{"meta":{"page":1,"per_page":50,"max_per_page":100,"total":125,"total_is_capped":false,"direct_labels_cover":115,"predictions_cover":125,"direct_label_status":"direct model label, unvalidated","prediction_status":"machine_predicted_unvalidated (Codex and Gemma teacher distillation)","score_status":"score_only:v0-immature-baseline (scores rank; they never assert a category)","snapshot":{"source":"OpenAlex, pinned release, all 482 partitions","release":"2026-06-24","frame_built":"2026-07-12","author_layer_release":"2026-06-26"},"query_hash":"d263e7495388","filters":{"venue":"BMJ evidence-based medicine"}},"results":[{"id":"W2901928126","doi":"10.1136/bmjebm-2018-111019","title":"Evidence-based practice education for healthcare professions: an expert view","year":2018,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":317,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Curriculum; Health care; Stakeholder; Medical education; Professional development; Stakeholder engagement; Best practice; Sociology; Pedagogy; Medicine; Psychology; Political science; Nursing; Public relations","authors":[{"name":"Elaine Lehane","is_ca":false},{"name":"Patricia Leahy‐Warren","is_ca":false},{"name":"Cliona O’Riordan","is_ca":false},{"name":"Eileen Savage","is_ca":false},{"name":"Jonathan Drennan","is_ca":false},{"name":"Colm O’Tuathaigh","is_ca":false},{"name":"Michael K. O’Connor","is_ca":false},{"name":"Mark Corrigan","is_ca":false},{"name":"Francis Burke","is_ca":false},{"name":"Martina Hayes","is_ca":false},{"name":"Helen Lynch","is_ca":false},{"name":"Laura J. Sahm","is_ca":false},{"name":"Elizabeth Heffernan","is_ca":false},{"name":"Elizabeth O’Keeffe","is_ca":false},{"name":"Catherine Blake","is_ca":false},{"name":"Frances Horgan","is_ca":false},{"name":"Josephine Hegarty","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.5722276379167234,"gpt":0.6578693978875588,"spread":0.08564175997083545,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.02016616,0.0003875136,0.0006490778,0.0006079188,0.00317977,0.00003706049,0.0008622458,0.0003962928,0.002112177],"category_scores_gemma":[0.09352127,0.0002787618,0.00009256976,0.001684871,0.0007924275,0.00260702,0.00006777501,0.0009310627,0.0005133643],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001210096,"about_ca_system_score_gemma":0.04385397,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.006978742,"about_ca_topic_score_gemma":0.0009931956,"domain_scores_codex":[0.9887704,0.004990166,0.001726051,0.00107147,0.001788898,0.001653066],"domain_scores_gemma":[0.970828,0.0186473,0.001052281,0.001486461,0.006127128,0.001858809],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.003374484,0.0007681067,0.009323052,0.006068076,0.00001518285,0.000004364158,0.01196185,0.000004334921,0.001429285,0.002251443,0.7812791,0.1835207],"study_design_scores_gemma":[0.002906409,0.01141549,0.0134645,0.1343673,0.0001017312,0.000009280474,0.05321521,0.005130881,0.0006075369,0.001814172,0.7761946,0.0007728825],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.01285277,0.02996219,0.01721189,0.9129424,0.01095151,0.01494967,0.00001474172,0.000299648,0.0008151939],"genre_scores_gemma":[0.4999389,0.007722046,0.04878257,0.3840789,0.03250737,0.02348084,0.0002554788,0.000129563,0.003104284],"genre_candidate":"commentary","genre_consensus":null,"teacher_disagreement_score":0.5288634,"threshold_uncertainty_score":0.9999664,"prediction_status":"machine_predicted_unvalidated"},"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":"W4224095215","doi":"10.1136/bmjebm-2022-111962","title":"Good or best practice statements: proposal for the operationalisation and implementation of GRADE guidance","year":2022,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":164,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Université Laval; Ottawa Hospital; Centre hospitalier universitaire de Québec; McMaster University; Impact; Cochrane; Bruyère; Hôpital de l'Enfant-Jésus; University of Ottawa","funders":"Canadian Institutes of Health Research; World Health Organization","keywords":"Grading (engineering); Guideline; Global Positioning System; Computer science; Fidelity; Process management; Best practice; Medical education; Knowledge management; Medicine; Business; Engineering; Political science","authors":[{"name":"Omar Dewidar","is_ca":true},{"name":"Tamara Lotfi","is_ca":true},{"name":"Miranda Langendam","is_ca":false},{"name":"Elena Parmelli","is_ca":false},{"name":"Zuleika Saz Parkinson","is_ca":false},{"name":"Karla Solo","is_ca":true},{"name":"Derek K. Chu","is_ca":true},{"name":"Joseph L. Mathew","is_ca":false},{"name":"Elie A. Akl","is_ca":true},{"name":"Romina Brignardello‐Petersen","is_ca":true},{"name":"Reem A. Mustafa","is_ca":false},{"name":"Lorenzo Moja","is_ca":false},{"name":"Alfonso Iorio","is_ca":true},{"name":"Yuan Chi","is_ca":false},{"name":"Carlos Canelo‐Aybar","is_ca":false},{"name":"Tamara Kredo","is_ca":false},{"name":"Justine Karpusheff","is_ca":false},{"name":"Alexis F. Turgeon","is_ca":true},{"name":"Pablo Alonso‐Coello","is_ca":false},{"name":"Wojtek Wiercioch","is_ca":true},{"name":"Annette Gerritsen","is_ca":false},{"name":"Miloslav Klugar","is_ca":false},{"name":"María Ximena Rojas","is_ca":false},{"name":"Peter Tugwell","is_ca":true},{"name":"Vivian Welch","is_ca":true},{"name":"Kevin Pottie","is_ca":true},{"name":"Zachary Munn","is_ca":false},{"name":"Robby Nieuwlaat","is_ca":true},{"name":"Nathan Ford","is_ca":false},{"name":"Adrienne Stevens","is_ca":true},{"name":"Joanne Khabsa","is_ca":false},{"name":"Zil Nasir","is_ca":true},{"name":"Grigorios I. Leontiadis","is_ca":true},{"name":"Joerg J Meerpohl","is_ca":false},{"name":"Thomas Piggott","is_ca":true},{"name":"Amir Qaseem","is_ca":false},{"name":"Micayla Matthews","is_ca":true},{"name":"Holger J. Schünemann","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.5414977190713219,"gpt":0.5727772921882497,"spread":0.03127957311692775,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.03090917,0.0001258853,0.000445667,0.0001651625,0.000449919,0.00002883893,0.0002039816,0.00002886714,0.001054842],"category_scores_gemma":[0.01242347,0.0001064041,0.00003918053,0.0002322727,0.0001173055,0.000895394,0.00004391727,0.0001021971,0.00001433319],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004153823,"about_ca_system_score_gemma":0.0008746596,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002609449,"about_ca_topic_score_gemma":0.0003565639,"domain_scores_codex":[0.9957404,0.0005720393,0.002860063,0.0003561192,0.0002685684,0.0002028763],"domain_scores_gemma":[0.9899577,0.006935209,0.002501218,0.0003223797,0.0002128591,0.00007062926],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","study_design_scores_codex":[0.001127651,0.0005499142,0.05465704,0.002885794,0.0003893438,0.000003499483,0.01363771,0.00722677,0.0002044199,0.7482204,0.1665387,0.004558735],"study_design_scores_gemma":[0.01778954,0.01012298,0.03524067,0.001760195,0.000394723,0.00006452195,0.1307781,0.06329753,0.0003669008,0.03022794,0.7088756,0.001081271],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.02243711,0.01088361,0.06795789,0.8902171,0.0008955941,0.006840526,0.0006155295,0.00002534518,0.0001272866],"genre_scores_gemma":[0.8931047,0.001237923,0.03723871,0.05731919,0.001436243,0.008106401,0.0003761333,0.00006406169,0.001116574],"genre_candidate":"commentary","genre_consensus":null,"teacher_disagreement_score":0.8706676,"threshold_uncertainty_score":0.9998583,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W4285794520","doi":"10.1136/bmjebm-2022-111952","title":"Adapt or die: how the pandemic made the shift from EBM to EBM+ more urgent","year":2022,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Healthcare cost, quality, practices","field":"Health Professions","cited_by":123,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"CARE Canada; Association of Professional Engineers and Geoscientists of Manitoba; Public Health Ontario; University of Toronto","funders":"National Institute for Health and Care Research; Canadian Institutes of Health Research; NIHR Oxford Biomedical Research Centre; National Medical Research Council; National Health and Medical Research Council; Wellcome Trust","keywords":"Context (archaeology); Checklist; Evidence-based medicine; Psychological intervention; Quality (philosophy); Yesterday; Evidence-based practice; Health care; Scientific evidence; Intervention (counseling); Psychology; Engineering ethics; MEDLINE; Management science; Medicine; Political science; Engineering; Alternative medicine; Nursing; Epistemology; History","authors":[{"name":"Trisha Greenhalgh","is_ca":false},{"name":"David N. Fisman","is_ca":true},{"name":"Danielle J. Cane","is_ca":true},{"name":"Matthew Oliver","is_ca":true},{"name":"C. Raina MacIntyre","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.7945608969354264,"gpt":0.5892763489863074,"spread":0.205284547949119,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","sts","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0194685,0.0004345082,0.0007741873,0.0001800305,0.003261043,0.00002909983,0.001643767,0.0002048537,0.01123932],"category_scores_gemma":[0.04954778,0.0002121302,0.000122944,0.001129354,0.0004194011,0.0003033144,0.0007553914,0.003281813,0.0003089907],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0009637572,"about_ca_system_score_gemma":0.005467749,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01650484,"about_ca_topic_score_gemma":0.006588056,"domain_scores_codex":[0.9820834,0.0119859,0.001389218,0.0008380513,0.002528933,0.001174455],"domain_scores_gemma":[0.9568813,0.0385507,0.001050513,0.002217322,0.0004068607,0.000893241],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.004640232,0.0001949398,0.1436288,0.000857466,0.00017085,0.0003430047,0.05556172,0.0001823491,0.0006965786,0.0006629606,0.7661598,0.02690128],"study_design_scores_gemma":[0.001685091,0.0009553066,0.1502526,0.003417216,0.0002557151,0.00001073332,0.0682379,0.0003728149,0.0000253186,0.0005634782,0.7738438,0.0003800507],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.1638168,0.007491655,0.0008456003,0.8189709,0.003106299,0.005269233,0.0001191633,0.0002130867,0.0001672187],"genre_scores_gemma":[0.7035744,0.0007164622,0.0002117662,0.2841518,0.003480951,0.004016275,0.00006192007,0.00006993154,0.003716565],"genre_candidate":"commentary","genre_consensus":null,"teacher_disagreement_score":0.5397575,"threshold_uncertainty_score":0.9990177,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4283587010","doi":"10.1136/bmjebm-2022-111928","title":"Theory and practice of Bayesian and frequentist frameworks for network meta-analysis","year":2022,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":114,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"St. Joseph’s Healthcare Hamilton; McMaster University; Ted Rogers Centre for Heart Research; University Health Network; Impact","funders":"","keywords":"Frequentist inference; Bayesian probability; Computer science; Bayesian network; Bayesian statistics; Frequentist probability; Machine learning; Bayesian inference; Artificial intelligence; Data science","authors":[{"name":"Behnam Sadeghirad","is_ca":true},{"name":"Farid Foroutan","is_ca":true},{"name":"Michael J. Zoratti","is_ca":true},{"name":"Jason W. Busse","is_ca":true},{"name":"Romina Brignardello‐Petersen","is_ca":true},{"name":"Gordon Guyatt","is_ca":true},{"name":"Lehana Thabane","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.7536134177763099,"gpt":0.5727628547311456,"spread":0.1808505630451642,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.3995094,0.0003369952,0.006762949,0.0006453599,0.0003869493,0.000186124,0.001029205,0.0001108922,0.02172274],"category_scores_gemma":[0.3029681,0.000161501,0.002438954,0.003447274,0.0003639947,0.0003326251,0.000197115,0.0003707439,0.000008795756],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0000288018,"about_ca_system_score_gemma":0.00012681,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00005985098,"about_ca_topic_score_gemma":0.00002741304,"domain_scores_codex":[0.9439968,0.03885422,0.008892325,0.00142,0.006467069,0.0003696226],"domain_scores_gemma":[0.8230204,0.1642877,0.007785227,0.003217987,0.00141853,0.0002700944],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"meta_analysis","study_design_scores_codex":[0.002101839,0.0003347886,0.02375129,0.001086305,0.2366084,0.0001118155,0.003197683,0.02543274,0.0001810917,0.2025056,0.4678226,0.03686589],"study_design_scores_gemma":[0.0009523735,0.001043653,0.002480156,0.0001831842,0.6646189,0.00008877558,0.006172001,0.03261759,0.00001298239,0.1161701,0.175134,0.000526203],"study_design_candidate":"meta_analysis","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.0008163686,0.1052002,0.8445493,0.04590775,0.0002561488,0.002326816,0.00004457649,0.000007041049,0.000891779],"genre_scores_gemma":[0.9014411,0.0003456014,0.07948804,0.01290115,0.0003853842,0.001041786,0.00002598918,0.00002764005,0.004343335],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.9006247,"threshold_uncertainty_score":0.9791715,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W3112825017","doi":"10.1136/bmjebm-2020-111493","title":"Framework for the synthesis of non-randomised studies and randomised controlled trials: a guidance on conducting a systematic review and meta-analysis for healthcare decision making","year":2020,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":103,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Canadian Agency for Drugs and Technologies in Health","funders":"Sanofi Pasteur; Sanofi","keywords":"Health care; Randomized controlled trial; Evidence-based medicine; Systematic review; Delphi method; Management science; Medicine; Quality (philosophy); Transparency (behavior); Best practice; MEDLINE; Computer science; Alternative medicine; Engineering; Artificial intelligence; Political science","authors":[{"name":"Grammati Sarri","is_ca":false},{"name":"Elisabetta Patorno","is_ca":false},{"name":"Hongbo Yuan","is_ca":true},{"name":"Jianfei Guo","is_ca":false},{"name":"Dimitri Bennett","is_ca":false},{"name":"Xuerong Wen","is_ca":false},{"name":"Andrew R. Zullo","is_ca":false},{"name":"Joan Largent","is_ca":false},{"name":"Mary Panaccio","is_ca":false},{"name":"Mugdha Gokhale","is_ca":false},{"name":"Daniela C. Moga","is_ca":false},{"name":"M. Sanni Ali","is_ca":false},{"name":"Thomas P. A. Debray","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9173359485862664,"gpt":0.6565785159814549,"spread":0.2607574326048115,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"category_scores_codex":[0.5970172,0.001718039,0.1738865,0.001354511,0.0005178866,0.0003912935,0.00236622,0.0003923773,0.0004501731],"category_scores_gemma":[0.957992,0.0005147084,0.03550149,0.003912121,0.0004268749,0.0001719973,0.0001578044,0.0004823596,0.00001146778],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001194312,"about_ca_system_score_gemma":0.0006336566,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00000823246,"about_ca_topic_score_gemma":0.00001172904,"domain_scores_codex":[0.8513703,0.07963595,0.05848384,0.002940883,0.006974075,0.0005949428],"domain_scores_gemma":[0.05543227,0.8646964,0.06914026,0.005479185,0.004909758,0.0003420697],"domain_codex":null,"domain_gemma":"methods","domain_candidate":"methods","domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","study_design_scores_codex":[0.003147752,0.00001506959,5.628877e-7,0.7209207,0.2279678,0.000003687289,0.0001774767,0.00001346776,1.068129e-7,0.0002152411,0.002917705,0.04462042],"study_design_scores_gemma":[0.008844359,0.0001962275,1.085854e-7,0.3787683,0.6059299,0.000006649555,0.0004453892,0.001099302,2.818342e-7,0.0007775103,0.003667225,0.0002647705],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[1.941899e-7,0.7137744,0.1917673,0.01441355,0.000172942,0.07974252,0.0001181433,0.000007243581,0.000003700114],"genre_scores_gemma":[0.0001700166,0.913206,0.02485973,0.003023552,0.0002888567,0.05827443,0.000006144409,0.00005968208,0.0001115576],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.7959381,"threshold_uncertainty_score":0.9997305,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2809195898","doi":"10.1136/bmjebm-2018-110968","title":"Reflections on the history of systematic reviews","year":2018,"lang":"en","type":"editorial","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":81,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University","funders":"","keywords":"Psychological intervention; Systematic review; MEDLINE; Medicine; Value (mathematics); Evidence-based medicine; Alternative medicine; Intensive care medicine; Computer science; Political science; Pathology; Psychiatry","authors":[{"name":"Mike Clarke","is_ca":true},{"name":"Iain