0.0033132145],"category_scores_gemma":[0.52469313,0.0022317285,0.037087385,0.011975008,0.0034589185,0.0041290005,0.004544377,0.0023969428,0.00038761023],"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":[0.011474966,0.00028004352,0.08554618,0.4037267,0.4152041,0.00037749723,0.0032283373,0.0025063504,0.0013738769,0.004901564,0.003167567,0.06821278],"study_design_scores_gemma":[0.020869916,0.0068509937,0.11445001,0.12947942,0.6773581,0.0009300949,0.0019740989,0.008726191,0.0036801912,0.012654768,0.022185892,0.0008403613],"about_ca_topic_score_codex":0.002121077,"about_ca_topic_score_gemma":0.005307216,"teacher_disagreement_score":0.9872421,"about_ca_system_score_codex":0.004549246,"about_ca_system_score_gemma":0.010852251,"threshold_uncertainty_score":0.7985132},"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":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.029596044,0.93554205,0.018703764,0.0020407569,0.0024682647,0.008221055,0.0013075905,0.00015556517,0.0019649111],"genre_scores_gemma":[0.66166055,0.24767186,0.059430774,0.0034618324,0.0016676177,0.023465475,0.0019177385,0.00012316281,0.00060105504],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.48129758,0.41401604,0.07632426,0.008871492,0.018222293,0.001268313],"domain_scores_gemma":[0.4258801,0.51471525,0.039146062,0.010481659,0.008878917,0.00089801475],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.21822202,0.004074464,0.02588474,0.016527867,0.00087676244,0.0084884735,0.0034595036,0.0057750256,0.0030829972],"category_scores_gemma":[0.4802741,0.0028509384,0.046313316,0.009247375,0.0032292702,0.0048768986,0.0040208986,0.004018625,0.00030911746],"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":[0.007951856,0.00013880145,0.004675395,0.16611464,0.7944062,0.00010030986,0.0002634486,0.0016033401,0.00019797547,0.0010873834,0.000406333,0.023054365],"study_design_scores_gemma":[0.013489528,0.002488114,0.006451553,0.04992105,0.9172814,0.0001401345,0.00018082547,0.0020985967,0.00052152807,0.0049681976,0.0023191394,0.00013999759],"about_ca_topic_score_codex":0.0025058386,"about_ca_topic_score_gemma":0.0027951088,"teacher_disagreement_score":0.781778,"about_ca_system_score_codex":0.0050295424,"about_ca_system_score_gemma":0.005046077,"threshold_uncertainty_score":0.96407145},"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":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0015328694,0.9163721,0.039908096,0.037158772,0.0023161185,0.00034269644,0.0005412685,0.00020253405,0.001625478],"genre_scores_gemma":[0.10747229,0.7111391,0.14529571,0.02084566,0.008863444,0.0040511093,0.0011237396,0.00056367397,0.0006452522],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.44991183,0.43850255,0.06310242,0.016077992,0.031084832,0.0013202693],"domain_scores_gemma":[0.13876425,0.8025459,0.016432872,0.01599849,0.025250679,0.0010077778],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.55948126,0.0039045543,0.018243052,0.021515502,0.002599632,0.014077432,0.0068207826,0.0033848672,0.0036025415],"category_scores_gemma":[0.72254866,0.0029178371,0.015313794,0.02222906,0.005305193,0.017292194,0.009183818,0.007296047,0.0009832395],"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":[0.0015522325,0.000086602005,0.014548394,0.22777954,0.11209648,0.00020619895,0.0020632553,0.0042544277,0.00029556415,0.052758574,0.022076258,0.56228256],"study_design_scores_gemma":[0.001275998,0.0006929721,0.013146586,0.26536456,0.15264452,0.00079394126,0.0025955904,0.024608145,0.00122768,0.34209102,0.19491768,0.00064135424],"about_ca_topic_score_codex":0.0090999575,"about_ca_topic_score_gemma":0.017365059,"teacher_disagreement_score":0.44051874,"about_ca_system_score_codex":0.010747992,"about_ca_system_score_gemma":0.025062196,"threshold_uncertainty_score":0.54323804},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0003404499,0.99794203,0.00018681347,0.00031602633,0.00022929788,0.00042378422,0.00012311118,0.0000055933647,0.00043281948],"genre_scores_gemma":[0.002225537,0.9962289,0.00044642718,0.0002541825,0.000105761654,0.0004907268,0.00012250854,0.000003177057,0.00012278082],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.985559,0.0048810435,0.0051324116,0.00081576256,0.0031758435,0.00043588734],"domain_scores_gemma":[0.9602315,0.02862506,0.005901318,0.0006719904,0.0042659347,0.00030426003],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.015147087,0.0022063367,0.00659489,0.018446278,0.0014846884,0.005158299,0.0021737807,0.0040684585,0.0057009426],"category_scores_gemma":[0.049583506,0.0015141757,0.006675866,0.017905686,0.0014549417,0.003890936,0.0027210047,0.0019077656,0.0009495129],"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":[0.00011460848,0.000035355115,0.000400067,0.86984414,0.0015232062,0.00022782428,0.00038828922,0.00015988004,0.00028167735,0.0007203012,0.004357891,0.121946745],"study_design_scores_gemma":[0.000030938987,0.00009646325,0.0010788031,0.95057,0.0061398577,0.00044286723,0.00035998502,0.0000690926,0.00017696225,0.00044414605,0.04056696,0.000023861214],"about_ca_topic_score_codex":0.00619088,"about_ca_topic_score_gemma":0.012437629,"teacher_disagreement_score":0.018446278,"about_ca_system_score_codex":0.0046329317,"about_ca_system_score_gemma":0.01932335,"threshold_uncertainty_score":0.08010644},"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":"commentary","genre_gemma":"methods","domain_codex":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015472016,0.014304987,0.38900483,0.39188474,0.03689528,0.06619518,0.005821775,0.0034900347,0.07693121],"genre_scores_gemma":[0.049607925,0.0057686768,0.83173597,0.052690577,0.0031363918,0.045106553,0.0035917966,0.0005866138,0.007775414],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.34265476,0.41998705,0.16355477,0.0059255823,0.06161302,0.0062647294],"domain_scores_gemma":[0.14211923,0.46671346,0.06786048,0.053103447,0.25944707,0.010756282],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.51357454,0.001981452,0.0030242756,0.011251641,0.004982823,0.010423615,0.0050550457,0.013618882,0.005777271],"category_scores_gemma":[0.65451545,0.003269722,0.006064419,0.0054239766,0.009484428,0.016152175,0.018310845,0.019437948,0.0043611396],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008460834,0.000675278,0.0078034364,0.041335408,0.00040062945,0.0012797201,0.033268284,0.0017515755,0.00452995,0.10932029,0.39504442,0.4037449],"study_design_scores_gemma":[0.0007552167,0.0005415834,0.0069108014,0.05406254,0.00044380457,0.0020924525,0.0077651963,0.0022130462,0.0029171144,0.03643822,0.8851721,0.0006880013],"about_ca_topic_score_codex":0.004337301,"about_ca_topic_score_gemma":0.006444612,"teacher_disagreement_score":0.48642546,"about_ca_system_score_codex":0.009171888,"about_ca_system_score_gemma":0.08914979,"threshold_uncertainty_score":0.5998492},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.39388093,0.5864676,0.00041764256,0.006643466,0.0015376173,0.00028176303,0.0010726025,0.000075227246,0.009623098],"genre_scores_gemma":[0.7994422,0.19204256,0.0019402223,0.0025177558,0.0011298044,0.000114381626,0.0005224324,0.000005976724,0.002284738],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.999554,0.00011877491,0.000082838895,0.000039804498,0.00014951311,0.00005504825],"domain_scores_gemma":[0.9991647,0.00024132337,0.00032352438,0.000021288439,0.00012002741,0.00012915755],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00080149487,0.00038013177,0.000966408,0.0007926324,0.00029303905,0.00051358016,0.00043318304,0.0011183146,0.003538599],"category_scores_gemma":[0.0035187756,0.0002257792,0.0012991435,0.0004578747,0.0001727128,0.00032559846,0.0003420134,0.0011080698,0.00029165187],"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":[0.049699526,0.0059686145,0.12820673,0.016673334,0.008169363,0.00061834004,0.0003664017,0.00034985476,0.0049152067,0.00074316066,0.011608975,0.7726806],"study_design_scores_gemma":[0.036947418,0.055746496,0.79353046,0.023943605,0.034169193,0.0025887345,0.0013762347,0.0008820258,0.0048378618,0.0022802392,0.04356969,0.00012795533],"about_ca_topic_score_codex":0.0044398187,"about_ca_topic_score_gemma":0.012515703,"teacher_disagreement_score":0.0044398187,"about_ca_system_score_codex