{"meta":{"query_hash":"4d5d67e28eb3","filters":{"category":"metaepi_broad"},"cohort_total":233,"direct_labels_cover":233,"predictions_cover":233,"exported":233,"export_cap":100000,"truncated":false,"label_status":"direct model label, unvalidated","prediction_status":"machine_predicted_unvalidated (Codex and Gemma teacher distillation)","score_status":"score_only:v0-immature-baseline","snapshot":{"source":"OpenAlex, pinned release, all 482 partitions","release":"2026-06-24","frame_built":"2026-07-12"},"permalink":"https://metacan.xera.ac/q/4d5d67e28eb3","api":"https://metacan.xera.ac/api/v1/cohort?category=metaepi_broad"},"results":[{"id":"W1493255413","doi":"10.3138/cjpe.027.003","title":"Capturing the Complexity of Evaluations of Health Promotion Interventions: A Scoping Review","year":2012,"lang":"en","type":"review","venue":"Canadian Journal of Program Evaluation","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":27,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Lloydminster Hospital; University of Calgary","funders":"","keywords":"Psychological intervention; Perspective (graphical); Promotion (chess); Grey literature; Health promotion; Management science; Psychology; Applied psychology; Computer science; MEDLINE; Medicine; Public health; Nursing; Political science; Engineering; Artificial intelligence","score_opus":0.9727255468434597,"score_gpt":0.7916901593614977,"score_spread":0.18103538748196202,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1493255413","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.021664474,0.8879426,0.048908733,0.009747835,0.0014073576,0.01962988,0.0010717901,0.00013745273,0.00948988],"genre_scores_gemma":[0.2338677,0.5744503,0.14948894,0.0035081229,0.000578216,0.035890885,0.0012767209,0.0000896451,0.00084947277],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.6193001,0.252163,0.07159695,0.0051159984,0.050239112,0.0015848039],"domain_scores_gemma":[0.26970792,0.6151707,0.041263677,0.015418152,0.05728308,0.0011565087],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.32814425,0.002069994,0.0072532436,0.02118894,0.0021959038,0.010471111,0.0031534738,0.0027899514,0.0019880203],"category_scores_gemma":[0.59623456,0.002361517,0.006069713,0.021134507,0.0031766575,0.01003801,0.0042791734,0.0028236278,0.000271761],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00041120258,0.00022174223,0.0072784983,0.46892545,0.0113006495,0.0001418584,0.0052953186,0.002130482,0.0007876255,0.008621112,0.0041911397,0.4906949],"study_design_scores_gemma":[0.00030847354,0.00074942276,0.011440522,0.9145831,0.020570993,0.00021796512,0.0041428637,0.0021982922,0.00159371,0.010395952,0.033613708,0.00018497613],"about_ca_topic_score_codex":0.011184769,"about_ca_topic_score_gemma":0.03301184,"teacher_disagreement_score":0.67185575,"about_ca_system_score_codex":0.013906963,"about_ca_system_score_gemma":0.057181578,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1544000882","doi":"10.1111/j.1756-5391.2012.01173.x","title":"Systematic reviews of TCM trials: how does inclusion of Chinese trials affect outcome?","year":2012,"lang":"en","type":"article","venue":"Journal of Evidence-Based Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Inclusion (mineral); English language; Clinical trial; Chinese language; Systematic review; MEDLINE; Randomized controlled trial; China; Alternative medicine; Medicine; Computer science; Medical education; Psychology; Pathology; Linguistics; Social psychology; Political science","score_opus":0.9254087096431709,"score_gpt":0.6418932184013989,"score_spread":0.283515491241772,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1544000882","genre_codex":"review","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015923152,0.9094245,0.020600228,0.027582252,0.003281126,0.014769997,0.0036806562,0.00018611533,0.004551968],"genre_scores_gemma":[0.5286565,0.28856152,0.07050012,0.026325062,0.0032435143,0.07867394,0.0027117513,0.00033166012,0.0009960318],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.20404193,0.51803386,0.21192892,0.018774731,0.04504545,0.002175119],"domain_scores_gemma":[0.11560245,0.74767303,0.09019625,0.02031721,0.025179023,0.0010321158],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.5837778,0.0032627408,0.01802771,0.018407097,0.0042729974,0.014767172,0.006378443,0.006267817,0.006144984],"category_scores_gemma":[0.82465434,0.0033447582,0.014541748,0.025135128,0.008381182,0.019385848,0.008650056,0.004436077,0.0007784676],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0027009032,0.00011666533,0.025359116,0.7685531,0.070980445,0.00033589697,0.0067268577,0.0009698062,0.00049588847,0.007162373,0.00543368,0.11116525],"study_design_scores_gemma":[0.005131751,0.0014183922,0.042080276,0.7079582,0.17876682,0.0007740536,0.003589564,0.0035249633,0.0020935666,0.025523793,0.028619893,0.0005187931],"about_ca_topic_score_codex":0.009148925,"about_ca_topic_score_gemma":0.015064589,"teacher_disagreement_score":0.9819723,"about_ca_system_score_codex":0.01910528,"about_ca_system_score_gemma":0.035429824,"threshold_uncertainty_score":0.5132761},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W1561379059","doi":"10.1371/journal.pone.0132634","title":"PEDro or Cochrane to Assess the Quality of Clinical Trials? A Meta-Epidemiological Study","year":2015,"lang":"en","type":"review","venue":"PLoS ONE","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":173,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"Fondation pour la Recherche Médicale; Physiotherapy Foundation of Canada; Alberta Innovates; Killam Trusts; National Institutes of Health; Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung; University of Alberta; Alberta Innovates - Health Solutions; Women and Children's Health Research Institute; Children's Health Research Institute; Bill and Melinda Gates Foundation; Canadian Institutes of Health Research; National Science Foundation","keywords":"Blinding; Medicine; Meta-analysis; Systematic review; Clinical trial; MEDLINE; Physical therapy; Internal medicine","score_opus":0.9978832001452151,"score_gpt":0.8031065913405215,"score_spread":0.1947766088046936,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1561379059","genre_codex":"review","genre_gemma":"review","domain_codex":"methods","domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0025780934,0.95136917,0.020833287,0.00955124,0.0054785134,0.006427243,0.00167041,0.00025818552,0.0018337592],"genre_scores_gemma":[0.22744308,0.4730799,0.1798089,0.029685592,0.009049092,0.07649571,0.0024906562,0.0006561258,0.0012909665],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.38489363,0.47613102,0.091194145,0.018242396,0.028290313,0.0012485426],"domain_scores_gemma":[0.34559986,0.5301894,0.07981639,0.024097791,0.0179605,0.0023360779],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.32634926,0.0039106905,0.03836769,0.027686872,0.001516888,0.01896381,0.006310943,0.010074567,0.0092360405],"category_scores_gemma":[0.6629876,0.004238889,0.032025058,0.024256915,0.005144932,0.012246159,0.0066036712,0.008317989,0.0015313275],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0052369256,0.00008117652,0.005581405,0.4866207,0.42142698,0.00015429444,0.0006357697,0.00095673034,0.00021386553,0.0063588494,0.008775158,0.06395814],"study_design_scores_gemma":[0.017812662,0.0014153827,0.009555934,0.37269092,0.52030087,0.00062207057,0.0006503961,0.004487571,0.000537085,0.03699398,0.03446099,0.00047219367],"about_ca_topic_score_codex":0.0022424797,"about_ca_topic_score_gemma":0.0022987581,"teacher_disagreement_score":0.9616323,"about_ca_system_score_codex":0.007150777,"about_ca_system_score_gemma":0.009667874,"threshold_uncertainty_score":0.8307313},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"evaluation","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W1570138353","doi":"10.1111/j.1365-2753.2007.00817.x","title":"Accuracy assessment of prediction in patient outcomes","year":2008,"lang":"en","type":"review","venue":"Journal of Evaluation in Clinical Practice","topic":"Machine Learning in Healthcare","field":"Computer Science","cited_by":25,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ontario Tech University","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Medicine; Outcome (game theory); Clinical Practice; Health care; Precision medicine; MEDLINE; Disease; Intensive care medicine; Physical therapy; Internal medicine","score_opus":0.268524647334157,"score_gpt":0.6142842616869206,"score_spread":0.34575961435276364,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1570138353","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0057321806,0.9225444,0.052723866,0.0079568,0.0013154512,0.0001809925,0.0004736659,0.00016493627,0.008907667],"genre_scores_gemma":[0.1970644,0.7255137,0.06863611,0.002502133,0.0030689037,0.00054729055,0.00079335365,0.00008648046,0.0017876028],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.95853573,0.021630252,0.0043613035,0.0019327116,0.013118018,0.00042195036],"domain_scores_gemma":[0.8343246,0.13905512,0.010705667,0.0035884685,0.011909708,0.00041636336],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.05176748,0.0014235706,0.0033996096,0.0084482385,0.00037076796,0.0032206606,0.002441588,0.0017977887,0.0013664373],"category_scores_gemma":[0.14326034,0.00049661763,0.0016254522,0.0057531046,0.0020098963,0.0037807995,0.0016616305,0.002526637,0.0009629256],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018119405,0.0000696628,0.021692412,0.00935693,0.0012113085,0.000080146376,0.00038958568,0.0054366197,0.00027181467,0.015832419,0.010549615,0.93492824],"study_design_scores_gemma":[0.0002889056,0.0016820438,0.15209445,0.11379342,0.0083499765,0.005758218,0.001720495,0.06696279,0.0083261505,0.24743219,0.39272472,0.0008666713],"about_ca_topic_score_codex":0.0021545454,"about_ca_topic_score_gemma":0.0015970552,"teacher_disagreement_score":0.94823253,"about_ca_system_score_codex":0.0018569565,"about_ca_system_score_gemma":0.0022523392,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W1821441292","doi":"10.1186/s13012-015-0286-1","title":"The effectiveness of knowledge translation interventions for promoting evidence-informed decision-making among nurses in tertiary care: a systematic review and meta-analysis","year":2015,"lang":"en","type":"review","venue":"Implementation Science","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":103,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"St. Peter's Hospital; Hamilton Health Sciences; Toronto Metropolitan University; Hamilton General Hospital; Lakehead University; McMaster Children's Hospital; McMaster University","funders":"McMaster University; London Health Sciences Centre; Canadian Institutes of Health Research; Hamilton Health Sciences","keywords":"CINAHL; Psychological intervention; Knowledge translation; Medicine; Data extraction; MEDLINE; Meta-analysis; Nursing research; Systematic review; Health care; Health services research; Nursing; Health administration; Public health; Knowledge management; Pathology","score_opus":0.5148330095571396,"score_gpt":0.701256680182869,"score_spread":0.1864236706257294,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1821441292","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0065634088,0.98720646,0.0016279989,0.00037856188,0.00031741415,0.003117485,0.000324929,0.000051945768,0.00041185526],"genre_scores_gemma":[0.17722946,0.79621303,0.010591542,0.0011837255,0.0003090045,0.013473215,0.0005909111,0.000050427723,0.00035868026],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9458029,0.028817097,0.015285569,0.002770065,0.006511824,0.0008125668],"domain_scores_gemma":[0.8907951,0.083512835,0.01385359,0.002983146,0.0078275055,0.001027777],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.06525699,0.0044650445,0.029300524,0.016631737,0.0015040826,0.005766977,0.0039383373,0.0036992144,0.0039055892],"category_scores_gemma":[0.14280139,0.0027084567,0.05530354,0.011913084,0.0018151298,0.0047832397,0.0037876663,0.0031939598,0.00034318696],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0021967564,0.000118998585,0.0011603106,0.605655,0.36802194,0.000077739,0.00028228897,0.0004360416,0.00012834878,0.00016948988,0.0002979505,0.021455184],"study_design_scores_gemma":[0.0015140586,0.00067104265,0.0015187159,0.105861455,0.8880628,0.0000646985,0.0001420287,0.00028825962,0.00024458935,0.00027276756,0.0013141441,0.00004541987],"about_ca_topic_score_codex":0.009189986,"about_ca_topic_score_gemma":0.01671228,"teacher_disagreement_score":0.934743,"about_ca_system_score_codex":0.010664278,"about_ca_system_score_gemma":0.01608227,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1837422480","doi":"10.1111/j.1360-0443.2012.03867.x","title":"Assessment of generalizability, applicability and predictability (GAP) for evaluating external validity in studies of universal family‐based prevention of alcohol misuse in young people: systematic methodological review of randomized controlled trials","year":2012,"lang":"en","type":"review","venue":"Addiction","topic":"Substance Abuse Treatment and Outcomes","field":"Medicine","cited_by":33,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital; University of Ottawa","funders":"Plan Nacional sobre Drogas; National Institutes of Health; University of Oxford; National Institute for Health and Care Research; Oxford Brookes University","keywords":"Generalizability theory; External validity; Internal validity; Randomized controlled trial; Psychology; Observational study; Systematic review; Population; Criterion validity; Clinical psychology; PsycINFO; Medicine; MEDLINE; Construct validity; Psychometrics; Social psychology; Developmental psychology; Environmental health","score_opus":0.5351974563246056,"score_gpt":0.550507421152536,"score_spread":0.015309964827930456,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1837422480","genre_codex":"review","genre_gemma":"review","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.023631983,0.8366884,0.056607142,0.0070389896,0.0036500562,0.06360665,0.0035900918,0.00032348253,0.0048631458],"genre_scores_gemma":[0.41779754,0.15935208,0.19559103,0.006641824,0.001730145,0.21474493,0.0029792334,0.00027700965,0.00088629045],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.25791776,0.35673198,0.3231476,0.015455172,0.045115132,0.0016323563],"domain_scores_gemma":[0.106242724,0.7391557,0.09419956,0.029149415,0.029926183,0.0013263905],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.6090625,0.0033274717,0.02098133,0.02525686,0.0028756396,0.009134623,0.0044846386,0.0067510074,0.0032788771],"category_scores_gemma":[0.78007686,0.0038452079,0.033012897,0.020435348,0.009384346,0.00915663,0.009075537,0.004875253,0.0003923123],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0023847898,0.00010013316,0.030605746,0.738309,0.13404855,0.00027191773,0.0037205613,0.001226873,0.0005190507,0.004446148,0.0031036707,0.08126359],"study_design_scores_gemma":[0.006197025,0.0020220485,0.042865314,0.5982385,0.28670287,0.0008928942,0.0028597969,0.005311479,0.0024505549,0.023344262,0.02851076,0.0006044991],"about_ca_topic_score_codex":0.0031980628,"about_ca_topic_score_gemma":0.0057708328,"teacher_disagreement_score":0.3909375,"about_ca_system_score_codex":0.011225374,"about_ca_system_score_gemma":0.018425107,"threshold_uncertainty_score":0.4820956},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"systematic_review","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1850986795","doi":"10.1186/1745-6215-16-s1-p24","title":"Completeness of outcome description in studies for low back pain rehabilitation interventions: a survey of trials included in Cochrane reviews","year":2015,"lang":"en","type":"article","venue":"Trials","topic":"Musculoskeletal pain and rehabilitation","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Arthritis Research Centre of Canada; University of British Columbia","funders":"","keywords":"Medicine; Psychological intervention; Rehabilitation; Physical therapy; Alternative medicine; Completeness (order theory); Systematic review; MEDLINE; Physical medicine and rehabilitation; Nursing; Pathology","score_opus":0.6664556406432591,"score_gpt":0.5615910426219709,"score_spread":0.10486459802128822,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1850986795","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.007348967,0.9751593,0.006654258,0.0012533758,0.00039106517,0.0038868561,0.0031130246,0.00011921356,0.0020739322],"genre_scores_gemma":[0.25552037,0.682385,0.024835397,0.0040912535,0.0006093049,0.02575386,0.005279405,0.0005230552,0.0010023317],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.5220892,0.18195745,0.23598483,0.010409673,0.04802906,0.0015299066],"domain_scores_gemma":[0.25657216,0.56028605,0.13816415,0.015332237,0.02845002,0.0011953661],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.25299734,0.0022998815,0.014001869,0.021788437,0.0014159624,0.0059095756,0.0027356797,0.0031063214,0.0053756116],"category_scores_gemma":[0.63279176,0.0032642845,0.022506058,0.022940623,0.0037672045,0.0078120735,0.0058133197,0.0036583147,0.00075334276],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0032056915,0.000044902852,0.014805506,0.8262824,0.07859698,0.00016626001,0.0011739045,0.0003735193,0.0005812934,0.0013588285,0.004210572,0.06920015],"study_design_scores_gemma":[0.0024697834,0.0006679843,0.047211144,0.6706632,0.22826792,0.0015720533,0.0010718004,0.0007667709,0.0016338929,0.0031698982,0.042285025,0.00022046891],"about_ca_topic_score_codex":0.0045998967,"about_ca_topic_score_gemma":0.007841738,"teacher_disagreement_score":0.74700266,"about_ca_system_score_codex":0.0062314006,"about_ca_system_score_gemma":0.0058586523,"threshold_uncertainty_score":0.9211873},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1966512523","doi":"10.1186/1472-6963-13-528","title":"A governance model for integrated primary/secondary care for the health-reforming first world – results of a systematic review","year":2013,"lang":"en","type":"review","venue":"BMC Health Services Research","topic":"Interprofessional Education and Collaboration","field":"Health Professions","cited_by":116,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Health informatics; Integrated care; Health administration; Nursing research; Nursing; Health care; Grey literature; CINAHL; Corporate governance; Population health; Health services research; Health policy; Inclusion (mineral); Public health; MEDLINE; Business; Political science; Psychological intervention; Psychology","score_opus":0.18176772037647074,"score_gpt":0.5574195498135738,"score_spread":0.37565182943710307,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1966512523","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.012256056,0.9708011,0.0020385738,0.0019605218,0.00043488407,0.009364753,0.0009748627,0.000034627057,0.0021345504],"genre_scores_gemma":[0.1450552,0.8279916,0.014043514,0.0019009236,0.00016316725,0.00947016,0.0010081702,0.000023041794,0.00034427337],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9538023,0.023944078,0.0120930085,0.0019668746,0.0071813674,0.0010123077],"domain_scores_gemma":[0.9475396,0.028753346,0.013104334,0.0021423327,0.007637787,0.00082252943],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.05853456,0.0015568124,0.006583753,0.0138560105,0.0011010205,0.006583616,0.0022143798,0.0019722034,0.0029259766],"category_scores_gemma":[0.097622864,0.0012099578,0.0099746715,0.015401013,0.0018848771,0.006223427,0.0036177214,0.0014672135,0.00027756984],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003951418,0.000079935446,0.0040547713,0.8780583,0.012008388,0.00021909362,0.0029748392,0.00058998604,0.00018850341,0.0029863715,0.0031055245,0.0953393],"study_design_scores_gemma":[0.00052178773,0.00039077818,0.0073592695,0.92694485,0.0377443,0.00035737432,0.0043209046,0.00030997826,0.00022936786,0.0010269823,0.020711701,0.00008266461],"about_ca_topic_score_codex":0.011122202,"about_ca_topic_score_gemma":0.026344534,"teacher_disagreement_score":0.05853456,"about_ca_system_score_codex":0.013737112,"about_ca_system_score_gemma":0.047608245,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"opus","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W1974429147","doi":"10.1136/annrheumdis-2014-eular.2210","title":"THU0416 Attrition Bias in Rheumatoid Arthritis Randomised Trials with Different Modified Intention-To-Treat Approaches: A Meta-Epidemiological Study","year":2014,"lang":"en","type":"article","venue":"Annals of the Rheumatic Diseases","topic":"Advanced Causal Inference Techniques","field":"Mathematics","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Randomized controlled trial; Population; Rheumatoid arthritis; Meta-analysis; Internal medicine; Intention-to-treat analysis; Sample size determination; Physical therapy; Imputation (statistics); Study heterogeneity; Missing data; Statistics","score_opus":0.6572351615281036,"score_gpt":0.44747438123377986,"score_spread":0.2097607802943237,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1974429147","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2201201,0.53177387,0.21544708,0.006859262,0.007442755,0.0066567957,0.0038825967,0.0018694941,0.005948176],"genre_scores_gemma":[0.9259026,0.019297602,0.040109705,0.0042571756,0.0017780937,0.004320605,0.0012725047,0.0004284045,0.002633349],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.71683407,0.24901652,0.017323535,0.009084807,0.0054378677,0.0023032199],"domain_scores_gemma":[0.7330773,0.2265989,0.012093969,0.022740567,0.0043926258,0.0010966012],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.15971294,0.0043580392,0.015147193,0.0037484374,0.0014220421,0.006236112,0.0053462167,0.009724429,0.010320503],"category_scores_gemma":[0.26303396,0.003115857,0.05516843,0.00518877,0.0025584265,0.005347708,0.0029046931,0.008572807,0.0010380801],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.09953756,0.0003573877,0.008694963,0.025308566,0.8238805,0.0004406815,0.00028900022,0.0029145805,0.0004368626,0.0039619417,0.0021779842,0.032],"study_design_scores_gemma":[0.07975696,0.004145932,0.0066345595,0.0035629068,0.8773566,0.00039666987,0.0001588531,0.011368132,0.00052467646,0.013089339,0.002778087,0.00022737173],"about_ca_topic_score_codex":0.003990686,"about_ca_topic_score_gemma":0.0030434395,"teacher_disagreement_score":0.99564195,"about_ca_system_score_codex":0.0019539236,"about_ca_system_score_gemma":0.0031206745,"threshold_uncertainty_score":0.8446529},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W1980348059","doi":"10.1177/000841740807500513","title":"Qualitative Meta-Synthesis: Reflections on the Utility and Challenges in Occupational Therapy","year":2008,"lang":"en","type":"article","venue":"Canadian Journal of Occupational Therapy","topic":"Occupational Therapy Practice and Research","field":"Health Professions","cited_by":53,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"McMaster University; Toronto Rehabilitation Institute; University of Toronto","funders":"","keywords":"Qualitative research; Phenomenon; Interpretation (philosophy); Computer science; Management science; Engineering ethics; Knowledge management; Epistemology; Psychology; Sociology; Engineering; Social science","score_opus":0.901061825158874,"score_gpt":0.6193784157795532,"score_spread":0.2816834093793208,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1980348059","genre_codex":"commentary","genre_gemma":"review","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.007016596,0.08268706,0.18900546,0.70029324,0.008607827,0.0029790143,0.00058973266,0.0003924815,0.0084286155],"genre_scores_gemma":[0.31557488,0.07486919,0.44173846,0.1356905,0.0035110302,0.023692502,0.00032196863,0.0009108244,0.003690672],"study_design_codex":"qualitative","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.18898946,0.7502041,0.028045855,0.006584899,0.0232317,0.0029439945],"domain_scores_gemma":[0.054179683,0.9001798,0.007128169,0.016803447,0.020731641,0.0009772844],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.7583464,0.0029520572,0.0070356005,0.010576984,0.010304954,0.030613378,0.01231442,0.02132251,0.00838291],"category_scores_gemma":[0.77414715,0.004411864,0.009758052,0.015203785,0.05042027,0.045550916,0.02456382,0.028826501,0.0010425702],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00053077575,0.00010538244,0.0024353347,0.059149444,0.0020166428,0.0011952294,0.4788374,0.0040301178,0.00085032877,0.26909435,0.045569737,0.13618529],"study_design_scores_gemma":[0.00034621914,0.00028231725,0.0012553729,0.12697549,0.001199422,0.0014767776,0.21693501,0.006145403,0.0024338244,0.40131712,0.24111217,0.0005208946],"about_ca_topic_score_codex":0.018436624,"about_ca_topic_score_gemma":0.018216139,"teacher_disagreement_score":0.24165362,"about_ca_system_score_codex":0.04351248,"about_ca_system_score_gemma":0.082309164,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W1985235854","doi":"10.1177/135245850000600509","title":"Multiple sclerosis, disease modifying treatments and depression: a critical methodological review","year":2000,"lang":"en","type":"review","venue":"Multiple Sclerosis Journal","topic":"Multiple Sclerosis Research Studies","field":"Medicine","cited_by":68,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Sunnybrook Health Science Centre","funders":"","keywords":"Depression (economics); Mood; Psychiatry; Disease; Psychology; Clinical psychology; Mood disorders; Multiple sclerosis; Major depressive disorder; Placebo; Medicine; Anxiety; Alternative medicine; Internal medicine","score_opus":0.5577539936427163,"score_gpt":0.46577594370169356,"score_spread":0.09197804994102271,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1985235854","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000093360555,0.9972893,0.00027861708,0.0012171492,0.0005425433,0.00041068284,0.000058748774,0.0000043217633,0.00010539722],"genre_scores_gemma":[0.003072106,0.9888087,0.0026907532,0.002347006,0.0005781462,0.0023066925,0.000092476206,0.000007878204,0.00009623316],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9142082,0.04485758,0.028403152,0.0022583588,0.009570389,0.0007023719],"domain_scores_gemma":[0.6559005,0.2778032,0.038017265,0.005043232,0.021934364,0.0013015158],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.09981266,0.0020284487,0.009447256,0.020139711,0.0019934087,0.007428597,0.0037021646,0.0059355097,0.004205014],"category_scores_gemma":[0.26650742,0.0019273004,0.00901687,0.016460909,0.0032164033,0.005246461,0.0038838105,0.003992128,0.0006078488],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00042705302,0.00005281323,0.00074556336,0.844797,0.0077750334,0.00009413467,0.00041106687,0.00022072432,0.0001409156,0.002367524,0.008438999,0.13452913],"study_design_scores_gemma":[0.00033800068,0.0002743956,0.0016522792,0.93016803,0.0205412,0.00024147135,0.00036054762,0.00013003453,0.00010109064,0.0026556177,0.04348656,0.0000506631],"about_ca_topic_score_codex":0.006507555,"about_ca_topic_score_gemma":0.014407136,"teacher_disagreement_score":0.9001874,"about_ca_system_score_codex":0.012125715,"about_ca_system_score_gemma":0.031193653,"threshold_uncertainty_score":0.5278661},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W1998831987","doi":"10.1136/bmjopen-2014-005491","title":"Assessing bias in osteoarthritis trials included in Cochrane reviews: protocol for a meta-epidemiological study","year":2014,"lang":"en","type":"article","venue":"BMJ Open","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; Ottawa Hospital; Cochrane","funders":"National Health and Medical Research Council; University of Ottawa; Medical Research Council; Parker Institute for Cancer Immunotherapy; Canadian Institutes of Health Research; Oak Foundation","keywords":"Medicine; Epidemiology; Protocol (science); Meta-analysis; Osteoarthritis; Alternative medicine; Cochrane collaboration; Systematic review; MEDLINE; Physical therapy; Family medicine; Internal medicine; Pathology; Cochrane Library","score_opus":0.9811986829709822,"score_gpt":0.7555210505783807,"score_spread":0.22567763239260152,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1998831987","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000293078,0.0021978428,0.005862086,0.0005215843,0.00070727145,0.984781,0.0044291774,0.00052323093,0.00068469887],"genre_scores_gemma":[0.00031613826,0.0004153089,0.008560316,0.000116169416,0.00003159844,0.9900422,0.0003138624,0.00003757102,0.00016679728],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.79335636,0.12126401,0.055134997,0.013115033,0.013342416,0.0037871257],"domain_scores_gemma":[0.74238133,0.13411725,0.057865333,0.024077797,0.035683505,0.0058747097],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.15985464,0.01115771,0.027173933,0.015124306,0.004366661,0.011770713,0.0075557353,0.012658285,0.1066387],"category_scores_gemma":[0.30686125,0.009581524,0.02960055,0.019560054,0.008559832,0.0111901965,0.008432344,0.02088894,0.022881938],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.019273246,0.00043882112,0.00088731165,0.84144217,0.011363325,0.00039702904,0.0020133248,0.002304174,0.0017314034,0.012165494,0.059973907,0.048009798],"study_design_scores_gemma":[0.13055027,0.003589067,0.0068713115,0.47477844,0.020799754,0.0009748915,0.0014958025,0.009519718,0.0055400627,0.054570727,0.2898845,0.0014255052],"about_ca_topic_score_codex":0.0033235066,"about_ca_topic_score_gemma":0.0052416953,"teacher_disagreement_score":0.9888423,"about_ca_system_score_codex":0.016465217,"about_ca_system_score_gemma":0.035223015,"threshold_uncertainty_score":0.84540224},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"design_other","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2001201455","doi":"10.1136/bmjopen-2013-004527","title":"Methodological characteristics and treatment effect sizes in oral health randomised controlled trials: Is there a relationship? Protocol for a meta-epidemiological study","year":2014,"lang":"en","type":"article","venue":"BMJ Open","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"Canadian Institutes of Health Research; Alberta Innovates; Killam Trusts; University of Alberta; Alberta Innovates - Health Solutions; Women and Children's Health Research Institute; Children's Health Research Institute; Fondation pour la Recherche Médicale; Alpha Omega Foundation","keywords":"Medicine; Sample size determination; Protocol (science); Randomized controlled trial; Meta-analysis; Epidemiology; Intervention (counseling); Research design; MEDLINE; Clinical trial; Placebo; Alternative medicine; Physical therapy; Statistics; Psychiatry; Internal medicine; Pathology","score_opus":0.9742725078857929,"score_gpt":0.7329682702713417,"score_spread":0.2413042376144512,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2001201455","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005640363,0.00322192,0.009150032,0.001212458,0.0016340048,0.9780352,0.0048578633,0.00042614157,0.00089840655],"genre_scores_gemma":[0.00075296784,0.00046708694,0.00794223,0.00028443956,0.000039016133,0.990006,0.00031753295,0.00003266244,0.0001580556],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.73436505,0.15712598,0.075865924,0.011285101,0.017928123,0.0034298196],"domain_scores_gemma":[0.686529,0.18681446,0.054424837,0.030959483,0.036512896,0.0047593485],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.20173547,0.006689571,0.021561213,0.01164611,0.0033381076,0.010337822,0.005941013,0.017658709,0.05366499],"category_scores_gemma":[0.42352897,0.0069154343,0.032627273,0.015105629,0.007445401,0.010457652,0.005430468,0.020011075,0.010857642],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.031955242,0.0007377188,0.0020597756,0.7996633,0.028237792,0.00047248052,0.0020912208,0.0032335145,0.0014346902,0.014582697,0.066073,0.049458563],"study_design_scores_gemma":[0.1545058,0.0050553065,0.012239666,0.4491963,0.04806022,0.0007783745,0.0015670656,0.01358191,0.0055414294,0.08280768,0.225352,0.001314361],"about_ca_topic_score_codex":0.0020284334,"about_ca_topic_score_gemma":0.0027651398,"teacher_disagreement_score":0.99331045,"about_ca_system_score_codex":0.012503481,"about_ca_system_score_gemma":0.021517139,"threshold_uncertainty_score":0.9844023},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2007317588","doi":"10.1007/s13187-010-0183-x","title":"Effective Teaching Strategies and Methods of Delivery for Patient Education: A Systematic Review and Practice Guideline Recommendations","year":2010,"lang":"en","type":"review","venue":"Journal of Cancer Education","topic":"Patient-Provider Communication in Healthcare","field":"Health Professions","cited_by":334,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"London Health Sciences Centre; McMaster University; Grand River Hospital; Sudbury Regional Hospital; Ontario Clinical Oncology Group; University Health Network","funders":"Ontario Ministry of Health and Long-Term Care; University Health Network; Cancer Care Ontario","keywords":"Medicine; Guideline; Medical education; Teaching method; Anxiety; Recall; Multimedia; Psychology; Pedagogy; Computer science; Pathology","score_opus":0.2071370985328947,"score_gpt":0.6437488558854468,"score_spread":0.43661175735255214,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2007317588","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0045747217,0.95245075,0.005803994,0.027476821,0.0021744822,0.004000946,0.0010038603,0.00013455025,0.0023799222],"genre_scores_gemma":[0.0355959,0.86705494,0.08128903,0.008187011,0.00069083215,0.005307678,0.0005395963,0.000054854525,0.0012801262],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9820117,0.0074638594,0.005639858,0.0010357079,0.0033693833,0.00047954882],"domain_scores_gemma":[0.92733043,0.047096536,0.011721397,0.0010845824,0.011238598,0.0015284871],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0342564,0.0016730385,0.006636971,0.006972067,0.001218173,0.0026413074,0.0034672604,0.0033955083,0.0044673663],"category_scores_gemma":[0.0925918,0.0015273425,0.006021492,0.0046349396,0.0011964815,0.0037587155,0.0019126767,0.0035357769,0.00075480586],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006668092,0.00048779498,0.003536053,0.36026776,0.0044197715,0.00025769026,0.0012057112,0.0004943703,0.0006197157,0.001248114,0.03377216,0.59302413],"study_design_scores_gemma":[0.002963288,0.00066821027,0.009550867,0.8702536,0.037605923,0.00043566484,0.0016100857,0.00083953934,0.0012669571,0.0027135008,0.0719388,0.00015343368],"about_ca_topic_score_codex":0.01655979,"about_ca_topic_score_gemma":0.073064186,"teacher_disagreement_score":0.0342564,"about_ca_system_score_codex":0.0065265885,"about_ca_system_score_gemma":0.022704618,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2007795708","doi":"10.1371/journal.pone.0080233","title":"Knowledge