Chalmers","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9165977472242489,"gpt":0.6280865975200167,"spread":0.2885111497042322,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","open_science","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.4978683,0.0009659021,0.01584924,0.0012657,0.0001909324,0.0001590277,0.006100334,0.0006462663,0.0314949],"category_scores_gemma":[0.8868983,0.0003152963,0.003967654,0.002288128,0.0009207086,0.0001928746,0.0001342893,0.001040751,0.008437886],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0009725193,"about_ca_system_score_gemma":0.002434244,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001005188,"about_ca_topic_score_gemma":0.0000941442,"domain_scores_codex":[0.8589927,0.07433193,0.03590377,0.002228125,0.02803316,0.0005103022],"domain_scores_gemma":[0.5640306,0.3390217,0.06010138,0.02475624,0.01162734,0.000462685],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00002920497,0.0000353239,0.000004243936,0.02003497,0.0001824332,0.000005080714,0.000378579,0.000003418317,0.00001965754,0.0003243017,0.9787899,0.0001929178],"study_design_scores_gemma":[0.0001615076,0.0004953168,0.000003405079,0.1277397,0.002069153,0.00000266049,0.0001986273,0.0001564648,0.000002903075,0.0005359518,0.868398,0.0002363605],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"editorial","genre_gemma":"editorial","genre_scores_codex":[0.00000378311,0.09123849,0.003996944,0.008269978,0.8676074,0.009696753,0.00002418527,0.00001452642,0.01914797],"genre_scores_gemma":[0.0001544413,0.003945023,0.0008806289,0.002722224,0.778227,0.002547956,0.00003986677,0.00007143576,0.2114115],"genre_candidate":"editorial","genre_consensus":"editorial","teacher_disagreement_score":0.3890301,"threshold_uncertainty_score":0.9999299,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4366386735","doi":"10.1136/bmjebm-2022-112185","title":"Rapid reviews methods series: Guidance on team considerations, study selection, data extraction and risk of bias assessment","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":78,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Data extraction; Correctness; Computer science; Systematic review; Completeness (order theory); Selection (genetic algorithm); Data mining; Information retrieval; MEDLINE; Artificial intelligence; Mathematics; Algorithm","authors":[{"name":"Barbara Nußbaumer-Streit","is_ca":false},{"name":"Isolde Sommer","is_ca":false},{"name":"Candyce Hamel","is_ca":true},{"name":"Declan Devane","is_ca":false},{"name":"Anna H Noel-Storr","is_ca":false},{"name":"Livia Puljak","is_ca":false},{"name":"Marialena Trivella","is_ca":false},{"name":"Gerald Gartlehner","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9188994641379558,"gpt":0.6765426507253051,"spread":0.2423568134126507,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.464401,0.0003776993,0.003930617,0.0009147648,0.0003361525,0.0002307662,0.0009927761,0.00007931896,0.004917746],"category_scores_gemma":[0.4995241,0.0001792641,0.0003461291,0.003664662,0.0001873903,0.0007935393,0.0001551671,0.0003018303,0.000280577],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00006013089,"about_ca_system_score_gemma":0.0003792458,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001754393,"about_ca_topic_score_gemma":0.0002105793,"domain_scores_codex":[0.9202019,0.06005432,0.01233125,0.001732664,0.005403225,0.0002766461],"domain_scores_gemma":[0.9172035,0.06180756,0.01117355,0.007394707,0.002166582,0.0002541417],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00005033608,0.0001709282,0.1395222,0.0002974029,0.0003053707,0.000008573134,0.0005091123,0.0003142093,0.0005449297,0.0002723169,0.7208855,0.1371191],"study_design_scores_gemma":[0.001550535,0.003049057,0.262053,0.004479408,0.003245573,0.00006005912,0.006860294,0.06030317,0.0003596329,0.003427356,0.6540655,0.0005464271],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.09856418,0.07306467,0.7290888,0.05918693,0.004163868,0.03103115,0.000182754,0.0001877888,0.004529863],"genre_scores_gemma":[0.3311495,0.06655069,0.5840201,0.003250472,0.001898042,0.002014414,0.0001478909,0.00009240501,0.01087651],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.2325853,"threshold_uncertainty_score":0.9959919,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3003086189","doi":"10.1136/bmjebm-2019-111268","title":"Assessment of assumptions of statistical analysis methods in randomised clinical trials: the what and how","year":2020,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":68,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"Novo Nordisk Fonden","keywords":"Statistical analysis; Clinical trial; Statistics; Medical physics; Medicine; Econometrics; Mathematics; Internal medicine","authors":[{"name":"Anders Kehlet Nørskov","is_ca":false},{"name":"Theis Lange","is_ca":false},{"name":"Emil Eik Nielsen","is_ca":false},{"name":"Christian Gluud","is_ca":false},{"name":"Per Winkel","is_ca":false},{"name":"Jan Beyersmann","is_ca":false},{"name":"Jacobo de Uña‐Álvarez","is_ca":false},{"name":"Valter Torri","is_ca":false},{"name":"Laurent Billot","is_ca":false},{"name":"Hein Putter","is_ca":false},{"name":"Jørn Wetterslev","is_ca":false},{"name":"Lehana Thabane","is_ca":true},{"name":"Janus Christian Jakobsen","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9498731355574201,"gpt":0.7462174734380085,"spread":0.2036556621194117,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.6434263,0.0002453778,0.01022185,0.0005542534,0.00004836503,0.0001991041,0.00103323,0.0001021903,0.005867246],"category_scores_gemma":[0.7123132,0.00008922105,0.001955482,0.003718926,0.0006130015,0.0002924211,0.00008521668,0.000292823,0.000008487899],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001869148,"about_ca_system_score_gemma":0.0003366075,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002597173,"about_ca_topic_score_gemma":0.00003270548,"domain_scores_codex":[0.6728549,0.2803615,0.03481843,0.001756048,0.00983269,0.0003764261],"domain_scores_gemma":[0.4428292,0.5352082,0.01615646,0.003472601,0.00182777,0.0005057862],"domain_codex":null,"domain_gemma":"methods","domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.001635125,0.0002971468,0.5479127,0.0008495338,0.005498344,0.00002875154,0.001394586,0.002040623,0.00174814,0.006453751,0.05793776,0.3742035],"study_design_scores_gemma":[0.01229717,0.0007945096,0.470051,0.001520293,0.01738496,0.000002113488,0.006365019,0.483827,0.00007667261,0.00215421,0.005281995,0.0002450511],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.01128177,0.007559923,0.8924442,0.08633724,0.0001765913,0.002086048,0.00001345729,0.000002287615,0.00009846483],"genre_scores_gemma":[0.8445281,0.001686138,0.1514774,0.001785484,0.0002271925,0.000142999,0.000009373009,0.000007835277,0.0001354296],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.8332464,"threshold_uncertainty_score":0.9950415,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"design_other","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2846027767","doi":"10.1136/bmjebm-2018-110922","title":"Shared decision is the only outcome that matters when it comes to evaluating evidence-based practice","year":2018,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":60,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Population; Argument (complex analysis); Psychological intervention; Medicine; Health care; Incidence (geometry); Psychology; Nursing; Political science","authors":[{"name":"James McCormack","is_ca":true},{"name":"Glyn Elwyn","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.7238961987099377,"gpt":0.5691387085368007,"spread":0.154757490173137,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.090679,0.0005828833,0.001827252,0.000835729,0.0007577181,0.0002856905,0.001521968,0.0002545136,0.007111953],"category_scores_gemma":[0.1466001,0.0004917555,0.0002855518,0.0008185568,0.0005085968,0.00221401,0.0001860739,0.0004870803,0.008651549],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001417463,"about_ca_system_score_gemma":0.001239736,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003973341,"about_ca_topic_score_gemma":0.0005198231,"domain_scores_codex":[0.9885206,0.001392908,0.006489664,0.001485459,0.001163092,0.0009483358],"domain_scores_gemma":[0.951055,0.04023491,0.004650254,0.002478784,0.0009076117,0.0006734414],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0007681185,0.0001017519,0.1353457,0.0006452869,0.0001110175,0.00001008381,0.005650412,0.0004150023,0.00003343544,0.001911404,0.8515965,0.003411225],"study_design_scores_gemma":[0.01014141,0.00661206,0.1826064,0.05444084,0.0005763719,0.00009680137,0.01658784,0.09211999,0.0003331728,0.01416867,0.6190795,0.003236844],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.02952231,0.005555323,0.05282738,0.9051926,0.002170251,0.003857136,0.00008643252,0.0001251412,0.0006634613],"genre_scores_gemma":[0.2469489,0.0001413833,0.04486394,0.7042355,0.002259431,0.0005457837,0.00002211714,0.00009148074,0.0008914387],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.232517,"threshold_uncertainty_score":0.9997534,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"theoretical_or_conceptual","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4400879434","doi":"10.1136/bmjebm-2024-112899","title":"Rapid review method series: interim guidance for the reporting of rapid reviews","year":2024,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":55,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Public Health Ontario; University of Toronto; St. Michael's Hospital; University of Alberta; Public Health Agency of Canada; Bruyère; University of Ottawa","funders":"Canadian Institutes of Health Research; Canada Research Chairs","keywords":"Interim; Systematic review; Set (abstract data type); Computer science; Psychology; MEDLINE; Political science","authors":[{"name":"Adrienne Stevens","is_ca":true},{"name":"Mona Hersi","is_ca":true},{"name":"Chantelle Garritty","is_ca":true},{"name":"Lisa Hartling","is_ca":true},{"name":"Beverley Shea","is_ca":true},{"name":"Lesley Stewart","is_ca":false},{"name":"Vivian Welch","is_ca":true},{"name":"Andrea C. Tricco","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9004752747875066,"gpt":0.647176774034603,"spread":0.2532985007529035,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.6212929,0.0005126928,0.008255812,0.0003724203,0.0001626557,0.0002288833,0.002523698,0.00008886765,0.01805723],"category_scores_gemma":[0.7240558,0.0001770049,0.003524217,0.003375835,0.0002885413,0.0004984011,0.0001321225,0.0002813146,0.0004013348],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00006049069,"about_ca_system_score_gemma":0.0004617695,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0000241509,"about_ca_topic_score_gemma":0.000009566395,"domain_scores_codex":[0.9181783,0.01889427,0.05374631,0.001955569,0.006711117,0.0005144303],"domain_scores_gemma":[0.7314931,0.1577007,0.08257787,0.01996645,0.007655666,0.0006062778],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00001939963,0.000008096014,0.00006162098,0.01072564,0.0001251199,0.00001076435,0.0001664192,0.000007610145,0.0003821898,0.0005594808,0.5713201,0.4166135],"study_design_scores_gemma":[0.0001166431,0.0003165945,0.00008848333,0.09394831,0.001492092,0.0000775343,0.0002625198,0.005205038,0.0002782673,0.001005659,0.8970376,0.0001712299],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.0000123077,0.5960172,0.3266341,0.07120948,0.0009676515,0.00426934,0.000006935545,0.00001305338,0.0008699562],"genre_scores_gemma":[0.006081666,0.6034282,0.3036647,0.03809823,0.004126733,0.006502778,0.0000360153,0.0001355678,0.03792608],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.4164423,"threshold_uncertainty_score":0.9828404,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","research_integrity"],"domain":"reporting","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2925846695","doi":"10.1136/bmjebm-2019-111174","title":"Assessing assumptions for statistical analyses in randomised clinical trials","year":2019,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":53,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"Novo Nordisk Fonden","keywords":"Statistical power; Statistical hypothesis testing; Clinical trial; Statistical analysis; Computer science; Statistical model; Econometrics; Medical physics; Management science; Data science; Medicine; Statistics; Machine learning; Mathematics; Engineering","authors":[{"name":"Emil Eik Nielsen","is_ca":false},{"name":"Anders Kehlet Nørskov","is_ca":false},{"name":"Theis Lange","is_ca":false},{"name":"Lehana Thabane","is_ca":true},{"name":"Jørn Wetterslev","is_ca":false},{"name":"Jan Beyersmann","is_ca":false},{"name":"Jacobo de Uña‐Álvarez","is_ca":false},{"name":"Valter Torri","is_ca":false},{"name":"Laurent Billot","is_ca":false},{"name":"Hein Putter","is_ca":false},{"name":"Per Winkel","is_ca":false},{"name":"Christian Gluud","is_ca":false},{"name":"Janus Christian Jakobsen","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9827549661353883,"gpt":0.784218247525338,"spread":0.1985367186100503,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.6514105,0.000372738,0.01069666,0.0009201188,0.000097782,0.0005706621,0.001273046,0.0001854335,0.03049665],"category_scores_gemma":[0.8360813,0.0001593237,0.00260415,0.001690995,0.0002273395,0.0005737398,0.0000488186,0.000302672,0.001267043],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00006789946,"about_ca_system_score_gemma":0.0006219662,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00005494031,"about_ca_topic_score_gemma":0.00004675775,"domain_scores_codex":[0.8520341,0.09455587,0.04235888,0.002064406,0.008381313,0.0006054457],"domain_scores_gemma":[0.4141476,0.5654708,0.01300244,0.004657028,0.002177734,0.0005444282],"domain_codex":null,"domain_gemma":"methods","domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"observational","study_design_scores_codex":[0.002134494,0.0002735889,0.1237971,0.0004479366,0.0005174236,0.00003309165,0.0001070023,0.00130782,0.001396618,0.00159627,0.7974027,0.07098584],"study_design_scores_gemma":[0.08785263,0.001768971,0.3532095,0.0095987,0.007833742,0.0000219783,0.003672826,0.339555,0.000139127,0.02301883,0.1719331,0.001395552],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.07567447,0.003354019,0.8854598,0.02059638,0.001981914,0.0117344,0.000338734,0.00001682122,0.0008434198],"genre_scores_gemma":[0.9425531,0.0001409314,0.04856948,0.003853515,0.001492465,0.0009222199,0.0003969778,0.00002918197,0.002042103],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8668786,"threshold_uncertainty_score":0.9995106,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W3136262608","doi":"10.1136/bmjebm-2020-111542","title":"Understanding of research results, evidence summaries and their applicability—not critical appraisal—are core skills of medical curriculum","year":2021,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":51,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"Terveyden Tutkimuksen Toimikunta; Sigrid Juséliuksen Säätiö; Helsingin ja Uudenmaan Sairaanhoitopiiri","keywords":"Critical appraisal; Rigour; Evidence-based medicine; Observational study; Systematic review; Medical education; Psychology; Medicine; Critical reading; Evidence-based practice; Psychological intervention; Medical literature; Curriculum; Reading (process); MEDLINE; Alternative medicine; Nursing; Pedagogy; Epistemology; Pathology","authors":[{"name":"Kari A.O. Tikkinen","is_ca":false},{"name":"Gordon Guyatt","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.8221500894503884,"gpt":0.6922985889787482,"spread":0.1298515004716402,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.07394298,0.0005499406,0.003780675,0.0009332692,0.001066551,0.00002774663,0.001513867,0.001239412,0.0009825596],"category_scores_gemma":[0.5269768,0.0003388712,0.0002744907,0.0038867,0.007654679,0.0003541876,0.0007544247,0.004148264,0.00002323931],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001055662,"about_ca_system_score_gemma":0.02870036,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00249632,"about_ca_topic_score_gemma":0.0004481111,"domain_scores_codex":[0.9768684,0.009646107,0.00438275,0.001436877,0.005947994,0.001717904],"domain_scores_gemma":[0.6415027,0.3475195,0.002026903,0.00229638,0.004182757,0.002471743],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","study_design_scores_codex":[0.0004158938,0.0006127253,0.00499666,0.5607635,0.0001014972,0.0001054019,0.005441548,7.242642e-7,0.000005485173,0.01114153,0.2673598,0.1490552],"study_design_scores_gemma":[0.0005643659,0.0007082992,0.0002905758,0.7038736,0.0001355456,0.00002052604,0.02456335,0.0001544014,0.000003170524,0.0005228969,0.268882,0.0002813341],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.0001708024,0.9372897,0.0009864394,0.05539135,0.001681621,0.004144358,0.0001684565,0.0000429409,0.000124374],"genre_scores_gemma":[0.004303672,0.9915509,0.0003211043,0.0003476492,0.00212771,0.001098503,0.000113698,0.00004427284,0.0000925021],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.4530338,"threshold_uncertainty_score":0.9999307,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4410343249","doi":"10.1136/bmjebm-2025-113825","title":"Reporting guideline for the use of Generative Artificial intelligence tools in MEdical