":0.0002993374,"about_ca_system_score_gemma":0.00094174885,"threshold_uncertainty_score":0.01183784},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0024269114,0.9962549,0.000053692067,0.00020754583,0.00011775942,0.0001265155,0.00018463048,0.0000042389083,0.0006237849],"genre_scores_gemma":[0.058804233,0.93803346,0.0006801671,0.0011806199,0.00020458852,0.0003548148,0.00023065461,0.0000073037063,0.0005042339],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9981001,0.00071087835,0.0007172885,0.00015595282,0.0002592941,0.00005653119],"domain_scores_gemma":[0.99719363,0.002011355,0.0005981928,0.000026410977,0.00011540958,0.000055028882],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0017588255,0.00055470166,0.0033951083,0.002194592,0.00028325967,0.0010620615,0.0005599577,0.001349681,0.006257831],"category_scores_gemma":[0.0069449986,0.000298921,0.0041005677,0.0021393062,0.00040643956,0.0012232314,0.00049032056,0.00091083767,0.00024346488],"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":[0.0041698343,0.000095504955,0.0009079224,0.8142986,0.011766416,0.00013900023,0.000180258,0.00016222592,0.00039126677,0.00086394633,0.002838462,0.16418651],"study_design_scores_gemma":[0.008283003,0.00498903,0.01698095,0.76565284,0.117336705,0.0017170092,0.000748645,0.0002723202,0.0008339382,0.0014681785,0.08164227,0.00007510885],"about_ca_topic_score_codex":0.0020022297,"about_ca_topic_score_gemma":0.0053709885,"teacher_disagreement_score":0.006257831,"about_ca_system_score_codex":0.00093094737,"about_ca_system_score_gemma":0.0010318235,"threshold_uncertainty_score":0.020934522},"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":"commentary","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0062749884,0.9391015,0.018455481,0.025884368,0.0015701938,0.0011321714,0.0036787875,0.00009970508,0.0038028308],"genre_scores_gemma":[0.29024234,0.56711304,0.0925007,0.032768056,0.005256535,0.0056031374,0.00523772,0.0003996957,0.0008788355],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.78308,0.16846325,0.027241718,0.0050638383,0.014853266,0.0012979389],"domain_scores_gemma":[0.22743665,0.685535,0.044131074,0.013360228,0.028522307,0.0010147586],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.21286198,0.0023581614,0.011455339,0.009850861,0.00061807124,0.011017602,0.004686438,0.0031645044,0.013339581],"category_scores_gemma":[0.53261036,0.0015199012,0.026395382,0.012289946,0.0019268154,0.0115373805,0.0033061143,0.0063560037,0.0009867455],"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":[0.0025021979,0.00032543996,0.014864556,0.3114589,0.08403574,0.00011846375,0.00061364745,0.012557276,0.00039468304,0.018224917,0.021192558,0.5337117],"study_design_scores_gemma":[0.002277358,0.002395144,0.050821707,0.6453115,0.071444176,0.0007296568,0.0018065501,0.01128462,0.00097087515,0.06921285,0.14314067,0.00060489925],"about_ca_topic_score_codex":0.012674314,"about_ca_topic_score_gemma":0.01458926,"teacher_disagreement_score":0.787138,"about_ca_system_score_codex":0.01130799,"about_ca_system_score_gemma":0.014764599,"threshold_uncertainty_score":0.9706813},"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":"methods","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0031961293,0.0816643,0.030635769,0.7862438,0.025004938,0.008778312,0.007978019,0.0014315025,0.055067245],"genre_scores_gemma":[0.030101081,0.14724536,0.33054578,0.403562,0.018594297,0.018122444,0.021245545,0.00054263935,0.030040875],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.98697114,0.0049394574,0.0034954587,0.00058328337,0.003491231,0.000519449],"domain_scores_gemma":[0.94707215,0.011959896,0.0037335933,0.0010804788,0.033218402,0.0029355239],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.013314325,0.0011847013,0.0023941796,0.0039478755,0.00140628,0.0030638946,0.0045185373,0.01190072,0.007247066],"category_scores_gemma":[0.061552346,0.0008490256,0.004391371,0.0027993785,0.0014151172,0.0033948228,0.0027661156,0.012060345,0.008657612],"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":[0.000079754915,0.00016952526,0.0020613512,0.0062723374,0.00018095624,0.0005374421,0.00042747275,0.00057791517,0.0005297248,0.003859351,0.79330885,0.19199543],"study_design_scores_gemma":[0.0007109212,0.00020454974,0.008068902,0.069189705,0.0008745847,0.0018681487,0.0005744132,0.001977397,0.00054173596,0.014114813,0.9016715,0.0002032562],"about_ca_topic_score_codex":0.01559466,"about_ca_topic_score_gemma":0.023904562,"teacher_disagreement_score":0.01559466,"about_ca_system_score_codex":0.004726025,"about_ca_system_score_gemma":0.017846648,"threshold_uncertainty_score":0.07041371},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0121536665,0.899161,0.010500212,0.019574514,0.0021422892,0.039161976,0.013844163,0.00035707882,0.0031051356],"genre_scores_gemma":[0.13213088,0.61393225,0.12165982,0.01605851,0.0016712846,0.10041137,0.012450339,0.00023391063,0.0014516398],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.89272535,0.029500686,0.059273504,0.0052744253,0.01198425,0.0012417083],"domain_scores_gemma":[0.66490436,0.20358284,0.07545489,0.00929095,0.043723512,0.0030434786],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.122311674,0.0023238566,0.012542457,0.021513194,0.0018657465,0.0035991163,0.0050890055,0.0040404005,0.0032919482],"category_scores_gemma":[0.26799616,0.0026924915,0.013418562,0.013582349,0.0020574557,0.005679936,0.004539586,0.0026503748,0.000572213],"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":[0.0003956536,0.000034575634,0.014892761,0.8649564,0.018402902,0.00023568522,0.0010590602,0.00035160407,0.00048401172,0.0007231095,0.013083805,0.08538048],"study_design_scores_gemma":[0.0009910807,0.00021972376,0.020907395,0.85116696,0.08226376,0.0005628362,0.00076893024,0.0006997302,0.0005592924,0.0014745337,0.040213455,0.0001722603],"about_ca_topic_score_codex":0.009567177,"about_ca_topic_score_gemma":0.018116876,"teacher_disagreement_score":0.122311674,"about_ca_system_score_codex":0.006863772,"about_ca_system_score_gemma":0.039853517,"threshold_uncertainty_score":0.6468537},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0017351091,0.994031,0.00048056984,0.00056990486,0.00014551873,0.0010854179,0.0015596374,0.00003294218,0.00035985708],"genre_scores_gemma":[0.024792517,0.96628344,0.0029177573,0.00097052177,0.000111277404,0.0035789823,0.0011442355,0.00002847483,0.0001727795],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.96313596,0.01078774,0.01924582,0.0019369227,0.0041428385,0.00075074215],"domain_scores_gemma":[0.904565,0.06856245,0.017726729,0.0016697984,0.0068137874,0.00066226034],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.028363481,0.0030947884,0.0120685585,0.031625587,0.0012479561,0.0042780214,0.0032955268,0.0025149407,0.0051579783],"category_scores_gemma":[0.104604565,0.0022989758,0.014463647,0.028613457,0.0015279759,0.003719105,0.0037581243,0.0018338392,0.00052055845],"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":[0.0001769263,0.000010146141,0.00093950785,0.96999884,0.0116418945,0.000101452955,0.00022793545,0.000086134445,0.00014727331,0.00019428994,0.0011020573,0.015373588],"study_design_scores_gemma":[0.00020551588,0.00012891027,0.0041445526,0.8946982,0.08868588,0.0003052675,0.00034834078,0.00008508584,0.0001703632,0.00029943278,0.010889133,0.000039308095],"about_ca_topic_score_codex":0.010219863,"about_ca_topic_score_gemma":0.029340103,"teacher_disagreement_score":0.031625587,"about_ca_system_score_codex":0.008015777,"about_ca_system_score_gemma":0.016576977,"threshold_uncertainty_score":0.15000224},"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":"review","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.046502728,0.42441884,0.25262126,0.16554509,0.032038584,0.04356848,0.018910762,0.0020372616,0.014356882],"genre_scores_gemma":[0.39737418,0.11585994,0.3258863,0.044520114,0.008641188,0.0919067,0.011729582,0.0016174988,0.0024644572],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.120627515,0.47958997,0.3192907,0.014078358,0.06464468,0.0017688521],"domain_scores_gemma":[0.022508925,0.67957544,0.15564942,0.04378177,0.09621815,0.0022662396],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.79620206,0.0021386