Transfer on Complex Social Interventions in Public Health: A Scoping Study","year":2013,"lang":"en","type":"review","venue":"PLoS ONE","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":39,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Montréal; Centre de Liaison Sur l'Intervention et la Prévention Psychosociales; Université Laval; Centre Hospitalier de l’Université de Montréal; Université de Montréal","funders":"Canadian Institutes of Health Research","keywords":"Psychological intervention; Public health; Knowledge transfer; Field (mathematics); Knowledge management; Exploratory research; Data science; Conceptual framework; Computer science; Management science; Psychology; Medicine; Sociology; Social science; Engineering","score_opus":0.9795260618445067,"score_gpt":0.7529023926355171,"score_spread":0.2266236692089897,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2007795708","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0031325212,0.9885385,0.001494887,0.0013468856,0.00022738277,0.0017139913,0.00008095792,0.000010115644,0.003454824],"genre_scores_gemma":[0.034354556,0.9571673,0.0037595672,0.00061739766,0.00014459052,0.003521243,0.000113730384,0.0000071827662,0.00031435522],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9558957,0.026149148,0.008000967,0.0015713371,0.0074810754,0.000901829],"domain_scores_gemma":[0.84553343,0.1340765,0.007963097,0.002638039,0.009225944,0.00056297996],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.06606125,0.0016385666,0.0060614473,0.018593544,0.002283935,0.0067735934,0.0024057443,0.003777437,0.003706749],"category_scores_gemma":[0.15455173,0.0011355052,0.0044873175,0.020641997,0.003060277,0.007274714,0.0047749956,0.0021492457,0.0004114653],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020289622,0.00023509533,0.0012274785,0.47742414,0.0025638118,0.00020199017,0.0044728136,0.0005561258,0.00028302564,0.0064562126,0.0019643756,0.50441194],"study_design_scores_gemma":[0.000107883956,0.0003294641,0.0031392714,0.95260316,0.0066585094,0.0002674185,0.0036402086,0.00032232475,0.0005082582,0.0032779179,0.029106658,0.00003909907],"about_ca_topic_score_codex":0.0072912346,"about_ca_topic_score_gemma":0.010812138,"teacher_disagreement_score":0.93393874,"about_ca_system_score_codex":0.009101682,"about_ca_system_score_gemma":0.03618647,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2009497067","doi":"10.1002/ebch.282","title":"<i>The Cochrane Library</i> and Long‐Acting Beta‐agonist Treatment for Childhood Asthma: An Overview of Reviews","year":2008,"lang":"en","type":"article","venue":"Evidence-Based Child Health A Cochrane Review Journal","topic":"Asthma and respiratory diseases","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary; University of Alberta; University of Ottawa","funders":"","keywords":"Asthma; Medicine; Cochrane Library; Systematic review; Placebo; Inhaled corticosteroids; Formoterol; B2 receptor; Meta-analysis; Pediatrics; Physical therapy; MEDLINE; Internal medicine; Alternative medicine; Budesonide","score_opus":0.09366244879714936,"score_gpt":0.3933631038727442,"score_spread":0.29970065507559485,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2009497067","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00017880561,0.9976417,0.000105741885,0.000392605,0.00026295232,0.00041420752,0.0004376923,0.000022715894,0.0005435804],"genre_scores_gemma":[0.002275647,0.99466646,0.0006566256,0.0007524794,0.00015555126,0.00096990983,0.00027087744,0.000009713074,0.00024270464],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99102306,0.002697508,0.003279602,0.00059310155,0.0020162058,0.00039064582],"domain_scores_gemma":[0.9750806,0.015378906,0.005747469,0.00033394113,0.0031264187,0.0003326286],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0074917343,0.0026255795,0.010631064,0.019290289,0.0006896921,0.0048412844,0.0023313952,0.0029590942,0.013331494],"category_scores_gemma":[0.033389717,0.0014473413,0.008129256,0.02619232,0.0009153081,0.0037960093,0.0022407668,0.0019191023,0.0011645728],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010381148,0.000008655144,0.00014127491,0.96176183,0.0019303963,0.000047698366,0.00007121466,0.00006331662,0.0001159783,0.00035278316,0.004569182,0.030833758],"study_design_scores_gemma":[0.00013932129,0.00008519164,0.001130207,0.95435613,0.011057755,0.0002199997,0.00010676044,0.000060442424,0.0000926171,0.0002744125,0.032454688,0.000022509777],"about_ca_topic_score_codex":0.012659214,"about_ca_topic_score_gemma":0.025131186,"teacher_disagreement_score":0.019290289,"about_ca_system_score_codex":0.006858706,"about_ca_system_score_gemma":0.015080268,"threshold_uncertainty_score":0.04976362},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2018803201","doi":"10.1016/s0895-4356(99)00160-2","title":"Randomized trials versus observational studies in adolescent pregnancy prevention","year":2000,"lang":"en","type":"article","venue":"Journal of Clinical Epidemiology","topic":"Adolescent Sexual and Reproductive Health","field":"Health Professions","cited_by":94,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Observational study; Randomized controlled trial; Medicine; Psychological intervention; CINAHL; PsycINFO; Pregnancy; MEDLINE; Family medicine; Psychiatry; Surgery; Internal medicine","score_opus":0.9380215926437274,"score_gpt":0.7360098352874899,"score_spread":0.20201175735623755,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2018803201","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.33261433,0.5481998,0.034624882,0.012003689,0.03519018,0.024664547,0.0034603286,0.00064586714,0.008596439],"genre_scores_gemma":[0.90561897,0.050886925,0.012567434,0.0068005864,0.0062354854,0.014771953,0.0010260371,0.00010923848,0.0019833774],"study_design_codex":"randomized_trial","study_design_gemma":"observational","domain_scores_codex":[0.81200635,0.15748622,0.020251544,0.005274944,0.0028028297,0.002178084],"domain_scores_gemma":[0.70182675,0.25553134,0.025671184,0.009672299,0.004007776,0.0032906537],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.098213315,0.0042029764,0.01644989,0.002148308,0.0016017769,0.0038946616,0.0026251625,0.009653111,0.010098281],"category_scores_gemma":[0.18574749,0.002445183,0.015781902,0.0023796253,0.005349136,0.005887479,0.002725186,0.009433651,0.00106984],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.84612775,0.0014297991,0.00317405,0.034553107,0.08827582,0.00011860259,0.00026340247,0.0009625003,0.000521712,0.0019220473,0.0011790254,0.021472238],"study_design_scores_gemma":[0.89075094,0.024985746,0.0026111526,0.0046197707,0.06815552,0.00009881542,0.0001538705,0.0012217731,0.000339301,0.004352837,0.0026343428,0.000075971264],"about_ca_topic_score_codex":0.0022300698,"about_ca_topic_score_gemma":0.0022073868,"teacher_disagreement_score":0.9017867,"about_ca_system_score_codex":0.0023733815,"about_ca_system_score_gemma":0.0031523127,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2022418062","doi":"10.1186/1471-2288-14-127","title":"Methodological and ethical issues in research using social media: a metamethod of Human Papillomavirus vaccine studies","year":2014,"lang":"en","type":"article","venue":"BMC Medical Research Methodology","topic":"Social Media in Health Education","field":"Social Sciences","cited_by":43,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Memorial University of Newfoundland","funders":"","keywords":"Social media; Credibility; Research ethics; Transparency (behavior); Generalizability theory; CLARITY; Psychology; Internet privacy; Informed consent; Public relations; Medicine; Computer science; World Wide Web; Political science; Alternative medicine","score_opus":0.9767535309614988,"score_gpt":0.79382938185026,"score_spread":0.18292414911123878,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2022418062","genre_codex":"methods","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.03375107,0.03662799,0.74539274,0.12988779,0.004648561,0.034612678,0.000706009,0.00024144548,0.014131786],"genre_scores_gemma":[0.21734436,0.008372166,0.69349843,0.020184306,0.001328655,0.058082078,0.00021707597,0.000106514366,0.00086631626],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"qualitative","domain_scores_codex":[0.09310271,0.8134896,0.062041633,0.008601106,0.021289552,0.0014753545],"domain_scores_gemma":[0.09317434,0.75863314,0.03468526,0.07451852,0.036721483,0.002267276],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.7676581,0.0022940212,0.0050289626,0.026782144,0.011197845,0.02040957,0.0066530965,0.011903587,0.0017120196],"category_scores_gemma":[0.7283162,0.0024771376,0.0045507923,0.01846249,0.07117925,0.02338648,0.021183848,0.013789255,0.00062712736],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007654387,0.00030044626,0.012751235,0.026544746,0.0011711288,0.0006879268,0.1957578,0.0023754002,0.0023993663,0.5520333,0.00846786,0.1967453],"study_design_scores_gemma":[0.0005420906,0.0010490821,0.0043202443,0.078394495,0.0009888115,0.002108497,0.079792775,0.0070215096,0.0037387153,0.69095874,0.13068467,0.00040046766],"about_ca_topic_score_codex":0.0016853422,"about_ca_topic_score_gemma":0.0020565866,"teacher_disagreement_score":0.23234189,"about_ca_system_score_codex":0.016151251,"about_ca_system_score_gemma":0.044984818,"threshold_uncertainty_score":0.286519},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2026016649","doi":"10.1016/s2352-3026(15)00003-4","title":"Rituximab plus standard of care for treatment of primary immune thrombocytopenia: a systematic review and meta-analysis","year":2015,"lang":"en","type":"review","venue":"The Lancet Haematology","topic":"Platelet Disorders and Treatments","field":"Medicine","cited_by":138,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Canadian Blood Services; McMaster University","funders":"","keywords":"Medicine; Rituximab; Internal medicine; Splenectomy; Meta-analysis; Immune thrombocytopenia; Relative risk; Randomized controlled trial; Confidence interval; Gastroenterology; Platelet; Lymphoma","score_opus":0.15312143755456406,"score_gpt":0.40161448540477473,"score_spread":0.24849304785021067,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2026016649","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0061052646,0.99273187,0.000315658,0.00011871906,0.00018206531,0.00015463312,0.0001735711,0.000018064475,0.00020006625],"genre_scores_gemma":[0.26090074,0.73386294,0.0023386222,0.0013052721,0.00030959825,0.00043963332,0.00046332105,0.000025925692,0.00035403404],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.99583167,0.0017475903,0.0012462166,0.0005036743,0.00046829667,0.00020251224],"domain_scores_gemma":[0.9941136,0.0040389528,0.0012014374,0.00019405597,0.0003231684,0.0001287895],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005591081,0.0019499849,0.016723856,0.002839111,0.00042179198,0.0022989276,0.0020135744,0.0021376284,0.004447696],"category_scores_gemma":[0.014560119,0.0012128784,0.024880279,0.003318863,0.000672341,0.0016962399,0.0013153796,0.002305974,0.00031527574],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.009614174,0.00008910162,0.0022828088,0.3501694,0.61179644,0.00012805927,0.000091382644,0.0005417143,0.00033258172,0.00015650771,0.00071915874,0.024078688],"study_design_scores_gemma":[0.004251011,0.0006070558,0.0027876259,0.014811592,0.9755853,0.000103850696,0.000056658442,0.00017897868,0.00009902539,0.00018350649,0.0013054709,0.000029919724],"about_ca_topic_score_codex":0.005735022,"about_ca_topic_score_gemma":0.015808752,"teacher_disagreement_score":0.016723856,"about_ca_system_score_codex":0.00212583,"about_ca_system_score_gemma":0.0019696061,"threshold_uncertainty_score":0.029568791},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2028540474","doi":"10.1371/journal.pone.0003684","title":"Following 411 Cochrane Protocols to Completion: A Retrospective Cohort Study","year":2008,"lang":"en","type":"article","venue":"PLoS ONE","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":62,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Institute of Population and Public Health; Children's Hospital of Eastern Ontario; University of Ottawa","funders":"Canadian Institutes of Health Research; University of Ottawa","keywords":"Cochrane Library; Medicine; Systematic review; MEDLINE; Protocol (science); Confidence interval; Cohort study; Hazard ratio; Internal medicine; Alternative medicine; Pathology","score_opus":0.7904036982622812,"score_gpt":0.5202770186509911,"score_spread":0.2701266796112901,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2028540474","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.84439963,0.034932006,0.014174556,0.002064317,0.00059349596,0.05102796,0.04552805,0.0003224585,0.006957545],"genre_scores_gemma":[0.8732538,0.0117858285,0.014230095,0.004630201,0.0005236677,0.07616981,0.014279981,0.00038733124,0.004739397],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.89158005,0.01957796,0.04601905,0.015978916,0.02265781,0.0041863075],"domain_scores_gemma":[0.67794985,0.092873946,0.12816471,0.04624084,0.04941511,0.005355593],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.07514405,0.0015293143,0.004180209,0.012239558,0.004061538,0.005780926,0.0034079736,0.0023455725,0.009794035],"category_scores_gemma":[0.2345552,0.0056437897,0.0067374287,0.011902336,0.0020037997,0.0071138265,0.004404811,0.0038926401,0.0059825014],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.010541848,0.001411151,0.90144974,0.01681558,0.0045911577,0.0035054812,0.008817171,0.00031252447,0.00087499275,0.0015712042,0.012747264,0.037361935],"study_design_scores_gemma":[0.0037842207,0.0068389685,0.8836137,0.014605677,0.011609289,0.010586671,0.0067265094,0.0015346542,0.0029543908,0.002499254,0.05446931,0.0007773341],"about_ca_topic_score_codex":0.011171819,"about_ca_topic_score_gemma":0.01140409,"teacher_disagreement_score":0.9958198,"about_ca_system_score_codex":0.0056109405,"about_ca_system_score_gemma":0.01665052,"threshold_uncertainty_score":0.3974045},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2045910459","doi":"10.1016/j.jclinepi.2009.06.007","title":"A scale for distinguishing efficacy from effectiveness was adapted and applied to stroke rehabilitation studies","year":2009,"lang":"en","type":"review","venue":"Journal of Clinical Epidemiology","topic":"Stroke Rehabilitation and Recovery","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St Joseph's Health Care; Lawson Health Research Institute; Western University","funders":"Canadian Institutes of Health Research","keywords":"Confidence interval; Physical therapy; Medicine; Kappa; Rehabilitation; Randomized controlled trial; Reliability (semiconductor); Stroke (engine); Operationalization; Internal medicine; Mathematics","score_opus":0.3121866161788805,"score_gpt":0.5622545008997024,"score_spread":0.2500678847208219,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2045910459","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.07919216,0.562416,0.14394279,0.009367411,0.007172824,0.10962049,0.032861434,0.0011976536,0.05422928],"genre_scores_gemma":[0.22352967,0.18915302,0.41465896,0.0037540023,0.0011944511,0.14790264,0.010549872,0.00022180924,0.009035655],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9446862,0.019470567,0.02335396,0.0013803181,0.01034987,0.00075915584],"domain_scores_gemma":[0.9254769,0.055896234,0.005458319,0.0014507726,0.011234971,0.00048270667],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.04530444,0.002111414,0.006080781,0.018261671,0.0018308809,0.0030476972,0.0017979848,0.0014220413,0.0051985187],"category_scores_gemma":[0.07798537,0.0011273216,0.0149246985,0.0133349,0.0020986032,0.0032496373,0.00402789,0.0024767476,0.00081515237],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018827928,0.00057511986,0.051200904,0.13197462,0.017788531,0.00042967752,0.00460818,0.0015691189,0.0013648164,0.018491792,0.038623847,0.73149055],"study_design_scores_gemma":[0.0056384285,0.0057107797,0.30670583,0.13525942,0.044055246,0.0036677793,0.010572752,0.006159906,0.0042036064,0.057110637,0.41967574,0.0012399817],"about_ca_topic_score_codex":0.0022933949,"about_ca_topic_score_gemma":0.008087569,"teacher_disagreement_score":0.9546956,"about_ca_system_score_codex":0.0034221185,"about_ca_system_score_gemma":0.007014369,"threshold_uncertainty_score":0.23959559},"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":["metaepi_broad"],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2059874779","doi":"10.1371/journal.pone.0020185","title":"Epidemiology, Quality and Reporting Characteristics of Systematic Reviews of Traditional Chinese Medicine Interventions Published in Chinese Journals","year":2011,"lang":"en","type":"review","venue":"PLoS ONE","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":58,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"University of Alberta; Lanzhou University","keywords":"Systematic review; Medicine; Psychological intervention; Epidemiology; MEDLINE; Alternative medicine; Specialty; Meta-analysis; Data extraction; Family medicine; Publication bias; China; Quality (philosophy); Pathology; Psychiatry","score_opus":0.9767140384802047,"score_gpt":0.6506685977850202,"score_spread":0.3260454406951845,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2059874779","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.16493022,0.74237084,0.020042324,0.007316955,0.0012895142,0.0146677,0.044991747,0.00028638536,0.0041043633],"genre_scores_gemma":[0.7375803,0.1557039,0.032727808,0.0035138465,0.0005651053,0.05056878,0.018593196,0.0001655572,0.0005814539],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.43300432,0.18713559,0.31155512,0.01705563,0.04899803,0.0022512693],"domain_scores_gemma":[0.19533685,0.4905107,0.2426273,0.01809739,0.051721934,0.0017059433],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.2641658,0.0014322795,0.0075837877,0.026786948,0.0016918667,0.0044487235,0.002703051,0.0022988643,0.002482377],"category_scores_gemma":[0.6263713,0.0019132822,0.011422614,0.042867564,0.002775518,0.0045463773,0.0037036636,0.0017275714,0.00030585093],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018380771,0.00014596841,0.40524164,0.43539143,0.081664786,0.00040453044,0.004941975,0.00090597686,0.00079984975,0.0017958687,0.007917139,0.05895275],"study_design_scores_gemma":[0.0021131888,0.001511029,0.6306634,0.2047367,0.117167674,0.002139208,0.00493693,0.0025942568,0.0023867413,0.0047388086,0.026472518,0.00053970475],"about_ca_topic_score_codex":0.0068549686,"about_ca_topic_score_gemma":0.009706409,"teacher_disagreement_score":0.9924162,"about_ca_system_score_codex":0.0076118214,"about_ca_system_score_gemma":0.016802717,"threshold_uncertainty_score":0.90741456},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2068264290","doi":"10.1002/jrsm.1123","title":"A scoping review of scoping reviews: advancing the approach and enhancing the consistency","year":2014,"lang":"en","type":"review","venue":"Research Synthesis Methods","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":3082,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Guelph; Public Health Agency of Canada","funders":"Ontario Ministry of Agriculture, Food and Rural Affairs","keywords":"Consistency (knowledge bases); Computer science; Management science; Data science; Artificial intelligence; Engineering","score_opus":0.923851977769396,"score_gpt":0.7277039479738097,"score_spread":0.19614802979558632,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2068264290","genre_codex":"methods","genre_gemma":"review","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.006441922,0.38655174,0.40669885,0.09350292,0.019493524,0.057676304,0.0019270387,0.001693878,0.026013883],"genre_scores_gemma":[0.027295852,0.18585876,0.7233301,0.0070421016,0.0032215642,0.050360754,0.0009764163,0.00042185848,0.0014925876],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.2911311,0.5107797,0.13201867,0.010007797,0.05435693,0.0017058355],"domain_scores_gemma":[0.1574939,0.5370502,0.070502825,0.043512326,0.18711802,0.0043227575],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.6293708,0.0037960547,0.01051723,0.072989285,0.008799515,0.030502886,0.0066136005,0.01221118,0.008340462],"category_scores_gemma":[0.78567797,0.00714655,0.010675426,0.057923447,0.013806924,0.028670564,0.02325455,0.012694324,0.003836775],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003331086,0.0001381601,0.0028688838,0.24194868,0.0030583323,0.0005756556,0.019093871,0.001052052,0.0015124388,0.046386607,0.053130455,0.6299017],"study_design_scores_gemma":[0.00051700155,0.00037852334,0.0035125974,0.7010743,0.004734544,0.00092283695,0.0064571137,0.002442288,0.00091081916,0.07201738,0.20656376,0.00046881486],"about_ca_topic_score_codex":0.00451733,"about_ca_topic_score_gemma":0.008169201,"teacher_disagreement_score":0.3706292,"about_ca_system_score_codex":0.01594701,"about_ca_system_score_gemma":0.09835402,"threshold_uncertainty_score":0.4570518},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2081069584","doi":"10.1097/01.sla.0000154258.30305.df","title":"Evaluating Meta-analyses in the General Surgical Literature","year":2005,"lang":"en","type":"article","venue":"Annals of Surgery","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":131,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; University of Calgary","funders":"","keywords":"Medicine; Meta-analysis; MEDLINE; Concordance; Critical appraisal; Systematic review; Evidence-based medicine; Family medicine; Alternative medicine; Internal medicine; Pathology","score_opus":0.9842320020590651,"score_gpt":0.6970802859876226,"score_spread":0.2871517160714425,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2081069584","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0181521,0.9118978,0.057526354,0.0017995328,0.0014408062,0.005342651,0.0022688354,0.00043226453,0.0011397927],"genre_scores_gemma":[0.5448792,0.2655601,0.1575279,0.002667588,0.00274993,0.021727607,0.00369929,0.00046707186,0.00072134],"study_design_codex":"meta_analysis","study_design_gemma":"observational","domain_scores_codex":[0.483548,0.4219535,0.063845485,0.012306376,0.017496979,0.0008496741],"domain_scores_gemma":[0.27533334,0.6573957,0.04118503,0.013373828,0.011973936,0.00073817914],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.35842234,0.0076727117,0.024135396,0.02801285,0.0010995803,0.008748183,0.0056518954,0.0043513365,0.0041513382],"category_scores_gemma":[0.6302723,0.0035196403,0.052261293,0.018770734,0.00215625,0.004355123,0.004506005,0.0027026667,0.0004557875],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002264504,0.000042241812,0.011558963,0.11278052,0.84704494,0.00025891527,0.00019741335,0.0038638155,0.00022062805,0.00052381895,0.00068696926,0.020557314],"study_design_scores_gemma":[0.0024684502,0.0010484961,0.010450013,0.024795199,0.9469015,0.00032528647,0.00017002775,0.0046562026,0.00056647713,0.005205336,0.0032683162,0.00014475718],"about_ca_topic_score_codex":0.003240314,"about_ca_topic_score_gemma":0.004826736,"teacher_disagreement_score":0.9758646,"about_ca_system_score_codex":0.0061507146,"about_ca_system_score_gemma":0.0057748533,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2081631580","doi":"10.1016/j.jmpt.2013.04.001","title":"Manipulative and Multimodal Therapy for Upper Extremity and Temporomandibular Disorders: A Systematic Review","year":2013,"lang":"en","type":"review","venue":"Journal of Manipulative and Physiological Therapeutics","topic":"Temporomandibular Joint Disorders","field":"Health Professions","cited_by":62,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Canadian Memorial Chiropractic College","funders":"","keywords":"Medicine; Chiropractic; Manual therapy; Physical therapy; Randomized controlled trial; MEDLINE; Orthopedic surgery; Physical medicine and rehabilitation; Alternative medicine; Surgery; Pathology","score_opus":0.4647909135108516,"score_gpt":0.4799292772683735,"score_spread":0.015138363757521867,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2081631580","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0015070369,0.9972166,0.00013054743,0.00021225614,0.00009479194,0.00046340298,0.00014063943,0.0000044992344,0.00023025821],"genre_scores_gemma":[0.01771733,0.9782808,0.0015618611,0.00059546507,0.00014469927,0.0013260654,0.00017351926,0.0000035918006,0.0001966274],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9877868,0.004589035,0.004190713,0.0007336174,0.0024009924,0.00029874037],"domain_scores_gemma":[0.9687553,0.019925987,0.0076432005,0.00043215332,0.002923262,0.00032008014],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01110537,0.0014878028,0.007553093,0.01406815,0.00093846826,0.0025944752,0.002013242,0.002559384,0.003842407],"category_scores_gemma":[0.034021717,0.001130486,0.005403862,0.012493005,0.00096165034,0.0029901124,0.001454551,0.00094637624,0.00029946928],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011150245,0.000026401716,0.000743955,0.96477383,0.004276999,0.0001472481,0.00022759345,0.00006526242,0.00020898557,0.00014908856,0.0009829783,0.028286226],"study_design_scores_gemma":[0.0002359408,0.0003406393,0.004840669,0.94529146,0.03509478,0.0007682822,0.00043710344,0.00010805927,0.00019473462,0.00016366584,0.012491669,0.00003305379],"about_ca_topic_score_codex":0.0041283886,"about_ca_topic_score_gemma":0.021136587,"teacher_disagreement_score":0.01406815,"about_ca_system_score_codex":0.0031669454,"about_ca_system_score_gemma":0.016426938,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2096833901","doi":"10.1371/journal.pone.0088008","title":"Quantifying Bias in Randomized Controlled Trials in Child Health: A Meta-Epidemiological Study","year":2014,"lang":"en","type":"review","venue":"PLoS ONE","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":43,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"Canadian Institutes of Health Research","keywords":"Epidemiology; Randomized controlled trial; Meta-analysis; Medicine; MEDLINE; Environmental health; Biology; Internal medicine","score_opus":0.9817470593674227,"score_gpt":0.6519590330899844,"score_spread":0.3297880262774383,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2096833901","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.006130174,0.940438,0.03795591,0.0021190536,0.0019447374,0.008888373,0.0014823488,0.00019421407,0.00084726705],"genre_scores_gemma":[0.33332133,0.4003554,0.18528965,0.007984627,0.0030416558,0.0657717,0.0029564346,0.0003350667,0.00094413967],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.4616793,0.40944183,0.08712559,0.013251381,0.027090672,0.0014112779],"domain_scores_gemma":[0.31093597,0.5739766,0.07886933,0.019260075,0.015947806,0.0010102844],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.34036097,0.005669951,0.025653552,0.020020392,0.0015299567,0.011128466,0.005363339,0.0075594517,0.004413481],"category_scores_gemma":[0.6200874,0.003485277,0.059141044,0.0146045275,0.0034371123,0.009317963,0.006061543,0.0055970596,0.00056206423],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0031755727,0.000086519176,0.010861536,0.43097952,0.5127032,0.00020096476,0.0004910898,0.0020172824,0.0003232377,0.0023566366,0.0021111441,0.034693345],"study_design_scores_gemma":[0.0054468536,0.0013580457,0.010424412,0.20892811,0.7509107,0.0006837572,0.00027660938,0.003011194,0.0008327457,0.010264837,0.007593318,0.00026934542],"about_ca_topic_score_codex":0.002711654,"about_ca_topic_score_gemma":0.0038275586,"teacher_disagreement_score":0.99433005,"about_ca_system_score_codex":0.0071727172,"about_ca_system_score_gemma":0.00971658,"threshold_uncertainty_score":0.81345236},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2102685275","doi":"10.1186/2046-4053-2-24","title":"A protocol for a systematic review on the impact of unpublished studies and studies published in the gray literature in meta-analyses","year":2013,"lang":"en","type":"review","venue":"Systematic Reviews","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":41,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"European Commission; McMaster University","keywords":"Medicine; Grey literature; Meta-analysis; Systematic review; MEDLINE; Cochrane Library; Publication bias; Gray (unit); Protocol (science); Inclusion and exclusion criteria; Gerontology; Alternative medicine; Pathology","score_opus":0.9180444448269799,"score_gpt":0.6734652623171064,"score_spread":0.24457918250987354,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2102685275","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":"reproducibility","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":"reproducibility","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0003097001,0.00085060013,0.006112371,0.0007059421,0.00077828416,0.9820287,0.0076718605,0.00050703547,0.0010355135],"genre_scores_gemma":[0.00030367414,0.00024901467,0.011893384,0.00018182148,0.000035325436,0.98639745,0.0005484026,0.000036271715,0.00035462008],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.83207244,0.09293342,0.052360065,0.008561182,0.010814992,0.0032579482],"domain_scores_gemma":[0.6479911,0.17630118,0.044095084,0.042972002,0.0817238,0.0069168326],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.18266788,0.007809128,0.016606184,0.017619016,0.0051299697,0.007837617,0.007398313,0.013803445,0.14008932],"category_scores_gemma":[0.2992265,0.009284394,0.018966366,0.019400254,0.0070434366,0.010090238,0.0063706897,0.014677461,0.028634207],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.016322589,0.00083393644,0.0010489795,0.6758685,0.004350846,0.0011839081,0.006222287,0.0031975857,0.003808524,0.017262366,0.18028294,0.089617625],"study_design_scores_gemma":[0.06716193,0.002404804,0.005744989,0.3797173,0.0064353403,0.0008537641,0.003407267,0.0053569614,0.005952225,0.05232639,0.4692749,0.001364252],"about_ca_topic_score_codex":0.0032669103,"about_ca_topic_score_gemma":0.0061354428,"teacher_disagreement_score":0.81733215,"about_ca_system_score_codex":0.014105277,"about_ca_system_score_gemma":0.0665527,"threshold_uncertainty_score":0.96605164},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2102903042","doi":"10.1186/2046-4053-2-88","title":"How should we evaluate the risk of bias of physical therapy trials?: a psychometric and meta-epidemiological approach towards developing guidelines for the design, conduct, and reporting of RCTs in Physical Therapy (PT) area: a study protocol","year":2013,"lang":"en","type":"article","venue":"Systematic Reviews","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"Alberta Innovates; Killam Trusts; Fondation pour la Recherche Médicale; University of Alberta; Canadian Institutes of Health Research; Physiotherapy Foundation of Canada; Alberta Innovates - Health Solutions; Women and Children's Health Research Institute; Children's Health Research Institute","keywords":"Medicine; Epidemiology; Randomized controlled trial; Meta-analysis; Physical therapy; Research design; Pathology","score_opus":0.9844898175982594,"score_gpt":0.6756415085022591,"score_spread":0.30884830909600025,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2102903042","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.001863103,0.030336808,0.3020448,0.019426148,0.004879871,0.63449985,0.0035852247,0.0009192904,0.0024450566],"genre_scores_gemma":[0.0045135175,0.0041653817,0.3376359,0.002450912,0.0003051648,0.65024734,0.00033955462,0.00007362436,0.00026863377],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.15907481,0.66717416,0.13673091,0.009322339,0.02627817,0.0014196612],"domain_scores_gemma":[0.18704502,0.6071325,0.06934068,0.05581519,0.0767237,0.003942981],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.7681447,0.008192635,0.020911867,0.022316914,0.0046431324,0.01612566,0.010608668,0.022420341,0.006627773],"category_scores_gemma":[0.8262535,0.0077023427,0.030500459,0.017173659,0.010028827,0.016841281,0.011334057,0.015659532,0.003375507],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008997284,0.0007992442,0.0075340588,0.46710053,0.04764693,0.00075780763,0.0068210918,0.008259258,0.0019457376,0.06502827,0.046931785,0.33817792],"study_design_scores_gemma":[0.022439374,0.0042250506,0.00846829,0.5082804,0.06114203,0.0010184557,0.0020332036,0.031763136,0.0071940757,0.17764623,0.17442355,0.001366173],"about_ca_topic_score_codex":0.0023515788,"about_ca_topic_score_gemma":0.00336157,"teacher_disagreement_score":0.9790881,"about_ca_system_score_codex":0.012972182,"about_ca_system_score_gemma":0.054435827,"threshold_uncertainty_score":0.2859189},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2103669116","doi":"10.1186/1748-5908-9-54","title":"Applications of social constructivist learning theories in knowledge translation for healthcare professionals: a scoping review","year":2014,"lang":"en","type":"review","venue":"Implementation Science","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":326,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Interdisciplinary Research in Rehabilitation; McGill University","funders":"McGill University","keywords":"CINAHL; Constructivism (international relations); Social constructivism; Knowledge translation; PsycINFO; Health care; Constructivist teaching methods; Psychological intervention; Context (archaeology); Epistemology; MEDLINE; Medicine; Psychology; Sociology; Knowledge management; Social science; Nursing; Computer science; Pedagogy; Politics; Teaching method; Political science","score_opus":0.7546677192685569,"score_gpt":0.793139881474038,"score_spread":0.03847216220548111,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2103669116","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0015955426,0.97843677,0.005891298,0.004822183,0.00094114756,0.0033938766,0.0001476115,0.000033805867,0.0047377385],"genre_scores_gemma":[0.014250434,0.96448565,0.013706057,0.0010812738,0.00027879103,0.005651288,0.00017300229,0.000017562039,0.00035593344],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.90522933,0.050632395,0.022311289,0.0027407706,0.017813528,0.001272714],"domain_scores_gemma":[0.6564318,0.29425025,0.017282348,0.00495603,0.025889013,0.0011906064],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.10692072,0.0024386367,0.006365612,0.056694053,0.0042364006,0.01620164,0.00422837,0.006640937,0.0045891907],"category_scores_gemma":[0.23974833,0.0022587455,0.0066211703,0.05185099,0.00670621,0.017442206,0.007712543,0.004174962,0.0010101815],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000093301234,0.000106073785,0.0012504394,0.5219729,0.0013364612,0.00042096112,0.00990647,0.0007309541,0.0002816698,0.014857983,0.0059713526,0.4430714],"study_design_scores_gemma":[0.000050835562,0.000081445294,0.0008899752,0.95212406,0.001409944,0.00023788266,0.0037802274,0.0004251671,0.00020142166,0.0066223657,0.034131918,0.00004481852],"about_ca_topic_score_codex":0.007702711,"about_ca_topic_score_gemma":0.015412355,"teacher_disagreement_score":0.10692072,"about_ca_system_score_codex":0.017142916,"about_ca_system_score_gemma":0.06359052,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"opus","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2103936380","doi":"10.1146/annurev-publhealth-031811-124540","title":"Advertising