Research: the GAMER Statement","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Artificial Intelligence in Healthcare and Education","field":"Medicine","cited_by":49,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University","funders":"Chinese Academy of Medical Sciences; Lanzhou University","keywords":"Declaration; Delphi method; Checklist; Transparency (behavior); Delphi; Computer science; Guideline; Artificial intelligence; Medical education; Psychology; Data science; Medicine; Computer security","authors":[{"name":"Xufei Luo","is_ca":false},{"name":"Yih Chung Tham","is_ca":false},{"name":"Mauro Giuffrè","is_ca":false},{"name":"Robert Ranisch","is_ca":false},{"name":"Mohammad Daher","is_ca":false},{"name":"Kyle Lam","is_ca":false},{"name":"Alexander Viktor Eriksen","is_ca":false},{"name":"Che-Wei Hsu","is_ca":false},{"name":"Akihiko Ozaki","is_ca":false},{"name":"Fábio Ynoe de Moraes","is_ca":true},{"name":"Sahil Khanna","is_ca":false},{"name":"Kuan‐Pin Su","is_ca":false},{"name":"Emir Begagić","is_ca":false},{"name":"Zhaoxiang Bian","is_ca":false},{"name":"Yaolong Chen","is_ca":false},{"name":"Janne Estill","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.8920578848643578,"gpt":0.6593057290771102,"spread":0.2327521557872476,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.03267658,0.0001886593,0.0005692339,0.0004007065,0.0003086395,0.00003636916,0.0003513422,0.0001562204,0.0004641403],"category_scores_gemma":[0.2202984,0.00009697883,0.0001208694,0.001528944,0.001116831,0.0001806717,0.00007598887,0.0007014473,0.000007193173],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003080686,"about_ca_system_score_gemma":0.004010554,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.005603437,"about_ca_topic_score_gemma":0.001736134,"domain_scores_codex":[0.9914873,0.0006849143,0.005233718,0.0004609276,0.001592694,0.000540486],"domain_scores_gemma":[0.9611146,0.03272915,0.001488466,0.001176442,0.003263374,0.000228007],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.001797611,0.0002675239,0.005676921,0.0005231568,0.00007583675,0.00005121785,0.003036028,0.002491238,0.001563789,0.008249762,0.09826717,0.8779997],"study_design_scores_gemma":[0.0005706488,0.004319024,0.006666331,0.04064702,0.0006026442,0.00005783594,0.0560716,0.4242367,0.1836656,0.03267629,0.2499284,0.0005579235],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.04155162,0.003629153,0.2471075,0.7017167,0.00111461,0.004775219,0.000005019756,0.00002450985,0.00007568073],"genre_scores_gemma":[0.9635655,0.001974567,0.006678875,0.02367564,0.00201953,0.001515029,0.00004824861,0.00002607194,0.0004965357],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.9220139,"threshold_uncertainty_score":0.9960631,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4366386724","doi":"10.1136/bmjebm-2022-112070","title":"Rapid Reviews Methods Series: Involving patient and public partners, healthcare providers and policymakers as knowledge users","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Mental Health and Patient Involvement","field":"Health Professions","cited_by":48,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Canadian Agency for Drugs and Technologies in Health; Public Health Ontario; University of Toronto; St. Michael's Hospital; Cochrane; Public Health Agency of Canada; University of Ottawa","funders":"","keywords":"Timeline; Health care; Commission; Knowledge translation; Psychology; Knowledge management; Public relations; Management science; Computer science; Business; Political science; Engineering; Mathematics","authors":[{"name":"Chantelle Garritty","is_ca":true},{"name":"Andrea C. Tricco","is_ca":true},{"name":"Maureen Smith","is_ca":true},{"name":"Danielle Pollock","is_ca":false},{"name":"Chris Kamel","is_ca":true},{"name":"Valerie King","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.6031864510411815,"gpt":0.5799916421850461,"spread":0.02319480885613545,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.009362301,0.0004527502,0.0009664492,0.0005871428,0.001303856,0.00002122897,0.0002007456,0.000259814,0.0003374251],"category_scores_gemma":[0.005713233,0.000336601,0.00007437514,0.001264432,0.0004517699,0.0005382662,0.0002350314,0.0007167681,0.0001885769],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004854527,"about_ca_system_score_gemma":0.001840606,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00330756,"about_ca_topic_score_gemma":0.001184407,"domain_scores_codex":[0.9904715,0.004923746,0.001851826,0.000790769,0.0005938545,0.001368308],"domain_scores_gemma":[0.9926904,0.003935236,0.0007924737,0.0006636499,0.0003240466,0.001594149],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0009524396,0.0001360154,0.1463707,0.05403343,0.00006833413,0.00004701844,0.05592483,0.000001567207,0.0003208958,0.004351052,0.247005,0.4907887],"study_design_scores_gemma":[0.002848977,0.004454576,0.004048455,0.04386346,0.00009513939,0.000009804124,0.05980671,0.0004965396,0.00007652136,0.00098081,0.8827592,0.0005598775],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"review","genre_scores_codex":[0.258769,0.3104999,0.0003951119,0.3945131,0.005428823,0.02710076,0.00006131089,0.0008907566,0.002341158],"genre_scores_gemma":[0.2743151,0.4042268,0.02263191,0.2698857,0.004321454,0.0167245,0.001344849,0.0003951873,0.006154581],"genre_candidate":"review","genre_consensus":null,"teacher_disagreement_score":0.6357541,"threshold_uncertainty_score":0.9999963,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"design_other","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2855200104","doi":"10.1136/bmjebm-2018-110972","title":"‘Depletion of the susceptibles’ taught through a story, a table and basic arithmetic","year":2018,"lang":"en","type":"letter","venue":"BMJ evidence-based medicine","topic":"Advanced Causal Inference Techniques","field":"Mathematics","cited_by":42,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McGill University","funders":"","keywords":"Phrase; Arithmetic; Table (database); Psychology; Statistics; Computer science; Natural language processing; Mathematics; Data mining","authors":[{"name":"Steven D. Stovitz","is_ca":false},{"name":"Hailey R. Banack","is_ca":false},{"name":"Jay S. Kaufman","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2259520065183394,"gpt":0.4243785418092606,"spread":0.1984265352909212,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00224555,0.0006224758,0.001250214,0.0002518247,0.0001756752,0.00002230602,0.0007671597,0.0007838411,0.0005944971],"category_scores_gemma":[0.01068677,0.0003923071,0.0001568517,0.0006573657,0.001913483,0.0003090438,0.0001882994,0.00182906,0.00001211769],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002719123,"about_ca_system_score_gemma":0.0005506938,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002332983,"about_ca_topic_score_gemma":0.00005771108,"domain_scores_codex":[0.9956594,0.0006976739,0.0009647865,0.0007135736,0.001398405,0.0005661416],"domain_scores_gemma":[0.9914182,0.004784068,0.001329363,0.001740366,0.0006493857,0.00007857469],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00004552162,0.00003022051,0.0001951014,0.002440853,0.00004402883,0.00005014249,0.0002702944,6.920823e-7,0.001306378,0.001294061,0.9938762,0.0004464611],"study_design_scores_gemma":[0.001476984,0.002100405,0.0001800587,0.05714783,0.001026497,0.0001366508,0.0002232203,0.0001232521,0.01144751,0.2506336,0.6744613,0.001042731],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.001491682,0.003635163,0.06693821,0.9218978,0.0008883777,0.003689918,0.00006057312,0.0003867279,0.001011554],"genre_scores_gemma":[0.01341116,0.002276125,0.1433267,0.809821,0.01770033,0.001696449,0.0001552734,0.0005545425,0.01105846],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.319415,"threshold_uncertainty_score":0.9998529,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W3028006966","doi":"10.1136/bmjebm-2020-111386","title":"Thrombolysis with alteplase 3–4.5 hours after acute ischaemic stroke: trial reanalysis adjusted for baseline imbalances","year":2020,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Acute Ischemic Stroke Management","field":"Medicine","cited_by":41,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; St. Joseph’s Healthcare Hamilton; Impact","funders":"","keywords":"Medicine; Thrombolysis; Stroke (engine); Modified Rankin Scale; Ischaemic stroke; Clinical trial; Placebo; Internal medicine; Physical therapy; Randomized controlled trial; Tissue plasminogen activator; Ischemic stroke; Ischemia; Myocardial infarction","authors":[{"name":"Brian S. Alper","is_ca":false},{"name":"Gary Foster","is_ca":true},{"name":"Lehana Thabane","is_ca":true},{"name":"Alex Rae-Grant","is_ca":false},{"name":"Meghan Malone-Moses","is_ca":false},{"name":"Eric Manheimer","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.06341808429054929,"gpt":0.3372272680659038,"spread":0.2738091837753545,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00170931,0.0007210145,0.001963613,0.0004357458,0.0001182858,0.00003132044,0.0004434291,0.0002034978,0.001395633],"category_scores_gemma":[0.00379376,0.000482986,0.0004926598,0.001315096,0.0004754497,0.0003194563,0.0001038232,0.0004788383,0.00005300624],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002220967,"about_ca_system_score_gemma":0.0005975784,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001291727,"about_ca_topic_score_gemma":0.00003554003,"domain_scores_codex":[0.9945742,0.000208488,0.001295673,0.001284295,0.001825213,0.0008121161],"domain_scores_gemma":[0.9952592,0.001379615,0.0006732777,0.001203688,0.0006052031,0.0008790572],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"randomized_trial","study_design_scores_codex":[0.4644588,0.0005015442,0.02318093,0.001756624,0.00633928,0.001953696,0.000753071,0.0003550365,0.01611804,0.00001853504,0.4743387,0.01022572],"study_design_scores_gemma":[0.5757156,0.06431678,0.01045601,0.02357329,0.1086551,0.0002437822,0.005985648,0.06112333,0.05237218,0.00002239158,0.0937123,0.003823573],"study_design_candidate":"randomized_trial","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.4838958,0.006351255,0.1616375,0.3323441,0.0008662692,0.01275979,0.0003877984,0.0007093294,0.001048106],"genre_scores_gemma":[0.9359354,0.0008814654,0.02016404,0.03276777,0.005615483,0.002276519,0.0005951154,0.0001572986,0.001606913],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.4520396,"threshold_uncertainty_score":0.9997622,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2911416111","doi":"10.1136/bmjebm-2018-111069","title":"Causal inference for clinicians","year":2019,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Advanced Causal Inference Techniques","field":"Mathematics","cited_by":34,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Jewish General Hospital","funders":"","keywords":"Inference; Causal inference; Psychology; Computer science; Artificial intelligence; Econometrics; Mathematics","authors":[{"name":"Steven D. Stovitz","is_ca":false},{"name":"Ian Shrier","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3831093233932655,"gpt":0.5368590741435255,"spread":0.15374975075026,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002501458,0.0003419304,0.0007692828,0.0002030403,0.0000736826,0.00001708696,0.0004804468,0.0002065427,0.0009223947],"category_scores_gemma":[0.0221725,0.000261971,0.0001306722,0.0003309873,0.0002545741,0.000381179,0.00006178846,0.0003216556,0.0001259651],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001473677,"about_ca_system_score_gemma":0.0004075431,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00005015052,"about_ca_topic_score_gemma":0.00003784749,"domain_scores_codex":[0.99726,0.0001358007,0.0009326602,0.000532308,0.0006118934,0.0005272892],"domain_scores_gemma":[0.9847475,0.01307064,0.0004734812,0.001016879,0.0005010071,0.0001904474],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.001739654,0.0005134828,0.03487584,0.004387754,0.0001602306,0.000081536,0.001117931,0.0001783303,0.04534408,0.7144293,0.1476591,0.04951271],"study_design_scores_gemma":[0.004288103,0.006458261,0.001550401,0.01726535,0.0003325635,0.00002093243,0.0006101922,0.004460077,0.03140363,0.9084217,0.02393658,0.001252198],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.06182809,0.0006471072,0.907122,0.01876421,0.001098521,0.006100208,0.00002556312,0.001178003,0.003236284],"genre_scores_gemma":[0.8122418,0.0001771011,0.1784278,0.004634079,0.0009998265,0.000909318,0.00002463204,0.00009867409,0.002486749],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.7504137,"threshold_uncertainty_score":0.9999909,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W3098080120","doi":"10.1136/bmjebm-2020-111503","title":"Ten years later: a review of the US 2009 institute of medicine report on conflicts of interest and solutions for further reform","year":2020,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Pharmaceutical industry and healthcare","field":"Pharmacology, Toxicology and Pharmaceutics","cited_by":33,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Toronto; St. Michael's Hospital","funders":"","keywords":"Conflict of interest; Guideline; Alternative medicine; Political science; Health care; Medical research; Public relations; Medical education; Engineering ethics; Medicine; Engineering; Law","authors":[{"name":"Trevor Torgerson","is_ca":false},{"name":"Cole Wayant","is_ca":false},{"name":"Lisa Cosgrove","is_ca":false},{"name":"Elie A. Akl","is_ca":false},{"name":"Jake X. Checketts","is_ca":false},{"name":"Rafael Dal‐Ré","is_ca":false},{"name":"Jennifer Gill","is_ca":false},{"name":"Samir C. Grover","is_ca":true},{"name":"Nasim A. Khan","is_ca":false},{"name":"Rishad Khan","is_ca":true},{"name":"Ana Marušić","is_ca":false},{"name":"Matthew S. McCoy","is_ca":false},{"name":"Aaron Philip Mitchell","is_ca":false},{"name":"Vinay Prasad","is_ca":false},{"name":"Matt Vassar","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.7738178979150093,"gpt":0.6174324461110864,"spread":0.1563854518039229,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.004985844,0.0004413076,0.002882897,0.000186711,0.00009693319,0.000001492131,0.0005755045,0.0007329551,0.0005408337],"category_scores_gemma":[0.005762634,0.0002512517,0.0003658076,0.0006216591,0.002210347,0.00007586263,0.0001486245,0.001691395,0.000004559824],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001586267,"about_ca_system_score_gemma":0.001187228,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001309071,"about_ca_topic_score_gemma":0.00002635876,"domain_scores_codex":[0.9953872,0.0008685912,0.002461985,0.0004707728,0.0003898845,0.0004215945],"domain_scores_gemma":[0.9945471,0.001710738,0.002215504,0.0006390502,0.0004618746,0.0004256877],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0006675711,0.0002030434,0.0001314718,0.3786568,0.0007300445,0.000197521,0.0002556672,0.00000376494,0.0001110453,0.002170018,0.03966544,0.5772076],"study_design_scores_gemma":[0.00100796,0.0007497311,0.00005503774,0.2877445,0.002853624,0.0001213375,0.00001764541,0.00002276207,0.00008423722,0.0000545719,0.7071473,0.0001412601],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00005984295,0.93055,0.00001394017,0.06342687,0.001138874,0.003924965,0.0002932059,0.00001909156,0.0005731732],"genre_scores_gemma":[0.001142657,0.9873464,0.00003284589,0.01000314,0.0007883641,0.0002698097,0.0001006969,0.00003320749,0.0002829444],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.6674818,"threshold_uncertainty_score":0.999994,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","research_integrity"],"domain":"incentives","study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","research_integrity"],"domain":"methods","study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2981608944","doi":"10.1136/bmjebm-2019-111273","title":"Challenges facing early-career and mid-career researchers: potential solutions to safeguard the future of evidence-based medicine","year":2019,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":32,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Hospital for Sick Children","funders":"National Institute for Health and Care Research","keywords":"Globe; Attendance; Career development; Career Pathways; General partnership; Public relations; Medical education; Psychology; Management; Political science; Medicine","authors":[{"name":"Georgia C. Richards","is_ca":false},{"name":"Stephen H Bradley","is_ca":false},{"name":"Andrew Dagens","is_ca":false},{"name":"Christoffer Bjerre Haase","is_ca":false},{"name":"Brennan C Kahan","is_ca":false},{"name":"Tanja Rombey","is_ca":false},{"name":"Cole Wayant","is_ca":false},{"name":"Logan Z. J. Williams","is_ca":false},{"name":"Peter