202,0.00819412,0.015061016,0.0031538247,0.016023064,0.007924255,0.005790908,0.0037681658],"category_scores_gemma":[0.9131887,0.002950985,0.009830436,0.01591414,0.00951567,0.016894057,0.0066290507,0.007186222,0.0015163616],"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":[0.0050024805,0.00029858074,0.067362696,0.37487823,0.016963687,0.00032549247,0.01223173,0.0027834063,0.0022696885,0.02048147,0.05926602,0.43813655],"study_design_scores_gemma":[0.004599124,0.0019404333,0.06872976,0.57266736,0.022167768,0.001408177,0.0068898727,0.009370343,0.005723239,0.057423044,0.2477381,0.0013428228],"about_ca_topic_score_codex":0.0048395707,"about_ca_topic_score_gemma":0.0036871573,"teacher_disagreement_score":0.20379794,"about_ca_system_score_codex":0.014096446,"about_ca_system_score_gemma":0.03986675,"threshold_uncertainty_score":0.25131917},"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":"commentary","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00019587496,0.003528747,0.00019684873,0.98063445,0.01447608,0.000012201834,0.0000068778877,0.000016368127,0.00093254866],"genre_scores_gemma":[0.0049184673,0.0029551147,0.0005930032,0.9729126,0.015987603,0.00006056023,0.000008626067,0.00004929126,0.002514653],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.94268894,0.04113301,0.0021699714,0.0033978086,0.007320497,0.0032897643],"domain_scores_gemma":[0.8722584,0.079335906,0.0043054377,0.0018876622,0.009624415,0.032588154],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.06109126,0.0016282821,0.0028561673,0.0021889023,0.01733122,0.0154744685,0.0043463875,0.038103525,0.012363319],"category_scores_gemma":[0.10629475,0.0015844966,0.0025279557,0.0015579439,0.019722298,0.034002054,0.012867701,0.112641804,0.003425283],"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":[0.00007468653,0.00011316999,0.0001770144,0.00029027188,0.000027562483,0.0006523601,0.016830768,0.00005042928,0.00026220854,0.013173909,0.95620984,0.012137867],"study_design_scores_gemma":[0.00006229637,0.00015958582,0.00028464475,0.0016046709,0.000017926099,0.0013646318,0.02278508,0.000238335,0.00014278071,0.014208166,0.9590079,0.00012395282],"about_ca_topic_score_codex":0.0042237425,"about_ca_topic_score_gemma":0.0059646606,"teacher_disagreement_score":0.93890876,"about_ca_system_score_codex":0.0125821335,"about_ca_system_score_gemma":0.01774189,"threshold_uncertainty_score":0.3230853},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.007395657,0.98993486,0.00067083753,0.0002752453,0.00014903,0.0006378872,0.0006083013,0.000030986892,0.0002972163],"genre_scores_gemma":[0.21143892,0.7794457,0.0038098926,0.0010323254,0.00026643364,0.0026907572,0.0009822291,0.000029474697,0.00030431984],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.993204,0.0035275437,0.0015926019,0.0006773181,0.00079802546,0.00020052902],"domain_scores_gemma":[0.98845226,0.0084482385,0.0019063809,0.00029265368,0.0007406981,0.00015972665],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.010817004,0.0022076713,0.0143386945,0.0052777077,0.00062324957,0.0019873288,0.0018413083,0.0014968159,0.003250175],"category_scores_gemma":[0.024185508,0.001011524,0.026634293,0.005525362,0.00053201296,0.0012934018,0.0012340747,0.0016026542,0.00019324044],"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":[0.0024002087,0.00006615132,0.0033522402,0.39893726,0.5758893,0.00007742344,0.00008741113,0.0005231982,0.00015016815,0.00013026941,0.00062483124,0.017761566],"study_design_scores_gemma":[0.0014796777,0.0002816892,0.0037296158,0.030672474,0.9619128,0.000069679816,0.000036191705,0.00025366963,0.000074160555,0.0001629454,0.001309119,0.000017896078],"about_ca_topic_score_codex":0.008731696,"about_ca_topic_score_gemma":0.018693335,"teacher_disagreement_score":0.0143386945,"about_ca_system_score_codex":0.0027906601,"about_ca_system_score_gemma":0.0037337914,"threshold_uncertainty_score":0.057206452},"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":"commentary","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0067101284,0.37840575,0.05971322,0.45933494,0.019183733,0.0103596235,0.00824366,0.0013518368,0.056697156],"genre_scores_gemma":[0.06374853,0.37845996,0.33659664,0.16232096,0.012018164,0.013467386,0.013616309,0.00042632976,0.019345785],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.98435336,0.005509118,0.0061913724,0.0008098262,0.0027372611,0.00039900083],"domain_scores_gemma":[0.969376,0.009363374,0.0043771765,0.0007940509,0.014860942,0.0012284217],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.014833781,0.0018256205,0.0039380267,0.0037745938,0.001242567,0.0032861847,0.005793019,0.008738657,0.0028479346],"category_scores_gemma":[0.042719565,0.0012946805,0.005169244,0.0033385172,0.0012034866,0.002444426,0.0017184329,0.008916043,0.003111677],"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":[0.00032838716,0.00030586668,0.0028234238,0.029676938,0.0023163357,0.0009505279,0.00058583956,0.0029983604,0.0014132845,0.0066246325,0.6070175,0.34495887],"study_design_scores_gemma":[0.0025200795,0.0005105449,0.0113075655,0.13726597,0.007480868,0.001834204,0.00044564356,0.0061476445,0.0013445956,0.021115497,0.80974513,0.00028229124],"about_ca_topic_score_codex":0.016871374,"about_ca_topic_score_gemma":0.02308671,"teacher_disagreement_score":0.016871374,"about_ca_system_score_codex":0.0041733803,"about_ca_system_score_gemma":0.012149344,"threshold_uncertainty_score":0.07844943},"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":"methods","genre_gemma":"methods","domain_codex":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00486777,0.009213272,0.5163251,0.052428745,0.026609369,0.3642601,0.013201497,0.004457482,0.008636713],"genre_scores_gemma":[0.01327543,0.0028313005,0.47618794,0.008722641,0.0022625101,0.49098894,0.0041754236,0.0005286935,0.001027132],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.08867853,0.6165675,0.2626482,0.0068031107,0.023189325,0.002113345],"domain_scores_gemma":[0.066947795,0.6356392,0.0955529,0.07320289,0.12372501,0.004932222],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.7794934,0.004890652,0.012060112,0.022966,0.0066735293,0.013242485,0.01137781,0.019503156,0.006112579],"category_scores_gemma":[0.82851744,0.00743671,0.021601694,0.021872574,0.013325644,0.010267967,0.013632879,0.029406786,0.0052125417],"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":[0.007007771,0.0009620847,0.014447655,0.16413517,0.010086983,0.001027675,0.023081535,0.0070860214,0.0041288678,0.11558112,0.34835336,0.30410182],"study_design_scores_gemma":[0.015490001,0.0023225313,0.027908865,0.18354216,0.00759928,0.002056514,0.004962572,0.0328252,0.008411086,0.14775805,0.5651293,0.001994413],"about_ca_topic_score_codex":0.002467848,"about_ca_topic_score_gemma":0.0028520839,"teacher_disagreement_score":0.22050661,"about_ca_system_score_codex":0.01604403,"about_ca_system_score_gemma":0.056900676,"threshold_uncertainty_score":0.27192396},"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":"review","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0015898756,0.5437261,0.017235732,0.34887108,0.027208658,0.015002176,0.008015467,0.00068875495,0.037662223],"genre_scores_gemma":[0.011729942,0.64712924,0.1429369,0.140117,0.007413827,0.018954027,0.011483274,0.00017777368,0.02005805],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9884531,0.0039091236,0.0037421428,0.0005932092,0.0029043923,0.00039810306],"domain_scores_gemma":[0.9749255,0.0077100187,0.0027826182,0.00055771007,0.012817451,0.0012067107],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.011420671,0.002025366,0.00443338,0.0054234597,0.0015872326,0.003255583,0.0071476195,0.01349561,0.009720592],"category_scores_gemma":[0.035574306,0.0014682519,0.006798121,0.0054842294,0.001524001,0.0037918575,0.0024393653,0.00931633,0.008053184],"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":[0.0001875358,0.00022227387,0.00063060236,0.090091534,0.00073683896,0.0006908759,0.0006375387,0.0007306974,0.00088438217,0.0045983386,0.64201814,0.2585713],"study_design_scores_gemma":[0.001176249,0.00018365901,0.003109424,0.2929554,0.0023590527,0.0007171769,0.00031928698,0.00094542303,0.0003650942,0.010229834,0.6874838,0.00015557889],"about_ca