of Prescription-Only Medicines to the Public: Does Evidence of Benefit Counterbalance Harm?","year":2012,"lang":"en","type":"review","venue":"Annual Review of Public Health","topic":"Pharmaceutical industry and healthcare","field":"Pharmacology, Toxicology and Pharmaceutics","cited_by":76,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Harm; Scrutiny; Medical prescription; Direct-to-consumer advertising; Causality (physics); Medicine; Public health; Public economics; Health care; Business; Family medicine; Psychology; Nursing; Economics; Political science","score_opus":0.5564649027992554,"score_gpt":0.6081295504974032,"score_spread":0.051664647698147825,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2103936380","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00044205764,0.99269485,0.00013815178,0.004718446,0.0003053431,0.000024137853,0.000040844287,0.000003321464,0.0016328392],"genre_scores_gemma":[0.012498842,0.9825117,0.0005174323,0.0034654797,0.0006629531,0.000041514722,0.000039221064,0.0000056196286,0.0002573644],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9878853,0.00651639,0.0020186855,0.000720733,0.0025702955,0.00028850732],"domain_scores_gemma":[0.9217529,0.06666159,0.006488428,0.0012187593,0.0034705088,0.00040777022],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.022278467,0.0010545112,0.0066994447,0.0036126184,0.00049771584,0.0038946357,0.0020662143,0.004242268,0.011503995],"category_scores_gemma":[0.060400743,0.0007178789,0.0029321692,0.004396198,0.002645354,0.0046722796,0.0016074751,0.003388994,0.0012610838],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012140204,0.000106542095,0.0008448517,0.23422511,0.0066843475,0.0001925965,0.00030736503,0.0001743263,0.00034734057,0.0126473075,0.009620404,0.73363584],"study_design_scores_gemma":[0.0020198692,0.001463852,0.009564488,0.48797497,0.02839549,0.0013608838,0.0012951904,0.0002940357,0.0011767453,0.030087011,0.4362551,0.00011240598],"about_ca_topic_score_codex":0.0023777662,"about_ca_topic_score_gemma":0.0037696054,"teacher_disagreement_score":0.022278467,"about_ca_system_score_codex":0.0022639881,"about_ca_system_score_gemma":0.004458663,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2111376665","doi":"10.1177/0306624x12447771","title":"A Meta-Analysis of Moral Reconation Therapy","year":2012,"lang":"en","type":"review","venue":"International Journal of Offender Therapy and Comparative Criminology","topic":"Psychology of Moral and Emotional Judgment","field":"Neuroscience","cited_by":76,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Saskatchewan","funders":"","keywords":"Psychology; Medicine; Psychotherapist","score_opus":0.9301869343671447,"score_gpt":0.47731096573716825,"score_spread":0.4528759686299765,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2111376665","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0050451024,0.9916971,0.00081371155,0.00036880496,0.00042259792,0.00038178157,0.00054608076,0.000041779298,0.00068314996],"genre_scores_gemma":[0.2002918,0.78915673,0.004532577,0.0020668334,0.0005342882,0.0013106919,0.0012444308,0.00007873486,0.000784011],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.98665327,0.0070603425,0.0033093146,0.0011696503,0.0015076348,0.00029981125],"domain_scores_gemma":[0.97637886,0.017296702,0.0029988901,0.0010738849,0.0019377965,0.00031382646],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.017482843,0.002575074,0.016428802,0.009863012,0.0008854513,0.0043786997,0.0017161351,0.0020275742,0.0041433442],"category_scores_gemma":[0.04541466,0.0013024762,0.037348963,0.008093194,0.0006719473,0.0022396871,0.0015463023,0.002472281,0.0003860627],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0021928293,0.00003705094,0.0018128649,0.23540394,0.7421693,0.00014348888,0.00008752424,0.00038085296,0.0002850806,0.00021127255,0.0008893239,0.01638638],"study_design_scores_gemma":[0.000635297,0.00023047466,0.002507068,0.019275013,0.9742621,0.000110866364,0.00004579322,0.00009697836,0.00016625816,0.00028133567,0.0023672422,0.000021576545],"about_ca_topic_score_codex":0.0062565263,"about_ca_topic_score_gemma":0.014815534,"teacher_disagreement_score":0.017482843,"about_ca_system_score_codex":0.0035855426,"about_ca_system_score_gemma":0.0040900116,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaepi_narrow","metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2125820775","doi":"10.1002/j.0022-0337.2004.68.9.tb03849.x","title":"Implementing Evidence‐Based Practice in Undergraduate Teaching Clinics: A Systematic Review and Recommendations","year":2004,"lang":"en","type":"review","venue":"Journal of Dental Education","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":55,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Evidence-based dentistry; Medical education; Curriculum; Systematic review; Evidence-based practice; Psychological intervention; Evidence-based medicine; MEDLINE; Quality (philosophy); Psychology; Health care; Medicine; Alternative medicine; Nursing; Pedagogy","score_opus":0.25863639948707423,"score_gpt":0.6457552192344075,"score_spread":0.38711881974733325,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2125820775","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0043694917,0.9491611,0.005207142,0.030533064,0.0021491242,0.0065812264,0.00031080263,0.00009962387,0.0015884537],"genre_scores_gemma":[0.0296523,0.8949983,0.060373444,0.0053810803,0.0004564536,0.008391188,0.0002831205,0.000018112722,0.000445942],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.92870903,0.03887318,0.020216756,0.0018978879,0.009255877,0.0010473005],"domain_scores_gemma":[0.7767686,0.16285913,0.024851756,0.0028519388,0.028651481,0.0040170196],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.10594417,0.0027596622,0.007956156,0.016114913,0.00184063,0.0070314347,0.0061067934,0.007584947,0.0039372975],"category_scores_gemma":[0.21550952,0.0028621817,0.0064793704,0.012156516,0.0020849372,0.010975058,0.0036470732,0.0045558307,0.0007701713],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00040668715,0.00032226354,0.0037005355,0.7164286,0.00474873,0.00039524978,0.0022723288,0.0008617117,0.00038166268,0.0018261357,0.009540971,0.2591151],"study_design_scores_gemma":[0.00054272584,0.00028916544,0.0035099748,0.9600443,0.01077395,0.000208736,0.0021862206,0.00075786316,0.00033860782,0.0018688969,0.019394284,0.00008524798],"about_ca_topic_score_codex":0.012737942,"about_ca_topic_score_gemma":0.04836546,"teacher_disagreement_score":0.89405584,"about_ca_system_score_codex":0.011464208,"about_ca_system_score_gemma":0.061023295,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2146200566","doi":"10.1016/j.annemergmed.2008.05.036","title":"The Relative Efficacy of Meperidine for the Treatment of Acute Migraine: A Meta-analysis of Randomized Controlled Trials","year":2008,"lang":"en","type":"review","venue":"Annals of Emergency Medicine","topic":"Migraine and Headache Studies","field":"Medicine","cited_by":49,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"National Institute of Neurological Disorders and Stroke","keywords":"Medicine; Dihydroergotamine; Migraine; Randomized controlled trial; Odds ratio; Relative risk; Anesthesia; Adverse effect; Clinical trial; Antiemetic; Confidence interval; Meta-analysis; Ketorolac; Triptans; Nausea; Internal medicine; Analgesic","score_opus":0.5200886805233103,"score_gpt":0.5377539528548487,"score_spread":0.017665272331538473,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2146200566","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.013624568,0.9844342,0.00081315945,0.00015158857,0.00024616846,0.00020387015,0.00024718803,0.000031880136,0.00024738946],"genre_scores_gemma":[0.27943632,0.7143302,0.0035392821,0.0007300561,0.00044206955,0.0005088776,0.0005902245,0.00004648806,0.00037645514],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.99053407,0.005168496,0.0023776996,0.0009111282,0.0008162305,0.00019242331],"domain_scores_gemma":[0.9737959,0.022072477,0.0025402687,0.00062440155,0.000706354,0.00026053912],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01562988,0.0031694225,0.020565866,0.0034457406,0.00052065746,0.0027537213,0.0024885044,0.0023231315,0.0031994844],"category_scores_gemma":[0.03378471,0.0016830339,0.029218644,0.0034955568,0.0009720687,0.0020604038,0.0013574001,0.0028042987,0.00023473352],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.018740648,0.00005244987,0.0019883954,0.1458704,0.8166054,0.000094993644,0.00007001508,0.0006206483,0.00033060578,0.0001297721,0.00036193003,0.015134768],"study_design_scores_gemma":[0.0046730856,0.0005375714,0.0021183419,0.004311289,0.9869271,0.000079297184,0.000032199405,0.00019187534,0.00014733973,0.00022613916,0.0007348185,0.00002098775],"about_ca_topic_score_codex":0.0030132355,"about_ca_topic_score_gemma":0.0055928067,"teacher_disagreement_score":0.020565866,"about_ca_system_score_codex":0.0018299051,"about_ca_system_score_gemma":0.0014814876,"threshold_uncertainty_score":0.08265966},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2150288182","doi":"10.1016/s0140-6736(14)61133-8","title":"Health interventions for people who are homeless","year":2014,"lang":"en","type":"review","venue":"The Lancet","topic":"Homelessness and Social Issues","field":"Health Professions","cited_by":314,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; St. Michael's Hospital","funders":"","keywords":"Respite care; Psychological intervention; Assertive community treatment; Mental health; Medicine; Intervention (counseling); Nursing; Case management; Health care; Mental illness; Psychiatry","score_opus":0.23725616397547042,"score_gpt":0.5367720883114878,"score_spread":0.29951592433601737,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2150288182","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00006915737,0.99865013,0.000026277105,0.00056653767,0.00024835105,0.000012468793,0.000024730627,0.0000028363147,0.00039942758],"genre_scores_gemma":[0.0023494614,0.99589396,0.00018913383,0.0010383019,0.00020417348,0.00003451627,0.000019577628,0.0000017690734,0.00026910193],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9990742,0.00035118306,0.00015819825,0.00009491971,0.0002325226,0.00008899653],"domain_scores_gemma":[0.9967367,0.0025360244,0.00039245875,0.000037423415,0.00018704607,0.00011033498],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0024555216,0.0011435879,0.00265921,0.0027244526,0.00039288244,0.0012958063,0.0009854457,0.002551945,0.009433946],"category_scores_gemma":[0.007888625,0.00042535935,0.001562252,0.002401098,0.0006758644,0.0010004968,0.0010297965,0.0024118493,0.0005834183],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00023041929,0.00009866605,0.0002217206,0.24384628,0.001131645,0.00008246223,0.00016557479,0.00022318015,0.00015894877,0.0023137366,0.028469423,0.723058],"study_design_scores_gemma":[0.0011946793,0.00045442054,0.003925896,0.4783777,0.008829359,0.0005537494,0.0008439761,0.00015049202,0.00027677874,0.004226772,0.5011221,0.000044136388],"about_ca_topic_score_codex":0.008188681,"about_ca_topic_score_gemma":0.0253768,"teacher_disagreement_score":0.009433946,"about_ca_system_score_codex":0.0019278058,"about_ca_system_score_gemma":0.0045942366,"threshold_uncertainty_score":0.031559706},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2151318777","doi":"10.1177/1740774509344440","title":"Constructing common cohorts from trials with overlapping eligibility criteria: implications for comparing effect sizes between trials","year":2009,"lang":"en","type":"article","venue":"Clinical Trials","topic":"Diabetes, Cardiovascular Risks, and Lipoproteins","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St Joseph's Health Care","funders":"National Institute of Diabetes and Digestive and Kidney Diseases; National Heart, Lung, and Blood Institute; National Institutes of Health","keywords":"Medicine; Glycemic; Clinical trial; Psychological intervention; Disease; Type 2 diabetes; Randomized controlled trial; Diabetes mellitus; Research design; Physical therapy; Gerontology; Internal medicine","score_opus":0.3252983017345073,"score_gpt":0.5344054550213949,"score_spread":0.20910715328688761,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2151318777","genre_codex":"methods","genre_gemma":"methods","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.07530681,0.032580543,0.76833683,0.021340825,0.010827041,0.07204284,0.0051482837,0.0024284332,0.011988397],"genre_scores_gemma":[0.35202456,0.0026631446,0.5290017,0.009448619,0.0019983046,0.100907914,0.0025846965,0.00054933416,0.0008217051],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.05431388,0.82251203,0.058459196,0.030725304,0.03166127,0.002328302],"domain_scores_gemma":[0.022402681,0.8696253,0.03159876,0.06652544,0.008333332,0.0015144512],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.8684059,0.005842921,0.018765664,0.02053959,0.0062790625,0.018658021,0.017541077,0.013095228,0.006987144],"category_scores_gemma":[0.9482622,0.0056001074,0.03273985,0.017662432,0.02413386,0.0187328,0.024980582,0.017388793,0.0012408562],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.03048276,0.0015053982,0.23055921,0.033762414,0.17120332,0.0027580247,0.020602386,0.02321961,0.0016383987,0.115111716,0.024181252,0.34497547],"study_design_scores_gemma":[0.040594127,0.016795378,0.1922071,0.03762732,0.09021999,0.0029629422,0.006457715,0.12898892,0.0049653808,0.40170738,0.075495906,0.0019777743],"about_ca_topic_score_codex":0.0071663563,"about_ca_topic_score_gemma":0.0054310113,"teacher_disagreement_score":0.13159412,"about_ca_system_score_codex":0.01057899,"about_ca_system_score_gemma":0.01686413,"threshold_uncertainty_score":0.16227907},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2160194973","doi":"10.1586/14737167.8.3.273","title":"Systematic review of economic evaluations of Alzheimer’s disease medications","year":2008,"lang":"en","type":"article","venue":"Expert Review of Pharmacoeconomics & Outcomes Research","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":21,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; McMaster University Medical Centre","funders":"","keywords":"Pharmacoeconomics; Medicine; Disease; Economic evaluation; Intensive care medicine; Cost effectiveness; Health economics; Actuarial science; Economics; Risk analysis (engineering); Public health; Pathology","score_opus":0.3929277369869546,"score_gpt":0.6398002024196178,"score_spread":0.24687246543266317,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2160194973","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0030395084,0.9920123,0.00074202183,0.0005717064,0.000229173,0.0009876495,0.0013445682,0.000016103657,0.0010568689],"genre_scores_gemma":[0.05019211,0.94431514,0.0023411806,0.0005062719,0.00015001536,0.0016035276,0.0006373652,0.000012421471,0.00024186914],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9647645,0.020011837,0.008156206,0.0009779605,0.0055312403,0.0005582988],"domain_scores_gemma":[0.91256773,0.07044881,0.0103192,0.0009244546,0.0052621574,0.0004777952],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.026150348,0.0022710098,0.0111964345,0.015851995,0.00062538334,0.0032996156,0.0017477126,0.0018862466,0.00637399],"category_scores_gemma":[0.13757251,0.0012276828,0.009837596,0.01430091,0.00092910475,0.0020382535,0.0014726537,0.0015303245,0.0003462472],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008272488,0.000037913855,0.0009584629,0.90750486,0.0377332,0.00016810834,0.000092179405,0.0008292879,0.000083130755,0.0008192001,0.0020925642,0.04885379],"study_design_scores_gemma":[0.0016383672,0.0005416903,0.0071939235,0.7880786,0.17517579,0.0005458473,0.00023316414,0.000787268,0.00030101454,0.0018713498,0.023566077,0.0000668551],"about_ca_topic_score_codex":0.0075987424,"about_ca_topic_score_gemma":0.019426337,"teacher_disagreement_score":0.026150348,"about_ca_system_score_codex":0.007515929,"about_ca_system_score_gemma":0.011966447,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2164404559","doi":"10.1136/bmjopen-2014-005297","title":"Interpreting trial results following use of different intention-to-treat approaches for preventing attrition bias: a meta-epidemiological study protocol","year":2014,"lang":"en","type":"article","venue":"BMJ Open","topic":"Advanced Causal Inference Techniques","field":"Mathematics","cited_by":21,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"Parker Institute for Cancer Immunotherapy; Oak Foundation","keywords":"Medicine; Protocol (science); Attrition; Epidemiology; Selection bias; Publication bias; Alternative medicine; Meta-analysis; Environmental health; Pathology","score_opus":0.8468178363736624,"score_gpt":0.5865171235954936,"score_spread":0.2603007127781688,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2164404559","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0010408907,0.0010347948,0.04775416,0.0009540493,0.0018322314,0.9420783,0.003314861,0.00081436307,0.0011763771],"genre_scores_gemma":[0.0014028169,0.00017918232,0.027719384,0.00026143002,0.00006183653,0.9698574,0.00025840692,0.00004161815,0.000217945],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.54440296,0.34130138,0.076611534,0.017654557,0.017225143,0.0028044123],"domain_scores_gemma":[0.60683924,0.21771897,0.050441064,0.07725621,0.043380693,0.0043638097],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.36633945,0.008920852,0.015082846,0.0118787335,0.0026640175,0.00762315,0.005937747,0.019838732,0.03449022],"category_scores_gemma":[0.5241475,0.0060673426,0.022887312,0.011838195,0.0068304506,0.0071342457,0.004395207,0.017662292,0.013579774],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.12153967,0.0026936934,0.005643968,0.42069325,0.055879924,0.0016332599,0.007095133,0.014072113,0.006710199,0.07735243,0.10666287,0.18002355],"study_design_scores_gemma":[0.25204304,0.013221734,0.014405627,0.12759033,0.05286732,0.0013696677,0.0010068043,0.07255606,0.017332098,0.19788301,0.24734941,0.002374888],"about_ca_topic_score_codex":0.00096050743,"about_ca_topic_score_gemma":0.001170588,"teacher_disagreement_score":0.98491716,"about_ca_system_score_codex":0.0062880307,"about_ca_system_score_gemma":0.019142982,"threshold_uncertainty_score":0.7814163},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2165271154","doi":"10.1093/pubmed/fdu050","title":"The benefits and challenges of conducting an overview of systematic reviews in public health: a focus on physical activity","year":2014,"lang":"en","type":"article","venue":"Journal of Public Health","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":53,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Systematic review; Public health; MEDLINE; Alternative medicine; Medicine; Scientific literature; Cochrane collaboration; Management science; Psychology; Engineering ethics; Political science; Nursing; Engineering","score_opus":0.9583424756036388,"score_gpt":0.5984104860565596,"score_spread":0.3599319895470793,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2165271154","genre_codex":"review","genre_gemma":"review","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000999331,0.81952876,0.037813637,0.12884018,0.00791604,0.0024317927,0.0003604516,0.00031984778,0.0017898965],"genre_scores_gemma":[0.036913525,0.6268124,0.27074078,0.040183377,0.01337565,0.010586238,0.0003202841,0.00021653264,0.00085113035],"study_design_codex":"systematic_review","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.2066208,0.63962835,0.09775721,0.0083788745,0.045690924,0.0019238835],"domain_scores_gemma":[0.047844246,0.8716051,0.034509122,0.012582793,0.031148266,0.0023104816],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.60692066,0.0060727946,0.018287532,0.03337557,0.004204603,0.023403363,0.006926112,0.021607261,0.004830643],"category_scores_gemma":[0.8219074,0.007532091,0.013997399,0.03153565,0.0144702215,0.031367272,0.012304644,0.01581724,0.001590015],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0015481346,0.0001410618,0.0029468408,0.5570902,0.016893683,0.00049703405,0.0046837595,0.0020734475,0.0015150964,0.042167168,0.039785765,0.3306578],"study_design_scores_gemma":[0.00090501463,0.0007590045,0.0034746674,0.8052489,0.017847285,0.0009804523,0.0019032954,0.0031444042,0.00064003945,0.08843543,0.07620823,0.0004532214],"about_ca_topic_score_codex":0.007016937,"about_ca_topic_score_gemma":0.019762712,"teacher_disagreement_score":0.98171246,"about_ca_system_score_codex":0.01603956,"about_ca_system_score_gemma":0.052651316,"threshold_uncertainty_score":0.48473686},"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":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2167473079","doi":"10.1093/humrep/des342","title":"Methodological quality of systematic reviews in subfertility: a comparison of Cochrane and non-Cochrane systematic reviews in assisted reproductive technologies","year":2012,"lang":"en","type":"article","venue":"Human Reproduction","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":50,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"University of Auckland","keywords":"Systematic review; Data extraction; Cochrane Library; Medicine; MEDLINE; Meta-analysis; Biology; Pathology","score_opus":0.9080448907257459,"score_gpt":0.6256176361119402,"score_spread":0.2824272546138057,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2167473079","genre_codex":"review","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.012531347,0.96600336,0.006695448,0.0033688522,0.0012641343,0.0076105935,0.0014271301,0.0000677813,0.0010313333],"genre_scores_gemma":[0.40400854,0.46793228,0.05539739,0.00657769,0.0018972973,0.06072375,0.0026451286,0.00021122042,0.000606701],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.28344116,0.3696002,0.26293993,0.01722421,0.06441195,0.0023825783],"domain_scores_gemma":[0.16681337,0.6417226,0.13040279,0.013813671,0.045219198,0.0020282974],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.44272968,0.0027710842,0.024128508,0.04279866,0.0025618332,0.014578125,0.004940976,0.005219908,0.003906506],"category_scores_gemma":[0.7675548,0.0039727716,0.03675707,0.04055701,0.005892677,0.011739422,0.0102426205,0.004513194,0.00038020464],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0024007442,0.000066318295,0.016076732,0.8010561,0.124498256,0.00015678888,0.0039439737,0.00038961347,0.00027957145,0.0019451892,0.002163959,0.04702278],"study_design_scores_gemma":[0.0052638333,0.0016708537,0.049964823,0.694963,0.22046854,0.0011234824,0.0041360594,0.0012860781,0.00066335103,0.006370384,0.013670462,0.0004191518],"about_ca_topic_score_codex":0.008197205,"about_ca_topic_score_gemma":0.015069549,"teacher_disagreement_score":0.9758715,"about_ca_system_score_codex":0.019254113,"about_ca_system_score_gemma":0.025823258,"threshold_uncertainty_score":0.68721354},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2204701188","doi":"10.7554/elife.08351","title":"A meta-analysis of threats to valid clinical inference in preclinical research of sunitinib","year":2015,"lang":"en","type":"review","venue":"eLife","topic":"Colorectal Cancer Treatments and Studies","field":"Medicine","cited_by":42,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital; McGill University","funders":"Canadian Institutes of Health Research","keywords":"Sunitinib; Meta-analysis; Inference; Preclinical research; Computational biology; Research integrity; Medicine; Bioinformatics; Biology; Computer science; Internal medicine; Artificial intelligence; Engineering ethics; Cancer","score_opus":0.9145509575040137,"score_gpt":0.7022478235203603,"score_spread":0.2123031339836534,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2204701188","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.011691997,0.8652422,0.09817827,0.0072628423,0.002739592,0.0060928147,0.0043282183,0.00044765152,0.004016421],"genre_scores_gemma":[0.5944965,0.20143928,0.1392537,0.017411951,0.0024021647,0.036349043,0.0053735324,0.0006084258,0.0026652985],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.55211943,0.34026462,0.07111028,0.014369373,0.020607663,0.0015285987],"domain_scores_gemma":[0.30931738,0.62093294,0.031010047,0.029978313,0.008152569,0.00060876587],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.29784873,0.0030444087,0.013197815,0.007467842,0.0012891699,0.00651203,0.003944298,0.0049849767,0.0047526504],"category_scores_gemma":[0.5620173,0.002253427,0.034802843,0.008327059,0.0024108216,0.0045262203,0.0035793772,0.004449264,0.0005722983],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0033854635,0.00006839453,0.011496456,0.23847015,0.64934486,0.0005233676,0.0005096432,0.0029410722,0.0005786106,0.007473549,0.0054309606,0.07977755],"study_design_scores_gemma":[0.0030399174,0.0007837942,0.012133531,0.070256114,0.84573376,0.00053778273,0.00024414907,0.0041464847,0.0016454308,0.03346519,0.027817037,0.0001967984],"about_ca_topic_score_codex":0.0026367202,"about_ca_topic_score_gemma":0.0042414884,"teacher_disagreement_score":0.98680216,"about_ca_system_score_codex":0.003666808,"about_ca_system_score_gemma":0.0059971283,"threshold_uncertainty_score":0.8658775},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2287844929","doi":"10.1136/bmjopen-2015-008562","title":"What is the influence of randomisation sequence generation and allocation concealment on treatment effects of physical therapy trials? A meta-epidemiological study","year":2015,"lang":"en","type":"review","venue":"BMJ Open","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":81,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Alberta Health; University of Alberta","funders":"Canadian Institutes of Health Research; Alberta Innovates; University of Alberta; Physiotherapy Foundation of Canada","keywords":"Medicine; Epidemiology; Meta-analysis; Alternative medicine; Sequence (biology); Public health; MEDLINE; Physical therapy; Gerontology; Internal medicine; Pathology","score_opus":0.9578471903528902,"score_gpt":0.6968038048601565,"score_spread":0.26104338549273376,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2287844929","genre_codex":"review","genre_gemma":"review","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015737727,0.9198148,0.047500648,0.006998532,0.0042574103,0.0031674418,0.00070140336,0.0002726943,0.0015493124],"genre_scores_gemma":[0.6447996,0.23447321,0.081069686,0.013900658,0.0065590492,0.016524483,0.0011288385,0.0005556799,0.0009887764],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.3308317,0.5555229,0.07079982,0.01606487,0.024968108,0.0018125728],"domain_scores_gemma":[0.24725217,0.6553109,0.059141695,0.026970878,0.01034998,0.00097434054],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.45749414,0.004957609,0.024495222,0.0068325954,0.0015173503,0.009728903,0.005008601,0.010862487,0.002878586],"category_scores_gemma":[0.66299635,0.004271474,0.060000505,0.008313242,0.005012969,0.012107935,0.0036267736,0.0078092725,0.00062806974],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.014915297,0.0001558367,0.015954033,0.12495721,0.76926947,0.0002375494,0.00067080895,0.0035202121,0.0004354298,0.0031851244,0.0022467626,0.064452335],"study_design_scores_gemma":[0.0139469765,0.0026865022,0.015490532,0.07194759,0.86672336,0.00045941045,0.00024107097,0.0067670564,0.0012835673,0.012716955,0.0074215154,0.0003154922],"about_ca_topic_score_codex":0.003276558,"about_ca_topic_score_gemma":0.0035138435,"teacher_disagreement_score":0.9950424,"about_ca_system_score_codex":0.0050606276,"about_ca_system_score_gemma":0.0051739053,"threshold_uncertainty_score":0.66900635},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2316188450","doi":"10.9778/cmajo.20150036","title":"Current use of routinely collected health data to complement randomized controlled trials: a meta-epidemiological survey","year":2016,"lang":"en","type":"article","venue":"CMAJ Open","topic":"Advanced Causal Inference Techniques","field":"Mathematics","cited_by":27,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Epidemiology; Randomized controlled trial; Complement (music); Medicine; Environmental health; Internal medicine; Biology","score_opus":0.9025846780532691,"score_gpt":0.6125923435989659,"score_spread":0.2899923344543033,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2316188450","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.040683992,0.9160859,0.022336142,0.007723218,0.0010411717,0.0026157862,0.0068388204,0.00021826707,0.0024566853],"genre_scores_gemma":[0.6350433,0.2813197,0.057925865,0.008044317,0.0014813623,0.0076423273,0.00804446,0.00024557026,0.0002530879],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.37612823,0.3350783,0.20407149,0.017103568,0.06589002,0.0017283483],"domain_scores_gemma":[0.07759088,0.7120894,0.13308586,0.042112235,0.033828266,0.0012932689],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.44227636,0.0024604497,0.01018016,0.05129476,0.0013623617,0.013055324,0.0056802784,0.005022158,0.0034323137],"category_scores_gemma":[0.7725924,0.0032451877,0.024042774,0.046114106,0.0039747213,0.011569671,0.005183675,0.004218894,0.0005526308],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0024656905,0.00019481458,0.24026443,0.3167334,0.3139064,0.00029843504,0.002294657,0.001338304,0.00046963678,0.0045481506,0.005190829,0.1122952],"study_design_scores_gemma":[0.001867309,0.0013516353,0.18472944,0.3307624,0.42760217,0.0021654824,0.0029927462,0.0037729924,0.0016342582,0.011625914,0.030892754,0.0006029704],"about_ca_topic_score_codex":0.003781374,"about_ca_topic_score_gemma":0.0042579677,"teacher_disagreement_score":0.9898198,"about_ca_system_score_codex":0.0066612344,"about_ca_system_score_gemma":0.010437445,"threshold_uncertainty_score":0.6877726},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2319079108","doi":"10.1016/j.joca.2016.01.101","title":"Assessing bias in osteoarthritis trials included in cochrane reviews: A meta-epidemiological study","year":2016,"lang":"en","type":"article","venue":"Osteoarthritis and Cartilage","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital; Bruyère; University of Ottawa; Cochrane","funders":"","keywords":"Blinding; Meta-analysis; Medicine; Systematic review; Publication bias; Reporting bias; MEDLINE; Study heterogeneity; Physical therapy; Sample size determination; Randomized controlled trial; Clinical trial; Psychological intervention; Epidemiology; Random effects model; Cochrane Library; Internal medicine; Statistics; Psychiatry","score_opus":0.8062375650609488,"score_gpt":0.5538883201150719,"score_spread":0.25234924494587685,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2319079108","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.03635264,0.9394122,0.017480638,0.00068094407,0.0011326923,0.0034016955,0.00082947465,0.00013932819,0.00057040167],"genre_scores_gemma":[0.7246042,0.21702813,0.041196913,0.0033868102,0.0011954657,0.010272186,0.0012899792,0.00028391238,0.0007424698],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.57666814,0.308152,0.07943793,0.01756638,0.015785918,0.0023896676],"domain_scores_gemma":[0.46784967,0.4589836,0.042962156,0.018499726,0.01017306,0.0015317794],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.22346777,0.0071780575,0.03690466,0.013708549,0.0016334265,0.009948881,0.0051200646,0.009356797,0.004065293],"category_scores_gemma":[0.46045542,0.0064588613,0.08900198,0.01112323,0.0032612744,0.008969825,0.0071886983,0.007197494,0.00048082403],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.007142019,0.000050167368,0.0060812035,0.06648231,0.9148422,0.00011974229,0.00021622829,0.00033863413,0.00011632536,0.00023723516,0.00024287424,0.004131103],"study_design_scores_gemma":[0.004790759,0.00043147538,0.003172167,0.0074155065,0.98186594,0.00015051635,0.0001043409,0.0005568954,0.000113031565,0.0008698546,0.0004569306,0.00007253023],"about_ca_topic_score_codex":0.0053911093,"about_ca_topic_score_gemma":0.009939843,"teacher_disagreement_score":0.99282193,"about_ca_system_score_codex":0.004999755,"about_ca_system_score_gemma":0.005983931,"threshold_uncertainty_score":0.9576025},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W233084541","doi":"10.3138/cjpe.017.006","title":"Public Training Programs in Canada: A Meta-evaluation","year":2002,"lang":"en","type":"article","venue":"Canadian Journal of Program Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"University of Manitoba","funders":"","keywords":"Summative assessment; Formative assessment; Training (meteorology); Program evaluation; Medical education; Political science; Public relations; Psychology; Pedagogy; Public administration; Medicine; Geography","score_opus":0.9006990989952361,"score_gpt":0.5282974646477141,"score_spread":0.37240163434752205,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W233084541","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3116767,0.6228138,0.008050878,0.0064136717,0.001709285,0.022735469,0.0180766,0.00034951596,0.008174106],"genre_scores_gemma":[0.88991684,0.08431486,0.00871772,0.0014911786,0.00022743658,0.010221937,0.0037806022,0.000059682963,0.0012697994],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.934086,0.043896284,0.0057422477,0.002319257,0.012267266,0.0016888313],"domain_scores_gemma":[0.89746624,0.05419417,0.0132257575,0.0033868963,0.028053442,0.003673575],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.059121016,0.0022568111,0.0077091428,0.0060642064,0.0030198595,0.0026265113,0.0030676136,0.0018441657,0.0035732293],"category_scores_gemma":[0.13470954,0.0010837994,0.013379367,0.011514155,0.0018507369,0.0010837152,0.0019073308,0.0018906947,0.00014161416],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.04153133,0.0014608236,0.07693619,0.1886291,0.551327,0.00037737217,0.0018841375,0.007328368,0.0004384176,0.0012158782,0.011884665,0.11698665],"study_design_scores_gemma":[0.023381345,0.004042233,0.10628773,0.054764505,0.7941281,0.0001302297,0.0013103024,0.002843445,0.00090030214,0.0007556625,0.011236363,0.00021974198],"about_ca_topic_score_codex":0.8252498,"about_ca_topic_score_gemma":0.8805843,"teacher_disagreement_score":0.940879,"about_ca_system_score_codex":0.059720833,"about_ca_system_score_gemma":0.091916814,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":true,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":true,"confidence":"high"},{"model":"opus","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":true,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2332511819","doi":"10.1016/j.jclinepi.2016.03.025","title":"A