J. Gill","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.6353307993045753,"gpt":0.534381934536827,"spread":0.1009488647677482,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.02214429,0.0003572167,0.0007873347,0.0007671003,0.001201564,0.00002151003,0.001001187,0.0003324717,0.002619265],"category_scores_gemma":[0.01568683,0.0001918847,0.00009713304,0.00153293,0.001178742,0.0006477709,0.0002420795,0.001408827,0.0002993405],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000453102,"about_ca_system_score_gemma":0.00438723,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.004581213,"about_ca_topic_score_gemma":0.0008243399,"domain_scores_codex":[0.989946,0.002960129,0.001329352,0.0008722679,0.003174365,0.001717904],"domain_scores_gemma":[0.9852505,0.009722828,0.0004639832,0.001352039,0.001870955,0.001339684],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"observational","study_design_scores_codex":[0.006637536,0.0003493175,0.2015357,0.01800258,0.0002199202,0.00009685228,0.07406697,0.001245221,0.03295636,0.003201695,0.4252786,0.2364092],"study_design_scores_gemma":[0.00583033,0.008597729,0.7054395,0.08417825,0.0002142952,0.00001819607,0.113267,0.003447521,0.000472868,0.0009583603,0.07676794,0.0008079614],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.2265831,0.05042278,0.001219675,0.712868,0.00263343,0.005294445,0.00001426395,0.00007064615,0.0008936542],"genre_scores_gemma":[0.9792147,0.00742354,0.0005035803,0.006093552,0.004683883,0.0006423045,0.00001541051,0.00003730603,0.001385779],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.7526315,"threshold_uncertainty_score":0.9982924,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"incentives","study_design":"theoretical_or_conceptual","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"incentives","study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3039340826","doi":"10.1136/bmjebm-2020-111412","title":"Dietary Recommendations for Familial Hypercholesterolaemia: an Evidence-Free Zone","year":2020,"lang":"en","type":"letter","venue":"BMJ evidence-based medicine","topic":"Diet and metabolism studies","field":"Medicine","cited_by":30,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Canadian VIGOUR Centre","funders":"","keywords":"Medicine; Insulin resistance; Saturated fat; Internal medicine; Cholesterol; Obesity; Endocrinology; Coronary heart disease; Low carbohydrate; Cardiology; Weight loss","authors":[{"name":"David M. Diamond","is_ca":false},{"name":"Abdullah Alabdulgader","is_ca":false},{"name":"Michel de Lorgeril","is_ca":false},{"name":"Zoë Harcombe","is_ca":false},{"name":"Malcolm Kendrick","is_ca":false},{"name":"Aseem Malhotra","is_ca":false},{"name":"Blair J. O’Neill","is_ca":true},{"name":"Uffe Ravnskov","is_ca":false},{"name":"Sherif Sultan","is_ca":false},{"name":"Jeff S. Volek","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2491908472061366,"gpt":0.4015866636004412,"spread":0.1523958163943046,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.00184483,0.001238929,0.003056668,0.0006684986,0.0004180135,0.00006459336,0.001178535,0.001058474,0.0005073479],"category_scores_gemma":[0.01740431,0.0009565275,0.0006888177,0.0007509377,0.000886618,0.0008071825,0.0002814446,0.003148615,0.0001099748],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003210192,"about_ca_system_score_gemma":0.00147237,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003953366,"about_ca_topic_score_gemma":0.0000729819,"domain_scores_codex":[0.9926295,0.0005317016,0.001879415,0.001790869,0.001948038,0.00122048],"domain_scores_gemma":[0.9901466,0.004649153,0.0007749545,0.002513046,0.001170905,0.0007453238],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.001793844,0.0001249053,0.000489963,0.003607178,0.0003735473,0.0007781861,0.0006493289,0.000001538111,0.001090749,0.00001008287,0.9626548,0.02842593],"study_design_scores_gemma":[0.005911354,0.005392734,0.002022181,0.02677923,0.004897121,0.00005615503,0.0004679397,0.0001491346,0.0001250045,0.0003320266,0.9529607,0.0009064409],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.0004310044,0.01912129,0.001797652,0.9671151,0.004889767,0.005539749,0.0002813802,0.0004458301,0.0003782379],"genre_scores_gemma":[0.0008476671,0.009163685,0.01174323,0.9118255,0.05697049,0.002133334,0.002365826,0.0002593265,0.004690901],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.05528955,"threshold_uncertainty_score":0.9992885,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4366386744","doi":"10.1136/bmjebm-2022-112111","title":"Rapid reviews methods series: Guidance on assessing the certainty of evidence","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":30,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"Public Health Agency","keywords":"Grading (engineering); Systematic review; Computer science; Rating system; Certainty; Management science; Medicine; Actuarial science; MEDLINE; Engineering; Mathematics; Business; Political science","authors":[{"name":"Gerald Gartlehner","is_ca":false},{"name":"Barbara Nußbaumer-Streit","is_ca":false},{"name":"Declan Devane","is_ca":false},{"name":"Leila C. Kahwati","is_ca":false},{"name":"Meera Viswanathan","is_ca":false},{"name":"Valerie King","is_ca":false},{"name":"Amir Qaseem","is_ca":false},{"name":"Elie A. Akl","is_ca":false},{"name":"Holger J Schuenemann","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9546345879208633,"gpt":0.6949051985894442,"spread":0.2597293893314191,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.5669478,0.0005660722,0.005928846,0.0007524002,0.0003408121,0.0003460063,0.003960373,0.0001704124,0.009986703],"category_scores_gemma":[0.7011788,0.0001853313,0.001808796,0.007838656,0.0006458579,0.0007350892,0.0002056498,0.0005176863,0.002264584],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00009005307,"about_ca_system_score_gemma":0.0004389672,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00005547887,"about_ca_topic_score_gemma":0.00001238749,"domain_scores_codex":[0.881878,0.08236209,0.02071065,0.002047578,0.01226977,0.0007318844],"domain_scores_gemma":[0.7892361,0.1769673,0.01677387,0.01309238,0.003497257,0.0004330639],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00005273949,0.00002156759,0.003209366,0.0007753078,0.00008857375,0.00002065081,0.0006504897,0.000473132,0.00250013,0.0009153329,0.6121005,0.3791922],"study_design_scores_gemma":[0.0003842938,0.0006628396,0.01737933,0.03232053,0.000745789,0.00002351068,0.002297088,0.01545462,0.001448854,0.004735726,0.9241154,0.000432001],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"methods","genre_scores_codex":[0.00785054,0.1790927,0.3232855,0.4683611,0.004029092,0.01049595,0.00001284842,0.0001112488,0.006761008],"genre_scores_gemma":[0.3598139,0.07536746,0.3703447,0.0829935,0.007066392,0.004559278,0.00004340591,0.0002873362,0.09952403],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.3853676,"threshold_uncertainty_score":0.9985123,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W3094203153","doi":"10.1136/bmjebm-2020-111452","title":"Optimising the process for conducting scoping reviews","year":2020,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":28,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Izaak Walton Killam Health Centre; BC Children's Hospital; University of Toronto; Dalhousie University; York University; University of British Columbia","funders":"Canadian Institutes of Health Research","keywords":"Process (computing); Systematic review; Resource (disambiguation); Knowledge management; Process management; Best practice; Key (lock); Computer science; Management science; Data science; Engineering; MEDLINE; Political science","authors":[{"name":"Colleen Pawliuk","is_ca":true},{"name":"Helen Brown","is_ca":true},{"name":"Kim Widger","is_ca":true},{"name":"Tammie Dewan","is_ca":true},{"name":"Anne‐Mette Hermansen","is_ca":true},{"name":"Marie-Claude Grégoire","is_ca":true},{"name":"Rose Steele","is_ca":true},{"name":"Harold Siden","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.895103126102174,"gpt":0.7197301643337385,"spread":0.1753729617684355,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","sts"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.02899186,0.000487123,0.002727547,0.0003192126,0.002640605,0.0000338156,0.001181913,0.0003581536,0.0006569694],"category_scores_gemma":[0.1017964,0.0002434368,0.0003006394,0.002010405,0.0004185202,0.0003438302,0.0001097464,0.00193543,0.0003247589],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000624968,"about_ca_system_score_gemma":0.02373371,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001640911,"about_ca_topic_score_gemma":0.00003315778,"domain_scores_codex":[0.9900265,0.003605834,0.00292172,0.0008479849,0.001111593,0.001486339],"domain_scores_gemma":[0.9722173,0.02297563,0.002349066,0.0008300898,0.0008015365,0.0008263656],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00001973959,0.000008051316,0.000008112787,0.3193465,0.00001120238,0.000002662455,0.00129376,0.000002146368,4.862236e-7,0.00002994522,0.07017747,0.6091],"study_design_scores_gemma":[0.0001622236,0.0001432247,3.804229e-7,0.4348913,0.0001590965,0.000003138091,0.001092314,0.0001210392,2.872054e-7,0.00002432111,0.5632669,0.0001357563],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.000001111387,0.9200343,0.002996006,0.04882947,0.002630273,0.02513527,0.000008587311,0.00007246426,0.0002924563],"genre_scores_gemma":[0.000002563164,0.9695638,0.001170231,0.007254298,0.009902827,0.01154537,0.0000730083,0.00005790238,0.0004300384],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.6089642,"threshold_uncertainty_score":0.9998572,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2993142824","doi":"10.1136/bmjebm-2019-111300","title":"The need for clinical judgement in the application of evidence-based medicine","year":2019,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":26,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Robarts Clinical Trials; Western University","funders":"","keywords":"Judgement; Evidence-based medicine; Medicine; Intensive care medicine; MEDLINE; Alternative medicine; Psychotherapist; Psychology; Epistemology; Pathology","authors":[{"name":"J. David Spence","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9715138239990201,"gpt":0.7161295285443602,"spread":0.2553842954546599,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","open_science","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.553395,0.001159156,0.01808487,0.001177975,0.0002792509,0.0002374406,0.00953487,0.0005259661,0.001395011],"category_scores_gemma":[0.4358664,0.0003213119,0.005459218,0.005338728,0.001253553,0.0002861917,0.0001763135,0.0009120569,0.0007813229],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002167853,"about_ca_system_score_gemma":0.002165484,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001252336,"about_ca_topic_score_gemma":0.00009015753,"domain_scores_codex":[0.8659865,0.05989425,0.05094269,0.003321383,0.01887845,0.0009767798],"domain_scores_gemma":[0.5949908,0.3576071,0.02992976,0.01363928,0.003528343,0.0003047538],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00008126811,0.00007264176,0.0005150688,0.01104885,0.0002097771,0.000003106947,0.00008087335,0.0000268673,5.47101e-7,0.001137133,0.2112202,0.7756037],"study_design_scores_gemma":[0.0008746986,0.0008419699,0.0002240688,0.0852736,0.005338903,0.000005859733,0.0004278037,0.003509786,3.067412e-7,0.0007357417,0.9024619,0.0003053897],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00000388949,0.8704239,0.05103543,0.05753536,0.001300876,0.01943649,0.00001753179,0.000008619724,0.0002379057],"genre_scores_gemma":[0.0002950188,0.9858006,0.0009502739,0.004218432,0.002046082,0.004878618,0.00007555239,0.00005402694,0.001681437],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.7752983,"threshold_uncertainty_score":0.9999967,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4306181334","doi":"10.1136/bmjebm-2022-112065","title":"The Pandora’s Box of Evidence Synthesis and the case for a living Evidence Synthesis Taxonomy","year":2022,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Healthcare cost, quality, practices","field":"Health Professions","cited_by":24,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Public Health Ontario; St. Michael's Hospital; University of Toronto; McMaster University; Queen's University; Impact","funders":"National Health and Medical Research Council; University of Adelaide","keywords":"Context (archaeology); Pandemic; Nursing; Scope (computer science); Public relations; Psychology; Medical education; Business; Medicine; Coronavirus disease 2019 (COVID-19); Political science; Geography; Computer science","authors":[{"name":"Zachary Munn","is_ca":false},{"name":"Danielle Pollock","is_ca":false},{"name":"Timothy Hugh Barker","is_ca":false},{"name":"Jennifer Stone","is_ca":false},{"name":"Cindy Stern","is_ca":false},{"name":"Edoardo Aromataris","is_ca":false},{"name":"Holger J. Schünemann","is_ca":true},{"name":"Barbara Clyne","is_ca":false},{"name":"Hanan Khalil","is_ca":false},{"name":"Reem A. Mustafa","is_ca":false},{"name":"Christina Godfrey","is_ca":true},{"name":"Andrew Booth","is_ca":false},{"name":"Andrea C. Tricco","is_ca":true},{"name":"Alan Pearson","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.7695973013225446,"gpt":0.5618726478439796,"spread":0.2077246534785649,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","sts","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.08895037,0.000451666,0.001244975,0.0002521058,0.005522552,0.00004450409,0.001224693,0.000182359,0.001028091],"category_scores_gemma":[0.5281563,0.0002359309,0.0002301779,0.0008683498,0.001749158,0.0008889062,0.0007103605,0.001494992,0.00001814769],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0006320087,"about_ca_system_score_gemma":0.003719024,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01488642,"about_ca_topic_score_gemma":0.002445777,"domain_scores_codex":[0.9701009,0.02390796,0.002525732,0.0008852295,0.001479157,0.001101035],"domain_scores_gemma":[0.4435838,0.5511241,0.002320067,0.001669548,0.0009244581,0.0003780671],"domain_codex":null,"domain_gemma":"methods","domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"systematic_review","study_design_scores_codex":[0.04064043,0.0006093096,0.1640456,0.0418591,0.00106093,0.001641681,0.0379085,0.0002609977,0.001410557,0.01555755,0.1904767,0.5045286],"study_design_scores_gemma":[0.007524949,0.005039957,0.02187499,0.3283021,0.007608755,0.001765634,0.3118049,0.01590508,0.001289133,0.00269304,0.2934649,0.00272655],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.09481427,0.104466,0.005204379,0.7740754,0.002053749,0.01891594,0.00009224699,0.0001951976,0.0001828316],"genre_scores_gemma":[0.9444512,0.009312398,0.001110325,0.0113905,0.0009802452,0.03221746,9.415832e-7,0.00006326784,0.0004736346],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.849637,"threshold_uncertainty_score":0.9998851,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2887866346","doi":"10.1136/bmjebm-2018-111040","title":"More than 2 billion pairs of eyeballs: Why aren’t you sharing medical knowledge on Wikipedia?","year":2018,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Wikis in Education and Collaboration","field":"Social Sciences","cited_by":22,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University","funders":"","keywords":"Reading (process); Knowledge translation; Internet privacy; World Wide Web; Public relations; Knowledge sharing; Encyclopedia; Resource (disambiguation); Information Dissemination; Computer science; Medical education; Political science; Library science; Knowledge management; Medicine; Law","authors":[{"name":"Heather Murray","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.1161487099793547,"gpt":0.4504509774698217,"spread":0.3343022674904669,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004630015,0.0001830207,0.0003254802,0.0003110839,0.0003543466,0.00002860558,0.0005918871,0.0002250205,0.003360987],"category_scores_gemma":[0.01405614,0.0001423936,0.00006670058,0.001382626,0.001578699,0.0003006567,0.00004351611,0.0002077471,0.000108119],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002458562,"about_ca_system_score_gemma":0.001977165,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001644031,"about_ca_topic_score_gemma":0.003296679,"domain_scores_codex":[0.9965275,0.0003873924,0.0005920181,0.0004264929,0.001709188,0.0003574172],"domain_scores_gemma":[0.9968806,0.001008123,0.0003086268,0.0004823304,0.0009118203,0.0004084958],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0009796361,0.001342618,0.117814,0.0005398632,0.00009314148,0.00003157141,0.1220384,0.0000374425,0.001006995,0.06455932,0.5705924,0.1209645],"study_design_scores_gemma":[0.004839323,0.005114344,0.05896708,0.02820513,0.0003156303,0.000005791457,0.09055906,0.004160958,0.006702699,0.005208592,0.7945506,0.001370803],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.4803763,0.003709541,0.003065245,0.4317496,0.009758262,0.002232896,0.000009810873,0.0004067124,0.06869159],"genre_scores_gemma":[0.9864697,0.0004470129,0.000197874,0.004389865,0.0069116,0.00008565325,0.00001477357,0.00001808262,0.001465396],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5060934,"threshold_uncertainty_score":0.9975501,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["scholarly_communication","open_science"],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["scholarly_communication"],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4392797608","doi":"10.1136/bmjebm-2023-112722","title":"Rapid