_topic_score_codex":0.025513865,"about_ca_topic_score_gemma":0.043271717,"teacher_disagreement_score":0.025513865,"about_ca_system_score_codex":0.006074035,"about_ca_system_score_gemma":0.023374412,"threshold_uncertainty_score":0.060398996},"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":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06690415,0.7764965,0.105290756,0.006380104,0.0021423902,0.027450433,0.0104771,0.00044246993,0.004416052],"genre_scores_gemma":[0.6902185,0.106315345,0.101697475,0.0072316546,0.0013134882,0.08649743,0.0055467957,0.00041095293,0.0007683346],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.31581056,0.30034894,0.30655462,0.020861631,0.054836046,0.0015882236],"domain_scores_gemma":[0.12825787,0.62538433,0.177577,0.028551385,0.039100602,0.0011287753],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.4915414,0.002375418,0.010538365,0.0145284645,0.0020526748,0.008582544,0.004090627,0.004776394,0.0044283736],"category_scores_gemma":[0.7407863,0.0029808925,0.022411589,0.0160558,0.0042593754,0.00918075,0.0049046236,0.0034997715,0.0005401907],"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":[0.005735019,0.00026745367,0.22848149,0.4400532,0.20797722,0.0005255956,0.005571327,0.0017877101,0.0007852782,0.0061305733,0.0070033977,0.095681764],"study_design_scores_gemma":[0.005962414,0.0035715508,0.18658437,0.34809032,0.36953837,0.0050523654,0.0038356541,0.0075603495,0.0044631916,0.02175317,0.04238454,0.0012037781],"about_ca_topic_score_codex":0.0030011453,"about_ca_topic_score_gemma":0.0032127483,"teacher_disagreement_score":0.98946166,"about_ca_system_score_codex":0.007100757,"about_ca_system_score_gemma":0.009864441,"threshold_uncertainty_score":0.62702},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00033338205,0.9966719,0.000728612,0.0010294875,0.00022031425,0.00019358889,0.000042370502,0.0000048285397,0.0007755365],"genre_scores_gemma":[0.0056659593,0.99111265,0.0020805628,0.0004443845,0.00011796507,0.00045114776,0.000055302302,0.0000035550283,0.000068571935],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9610356,0.021092473,0.009128591,0.0013863445,0.0067023872,0.00065461],"domain_scores_gemma":[0.82970124,0.144836,0.0100199,0.0018975363,0.013060734,0.00048459216],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.04659396,0.001997125,0.004445091,0.026415229,0.0013652868,0.0071682604,0.0025278653,0.0037340345,0.0027241977],"category_scores_gemma":[0.13692597,0.0016480631,0.004900797,0.027355013,0.0025594938,0.007988709,0.003616508,0.003081576,0.00045195466],"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":[0.000089940244,0.000068645284,0.0011044155,0.5849902,0.0018661495,0.00012594917,0.0017034203,0.00060529384,0.00019375749,0.009446788,0.0049924604,0.39481303],"study_design_scores_gemma":[0.000026191447,0.000063239815,0.001245007,0.95497125,0.0026644617,0.00014680845,0.00094678195,0.00020961961,0.00016803843,0.0024132398,0.03711146,0.000033909706],"about_ca_topic_score_codex":0.008942995,"about_ca_topic_score_gemma":0.014446943,"teacher_disagreement_score":0.04659396,"about_ca_system_score_codex":0.008567026,"about_ca_system_score_gemma":0.026872007,"threshold_uncertainty_score":0.24641538},"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":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99770194,0.0006997685,0.0003669078,0.00006045645,0.00000776107,0.00007305428,0.00041219342,0.0000040314117,0.00067392184],"genre_scores_gemma":[0.99790937,0.00039356702,0.0011061665,0.000016589554,0.000011485573,0.00007411419,0.00035370787,0.0000018835619,0.00013298624],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.999458,0.00022765968,0.00007103619,0.00007922,0.00007205769,0.00009203889],"domain_scores_gemma":[0.99925333,0.00021262883,0.0003179738,0.000030098137,0.00009000602,0.00009593763],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010620139,0.0005478149,0.00039911928,0.0028505898,0.00047930982,0.00064290053,0.00040134115,0.0003349341,0.0010760864],"category_scores_gemma":[0.0033929986,0.00032530678,0.00046210972,0.0017204981,0.00029102358,0.0006958363,0.00072299654,0.0003617283,0.000150881],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000051373485,0.000031273496,0.99600285,0.000034557692,0.000046719917,0.0001575787,0.00051788834,0.00016414479,0.00018129087,0.000047922906,0.00006989479,0.0026944866],"study_design_scores_gemma":[0.0000040698155,0.00007627009,0.995826,0.000036252804,0.000023809389,0.0002758188,0.002231674,0.0009676086,0.00008121696,0.00008361625,0.00038590748,0.000007787661],"about_ca_topic_score_codex":0.014724187,"about_ca_topic_score_gemma":0.014401978,"teacher_disagreement_score":0.014724187,"about_ca_system_score_codex":0.0006859067,"about_ca_system_score_gemma":0.0005168755,"threshold_uncertainty_score":0.029276967},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0011581616,0.99736965,0.00009636829,0.00040728212,0.000085330685,0.00041129038,0.00015404604,0.000005868684,0.00031197222],"genre_scores_gemma":[0.015982552,0.9817092,0.00064306444,0.00059709337,0.00008638734,0.00078216055,0.000107882996,0.000003053781,0.00008866499],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9908406,0.0042119576,0.0026543792,0.0004684947,0.00145566,0.00036895042],"domain_scores_gemma":[0.9642606,0.028332578,0.004351326,0.0003418388,0.002283482,0.00043029885],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.010398044,0.0012895598,0.005978978,0.009668462,0.00092782016,0.0041543893,0.0017916749,0.0026229562,0.005541523],"category_scores_gemma":[0.040734403,0.0010607571,0.0061489344,0.01158423,0.0009567469,0.0028774862,0.0016950655,0.0014485788,0.0003234815],"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":[0.00013334714,0.00003319988,0.000676692,0.9597421,0.0037337034,0.00006649479,0.0002606494,0.00007555397,0.000055775603,0.00023460874,0.0009099697,0.03407788],"study_design_scores_gemma":[0.00019605055,0.0002470757,0.0038047766,0.9541539,0.02800266,0.00020229854,0.0004993851,0.000114416034,0.00010353576,0.00020589387,0.012441841,0.000028251625],"about_ca_topic_score_codex":0.010711439,"about_ca_topic_score_gemma":0.027516827,"teacher_disagreement_score":0.010711439,"about_ca_system_score_codex":0.006139026,"about_ca_system_score_gemma":0.021503592,"threshold_uncertainty_score":0.05499077},"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":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00069705927,0.4282897,0.04809996,0.49244943,0.0269906,0.00043818465,0.00028748458,0.00032198144,0.0024256606],"genre_scores_gemma":[0.07799948,0.3621618,0.15658839,0.2996303,0.09579958,0.0046398425,0.00065201573,0.00097351044,0.001555115],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.22460325,0.48261276,0.19621243,0.01737795,0.0765167,0.0026768947],"domain_scores_gemma":[0.08362361,0.7502654,0.06230076,0.028431842,0.072278835,0.0030996238],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.5740926,0.0035498857,0.015561818,0.027178032,0.0060636834,0.033747032,0.01516164,0.029505422,0.0035469197],"category_scores_gemma":[0.77371395,0.003555238,0.0071394662,0.031815164,0.055328242,0.035902757,0.01769389,0.040957615,0.0030920603],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00046948149,0.00010205787,0.0034239662,0.09020444,0.0030466805,0.0005129133,0.019186074,0.00076165126,0.000647863,0.30073267,0.14811607,0.4327961],"study_design_scores_gemma":[0.00043541958,0.00019175709,0.0027623714,0.17768073,0.0015186801,0.0025342116,0.005203588,0.0028053198,0.0006277264,0.56878024,0.23668818,0.00077175815],"about_ca_topic_score_codex":0.006270232,"about_ca_topic_score_gemma":0.0038600434,"teacher_disagreement_score":0.42590737,"about_ca_system_score_codex":0.022217922,"about_ca_system_score_gemma":0.031789847,"threshold_uncertainty_score":0.5252197},"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":"review","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.043052416,0.86641395,0.042143617,0.008569231,0.0024880255,0.0073265596,0.025683438,0.00020977744,0.0041129715],"genre_scores_gemma":[0.5763487,0.2771503,0.08445607,0.012616952,0.0016822875,0.029982325,0.016608825,0.0002931426,0.00086135906],"study_design_codex":"systematic_review","