third of systematic reviews changed or did not specify the primary outcome: a PROSPERO register study","year":2016,"lang":"en","type":"article","venue":"Journal of Clinical Epidemiology","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":75,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital; Canadian Agency for Drugs and Technologies in Health; Canada Research Chairs; University of Ottawa; University of Toronto; St. Michael's Hospital","funders":"National Institute for Health and Care Research; Canadian Institutes of Health Research; University of Ottawa","keywords":"Register (sociolinguistics); Outcome (game theory); Medicine; Family medicine; Linguistics; Mathematics; Philosophy","score_opus":0.973059586795305,"score_gpt":0.7017350610403976,"score_spread":0.2713245257549074,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2332511819","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.20016871,0.6458673,0.022162398,0.036307584,0.010215448,0.015587184,0.04859521,0.0006929748,0.0204032],"genre_scores_gemma":[0.7468105,0.11808661,0.024889726,0.07213236,0.0030314692,0.017002383,0.013661473,0.0008534224,0.0035321077],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.64888906,0.12638584,0.18015328,0.01217241,0.027678246,0.004721155],"domain_scores_gemma":[0.4164917,0.41804832,0.08067246,0.053532276,0.028840076,0.0024152796],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.17892097,0.0023895146,0.009530431,0.0127740195,0.0026957442,0.00601891,0.002578758,0.007689412,0.005347127],"category_scores_gemma":[0.4750312,0.0029022691,0.019826727,0.016545314,0.0028621613,0.005751438,0.00531738,0.0058313017,0.0014608323],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.023856925,0.00056935113,0.1681797,0.38934076,0.18416168,0.004320619,0.00870246,0.00034851214,0.0039208666,0.007520049,0.03889478,0.17018428],"study_design_scores_gemma":[0.013120345,0.0015344373,0.1303794,0.17887124,0.50542194,0.00594504,0.002636731,0.0007826074,0.008043812,0.007664753,0.14489838,0.0007012816],"about_ca_topic_score_codex":0.017787402,"about_ca_topic_score_gemma":0.03186604,"teacher_disagreement_score":0.9904696,"about_ca_system_score_codex":0.0046458663,"about_ca_system_score_gemma":0.014968098,"threshold_uncertainty_score":0.94623584},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2528175253","doi":"10.7717/peerj.2429","title":"Prevention of necrotizing enterocolitis with probiotics: a systematic review and meta-analysis","year":2016,"lang":"en","type":"review","venue":"PeerJ","topic":"Infant Nutrition and Health","field":"Nursing","cited_by":134,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; London Health Sciences Centre","funders":"London Health Sciences Centre","keywords":"Necrotizing enterocolitis; Medicine; Meta-analysis; Randomized controlled trial; Internal medicine; Enterocolitis; Incidence (geometry); Clinical trial; Probiotic; Relative risk; Cochrane Library; Sepsis; Gestational age; MEDLINE; Systematic review; Pediatrics; Intensive care medicine; Confidence interval; Pregnancy","score_opus":0.09907042979141645,"score_gpt":0.39237747844070814,"score_spread":0.2933070486492917,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2528175253","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0025929806,0.9963013,0.0002647149,0.00015744609,0.00011280842,0.00019398252,0.00022113479,0.000015813854,0.00013980713],"genre_scores_gemma":[0.08592905,0.91004026,0.0016841983,0.00063821697,0.00024775442,0.000797226,0.0004456886,0.000018687319,0.00019892918],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9914215,0.004261136,0.0023254387,0.0007564829,0.0009305822,0.0003048043],"domain_scores_gemma":[0.9856264,0.010172572,0.0022496548,0.00042182076,0.0012681249,0.00026144568],"candidate_categories":["metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.012843582,0.002759838,0.021915523,0.008116797,0.0007148175,0.0032589578,0.0022932792,0.0023691626,0.0041493163],"category_scores_gemma":[0.027470525,0.0016207765,0.039611403,0.008377217,0.000650944,0.0019654601,0.0015879424,0.0022862267,0.0003246218],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0022762218,0.00004485116,0.002429998,0.45974326,0.5201869,0.000121597026,0.00006659372,0.00053434924,0.00019041075,0.000109886874,0.0005806001,0.013715293],"study_design_scores_gemma":[0.0006938129,0.00024223562,0.0023547742,0.03386853,0.9610407,0.00008327319,0.000038111648,0.00019573653,0.00009606895,0.0001535573,0.001211505,0.000021624595],"about_ca_topic_score_codex":0.0065157637,"about_ca_topic_score_gemma":0.0137756495,"teacher_disagreement_score":0.9780845,"about_ca_system_score_codex":0.0027094563,"about_ca_system_score_gemma":0.0039637703,"threshold_uncertainty_score":0.06792414},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2536368941","doi":"10.1093/ejo/cjw062","title":"Historical controls in orthodontics: need for larger meta-epidemiological studies","year":2016,"lang":"en","type":"letter","venue":"European Journal of Orthodontics","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"University of Alberta; Alberta Innovates; Killam Trusts; Alberta Innovates - Health Solutions; Women and Children's Health Research Institute; Children's Health Research Institute","keywords":"Orthodontics; Epidemiology; Medicine; Dentistry; Pathology","score_opus":0.9015952426937508,"score_gpt":0.5451367677650194,"score_spread":0.3564584749287314,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2536368941","genre_codex":"review","genre_gemma":"commentary","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0013389454,0.53360826,0.009389328,0.4164677,0.035805684,0.00044592647,0.0008540149,0.00014701855,0.0019432071],"genre_scores_gemma":[0.05356685,0.28848678,0.056558542,0.4745042,0.12077149,0.0031008685,0.0015246083,0.00021680638,0.0012697898],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.7138021,0.2082761,0.046864495,0.008292837,0.021726226,0.0010382435],"domain_scores_gemma":[0.13917352,0.781185,0.02516804,0.024480155,0.027686952,0.00230634],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.31571636,0.0024892823,0.014223556,0.0067380867,0.0019613376,0.0084643075,0.007808179,0.01594719,0.010232179],"category_scores_gemma":[0.6278673,0.002665547,0.013076334,0.007061832,0.004136171,0.01737739,0.0043112664,0.022992706,0.0016907878],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008331336,0.0004815654,0.018548576,0.1803432,0.08959745,0.0023859686,0.0015276972,0.0012861383,0.0007813798,0.029133528,0.3521517,0.31543145],"study_design_scores_gemma":[0.010908121,0.0017932254,0.024567435,0.21717292,0.09289901,0.003957992,0.0021776948,0.004804736,0.0009359845,0.14775051,0.4921061,0.0009263177],"about_ca_topic_score_codex":0.0037080843,"about_ca_topic_score_gemma":0.007016688,"teacher_disagreement_score":0.9857764,"about_ca_system_score_codex":0.004137568,"about_ca_system_score_gemma":0.008509175,"threshold_uncertainty_score":0.8438436},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2612870289","doi":"10.15173/ijsap.v1i1.3119","title":"A Systematic Literature Review of Students as Partners in Higher Education","year":2017,"lang":"en","type":"article","venue":"International Journal for Students as Partners","topic":"Higher Education Practises and Engagement","field":"Social Sciences","cited_by":597,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"General partnership; Reciprocity (cultural anthropology); Systematic review; Higher education; Space (punctuation); Scale (ratio); Pedagogy; Psychology; Medical education; Sociology; Political science; Medicine; MEDLINE; Computer science; Social psychology","score_opus":0.10269377469213052,"score_gpt":0.6131118316714596,"score_spread":0.5104180569793291,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2612870289","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0035291028,0.99244916,0.0005844815,0.0008099457,0.00023279143,0.00068163755,0.0006297909,0.000011394918,0.0010717473],"genre_scores_gemma":[0.029875651,0.96465087,0.0024517218,0.0009168145,0.00009199319,0.0012485466,0.00050905056,0.000009647707,0.00024569448],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.97453135,0.010802937,0.009626358,0.0012207527,0.0033410115,0.00047754624],"domain_scores_gemma":[0.910288,0.06972417,0.008986656,0.0013371021,0.008486086,0.001177996],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.02230247,0.0011189007,0.0042599295,0.022857942,0.0012972511,0.0039178277,0.0017689132,0.002387149,0.0036034249],"category_scores_gemma":[0.090418115,0.0010461056,0.0037694476,0.022177635,0.0015238414,0.005282759,0.0030762882,0.0013704357,0.0004401953],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012424955,0.00003066905,0.0019743242,0.87702084,0.0019427822,0.0003132186,0.0021656503,0.00010811155,0.00020186976,0.0009644831,0.003355324,0.11179851],"study_design_scores_gemma":[0.000046280256,0.00012650514,0.003261393,0.9446686,0.0075720497,0.0004546732,0.0024250196,0.00004195628,0.00018476123,0.00053189794,0.040656477,0.000030357587],"about_ca_topic_score_codex":0.008338092,"about_ca_topic_score_gemma":0.03619833,"teacher_disagreement_score":0.022857942,"about_ca_system_score_codex":0.00490152,"about_ca_system_score_gemma":0.035268582,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2615663954","doi":"10.1200/jco.2014.32.15_suppl.3057","title":"A systematic review of immune-related adverse event (irAE) reporting in clinical trials of immune checkpoint inhibitors (ICIs).","year":2014,"lang":"en","type":"review","venue":"Journal of Clinical Oncology","topic":"Cancer Immunotherapy and Biomarkers","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Princess Margaret Cancer Centre; University Health Network","funders":"","keywords":"Medicine; Clinical trial; Adverse effect; Consolidated Standards of Reporting Trials; Internal medicine; MEDLINE; Confidence interval; Systematic review","score_opus":0.27038554086918604,"score_gpt":0.5699617890490783,"score_spread":0.2995762481798923,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2615663954","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0036476399,0.9869832,0.0010383188,0.0008689971,0.0004168193,0.0042241076,0.0022410764,0.00003783173,0.0005420502],"genre_scores_gemma":[0.05932603,0.9097194,0.008329365,0.002180759,0.00037266733,0.017625166,0.0019884056,0.000049646107,0.00040856883],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.87836,0.041906413,0.062091686,0.0038218868,0.012997489,0.00082256994],"domain_scores_gemma":[0.73871255,0.1651644,0.06939723,0.003859474,0.021395462,0.0014707991],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.06850864,0.002330691,0.013567313,0.024989123,0.0013343545,0.0047612675,0.0027799164,0.0026795364,0.0041675144],"category_scores_gemma":[0.22890773,0.0019118458,0.0115082525,0.02047184,0.0022441372,0.004780401,0.0029495277,0.0017327954,0.0004859396],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00026633285,0.000008423714,0.0009086412,0.9771778,0.007422365,0.000052845546,0.00019907863,0.000052515872,0.00012794607,0.00012246294,0.00080523995,0.012856363],"study_design_scores_gemma":[0.00061683723,0.00027242102,0.004362252,0.91088957,0.07181907,0.00028789975,0.00029071892,0.000138975,0.0002566354,0.0003463717,0.010665465,0.000053779593],"about_ca_topic_score_codex":0.005431398,"about_ca_topic_score_gemma":0.016778575,"teacher_disagreement_score":0.9314914,"about_ca_system_score_codex":0.0072258776,"about_ca_system_score_gemma":0.023055356,"threshold_uncertainty_score":0.36231267},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2620826610","doi":"10.1016/j.jclinepi.2017.05.017","title":"A systematic survey on reporting and methods for handling missing participant data for continuous outcomes in randomized controlled trials","year":2017,"lang":"en","type":"review","venue":"Journal of Clinical Epidemiology","topic":"Statistical Methods and Bayesian Inference","field":"Mathematics","cited_by":34,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; McMaster University; Impact","funders":"","keywords":"Missing data; Randomized controlled trial; Medicine; Interquartile range; Imputation (statistics); MEDLINE; Clinical trial; Statistics; Surgery; Internal medicine; Mathematics","score_opus":0.9405744797648787,"score_gpt":0.763690149918479,"score_spread":0.17688432984639968,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2620826610","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0011711699,0.95231074,0.03015415,0.007993059,0.0012352259,0.0043893247,0.0016295796,0.00017952482,0.0009372088],"genre_scores_gemma":[0.020068366,0.8556011,0.09838827,0.008693773,0.00094079034,0.014239129,0.0014916637,0.0001901863,0.0003868513],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.49246556,0.2835225,0.17025456,0.00766639,0.044246066,0.0018449712],"domain_scores_gemma":[0.16136847,0.735537,0.057616334,0.015488104,0.028476918,0.001513101],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.40741128,0.002865523,0.015702441,0.014906944,0.0015055757,0.007779796,0.00629167,0.008212206,0.0046298676],"category_scores_gemma":[0.7438943,0.004556388,0.018083973,0.014507164,0.005241066,0.010428471,0.00578855,0.013704958,0.0009262856],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018114657,0.00008485567,0.00199144,0.5956739,0.029692924,0.00008400899,0.0010026466,0.0011887917,0.000292501,0.008132471,0.018604884,0.34144014],"study_design_scores_gemma":[0.0023056108,0.0006070937,0.0039115716,0.8570258,0.069324374,0.00045117844,0.00034559748,0.0013337134,0.0006350366,0.018702473,0.0450834,0.00027422988],"about_ca_topic_score_codex":0.009930172,"about_ca_topic_score_gemma":0.013792783,"teacher_disagreement_score":0.5925887,"about_ca_system_score_codex":0.010877223,"about_ca_system_score_gemma":0.04139325,"threshold_uncertainty_score":0.73076737},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2738634406","doi":"10.1177/1049731517718940","title":"A Systematic Survey of Control Groups in Behavioral and Social Science Trials","year":2017,"lang":"en","type":"article","venue":"Research on Social Work Practice","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Joseph’s Healthcare Hamilton; McMaster University; McMaster University Medical Centre; Impact","funders":"","keywords":"Data extraction; Randomized controlled trial; MEDLINE; Psychological intervention; Intervention (counseling); Cochrane Library; Meta-analysis; Behavioural sciences; Medicine; Treatment and control groups; Psychology; Clinical psychology; Psychiatry; Psychotherapist; Surgery; Biology; Internal medicine","score_opus":0.9478805604741037,"score_gpt":0.7232155535404406,"score_spread":0.2246650069336631,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2738634406","genre_codex":"review","genre_gemma":"review","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.026149832,0.8119214,0.049033746,0.009066577,0.0028868807,0.086632445,0.008342046,0.00030131053,0.0056657675],"genre_scores_gemma":[0.24236217,0.2576494,0.14745794,0.011672512,0.00086964265,0.33181316,0.007018731,0.00032543854,0.00083098403],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.15932903,0.41588032,0.36721426,0.013327563,0.04297878,0.0012700789],"domain_scores_gemma":[0.13794544,0.69118685,0.09909664,0.028514648,0.041624,0.0016323437],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.54040414,0.003069144,0.014205975,0.027592856,0.0027866368,0.008834832,0.0041907346,0.0061921477,0.0050441385],"category_scores_gemma":[0.8082518,0.004429038,0.015171435,0.02974222,0.006095288,0.01090113,0.0071745683,0.0046816003,0.00086497],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004138036,0.000116312476,0.010042352,0.81021297,0.03590331,0.00014633079,0.0030088348,0.00051429705,0.0008382712,0.004987415,0.004578334,0.12551354],"study_design_scores_gemma":[0.008541906,0.0018572753,0.015052505,0.8325573,0.079235844,0.0005333182,0.002049417,0.0010700481,0.0018349197,0.010712402,0.046278514,0.0002765935],"about_ca_topic_score_codex":0.002833085,"about_ca_topic_score_gemma":0.004745508,"teacher_disagreement_score":0.45959586,"about_ca_system_score_codex":0.008931414,"about_ca_system_score_gemma":0.022854831,"threshold_uncertainty_score":0.5667635},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2749388036","doi":"10.1177/0022034517725049","title":"Impact of Selection Bias on Treatment Effect Size Estimates in Randomized Trials of Oral Health Interventions: A Meta-epidemiological Study","year":2017,"lang":"en","type":"review","venue":"Journal of Dental Research","topic":"Statistical Methods in Clinical Trials","field":"Mathematics","cited_by":41,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute of Health Economics; University of Alberta","funders":"","keywords":"Selection bias; Meta-analysis; Epidemiology; Randomized controlled trial; Psychological intervention; Medicine; Sample size determination; Oral health; Effect modification; Environmental health; Selection (genetic algorithm); Statistics; Dentistry; Confidence interval; Internal medicine; Computer science; Mathematics; Psychiatry; Pathology","score_opus":0.9859906141623999,"score_gpt":0.8396295782838669,"score_spread":0.146361035878533,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2749388036","genre_codex":"review","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.018494688,0.85855687,0.108771645,0.0046599796,0.0028044248,0.0035279617,0.0011086809,0.00035911085,0.0017167246],"genre_scores_gemma":[0.64948475,0.1918979,0.13026232,0.009423707,0.002338042,0.013826187,0.0014350716,0.000445777,0.00088618847],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.41291055,0.460783,0.08276537,0.016014628,0.026000556,0.0015259478],"domain_scores_gemma":[0.24845886,0.68012035,0.038896546,0.02190562,0.009942393,0.0006761646],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.42773893,0.004418932,0.018342838,0.012135994,0.0012710043,0.011325981,0.0039673112,0.0060396795,0.0023985694],"category_scores_gemma":[0.65088445,0.003276793,0.05145034,0.012460903,0.003986713,0.008142258,0.0051821577,0.0069881435,0.00033865724],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004387647,0.000094150855,0.023587005,0.11599486,0.77849084,0.0004259996,0.0007456318,0.004926826,0.0004443965,0.005370085,0.0017611609,0.06377131],"study_design_scores_gemma":[0.005307195,0.0010996358,0.015526982,0.04889485,0.88069683,0.0008193623,0.00028911926,0.0106883235,0.0013200268,0.027910315,0.0071395165,0.00030787126],"about_ca_topic_score_codex":0.002595644,"about_ca_topic_score_gemma":0.0032475141,"teacher_disagreement_score":0.9955811,"about_ca_system_score_codex":0.004603424,"about_ca_system_score_gemma":0.0062040635,"threshold_uncertainty_score":0.7056998},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2750781151","doi":"10.1136/bmjopen-2017-015886","title":"Scoping review protocol: education initiatives for medical psychiatry collaborative care","year":2017,"lang":"en","type":"article","venue":"BMJ Open","topic":"Mental Health and Patient Involvement","field":"Health Professions","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Centre for Addiction and Mental Health; University Health Network; Trillium Health Centre; University of Toronto","funders":"Hospital for Sick Children; University of Toronto; Ontario Ministry of Health and Long-Term Care; Medical Psychiatry Alliance; Centre for Addiction and Mental Health","keywords":"Medicine; Health care; Protocol (science); Medical education; Collaborative Care; Descriptive statistics; Mental health; Nursing; Alternative medicine; Psychiatry","score_opus":0.41217486699205347,"score_gpt":0.6752642165252536,"score_spread":0.2630893495332001,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2750781151","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00018664407,0.00039711857,0.0014668566,0.0003727031,0.00023080171,0.9913378,0.0046530818,0.00010986017,0.001245304],"genre_scores_gemma":[0.00018817221,0.0003550801,0.0036570283,0.00012367673,0.000017791905,0.9946838,0.0005665453,0.000011512486,0.0003963102],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.9327435,0.028831353,0.023086103,0.0051326943,0.007594143,0.0026121826],"domain_scores_gemma":[0.8513313,0.047805727,0.019902784,0.014864013,0.059914622,0.0061815516],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.1552128,0.004855572,0.010768007,0.017858012,0.006438832,0.013996077,0.007378897,0.01102909,0.17474909],"category_scores_gemma":[0.16941564,0.0061313384,0.009243238,0.020666767,0.0065330137,0.010585415,0.00984569,0.0098060565,0.044552453],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0075457934,0.0007151156,0.00095899595,0.68351024,0.0010066974,0.000708193,0.0065033105,0.0012851829,0.0018151279,0.016460016,0.19044934,0.08904197],"study_design_scores_gemma":[0.021453075,0.0015347584,0.0041155624,0.45180005,0.0020196436,0.0004389552,0.0068678977,0.0013700224,0.0020685527,0.015170244,0.4926896,0.00047170912],"about_ca_topic_score_codex":0.0049159797,"about_ca_topic_score_gemma":0.009072014,"teacher_disagreement_score":0.17474909,"about_ca_system_score_codex":0.01759017,"about_ca_system_score_gemma":0.09909953,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":[],"domain":null,"study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2756094298","doi":"10.1136/ebmed-2017-110798","title":"What does expert opinion in guidelines mean? a meta-epidemiological study","year":2017,"lang":"en","type":"review","venue":"Evidence-Based Medicine","topic":"Clinical practice guidelines implementation","field":"Medicine","cited_by":38,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Guideline; Observational study; CLARITY; Medicine; Expert opinion; Evidence-based medicine; Evidence-based practice; Family medicine; MEDLINE; Alternative medicine; Pathology; Intensive care medicine","score_opus":0.934234501038434,"score_gpt":0.7009597189922305,"score_spread":0.23327478204620344,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2756094298","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.002505329,0.9931096,0.0009176985,0.0020179916,0.0003844374,0.00016514116,0.00024420483,0.0000080222635,0.0006476319],"genre_scores_gemma":[0.16600074,0.8218055,0.004130989,0.004545182,0.001581101,0.0009429006,0.0006865091,0.000049653903,0.0002574737],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.85514176,0.09441846,0.028280536,0.0075987363,0.013389032,0.0011714653],"domain_scores_gemma":[0.45890886,0.4699189,0.04572413,0.0066956435,0.017372588,0.0013798799],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.086951666,0.0016642761,0.010420713,0.0141925,0.0010554986,0.0075931177,0.003315478,0.0047935224,0.007043593],"category_scores_gemma":[0.3649171,0.00158146,0.015374636,0.016785018,0.001972853,0.007856677,0.0025077309,0.0039122677,0.00061555015],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00213779,0.000114615286,0.033825286,0.57405925,0.22832276,0.00034564527,0.0016941361,0.00061511865,0.00013416009,0.0031582653,0.0068427515,0.14875014],"study_design_scores_gemma":[0.0015576485,0.000609621,0.036204435,0.5032052,0.42108357,0.0008123619,0.001190456,0.0010286272,0.00025319407,0.007994121,0.025902588,0.00015815857],"about_ca_topic_score_codex":0.0059412494,"about_ca_topic_score_gemma":0.009110675,"teacher_disagreement_score":0.98957926,"about_ca_system_score_codex":0.0057800054,"about_ca_system_score_gemma":0.0062402748,"threshold_uncertainty_score":0.45984983},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2763139081","doi":"10.1186/s13012-017-0648-y","title":"Organizational participatory research: a systematic mixed studies review exposing its extra benefits and the key factors associated with them","year":2017,"lang":"en","type":"review","venue":"Implementation Science","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":94,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Canadian Pharmacists Association; Université de Montréal; Université de Sherbrooke; Douglas Mental Health University Institute; McGill University","funders":"Canadian Institutes of Health Research; McGill University","keywords":"Health services research; Health administration; Medicine; Health informatics; Citizen journalism; Participatory action research; Nursing research; Value (mathematics); Public health; Key (lock); Public relations; Knowledge management; Sociology; Nursing; Political science; Computer science","score_opus":0.9823586787688844,"score_gpt":0.7849291663275012,"score_spread":0.19742951244138318,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2763139081","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0053963163,0.98055077,0.0056508156,0.0022141566,0.0004707859,0.004052554,0.0004489157,0.000029041243,0.0011865224],"genre_scores_gemma":[0.11095595,0.8418579,0.026554113,0.0029365094,0.00053113484,0.016159603,0.0005363707,0.00005418016,0.00041428677],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.74050087,0.18397094,0.044395283,0.0068167783,0.022598509,0.0017175887],"domain_scores_gemma":[0.40846145,0.5210779,0.036745146,0.009333375,0.02321467,0.0011673926],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.20155142,0.002099616,0.0071624294,0.022283204,0.0027843388,0.011054062,0.002898109,0.0036709774,0.003979999],"category_scores_gemma":[0.4213026,0.002040632,0.010466617,0.025837753,0.0044788728,0.009984857,0.005622523,0.0024840059,0.0004299839],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00036620456,0.00008318181,0.0049564,0.80277187,0.009999434,0.00023239892,0.007296087,0.0002788762,0.00021539743,0.0054120524,0.0021850897,0.16620311],"study_design_scores_gemma":[0.0001567186,0.00035303543,0.0044424273,0.9438422,0.017754726,0.00037943394,0.0058579114,0.00029846,0.0003179491,0.003211236,0.023302458,0.00008334796],"about_ca_topic_score_codex":0.00454174,"about_ca_topic_score_gemma":0.011270419,"teacher_disagreement_score":0.20155142,"about_ca_system_score_codex":0.011323782,"about_ca_system_score_gemma":0.0338006,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2766184587","doi":"10.1186/s40510-017-0196-7","title":"Effectiveness of en masse versus two-step retraction: a systematic review and meta-analysis","year":2017,"lang":"en","type":"review","venue":"Progress in Orthodontics","topic":"Orthodontics and Dentofacial Orthopedics","field":"Dentistry","cited_by":53,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Meta-analysis; Dentistry; Randomized controlled trial; Orthodontics; Surgery; Internal medicine","score_opus":0.18654494011674747,"score_gpt":0.4686407095626851,"score_spread":0.2820957694459376,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2766184587","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0016785776,0.9971794,0.00026074532,0.00015404704,0.000103655424,0.00031565715,0.0001380558,0.000008427894,0.00016147656],"genre_scores_gemma":[0.067732334,0.9261234,0.0024077368,0.0010143479,0.00032587367,0.0018890182,0.00031692325,0.000018011613,0.00017236415],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.978512,0.0105345715,0.0061799516,0.0015767459,0.0026943113,0.00050240033],"domain_scores_gemma":[0.95965916,0.030764515,0.0064273193,0.00072723074,0.0020603002,0.0003614281],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.024301816,0.0030991414,0.019357681,0.00835547,0.00076406804,0.0047151884,0.0026790868,0.0029599974,0.004092792],"category_scores_gemma":[0.059031017,0.0015545683,0.032144766,0.006645446,0.0011890722,0.0029582123,0.0016462345,0.0025417395,0.00030116283],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0020506964,0.00005540216,0.0013033878,0.7932136,0.18097232,0.00008575783,0.000102290796,0.0002520452,0.0001703164,0.00018762804,0.00038482586,0.0212218],"study_design_scores_gemma":[0.0017967196,0.0009111177,0.0035211022,0.18597637,0.8025201,0.00018415821,0.00014089647,0.00019111348,0.00030976167,0.00036347366,0.0040381127,0.000046985362],"about_ca_topic_score_codex":0.0043136063,"about_ca_topic_score_gemma":0.009904828,"teacher_disagreement_score":0.024301816,"about_ca_system_score_codex":0.004149289,"about_ca_system_score_gemma":0.0071315845,"threshold_uncertainty_score":0.1285218},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"evaluation","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","research_integrity","scholarly_communication"],"domain":"evaluation","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2767159407","doi":"10.1136/bmjopen-2017-018856","title":"Association between statistical significance and time to publication among systematic reviews: a study protocol for a meta-epidemiological investigation","year":2017,"lang":"en","type":"article","venue":"BMJ Open","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"Japan Society for the Promotion of Science","keywords":"Medicine; Publication bias; Protocol (science); Family medicine; Clinical trial; Systematic review; Meta-analysis; Epidemiology; Publication; Peer review; Psychological intervention; MEDLINE; Medical education; Alternative medicine; Pathology; Psychiatry","score_opus":0.9313157082465333,"score_gpt":0.6675016834766426,"score_spread":0.2638140247698907,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2767159407","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00089060486,0.0008364855,0.008580626,0.00047052468,0.00048279675,0.9857319,0.0023507285,0.00021970403,0.0004365596],"genre_scores_gemma":[0.00062913384,0.00015296375,0.0069511407,0.00007569106,0.00001873478,0.99189174,0.0001854158,0.000009838699,0.00008536324],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.75370055,0.14734884,0.07143454,0.009996256,0.014528178,0.002991644],"domain_scores_gemma":[0.7235413,0.14993748,0.04262769,0.04064716,0.038662877,0.0045835134],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.23763415,0.0048972075,0.010712394,0.011477964,0.0035419283,0.0050967494,0.004025607,0.010754651,0.043879565],"category_scores_gemma":[0.33976465,0.0048500495,0.018109893,0.013949683,0.0059777456,0.0075559714,0.0037786663,0.010389922,0.0094770305],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.077253245,0.0031082935,0.006166592,0.63751537,0.023840224,0.0014725785,0.007987019,0.005760844,0.004865429,0.023381999,0.082483724,0.12616475],"study_design_scores_gemma":[0.25949717,0.015025146,0.032778982,0.29346853,0.040870644,0.0015827996,0.003285394,0.025788182,0.009351021,0.07820796,0.23814565,0.0019984767],"about_ca_topic_score_codex":0.0011749157,"about_ca_topic_score_gemma":0.0026601586,"teacher_disagreement_score":0.9892876,"about_ca_system_score_codex":0.007614664,"about_ca_system_score_gemma":0.027774857,"threshold_uncertainty_score":0.9401328},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"evaluation","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_narrow","metaepi_broad","metaresearch"],"domain":"evaluation","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2769378808","doi":"10.1016/j.jclinepi.2017.10.014","title":"Same family, different species: methodological conduct and quality varies according to purpose for five types of knowledge synthesis","year":2017,"lang":"en","type":"review","venue":"Journal of Clinical Epidemiology","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":126,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Canada Research Chairs; Ottawa Hospital; University of Toronto; University of Alberta; St. Michael's Hospital","funders":"National Health and Medical Research Council; Medical Research Council; Canadian Institutes of Health Research; University of Ottawa","keywords":"Systematic review; Medicine; Protocol (science); Citation; Missing data; MEDLINE; Medical physics; Statistics; Alternative medicine; Computer science; Mathematics; Pathology; Library science","score_opus":0.9944644201276043,"score_gpt":0.7932841591772162,"score_spread":0.2011802609503881,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2769378808","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.014620837,0.6524433,0.20411824,0.07353954,0.025351712,0.00777324,0.0073244977,0.001072924,0.013755809],"genre_scores_gemma":[0.24209733,0.1790945,0.47030753,0.068598986,0.0064158095,0.02397451,0.0055180555,0.0017171422,0.0022761677],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.5014104,0.3146493,0.12685852,0.018965712,0.035716727,0.0023992597],"domain_scores_gemma":[0.20855165,0.64723593,0.058690663,0.049083512,0.034078166,0.0023600983],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.35998967,0.0019976308,0.009923826,0.02611766,0.0043198857,0.01731975,0.00448033,0.009924623,0.0072161877],"category_scores_gemma":[0.7728836,0.0025945099,0.016898785,0.031356625,0.010960833,0.015725171,0.009871923,0.009822418,0.0016917919],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004206659,0.00009790054,0.014542441,0.3447362,0.092897706,0.00049551803,0.01574446,0.00071965007,0.001820528,0.081883386,0.06665721,0.37619835],"study_design_scores_gemma":[0.0034030252,0.0008230261,0.025575222,0.37699398,0.14160372,0.0025994787,0.008613416,0.0030348438,0.002463472,0.25919285,0.17485616,0.0008407717],"about_ca_topic_score_codex":0.00517073,"about_ca_topic_score_gemma":0.008057565,"teacher_disagreement_score":0.9900762,"about_ca_system_score_codex":0.006762432,"about_ca_system_score_gemma":0.01696075,"threshold_uncertainty_score":0.7892467},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2772564313","doi":"10.1016/j.jclinepi.2017.11.022","title":"Flaws in the application and interpretation of statistical analyses in systematic reviews of therapeutic interventions were common: a cross-sectional analysis","year":2017,"lang":"en","type":"article","venue":"Journal of Clinical Epidemiology","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":85,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Public Health Ontario; University of Toronto; St. Michael's Hospital; Ottawa Hospital; University of Ottawa","funders":"National