reviews methods series: assessing the appropriateness of conducting a rapid review","year":2024,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":22,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Ottawa Public Health; Public Health Agency of Canada; University of Ottawa","funders":"","keywords":"Scope (computer science); Context (archaeology); Risk analysis (engineering); Resource (disambiguation); Process (computing); Management science; Computer science; Systematic review; Operations research; Medicine; MEDLINE; Political science; Engineering","authors":[{"name":"Chantelle Garritty","is_ca":true},{"name":"Barbara Nußbaumer-Streit","is_ca":false},{"name":"Candyce Hamel","is_ca":true},{"name":"Declan Devane","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9588174281060408,"gpt":0.7032138568070663,"spread":0.2556035712989745,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","open_science","insufficient_payload"],"consensus_categories":["metaresearch","metaepi_narrow","metaepi_broad","insufficient_payload"],"category_scores_codex":[0.6782458,0.002037087,0.04469907,0.001365592,0.0003397198,0.0008481729,0.007519968,0.0005075359,0.02235428],"category_scores_gemma":[0.4678687,0.0006429119,0.01209075,0.01182714,0.0009380718,0.0008872623,0.0005743366,0.001664782,0.002120961],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002014159,"about_ca_system_score_gemma":0.002185916,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001732368,"about_ca_topic_score_gemma":0.000002555329,"domain_scores_codex":[0.7213483,0.1963942,0.06369156,0.004234786,0.01332849,0.001002725],"domain_scores_gemma":[0.7686756,0.1203081,0.07870916,0.02561792,0.005898513,0.0007906923],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.000001692046,0.00001289751,6.55416e-7,0.18709,0.0003458699,0.00002052526,0.00007450215,3.700755e-7,0.000002621218,0.0001238792,0.1797728,0.6325542],"study_design_scores_gemma":[0.00006310485,0.00007756165,2.321228e-7,0.4149204,0.0134387,0.000182316,0.0001832487,0.00004236983,0.00000160592,0.0002219222,0.5705234,0.00034517],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[9.843276e-8,0.9468769,0.02711616,0.009356133,0.002685509,0.01271871,0.00002210129,0.00002715064,0.001197175],"genre_scores_gemma":[2.078918e-7,0.9540018,0.03690208,0.002085582,0.001515299,0.002108539,0.00005331128,0.0001102809,0.003222927],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.632209,"threshold_uncertainty_score":0.9996022,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4408555504","doi":"10.1136/bmjebm-2024-112974","title":"Analgesic effects of non-surgical and non-interventional treatments for low back pain: a systematic review and meta-analysis of placebo-controlled randomised trials","year":2025,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Musculoskeletal pain and rehabilitation","field":"Medicine","cited_by":22,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Memorial University of Newfoundland","funders":"","keywords":"Medicine; Placebo; Meta-analysis; Analgesic; Randomized controlled trial; Low back pain; Chronic pain; Physical therapy; Confidence interval; Strictly standardized mean difference; Relative risk; Clinical trial; Anesthesia; Internal medicine; Alternative medicine","authors":[{"name":"Aidan G Cashin","is_ca":false},{"name":"Bradley Furlong","is_ca":true},{"name":"Steven J. Kamper","is_ca":false},{"name":"Diana De Carvalho","is_ca":true},{"name":"Luciana A. C. Machado","is_ca":false},{"name":"Simon Davidson","is_ca":false},{"name":"Krystal Bursey","is_ca":true},{"name":"Christina Abdel Shaheed","is_ca":false},{"name":"Amanda Häll","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.09458008446055835,"gpt":0.4391426482895198,"spread":0.3445625638289614,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaepi_broad"],"category_scores_codex":[0.02833318,0.0007924766,0.03770564,0.001311337,0.00004538558,0.000009184338,0.0001700289,0.0003304277,0.0003926552],"category_scores_gemma":[0.06302699,0.0003806059,0.01010285,0.001233487,0.0003295249,0.0000591378,0.00003080135,0.0001881513,0.000001301921],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001020083,"about_ca_system_score_gemma":0.0005488474,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003225831,"about_ca_topic_score_gemma":0.000003485007,"domain_scores_codex":[0.9856091,0.006649978,0.005923774,0.0007302098,0.0008271882,0.0002598126],"domain_scores_gemma":[0.9092165,0.08532017,0.003889736,0.0007747065,0.0005493782,0.0002494694],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","study_design_scores_codex":[0.004701597,0.0002053033,0.000001619977,0.8040515,0.1879801,0.000008821257,0.00001564911,8.601828e-7,0.000003342533,0.000004844254,0.001102799,0.001923579],"study_design_scores_gemma":[0.04143145,0.001389476,0.000005629897,0.4447382,0.5120553,0.000002836715,0.000006444274,0.0001377935,0.000001211855,0.000003911935,0.0001008081,0.0001269609],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.000007299719,0.9334508,0.002087162,0.001052647,0.00006641797,0.06297216,0.0003415567,0.00000903516,0.00001293756],"genre_scores_gemma":[0.0001378326,0.9859987,0.0003635892,0.0002567147,0.0000781994,0.01095227,0.001207812,0.00002719236,0.0009777002],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.3593133,"threshold_uncertainty_score":0.9998646,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W4384663396","doi":"10.1136/bmjebm-2022-111921","title":"Reducing the residue of retractions in evidence synthesis: ways to minimise inappropriate citation and use of retracted data","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Academic integrity and plagiarism","field":"Social Sciences","cited_by":20,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Regina","funders":"Alfred P. Sloan Foundation","keywords":"Citation; Systematic review; MEDLINE; Computer science; Medical literature; Data science; Medicine; Political science; Library science; Law","authors":[{"name":"Caitlin Bakker","is_ca":true},{"name":"Stephanie Boughton","is_ca":false},{"name":"Clóvis Mariano Faggion","is_ca":false},{"name":"Daniele Fanelli","is_ca":false},{"name":"Kathryn A. Kaiser","is_ca":false},{"name":"Jodi Schneider","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3537610203336972,"gpt":0.4262577316876646,"spread":0.0724967113539674,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01768845,0.0001224214,0.0003404338,0.000419445,0.0002203311,0.00002187372,0.0006310099,0.0003694937,0.00007860528],"category_scores_gemma":[0.2119271,0.00008841907,0.00002738354,0.002273639,0.0006467627,0.0009975825,0.000100399,0.0008216669,0.000007370555],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00008541364,"about_ca_system_score_gemma":0.0006781215,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.02823934,"about_ca_topic_score_gemma":0.003304774,"domain_scores_codex":[0.9966491,0.0009135688,0.0007529228,0.0003832615,0.001017754,0.0002833479],"domain_scores_gemma":[0.9761715,0.02223604,0.0004138748,0.0006709268,0.0003791818,0.0001285061],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"observational","study_design_scores_codex":[0.00619078,0.0003228949,0.0907215,0.002982691,0.0002165735,0.0001962491,0.2365932,0.002952592,0.1236936,0.02283537,0.3811357,0.1321588],"study_design_scores_gemma":[0.001567344,0.001156242,0.7119565,0.1270762,0.0008772072,0.00001594777,0.09289613,0.01843221,0.02252711,0.006734299,0.01573985,0.001020934],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8193047,0.001270653,0.001478944,0.1755483,0.0004769405,0.001500414,0.0000482213,0.00007336167,0.000298393],"genre_scores_gemma":[0.9935324,0.003754583,0.001726786,0.0004545025,0.000320808,0.00005493279,0.00001273832,0.0000100277,0.0001332798],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.621235,"threshold_uncertainty_score":0.9782317,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","research_integrity"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","research_integrity"],"domain":"methods","study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4398251068","doi":"10.1136/bmjebm-2023-112681","title":"Guidelines for the development and validation of patient-reported outcome measures: a scoping review","year":2024,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Cancer survivorship and care","field":"Medicine","cited_by":20,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"","keywords":"Prom; PsycINFO; CLARITY; Guideline; MEDLINE; Scopus; Medicine; Data extraction; Computer science; Medical physics; Pathology","authors":[{"name":"Andrés Jung","is_ca":false},{"name":"Dimitris Challoumas","is_ca":false},{"name":"Larissa Pagels","is_ca":false},{"name":"Susan Armijo‐Olivo","is_ca":true},{"name":"Tobias Braun","is_ca":false},{"name":"Kerstin Luedtke","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.5391998939064311,"gpt":0.5249969430455457,"spread":0.01420295086088541,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.004998679,0.0005595496,0.003208339,0.0003051165,0.00008850591,0.00001635057,0.0002005748,0.0002002816,0.00006845484],"category_scores_gemma":[0.0158627,0.0002769219,0.0005170788,0.0007335463,0.000176206,0.00006522065,0.00006941367,0.0003214736,0.000006076746],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002092991,"about_ca_system_score_gemma":0.003621193,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001177063,"about_ca_topic_score_gemma":0.0001244456,"domain_scores_codex":[0.9940474,0.0001938201,0.003673286,0.0006180631,0.001186094,0.0002813491],"domain_scores_gemma":[0.9939337,0.002169885,0.001376794,0.0008243536,0.001513543,0.0001817478],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"systematic_review","study_design_scores_codex":[0.00003300675,0.000009073226,0.00002543631,0.4279388,0.0002312245,0.00001349714,0.0001225559,2.480565e-7,4.199811e-7,0.000006732953,0.005341461,0.5662775],"study_design_scores_gemma":[0.0002514254,0.0001689933,0.000001296402,0.5639592,0.008385647,0.00005506996,0.0000734424,0.000003345672,0.00001464833,0.000003640488,0.4269304,0.0001529482],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00000451222,0.9769788,0.001513308,0.009339353,0.001177671,0.01088226,0.00001192848,0.000061181,0.0000309931],"genre_scores_gemma":[0.00001411914,0.9924526,0.0008717243,0.002583989,0.0007209881,0.002910578,0.0001521388,0.00007237339,0.0002214416],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.5661246,"threshold_uncertainty_score":0.9999683,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4291021075","doi":"10.1136/bmjebm-2022-111944","title":"Knowledge user survey and Delphi process to inform development of a new risk of bias tool to assess systematic reviews with network meta-analysis (RoB NMA tool)","year":2022,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":19,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"McGill University; Simon Fraser University; Ottawa Hospital; University of Ottawa; University of Toronto; University of British Columbia; Cochrane; St. Michael's Hospital","funders":"Canadian Institutes of Health Research; Medical Research Council; National Institute for Health and Care Research; World Health Organization","keywords":"Delphi method; Knowledge translation; Systematic review; Delphi; Psychology; Process (computing); Quality (philosophy); Knowledge management; Computer science; MEDLINE; Political science; Artificial intelligence","authors":[{"name":"Carole Lunny","is_ca":true},{"name":"Areti Angeliki Veroniki","is_ca":true},{"name":"Brian Hutton","is_ca":true},{"name":"Ian R. White","is_ca":false},{"name":"James M Wright","is_ca":true},{"name":"Ji Yoon Kim","is_ca":true},{"name":"Sai Surabi Thirugnanasampanthar","is_ca":true},{"name":"Shazia Siddiqui","is_ca":true},{"name":"Jennifer Watt","is_ca":true},{"name":"Lorenzo Moja","is_ca":false},{"name":"Nichole Taske","is_ca":false},{"name":"Robert C. Lorenz","is_ca":false},{"name":"Savannah Gerrish","is_ca":true},{"name":"Sharon E. Straus","is_ca":true},{"name":"Virginia Minogue","is_ca":false},{"name":"Franklin Hu","is_ca":true},{"name":"Kevin Lin","is_ca":true},{"name":"Ayah Kapani","is_ca":true},{"name":"Samin Nagi","is_ca":true},{"name":"Lillian Chen","is_ca":true},{"name":"Mona Akbar-nejad","is_ca":true},{"name":"Andrea C. Tricco","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9019285677618601,"gpt":0.5686405846106061,"spread":0.3332879831512541,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.4170017,0.0007063499,0.01745882,0.001634156,0.0003309211,0.000178627,0.002442498,0.00007214879,0.006381604],"category_scores_gemma":[0.2314209,0.0002964036,0.002192884,0.01764683,0.00009410682,0.0003419352,0.0004016459,0.0002540807,0.0001455415],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000126098,"about_ca_system_score_gemma":0.00136371,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002568571,"about_ca_topic_score_gemma":0.003164535,"domain_scores_codex":[0.9250226,0.03330708,0.02692805,0.001812385,0.01228942,0.0006404861],"domain_scores_gemma":[0.9333867,0.03699624,0.01813498,0.005770259,0.005008185,0.0007036795],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"meta_analysis","study_design_scores_codex":[0.001194922,0.000543292,0.3931358,0.06397209,0.1423986,0.00002918534,0.0300168,0.1150796,0.0000438568,0.0004340581,0.2310738,0.02207808],"study_design_scores_gemma":[0.00313678,0.004124143,0.188556,0.02817608,0.6358028,0.00006030134,0.01296723,0.016566,0.0002083226,0.0004978776,0.1066655,0.003238913],"study_design_candidate":"meta_analysis","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.3956903,0.07893616,0.4920147,0.003545968,0.0003716908,0.02901131,0.00009674081,0.00002532856,0.000307838],"genre_scores_gemma":[0.8768491,0.0002309227,0.1054204,0.002384062,0.0001867432,0.005727638,0.00005084358,0.00006381579,0.009086456],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.4934042,"threshold_uncertainty_score":0.9999488,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3027820216","doi":"10.1136/bmjebm-2020-111371","title":"Problem with patient decision aids","year":2020,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Patient-Provider Communication in Healthcare","field":"Health Professions","cited_by":19,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Université Laval","funders":"","keywords":"Decision aids; Certainty; Presentation (obstetrics); Health care; Decision support system; Medicine; Decision quality; Decision analysis; Value (mathematics); Quality (philosophy); Medical emergency; Patient satisfaction; Nursing; Computer science; Alternative medicine; Artificial intelligence; Surgery","authors":[{"name":"Joshua R Zadro","is_ca":false},{"name":"Adrian C. Traeger","is_ca":false},{"name":"Simon Décary","is_ca":true},{"name":"Mary O’Keeffe","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3859947806013224,"gpt":0.470309984312354,"spread":0.08431520371103168,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0007805927,0.0003124363,0.0005946803,0.0001704879,0.0007077033,0.000006938653,0.0006362555,0.0002020757,0.001328626],"category_scores_gemma":[0.006859427,0.0002070662,0.00005790198,0.0009849756,0.0002501572,0.0003184594,0.0002225159,0.001067466,0.0004951021],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000333158,"about_ca_system_score_gemma":0.001762201,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0006745636,"about_ca_topic_score_gemma":0.0001200409,"domain_scores_codex":[0.9945966,0.001409005,0.001464405,0.0005919122,0.00129324,0.0006448656],"domain_scores_gemma":[0.9906505,0.005499652,0.0007872405,0.001288968,0.0009895368,0.0007841325],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"not_applicable","study_design_scores_codex":[0.008716326,0.0002394189,0.4662617,0.006128352,0.00007938386,0.0001341551,0.06394558,0.0009885655,0.0009021353,0.002993146,0.2395852,0.2100261],"study_design_scores_gemma":[0.02017377,0.03000679,0.03109853,0.1315081,0.0003778601,0.00001981155,0.06630117,0.007841729,0.0006853073,0.002732418,0.7074426,0.001811942],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.1779008,0.02431805,0.02882199,0.7401028,0.001731253,0.0185566,0.00004251737,0.001267518,0.007258448],"genre_scores_gemma":[0.9026312,0.0002476676,0.03980089,0.0548984,0.0006767155,0.001552653,0.00004544581,0.00006234642,0.00008472855],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.7247304,"threshold_uncertainty_score":0.9995843,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W4407132017","doi":"10.1136/bmjebm-2023-112389","title":"Rapid reviews methods series: guidance on rapid scoping, mapping and evidence and gap map (‘Big Picture Reviews’)","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Delphi Technique in Research","field":"Social Sciences","cited_by":18,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"St. Michael's Hospital; Ottawa