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.4059044,0.3280374,0.21469456,0.015135176,0.034487303,0.001741202],"domain_scores_gemma":[0.16649874,0.63291824,0.14627063,0.03273306,0.020900974,0.00067837187],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.2865461,0.0019062598,0.005448453,0.012177957,0.0016794205,0.00560019,0.004313209,0.0032057606,0.002577228],"category_scores_gemma":[0.66567695,0.0023703272,0.00741293,0.018788746,0.0036841896,0.0060474216,0.003673289,0.002273928,0.0004711571],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018483782,0.000102869104,0.100245714,0.74553084,0.045094892,0.00050290243,0.006424033,0.0013769668,0.001015027,0.0061135585,0.008466169,0.083278716],"study_design_scores_gemma":[0.0013358698,0.00077962776,0.10010474,0.7212529,0.073147856,0.0013477815,0.004800073,0.0023334594,0.0037536477,0.016294675,0.07420982,0.00063953624],"about_ca_topic_score_codex":0.011758771,"about_ca_topic_score_gemma":0.011242361,"teacher_disagreement_score":0.7134539,"about_ca_system_score_codex":0.004823354,"about_ca_system_score_gemma":0.015527216,"threshold_uncertainty_score":0.8798157},"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":"other","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.23595001,0.046780106,0.14481483,0.14356165,0.0047641266,0.10386728,0.03767944,0.0026347295,0.2799478],"genre_scores_gemma":[0.4436017,0.008319361,0.50705093,0.006335812,0.0002125259,0.017164996,0.010633245,0.00019795002,0.0064833774],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9071049,0.04114401,0.013821141,0.0033816393,0.030528909,0.00401943],"domain_scores_gemma":[0.8413772,0.039307043,0.0069469404,0.0024492068,0.10053906,0.009380503],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.10456922,0.0010373199,0.001558852,0.0094376,0.007993872,0.0069540488,0.005392539,0.0020161604,0.0061870473],"category_scores_gemma":[0.15599544,0.0009547403,0.0031524294,0.008319637,0.0022034864,0.0018318745,0.0060354644,0.002695393,0.0008309354],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009851545,0.0007752148,0.116719164,0.01696604,0.0012330831,0.0024133308,0.026204553,0.008853492,0.0015435022,0.013173483,0.22420035,0.58693266],"study_design_scores_gemma":[0.0027679056,0.0011498044,0.27080742,0.06345871,0.0027835902,0.0020748605,0.0636645,0.03810914,0.0031861425,0.023629181,0.5271228,0.0012460375],"about_ca_topic_score_codex":0.86296576,"about_ca_topic_score_gemma":0.9232124,"teacher_disagreement_score":0.13703424,"about_ca_system_score_codex":0.10375707,"about_ca_system_score_gemma":0.40293065,"threshold_uncertainty_score":0.75281376},"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":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0026893467,0.4821428,0.10469786,0.32699344,0.029451786,0.042049717,0.0039310157,0.0012427593,0.006801251],"genre_scores_gemma":[0.015864378,0.36284536,0.49387354,0.052948043,0.0035985457,0.06478725,0.0032150815,0.0002882802,0.0025795747],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.60947543,0.22507839,0.12131178,0.010054047,0.030534467,0.0035459178],"domain_scores_gemma":[0.25608787,0.517913,0.05654802,0.021545364,0.14237739,0.0055283196],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.42446387,0.0052924654,0.009794344,0.03476973,0.005772929,0.015276592,0.0137451915,0.017611772,0.009018549],"category_scores_gemma":[0.6793048,0.005470053,0.017258799,0.024681022,0.0073821703,0.03165105,0.011461171,0.015448852,0.0041246596],"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":[0.00049643015,0.00022842147,0.0023953954,0.48936394,0.0038983892,0.0006411734,0.00854639,0.0017375791,0.00090825005,0.017274784,0.11815655,0.35635278],"study_design_scores_gemma":[0.00040483117,0.00010981555,0.0012062944,0.8716496,0.0055948277,0.00022055623,0.0031106307,0.0014153321,0.0005430106,0.019734666,0.09578035,0.00022999445],"about_ca_topic_score_codex":0.020243585,"about_ca_topic_score_gemma":0.03498764,"teacher_disagreement_score":0.57553613,"about_ca_system_score_codex":0.020795336,"about_ca_system_score_gemma":0.118887514,"threshold_uncertainty_score":0.7097385},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00196175,0.9969181,0.00009853758,0.00019362044,0.00006199107,0.00029020806,0.00027673633,0.000005966472,0.0001931658],"genre_scores_gemma":[0.032844618,0.9650373,0.000657474,0.00044337974,0.00008542322,0.0006084384,0.0002232649,0.0000037890984,0.00009630852],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9927712,0.0021872083,0.0027220026,0.00054651557,0.0015814384,0.00019162222],"domain_scores_gemma":[0.96599126,0.022975016,0.008259636,0.00035615114,0.0020860583,0.00033197572],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.006925619,0.0012275635,0.0068681547,0.008062121,0.0005899421,0.0022353546,0.0014159139,0.0019389716,0.0028097765],"category_scores_gemma":[0.04019035,0.0011338708,0.007270548,0.008678705,0.00073580985,0.0018337553,0.0013324262,0.0010992326,0.00017111473],"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":[0.00017610376,0.000014966077,0.0019191424,0.96430403,0.010728838,0.0000927983,0.00016707438,0.0000755976,0.00006167932,0.00010380386,0.00077020144,0.02158586],"study_design_scores_gemma":[0.00029156753,0.00023998972,0.012131176,0.8972468,0.0804035,0.0005594691,0.0002513287,0.00014070603,0.000108130174,0.00027139575,0.008314216,0.00004171701],"about_ca_topic_score_codex":0.013323881,"about_ca_topic_score_gemma":0.031930123,"teacher_disagreement_score":0.013323881,"about_ca_system_score_codex":0.003595495,"about_ca_system_score_gemma":0.009873089,"threshold_uncertainty_score":0.036626577},"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":"methods","genre_gemma":"editorial","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"editorial","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0003372232,0.008207793,0.87023294,0.056099236,0.028838722,0.010970992,0.0026621947,0.010963164,0.011687708],"genre_scores_gemma":[0.002095672,0.0047454606,0.9627502,0.0065022046,0.009362094,0.007492158,0.0004037118,0.0016239731,0.0050245556],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.83714074,0.11671359,0.021536596,0.0038677272,0.02002957,0.0007118182],"domain_scores_gemma":[0.4531566,0.4442467,0.013230494,0.014140629,0.07004215,0.005183585],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.14705855,0.00480204,0.0036520842,0.018467631,0.0020235444,0.011652472,0.0069068857,0.0055807563,0.050072942],"category_scores_gemma":[0.40014294,0.0044457363,0.005174136,0.006626806,0.006673279,0.0071569285,0.0056757717,0.015386327,0.02548444],"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":[0.00031642616,0.00009044336,0.00021893777,0.008519151,0.00038179822,0.00045903385,0.0012272798,0.0011779334,0.0009507843,0.032058176,0.7195992,0.23500086],"study_design_scores_gemma":[0.0010615555,0.00028654168,0.0006839272,0.013034108,0.0004280001,0.0017895283,0.00070478994,0.014265921,0.0008727663,0.16411243,0.80236757,0.00039285942],"about_ca_topic_score_codex":0.0019497083,"about_ca_topic_score_gemma":0.0039074393,"teacher_disagreement_score":0.85294145,"about_ca_system_score_codex":0.0038557881,"about_ca_system_score_gemma":0.016602026,"threshold_uncertainty_score":0.7777292},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0014952171,0.9978625,0.000064178166,0.00009546546,0.000070821276,0.00016869008,0.00011814724,0.0000040582004,0.00012082889],"genre_scores_gemma":[0.032317225,0.9658223,0.00067688595,0.00042509547,0.00012099379,0.00042706582,0.00013724336,0.000004238984,0.0000690034],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9938414,0.0026785098,0.002048167,0.00040978304,0.0008299065,0.00019221756],"domain_scores_gemma":[0.9849946,0.011097786,0.0027729175,0.00016994005,0.0007765778,0.0001882208],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.007255884,0.0013925366,0.011962084,0.0049302992,0.00047234516,0.0022707442,0.0014242933,0.0020780584,0.0034771452],"category_scores_gemma":[0.023266185,0.0008022395,0.012263603,0.004712228,0.00074168824,0.001601523,0.00094095466,0.0014117681,0.0001778165],"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":[0.001494589,0.000056416116,0.0009491986,0.9205347,0.041880894,0.00007108436,0.000086133645,0.00015424876