Health and Medical Research Council; Generalitat Valenciana; University of Bristol; National Institute for Health and Care Research; Canada Research Chairs; Cancer Research UK; Medical Research Council; University of Ottawa","keywords":"Meta-analysis; Funnel plot; Interquartile range; Medicine; Random effects model; Sample size determination; Statistics; Publication bias; Subgroup analysis; Confounding; Psychological intervention; Study heterogeneity; MEDLINE; Mathematics; Internal medicine","score_opus":0.9443845471738417,"score_gpt":0.7567789989945132,"score_spread":0.1876055481793285,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2772564313","genre_codex":"methods","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.36760122,0.10853541,0.48006496,0.01620118,0.0051656673,0.010954319,0.004816851,0.0012708444,0.0053895623],"genre_scores_gemma":[0.8711884,0.004477173,0.11299848,0.0042331475,0.00050487847,0.0049583786,0.00083401176,0.00043636086,0.00036922167],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.16637015,0.5756315,0.20298019,0.019684592,0.033839915,0.0014936646],"domain_scores_gemma":[0.035792604,0.8431635,0.054867215,0.04759827,0.018105866,0.00047257525],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.725693,0.0020866285,0.0053404337,0.005543636,0.002107308,0.006641176,0.003620342,0.0051339557,0.002129825],"category_scores_gemma":[0.89225864,0.0039073653,0.014071035,0.008152404,0.00567793,0.007909955,0.004269185,0.00560512,0.00031408563],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.009039705,0.00071186846,0.54079646,0.04925933,0.20068848,0.0011789408,0.019248953,0.0074166376,0.0021502965,0.019274656,0.009029775,0.14120491],"study_design_scores_gemma":[0.0039082333,0.006823068,0.55260426,0.053219844,0.21148735,0.0045975377,0.0050918604,0.044709474,0.012240415,0.06637688,0.03757209,0.0013691003],"about_ca_topic_score_codex":0.003644934,"about_ca_topic_score_gemma":0.005357423,"teacher_disagreement_score":0.274307,"about_ca_system_score_codex":0.003343997,"about_ca_system_score_gemma":0.006823402,"threshold_uncertainty_score":0.3382694},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2776920184","doi":"10.2106/jbjs.16.01408","title":"A Historical Analysis of Randomized Controlled Trials in Anterior Cruciate Ligament Surgery","year":2017,"lang":"en","type":"review","venue":"Journal of Bone and Joint Surgery","topic":"Knee injuries and reconstruction techniques","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Consolidated Standards of Reporting Trials; Randomized controlled trial; Confidence interval; Checklist; Anterior cruciate ligament reconstruction; MEDLINE; Randomization; Anterior cruciate ligament; Physical therapy; Internal medicine; Surgery; Psychology","score_opus":0.1492740404499428,"score_gpt":0.3919929937555513,"score_spread":0.2427189533056085,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2776920184","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0056284247,0.96222806,0.017706791,0.0037447296,0.0022251501,0.003730653,0.0021866772,0.00009706352,0.002452606],"genre_scores_gemma":[0.3588516,0.5362498,0.064927205,0.010833278,0.0039941743,0.020111796,0.0036850844,0.00027074732,0.0010763181],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.38132778,0.32549298,0.2037134,0.02184571,0.06552238,0.0020976556],"domain_scores_gemma":[0.11547967,0.6890557,0.13378814,0.018790266,0.041719038,0.0011671587],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.42730713,0.0021666195,0.011790108,0.020597458,0.0018021847,0.008642368,0.002691873,0.004063523,0.006520707],"category_scores_gemma":[0.7523092,0.002510323,0.021645313,0.017186057,0.0039636986,0.008499344,0.0037675628,0.0046575502,0.00080536574],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.005354417,0.00010356888,0.028110683,0.6587711,0.15894406,0.00038114213,0.0012010419,0.001954478,0.0005361297,0.009373844,0.006707316,0.12856227],"study_design_scores_gemma":[0.003516871,0.0038132991,0.05234159,0.5511458,0.29544437,0.0018257165,0.000852708,0.00363022,0.0013411718,0.017859807,0.067894585,0.0003339261],"about_ca_topic_score_codex":0.0021413828,"about_ca_topic_score_gemma":0.0034596848,"teacher_disagreement_score":0.5726929,"about_ca_system_score_codex":0.009031266,"about_ca_system_score_gemma":0.012415994,"threshold_uncertainty_score":0.7062323},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad","bibliometrics"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2790296943","doi":"10.1177/1049732318759668","title":"A Meta-Study of Qualitative Research Examining Sport and Recreation Experiences of Indigenous Youth","year":2018,"lang":"en","type":"review","venue":"Qualitative Health Research","topic":"Recreation, Leisure, Wilderness Management","field":"Psychology","cited_by":28,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Alberta","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"Recreation; Indigenous; Qualitative research; Situated; Psychology; Applied psychology; Sociology; Political science; Social science; Ecology","score_opus":0.9412498822152408,"score_gpt":0.7399334983437864,"score_spread":0.20131638387145434,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2790296943","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.013522408,0.97566813,0.0044583324,0.0013285252,0.00043447362,0.0021122708,0.0008423832,0.000025986155,0.0016075049],"genre_scores_gemma":[0.1974904,0.77698195,0.013897709,0.001732181,0.00013867204,0.008296406,0.0007698821,0.000046864054,0.0006460166],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9701357,0.016101088,0.007815505,0.0014704518,0.0037865776,0.0006907806],"domain_scores_gemma":[0.87856585,0.096924044,0.006474802,0.003245197,0.014219299,0.00057086005],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.057050105,0.0012790486,0.0049650315,0.012550119,0.0021680736,0.0035321584,0.0018889273,0.0012137636,0.001986387],"category_scores_gemma":[0.15132062,0.0011437682,0.006830318,0.014441424,0.0016200976,0.0030230712,0.0026117954,0.0015206019,0.00015846614],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00044107344,0.000069255824,0.0080962125,0.79232615,0.025463894,0.00040578056,0.020171296,0.00043382935,0.0006622591,0.0026138097,0.002654004,0.14666241],"study_design_scores_gemma":[0.00020069462,0.0003533406,0.011713567,0.81233644,0.10922544,0.0006304047,0.016547281,0.00027559535,0.0015497265,0.002424683,0.044632964,0.00010982527],"about_ca_topic_score_codex":0.037258357,"about_ca_topic_score_gemma":0.08688244,"teacher_disagreement_score":0.995035,"about_ca_system_score_codex":0.014865185,"about_ca_system_score_gemma":0.036210667,"threshold_uncertainty_score":0.3017134},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2802073505","doi":"10.1183/13993003.00376-2018","title":"The effect of anti-acid therapy on survival in idiopathic pulmonary fibrosis: a methodological review of observational studies","year":2018,"lang":"en","type":"review","venue":"European Respiratory Journal","topic":"Interstitial Lung Diseases and Idiopathic Pulmonary Fibrosis","field":"Medicine","cited_by":50,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Jewish General Hospital","funders":"","keywords":"Medicine; Observational study; Pulmonary fibrosis; Idiopathic pulmonary fibrosis; Intensive care medicine; Internal medicine; Fibrosis; Lung","score_opus":0.2886020444561501,"score_gpt":0.44182245393585945,"score_spread":0.15322040947970933,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2802073505","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0006334123,0.99806494,0.00044286996,0.00027475183,0.00010553605,0.00019461225,0.00015826573,0.000003232225,0.00012241054],"genre_scores_gemma":[0.017258849,0.9782445,0.0022557024,0.00072346086,0.00017953807,0.00088598236,0.00035707402,0.000009266897,0.00008567313],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.93850416,0.02655804,0.026417583,0.0025217216,0.0055286097,0.0004699036],"domain_scores_gemma":[0.7815145,0.1763232,0.027352277,0.0045526247,0.00973413,0.0005232738],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.06683796,0.0016669781,0.00770628,0.008253897,0.000731723,0.0030097738,0.0026822048,0.0021587713,0.0025352633],"category_scores_gemma":[0.19656609,0.0012970257,0.012715929,0.009151739,0.0013261682,0.0025829584,0.0022353958,0.0015180542,0.00025978265],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006797076,0.000032555352,0.0047524744,0.860606,0.041147266,0.000116128555,0.00029645534,0.00025095325,0.0001981075,0.000657824,0.0013300398,0.08993259],"study_design_scores_gemma":[0.0006266778,0.00047894375,0.016534902,0.66168445,0.2890676,0.0005832982,0.00037661768,0.00022775115,0.00050939363,0.002057317,0.027770324,0.000082746126],"about_ca_topic_score_codex":0.0046321317,"about_ca_topic_score_gemma":0.008718787,"teacher_disagreement_score":0.93316203,"about_ca_system_score_codex":0.003269977,"about_ca_system_score_gemma":0.0096302545,"threshold_uncertainty_score":0.35347712},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_narrow","metaepi_broad","metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2885318744","doi":"10.1016/j.jval.2018.06.010","title":"Valuation of Health and Nonhealth Outcomes from Next-Generation Sequencing: Approaches, Challenges, and Solutions","year":2018,"lang":"en","type":"review","venue":"Value in Health","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":76,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary; Canadian Centre for Applied Research in Cancer Control; University of British Columbia","funders":"Canadian Cancer Society Research Institute; Canadian Centre for Applied Research in Cancer Control; Cancer Research Institute","keywords":"Valuation (finance); MEDLINE; Health care; Preference; Actuarial science; Computer science; Data science; Management science; Medicine; Economics; Biology; Accounting; Microeconomics","score_opus":0.9440262427380309,"score_gpt":0.4852302784682474,"score_spread":0.45879596426978353,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2885318744","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0013019594,0.9826095,0.009822879,0.0041295546,0.00026204588,0.000031206433,0.00021273726,0.000013840958,0.0016163575],"genre_scores_gemma":[0.030060645,0.9570922,0.010299772,0.00100057,0.00078406057,0.00011037935,0.0002616408,0.000014711662,0.00037606095],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9922105,0.00492079,0.0005045806,0.0004971983,0.0017083852,0.00015857184],"domain_scores_gemma":[0.928922,0.06574709,0.0021203049,0.0009005295,0.0020716623,0.00023842957],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.024940474,0.001896271,0.004630764,0.004504889,0.0003154784,0.0047110436,0.0028058258,0.0023389682,0.003710939],"category_scores_gemma":[0.05606315,0.00049757096,0.0024877517,0.0063326624,0.0019984029,0.005636149,0.002095528,0.0035196117,0.00038072473],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015987345,0.000104735445,0.004101708,0.01836678,0.0018564021,0.00011738878,0.00019798355,0.010299402,0.00019003247,0.12379902,0.009569537,0.83123714],"study_design_scores_gemma":[0.00014851533,0.00026615264,0.008567069,0.07042096,0.0029031676,0.0009763845,0.0008339596,0.02111234,0.0006828966,0.67316914,0.22066833,0.00025108652],"about_ca_topic_score_codex":0.0036824427,"about_ca_topic_score_gemma":0.004870058,"teacher_disagreement_score":0.024940474,"about_ca_system_score_codex":0.0025816606,"about_ca_system_score_gemma":0.004672322,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2886516149","doi":"10.1136/bmjebm-2018-111024.40","title":"40 How fragile is the evidence base? a meta-epidemiologic study of the fragilityindex derived from 374 randomised trials","year":2018,"lang":"en","type":"article","venue":"","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Fragility; Randomized controlled trial; Meta-analysis; Sample size determination; False positive paradox; Medicine; Statistics; Index (typography); Computer science; Mathematics; Internal medicine","score_opus":0.7901114122054491,"score_gpt":0.49293071844313197,"score_spread":0.29718069376231715,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2886516149","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.030007564,0.8991595,0.04313177,0.009415833,0.0027701405,0.006453842,0.004609424,0.00029870024,0.00415327],"genre_scores_gemma":[0.8149791,0.1018795,0.052218217,0.011950079,0.0014459175,0.013407292,0.002937233,0.0002224658,0.0009603087],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.5722841,0.29835483,0.093041904,0.013015358,0.021229027,0.0020747262],"domain_scores_gemma":[0.25880063,0.6447164,0.05275673,0.030004812,0.012643901,0.0010774959],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.29901454,0.002940666,0.01174985,0.013807865,0.00156379,0.0074028918,0.00431477,0.0046637855,0.0073329275],"category_scores_gemma":[0.6141738,0.0026376985,0.04578748,0.012579469,0.0033150732,0.007549199,0.0042923614,0.005226632,0.0009216615],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.005720903,0.00006533037,0.024148833,0.19135165,0.733145,0.00039339127,0.0005020685,0.0016056441,0.00037700808,0.0026834311,0.002666088,0.037340596],"study_design_scores_gemma":[0.003717511,0.0009780857,0.018039325,0.06694474,0.88646466,0.00042250502,0.00032204806,0.0017796482,0.0004922786,0.010867459,0.009803938,0.00016786433],"about_ca_topic_score_codex":0.0038172617,"about_ca_topic_score_gemma":0.006314867,"teacher_disagreement_score":0.98825014,"about_ca_system_score_codex":0.0045862948,"about_ca_system_score_gemma":0.00559014,"threshold_uncertainty_score":0.86443985},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2888025021","doi":"10.1093/scan/nsy072","title":"Meta-analysis of ERP investigations of pain empathy underlines methodological issues in ERP research","year":2018,"lang":"en","type":"review","venue":"Social Cognitive and Affective Neuroscience","topic":"Neural and Behavioral Psychology Studies","field":"Neuroscience","cited_by":121,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Fonds de Recherche du Québec - Santé; Canadian Institutes of Health Research","keywords":"Empathy; Psychology; Flexibility (engineering); Cognitive psychology; Reliability (semiconductor); Cognition; Meta-analysis; Social psychology; Neuroscience","score_opus":0.9030527721176843,"score_gpt":0.6320929554922415,"score_spread":0.2709598166254428,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2888025021","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0018754725,0.9954194,0.0015621762,0.00042646818,0.0001492367,0.000049453793,0.00018226524,0.000015508656,0.00031996536],"genre_scores_gemma":[0.12560502,0.8638778,0.0067283064,0.0017364937,0.00048312772,0.00053607236,0.0006974773,0.000067614295,0.00026817466],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.98145515,0.008510297,0.005925768,0.0020856455,0.001780754,0.00024244077],"domain_scores_gemma":[0.8801349,0.10622151,0.0061650565,0.0041526044,0.0030727324,0.00025326485],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.032787926,0.0018086996,0.007772412,0.0043611955,0.00052516366,0.0029663201,0.0021519123,0.0015057342,0.0020326453],"category_scores_gemma":[0.08771399,0.0008096583,0.011296459,0.005873376,0.0009820954,0.0021144706,0.0015335483,0.0016582175,0.00024740832],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00097289734,0.000047550962,0.015520223,0.4439362,0.3311031,0.0004811141,0.00041931556,0.0010517221,0.0010705311,0.0021837717,0.002485776,0.20072785],"study_design_scores_gemma":[0.00039162568,0.0005822287,0.060393628,0.112851806,0.76820093,0.0017210508,0.0005356918,0.0007371054,0.0017469752,0.0107044345,0.04202999,0.000104607105],"about_ca_topic_score_codex":0.0020512696,"about_ca_topic_score_gemma":0.004995257,"teacher_disagreement_score":0.9922276,"about_ca_system_score_codex":0.0015528759,"about_ca_system_score_gemma":0.0026999947,"threshold_uncertainty_score":0.17340124},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W2889636695","doi":"10.1002/lom3.10273","title":"Methodological review and meta‐analysis of dilution assays for estimates of virus‐ and grazer‐mediated phytoplankton mortality","year":2018,"lang":"en","type":"article","venue":"Limnology and Oceanography Methods","topic":"Microbial Community Ecology and Physiology","field":"Environmental Science","cited_by":22,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Dilution; Virus; Mortality rate; Biology; Meta-analysis; Demography; Virology; Medicine; Internal medicine","score_opus":0.187349636260773,"score_gpt":0.43220506717327944,"score_spread":0.24485543091250644,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2889636695","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0035027955,0.9840013,0.009339394,0.001080606,0.0005286198,0.00033710254,0.00080926047,0.00007106086,0.0003298848],"genre_scores_gemma":[0.28609222,0.67027444,0.030939896,0.0036809375,0.0013976458,0.0037529771,0.0028285868,0.00034973133,0.0006835735],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.8950748,0.07452492,0.01740714,0.006499565,0.005861713,0.00063188217],"domain_scores_gemma":[0.70325774,0.25724646,0.017968925,0.012333344,0.008559522,0.0006340066],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.12487763,0.0035833507,0.011805563,0.0066356547,0.00071656826,0.004723509,0.0040424177,0.0027762754,0.0027932145],"category_scores_gemma":[0.28152642,0.001984845,0.035787217,0.008416833,0.0015906631,0.002326951,0.0021441795,0.0025696775,0.0003519209],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012270146,0.000023677041,0.009474565,0.20686096,0.74387014,0.00011318371,0.00016848282,0.0012682814,0.00027887532,0.0010066106,0.0015702558,0.03413796],"study_design_scores_gemma":[0.000504902,0.00032541118,0.0126852,0.045518074,0.927867,0.0001751779,0.00011440134,0.0009364726,0.00055964355,0.0024624884,0.008776509,0.000074753756],"about_ca_topic_score_codex":0.0063939407,"about_ca_topic_score_gemma":0.009638599,"teacher_disagreement_score":0.12487763,"about_ca_system_score_codex":0.0034664662,"about_ca_system_score_gemma":0.0043130005,"threshold_uncertainty_score":0.66042393},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2890588440","doi":"10.1016/j.jclinepi.2018.08.019","title":"Lack of transparency in reporting narrative synthesis of quantitative data: a methodological assessment of systematic reviews","year":2018,"lang":"en","type":"article","venue":"Journal of Clinical Epidemiology","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":162,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Medical Research Council; McMaster University","keywords":"Transparency (behavior); Systematic review; Narrative review; Medicine; Narrative; MEDLINE; Accounting; Computer science; Political science; Business; Intensive care medicine; Computer security","score_opus":0.9948432497923828,"score_gpt":0.8785193252618085,"score_spread":0.11632392453057427,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2890588440","genre_codex":"methods","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.041725565,0.28538993,0.45900568,0.10608516,0.012898876,0.071959674,0.007494819,0.0011995095,0.0142408],"genre_scores_gemma":[0.39609882,0.0657364,0.3601454,0.016700866,0.004456636,0.15141842,0.0036228169,0.0007235345,0.0010971215],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.026800634,0.6523475,0.25731817,0.007538378,0.054949783,0.0010454508],"domain_scores_gemma":[0.013554587,0.7413421,0.14482602,0.03629844,0.06304873,0.0009300756],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.871401,0.003582998,0.008868902,0.027223354,0.005698867,0.022195619,0.00877754,0.008696199,0.0040215785],"category_scores_gemma":[0.96676666,0.008105522,0.014492045,0.024828244,0.015351042,0.024049886,0.01783698,0.009732772,0.001156451],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00543992,0.0002764854,0.047997516,0.42597234,0.027843388,0.00068563205,0.06696896,0.005778517,0.001985321,0.04756487,0.025033299,0.34445375],"study_design_scores_gemma":[0.0037047644,0.0022493282,0.025233116,0.71614724,0.02376121,0.0019475179,0.01282619,0.016991405,0.0066266456,0.101414345,0.08757298,0.0015252278],"about_ca_topic_score_codex":0.0074401386,"about_ca_topic_score_gemma":0.006211009,"teacher_disagreement_score":0.12859899,"about_ca_system_score_codex":0.021472897,"about_ca_system_score_gemma":0.05737384,"threshold_uncertainty_score":0.15858549},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"systematic_review","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2892809183","doi":"10.1016/j.apmr.2018.07.415","title":"Trends in Randomized Controlled Trials Of Technology-Based Interventions For Upper Limb Motor Recovery Post Stroke","year":2018,"lang":"en","type":"article","venue":"Archives of Physical Medicine and Rehabilitation","topic":"Stroke Rehabilitation and Recovery","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Lawson Health Research Institute","funders":"","keywords":"Physical medicine and rehabilitation; Rehabilitation; Stroke (engine); Psychological intervention; Upper limb; Motor function; Randomized controlled trial; Physical therapy; Medicine; Surgery; Engineering","score_opus":0.02007942061728634,"score_gpt":0.35263441828369424,"score_spread":0.3325549976664079,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2892809183","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.13395679,0.82997715,0.0038830964,0.010029142,0.0031565996,0.001643353,0.009171277,0.0002588648,0.007923801],"genre_scores_gemma":[0.77326554,0.19431171,0.005568698,0.009096191,0.0021877722,0.003603851,0.00941655,0.00026308163,0.0022866568],"study_design_codex":"randomized_trial","study_design_gemma":"systematic_review","domain_scores_codex":[0.9367489,0.0239634,0.02567295,0.0045321025,0.0070733447,0.0020092756],"domain_scores_gemma":[0.71914566,0.19179514,0.05762054,0.00559179,0.023098508,0.0027483846],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.061792642,0.0008720588,0.005661511,0.0033303816,0.0008031509,0.0056964443,0.0012186778,0.0037435738,0.0054145805],"category_scores_gemma":[0.2222583,0.001223792,0.008781537,0.0071276114,0.001640458,0.0056689805,0.0014449231,0.0054332525,0.00071658095],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.4662184,0.0011269113,0.07647726,0.10504995,0.11369056,0.00016529506,0.0010679759,0.0017816896,0.002117563,0.003664324,0.020880796,0.20775932],"study_design_scores_gemma":[0.21219571,0.03508341,0.40153325,0.058019824,0.21659617,0.0010304395,0.0013478453,0.0033805033,0.0033659164,0.008202543,0.058839615,0.00040481272],"about_ca_topic_score_codex":0.0054984638,"about_ca_topic_score_gemma":0.00742852,"teacher_disagreement_score":0.9382074,"about_ca_system_score_codex":0.0037325483,"about_ca_system_score_gemma":0.005684702,"threshold_uncertainty_score":0.32679462},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["bibliometrics"],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2901357801","doi":"10.1186/s12891-018-2337-7","title":"Pilot randomized controlled trials in the orthopaedic surgery literature: a systematic review","year":2018,"lang":"en","type":"review","venue":"BMC Musculoskeletal Disorders","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Impact; Western University; University of Toronto; McMaster University","funders":"","keywords":"Medicine; Randomized controlled trial; MEDLINE; Checklist; Physical therapy; Systematic review; Clinical trial; Orthopedic surgery; Consolidated Standards of Reporting Trials; Evidence-based medicine; Alternative medicine; Surgery; Internal medicine; Pathology","score_opus":0.4697155841341039,"score_gpt":0.4995283016012352,"score_spread":0.029812717467131267,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2901357801","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.001212313,0.99087954,0.0012917304,0.0011659539,0.00062632613,0.003739254,0.0004962643,0.000034899298,0.0005536609],"genre_scores_gemma":[0.030849872,0.9440744,0.008986608,0.0033281413,0.00079753296,0.011069065,0.00063140097,0.000036911115,0.00022604805],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.79036486,0.10231773,0.07412605,0.006587437,0.025209062,0.0013948692],"domain_scores_gemma":[0.3510825,0.5484289,0.0651579,0.009758064,0.024098625,0.0014739829],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.13958627,0.0023032178,0.013087361,0.020924436,0.0014488292,0.0075050066,0.0041511157,0.0053728665,0.0064293514],"category_scores_gemma":[0.4575284,0.0029511282,0.010211845,0.019755345,0.004045583,0.0105133355,0.003473086,0.0034276666,0.0010710413],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00026141584,0.000028388995,0.0007706673,0.961104,0.0060874694,0.00006417351,0.00029242475,0.00007803233,0.0001012135,0.0005090381,0.0015693165,0.029133825],"study_design_scores_gemma":[0.0005826973,0.00026396415,0.0018606184,0.9572649,0.027259314,0.00023782694,0.00035269678,0.000112863505,0.00013817726,0.0008443329,0.011039726,0.00004291969],"about_ca_topic_score_codex":0.003008585,"about_ca_topic_score_gemma":0.010796132,"teacher_disagreement_score":0.98691267,"about_ca_system_score_codex":0.0076233745,"about_ca_system_score_gemma":0.035877198,"threshold_uncertainty_score":0.7382116},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2901826044","doi":"10.1093/ije/dyy239","title":"How often can meta-analyses of individual-level data individualize treatment? A meta-epidemiologic study","year":2018,"lang":"en","type":"review","venue":"International Journal of Epidemiology","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":32,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital","funders":"Canadian Institutes of Health Research; Nederlandse Organisatie voor Wetenschappelijk Onderzoek; Laura and John Arnold Foundation","keywords":"Subgroup analysis; Meta-analysis; Medicine; Cochrane Library; MEDLINE; Randomized controlled trial; Internal medicine; Biology","score_opus":0.9907924205974797,"score_gpt":0.7172303309820024,"score_spread":0.27356208961547734,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2901826044","genre_codex":"review","genre_gemma":"review","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008740398,0.8814977,0.08333945,0.014640399,0.005641197,0.0017513251,0.0015935574,0.00054112915,0.0022549157],"genre_scores_gemma":[0.5269903,0.2597743,0.16143191,0.032120794,0.00757797,0.008206721,0.0017430376,0.0010025033,0.0011524623],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.46066448,0.4706745,0.0372767,0.01613431,0.014138237,0.0011117434],"domain_scores_gemma":[0.32519642,0.6025525,0.029185308,0.035708785,0.0067340457,0.00062308076],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.35538414,0.004620461,0.021549731,0.011621831,0.0011101094,0.008979626,0.0062183416,0.008204185,0.0030904291],"category_scores_gemma":[0.64210296,0.0039378214,0.04783996,0.012094789,0.003467173,0.00962794,0.0040533254,0.0066257417,0.0009014928],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002329902,0.000045808865,0.018138677,0.08345544,0.82120335,0.00023635713,0.00062170817,0.0032762298,0.00028749014,0.004935831,0.0069561037,0.05851313],"study_design_scores_gemma":[0.0032992626,0.00067601923,0.011065056,0.0552571,0.84000605,0.00044161742,0.00040524403,0.005452754,0.0012009057,0.05891973,0.022902159,0.00037407153],"about_ca_topic_score_codex":0.0026365872,"about_ca_topic_score_gemma":0.0033694776,"teacher_disagreement_score":0.9784503,"about_ca_system_score_codex":0.0034995768,"about_ca_system_score_gemma":0.0033870554,"threshold_uncertainty_score":0.7949261},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2902864468","doi":"10.1111/jebm.12332","title":"Traffic accidents, maxillofacial injuries and risk factors: A systematic review of observational studies","year":2018,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Facial Trauma and Fracture Management","field":"Medicine","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Observational study; Medicine; Web of science; Cochrane Library; Poison control; Demography; Injury prevention; Meta-analysis; Environmental health; Internal medicine","score_opus":0.40014623957982814,"score_gpt":0.4597323569938587,"score_spread":0.05958611741403058,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2902864468","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0042072684,0.9944622,0.00019913467,0.00022320004,0.0000969381,0.00030834522,0.00036316807,0.000004819185,0.00013501679],"genre_scores_gemma":[0.05861397,0.93750626,0.0014132726,0.0006230214,0.00018319713,0.0010999012,0.00045969486,0.0000058317733,0.0000948416],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9803388,0.0056278026,0.009830977,0.0012383267,0.0025996412,0.00036441715],"domain_scores_gemma":[0.9019985,0.0685286,0.021903317,0.0014637813,0.005533085,0.00057274307],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.016763926,0.0017084487,0.008548334,0.016315391,0.00090229744,0.003244498,0.0017196005,0.0024226902,0.0031128014],"category_scores_gemma":[0.0688396,0.0015633606,0.008524862,0.016925296,0.0012188769,0.0027920618,0.0015001813,0.0011384077,0.00023579206],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00022310358,0.000027749315,0.0069718305,0.9581485,0.019587394,0.00017810112,0.0002974256,0.00006776002,0.00014564498,0.00010942574,0.00044609673,0.013797138],"study_design_scores_gemma":[0.00031668536,0.00040188374,0.024449598,0.7961835,0.16843265,0.0010481314,0.0009602747,0.00012381127,0.00019904459,0.00031850068,0.0075062797,0.000059662892],"about_ca_topic_score_codex":0.005617793,"about_ca_topic_score_gemma":0.014399854,"teacher_disagreement_score":0.016763926,"about_ca_system_score_codex":0.0025394727,"about_ca_system_score_gemma":0.010279061,"threshold_uncertainty_score":0.0886572},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2911968494","doi":"10.1136/bmjebm-2018-111057","title":"Use of modified intention-to-treat analysis in studies of direct oral anticoagulants and risk of selection bias: a systematic review","year":2019,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Atrial Fibrillation Management and Outcomes","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre Hospitalier de l’Université de Montréal","funders":"Fonds de Recherche du Québec - Santé","keywords":"Selection (genetic algorithm); Risk analysis (engineering); Selection bias; Medicine; Intensive care medicine; Psychology; Actuarial science; Computer science; Business; Artificial intelligence; Pathology","score_opus":0.5686784527537067,"score_gpt":0.5045638600117248,"score_spread":0.06411459274198184,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2911968494","genre_codex":"review","genre_gemma":"review","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0058078594,0.8283394,0.045791768,0.003719161,0.004755591,0.10579068,0.0033360445,0.0005808411,0.0018786954],"genre_scores_gemma":[0.15070623,0.21847288,0.15787537,0.008169284,0.0018476213,0.4591471,0.002182031,0.00038465252,0.0012148943],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.37677386,0.3963726,0.15953498,0.020718534,0.045105726,0.0014943678],"domain_scores_gemma":[0.3592418,0.51686525,0.081887424,0.020096533,0.020613734,0.0012952499],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.33654216,0.004568919,0.035400894,0.02557627,0.0022527387,0.00951536,0.007784355,0.010070801,0.0061234515],"category_scores_gemma":[0.5951705,0.00403466,0.036942676,0.019918203,0.007163558,0.010913661,0.0045663808,0.0046970947,0.0009691797],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009495462,0.000045350975,0.0018898385,0.8853156,0.07637611,0.00012962325,0.00071381713,0.0003528645,0.00014050826,0.0016847361,0.0019249099,0.03047708],"study_design_scores_gemma":[0.0077197775,0.0012123806,0.0041318196,0.7441987,0.20964512,0.00045483292,0.0007709154,0.003594737,0.000925353,0.011015952,0.015921347,0.00040907625],"about_ca_topic_score_codex":0.0032180732,"about_ca_topic_score_gemma":0.0062249578,"teacher_disagreement_score":0.96459913,"about_ca_system_score_codex":0.013065534,"about_ca_system_score_gemma":0.01951845,"threshold_uncertainty_score":0.8181616},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_narrow","metaepi_broad","metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2918006074","doi":"10.14740/cr813","title":"Anti-Hypertensive Effect of Sacubitril/Valsartan: A Meta-Analysis of Randomized Controlled Trials","year":2019,"lang":"en","type":"article","venue":"Cardiology Research","topic":"Blood Pressure and Hypertension Studies","field":"Medicine","cited_by":22,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Medicine; Valsartan; Blood pressure; Sacubitril, Valsartan; Sacubitril; Ambulatory blood pressure; Cardiology; Internal medicine; Olmesartan; Randomized controlled trial; Heart failure; Calcium channel blocker; ACE inhibitor; Angiotensin-converting enzyme","score_opus":0.21121824656528804,"score_gpt":0.4462133553230914,"score_spread":0.23499510875780338,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2918006074","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.020111486,0.97479194,0.0022625488,0.00036029637,0.00038402173,0.0005428698,0.0010048874,0.000072382914,0.00046954662],"genre_scores_gemma":[0.58703345,0.3974941,0.008026094,0.0013886428,0.00065981573,0.0023911064,0.0022698555,0.00010307297,0.0006338733],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.97871196,0.012529876,0.004771166,0.0019359016,0.0015485203,0.0005025759],"domain_scores_gemma":[0.97263056,0.020533547,0.0037883604,0.0012911288,0.0014059508,0.00035041902],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.023544239,0.003849369,0.021007385,0.006358968,0.00062670343,0.0034307009,0.0022465095,0.002183428,0.0035069713],"category_scores_gemma":[0.03545982,0.001700741,0.060082823,0.0069644586,0.0006564892,0.001888779,0.001405276,0.0023294063,0.0003267973],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004007022,0.000025288036,0.0038279104,0.055387467,0.932566,0.000080804806,0.000026378862,0.00060428365,0.00014857217,0.000071140516,0.00020525676,0.003049979],"study_design_scores_gemma":[0.0006824743,0.0001373614,0.0020941538,0.002092838,0.9941309,0.00003481606,0.0000143533825,0.0002588049,0.000065007385,0.0001247871,0.00035339553,0.0000110997835],"about_ca_topic_score_codex":0.0036780003,"about_ca_topic_score_gemma":0.005729342,"teacher_disagreement_score":0.023544239,"about_ca_system_score_codex":0.0020015952,"about_ca_system_score_gemma":0.0019044655,"threshold_uncertainty_score":0.124515295},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2928825510","doi":"10.1097/corr.0000000000000690","title":"Is