Hospital","funders":"","keywords":"Series (stratigraphy); Data science; Computer science; Information retrieval; Biology","authors":[{"name":"Fiona Campbell","is_ca":false},{"name":"Anthea Sutton","is_ca":false},{"name":"Danielle Pollock","is_ca":false},{"name":"Chantelle Garritty","is_ca":true},{"name":"Andrea C. Tricco","is_ca":true},{"name":"Lena Schmidt","is_ca":false},{"name":"Hanan Khalil","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4793826492030341,"gpt":0.5678402709174526,"spread":0.08845762171441851,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.04251332,0.0004692966,0.001355304,0.0005603981,0.0007662417,0.0001447721,0.0009076035,0.0002960346,0.0005129893],"category_scores_gemma":[0.08226614,0.0003538545,0.000136855,0.001771834,0.002069876,0.0006488378,0.0002421949,0.0007341266,0.00003013176],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003082581,"about_ca_system_score_gemma":0.001006762,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0005267275,"about_ca_topic_score_gemma":0.0002362167,"domain_scores_codex":[0.9897276,0.005831035,0.001434542,0.001068813,0.001137768,0.0008002184],"domain_scores_gemma":[0.9897173,0.007634221,0.0006188886,0.001162811,0.000432323,0.0004344647],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0002045784,0.00003440785,0.001451274,0.01004258,0.00002526044,0.00004058604,0.002576847,0.000001065391,0.002700009,0.003137883,0.1624903,0.8172952],"study_design_scores_gemma":[0.0004146857,0.00048181,0.0006625156,0.2885278,0.00007469059,0.000005942105,0.001002382,0.00003896961,0.000922028,0.002175558,0.7053387,0.0003549462],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.0001822832,0.7794216,0.0339619,0.17462,0.0009805742,0.006910065,0.000001669404,0.0002114553,0.003710368],"genre_scores_gemma":[0.003193615,0.8652278,0.1048901,0.01800357,0.001677704,0.001818426,0.000003270697,0.00004216292,0.005143391],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.8169402,"threshold_uncertainty_score":0.9998913,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4224057184","doi":"10.1136/bmjebm-2021-111866","title":"Which actionable statements qualify as good practice statements In Covid-19 guidelines? A systematic appraisal","year":2022,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Clinical practice guidelines implementation","field":"Medicine","cited_by":17,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Université Laval; Ottawa Hospital; Centre hospitalier universitaire de Québec; McMaster University; Impact; Cochrane; Bruyère; Hôpital de l'Enfant-Jésus; University of Ottawa","funders":"Canadian Institutes of Health Research; World Health Organization","keywords":"Guideline; Global Positioning System; Critical appraisal; Systematic review; Grading (engineering); MEDLINE; Quality (philosophy); Medicine; Computer science; Alternative medicine; Engineering; Political science","authors":[{"name":"Omar Dewidar","is_ca":true},{"name":"Tamara Lotfi","is_ca":true},{"name":"Miranda Langendam","is_ca":false},{"name":"Elena Parmelli","is_ca":false},{"name":"Zuleika Saz Parkinson","is_ca":false},{"name":"Karla Solo","is_ca":true},{"name":"Derek K. Chu","is_ca":true},{"name":"Joseph L. Mathew","is_ca":false},{"name":"Elie A. Akl","is_ca":true},{"name":"Romina Brignardello‐Petersen","is_ca":true},{"name":"Reem A. Mustafa","is_ca":true},{"name":"Lorenzo Moja","is_ca":false},{"name":"Alfonso Iorio","is_ca":true},{"name":"Yuan Chi","is_ca":false},{"name":"Carlos Canelo‐Aybar","is_ca":false},{"name":"Tamara Kredo","is_ca":false},{"name":"Justine Karpusheff","is_ca":false},{"name":"Alexis F. Turgeon","is_ca":true},{"name":"Pablo Alonso‐Coello","is_ca":false},{"name":"Wojtek Wiercioch","is_ca":true},{"name":"Annette Gerritsen","is_ca":false},{"name":"Miloslav Klugar","is_ca":false},{"name":"María Ximena Rojas","is_ca":false},{"name":"Peter Tugwell","is_ca":true},{"name":"Vivian Welch","is_ca":true},{"name":"Kevin Pottie","is_ca":true},{"name":"Zachary Munn","is_ca":false},{"name":"Robby Nieuwlaat","is_ca":true},{"name":"Nathan Ford","is_ca":false},{"name":"Adrienne Stevens","is_ca":true},{"name":"Joanne Khabsa","is_ca":false},{"name":"Zil Nasir","is_ca":true},{"name":"Grigorios I. Leontiadis","is_ca":true},{"name":"Joerg J Meerpohl","is_ca":false},{"name":"Thomas Piggott","is_ca":true},{"name":"Amir Qaseem","is_ca":false},{"name":"Micayla Matthews","is_ca":true},{"name":"Holger J. Schünemann","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.7104620991194237,"gpt":0.7006955760346055,"spread":0.00976652308481818,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch","metaepi_narrow"],"category_scores_codex":[0.04843424,0.001500114,0.007043426,0.001956365,0.0003651251,0.0001324062,0.0009906235,0.0005960188,0.01061162],"category_scores_gemma":[0.6167063,0.001181108,0.0006590033,0.004322107,0.0002096756,0.002812982,0.0004472878,0.002643424,0.0007073942],"about_ca_system_candidate":true,"about_ca_system_consensus":true,"about_ca_system_score_codex":0.005739111,"about_ca_system_score_gemma":0.02052162,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.008832488,"about_ca_topic_score_gemma":0.000627561,"domain_scores_codex":[0.9668465,0.008379072,0.01496117,0.002204092,0.006317417,0.001291723],"domain_scores_gemma":[0.8865637,0.09363968,0.01219836,0.003162552,0.00255366,0.001882033],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","study_design_scores_codex":[0.002759177,0.001759108,0.0001444603,0.7828066,0.001994514,0.002134493,0.0005877278,0.0003290451,0.000003488953,0.0002716536,0.09092382,0.1162859],"study_design_scores_gemma":[0.006491621,0.001941015,0.000003274271,0.2518167,0.009215738,0.0008145485,0.004847692,0.0002633481,3.219398e-7,0.0000567266,0.7238622,0.0006868538],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.000008126289,0.9019367,0.00422244,0.06966313,0.002733854,0.0202664,0.0001310037,0.00026768,0.0007706812],"genre_scores_gemma":[0.00004165208,0.9543876,0.004118689,0.02728484,0.0014957,0.007922586,0.00340235,0.0002711223,0.001075442],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.6329383,"threshold_uncertainty_score":0.9997748,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"evaluation","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W4392583344","doi":"10.1136/bmjebm-2023-112695","title":"CARE-radiology statement explanation and elaboration: reporting guideline for radiological case reports","year":2024,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Delphi Technique in Research","field":"Social Sciences","cited_by":16,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"Fundamental Research Funds for the Central Universities; Lanzhou University","keywords":"Guideline; Radiological weapon; Checklist; Medicine; Delphi method; Multidisciplinary approach; Documentation; Medical physics; Radiology; Psychology; Pathology; Computer science","authors":[{"name":"Mengshu Wang","is_ca":false},{"name":"Xufei Luo","is_ca":false},{"name":"Xiaojuan Xiao","is_ca":false},{"name":"Linlin Zhang","is_ca":false},{"name":"Qi Wang","is_ca":true},{"name":"Shiyu Wang","is_ca":false},{"name":"Ximing Wang","is_ca":false},{"name":"Huadan Xue","is_ca":false},{"name":"Long Jiang Zhang","is_ca":false},{"name":"Yaolong Chen","is_ca":false},{"name":"Junqiang Lei","is_ca":false},{"name":"Tomaž Štupnik","is_ca":false},{"name":"Marco Scarci","is_ca":false},{"name":"Alfonso Fiorelli","is_ca":false},{"name":"Tanel Laisaar","is_ca":false},{"name":"Robert Fruscio","is_ca":false},{"name":"Hussein Elkhayat","is_ca":false},{"name":"Nuria Novoa","is_ca":false},{"name":"Fabio Davoli","is_ca":false},{"name":"Ryuichi Waseda","is_ca":false},{"name":"Janne Estill","is_ca":false},{"name":"Susan L. Norris","is_ca":false},{"name":"D. Riley","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3723114543237489,"gpt":0.5789761313316922,"spread":0.2066646770079433,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0242422,0.0001468444,0.0003435284,0.0002575278,0.0003861485,0.00007446273,0.00009382133,0.0001723943,0.0001974576],"category_scores_gemma":[0.05169165,0.0001112545,0.00005238361,0.0005087298,0.0005556772,0.0002966695,0.00003325521,0.0002025231,0.000001890105],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004129661,"about_ca_system_score_gemma":0.001147109,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001566143,"about_ca_topic_score_gemma":0.0006431842,"domain_scores_codex":[0.9957204,0.0004850577,0.002142678,0.0005783192,0.000690182,0.0003833338],"domain_scores_gemma":[0.9952175,0.002672487,0.0007263519,0.0003085143,0.0008690587,0.0002060728],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0002954032,0.00004612197,0.00561794,0.001192103,0.00008515822,0.1028695,0.01505372,0.0003205719,0.003191427,0.02165816,0.7202884,0.1293815],"study_design_scores_gemma":[0.0009132005,0.002511512,0.000320587,0.004187758,0.0002225861,0.01190859,0.04418162,0.009317409,0.001205145,0.01260216,0.911971,0.0006584729],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.1273246,0.04759983,0.3619925,0.4457431,0.003427709,0.009458978,0.00003022148,0.0012806,0.003142417],"genre_scores_gemma":[0.9749941,0.0008380647,0.01999598,0.0009797234,0.002014622,0.0008513239,0.00008601341,0.00001851829,0.0002216703],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8476694,"threshold_uncertainty_score":0.9562964,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3090996960","doi":"10.1136/bmjebm-2020-111499","title":"Quality of early evidence on the pathogenesis, diagnosis, prognosis and treatment of COVID-19","year":2020,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"COVID-19 Clinical Research Studies","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"University of Ottawa; McMaster University","funders":"","keywords":"Checklist; MEDLINE; Critical appraisal; Medicine; Rigour; Systematic review; Consolidated Standards of Reporting Trials; Declaration; Family medicine; Publication bias; Clinical study design; Data extraction; Clinical trial; Alternative medicine; Psychology; Meta-analysis; Pathology; Political science","authors":[{"name":"Sarah Yang","is_ca":true},{"name":"Allen Li","is_ca":true},{"name":"Ali Eshaghpour","is_ca":true},{"name":"Sofia Ivanisevic","is_ca":true},{"name":"Adrian Salopek","is_ca":true},{"name":"John W. Eikelboom","is_ca":true},{"name":"Mark Crowther","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.5759781474648025,"gpt":0.5831047699438358,"spread":0.007126622479033284,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.007532292,0.000947832,0.006883209,0.0004491408,0.0001653761,0.00001546115,0.0006558876,0.0003913466,0.0003122308],"category_scores_gemma":[0.4140433,0.0004683181,0.001169296,0.001531089,0.002361379,0.00009167012,0.0002557668,0.00060463,0.0000218179],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0009378795,"about_ca_system_score_gemma":0.007517273,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001887209,"about_ca_topic_score_gemma":0.00004390084,"domain_scores_codex":[0.9887285,0.003176233,0.00333794,0.001360137,0.002802623,0.0005945527],"domain_scores_gemma":[0.7915965,0.202785,0.002159286,0.001767313,0.0006748085,0.001016986],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.001744261,0.0007627913,0.02354101,0.1512704,0.0009768867,0.0002596352,0.0004027359,0.000001394813,0.00002653069,0.0001941436,0.003118599,0.8177016],"study_design_scores_gemma":[0.002118758,0.01847371,0.00896362,0.2642102,0.004894983,0.000005676441,0.00008383263,0.000008103561,0.00005865654,0.00005013584,0.7007102,0.0004220518],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.0003344711,0.8836675,0.0001062715,0.106363,0.0001139719,0.009200175,0.0001413567,0.00005784333,0.00001538191],"genre_scores_gemma":[0.005272352,0.9850683,0.0001536224,0.00398845,0.0007058112,0.004635254,0.00002839187,0.00007020654,0.00007761468],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.8172796,"threshold_uncertainty_score":0.9997768,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4376639745","doi":"10.1136/bmjebm-2022-112215","title":"Catalogue of bias: novelty bias","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Computational Drug Discovery Methods","field":"Computer Science","cited_by":16,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Toronto; St. Michael's Hospital","funders":"","keywords":"Novelty; Computer science; Psychology; Social psychology","authors":[{"name":"Yan Luo","is_ca":false},{"name":"Carl Heneghan","is_ca":false},{"name":"Nav Persaud","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3599031462954102,"gpt":0.4300701984081084,"spread":0.07016705211269819,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005793832,0.0002196741,0.0004659754,0.0006041042,0.00007606278,0.00002924485,0.001192197,0.00007005052,0.0000527004],"category_scores_gemma":[0.01145803,0.0001717121,0.0001207294,0.002986955,0.0002709154,0.0005902924,0.0002973562,0.0001706036,0.000156151],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0000788523,"about_ca_system_score_gemma":0.0008001684,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0005438872,"about_ca_topic_score_gemma":0.00003476648,"domain_scores_codex":[0.9965537,0.0005298821,0.0007415015,0.0005397237,0.001268864,0.0003663235],"domain_scores_gemma":[0.9895143,0.008581285,0.0003572307,0.001001945,0.0003599505,0.0001852935],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0003952733,0.0004846053,0.03104617,0.002126522,0.0002244814,0.0008965008,0.005065374,0.255799,0.01301651,0.1457829,0.231599,0.3135637],"study_design_scores_gemma":[0.00284256,0.001198821,0.1783829,0.006978045,0.00008137389,0.00007054281,0.0002559446,0.7564018,0.01739146,0.02523248,0.0103518,0.0008122888],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1007652,0.001008607,0.8470855,0.04773249,0.001681807,0.0007551684,0.00002410082,0.0005031031,0.0004440157],"genre_scores_gemma":[0.9334265,0.0001249948,0.06239153,0.002861409,0.0006511437,0.00009830346,0.00009084016,0.00002867484,0.000326588],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8326613,"threshold_uncertainty_score":0.9968688,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4323542543","doi":"10.1136/bmjebm-2022-112100","title":"Proposed triggers for retiring a living systematic review","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Statistical Methods in Clinical Trials","field":"Mathematics","cited_by":15,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Construct (python library); Computer science; Certainty; Actuarial science; Risk analysis (engineering); Business","authors":[{"name":"M. Hassan Murad","is_ca":false},{"name":"Zhen Wang","is_ca":false},{"name":"Haitao Chu","is_ca":false},{"name":"Lifeng Lin","is_ca":false},{"name":"Ibrahim K El Mikati","is_ca":false},{"name":"Joanne Khabsa","is_ca":false},{"name":"Elie A. Akl","is_ca":true},{"name":"Robby Nieuwlaat","is_ca":true},{"name":"Holger J Schuenemann","is_ca":true},{"name":"Irbaz Bin Riaz","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.8168270424800806,"gpt":0.6492158980082676,"spread":0.167611144471813,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.04519819,0.0003516391,0.002577207,0.0002344243,0.0001367039,0.00002366977,0.0005482389,0.0001609276,0.0003892443],"category_scores_gemma":[0.8755799,0.0002311395,0.0003409877,0.00120145,0.0002724007,0.0001076915,0.00007889716,0.0002872524,0.0001241308],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001427996,"about_ca_system_score_gemma":0.000292315,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000007094849,"about_ca_topic_score_gemma":0.000001583484,"domain_scores_codex":[0.9923989,0.002317822,0.002875955,0.0006157618,0.001194823,0.0005967488],"domain_scores_gemma":[0.676367,0.3209464,0.0009504853,0.00103998,0.0004232047,0.000272897],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","study_design_scores_codex":[0.000130078,0.00005597879,0.00002571996,0.9364212,0.0001105059,0.00006770914,0.0001032614,0.00000235164,0.0002719453,0.0101996,0.05135242,0.001259262],"study_design_scores_gemma":[0.0009989426,0.0006070661,0.0000210552,0.8616086,0.001750952,0.000009526775,0.0001098622,0.001887531,0.0002192663,0.1324111,0.0000876646,0.000288415],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.0007034876,0.03192959,0.811725,0.09688094,0.003684792,0.05197398,0.0000602215,0.002237922,0.0008041331],"genre_scores_gemma":[0.006429647,0.009068269,0.9586734,0.01251451,0.00281206,0.007239824,0.000009766056,0.0002659906,0.002986588],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.8303817,"threshold_uncertainty_score":0.9831694,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4391035776","doi":"10.1136/bmjebm-2023-112530","title":"Rapid review methods series: Guidance on the use of supportive software","year":2024,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":15,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Systematic review; Software; Computer science; Automation; Process (computing); Software engineering; Engineering; MEDLINE","authors":[{"name":"Lisa Affengruber","is_ca":false},{"name":"Barbara