,0.00015294783,0.00014453383,0.0005454171,0.033929747],"study_design_scores_gemma":[0.0028983937,0.0013160908,0.0059615015,0.6156521,0.36216745,0.00033596938,0.00022409142,0.00020150904,0.00028890616,0.000292934,0.010611468,0.000049553262],"about_ca_topic_score_codex":0.005804485,"about_ca_topic_score_gemma":0.012401035,"teacher_disagreement_score":0.011962084,"about_ca_system_score_codex":0.0026507857,"about_ca_system_score_gemma":0.0056422804,"threshold_uncertainty_score":0.038373232},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0048413295,0.99300224,0.00058137736,0.00019344974,0.00014566912,0.00036477676,0.00061017345,0.000027963642,0.00023297248],"genre_scores_gemma":[0.19205578,0.80003524,0.002816149,0.001102884,0.00035226968,0.002001432,0.0011619086,0.000043905584,0.0004304264],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.98980826,0.004536212,0.0033246812,0.0010270148,0.001025911,0.00027784245],"domain_scores_gemma":[0.97946036,0.015342024,0.0030403421,0.00059969915,0.0013575373,0.00019998469],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01410442,0.0028402684,0.020594161,0.0060323756,0.0006594484,0.0037236803,0.002192006,0.0024094607,0.0030117733],"category_scores_gemma":[0.037607547,0.0015444851,0.038606808,0.007508132,0.0008171448,0.0018968225,0.0014136389,0.0017701784,0.0003150369],"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":[0.0011867451,0.000017839864,0.0035355492,0.38631603,0.60034627,0.000098018165,0.00007472347,0.00046957433,0.00018492226,0.00010460961,0.00045662455,0.0072091296],"study_design_scores_gemma":[0.00024947664,0.00009618199,0.002148495,0.018465918,0.97774017,0.000059166166,0.000027576161,0.00015306937,0.00007817543,0.000115771705,0.0008489978,0.00001695548],"about_ca_topic_score_codex":0.006210338,"about_ca_topic_score_gemma":0.012397199,"teacher_disagreement_score":0.020594161,"about_ca_system_score_codex":0.0025079516,"about_ca_system_score_gemma":0.00344031,"threshold_uncertainty_score":0.07459223},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004620193,0.993178,0.0007694474,0.00018930499,0.00019689955,0.00038571173,0.0004378791,0.000025504094,0.00019706617],"genre_scores_gemma":[0.18242104,0.8087438,0.0036814052,0.0013359494,0.00037098362,0.0019980865,0.0009794168,0.00005140226,0.0004179471],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.98462427,0.0087017575,0.003744559,0.0012515573,0.0013672636,0.00031055],"domain_scores_gemma":[0.97069114,0.022454128,0.0040658684,0.00089031487,0.0016803358,0.00021824244],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.021244913,0.0031532897,0.021714965,0.0050703767,0.0006211421,0.0035574327,0.0023675044,0.0024742316,0.0026789594],"category_scores_gemma":[0.046958312,0.0017830771,0.04944241,0.006015818,0.00087041216,0.001680309,0.0015501479,0.0023918652,0.0002921503],"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":[0.001342244,0.000023643483,0.0023059142,0.26154336,0.7278611,0.00007163216,0.000052793617,0.00045171552,0.00014908653,0.000090468675,0.00031342445,0.0057946355],"study_design_scores_gemma":[0.00039477082,0.0001535768,0.0019837003,0.013760256,0.9823491,0.00005364215,0.000023234908,0.00015214279,0.000089347035,0.0001521686,0.00087130984,0.000016759863],"about_ca_topic_score_codex":0.00705966,"about_ca_topic_score_gemma":0.013069721,"teacher_disagreement_score":0.021714965,"about_ca_system_score_codex":0.0026791631,"about_ca_system_score_gemma":0.0036003378,"threshold_uncertainty_score":0.11235517},"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":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99100554,0.0006122552,0.005179576,0.0001193574,0.000018372237,0.0014154759,0.00028690405,0.000026687712,0.0013359397],"genre_scores_gemma":[0.9925799,0.0002537766,0.0048512053,0.000055856362,0.000015556823,0.0018547775,0.00021761216,0.000013836275,0.00015759391],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9233014,0.04075104,0.014518478,0.0025303327,0.01748438,0.0014143507],"domain_scores_gemma":[0.6687489,0.2434909,0.03659208,0.007013797,0.042567592,0.0015867477],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.082743816,0.00037978822,0.0010843416,0.006123594,0.00080411276,0.0017190531,0.00096559944,0.00047863382,0.0010018647],"category_scores_gemma":[0.19284134,0.00049746566,0.0015175699,0.004140131,0.0021604132,0.0020917035,0.0014991302,0.00063556974,0.00021441517],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00085491256,0.0007600893,0.89496434,0.002644132,0.00051958277,0.0003083533,0.041126657,0.0005064841,0.0014806367,0.00058905064,0.0006347163,0.05561113],"study_design_scores_gemma":[0.00013741427,0.004665602,0.9270495,0.0014658319,0.00044402265,0.0012624589,0.053783488,0.003359565,0.0025207025,0.00070677453,0.004431641,0.00017307722],"about_ca_topic_score_codex":0.0012087794,"about_ca_topic_score_gemma":0.0012990371,"teacher_disagreement_score":0.9172562,"about_ca_system_score_codex":0.0023341563,"about_ca_system_score_gemma":0.0026844493,"threshold_uncertainty_score":0.43759638},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99441564,0.00029622315,0.0014163272,0.00014029512,0.00005206239,0.0009559804,0.0014803055,0.000015621205,0.001227685],"genre_scores_gemma":[0.99080527,0.0001995139,0.0037731556,0.00023267306,0.00003571558,0.0013254929,0.003125493,0.000008807331,0.0004939733],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9959526,0.0015431137,0.0005806883,0.0005256686,0.0011651021,0.0002328807],"domain_scores_gemma":[0.9958407,0.0009742071,0.0010098249,0.00049079995,0.0014016501,0.00028286324],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.011307238,0.00068231946,0.00051827607,0.001269165,0.0006695263,0.00060363195,0.0007653411,0.0006958167,0.0009473722],"category_scores_gemma":[0.013237239,0.00025013,0.0010761674,0.0014337875,0.0006201007,0.0005715896,0.001080129,0.000627406,0.00024474494],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00033531373,0.0003867975,0.9779187,0.00013383091,0.00023055295,0.00007593699,0.0006143603,0.00030577462,0.0010549832,0.00014757372,0.00072513544,0.018071095],"study_design_scores_gemma":[0.00008486654,0.00027557745,0.99775773,0.00003171452,0.000058910886,0.000071411996,0.00017753447,0.00051159435,0.00023953318,0.00006114603,0.00072157895,0.000008268289],"about_ca_topic_score_codex":0.0063321856,"about_ca_topic_score_gemma":0.0075672218,"teacher_disagreement_score":0.011307238,"about_ca_system_score_codex":0.00072670256,"about_ca_system_score_gemma":0.0016164259,"threshold_uncertainty_score":0.059799135},"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":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0033658862,0.18597229,0.15005197,0.5248,0.042148296,0.014840431,0.022260338,0.013628833,0.04293199],"genre_scores_gemma":[0.029770428,0.13248093,0.66500497,0.07168629,0.02085326,0.026155207,0.039957926,0.00345598,0.010635041],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.8416474,0.09443816,0.030454654,0.007360838,0.02172105,0.0043779584],"domain_scores_gemma":[0.4892381,0.14876147,0.06364787,0.057516787,0.16416708,0.07666872],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.3182781,0.0021815826,0.0071032424,0.01714168,0.0025919222,0.023900818,0.0092257,0.011349755,0.025552943],"category_scores_gemma":[0.31523806,0.0018078699,0.0040554376,0.024115587,0.0057933982,0.02708843,0.014965034,0.014260018,0.02342122],"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":[0.0012982538,0.00066246465,0.005265776,0.017643753,0.0003785237,0.00030715263,0.0008789943,0.0006282281,0.0011086474,0.09001578,0.38463536,0.49717706],"study_design_scores_gemma":[0.00087434234,0.00068564265,0.0057695596,0.02767349,0.00053468684,0.0006006643,0.00085598987,0.0018181996,0.00079151284,0.033717528,0.9264559,0.00022235031],"about_ca_topic_score_codex":0.0023563784,"about_ca_topic_score_gemma":0.0012343639,"teacher_disagreement_score":0.6817219,"about_ca_system_score_codex":0.007598347,"about_ca_system_score_gemma":0.110327125,"threshold_uncertainty_score":0.8406845},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.010907395,0.98634607,0.00038971877,0.00023078317,0.00010706703,0.00018955412,0.0015266519,0.000022614175,0.00028012518],"genre_scores_gemma":[0.28036743,0.71393037,0.00151127,0.0008594401,0.00029221294,0.0008172182,0.0018155988,0.00002955733,0.00037690878],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9952708,0.0018832358,0.0014262977,0.00063562085,0.000570896,0.00021307207],"domain_scores_gemma":[0.98969847,0.007367513,0.0015867689,0.0003106347,0.0008669109,0.00016971987],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0057861023,0.0014122112,0.012270085,0.004584828,0.00050887023,0.0020023764,0.0017151664,0.0015653126,0.0033303006],"category_scores_gemma":[0.015236139,0.00084134383,0.020269258,0.00807402,0.0004212132,0.0010417685,0.0010411202,0.001176293,0.00023598029],"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":[0.0022272277,0.000034177865,0.018372392,0.46954948,0.4933098,0.00022620264,0.0001228161,0.00061694306,0.00029292793,0.00016990553,0.0015460356,0.013532051],"study_design_scores_gemma":[0.00055788975,0.00018322366,0.014820928,0.030790597,0.9503934,0.00017576104,0.000066366105,0.00034314112,0.00012008826,0.00021185672,0.0023016161,0.000035229197],"about_ca_topic_score_codex":0.0058256458,"about_ca_topic_score_gemma":0.0132424785,"teacher_disagreement_score":0.012270085,"about_ca_system_score_codex":0.001370069,"about_ca_system_score_gemma":0.0025186564,"threshold_uncertainty_score":0.03060019},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0047167377,0.99369437,0.00048344687,0.00016464255,0.00013626677,0.00029073292,0.00034587775,0.000014854759,0.00015315524],"genre_scores_gemma":[0.1692517,0.82429194,0.0027424311,0.00086979615,0.00035771218,0.0014281397,0.00071923336,0.000023443165,0.00031568256],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9903346,0.004380494,0.0028328078,0.0010262148,0.0010928141,0.00033310454],"domain_scores_gemma":[0.9838568,0.011330343,0.0027817232,0.0004643721,0.0013369677,0.00022979188],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.012909768,0.0029594852,0.019681469,0.008255034,0.0007110623,0.0040273294,0.0021120734,0.0024988367,0.0038093098],"category_scores_gemma":[0.0274508,0.00172749,0.03945018,0.008592819,0.00070286804,0.0019745184,0.001607821,0.0018461278,0.00027032138],"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":[0.0011298279,0.000036232515,0.005021218,0.37549475,0.6080173,0.00015381625,0.00007683976,0.00043953798,0.00019941194,0.00010654245,0.00033830298,0.008986152],"study_design_scores_gemma":[0.00032874048,0.00019145788,0.0037606535,0.026682965,0.9673899,0.00010420092,0.00004905289,0.00017766567,0.00008153754,0.00012874215,0.001083546,0.000021612406],"about_ca_topic_score_codex":0.0055830167,"about_ca_topic_score_gemma":0.01173347,"teacher_disagreement_score":0.019681469,"about_ca_system_score_codex":0.0021533975,"about_ca_system_score_gemma":0.004298753,"threshold_uncertainty_score":0.0682742},"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":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.14792736,0.649696,0.09754977,0.04056868,0.0059660287,0.017926216,0.026689067,0.0008208756,0.0128559405],"genre_scores_gemma":[0.7617657,0.089499705,0.09550637,0.013826878,0.0038265204,0.023630477,0.01077316,0.00037728748,0.00079391047],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.10007532,0.44980037,0.3390481,0.013613395,0.09548676,0.0019760397],"domain_scores_gemma":[0.016769396,0.69769454,0.21333018,0.026343673,0.044855215,0.0010069221],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.7016636,0.002315742,0.007934334,0.04406577,0.003429187,0.019633796,0.005580706,0.00548118,0.0034077857],"category_scores_gemma":[0.9251468,0.0029715693,0.009184477,0.04662166,0.008500361,0.015907213,0.012792867,0.0040790215,0.0007850373],"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":[0.0047979695,0.00026058048,0.25376606,0.40960073,0.04039723,0.00069182966,0.01477088,0.0016947411,0.0014081944,0.009071986,0.014034992,0.24950491],"study_design_scores_gemma":[0.0031232787,0.0024278013,0.30284217,0.48840442,0.05042178,0.0044492627,0.009051112,0.0047737635,0.005888495,0.051885933,0.07565788,0.0010741316],"about_ca_topic_score_codex":0.0025883215,"about_ca_topic_score_gemma":0.0030390904,"teacher_disagreement_score":0.2983364,"about_ca_system_score_codex":0.009459897,"about_ca_system_score_gemma":0.021068292,"threshold_uncertainty_score":0.36790192},"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":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0025632484,0.98073304,0.002831531,0.0058614593,0.00046065048,0.0019516156,0.0043587005,0.00014405408,0.0010956825],"genre_scores_gemma":[0.06729083,0.8880231,0.029368376,0.004682052,0.0004886092,0.006273586,0.003559109,0.00007736629,0.00023695678],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9484039,0.022230498,0.018214623,0.0038079235,0.006278288,0.0010647561],"domain_scores_gemma":[0.8173705,0.14494689,0.018264577,0.003421412,0.014792756,0.001203798],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.042548772,0.0035639124,0.013617524,0.066541485,0.0016424521,0.007101406,0.0038701259,0.0035593926,0.0042572687],"category_scores_gemma":[0.19981447,0.002033067,0.0107413065,0.042814806,0.0019457205,0.010120385,0.007634164,0.0033601215,0.00045818425],"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":[0.00023131633,0.000027793785,0.0021443423,0.8985544,0.014177686,0.00021394835,0.0010702029,0.00082249416,0.0002636591,0.0031810915,0.0054183784,0.073894754],"study_design_scores_gemma":[0.0002557826,0.00015422728,0.0033899925,0.8948238,0.060970612,0.00024100291,0.0008955198,0.0006770293,0.00019775603,0.00698986,0.03132646,0.000077933655],"about_ca_topic_score_codex":0.010925542,"about_ca_topic_score_gemma":0.019311007,"teacher_disagreement_score":0.9863825,"about_ca_system_score_codex":0.007779071,"about_ca_system_score_gemma":0.022199789,"threshold_uncertainty_score":0.22502214},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0011162199,0.984402,0.0030917304,0.00596478,0.00055424887,0.0004211461,0.0005275328,0.000029573344,0.0038927686],"genre_scores_gemma":[0.03170055,0.95180696,0.011804558,0.0027470263,0.00015098978,0.0008285343,0.00053105416,0.000023704006,0.0004066966],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.9861193,0.0065041077,0.003716331,0.0008925967,0.0024350616,0.00033264066],"domain_scores_gemma":[0.9113377,0.070485935,0.0058301073,0.0027841658,0.009004608,0.00055740937],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.031131243,0.0015732283,0.0031928562,0.015249193,0.0011381519,0.00582803,0.003225448,0.0026298424,0.0059233736],"category_scores_gemma":[0.095410876,0.0011110483,0.0037196002,0.013427421,0.0017399911,0.006880645,0.003981748,0.0031962045,0.0008144237],"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":[0.00010050305,0.000059587102,0.0008987534,0.58010715,0.0020685361,0.00022942759,0.00457421,0.0013024766,0.00047208156,0.029505098,0.010920383,0.36976174],"study_design_scores_gemma":[0.00004836145,0.00008661685,0.0017645753,0.7649665,0.0045760158,0.00021506104,0.004245151,0.0002638619,0.00049005327,0.01034621,0.21294935,0.000048128542],"about_ca_topic_score_codex":0.016193504,"about_ca_topic_score_gemma":0.031430103,"teacher_disagreement_score":0.031131243,"about_ca_system_score_codex":0.008946204,"about_ca_system_score_gemma":0.04931409,"threshold_uncertainty_score":0.16463971},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99932504,0.00007687252,0.0000623453,0.00007260265,0.0000020000539,0.000020629943,0.00005207426,9.609871e-7,0.00038752856],"genre_scores_gemma":[0.9994894,0.00012326999,0.00011013029,0.00007277627,0.0000036596527,0.000034849174,0.000053750515,6.070244e-7,0.00011148672],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99889237,0.0003939194,0.00015921991,0.00011819938,0.000270122,0.0001662186],"domain_scores_gemma":[0.9943982,0.0016864645,0.0024024944,0.00018542029,0.0006661121,0.00066128926],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003336525,0.00019523544,0.00032223001,0.001181893,0.0005783323,0.0009833768,0.00027727018,0.0007160048,0.0019940063],"category_scores_gemma":[0.007004482,0.00040748843,0.0004038633,0.0009604567,0.0005175903,0.0014834745,0.00093070307,0.00075444375,0.00033489612],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000039636463