the Risk of Infection Lower with Sutures than with Staples for Skin Closure After Orthopaedic Surgery? A Meta-analysis of Randomized Trials","year":2019,"lang":"en","type":"review","venue":"Clinical Orthopaedics and Related Research","topic":"Surgical Sutures and Adhesives","field":"Medicine","cited_by":48,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Medicine; Orthopedic surgery; Surgery; Randomized controlled trial; Meta-analysis; Sports medicine; Closure (psychology); MEDLINE; Physical therapy; Internal medicine","score_opus":0.3512710627236364,"score_gpt":0.5029919533418621,"score_spread":0.1517208906182257,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2928825510","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.009332663,0.9866579,0.0014998262,0.0006147859,0.00047693838,0.00041372527,0.0005556339,0.000038368806,0.00041013857],"genre_scores_gemma":[0.39209664,0.5919461,0.0070899283,0.003250502,0.0010370153,0.0026697032,0.001304337,0.00010994734,0.0004958017],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9476947,0.03141506,0.012567756,0.004298487,0.0032220983,0.00080181693],"domain_scores_gemma":[0.89994454,0.07986931,0.012367046,0.003833842,0.0032684295,0.00071684265],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.04737382,0.0036228232,0.024369827,0.00794409,0.00074855814,0.006024524,0.0028434505,0.0035494368,0.0037748234],"category_scores_gemma":[0.113254726,0.0022369951,0.07185629,0.007893975,0.001257906,0.0029107942,0.0018137014,0.0035397557,0.0004260245],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0034736819,0.000019456993,0.0050014034,0.08642709,0.8992638,0.000074285,0.00004542987,0.0003940396,0.000103124876,0.00013023928,0.0002913981,0.004775937],"study_design_scores_gemma":[0.00096451974,0.00021892824,0.0024835777,0.008109559,0.9869753,0.000056188434,0.000028357761,0.00016236535,0.0000893573,0.00034020646,0.0005533953,0.000018319206],"about_ca_topic_score_codex":0.0048776874,"about_ca_topic_score_gemma":0.005650044,"teacher_disagreement_score":0.04737382,"about_ca_system_score_codex":0.003402407,"about_ca_system_score_gemma":0.0028023536,"threshold_uncertainty_score":0.25053966},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2936104433","doi":"10.1136/bmjsem-2019-000512","title":"Population attributable fraction of leading non-communicable cardiovascular diseases due to leisure-time physical inactivity: a systematic review","year":2019,"lang":"en","type":"review","venue":"BMJ Open Sport & Exercise Medicine","topic":"Physical Activity and Health","field":"Medicine","cited_by":27,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Vancouver Coastal Health Research Institute; University of British Columbia","funders":"Institute of Aging; Canadian Institutes of Health Research; Michael Smith Health Research BC","keywords":"Attributable risk; Physical activity; Medicine; Population; Environmental health; Fraction (chemistry); Gerontology; Physical therapy; Chemistry","score_opus":0.11255515656908595,"score_gpt":0.4393836801993876,"score_spread":0.3268285236303017,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2936104433","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0012379079,0.99670357,0.00035575195,0.0002536535,0.00012599595,0.0004013885,0.00066752743,0.000013424207,0.00024076844],"genre_scores_gemma":[0.040299173,0.9552001,0.0014020337,0.0007493443,0.00015113795,0.0015336372,0.00052130787,0.000010925648,0.0001323547],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9840515,0.0064879945,0.0049493196,0.0014767895,0.0026796511,0.00035471097],"domain_scores_gemma":[0.9459838,0.041601785,0.008264034,0.0006902972,0.0031879547,0.00027203007],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.015517264,0.0019075839,0.011095272,0.011317147,0.0004783228,0.00300457,0.0021496324,0.0021450173,0.0033486853],"category_scores_gemma":[0.07833295,0.0016507376,0.013451344,0.009704163,0.0010140694,0.0022603679,0.0017163596,0.0017826799,0.00027056452],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000138465,0.000012323749,0.0024374686,0.94167745,0.036358003,0.00007724043,0.00013320774,0.00014917206,0.00006569078,0.00023222847,0.00097419316,0.01774438],"study_design_scores_gemma":[0.00028506535,0.00019564714,0.011464401,0.73003995,0.2464028,0.00040829834,0.00026004095,0.0001993213,0.00016398537,0.0004594331,0.01007063,0.000050413717],"about_ca_topic_score_codex":0.011583962,"about_ca_topic_score_gemma":0.023895567,"teacher_disagreement_score":0.015517264,"about_ca_system_score_codex":0.003673685,"about_ca_system_score_gemma":0.01044585,"threshold_uncertainty_score":0.08206409},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2938572552","doi":"10.1016/j.spinee.2019.04.004","title":"Measurement properties of walking outcome measures for neurogenic claudication: a systematic review and meta analysis","year":2019,"lang":"en","type":"review","venue":"The Spine Journal","topic":"Peripheral Artery Disease Management","field":"Medicine","cited_by":23,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Mount Royal University","funders":"","keywords":"Medicine; Physical therapy; Intraclass correlation; Neurogenic claudication; CINAHL; Meta-analysis; Context (archaeology); Intermittent claudication; Physical medicine and rehabilitation; MEDLINE; Quality of life (healthcare); Checklist; Claudication; Systematic review; Lumbar spinal stenosis; Psychometrics; Internal medicine; Surgery; Psychological intervention; Lumbar; Arterial disease","score_opus":0.31576574117060874,"score_gpt":0.3746876458398428,"score_spread":0.05892190466923408,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2938572552","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008915012,0.9897261,0.00061896996,0.00009619521,0.0001324473,0.00015342893,0.00021879161,0.000019555137,0.00011950242],"genre_scores_gemma":[0.34597415,0.64588237,0.004737647,0.001114156,0.000353937,0.0007988267,0.00074804167,0.000051937805,0.0003388615],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.983727,0.0072542657,0.004993874,0.0018579644,0.0017421698,0.00042470943],"domain_scores_gemma":[0.9631009,0.02857918,0.005204937,0.0013559074,0.001530196,0.00022882159],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.021343205,0.00345755,0.0217918,0.0043807975,0.0007723612,0.0041031507,0.0024689175,0.0029937255,0.0028622388],"category_scores_gemma":[0.052844718,0.0019377484,0.042711746,0.005851683,0.0012876146,0.0025244446,0.0023822756,0.0029510637,0.00029257455],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0024844273,0.00004157317,0.0066624717,0.15939802,0.8190387,0.000060786362,0.00009496556,0.0003167101,0.00014612028,0.000075418255,0.000278793,0.011402097],"study_design_scores_gemma":[0.00069568603,0.00020808702,0.0049929773,0.007760021,0.9853674,0.000060326554,0.000053456715,0.00015629786,0.000059220813,0.00013498927,0.00048458937,0.000026923031],"about_ca_topic_score_codex":0.0061798412,"about_ca_topic_score_gemma":0.0145214945,"teacher_disagreement_score":0.0217918,"about_ca_system_score_codex":0.002372501,"about_ca_system_score_gemma":0.0029998815,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2939565443","doi":"10.5716/wp19006.pdf","title":"What do we really know about the impacts of improved grain legumes and dryland cereals? A critical review of 18 impact studies","year":2019,"lang":"en","type":"review","venue":"","topic":"Food Security and Health in Diverse Populations","field":"Health Professions","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"York University; University of British Columbia","funders":"Universidade Estadual de Campinas","keywords":"Agronomy; Agroforestry; Environmental science; Agricultural engineering; Biology; Engineering","score_opus":0.35878354271755236,"score_gpt":0.581311468393108,"score_spread":0.22252792567555563,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2939565443","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00006995899,0.9993001,0.000030303552,0.00030877508,0.000084569845,0.000008045923,0.00003760578,0.0000015208904,0.00015896077],"genre_scores_gemma":[0.0006593005,0.99886644,0.00007009595,0.0002696774,0.000054681594,0.000012042406,0.000026423437,8.652566e-7,0.000040573566],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9980579,0.0005174012,0.0006126446,0.00022118227,0.00049935875,0.00009150494],"domain_scores_gemma":[0.98807794,0.009089479,0.0011978059,0.00014432095,0.0013240635,0.00016639319],"candidate_categories":["metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.0040507508,0.0010198341,0.0032116063,0.007864457,0.00046722827,0.0022171238,0.0014156348,0.0016675711,0.0051048505],"category_scores_gemma":[0.014886973,0.0006811561,0.0026439317,0.008174093,0.00082759047,0.0024079815,0.0010201533,0.0015681558,0.0006011165],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013601121,0.00003551511,0.00038742,0.5835484,0.0013255962,0.00010659671,0.00032458623,0.0001641488,0.0003008343,0.0014164222,0.012452779,0.39980173],"study_design_scores_gemma":[0.00005985173,0.00021920013,0.0038945335,0.5820114,0.008679642,0.00054684916,0.0006966594,0.00007057444,0.0002845197,0.001784809,0.4017123,0.000039615],"about_ca_topic_score_codex":0.0068628187,"about_ca_topic_score_gemma":0.013612069,"teacher_disagreement_score":0.9967884,"about_ca_system_score_codex":0.0018214189,"about_ca_system_score_gemma":0.00774918,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"opus","categories":["metaepi_broad"],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2964854296","doi":"10.14740/jocmr3887","title":"Simultaneous, Delayed and Liver-First Hepatic Resections for Synchronous Colorectal Liver Metastases: A Systematic Review and Network Meta-Analysis","year":2019,"lang":"en","type":"review","venue":"Journal of Clinical Medicine Research","topic":"Hepatocellular Carcinoma Treatment and Prognosis","field":"Medicine","cited_by":25,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Medicine; Meta-analysis; Hazard ratio; Confidence interval; Pairwise comparison; Ranking (information retrieval); Internal medicine; Randomized controlled trial; Oncology; Systematic review; Overall survival; Surgery; Gastroenterology; MEDLINE; Statistics; Artificial intelligence; Computer science; Mathematics","score_opus":0.4932523526304464,"score_gpt":0.5143128979229559,"score_spread":0.021060545292509447,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2964854296","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.009395028,0.9876546,0.0011925607,0.00030467074,0.00016232223,0.00029653648,0.00063608476,0.000024825258,0.00033328985],"genre_scores_gemma":[0.36156204,0.6291559,0.004882788,0.00095873204,0.00041767047,0.001440643,0.0012282241,0.0000358985,0.0003181601],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9885388,0.006584076,0.0022841142,0.0011791405,0.0010705227,0.00034330937],"domain_scores_gemma":[0.97569126,0.019093215,0.0030193306,0.0005795617,0.0013680883,0.0002486399],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0155331595,0.0026435645,0.0141969295,0.0061537763,0.00059404265,0.0028217097,0.0016431191,0.0018069707,0.0028242078],"category_scores_gemma":[0.034243967,0.0011463588,0.04029171,0.005673057,0.0006124556,0.0017039542,0.0013047633,0.0020486885,0.00022125074],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0028579878,0.000034202483,0.009153178,0.2253875,0.7468787,0.00012400278,0.00006757325,0.0015888312,0.00020959773,0.00019729114,0.00061035354,0.012890728],"study_design_scores_gemma":[0.00061052165,0.00018424328,0.0045016296,0.013852804,0.9790587,0.00007258474,0.000032635984,0.0004889122,0.00008188199,0.00024998892,0.0008472326,0.000018857643],"about_ca_topic_score_codex":0.007298645,"about_ca_topic_score_gemma":0.013737099,"teacher_disagreement_score":0.0155331595,"about_ca_system_score_codex":0.0023816626,"about_ca_system_score_gemma":0.003023182,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2974839292","doi":"10.1136/bmj.l5221","title":"Design characteristics, risk of bias, and reporting of randomised controlled trials supporting approvals of cancer drugs by European Medicines Agency, 2014-16: cross sectional analysis","year":2019,"lang":"en","type":"article","venue":"BMJ","topic":"Statistical Methods in Clinical Trials","field":"Mathematics","cited_by":244,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"Medical Research Council; Economic and Social Research Council; U.S. Department of the Treasury; National Institute for Health and Care Research; Bill and Melinda Gates Foundation; Harvard University; National Institutes of Health; National Science Foundation","keywords":"Medicine; Clinical trial; Randomized controlled trial; Reporting bias; Research design; Clinical endpoint; Publication bias; MEDLINE; Family medicine; Meta-analysis; Internal medicine; Statistics","score_opus":0.4469013499055617,"score_gpt":0.5635049915159488,"score_spread":0.11660364161038711,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2974839292","genre_codex":"methods","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.27301952,0.20656432,0.27778167,0.012474307,0.0048305104,0.16803256,0.045518912,0.0010500854,0.010728066],"genre_scores_gemma":[0.68414664,0.008750313,0.109876476,0.00598193,0.0007435187,0.18206275,0.007452783,0.00024340698,0.0007421763],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.121756405,0.4970197,0.32072902,0.018093497,0.04031717,0.0020842208],"domain_scores_gemma":[0.05955838,0.5664591,0.29469904,0.04283508,0.035164177,0.0012841924],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.64502853,0.0021913,0.006266205,0.011983756,0.00196811,0.005883213,0.0029970775,0.005164702,0.0062286607],"category_scores_gemma":[0.84496635,0.0026762113,0.012130256,0.015534275,0.005229427,0.006059894,0.0038309693,0.0028850466,0.0010072936],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.032352183,0.0007589188,0.5050488,0.17192246,0.11456495,0.000658588,0.0076038805,0.004748803,0.0012888012,0.015587304,0.021922128,0.123543136],"study_design_scores_gemma":[0.028748099,0.013651087,0.5120597,0.15448421,0.10966749,0.0036911792,0.0045440616,0.024652885,0.010135282,0.052331682,0.08415463,0.0018796813],"about_ca_topic_score_codex":0.001538478,"about_ca_topic_score_gemma":0.0015101538,"teacher_disagreement_score":0.35497147,"about_ca_system_score_codex":0.0056770984,"about_ca_system_score_gemma":0.008222014,"threshold_uncertainty_score":0.43774307},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W2978471806","doi":"10.1093/cdn/nzz104","title":"Methods for the Selection of Covariates in Nutritional Epidemiology Studies: A Meta-Epidemiological Review","year":2019,"lang":"en","type":"review","venue":"Current Developments in Nutrition","topic":"Nutrition and Health in Aging","field":"Medicine","cited_by":40,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University; Western University; University of Ottawa; Population Health Research Institute; McMaster University; Impact","funders":"Canadian Institutes of Health Research","keywords":"Epidemiology; Meta-analysis; Covariate; Selection (genetic algorithm); Nutritional epidemiology; Medicine; Computer science; Statistics; Internal medicine; Mathematics","score_opus":0.765916361887563,"score_gpt":0.6576582437527565,"score_spread":0.10825811813480646,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2978471806","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00013768497,0.99602616,0.0027732651,0.0004030431,0.00016244728,0.00018974576,0.00014532976,0.000022053033,0.00014026782],"genre_scores_gemma":[0.005871375,0.9811866,0.010881458,0.00052288786,0.0002544766,0.000842348,0.00029801764,0.000033169552,0.00010964397],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9327215,0.043429673,0.014239131,0.0026855406,0.0065102335,0.00041400647],"domain_scores_gemma":[0.7164512,0.2499243,0.016932853,0.006092228,0.010164826,0.00043466667],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.09448994,0.0033839536,0.013452408,0.014991973,0.00070614205,0.0049938094,0.0043754866,0.0033350268,0.004352041],"category_scores_gemma":[0.22115366,0.0022049127,0.019798866,0.011754616,0.0013714671,0.0036140808,0.002672968,0.0031477278,0.00083346176],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00055159914,0.000048729937,0.003120288,0.55906254,0.09500312,0.000106225845,0.00021342839,0.0011322494,0.00020856845,0.002716691,0.005174535,0.33266202],"study_design_scores_gemma":[0.00075807294,0.00036743146,0.006888408,0.5717888,0.32378137,0.00055497565,0.00022489174,0.0017678203,0.0007811203,0.0111305555,0.08175811,0.00019846788],"about_ca_topic_score_codex":0.004127456,"about_ca_topic_score_gemma":0.0062657744,"teacher_disagreement_score":0.9865476,"about_ca_system_score_codex":0.0036663918,"about_ca_system_score_gemma":0.008299247,"threshold_uncertainty_score":0.49971652},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W2987382174","doi":"10.7554/elife.08351.015","title":"Author response: A meta-analysis of threats to valid clinical inference in preclinical research of sunitinib","year":2015,"lang":"en","type":"peer-review","venue":"","topic":"Animal testing and alternatives","field":"Veterinary","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital; McGill University","funders":"","keywords":"Sunitinib; Inference; Meta-analysis; Computational biology; Psychology; Medicine; Bioinformatics; Oncology; Biology; Data science; Computer science; Internal medicine; Artificial intelligence; Cancer","score_opus":0.9606442935144826,"score_gpt":0.7394399394582948,"score_spread":0.22120435405618777,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2987382174","genre_codex":"commentary","genre_gemma":"commentary","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"commentary","genre_consensus":"commentary","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0017050424,0.012314173,0.011950181,0.8526405,0.08434144,0.0027070947,0.025221117,0.0009943395,0.008126169],"genre_scores_gemma":[0.023338743,0.0046976823,0.01434004,0.9112087,0.020791426,0.011823247,0.0030637323,0.00092135847,0.009815164],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.8275011,0.115020595,0.0292213,0.0076940195,0.018277727,0.0022853299],"domain_scores_gemma":[0.3686626,0.5381138,0.026053939,0.019485535,0.045850236,0.001833945],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.15244864,0.0019132394,0.004609351,0.0037620736,0.001962996,0.0042232336,0.0040364056,0.01886271,0.04856855],"category_scores_gemma":[0.67980844,0.0019649635,0.010776223,0.0060641286,0.0023566766,0.0050571878,0.0040790853,0.013306016,0.010950624],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012612572,0.000023812352,0.002109201,0.01475795,0.002669464,0.0002420187,0.0003922846,0.0005844697,0.00014139475,0.00394979,0.9464448,0.027423644],"study_design_scores_gemma":[0.0064826743,0.0003823968,0.0084644,0.052470353,0.0099126855,0.0006886127,0.0007377544,0.0029221901,0.001714287,0.049476918,0.86621296,0.00053473597],"about_ca_topic_score_codex":0.006276477,"about_ca_topic_score_gemma":0.007219703,"teacher_disagreement_score":0.84755135,"about_ca_system_score_codex":0.0065270187,"about_ca_system_score_gemma":0.00995779,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"grok","categories":[],"domain":null,"study_design":"not_applicable","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"opus","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2988645757","doi":"10.1186/s40814-019-0500-z","title":"The reporting of progression criteria in protocols of pilot trials designed to assess the feasibility of main trials is insufficient: a meta-epidemiological study","year":2019,"lang":"en","type":"article","venue":"Pilot and Feasibility Studies","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":65,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Population Health Research Institute; McMaster University; Hamilton Health Sciences; University Health Network; Toronto General Hospital; St. Joseph’s Healthcare Hamilton; Impact","funders":"","keywords":"Medicine; Odds ratio; Confidence interval; Clinical trial; Sample size determination; Randomized controlled trial; Meta-analysis; Intervention (counseling); Odds; MEDLINE; Family medicine; Internal medicine; Logistic regression; Psychiatry; Statistics","score_opus":0.9811647581245485,"score_gpt":0.7154181071887392,"score_spread":0.2657466509358093,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2988645757","genre_codex":"review","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.019601772,0.91324586,0.033338934,0.0051127165,0.0025924025,0.019275626,0.00337277,0.00030178597,0.0031581672],"genre_scores_gemma":[0.5774308,0.2637231,0.07168889,0.009146206,0.0017156767,0.070874356,0.004019377,0.00048557372,0.0009160718],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.266629,0.41586205,0.24853809,0.015797617,0.05111818,0.0020550871],"domain_scores_gemma":[0.08032564,0.7248053,0.13346696,0.034718458,0.025767447,0.00091611466],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.55264306,0.0032827037,0.017629154,0.018035736,0.0021399262,0.015006717,0.006945587,0.0070996173,0.004596664],"category_scores_gemma":[0.8291798,0.0046354327,0.039559964,0.020745555,0.005844292,0.01617985,0.006543749,0.006092507,0.0007464333],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004443611,0.00014831613,0.023604482,0.6542334,0.25491992,0.00017021451,0.0017763877,0.0007249569,0.00021683222,0.0030016967,0.0029053278,0.053854898],"study_design_scores_gemma":[0.0063207857,0.0016742777,0.033741623,0.4711591,0.45439157,0.0007138439,0.0010090843,0.002199721,0.0011358019,0.009699844,0.01755914,0.00039523654],"about_ca_topic_score_codex":0.0039314935,"about_ca_topic_score_gemma":0.0044948068,"teacher_disagreement_score":0.98237085,"about_ca_system_score_codex":0.010758806,"about_ca_system_score_gemma":0.013779829,"threshold_uncertainty_score":0.5516708},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W3004444768","doi":"10.1089/neu.2019.6844","title":"Interventions to Optimize Spinal Cord Perfusion in Patients with Acute Traumatic Spinal Cord Injuries: A Systematic Review","year":2020,"lang":"en","type":"review","venue":"Journal of Neurotrauma","topic":"Spinal Cord Injury Research","field":"Medicine","cited_by":44,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Population Health Research Institute; McMaster University; University of Calgary","funders":"","keywords":"Medicine; Spinal cord; Adverse effect; Anesthesia; Spinal cord injury; Randomized controlled trial; Perfusion; Meta-analysis; Surgery; Internal medicine","score_opus":0.18083616643952471,"score_gpt":0.4725873115280886,"score_spread":0.29175114508856387,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3004444768","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0012235186,0.99739826,0.00014175642,0.0002381983,0.00010561451,0.00040841792,0.00020025356,0.000010787186,0.0002732171],"genre_scores_gemma":[0.021259584,0.9761143,0.0007567593,0.0006664914,0.000117698495,0.000759602,0.00020703312,0.0000059828058,0.000112530295],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99473226,0.0020776899,0.0018240634,0.00033795784,0.00085658947,0.00017143278],"domain_scores_gemma":[0.9827837,0.012747659,0.0030724523,0.00018488342,0.0010178146,0.00019345188],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005942641,0.0014356566,0.008334856,0.0042048884,0.00050269667,0.0024238983,0.0017711958,0.002051093,0.004898449],"category_scores_gemma":[0.029404154,0.00083526457,0.007604732,0.004760786,0.00058268424,0.0017739303,0.0009172968,0.0014357269,0.0003073156],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00035745755,0.000029984692,0.00044966213,0.9539156,0.012299171,0.000049045695,0.00009048111,0.00011658582,0.000083104125,0.000154746,0.00090231735,0.03155191],"study_design_scores_gemma":[0.0014188931,0.00042528933,0.0039226133,0.8488002,0.13141856,0.00023840491,0.0001789561,0.00019586335,0.00019648299,0.00039519725,0.012767596,0.000042036925],"about_ca_topic_score_codex":0.0066821296,"about_ca_topic_score_gemma":0.01719633,"teacher_disagreement_score":0.008334856,"about_ca_system_score_codex":0.0031545218,"about_ca_system_score_gemma":0.007626752,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3008884145","doi":"10.2196/15467","title":"Patient-Reported Outcome Measures in Cystic Fibrosis: Protocol for a Systematic Review","year":2020,"lang":"en","type":"review","venue":"JMIR Research Protocols","topic":"Cystic Fibrosis Research Advances","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Medicine; CINAHL; Checklist; MEDLINE; Observational study; PsycINFO; Systematic review; Prom; Protocol (science); Cochrane Library; Guideline; Patient-reported outcome; Quality of life (healthcare); Family medicine; Strengthening the reporting of observational studies in epidemiology; Randomized controlled trial; Physical therapy; Alternative medicine; Pathology; Nursing; Psychology; Psychological intervention","score_opus":0.45367009772988054,"score_gpt":0.6189611267007018,"score_spread":0.16529102897082126,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3008884145","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00029826406,0.0028920935,0.0014150023,0.00041474585,0.00031520636,0.9893338,0.0047136997,0.00010301088,0.00051418133],"genre_scores_gemma":[0.00035452985,0.0009381877,0.002776211,0.00014749062,0.000019971672,0.9951509,0.00041892595,0.000009323871,0.0001843879],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.93872446,0.027425697,0.021594798,0.0040289215,0.0061069857,0.0021191654],"domain_scores_gemma":[0.92659557,0.03026616,0.016799273,0.0061891316,0.017734393,0.0024154426],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.07888905,0.0054520303,0.017513668,0.012026803,0.0039424733,0.007522622,0.0049206563,0.006563226,0.07201874],"category_scores_gemma":[0.11225427,0.0054609263,0.019468077,0.015111661,0.0051520583,0.008017578,0.005594698,0.0089808395,0.006985649],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0047619124,0.00023605379,0.0005203542,0.9409096,0.003117658,0.00028213972,0.0011211004,0.00093554624,0.0007954805,0.0034019377,0.021420123,0.02249793],"study_design_scores_gemma":[0.03780678,0.002197699,0.006106117,0.7281779,0.013251242,0.000596802,0.002185499,0.0018501575,0.0019116704,0.010631068,0.19476505,0.0005199446],"about_ca_topic_score_codex":0.005196499,"about_ca_topic_score_gemma":0.012035105,"teacher_disagreement_score":0.07888905,"about_ca_system_score_codex":0.015513469,"about_ca_system_score_gemma":0.048501737,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":[],"domain":null,"study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W3011157861","doi":"10.1136/bmjebm-2019-111331","title":"Adjuvant-containing control arms in pivotal quadrivalent human papillomavirus vaccine trials: restoration of previously unpublished methodology","year":2020,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Cervical Cancer and HPV Research","field":"Medicine","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"U.S. Food and Drug Administration; National Institute for Health and Care Research; Laura and John Arnold Foundation; European Respiratory Society; American Association of Colleges of Pharmacy","keywords":"Clinical trial; Medicine; Blinding; Family medicine; Informed consent; Placebo; Alternative medicine; Internal medicine; Pathology","score_opus":0.5405509864717235,"score_gpt":0.5166450541185875,"score_spread":0.023905932353135984,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3011157861","genre_codex":"methods","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.057043824,0.05281457,0.6750666,0.029014098,0.01505289,0.12927125,0.021831019,0.0015505877,0.018355154],"genre_scores_gemma":[0.17658828,0.00624786,0.5938154,0.012016861,0.0037846123,0.1960841,0.007787171,0.0010174185,0.002658372],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.059374038,0.6800893,0.18979937,0.018314598,0.050819885,0.0016027077],"domain_scores_gemma":[0.041855313,0.5239198,0.14018793,0.24360722,0.048840944,0.0015887412],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.8006513,0.003158124,0.0066258293,0.028342767,0.003647125,0.016119942,0.009494014,0.0059537296,0.0052141766],"category_scores_gemma":[0.9055918,0.0037527278,0.009191978,0.028081542,0.014337198,0.013423758,0.016744735,0.00798778,0.0022661812],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.01111284,0.00078238006,0.048827264,0.119112454,0.012022664,0.0008064749,0.046479587,0.0032460652,0.0031069908,0.10005361,0.03498355,0.6194661],"study_design_scores_gemma":[0.01072576,0.005550502,0.052033406,0.14385034,0.017695291,0.001016863,0.0062343306,0.010048673,0.013524662,0.1887992,0.54914045,0.0013805453],"about_ca_topic_score_codex":0.00166671,"about_ca_topic_score_gemma":0.0025264348,"teacher_disagreement_score":0.19934869,"about_ca_system_score_codex":0.0064255297,"about_ca_system_score_gemma":0.02846826,"threshold_uncertainty_score":0.2458325},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"not_applicable","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3011957638","doi":"10.1016/j.isjp.2020.03.002","title":"Characteristics, results, and reporting of contemporary surgical trials: A systematic review and analysis","year":2020,"lang":"en","type":"review","venue":"International Journal of Surgery Protocols","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Sunnybrook Health Science Centre; University of Toronto; Health Sciences Centre; Ottawa Hospital","funders":"","keywords":"Categorical variable; Medicine; Randomized controlled trial; Protocol (science); Normality; Descriptive statistics; Quartile; Continuous variable; Sample size determination; Medical physics; Statistics; Surgery; Alternative medicine; Confidence interval; Mathematics; Internal medicine","score_opus":0.9122179476274483,"score_gpt":0.6378596023219139,"score_spread":0.2743583453055345,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3011957638","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.011898899,0.9524234,0.006166396,0.001973784,0.0004154869,0.01949722,0.006542049,0.00006748901,0.0010151785],"genre_scores_gemma":[0.19972153,0.6258455,0.03470853,0.0033394068,0.00056701846,0.12944196,0.0057475017,0.00013441872,0.00049405644],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.5804964,0.16409875,0.20291182,0.011268226,0.039063025,0.002161797],"domain_scores_gemma":[0.33190078,0.48300585,0.14003597,0.013895603,0.029273273,0.001888538],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.20613278,0.0022251403,0.01613521,0.03373764,0.0015496477,0.0067094415,0.003655905,0.0033370166,0.0041459287],"category_scores_gemma":[0.52951235,0.0024630409,0.015965706,0.042581342,0.0034074055,0.007539277,0.0035128186,0.0021359476,0.0005537297],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008945728,0.000046909627,0.010923262,0.9199214,0.034972075,0.00017778036,0.00069725537,0.00032577344,0.00014985747,0.0010094228,0.0021457558,0.028735844],"study_design_scores_gemma":[0.0016446493,0.0007605006,0.025307536,0.80817956,0.13895577,0.0008649114,0.0011915112,0.0006779709,0.00046752964,0.0031560645,0.018634824,0.00015919498],"about_ca_topic_score_codex":0.0030965055,"about_ca_topic_score_gemma":0.0070803566,"teacher_disagreement_score":0.9838648,"about_ca_system_score_codex":0.00910093,"about_ca_system_score_gemma":0.024654353,"threshold_uncertainty_score":0.9789796},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W3028180553","doi":"10.3389/fmed.2020.00280","title":"Comparative Effects Between Direct Oral Anticoagulants for Acute Venous Thromboembolism: Indirect Comparison From Randomized Controlled Trials","year":2020,"lang":"en","type":"review","venue":"Frontiers in Medicine","topic":"Venous Thromboembolism Diagnosis and Management","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Joseph’s Healthcare Hamilton; McMaster University; Impact","funders":"Guangdong Medical Research Foundation","keywords":"Medicine; Apixaban; Rivaroxaban; Edoxaban; Dabigatran; Randomized controlled trial; Intensive care medicine; Clinical trial; Internal medicine; Warfarin; Atrial fibrillation","score_opus":0.11428727362436142,"score_gpt":0.43380823222900977,"score_spread":0.31952095860464835,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3028180553","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.013163187,0.9557688,0.012633581,0.0038348674,0.0032642707,0.0056518707,0.0021735711,0.00010041073,0.003409543],"genre_scores_gemma":[0.52018243,0.38053238,0.055948626,0.012475194,0.0029848053,0.021765767,0.004416356,0.00018115347,0.0015131588],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.66136396,0.26453653,0.04965248,0.009289651,0.013948996,0.0012084502],"domain_scores_gemma":[0.40778914,0.5407554,0.031324916,0.009420953,0.0096310275,0.0010786105],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.16897345,0.003594014,0.017808858,0.008966172,0.00074446865,0.007346741,0.0032948002,0.005117644,0.011367988],"category_scores_gemma":[0.40188584,0.0016600152,0.0331152,0.0068045114,0.0035181288,0.0058210967,0.0031312273,0.005110506,0.00080605637],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.012331169,0.00017013332,0.004102963,0.60261196,0.3299212,0.0001536112,0.00038690603,0.0014291613,0.00022703824,0.0039465376,0.0016727848,0.043046545],"study_design_scores_gemma":[0.025424417,0.004593138,0.0077308337,0.19989103,0.72366416,0.00042940944,0.00039662613,0.0018395816,0.0008395967,0.021765439,0.013204233,0.0002214718],"about_ca_topic_score_codex":0.0018754757,"about_ca_topic_score_gemma":0.002575182,"teacher_disagreement_score":0.16897345,"about_ca_system_score_codex":0.004893665,"about_ca_system_score_gemma":0.005956349,"threshold_uncertainty_score":0.89362764},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3028673208","doi":"10.2147/clep.s242080","title":"&lt;p&gt;Meta-Analyses Proved Inconsistent in How Missing Data Were Handled Across Their Included Primary Trials: A Methodological Survey&lt;/p&gt;","year":2020,"lang":"en","type":"article","venue":"Clinical Epidemiology","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Veterans Affairs Canada; Sunnybrook Health Science Centre; University of Toronto; McMaster University; Health Sciences Centre; Impact","funders":"","keywords":"Missing data; Meta-analysis; Systematic review; Medicine; MEDLINE; Statistics; Clinical trial; Data mining; Computer science; Internal