Nußbaumer-Streit","is_ca":false},{"name":"Candyce Hamel","is_ca":true},{"name":"Miriam Van der Maten","is_ca":false},{"name":"James Thomas","is_ca":false},{"name":"Chris Mavergames","is_ca":false},{"name":"René Spijker","is_ca":false},{"name":"Gerald Gartlehner","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.9106387882086024,"gpt":0.6251835614097659,"spread":0.2854552267988365,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.2806934,0.0004368053,0.004430105,0.0004070377,0.0001107407,0.0002459383,0.001925625,0.00008310277,0.04736626],"category_scores_gemma":[0.5873651,0.0001461946,0.001677199,0.003719645,0.0004524094,0.0005897171,0.000101809,0.0003120727,0.001293124],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005738764,"about_ca_system_score_gemma":0.0004900721,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002531484,"about_ca_topic_score_gemma":0.000006007078,"domain_scores_codex":[0.9504921,0.02836315,0.01204058,0.001279497,0.007487,0.0003376207],"domain_scores_gemma":[0.8342534,0.1470923,0.006474257,0.008666946,0.00319774,0.0003153326],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00002428002,0.00001520931,0.0003236651,0.002446122,0.0001502472,0.00003233772,0.0001929959,0.00003195032,0.00009784828,0.001876994,0.8139505,0.1808578],"study_design_scores_gemma":[0.00008170334,0.0003851155,0.0005554195,0.05720311,0.0008363036,0.00002309435,0.0001906728,0.001639244,0.0003771662,0.001819488,0.9367147,0.0001739974],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"review","genre_gemma":"methods","genre_scores_codex":[0.0003094034,0.4259602,0.3561747,0.211097,0.001306895,0.004203352,0.00003775429,0.00003865658,0.000871948],"genre_scores_gemma":[0.01677877,0.1301414,0.5938496,0.1526297,0.002765965,0.002875856,0.00006274723,0.0002059192,0.1006901],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.3066716,"threshold_uncertainty_score":0.9994845,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4394722187","doi":"10.1136/bmjebm-2023-112767","title":"Trivalent and quadrivalent seasonal influenza vaccine in adults aged 60 and older: a systematic review and network meta-analysis","year":2024,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Influenza Virus Research Studies","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":true},"ca_institutions":"Public Health Ontario; University of Ottawa; McMaster University; University of Toronto; Public Health Agency of Canada; Ottawa Hospital; Institute for Work & Health; St. Michael's Hospital","funders":"Canadian Institutes of Health Research; Canada Excellence Research Chairs, Government of Canada; Public Health Agency; Health Research Board","keywords":"Medicine; Meta-analysis; Confidence interval; Odds ratio; Influenza vaccine; Placebo; Randomized controlled trial; Systematic review; Internal medicine; MEDLINE; Pediatrics; Vaccination; Immunology; Alternative medicine","authors":[{"name":"Areti Angeliki Veroniki","is_ca":true},{"name":"Sai Surabi Thirugnanasampanthar","is_ca":true},{"name":"Menelaos Konstantinidis","is_ca":true},{"name":"Jasmeen Dourka","is_ca":true},{"name":"Marco Ghassemi","is_ca":true},{"name":"Dipika Neupane","is_ca":true},{"name":"Paul A. Khan","is_ca":true},{"name":"Vera Nincic","is_ca":true},{"name":"Margarita Corry","is_ca":false},{"name":"Reid Robson","is_ca":true},{"name":"Amanda Parker","is_ca":true},{"name":"Charlene Soobiah","is_ca":true},{"name":"Angela Sinilaité","is_ca":true},{"name":"Paméla Doyon-Plourde","is_ca":true},{"name":"Anabel Gil","is_ca":true},{"name":"Winnie Siu","is_ca":true},{"name":"Nasheed Moqueet","is_ca":true},{"name":"Adrienne Stevens","is_ca":true},{"name":"Kelly English","is_ca":true},{"name":"Iván D. Flórez","is_ca":true},{"name":"Juan José Yepes-Núñez","is_ca":false},{"name":"Brian Hutton","is_ca":true},{"name":"Matthew Muller","is_ca":true},{"name":"Lorenzo Moja","is_ca":false},{"name":"Sharon E. Straus","is_ca":true},{"name":"Andrea C. Tricco","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3326713281150513,"gpt":0.4941296525993264,"spread":0.1614583244842751,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.01137451,0.001605309,0.0219813,0.001738857,0.0001724055,0.00007709358,0.0003652012,0.0004296563,0.0003316414],"category_scores_gemma":[0.01209473,0.0008780344,0.001789003,0.003827814,0.0004951076,0.0002002793,0.0004929127,0.001665483,0.00003041861],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004103839,"about_ca_system_score_gemma":0.0007859378,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002291645,"about_ca_topic_score_gemma":0.0001790047,"domain_scores_codex":[0.98823,0.00252578,0.004221283,0.001878955,0.002106863,0.001037043],"domain_scores_gemma":[0.9902164,0.005597427,0.001284557,0.001553348,0.0004629144,0.0008853403],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","study_design_scores_codex":[0.0001883001,0.00005975048,0.0002292141,0.9195483,0.0679452,0.0009950805,0.0001732694,0.000002153395,7.675391e-8,0.00002335405,0.005570854,0.005264497],"study_design_scores_gemma":[0.001336956,0.0003934987,0.0002336784,0.5179011,0.4669777,0.0001828029,0.00003379921,0.00006808707,3.872714e-8,0.00001029344,0.01247557,0.0003865225],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00003045601,0.9755553,0.00002249302,0.006719634,0.00008691592,0.01739388,0.00007308129,0.00009467771,0.00002355817],"genre_scores_gemma":[0.00006442527,0.9880573,0.0002454287,0.005582029,0.0004415931,0.00503791,0.00007272122,0.0001161616,0.0003824695],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.4016472,"threshold_uncertainty_score":0.9996694,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":true,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4392545360","doi":"10.1136/bmjebm-2024-112868","title":"Overly complex methods may impair pragmatic use of core evidence-based medicine principles","year":2024,"lang":"en","type":"editorial","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":13,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Core (optical fiber); Psychology; Epistemology; Medicine; Computer science; Philosophy","authors":[{"name":"Rebecca Kuehn","is_ca":false},{"name":"Ying Wang","is_ca":true},{"name":"Gordon Guyatt","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.6553516264834451,"gpt":0.6381847810439233,"spread":0.01716684543952185,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","sts","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch","research_integrity"],"category_scores_codex":[0.05214505,0.001272848,0.003567502,0.002413937,0.001114655,0.00006311254,0.002152486,0.001990376,0.008138145],"category_scores_gemma":[0.3408563,0.0008408026,0.0003943113,0.003779893,0.00313547,0.001118236,0.0003887015,0.005618487,0.0004873938],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002578416,"about_ca_system_score_gemma":0.0469773,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01674113,"about_ca_topic_score_gemma":0.000969497,"domain_scores_codex":[0.9719639,0.009476703,0.005931622,0.002042124,0.007974083,0.002611498],"domain_scores_gemma":[0.8162705,0.1704461,0.0034747,0.002799177,0.004872151,0.002137375],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.001084006,0.0001210623,0.001762121,0.03992281,0.00008949859,0.00006476216,0.002221439,0.00009519998,0.0007943293,0.0002407933,0.949876,0.003728042],"study_design_scores_gemma":[0.001775356,0.003551978,0.002653259,0.2064351,0.0004970175,0.000002668911,0.001461397,0.00884538,0.00002723277,0.0004236939,0.7737373,0.0005896832],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"editorial","genre_gemma":"editorial","genre_scores_codex":[0.0009452309,0.02230209,0.008877452,0.1370451,0.817543,0.01181015,0.0003894636,0.0004918694,0.0005957061],"genre_scores_gemma":[0.002118349,0.008377424,0.03521528,0.007724916,0.9291053,0.003666135,0.001880101,0.0003074326,0.01160509],"genre_candidate":"editorial","genre_consensus":"editorial","teacher_disagreement_score":0.2887113,"threshold_uncertainty_score":0.9995774,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"editorial","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W3101218852","doi":"10.1136/bmjebm-2020-111577","title":"Blinding practices during acute point-of-care ultrasound research: the BLIND-US meta-research study","year":2020,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Ultrasound in Clinical Applications","field":"Medicine","cited_by":13,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; University of Ottawa","funders":"","keywords":"Blinding; Medicine; Subgroup analysis; Point of care ultrasound; Medical physics; Meta-analysis; Research design; Physical therapy; Clinical trial; Radiology; Ultrasound; Internal medicine; Statistics","authors":[{"name":"Ross Prager","is_ca":true},{"name":"Kay Wu","is_ca":true},{"name":"Roudi Bachar","is_ca":false},{"name":"Rudy R Unni","is_ca":true},{"name":"Joshua Bowdridge","is_ca":true},{"name":"Trevor A. McGrath","is_ca":true},{"name":"Rajiv Thavanathan","is_ca":true},{"name":"Michael Y. Woo","is_ca":true},{"name":"Matthew D. F. McInnes","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.717634259424709,"gpt":0.5996602644583762,"spread":0.1179739949663329,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","research_integrity","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.01973985,0.0004010341,0.001362504,0.0005822474,0.0008851021,0.00008927068,0.001319477,0.0002155865,0.001982408],"category_scores_gemma":[0.1652878,0.0002280737,0.000369047,0.004457091,0.002175352,0.0004593164,0.000312284,0.00315741,0.0002507689],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002456272,"about_ca_system_score_gemma":0.001278241,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0005116296,"about_ca_topic_score_gemma":0.00008839469,"domain_scores_codex":[0.988153,0.00275358,0.001769534,0.001197312,0.00515218,0.0009743707],"domain_scores_gemma":[0.9165762,0.07452274,0.001162189,0.002440504,0.004295807,0.001002547],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.03069509,0.005557392,0.5712839,0.006315762,0.01425827,0.001313239,0.06115841,0.0002234372,0.262733,0.001761222,0.04221661,0.002483668],"study_design_scores_gemma":[0.04658049,0.06672771,0.3735665,0.01300903,0.04284029,0.000636923,0.3232847,0.0005932156,0.08762902,0.001991912,0.04110939,0.002030803],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7751788,0.00559626,0.0001951852,0.2099956,0.0001557406,0.007554821,0.00002081906,0.0001308082,0.001171955],"genre_scores_gemma":[0.9909276,0.001058988,0.001472479,0.002354232,0.002198741,0.001539621,0.00003627242,0.00007546234,0.0003365937],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2621263,"threshold_uncertainty_score":0.9991423,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2990740094","doi":"10.1136/bmjebm-2019-111272","title":"Building capacity in evidence-based medicine in low-income and middle-income countries: problems and potential solutions","year":2019,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":13,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Hospital for Sick Children","funders":"National Institute for Health and Care Research","keywords":"Low and middle income countries; Income distribution; Middle income; Development economics; Economics; Demographic economics; Developing country; Economic growth; Mathematics; Inequality","authors":[{"name":"Peter J. Gill","is_ca":true},{"name":"Shabana M Ali","is_ca":false},{"name":"Yasmin Elsobky","is_ca":false},{"name":"Raymond C. Okechukwu","is_ca":false},{"name":"Tatiane Bomfim Ribeiro","is_ca":false},{"name":"Augusto César Soares dos Santos","is_ca":false},{"name":"Daniel Umpierre","is_ca":false},{"name":"Georgia C. Richards","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2712587169033339,"gpt":0.4607731902002766,"spread":0.1895144732969427,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0183651,0.0003722989,0.0009287625,0.00136404,0.000702323,0.00002614213,0.0004363678,0.0003370625,0.001165509],"category_scores_gemma":[0.01359961,0.000277599,0.00003879711,0.001813428,0.001294977,0.001074367,0.0001381029,0.001348432,0.00008261194],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001015284,"about_ca_system_score_gemma":0.003524941,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01849156,"about_ca_topic_score_gemma":0.002501452,"domain_scores_codex":[0.9922779,0.001760873,0.001806657,0.0009151775,0.001566897,0.001672447],"domain_scores_gemma":[0.988663,0.008694844,0.0005166727,0.0006497645,0.000537233,0.0009384628],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0006575641,0.00008140199,0.9762903,0.01407175,0.000008660743,0.00004903965,0.004797146,0.0006657092,0.001282315,0.0007544332,0.0008716058,0.0004700059],"study_design_scores_gemma":[0.006697354,0.001475872,0.7417037,0.2160323,0.00003147928,0.00001667505,0.002787903,0.02875029,0.00003854288,0.001335898,0.0007219985,0.0004079154],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8848535,0.006847908,0.001188235,0.1011431,0.001440147,0.004345217,0.000008450108,0.00005943325,0.0001139974],"genre_scores_gemma":[0.9917657,0.00233849,0.0008822408,0.003195782,0.0009295808,0.0006422365,0.000008832681,0.00002647164,0.0002106281],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2345866,"threshold_uncertainty_score":0.9999676,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"incentives","study_design":"theoretical_or_conceptual","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4200294250","doi":"10.1136/bmjebm-2021-111817","title":"Users' Guides to the Medical Literature series on social media (part 1): how to interpret healthcare information available on platforms","year":2021,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Social Media in Health Education","field":"Social Sciences","cited_by":13,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact; Université de Montréal","funders":"","keywords":"Social media; Health care; Medical information; Data science; Internet privacy; Computer science; Public relations; World Wide Web; Knowledge management; Political science","authors":[{"name":"Ryan S. D’Souza","is_ca":false},{"name":"Lubna Daraz","is_ca":true},{"name":"W. Michael Hooten","is_ca":false},{"name":"Gordon Guyatt","is_ca":true},{"name":"M. Hassan Murad","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2094278244877664,"gpt":0.434394690256741,"spread":0.2249668657689746,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.005749829,0.0002846805,0.0004787502,0.0003060875,0.001133574,0.000219837,0.0006772355,0.000417699,0.0014529],"category_scores_gemma":[0.1868077,0.0001973295,0.00009247343,0.002112227,0.0003927768,0.00100625,0.00007470173,0.0007138345,0.00057884],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0008248908,"about_ca_system_score_gemma":0.004542502,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001172763,"about_ca_topic_score_gemma":0.008668172,"domain_scores_codex":[0.9941326,0.0006749054,0.0007532372,0.0004265691,0.003268459,0.0007442747],"domain_scores_gemma":[0.9873769,0.009531721,0.0002382174,0.0005307341,0.00126604,0.001056433],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0004997365,0.00004755031,0.0009614557,0.0002315419,0.00001510392,0.00003035932,0.2043984,0.000008353196,0.000007502616,0.01079488,0.7264316,0.05657347],"study_design_scores_gemma":[0.0005718278,0.0007896817,0.004367244,0.01147282,0.00003048398,0.00000702715,0.04803704,0.00001394872,0.0003564388,0.0005849106,0.9334332,0.0003353525],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.04152169,0.001008515,0.0001056134,0.9397186,0.01366865,0.002557083,0.0000371548,0.0001507408,0.001231948],"genre_scores_gemma":[0.62822,0.0032872,0.001287761,0.3153249,0.04487125,0.00520304,0.0004708561,0.00007776519,0.001257179],"genre_candidate":"commentary","genre_consensus":null,"teacher_disagreement_score":0.6243937,"threshold_uncertainty_score":0.9994599,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4322758031","doi":"10.1136/bmjebm-2022-112098","title":"Decision Analysis in SHared decision making for Thromboprophylaxis during Pregnancy (DASH-TOP): a sequential explanatory mixed-methods pilot study","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Patient-Provider Communication in Healthcare","field":"Health Professions","cited_by":12,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"Mount Sinai Hospital; McMaster University; Impact","funders":"","keywords":"Decision aids; Intervention (counseling); Decision analysis; Medicine; Psychology; Nursing; Alternative medicine","authors":[{"name":"Brittany Humphries","is_ca":true},{"name":"Montserrat León‐García","is_ca":false},{"name":"Shannon M. Bates","is_ca":true},{"name":"Gordon Guyatt","is_ca":true},{"name":"Mark H. Eckman","is_ca":false},{"name":"Rohan D’Souza","is_ca":true},{"name":"Nadine Shehata","is_ca":true},{"name":"Susan M. Jack","is_ca":true},{"name":"Pablo Alonso‐Coello","is_ca":false},{"name":"Feng