,0.00029026726,0.9893694,0.00006452023,0.000023012464,0.00010643677,0.006599117,0.000021450853,0.00023934574,0.000039789167,0.00009211668,0.003114908],"study_design_scores_gemma":[0.0000070673777,0.0006540568,0.9792159,0.00008460227,0.000029929168,0.00040811865,0.01861517,0.00020106598,0.00016762722,0.00004153491,0.00056164124,0.000013325292],"about_ca_topic_score_codex":0.0015534667,"about_ca_topic_score_gemma":0.0015619856,"teacher_disagreement_score":0.003336525,"about_ca_system_score_codex":0.00036515653,"about_ca_system_score_gemma":0.0005958365,"threshold_uncertainty_score":0.017645419},"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":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.001146722,0.9754252,0.015515688,0.0038628748,0.00056213525,0.0019352733,0.00065447105,0.00009351719,0.0008041392],"genre_scores_gemma":[0.042985532,0.86950845,0.07079429,0.004306961,0.0004791321,0.010740184,0.000867817,0.00012829927,0.0001893894],"study_design_codex":"systematic_review","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.5361184,0.29659525,0.13750133,0.007942986,0.020847142,0.0009948594],"domain_scores_gemma":[0.1483331,0.786617,0.03512926,0.009879897,0.019243337,0.00079735403],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.31727886,0.003211208,0.015176622,0.019658366,0.0014365489,0.009897031,0.0050872676,0.006596493,0.005911999],"category_scores_gemma":[0.73009175,0.003970553,0.018294698,0.014980787,0.0040584086,0.010319865,0.0046966537,0.0054407557,0.0009225409],"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":[0.0008202936,0.000036821464,0.001601571,0.8140092,0.015631843,0.00007368203,0.00076218945,0.0019005563,0.00015429509,0.0061485483,0.0026837462,0.15617731],"study_design_scores_gemma":[0.0007802923,0.00034668576,0.0011404188,0.9429342,0.02541695,0.0002685042,0.00039765806,0.0017300879,0.00041044742,0.009033251,0.017428983,0.000112532754],"about_ca_topic_score_codex":0.0053202235,"about_ca_topic_score_gemma":0.007643208,"teacher_disagreement_score":0.68272114,"about_ca_system_score_codex":0.011357093,"about_ca_system_score_gemma":0.027520347,"threshold_uncertainty_score":0.8419167},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0063247657,0.9912847,0.0006715863,0.00020581385,0.00019382419,0.0005995053,0.0003774796,0.00003830061,0.00030406757],"genre_scores_gemma":[0.15723583,0.8339213,0.0034948073,0.0012736655,0.0003388579,0.002514593,0.00066655665,0.000034863802,0.0005194476],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9929055,0.0037960252,0.0016831447,0.0007112278,0.0006700964,0.0002340322],"domain_scores_gemma":[0.9887792,0.0082222065,0.0017238602,0.00034491374,0.00077245786,0.00015743951],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0104214605,0.0023583611,0.020033054,0.004155121,0.0006238605,0.0026416352,0.0019644042,0.0024188096,0.0044570225],"category_scores_gemma":[0.020768909,0.0012470334,0.03540437,0.004631094,0.00060128036,0.0013932654,0.0011888462,0.0020719466,0.00031862225],"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":[0.004087311,0.000108380715,0.0024681774,0.4305617,0.54231805,0.000092160895,0.00008773891,0.00046869274,0.00030714003,0.00015966782,0.0006315404,0.018709417],"study_design_scores_gemma":[0.001481168,0.0005143357,0.0029805754,0.02836707,0.9644375,0.00006049516,0.000039506627,0.00020135846,0.00013003877,0.000186227,0.0015794017,0.00002237698],"about_ca_topic_score_codex":0.0049116407,"about_ca_topic_score_gemma":0.009994515,"teacher_disagreement_score":0.020033054,"about_ca_system_score_codex":0.0020791742,"about_ca_system_score_gemma":0.002960448,"threshold_uncertainty_score":0.055114567},"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":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0016245069,0.017032994,0.02155854,0.94117004,0.008332465,0.00079598074,0.00023022703,0.0003670463,0.0088882325],"genre_scores_gemma":[0.16109781,0.032770164,0.4812506,0.2952994,0.015558061,0.00632057,0.001088463,0.0014540388,0.005160854],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.41915226,0.3874573,0.0955185,0.012875077,0.07561824,0.009378606],"domain_scores_gemma":[0.11392107,0.55490386,0.040785313,0.05362252,0.20367165,0.033095602],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.5891001,0.0019841809,0.0063627716,0.013791167,0.0075337845,0.04579047,0.013335533,0.019050827,0.013006893],"category_scores_gemma":[0.847019,0.00402389,0.003980133,0.010258599,0.016367598,0.04194397,0.04061167,0.0358435,0.004482438],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00069742603,0.00042180798,0.0039834003,0.023005487,0.0017651654,0.00033583553,0.018184386,0.0015109826,0.000571846,0.1814509,0.18388632,0.58418643],"study_design_scores_gemma":[0.00082198565,0.00042994888,0.0027219644,0.09533822,0.0013752959,0.00042934736,0.011106168,0.00511674,0.0009656291,0.5397175,0.34154487,0.00043222422],"about_ca_topic_score_codex":0.012886927,"about_ca_topic_score_gemma":0.034101814,"teacher_disagreement_score":0.41089988,"about_ca_system_score_codex":0.022574352,"about_ca_system_score_gemma":0.1757558,"threshold_uncertainty_score":0.50671273},"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":"protocol","genre_gemma":"methods","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.003903976,0.0198798,0.35600933,0.10186697,0.010685447,0.4503735,0.020971078,0.0045211716,0.031788763],"genre_scores_gemma":[0.009634706,0.01648865,0.4947835,0.018117951,0.0017946394,0.44525585,0.007492262,0.0006276928,0.005804697],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.33245805,0.32613918,0.30384508,0.004231297,0.03010244,0.003223955],"domain_scores_gemma":[0.22194748,0.41075096,0.117076755,0.04899982,0.19909473,0.0021303543],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.39223248,0.004071164,0.006157117,0.019118058,0.003674689,0.0075107696,0.008275276,0.014818012,0.013072237],"category_scores_gemma":[0.5876247,0.0043000085,0.010839774,0.013748307,0.006016628,0.009098157,0.009151436,0.010258812,0.013910501],"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":[0.00083803164,0.00041868727,0.003065331,0.10054082,0.00046560378,0.0015011989,0.01626099,0.0019029526,0.0033703805,0.031220712,0.5646835,0.27573186],"study_design_scores_gemma":[0.0011084208,0.00033867266,0.0036861484,0.13014479,0.0005261065,0.0019764977,0.004928996,0.002734165,0.0036041017,0.019286606,0.831223,0.00044245066],"about_ca_topic_score_codex":0.0057881135,"about_ca_topic_score_gemma":0.0061864755,"teacher_disagreement_score":0.6077675,"about_ca_system_score_codex":0.008426386,"about_ca_system_score_gemma":0.03500251,"threshold_uncertainty_score":0.7494856},"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":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5297196,0.0043549067,0.27769387,0.045600142,0.0016993715,0.11152581,0.00081047154,0.00035380755,0.02824207],"genre_scores_gemma":[0.74616456,0.0016462917,0.19946265,0.002348868,0.0001892395,0.04888627,0.00030570076,0.000067614565,0.0009289386],"study_design_codex":"design_other","study_design_gemma":"qualitative","domain_scores_codex":[0.49426067,0.43373188,0.032397736,0.005987179,0.025974525,0.007648125],"domain_scores_gemma":[0.527094,0.31232965,0.016273886,0.016079094,0.1173478,0.010875526],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.49670556,0.0012326398,0.0017863299,0.005759968,0.0049409103,0.006583204,0.0035113734,0.0031172289,0.0025981388],"category_scores_gemma":[0.42002308,0.0011986529,0.0027746707,0.002768379,0.0053474363,0.007931917,0.014012551,0.004967715,0.0006201521],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002255868,0.0014530191,0.058370132,0.018818652,0.0011778729,0.0017535525,0.3385253,0.0073084026,0.008252185,0.04271732,0.025457632,0.49390996],"study_design_scores_gemma":[0.0017498235,0.007087184,0.069878645,0.043282047,0.0018288272,0.0032076691,0.56582755,0.07321878,0.01484133,0.07446494,0.14306073,0.001552531],"about_ca_topic_score_codex":0.0023677915,"about_ca_topic_score_gemma":0.0018762178,"teacher_disagreement_score":0.49670556,"about_ca_system_score_codex":0.012770736,"about_ca_system_score_gemma":0.05556869,"threshold_uncertainty_score":0.6206516},"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"}]}