Xie","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4146435220118736,"gpt":0.5623066713106735,"spread":0.1476631492987999,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.005849603,0.0005007791,0.001422924,0.003085612,0.001332347,0.00002520779,0.001223864,0.0002695943,0.0005926988],"category_scores_gemma":[0.02629477,0.0004262105,0.0002201506,0.006153266,0.000158792,0.0005320439,0.000644527,0.00106713,0.0001050533],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001200705,"about_ca_system_score_gemma":0.001098091,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001274592,"about_ca_topic_score_gemma":0.004060547,"domain_scores_codex":[0.9881443,0.004758212,0.003112691,0.001260999,0.001544809,0.001178951],"domain_scores_gemma":[0.9686132,0.02629867,0.001164,0.002627309,0.0009432939,0.0003535312],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.006255574,0.0007843478,0.8581892,0.003973675,0.0003830999,0.0002184499,0.022741,0.002228941,0.002123302,0.00009038085,0.003468878,0.09954315],"study_design_scores_gemma":[0.009018635,0.003937401,0.8891975,0.05320401,0.0005485541,0.00000239935,0.02688409,0.01425494,0.0001113837,0.002028818,0.0001980614,0.0006141958],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.919865,0.008456768,0.04990045,0.003546375,0.003298047,0.01423529,0.00007427605,0.000568442,0.00005531347],"genre_scores_gemma":[0.956459,0.0003632088,0.03539705,0.0004403447,0.0003924824,0.006652152,0.0001202898,0.00009110058,0.00008431837],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.09892896,"threshold_uncertainty_score":0.9999678,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"qualitative","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2942870102","doi":"10.1136/bmjebm-2019-111186","title":"Training curriculum to help patient representatives participate meaningfully in the development of clinical practice guidelines","year":2019,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Mental Health and Patient Involvement","field":"Health Professions","cited_by":11,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Calgary","funders":"","keywords":"Guideline; Curriculum; Medical education; Grading (engineering); Likert scale; Medicine; Curriculum development; Face validity; Psychology; Scale (ratio); Readability; Pedagogy; Engineering; Computer science","authors":[{"name":"Lubna Daraz","is_ca":false},{"name":"Starr Webb","is_ca":false},{"name":"Rob Kunkle","is_ca":false},{"name":"M. Hassan Murad","is_ca":false},{"name":"Eddy Lang","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.787443666098379,"gpt":0.6554227771389366,"spread":0.1320208889594424,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.02233427,0.0002370025,0.0007406704,0.0001869974,0.0002548312,0.000003940413,0.000374806,0.000119356,0.0004961793],"category_scores_gemma":[0.02246301,0.0001395321,0.00007161838,0.0006285657,0.00009186492,0.0002147175,0.0001186176,0.0006028099,0.0002549595],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001484333,"about_ca_system_score_gemma":0.001871979,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0008229442,"about_ca_topic_score_gemma":0.0003376539,"domain_scores_codex":[0.9892868,0.00327385,0.00469376,0.0005194243,0.001508774,0.0007173659],"domain_scores_gemma":[0.9874505,0.009032088,0.001432337,0.0006353813,0.00108244,0.0003672281],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"qualitative","study_design_scores_codex":[0.002169866,0.000757034,0.6152365,0.002847183,0.00005374831,0.00004013034,0.2398238,0.0003753624,0.0002022389,0.0005968453,0.02532573,0.1125716],"study_design_scores_gemma":[0.00660765,0.007430283,0.09073158,0.07412777,0.0001346584,0.000003821739,0.6105456,0.00196968,0.0004478363,0.0002310888,0.2071811,0.0005889016],"study_design_candidate":"qualitative","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9417729,0.000715053,0.0009222085,0.04506941,0.002061918,0.006095177,0.000004770158,0.00003053351,0.003328028],"genre_scores_gemma":[0.8961114,0.0000962078,0.04237852,0.05920048,0.000639733,0.001332016,0.00004553554,0.00002317483,0.0001729314],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.524505,"threshold_uncertainty_score":0.9857712,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"nonrandomized_trial","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3121563112","doi":"10.1136/bmjebm-2020-111604","title":"Barriers to reporting guideline adherence in point-of-care ultrasound research: a cross-sectional survey of authors and journal editors","year":2021,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Ultrasound in Clinical Applications","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Ottawa Hospital; University of Ottawa","funders":"","keywords":"Likert scale; Guideline; Medicine; Family medicine; Cross-sectional study; Alternative medicine; Survey research; MEDLINE; Primary care; Medical education; Medical physics; Psychology; Pathology; Applied psychology","authors":[{"name":"Ross Prager","is_ca":true},{"name":"Luke B.-P. Gagnon","is_ca":true},{"name":"Joshua Bowdridge","is_ca":true},{"name":"Rudy R Unni","is_ca":true},{"name":"Trevor A. McGrath","is_ca":true},{"name":"Kelly D. Cobey","is_ca":true},{"name":"Patrick M. Bossuyt","is_ca":false},{"name":"Matthew D. F. McInnes","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3706017958214273,"gpt":0.5493210041737503,"spread":0.178719208352323,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.02854073,0.0002083276,0.0009060705,0.000505158,0.0001776379,0.00003158236,0.0002578545,0.0002372003,0.001321248],"category_scores_gemma":[0.5496205,0.00017002,0.0001219386,0.00219728,0.001069013,0.0001799798,0.00008949903,0.001099733,0.000005162568],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002563118,"about_ca_system_score_gemma":0.004447122,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001213826,"about_ca_topic_score_gemma":0.0004448232,"domain_scores_codex":[0.9915701,0.000814118,0.004640214,0.00064357,0.00187946,0.0004525688],"domain_scores_gemma":[0.9688174,0.01849238,0.001854235,0.0008758767,0.008746776,0.001213323],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.000757425,0.00007710012,0.9670929,0.0003398028,0.00003524978,0.00009415442,0.0005076342,0.0001093993,0.02033971,0.00003472104,0.009921089,0.0006908199],"study_design_scores_gemma":[0.001729787,0.00167355,0.9826974,0.004789034,0.00004290913,0.0003819652,0.001866636,0.00005942502,0.004945641,0.0004276769,0.001239804,0.0001461236],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9853976,0.002053054,0.001832517,0.008327161,0.001193986,0.0009268759,0.00004053752,0.00002085334,0.0002073964],"genre_scores_gemma":[0.9852462,0.0002577173,0.01137492,0.0007231087,0.002048727,0.00007218993,0.00007783312,0.000028082,0.0001712161],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5210798,"threshold_uncertainty_score":0.9995917,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4385863795","doi":"10.1136/bmjebm-2023-112271","title":"Reducing low value care: opportunities and challenges for Choosing Wisely campaigns","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Healthcare cost, quality, practices","field":"Health Professions","cited_by":11,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Toronto; St. Michael's Hospital; Ottawa Hospital; University of Ottawa","funders":"Canadian Institutes of Health Research","keywords":"Value (mathematics); Business; Computer science","authors":[{"name":"Jeremy Grimshaw","is_ca":true},{"name":"Wendy Levinson","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.8963123753308139,"gpt":0.6013933185560272,"spread":0.2949190567747867,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.01395188,0.0003665371,0.0008108861,0.000496393,0.001314404,0.00001931373,0.0002953638,0.0003503683,0.0001631149],"category_scores_gemma":[0.02947368,0.0003057126,0.0000796676,0.0004616075,0.0004050819,0.0006275942,0.0001285112,0.0007992597,0.00006812516],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004134361,"about_ca_system_score_gemma":0.002970639,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001649632,"about_ca_topic_score_gemma":0.000350425,"domain_scores_codex":[0.9930295,0.002936368,0.001331017,0.0007452568,0.0008626276,0.001095202],"domain_scores_gemma":[0.976226,0.02059427,0.0007627399,0.0007671585,0.0009768971,0.0006729073],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"qualitative","study_design_scores_codex":[0.002784233,0.0001896492,0.02000241,0.1464597,0.0002895127,0.0005525077,0.2404653,0.0002430518,0.002704825,0.03118231,0.1181109,0.4370157],"study_design_scores_gemma":[0.006650937,0.002779533,0.03550213,0.113761,0.0005455568,0.00002575841,0.523369,0.002924639,0.0003612742,0.002660724,0.3101038,0.001315669],"study_design_candidate":"qualitative","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.1958553,0.05272608,0.0008212518,0.7390133,0.003406809,0.006040175,0.00005992658,0.0008381318,0.001238958],"genre_scores_gemma":[0.8509769,0.09188624,0.002754292,0.0393884,0.008488831,0.003617849,0.0002246596,0.0002350457,0.002427781],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.699625,"threshold_uncertainty_score":0.9999858,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2935997213","doi":"10.1136/bmjebm-2019-111193","title":"Guidelines do not self-implement: time for a research paradigm shift from massive creation to effective implementation in evidence-based medicine research in China","year":2019,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Clinical practice guidelines implementation","field":"Medicine","cited_by":10,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Ottawa Hospital; University of Ottawa","funders":"European Society of Cardiology","keywords":"Guideline; Context (archaeology); Quality (philosophy); China; Paradigm shift; Engineering ethics; Chen; Clinical Practice; Quality management; Computer science; Psychology; Medical education; Management science; Medicine; Political science; Nursing; Engineering; Epistemology; Operations management","authors":[{"name":"Junqiang Zhao","is_ca":true},{"name":"Melissa Demery Varin","is_ca":true},{"name":"Ian D. Graham","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.5299034806332391,"gpt":0.6399346695960708,"spread":0.1100311889628317,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.06511396,0.0006977171,0.0018488,0.003946222,0.0002768219,0.00009037941,0.0007388811,0.0003823921,0.004218296],"category_scores_gemma":[0.08958434,0.0005697071,0.0002090547,0.004745918,0.0004016961,0.001238383,0.0002329261,0.001569393,0.0004633099],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.003252174,"about_ca_system_score_gemma":0.003158947,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0290545,"about_ca_topic_score_gemma":0.003519915,"domain_scores_codex":[0.9807048,0.004796259,0.005111491,0.00221202,0.005344219,0.001831227],"domain_scores_gemma":[0.9273381,0.06650263,0.0009192138,0.001662286,0.002801094,0.0007766383],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.07855631,0.001909053,0.5156892,0.004332963,0.0004999754,0.0005891853,0.01179579,0.003161571,0.06126717,0.0029139,0.1554365,0.1638484],"study_design_scores_gemma":[0.07640855,0.0336804,0.7258802,0.07670777,0.0008304742,0.00001592952,0.00816484,0.03633681,0.009953589,0.01229766,0.0183993,0.001324481],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.4792701,0.001511148,0.003959604,0.4893835,0.0004913461,0.0250661,0.00008296782,0.0001137261,0.000121511],"genre_scores_gemma":[0.9719641,0.0004670974,0.007652342,0.009648205,0.002970112,0.006037898,0.001003928,0.000135279,0.0001210925],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.4926939,"threshold_uncertainty_score":0.9996755,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4385063979","doi":"10.1136/bmjebm-2023-112270","title":"High-value care education can learn from the evidence-based medicine movement: moving beyond competencies and curricula to culture","year":2023,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Healthcare cost, quality, practices","field":"Health Professions","cited_by":9,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Health Sciences Centre; Women in Science and Engineering Newfoundland and Labrador; University of Toronto; Sunnybrook Health Science Centre","funders":"","keywords":"Curriculum; Movement (music); Value (mathematics); Medical education; Medicine; Pedagogy; Psychology; Computer science; Art","authors":[{"name":"Christopher Moriates","is_ca":false},{"name":"William K. Silverstein","is_ca":true},{"name":"Renato Gorga Bandeira de Mello","is_ca":false},{"name":"Lorette Stammen","is_ca":false},{"name":"Brian M. Wong","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4685115834507783,"gpt":0.5371929324341386,"spread":0.06868134898336031,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.009942932,0.0006401371,0.001103068,0.0004774486,0.001981825,0.00004472634,0.0008890289,0.0004041216,0.0008952202],"category_scores_gemma":[0.04945205,0.0003957339,0.00009615788,0.002214546,0.0007504699,0.0006596319,0.0002996608,0.001749475,0.00025456],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001020722,"about_ca_system_score_gemma":0.004877944,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.09940045,"about_ca_topic_score_gemma":0.009812913,"domain_scores_codex":[0.9875593,0.005916906,0.001863488,0.00118992,0.002323535,0.001146891],"domain_scores_gemma":[0.9723691,0.0221423,0.001112884,0.001536452,0.001845317,0.0009939008],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"observational","study_design_scores_codex":[0.0009755588,0.0001538928,0.3485786,0.00696191,0.0001453152,0.0001180474,0.1145646,0.0009039024,0.007728778,0.009799722,0.4499579,0.06011181],"study_design_scores_gemma":[0.0045432,0.002677251,0.4161595,0.1667744,0.0009058383,0.00000504482,0.2852409,0.001461898,0.0004369206,0.003965769,0.1165525,0.001276935],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.336037,0.02780373,0.0004452693,0.6266661,0.003501879,0.004830697,0.00008511438,0.0003928901,0.0002372888],"genre_scores_gemma":[0.7076984,0.004450806,0.002297402,0.2736393,0.007746851,0.001912269,0.0005871347,0.0001141581,0.001553722],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.3716614,"threshold_uncertainty_score":0.9998494,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4408245171","doi":"10.1136/bmjebm-2024-113077","title":"Rating certainty when the target threshold is the null and the point estimate is close to the null","year":2025,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":9,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Impact","funders":"Einstein Stiftung Berlin; National Natural Science Foundation of China","keywords":"Certainty; Null hypothesis; Null (SQL); Harm; Confidence interval; Point estimation; Point (geometry); Range (aeronautics); Grading (engineering); Econometrics; Mathematics; Statistics; Computer science; Psychology; Social psychology; Data mining; Engineering","authors":[{"name":"Linan Zeng","is_ca":true},{"name":"Monica Hultcrantz","is_ca":false},{"name":"David Tovey","is_ca":false},{"name":"Nancy Santesso","is_ca":true},{"name":"Philipp Dahm","is_ca":false},{"name":"Romina Brignardello‐Petersen","is_ca":true},{"name":"Reem A. Mustafa","is_ca":false},{"name":"M. Hassan Murad","is_ca":false},{"name":"Ariel Izcovich","is_ca":false},{"name":"Hans de Beer","is_ca":false},{"name":"Martín Ragusa","is_ca":false},{"name":"Bradley C. Johnston","is_ca":false},{"name":"Lingli Zhang","is_ca":false},{"name":"Alfonso Iorio","is_ca":true},{"name":"Gordon Guyatt","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.5544545072542626,"gpt":0.533779388139586,"spread":0.02067511911467657,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.2232665,0.0005042109,0.002601177,0.0002164196,0.001218216,0.00100396,0.004319274,0.00008785012,0.007346946],"category_scores_gemma":[0.1373863,0.0001154557,0.001013741,0.002073113,0.0009457282,0.0002189394,0.0004759413,0.0005140012,0.0006686706],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005274731,"about_ca_system_score_gemma":0.00031509,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0004395731,"about_ca_topic_score_gemma":0.0001612443,"domain_scores_codex":[0.9746692,0.01107965,0.006529301,0.001218525,0.005974186,0.0005291013],"domain_scores_gemma":[0.9257597,0.0599296,0.003498155,0.009029307,0.001537308,0.0002459388],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.000124449,0.00001526529,0.00393388,0.00008633265,0.0002127645,0.000006079075,0.008129672,0.0003654196,0.00003326464,0.009124847,0.967621,0.01034704],"study_design_scores_gemma":[0.002129341,0.0002357319,0.005747797,0.004217698,0.002196837,0.00003820795,0.01485491,0.2072164,0.0002951888,0.1286404,0.6339806,0.0004469466],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.009351308,0.02171089,0.02977515,0.9280016,0.000651152,0.005321885,0.00001414018,0.0000142868,0.00515957],"genre_scores_gemma":[0.7165158,0.0004154416,0.003598495,0.2558503,0.0009456702,0.00102852,0.000003970436,0.00003017279,0.02161164],"genre_candidate":"commentary","genre_consensus":null,"teacher_disagreement_score":0.7071645,"threshold_uncertainty_score":0.9935605,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"}]}