medicine; Mathematics","score_opus":0.9942150291509889,"score_gpt":0.7434143475047398,"score_spread":0.25080068164624914,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3028673208","genre_codex":"review","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08855238,0.7270705,0.103587426,0.028586881,0.0024551349,0.03243533,0.008852636,0.0005802753,0.007879504],"genre_scores_gemma":[0.5922279,0.1767203,0.15748104,0.01850974,0.0013824505,0.04792581,0.004516466,0.00046031718,0.00077600824],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.29775697,0.3901021,0.23480177,0.015808677,0.059284437,0.0022460471],"domain_scores_gemma":[0.062069513,0.7205859,0.1392382,0.028951775,0.048323087,0.0008316691],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.5665955,0.002186062,0.0077781333,0.027664343,0.0027299593,0.009694114,0.006004489,0.004783846,0.002932486],"category_scores_gemma":[0.87935793,0.0040526036,0.017567532,0.030993829,0.0068377065,0.010551514,0.007407129,0.0033943516,0.00091027026],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0028986894,0.00018441434,0.1551157,0.4735428,0.09982017,0.00061536115,0.008404603,0.0020730149,0.0014509871,0.008397388,0.016003096,0.23149385],"study_design_scores_gemma":[0.0027393373,0.0017219969,0.11547948,0.6023208,0.18903263,0.0023497231,0.0054984814,0.005519544,0.0054376256,0.024137875,0.045013946,0.00074859825],"about_ca_topic_score_codex":0.005443471,"about_ca_topic_score_gemma":0.00774561,"teacher_disagreement_score":0.9922219,"about_ca_system_score_codex":0.010362747,"about_ca_system_score_gemma":0.020422263,"threshold_uncertainty_score":0.53446496},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3030582648","doi":"10.1200/jco.2020.38.15_suppl.e19206","title":"Associations between safety and tolerability and reporting of health-related quality of life in phase III randomized trials in common solid tumors.","year":2020,"lang":"en","type":"article","venue":"Journal of Clinical Oncology","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Princess Margaret Cancer Centre; University of Toronto","funders":"","keywords":"Medicine; Tolerability; Internal medicine; Discontinuation; Adverse effect; Quality of life (healthcare); Clinical trial; Randomized controlled trial; Breast cancer; Oncology; Cancer","score_opus":0.8178381080444754,"score_gpt":0.6405963177570368,"score_spread":0.17724179028743858,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3030582648","genre_codex":"review","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.35344106,0.5512799,0.03353569,0.009646769,0.0011420322,0.015448268,0.024770064,0.00029925047,0.010436953],"genre_scores_gemma":[0.94943523,0.015482173,0.017230704,0.0029734743,0.00041447143,0.0096793305,0.0042264215,0.00011831427,0.00043986336],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.39159626,0.4223716,0.14272992,0.015187871,0.025960347,0.0021540949],"domain_scores_gemma":[0.10249631,0.66802746,0.20957202,0.01226656,0.0065744887,0.0010631619],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.41362906,0.0009298428,0.0031001803,0.005461243,0.0007039698,0.004669819,0.0020791187,0.0025699793,0.005917584],"category_scores_gemma":[0.5948677,0.00091750926,0.012160383,0.013991617,0.0028231295,0.0057917563,0.002593323,0.003113824,0.00048940396],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.042863723,0.00032000814,0.68263173,0.09204322,0.10608695,0.00062151975,0.0013552951,0.0017553078,0.0006436939,0.0039274176,0.0066943085,0.061056796],"study_design_scores_gemma":[0.022884484,0.009425066,0.6979864,0.06709485,0.12892205,0.004611736,0.0017119644,0.008833561,0.0025342728,0.017699974,0.037851434,0.00044417745],"about_ca_topic_score_codex":0.00046427586,"about_ca_topic_score_gemma":0.00097173394,"teacher_disagreement_score":0.58637094,"about_ca_system_score_codex":0.0019827103,"about_ca_system_score_gemma":0.0031930716,"threshold_uncertainty_score":0.72309977},"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","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3032356396","doi":"10.1136/bmjebm-2019-111326","title":"Evaluating the quality of systematic reviews and meta-analyses published on direct oral anticoagulants in the past 5 years","year":2020,"lang":"en","type":"article","venue":"BMJ evidence-based medicine","topic":"Venous Thromboembolism Diagnosis and Management","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; Public Health Ontario; University of Toronto; McMaster University","funders":"","keywords":"Systematic review; Checklist; Medicine; Data extraction; MEDLINE; Guideline; Meta-analysis; Quality (philosophy); Quality Score; Medical physics; Family medicine; Internal medicine; Pathology; Psychology","score_opus":0.7738296929297955,"score_gpt":0.5664963431092327,"score_spread":0.20733334982056284,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3032356396","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008509415,0.9672453,0.0075882305,0.003426699,0.0014208481,0.0047288444,0.0060615125,0.00013516947,0.00088399404],"genre_scores_gemma":[0.24458165,0.6457055,0.053174436,0.0057323882,0.0020777169,0.039430175,0.008373919,0.00024365366,0.0006806983],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.5075968,0.16642147,0.2642822,0.016138038,0.043651488,0.0019100404],"domain_scores_gemma":[0.24008578,0.53163373,0.1621177,0.014466413,0.049425747,0.0022706906],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.39873445,0.0036945553,0.025044065,0.049204286,0.0019883958,0.010292356,0.0056075924,0.004806061,0.003785017],"category_scores_gemma":[0.6656754,0.0039019294,0.03645596,0.03774399,0.0041229012,0.008385087,0.006643271,0.0033518362,0.00045980923],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006587032,0.000027362807,0.011100797,0.79267585,0.16717428,0.00022011044,0.0011515872,0.00060010253,0.00023069071,0.0006547856,0.0020384153,0.023467287],"study_design_scores_gemma":[0.0011740915,0.00061341276,0.024171915,0.4322419,0.51690197,0.0006629032,0.00093973556,0.0008564483,0.0006659415,0.002731532,0.018806554,0.0002336968],"about_ca_topic_score_codex":0.009241579,"about_ca_topic_score_gemma":0.022583868,"teacher_disagreement_score":0.9749559,"about_ca_system_score_codex":0.010880295,"about_ca_system_score_gemma":0.029267384,"threshold_uncertainty_score":0.7414675},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W3039344762","doi":"10.1136/annrheumdis-2020-217237","title":"Population characteristics as important contextual factors in rheumatological trials: an exploratory meta-epidemiological study from an OMERACT Working Group","year":2020,"lang":"en","type":"review","venue":"Annals of the Rheumatic Diseases","topic":"Rheumatoid Arthritis Research and Therapies","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital; Bruyère; University of Ottawa","funders":"Odense Universitetshospital; Oak Foundation","keywords":"Medicine; Epidemiology; Population; Meta-analysis; Physical therapy; Family medicine; Internal medicine; Gerontology; Environmental health","score_opus":0.4342632187980843,"score_gpt":0.4729137158865182,"score_spread":0.03865049708843388,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3039344762","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015446508,0.9799952,0.0022644703,0.0004282292,0.00031345995,0.0004097297,0.00068738207,0.000031186333,0.00042378783],"genre_scores_gemma":[0.56729865,0.4133268,0.010889818,0.0031917484,0.0006369197,0.0021072584,0.0018186191,0.000146261,0.00058379915],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.9533495,0.031068347,0.0090371035,0.003258367,0.0023589064,0.00092786923],"domain_scores_gemma":[0.9065724,0.080084175,0.007627961,0.0031513092,0.0019209429,0.0006432043],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.04006322,0.0023435794,0.010462247,0.0029017613,0.00067562325,0.004574848,0.0018600064,0.003048858,0.0051392354],"category_scores_gemma":[0.076674215,0.0015711592,0.030879822,0.0040065944,0.0008339774,0.0022010521,0.0028147749,0.003048077,0.00040726046],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00718409,0.00005654878,0.019917417,0.11978612,0.83092487,0.00014817045,0.00020176485,0.00037141214,0.00023409573,0.00039171462,0.0008419498,0.019941917],"study_design_scores_gemma":[0.0021970775,0.00026912033,0.011816732,0.010417211,0.9722367,0.00008027742,0.00009299653,0.00024039757,0.0000857792,0.00049950846,0.0020314648,0.000032736858],"about_ca_topic_score_codex":0.0034847818,"about_ca_topic_score_gemma":0.009662229,"teacher_disagreement_score":0.9895378,"about_ca_system_score_codex":0.001685501,"about_ca_system_score_gemma":0.0032715346,"threshold_uncertainty_score":0.21187705},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3048211970","doi":"10.1016/j.jclinepi.2020.08.002","title":"Abbreviated and comprehensive literature searches led to identical or very similar effect estimates: a meta-epidemiological study","year":2020,"lang":"en","type":"review","venue":"Journal of Clinical Epidemiology","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":20,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institute of Health Economics; University of Alberta","funders":"","keywords":"Interquartile range; Meta-analysis; Confidence interval; Medicine; Odds ratio; MEDLINE; Epidemiology; Statistics; Systematic review; Internal medicine; Mathematics; Biology","score_opus":0.9657825327774562,"score_gpt":0.7373159784020075,"score_spread":0.22846655437544872,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3048211970","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.043479796,0.8629765,0.062662676,0.0044389013,0.0030865786,0.010589498,0.007563755,0.00092008844,0.004282218],"genre_scores_gemma":[0.543724,0.19764066,0.20692855,0.012731254,0.0019059621,0.025361655,0.009284078,0.0012730781,0.0011507511],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.47367162,0.2919834,0.1668929,0.022594597,0.042952865,0.0019045412],"domain_scores_gemma":[0.21468516,0.6137548,0.07914864,0.057830498,0.0334484,0.0011325269],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.3605963,0.0056548147,0.016460778,0.027629806,0.0015299392,0.0143084675,0.0070448606,0.0073681935,0.007825898],"category_scores_gemma":[0.7473789,0.0053653517,0.04366566,0.03008031,0.0028646963,0.01669905,0.008424469,0.0060118134,0.0019104932],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008354227,0.00016893294,0.025898853,0.32372156,0.5497878,0.0011694899,0.0023217255,0.002474045,0.0024838585,0.0030026268,0.0056602075,0.07495671],"study_design_scores_gemma":[0.00516384,0.0014849678,0.034339786,0.130972,0.79451233,0.002017098,0.0008087293,0.002145492,0.0022810628,0.00846447,0.017228303,0.0005818951],"about_ca_topic_score_codex":0.0024479772,"about_ca_topic_score_gemma":0.0046967426,"teacher_disagreement_score":0.9835392,"about_ca_system_score_codex":0.005216361,"about_ca_system_score_gemma":0.007859205,"threshold_uncertainty_score":0.78849864},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3083911301","doi":"10.1093/cid/ciaa1353","title":"Methodological and Reporting Quality of Noninferiority Randomized Controlled Trials Comparing Antibiotic Therapies: A Systematic Review","year":2020,"lang":"en","type":"review","venue":"Clinical Infectious Diseases","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; McMaster University","funders":"","keywords":"Medicine; Blinding; Randomized controlled trial; Systematic review; MEDLINE; Confidence interval; Protocol (science); Meta-analysis; Odds ratio; Intensive care medicine; Food and drug administration; Internal medicine; Alternative medicine; Medical emergency","score_opus":0.9643438506930195,"score_gpt":0.7238925584481332,"score_spread":0.24045129224488637,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3083911301","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0042581176,0.96762913,0.008858588,0.004256465,0.0016213137,0.01006518,0.0021756876,0.000108244116,0.0010273646],"genre_scores_gemma":[0.20362651,0.6406773,0.06396513,0.009362478,0.0018426285,0.07598297,0.0036772264,0.00022952766,0.0006362655],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.32242477,0.3097359,0.29600483,0.01628785,0.05324772,0.0022988594],"domain_scores_gemma":[0.17035662,0.58124816,0.17585556,0.017941535,0.052627712,0.0019704488],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.37211257,0.0034286003,0.02023328,0.025605144,0.0027543826,0.011210698,0.0067070397,0.007241153,0.004091474],"category_scores_gemma":[0.7181575,0.004651243,0.020012263,0.022493562,0.0080146585,0.008766723,0.0054879864,0.004143701,0.0006207938],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004979599,0.000024813831,0.0043391557,0.94305253,0.025768142,0.00010267172,0.0007822847,0.00028091716,0.00017679349,0.0009261939,0.0023989608,0.021649476],"study_design_scores_gemma":[0.0011767953,0.0002142466,0.005489674,0.8995221,0.07755157,0.0003377426,0.0005705712,0.00060934364,0.0005528995,0.0030831995,0.010744457,0.0001473688],"about_ca_topic_score_codex":0.008582442,"about_ca_topic_score_gemma":0.01210005,"teacher_disagreement_score":0.6278874,"about_ca_system_score_codex":0.015832577,"about_ca_system_score_gemma":0.0410305,"threshold_uncertainty_score":0.774297},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3087727097","doi":"10.1186/s43058-020-00068-8","title":"Advancing the pragmatic measurement of sustainment: a narrative review of measures","year":2020,"lang":"en","type":"review","venue":"Implementation Science Communications","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":86,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Child, Adolescent and Family Mental Health","funders":"National Institute on Drug Abuse; National Institute of Mental Health; U.S. Department of Health and Human Services","keywords":"Narrative; History; Political science; Linguistics; Philosophy","score_opus":0.6523923922329052,"score_gpt":0.7300703640853611,"score_spread":0.07767797185245584,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3087727097","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0023925814,0.9631531,0.008755773,0.015065247,0.0014560627,0.0030926492,0.00086899963,0.000049792663,0.0051656696],"genre_scores_gemma":[0.052290495,0.8989874,0.027866498,0.0067305453,0.00063482655,0.012266729,0.0006558919,0.00007109825,0.00049648475],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.8127841,0.11404124,0.046272606,0.0045846547,0.020983474,0.0013338957],"domain_scores_gemma":[0.6526208,0.275919,0.03148104,0.0070391544,0.03169711,0.0012429261],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.13028583,0.001876516,0.0042125443,0.017414697,0.0019875169,0.008104832,0.0036394363,0.0029374044,0.003459844],"category_scores_gemma":[0.40748775,0.0015651613,0.0049906853,0.018933937,0.004859036,0.014195948,0.0064493697,0.0046528657,0.0005949915],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00028049437,0.00005734179,0.0012954988,0.633999,0.001634813,0.00021367629,0.010941108,0.00033372533,0.00031386252,0.036654063,0.016973019,0.29730338],"study_design_scores_gemma":[0.00009008662,0.00016551444,0.0017795753,0.84374636,0.002440171,0.0002856426,0.0045647547,0.00024421475,0.000488134,0.00767679,0.13843746,0.00008130091],"about_ca_topic_score_codex":0.0055448115,"about_ca_topic_score_gemma":0.008801915,"teacher_disagreement_score":0.86971414,"about_ca_system_score_codex":0.016119542,"about_ca_system_score_gemma":0.047490776,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"opus","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3089040936","doi":"10.1111/jep.13482","title":"Determinants that influence knowledge brokers' and opinion leaders' role to close knowledge practice gaps in rehabilitation: A realist review","year":2020,"lang":"en","type":"review","venue":"Journal of Evaluation in Clinical Practice","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":22,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Centre for Interdisciplinary Research in Rehabilitation","funders":"","keywords":"Context (archaeology); Knowledge translation; Mechanism (biology); Knowledge management; Psychological intervention; Outcome (game theory); Rehabilitation; Psychology; Computer science; Medicine; Nursing; Economics; Epistemology","score_opus":0.688466216630974,"score_gpt":0.7610447456251737,"score_spread":0.0725785289941997,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3089040936","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0050695045,0.9875238,0.0010401517,0.0030684227,0.0002955244,0.00085620326,0.00045314946,0.000012686241,0.001680472],"genre_scores_gemma":[0.123449594,0.863818,0.004496999,0.0035586776,0.000320139,0.0033744848,0.0005651924,0.000026019956,0.0003909041],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.94243413,0.030433431,0.016364012,0.003671086,0.0057748617,0.0013225258],"domain_scores_gemma":[0.6001621,0.34433332,0.035040725,0.0035861603,0.015623088,0.0012544995],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.05765055,0.0011779543,0.0046252683,0.013836388,0.0012916267,0.008515445,0.0028739704,0.0033090964,0.0053373165],"category_scores_gemma":[0.25436184,0.0013417694,0.005720537,0.0134090455,0.0029558728,0.0072426763,0.0033931208,0.0020833297,0.0006099647],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004478203,0.000084664374,0.009711152,0.80500764,0.00930184,0.00039048368,0.005397858,0.00022611082,0.00015665047,0.005427969,0.0037011,0.1601467],"study_design_scores_gemma":[0.00016431007,0.00019024416,0.012031487,0.91871065,0.02648577,0.00041047344,0.0046311542,0.0002010504,0.00027707464,0.003421048,0.033399213,0.00007753231],"about_ca_topic_score_codex":0.008299421,"about_ca_topic_score_gemma":0.02027234,"teacher_disagreement_score":0.05765055,"about_ca_system_score_codex":0.0076645827,"about_ca_system_score_gemma":0.026620759,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3093133284","doi":"10.5539/ijef.v12n11p1","title":"Governance and Performance in Insurance Companies: A Bibliometric Analysis and A Meta-Analysis","year":2020,"lang":"en","type":"article","venue":"International Journal of Economics and Finance","topic":"Insurance and Financial Risk Management","field":"Economics, Econometrics and Finance","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Corporate governance; Accounting; Meta-analysis; Point (geometry); Quantitative analysis (chemistry); Business; Insurance industry; Empirical research; Actuarial science; Finance; Statistics","score_opus":0.04692001812354021,"score_gpt":0.24029358453952596,"score_spread":0.19337356641598574,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3093133284","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0770715,0.9050697,0.0034212505,0.0017486055,0.00022951768,0.0005895885,0.00955676,0.00010515484,0.002207843],"genre_scores_gemma":[0.56813693,0.41403013,0.0069872327,0.0005823021,0.00043699823,0.001314049,0.008022921,0.00007878927,0.00041069556],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9763352,0.007912848,0.008368552,0.002180275,0.0047169994,0.0004860659],"domain_scores_gemma":[0.8840337,0.088528395,0.015423907,0.0047753314,0.006158191,0.0010804292],"candidate_categories":["metaresearch","bibliometrics"],"consensus_categories":[],"category_scores_codex":[0.022500964,0.0014295084,0.0066478075,0.11985775,0.0009860881,0.0054084565,0.0016090309,0.0014236539,0.0031246212],"category_scores_gemma":[0.076736234,0.000851149,0.0153303705,0.097683325,0.0011531607,0.0036645685,0.0029456194,0.0010596819,0.00032645973],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010243,0.00017220153,0.16666114,0.33769402,0.33938622,0.00053034275,0.0024220648,0.0017214827,0.00075375574,0.0025931736,0.004326169,0.14271511],"study_design_scores_gemma":[0.00036391016,0.00047081773,0.2159253,0.1106748,0.6362621,0.0012671435,0.0028086572,0.0019021687,0.000921087,0.004164054,0.02500558,0.00023437808],"about_ca_topic_score_codex":0.005690859,"about_ca_topic_score_gemma":0.008487498,"teacher_disagreement_score":0.977499,"about_ca_system_score_codex":0.0031831812,"about_ca_system_score_gemma":0.0061351317,"threshold_uncertainty_score":0.11899787},"labels":[{"model":"gemma","categories":["bibliometrics","metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["bibliometrics"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W3097681315","doi":"10.1186/s40814-020-00712-z","title":"More pilot trials could plan to use qualitative data: a meta-epidemiological study","year":2020,"lang":"en","type":"article","venue":"Pilot and Feasibility Studies","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":29,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Joseph’s Healthcare Hamilton; Population Health Research Institute; Laurentian University; McMaster University; Hamilton Health Sciences; Toronto General Hospital; University Health Network; Impact; Health Sciences Centre","funders":"","keywords":"Qualitative research; Medicine; Qualitative property; Psychological intervention; Clinical trial; Research design; Randomized controlled trial; Confidence interval; Data collection; Family medicine; Odds ratio; Clinical study design; Physical therapy; Nursing; Internal medicine","score_opus":0.9954596694666988,"score_gpt":0.7332717440502268,"score_spread":0.26218792541647207,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3097681315","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.027421102,0.86132836,0.049838457,0.010091655,0.0042424714,0.03900938,0.0043420475,0.0004444628,0.0032820348],"genre_scores_gemma":[0.56270087,0.18035807,0.10542065,0.018492961,0.0017088035,0.1263173,0.003622669,0.0002940302,0.0010846977],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.6467161,0.25285885,0.07077796,0.009215303,0.018240342,0.002191407],"domain_scores_gemma":[0.3188548,0.5689544,0.059337225,0.029640079,0.021722542,0.0014910108],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.31292766,0.00368507,0.014013268,0.008890427,0.0017071854,0.008342301,0.0042901062,0.006053237,0.008012274],"category_scores_gemma":[0.535077,0.003088128,0.03934707,0.010352506,0.0028374728,0.012742452,0.004126505,0.005103778,0.0007787614],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0048948256,0.00031876066,0.023121305,0.59423816,0.32205033,0.00019296914,0.00150743,0.00097523455,0.00023389273,0.0023451154,0.0038595432,0.04626245],"study_design_scores_gemma":[0.0079381,0.0021158434,0.014985596,0.35415286,0.59385514,0.00033603277,0.0012453506,0.0020378164,0.00071287784,0.007874913,0.014508167,0.00023719882],"about_ca_topic_score_codex":0.0042003505,"about_ca_topic_score_gemma":0.0063376212,"teacher_disagreement_score":0.99631494,"about_ca_system_score_codex":0.006589778,"about_ca_system_score_gemma":0.013296221,"threshold_uncertainty_score":0.8472825},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_narrow","metaepi_broad","metaresearch"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3108734079","doi":"10.21203/rs.3.rs-107409/v1","title":"Patient-important Outcomes Reported in Clinical Studies of Pharmacologic Treatments for COVID-19: A Protocol of a Meta-epidemiological Study","year":2020,"lang":"en","type":"preprint","venue":"Research Square","topic":"Pharmaceutical Practices and Patient Outcomes","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; McMaster University; Impact","funders":"","keywords":"Coronavirus disease 2019 (COVID-19); Epidemiology; Meta-analysis; Protocol (science); Medicine; 2019-20 coronavirus outbreak; Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); Intensive care medicine; Internal medicine; Virology; Pathology; Alternative medicine; Disease; Outbreak","score_opus":0.8994438648495541,"score_gpt":0.7345655858004917,"score_spread":0.16487827904906238,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3108734079","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0018963434,0.005001643,0.014388936,0.0009599921,0.0008924479,0.9689728,0.0065376563,0.00028491588,0.001065346],"genre_scores_gemma":[0.0021734461,0.0007522344,0.01238807,0.00026224504,0.000039249688,0.9834379,0.0007393301,0.000020516283,0.00018712161],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.85176176,0.08950051,0.042217124,0.0067276093,0.007789979,0.002003057],"domain_scores_gemma":[0.8609375,0.07658927,0.01739346,0.022908572,0.020058824,0.0021124491],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.17031556,0.0060825706,0.016364096,0.010686174,0.0023263276,0.006650799,0.0046265833,0.008353199,0.024130747],"category_scores_gemma":[0.2783918,0.004131662,0.04855377,0.010182525,0.0029807962,0.0043383823,0.0057509826,0.008734498,0.0035324777],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.06475381,0.0013016049,0.0072793327,0.67076933,0.13949993,0.0015131624,0.0029296444,0.008974832,0.0033077009,0.014652009,0.033551317,0.051467296],"study_design_scores_gemma":[0.16163646,0.008938741,0.022294104,0.23086053,0.26395115,0.0010217916,0.001834518,0.014957735,0.0063066217,0.053081974,0.23398075,0.0011357268],"about_ca_topic_score_codex":0.0024698444,"about_ca_topic_score_gemma":0.0038957812,"teacher_disagreement_score":0.9939174,"about_ca_system_score_codex":0.007147366,"about_ca_system_score_gemma":0.02120445,"threshold_uncertainty_score":0.90072554},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3110084249","doi":"10.1016/j.joca.2020.10.004","title":"Visual Analogue Scale has higher assay sensitivity than WOMAC pain in detecting between-group differences in treatment effects: a meta-epidemiological study","year":2020,"lang":"en","type":"review","venue":"Osteoarthritis and Cartilage","topic":"Osteoarthritis Treatment and Mechanisms","field":"Medicine","cited_by":85,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Impact; Dalhousie University; McMaster University; University of Toronto; St. Michael's Hospital","funders":"Arthritis Society","keywords":"WOMAC; Visual analogue scale; Medicine; Osteoarthritis; Meta-analysis; Randomized controlled trial; Placebo; Physical therapy; Internal medicine; Pathology","score_opus":0.08019099511676102,"score_gpt":0.32141410434270895,"score_spread":0.24122310922594792,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3110084249","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0074367435,0.9896953,0.0014494952,0.00016985081,0.00033948253,0.00014501906,0.00029166645,0.000029318537,0.0004431694],"genre_scores_gemma":[0.3199803,0.6657963,0.007306074,0.0031781115,0.0006587213,0.0005850048,0.0012348839,0.000118342985,0.0011423299],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.98431444,0.00856783,0.0036729784,0.0019213397,0.0012189605,0.00030449408],"domain_scores_gemma":[0.9582239,0.035868116,0.002645081,0.0020027836,0.0010796571,0.00018049369],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.023761472,0.0031932238,0.016425619,0.002494819,0.0005386229,0.0036822595,0.0022876433,0.003387957,0.00459011],"category_scores_gemma":[0.044662964,0.0017270509,0.03252099,0.003458475,0.00090428,0.0022821396,0.0013428931,0.0031785606,0.0005027911],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0063140583,0.0000739807,0.007289291,0.110703506,0.8480944,0.00013819318,0.00008809639,0.00028780804,0.0004298123,0.0001800072,0.0005186361,0.02588217],"study_design_scores_gemma":[0.0017484349,0.0004400168,0.009758625,0.005105528,0.98002267,0.00011187939,0.0000471119,0.00014808132,0.00012735979,0.00033466748,0.002124076,0.00003161728],"about_ca_topic_score_codex":0.0032859286,"about_ca_topic_score_gemma":0.008721779,"teacher_disagreement_score":0.9968068,"about_ca_system_score_codex":0.0011060906,"about_ca_system_score_gemma":0.0013104753,"threshold_uncertainty_score":0.12566417},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_narrow","metaepi_broad","metaresearch"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3110163884","doi":"10.1101/2020.11.29.20237875","title":"Clinical Trials in COVID-19 Management &amp; Prevention: A Meta-epidemiological Study examining methodological quality","year":2020,"lang":"en","type":"preprint","venue":"medRxiv","topic":"COVID-19 Clinical Research Studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia; Centre Hospitalier Universitaire de Sherbrooke; University of Ottawa; Western University; St. Michael's Hospital; Ted Rogers Centre for Heart Research; Impact; Toronto General Hospital; University Health Network; University of Toronto; McMaster University","funders":"Canadian Institutes of Health Research; Hamilton Health Sciences","keywords":"Medicine; Randomized controlled trial; Sample size determination; Blinding; Randomization; Clinical trial; Epidemiology; Consolidated Standards of Reporting Trials; Placebo; Internal medicine; Family medicine; Alternative medicine; Pathology","score_opus":0.9534335328935374,"score_gpt":0.7343115340039995,"score_spread":0.21912199888953787,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3110163884","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.013285297,0.9685357,0.009634572,0.0015993485,0.00078572787,0.0036717176,0.0014878678,0.000102770035,0.0008969983],"genre_scores_gemma":[0.653664,0.2748857,0.033477247,0.00539082,0.0012509061,0.027960096,0.0022771517,0.00023658371,0.00085749564],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.7231607,0.17508057,0.07239409,0.012582621,0.014943549,0.0018385218],"domain_scores_gemma":[0.45653775,0.44559816,0.070053674,0.013896405,0.012329716,0.0015843622],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.17461981,0.0031392558,0.020450331,0.010685286,0.0013484771,0.008047948,0.004097037,0.0048944666,0.0052466528],"category_scores_gemma":[0.40151054,0.002194403,0.046267718,0.013576058,0.0030087982,0.0050801286,0.0036602793,0.0035868336,0.000387606],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0037968282,0.000048177775,0.016509665,0.32720548,0.63726485,0.00013803526,0.0002669361,0.0007551679,0.00013825002,0.0008317664,0.0011294652,0.011915416],"study_design_scores_gemma":[0.0032798979,0.00054909487,0.012062222,0.08499266,0.8908574,0.00022916656,0.00019042926,0.0011318226,0.00027292277,0.0027798354,0.0035760936,0.00007840201],"about_ca_topic_score_codex":0.0040474506,"about_ca_topic_score_gemma":0.0048581897,"teacher_disagreement_score":0.99686074,"about_ca_system_score_codex":0.004984653,"about_ca_system_score_gemma":0.006235754,"threshold_uncertainty_score":0.92348886},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3116112122","doi":"10.1111/cdoe.12616","title":"Socioeconomic status and oral health‐related quality of life: A systematic review and meta‐analysis","year":2020,"lang":"en","type":"review","venue":"Community Dentistry And Oral Epidemiology","topic":"Dental Health and Care Utilization","field":"Dentistry","cited_by":162,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"Fundação de Amparo à Pesquisa do Estado do Rio Grande do Sul; Conselho Nacional de Desenvolvimento Científico e Tecnológico","keywords":"Socioeconomic status; Medicine; Observational study; Meta-analysis; Confidence interval; Scopus; Quality of life (healthcare); Critical appraisal; Demography; Study heterogeneity; Systematic review; MEDLINE; Gerontology; Environmental health; Population; Internal medicine; Pathology; Alternative medicine","score_opus":0.24911123320913814,"score_gpt":0.4792272198465582,"score_spread":0.23011598663742006,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3116112122","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0040587415,0.9939476,0.00046470453,0.00020241055,0.00012390585,0.00056936924,0.000458706,0.000019826655,0.00015462255],"genre_scores_gemma":[0.08491257,0.9071609,0.003175651,0.0007267609,0.0002808124,0.0026972727,0.00077239843,0.00002261002,0.0002510586],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9890864,0.004618544,0.0037685665,0.00085033977,0.0013197504,0.0003564276],"domain_scores_gemma":[0.9796662,0.013179441,0.004617341,0.00053617085,0.0017259314,0.00027495544],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.015317059,0.0031302972,0.020096578,0.011650622,0.0008049672,0.0035563987,0.0024136638,0.0026519934,0.00397477],"category_scores_gemma":[0.03886952,0.0018950746,0.029071558,0.010706715,0.00078112335,0.002349546,0.001960701,0.0014505321,0.0003225447],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00085303286,0.000035942357,0.0035049485,0.69617975,0.288038,0.0001536491,0.000097161974,0.00025812304,0.00019583675,0.0001009172,0.0004395247,0.010143163],"study_design_scores_gemma":[0.0005802621,0.0003469296,0.0061803726,0.112500794,0.8774579,0.00020904401,0.00009595472,0.0001846324,0.00013609826,0.00018674039,0.0020859912,0.000035380315],"about_ca_topic_score_codex":0.0053701173,"about_ca_topic_score_gemma":0.013431483,"teacher_disagreement_score":0.020096578,"about_ca_system_score_codex":0.0030456854,"about_ca_system_score_gemma":0.0064756456,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3134334065","doi":"10.1016/j.jebdp.2021.101544","title":"Influence of Sponsorship Bias on Treatment Effect Size Estimates in Randomized Trials of Oral Health Interventions: A Meta-epidemiological Study","year":2021,"lang":"en","type":"review","venue":"Journal of Evidence Based Dental Practice","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":4,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; St. Michael's Hospital; University of Alberta","funders":"Alberta Heritage Foundation for Medical Research; Alberta Innovates; Alberta Innovates - Health Solutions","keywords":"Meta-analysis; Publication bias; Randomized controlled trial; Medicine; Epidemiology; Clinical trial; Confidence interval; Sample size determination; Psychological intervention; Random effects model; Selection bias; Reporting bias; Relative risk; MEDLINE; Internal medicine; Statistics; Psychiatry; Pathology","score_opus":0.9725301236095575,"score_gpt":0.7303173162000859,"score_spread":0.2422128074094716,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3134334065","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.018441008,0.9523998,0.02205837,0.0019494775,0.0018412267,0.0012035711,0.0005962003,0.0001599246,0.0013504719],"genre_scores_gemma":[0.7095451,0.23600823,0.03786852,0.0066233943,0.0025134017,0.004555601,0.0011913261,0.00039415047,0.0013003432],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.5959351,0.31305262,0.06023706,0.0142950835,0.014588291,0.0018917908],"domain_scores_gemma":[0.28068197,0.6653673,0.026803331,0.01949228,0.0068580867,0.0007969587],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.2944574,0.003335493,0.021132069,0.0049315253,0.0009900923,0.009700516,0.0045128334,0.00811897,0.0041174423],"category_scores_gemma":[0.54864204,0.003377821,0.043465134,0.005404359,0.003901636,0.0071540354,0.0039670626,0.0086869635,0.00041914417],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.015461362,0.00013033552,0.0141380355,0.09921195,0.8132609,0.00030642218,0.00052466337,0.001629515,0.00029971867,0.0029939841,0.0016056015,0.050437555],"study_design_scores_gemma":[0.008080714,0.0009955582,0.008475207,0.015552772,0.95210385,0.00037484913,0.00016176194,0.0023214105,0.00048484982,0.007640546,0.0036939625,0.00011449939],"about_ca_topic_score_codex":0.0028011696,"about_ca_topic_score_gemma":0.0030449703,"teacher_disagreement_score":0.9966645,"about_ca_system_score_codex":0.003103421,"about_ca_system_score_gemma":0.0045563644,"threshold_uncertainty_score":0.8700596},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3142062327","doi":"10.9778/cmajo.20200140","title":"Completeness of reporting for COVID-19 case reports, January to April 2020: a meta-epidemiologic study","year":2021,"lang":"en","type":"article","venue":"CMAJ Open","topic":"COVID-19 and healthcare impacts","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"Queen's University; University of Toronto; St. Michael's Hospital","funders":"Amgen","keywords":"Coronavirus disease 2019 (COVID-19); Pandemic; Completeness (order theory); 2019-20 coronavirus outbreak; Medicine; Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); Coronavirus; Meta-analysis; MEDLINE; Disease; Infectious disease (medical specialty); Outbreak; Virology; Pathology; Political science; Mathematics","score_opus":0.5234770417953649,"score_gpt":0.546135499860114,"score_spread":0.022658458064749087,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3142062327","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.07483619,0.9111113,0.0020619712,0.0011960812,0.0004390655,0.0006536451,0.008478526,0.00006979644,0.001153415],"genre_scores_gemma":[0.79505056,0.19031806,0.0029177812,0.001861737,0.0006871073,0.0018247173,0.0069469176,0.000089672445,0.00030335394],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.87725186,0.04425632,0.050184764,0.011092336,0.01498371,0.0022309572],"domain_scores_gemma":[0.54305506,0.29291242,0.12870115,0.014976643,0.018650167,0.0017046534],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.09545379,0.0016204865,0.006434745,0.015246945,0.00085401314,0.0056640455,0.0034236328,0.0022554155,0.0025947986],"category_scores_gemma":[0.3088078,0.0016110007,0.024727225,0.020863876,0.0013769987,0.0053313915,0.0032507572,0.0020679317,0.00032493204],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0033349127,0.000088478846,0.39363784,0.2148505,0.35963964,0.00030044327,0.00079471833,0.00062649354,0.0001445473,0.000398103,0.002821413,0.023362847],"study_design_scores_gemma":[0.0009953789,0.0008323657,0.33071312,0.091569446,0.5639688,0.0011511573,0.0010595875,0.0009747642,0.00049509335,0.0010099306,0.0070508188,0.00017950976],"about_ca_topic_score_codex":0.007430887,"about_ca_topic_score_gemma":0.008184402,"teacher_disagreement_score":0.99356526,"about_ca_system_score_codex":0.0034161655,"about_ca_system_score_gemma":0.004234162,"threshold_uncertainty_score":0.5048139},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3154585211","doi":"10.1016/j.jid.2021.02.287","title":"265 A meta-epidemiologic study of power and sample size in randomized controlled trials published in dermatology journals","year":2021,"lang":"en","type":"article","venue":"Journal of Investigative Dermatology","topic":"COVID-19 and healthcare impacts","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Impact; Queen's University; Women's College Hospital; McMaster University; University of Toronto","funders":"","keywords":"Sample size determination; Medicine; Randomized controlled trial; Meta-analysis; Consolidated Standards of Reporting Trials; Sample (material); MEDLINE; Outcome (game theory); Statistics; Internal medicine; Mathematics","score_opus":0.2074584655249477,"score_gpt":0.44687235754155613,"score_spread":0.23941389201660843,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3154585211","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.18275765,0.7309811,0.062239528,0.0038404814,0.007307708,0.005602061,0.0021097653,0.0005257407,0.004636001],"genre_scores_gemma":[0.937376,0.032185297,0.019207124,0.003886299,0.0017455752,0.003543324,0.00089822087,0.00018806638,0.00097018306],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.59584343,0.31820732,0.054408792,0.018478379,0.011242805,0.001819251],"domain_scores_gemma":[0.35572967,0.57957745,0.026911616,0.029563786,0.00702381,0.0011936816],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.24553376,0.003513057,0.015223637,0.005924531,0.0014075622,0.0066857478,0.003414427,0.007221949,0.0045954906],"category_scores_gemma":[0.5028963,0.0041044625,0.051760904,0.0057038735,0.0038312515,0.0056853327,0.0032478075,0.006046287,0.0005447514],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.030760111,0.00015877529,0.022263937,0.02251687,0.91380346,0.00020841022,0.00022523002,0.00057832134,0.00029789738,0.00084395433,0.0005994923,0.0077435668],"study_design_scores_gemma":[0.02098178,0.0024584765,0.014805169,0.004096781,0.95040154,0.0002547059,0.00013984417,0.0012498152,0.00032774388,0.0035417052,0.0016493463,0.000093129835],"about_ca_topic_score_codex":0.0021201456,"about_ca_topic_score_gemma":0.0023773648,"teacher_disagreement_score":0.99648696,"about_ca_system_score_codex":0.0019787333,"about_ca_system_score_gemma":0.0020459897,"threshold_uncertainty_score":0.9303912},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W3158629300","doi":"10.1186/s12893-021-01278-0","title":"Bias estimation in study design: a meta-epidemiological analysis of transcatheter versus surgical aortic valve replacement","year":2021,"lang":"en","type":"article","venue":"BMC Surgery","topic":"Advanced Causal Inference Techniques","field":"Mathematics","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Michael's Hospital; Women's College Hospital; University of Toronto; Institute for Work & Health; Western University","funders":"University of Toronto; Canadian Institutes of Health Research; University of Ottawa","keywords":"Medicine; Aortic valve replacement; Propensity score matching; Stenosis; Randomized controlled trial; Aortic valve stenosis; Surgery; Aortic valve; Internal medicine; Epidemiology; Cardiology; Meta-analysis","score_opus":0.6141875518604735,"score_gpt":0.4766740758921051,"score_spread":0.13751347596836838,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3158629300","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.020234017,0.922825,0.048004396,0.0030548149,0.0019238403,0.001905623,0.0005388303,0.00017834353,0.0013352162],"genre_scores_gemma":[0.7426174,0.1665723,0.07295031,0.0061321026,0.0020130498,0.0074952347,0.0010253686,0.0002507127,0.00094350957],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.5743311,0.35355842,0.043111403,0.012283217,0.015303952,0.0014118548],"domain_scores_gemma":[0.42417705,0.5070992,0.038835097,0.019659044,0.009322677,0.0009068276],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.2694727,0.0037435547,0.014914433,0.010112368,0.001172255,0.008179554,0.0036659476,0.0060864203,0.0024657617],"category_scores_gemma":[0.42866236,0.002320188,0.053836115,0.007657735,0.0023637433,0.0047767907,0.004073156,0.004660575,0.00025251682],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00424488,0.000068678564,0.02567949,0.06419703,0.8628185,0.0002103521,0.0002885996,0.0030688387,0.00021016944,0.0019536633,0.0010554839,0.03620439],"study_design_scores_gemma":[0.0038834931,0.00071082,0.01294536,0.026789317,0.9344085,0.00039981358,0.00016648893,0.0057633747,0.000509357,0.009884291,0.0044103884,0.00012882064],"about_ca_topic_score_codex":0.0030532528,"about_ca_topic_score_gemma":0.0037618487,"teacher_disagreement_score":0.9962565,"about_ca_system_score_codex":0.0037538726,"about_ca_system_score_gemma":0.0039044877,"threshold_uncertainty_score":0.9008702},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3163788934","doi":"10.1186/s12874-021-01295-w","title":"Quality assessment tools used in systematic reviews of in vitro studies: A systematic review","year":2021,"lang":"en","type":"review","venue":"BMC Medical Research Methodology","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":126,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Al-Azhar University; Zagazig University; McMaster University","keywords":"Systematic review; Medical physics; Protocol (science); Guideline; Scopus; Quality assessment; Medicine; Computer science; MEDLINE; Pathology; External quality assessment; Alternative medicine; Biology","score_opus":0.9945041192763169,"score_gpt":0.8198052753117314,"score_spread":0.17469884396458546,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3163788934","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004433637,0.91720086,0.033840455,0.0055799256,0.0016015994,0.031113105,0.0044522784,0.0003811013,0.0013969778],"genre_scores_gemma":[0.13652094,0.47405216,0.20452604,0.0060405685,0.0010557381,0.17288554,0.004112321,0.0002819044,0.0005248342],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.34724346,0.33218178,0.25087005,0.012892027,0.054911442,0.0019012268],"domain_scores_gemma":[0.18967077,0.5961386,0.13215443,0.022814045,0.057496283,0.0017258797],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.41302642,0.0046065277,0.019734325,0.050911736,0.0037025532,0.011622723,0.006342082,0.0045710495,0.004971186],"category_scores_gemma":[0.68367755,0.0040730364,0.024514625,0.039071586,0.0057337983,0.011182875,0.0076477774,0.0038185187,0.0008231445],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00025189706,0.000030421448,0.0035316618,0.9291484,0.020859035,0.00011569315,0.0014131505,0.0004552136,0.0003314884,0.0012761584,0.0024081764,0.04017867],"study_design_scores_gemma":[0.0007505449,0.0002932533,0.0060378183,0.9068561,0.05937254,0.0003360983,0.0012296943,0.0009866821,0.0008052071,0.003983615,0.019103287,0.00024503682],"about_ca_topic_score_codex":0.0071916766,"about_ca_topic_score_gemma":0.015380093,"teacher_disagreement_score":0.58697355,"about_ca_system_score_codex":0.016842198,"about_ca_system_score_gemma":0.047394745,"threshold_uncertainty_score":0.723843},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W3172845662","doi":"10.1186/s12916-021-02008-2","title":"Utilization of the evidence from studies with no events in meta-analyses of adverse events: an empirical investigation","year":2021,"lang":"en","type":"article","venue":"BMC Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":53,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"Medical Research Council; National Health and Medical Research Council; Qatar National Library","keywords":"Systematic review; Medicine; Meta-analysis; Adverse effect; MEDLINE; Internal medicine","score_opus":0.951712016036363,"score_gpt":0.6398872374389903,"score_spread":0.3118247785973727,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3172845662","genre_codex":"review","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.020038746,0.9148308,0.05484042,0.0031203593,0.0009618667,0.0025942728,0.0013619807,0.00017679489,0.0020748423],"genre_scores_gemma":[0.74580103,0.17615798,0.064671606,0.0038579078,0.0012089512,0.0061395434,0.0014259471,0.00025253106,0.00048455852],"study_design_codex":"meta_analysis","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.3318441,0.5223263,0.09971047,0.020296765,0.023741081,0.0020812943],"domain_scores_gemma":[0.07766046,0.8433947,0.046914242,0.022501366,0.008867833,0.00066132506],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.4224945,0.0050057387,0.015897222,0.017202245,0.0017030436,0.010475604,0.0062694238,0.005847166,0.005429806],"category_scores_gemma":[0.779615,0.004657128,0.06649881,0.02302469,0.0054687327,0.01686984,0.0076879277,0.006555471,0.0006634507],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004190629,0.00013463858,0.07660557,0.20175457,0.6564378,0.0006975369,0.0015858109,0.0023449543,0.00024780928,0.003627419,0.0012894708,0.051083896],"study_design_scores_gemma":[0.0021352272,0.0017098491,0.030291785,0.120970525,0.8061093,0.0017373684,0.0012572025,0.004849798,0.00076447666,0.020024262,0.009873238,0.00027699396],"about_ca_topic_score_codex":0.0034392592,"about_ca_topic_score_gemma":0.0050342954,"teacher_disagreement_score":0.57750547,"about_ca_system_score_codex":0.005237917,"about_ca_system_score_gemma":0.0056870584,"threshold_uncertainty_score":0.71216714},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W3178874129","doi":"10.1186/s12916-021-02020-6","title":"Validity of observational evidence on putative risk and protective factors: appraisal of 3744 meta-analyses on 57 topics","year":2021,"lang":"en","type":"article","venue":"BMC Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":39,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Canada Research Chairs; University of Toronto; University of New Brunswick","funders":"Medical Research Council; Cancer Research UK; Laura and John Arnold Foundation","keywords":"Observational study; Medicine; Statistical significance; Sample size determination; Meta-analysis; Randomized controlled trial; Critical appraisal; Interquartile range; Null hypothesis; Statistics; Internal medicine; Pathology; Alternative medicine","score_opus":0.9664124400435993,"score_gpt":0.644063745652456,"score_spread":0.3223486943911432,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3178874129","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00783259,0.9798099,0.0033503578,0.0009350435,0.00030004457,0.0020984055,0.0046090977,0.00007166185,0.0009929001],"genre_scores_gemma":[0.1818099,0.78288585,0.01633853,0.0026425002,0.00085314456,0.010403439,0.0045374963,0.0001765009,0.000352573],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.81124616,0.07281296,0.0883838,0.007428508,0.018628234,0.0015003502],"domain_scores_gemma":[0.582619,0.343701,0.04974921,0.011042745,0.0117947655,0.0010932755],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.121917546,0.0036429374,0.017592559,0.036835678,0.0013384599,0.007244383,0.0041519967,0.0030118164,0.005028376],"category_scores_gemma":[0.3447808,0.0031902927,0.029153602,0.0318357,0.002714668,0.0041030515,0.006004808,0.0023457354,0.0009726245],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012585947,0.000030958337,0.014758676,0.6854373,0.25510716,0.00030578295,0.00041523168,0.00054375257,0.0003956431,0.0005386293,0.0016610107,0.039547242],"study_design_scores_gemma":[0.0007159456,0.00020995663,0.02044233,0.31748906,0.64762396,0.0003142701,0.00028372623,0.000265325,0.00055868266,0.001756717,0.010262558,0.000077588666],"about_ca_topic_score_codex":0.0032942255,"about_ca_topic_score_gemma":0.008217225,"teacher_disagreement_score":0.98240745,"about_ca_system_score_codex":0.004487865,"about_ca_system_score_gemma":0.00965106,"threshold_uncertainty_score":0.6447693},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W3178875310","doi":"10.1186/s13012-021-01140-0","title":"Barriers and facilitators to implementing evidence-based guidelines in long-term care: a qualitative evidence synthesis","year":2021,"lang":"en","type":"review","venue":"Implementation Science","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":291,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Research Institute for Aging; University of Waterloo; McMaster University; Canada Research Chairs; Centre for Excellence in Mining Innovation; University of Toronto; Dalhousie University","funders":"Norwegian Institute of Public Health; Canadian Institutes of Health Research; McMaster University","keywords":"CINAHL; Knowledge translation; Medicine; Nursing; MEDLINE; Staffing; Long-term care; Critical appraisal; Implementation research; Health administration; Health services research; Inclusion (mineral); Teamwork; Qualitative research; Health care; Nursing research; Public health; Psychological intervention; Psychology; Alternative medicine; Knowledge management","score_opus":0.8042982745295546,"score_gpt":0.7805708647425855,"score_spread":0.023727409786969078,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3178875310","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3350892,0.39500117,0.065870285,0.031266566,0.0015732166,0.13156618,0.011954236,0.0002361323,0.027443126],"genre_scores_gemma":[0.5503202,0.23082778,0.09461468,0.0066589373,0.00019033097,0.11217514,0.0019377263,0.00011961433,0.0031556236],"study_design_codex":"systematic_review","study_design_gemma":"qualitative","domain_scores_codex":[0.9230461,0.05058268,0.01577811,0.0020194342,0.0065012313,0.002072388],"domain_scores_gemma":[0.76724076,0.20052575,0.009641326,0.0030024333,0.018427528,0.0011621725],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.13127603,0.0013354898,0.0051738997,0.012945777,0.0033955376,0.006421692,0.0030723712,0.0023728136,0.004532683],"category_scores_gemma":[0.18787405,0.0015702372,0.004236477,0.017604893,0.0035393592,0.006765979,0.0072092693,0.002186887,0.00039437463],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00046965093,0.0001732988,0.0051618638,0.4298843,0.0016478276,0.0010472535,0.3876873,0.0006443899,0.0010232549,0.0056951386,0.0037496884,0.16281608],"study_design_scores_gemma":[0.00040614617,0.00056542683,0.0057417704,0.58638835,0.006148409,0.00046721107,0.34126183,0.00053618324,0.002068573,0.004746344,0.051511724,0.00015799662],"about_ca_topic_score_codex":0.008241514,"about_ca_topic_score_gemma":0.016973356,"teacher_disagreement_score":0.13127603,"about_ca_system_score_codex":0.015564223,"about_ca_system_score_gemma":0.04633583,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3182386410","doi":"10.1097/ccm.0000000000005195","title":"Coronavirus Disease 2019 as Cause of Viral Sepsis: A Systematic Review and Meta-Analysis*","year":2021,"lang":"en","type":"review","venue":"Critical Care Medicine","topic":"COVID-19 Clinical Research Studies","field":"Medicine","cited_by":179,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Meta-analysis; Coronavirus disease 2019 (COVID-19); Sepsis; Coronavirus; 2019-20 coronavirus outbreak; Disease; Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); Intensive care medicine; Virology; Pandemic; Immunology; Infectious disease (medical specialty); Internal medicine; Outbreak","score_opus":0.361154865756216,"score_gpt":0.5950817646424796,"score_spread":0.2339268988862636,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3182386410","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.005341089,0.9926002,0.00043070118,0.00023800225,0.0001668639,0.00043807845,0.00057665666,0.000017745391,0.00019074055],"genre_scores_gemma":[0.19764864,0.7929441,0.0028994493,0.0012102516,0.00040328177,0.0033013173,0.0011370444,0.000030804425,0.00042505024],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.98578703,0.0067099673,0.0041405745,0.0013160864,0.0015268537,0.0005195336],"domain_scores_gemma":[0.9735474,0.018866632,0.0044454695,0.00081724656,0.0019758,0.00034737217],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.016478343,0.0023026252,0.016502487,0.007838262,0.0007749123,0.0037620042,0.0020940516,0.0023903968,0.003711387],"category_scores_gemma":[0.04230245,0.0013735115,0.034155566,0.009225138,0.0007497604,0.0020596175,0.0018798099,0.0018102437,0.00029726696],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012947681,0.000023527342,0.0064208056,0.5411762,0.4424099,0.00015350024,0.00009616184,0.00033798502,0.00015523855,0.0001239061,0.0007297299,0.0070782737],"study_design_scores_gemma":[0.0005482861,0.00022989487,0.0061360192,0.058327887,0.9323917,0.0001482338,0.000069208225,0.0002337391,0.00008710228,0.00017794943,0.0016208084,0.000029038867],"about_ca_topic_score_codex":0.0075131906,"about_ca_topic_score_gemma":0.017873673,"teacher_disagreement_score":0.016502487,"about_ca_system_score_codex":0.0031361638,"about_ca_system_score_gemma":0.0052922494,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3192941963","doi":"10.1111/jebm.12440","title":"Comparing and combining evidence of treatment effects in randomized and nonrandomized studies on the use of misoprostol to prevent postpartum hemorrhage","year":2021,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"Maternal and fetal healthcare","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Joseph’s Healthcare Hamilton; McMaster University; Impact","funders":"","keywords":"Misoprostol; Medicine; Randomized controlled trial; Confidence interval; Odds ratio; Meta-analysis; Placebo; MEDLINE; Relative risk; Obstetrics; Pregnancy; Internal medicine; Abortion; Alternative medicine","score_opus":0.4126879839645758,"score_gpt":0.4530451666427994,"score_spread":0.040357182678223635,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3192941963","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.029596044,0.93554205,0.018703764,0.0020407569,0.0024682647,0.008221055,0.0013075905,0.00015556517,0.0019649111],"genre_scores_gemma":[0.66166055,0.24767186,0.059430774,0.0034618324,0.0016676177,0.023465475,0.0019177385,0.00012316281,0.00060105504],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.48129758,0.41401604,0.07632426,0.008871492,0.018222293,0.001268313],"domain_scores_gemma":[0.4258801,0.51471525,0.039146062,0.010481659,0.008878917,0.00089801475],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.21822202,0.004074464,0.02588474,0.016527867,0.00087676244,0.0084884735,0.0034595036,0.0057750256,0.0030829972],"category_scores_gemma":[0.4802741,0.0028509384,0.046313316,0.009247375,0.0032292702,0.0048768986,0.0040208986,0.004018625,0.00030911746],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.007951856,0.00013880145,0.004675395,0.16611464,0.7944062,0.00010030986,0.0002634486,0.0016033401,0.00019797547,0.0010873834,0.000406333,0.023054365],"study_design_scores_gemma":[0.013489528,0.002488114,0.006451553,0.04992105,0.9172814,0.0001401345,0.00018082547,0.0020985967,0.00052152807,0.0049681976,0.0023191394,0.00013999759],"about_ca_topic_score_codex":0.0025058386,"about_ca_topic_score_gemma":0.0027951088,"teacher_disagreement_score":0.781778,"about_ca_system_score_codex":0.0050295424,"about_ca_system_score_gemma":0.005046077,"threshold_uncertainty_score":0.96407145},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_narrow","metaepi_broad","metaresearch"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3193216869","doi":"10.1161/jaha.120.020513","title":"Representation of Women in Randomized Trials in Cardiac Surgery: A Meta‐Analysis","year":2021,"lang":"en","type":"review","venue":"Journal of the American Heart Association","topic":"Cardiac, Anesthesia and Surgical Outcomes","field":"Medicine","cited_by":61,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; University of Toronto; Sunnybrook Health Science Centre","funders":"","keywords":"Medicine; Randomized controlled trial; Cochrane Library; Meta-analysis; Internal medicine; Clinical trial; Sample size determination; MEDLINE; Psychological intervention; Inclusion and exclusion criteria; Surgery; Alternative medicine","score_opus":0.10371017795454268,"score_gpt":0.4127714158278478,"score_spread":0.30906123787330514,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3193216869","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.011974904,0.97846425,0.0057552406,0.00080133375,0.00051637733,0.0012339784,0.00062604644,0.000060170278,0.0005676452],"genre_scores_gemma":[0.6072252,0.35735944,0.01976076,0.0036536513,0.001017564,0.009027967,0.0012001618,0.00013885184,0.00061641086],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.76675117,0.16077545,0.051918067,0.008897267,0.01024738,0.0014107095],"domain_scores_gemma":[0.6913367,0.26528352,0.030689772,0.006940027,0.0048931325,0.00085679174],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.12848175,0.003013631,0.018412067,0.00850878,0.0007705156,0.005812019,0.002424589,0.0034362827,0.0029756324],"category_scores_gemma":[0.23742957,0.0019555935,0.046267908,0.009340121,0.0015211804,0.0036236837,0.0027060322,0.0022038575,0.00029139506],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004533026,0.000035092206,0.014662765,0.11929275,0.84800667,0.00013413925,0.000167747,0.0005076199,0.00016620889,0.000334776,0.00047241428,0.011686781],"study_design_scores_gemma":[0.0022599534,0.0004761714,0.007346534,0.021093974,0.96404225,0.00018208144,0.00010617233,0.0007478304,0.00019142631,0.0012848838,0.0022303853,0.000038355065],"about_ca_topic_score_codex":0.0019864521,"about_ca_topic_score_gemma":0.0040204325,"teacher_disagreement_score":0.98158795,"about_ca_system_score_codex":0.002324571,"about_ca_system_score_gemma":0.0027742444,"threshold_uncertainty_score":0.67948455},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W3196952617","doi":"10.21203/rs.3.rs-665009/v2","title":"Topic overlap and research waste at the ‘Overviews of systematic reviews’ level: a meta-research study","year":2021,"lang":"en","type":"preprint","venue":"Research Square","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Nova Scotia Health Authority; University of Ottawa; McMaster University; University of British Columbia","funders":"","keywords":"Scope (computer science); Systematic review; Psychological intervention; Data science; Computer science; MEDLINE; Psychology; Information retrieval; Management science; Engineering; Political science","score_opus":0.9763343927920256,"score_gpt":0.7055023385718725,"score_spread":0.2708320542201531,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3196952617","genre_codex":"review","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.05565216,0.8827646,0.039299104,0.0060970723,0.0012513512,0.0067780786,0.0041252295,0.00031381447,0.0037186919],"genre_scores_gemma":[0.72268474,0.17317657,0.06450222,0.007039013,0.0016937065,0.025484232,0.004155597,0.00045323212,0.0008106882],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.3004323,0.43587008,0.18463506,0.020601602,0.0553492,0.003111835],"domain_scores_gemma":[0.1116894,0.7439393,0.0884292,0.03240213,0.022372304,0.0011677399],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.4557027,0.0034029328,0.013124043,0.03902824,0.0026321947,0.011663142,0.0042804983,0.0045361617,0.0048376513],"category_scores_gemma":[0.71398365,0.0050768014,0.02181954,0.03889708,0.005117901,0.014260174,0.013285891,0.0032249389,0.0005327686],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0060709366,0.00026093633,0.104784176,0.4566904,0.2545345,0.00092503603,0.011842319,0.0018979622,0.00076433364,0.010720989,0.0058156983,0.14569272],"study_design_scores_gemma":[0.0033323965,0.0018795066,0.08444998,0.37891066,0.4459984,0.0029836218,0.008285716,0.0032282034,0.0023915318,0.027799,0.040244404,0.00049665745],"about_ca_topic_score_codex":0.0051066503,"about_ca_topic_score_gemma":0.007752312,"teacher_disagreement_score":0.54429734,"about_ca_system_score_codex":0.01018877,"about_ca_system_score_gemma":0.014062484,"threshold_uncertainty_score":0.6712155},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3202107667","doi":"10.1016/j.cjco.2021.09.015","title":"Sex, Race, and Age Differences in Cardiovascular Outcomes in Implantable Cardioverter–Defibrillator Randomized Controlled Trials: A Systematic Review and Meta-analysis","year":2021,"lang":"en","type":"review","venue":"CJC Open","topic":"Cardiac pacing and defibrillation studies","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta; St. Michael's Hospital; Toronto Rehabilitation Institute; University of Waterloo; McGill University Health Centre; Queen's University; University of Toronto; University Health Network; University of Ontario Institute of Technology","funders":"","keywords":"Medicine; Randomized controlled trial; CINAHL; Implantable cardioverter-defibrillator; Odds ratio; Meta-analysis; Sudden cardiac death; Internal medicine; MEDLINE; Psychological intervention","score_opus":0.14021283057226855,"score_gpt":0.3994996188610652,"score_spread":0.2592867882887967,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3202107667","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0010426431,0.99805033,0.00023676814,0.00012470035,0.000070521244,0.00017022263,0.00016031263,0.000009170872,0.00013528315],"genre_scores_gemma":[0.04036967,0.95606107,0.0015398589,0.00053037173,0.00016714438,0.00089528284,0.00027695976,0.000014349004,0.00014532407],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9856631,0.0067052343,0.004654992,0.0010437566,0.0015943964,0.0003385169],"domain_scores_gemma":[0.951228,0.040086564,0.0056016673,0.00075477624,0.0020498678,0.00027902154],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.020761404,0.0022777424,0.016655048,0.007369536,0.00058047497,0.0033294174,0.0018916855,0.0020717522,0.0040091393],"category_scores_gemma":[0.05274901,0.0012446311,0.024804723,0.008434215,0.0007676654,0.0020498633,0.0013750729,0.0015986672,0.0002704802],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011463581,0.000034325276,0.0028955904,0.7260227,0.239064,0.000087531625,0.00009858051,0.0003005746,0.00015132042,0.0002364666,0.0008521387,0.029110344],"study_design_scores_gemma":[0.0008291936,0.00026474544,0.004341682,0.122329585,0.86722916,0.00013687332,0.00008861605,0.00018394692,0.00016258062,0.00043147604,0.0039722836,0.000029922536],"about_ca_topic_score_codex":0.0046249432,"about_ca_topic_score_gemma":0.011460599,"teacher_disagreement_score":0.020761404,"about_ca_system_score_codex":0.0025951294,"about_ca_system_score_gemma":0.004859296,"threshold_uncertainty_score":0.109798074},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3203021509","doi":"10.1016/j.ocarto.2021.100215","title":"The state of trial registrations in the field of Orthopaedics in years 2015–2020. A meta-epidemiological study","year":2021,"lang":"en","type":"review","venue":"Osteoarthritis and Cartilage Open","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; Hand and Upper Limb Clinic; St. Joseph's Hospital; University of Toronto; Queen's University; Canadian Institute for Military and Veteran Health Research","funders":"","keywords":"Medicine; Randomized controlled trial; Epidemiology; Population; Meta-analysis; Orthopedic surgery; Trial registration; MEDLINE; Prospective cohort study; Surgery; Internal medicine; Environmental health","score_opus":0.6498741816734732,"score_gpt":0.5627773797430697,"score_spread":0.08709680193040348,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3203021509","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015485421,0.97764957,0.0020388472,0.0012207321,0.00047628224,0.0003044296,0.0022735475,0.000043400305,0.0005076868],"genre_scores_gemma":[0.50789744,0.4700512,0.009085977,0.0027354546,0.0011207941,0.002515747,0.005972543,0.000121553196,0.00049923954],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.79612136,0.09007979,0.07469391,0.01540847,0.02189679,0.0017997017],"domain_scores_gemma":[0.57715225,0.24000934,0.14164497,0.018229526,0.020729145,0.002234763],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.20573384,0.0013585058,0.00820613,0.0181867,0.0008502262,0.0068990043,0.0026111642,0.0032023948,0.0019283291],"category_scores_gemma":[0.29811412,0.0024958656,0.029006371,0.017375592,0.001505309,0.0065910956,0.0032764222,0.0026936375,0.00045998418],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0065400656,0.000091966074,0.17781134,0.29941678,0.4397282,0.00019358797,0.0005878034,0.0011259841,0.0005143505,0.0016682924,0.0034402418,0.0688814],"study_design_scores_gemma":[0.0017228829,0.0012535799,0.20610978,0.11587629,0.6494489,0.0011239423,0.00067489775,0.0018954349,0.0010285346,0.0023074935,0.01834854,0.00020971741],"about_ca_topic_score_codex":0.0032870737,"about_ca_topic_score_gemma":0.0069624963,"teacher_disagreement_score":0.9986415,"about_ca_system_score_codex":0.00287389,"about_ca_system_score_gemma":0.0057497993,"threshold_uncertainty_score":0.97947156},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3209288188","doi":"10.2196/28102","title":"The Application of Human-Centered Design Approaches in Health Research and Innovation: A Narrative Review of Current Practices","year":2021,"lang":"en","type":"review","venue":"JMIR mhealth and uhealth","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":357,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Radboud Universitair Medisch Centrum; Radboud Universiteit","keywords":"CINAHL; PsycINFO; Participatory design; User-centered design; Context (archaeology); Computer science; Knowledge management; Research design; Health care; Management science; MEDLINE; Medicine; Psychological intervention; Sociology; Engineering; Nursing; Social science; Human–computer interaction","score_opus":0.963688333660892,"score_gpt":0.7975905777322816,"score_spread":0.16609775592861042,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3209288188","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00055968517,0.993686,0.0013859808,0.0024582425,0.00030087348,0.00021062988,0.000044680222,0.000009124302,0.0013447582],"genre_scores_gemma":[0.011877918,0.98042756,0.0048679304,0.0016088594,0.0001822568,0.000784128,0.000041640393,0.000008949767,0.00020078672],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.93235534,0.041893285,0.014370494,0.0028509465,0.0076404754,0.0008893707],"domain_scores_gemma":[0.77231985,0.20368826,0.0108258,0.003428776,0.008956648,0.0007806843],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.058897275,0.0016463107,0.0034586163,0.017534716,0.00248845,0.00885491,0.002735045,0.004444544,0.0025663844],"category_scores_gemma":[0.13575235,0.0015094365,0.0033908933,0.01806012,0.0056907427,0.00867478,0.0051536183,0.0033104932,0.000606312],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020136718,0.000060382852,0.00075373653,0.5854797,0.0007750957,0.00055584154,0.020369755,0.00047914594,0.000574974,0.031228198,0.010248754,0.34927312],"study_design_scores_gemma":[0.000059727005,0.00015486039,0.001003579,0.79766566,0.0014652193,0.00096042594,0.007647849,0.00020006698,0.0004419794,0.006348292,0.18398535,0.00006699698],"about_ca_topic_score_codex":0.0046176305,"about_ca_topic_score_gemma":0.008198632,"teacher_disagreement_score":0.058897275,"about_ca_system_score_codex":0.009624258,"about_ca_system_score_gemma":0.03488212,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"opus","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W3214101616","doi":"10.1002/mpr.1900","title":"Selective outcome reporting across psychopharmacotherapy randomized controlled trials","year":2021,"lang":"en","type":"article","venue":"International Journal of Methods in Psychiatric Research","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Women's College Hospital; University of Toronto","funders":"","keywords":"Medicine; Odds; Randomized controlled trial; Logistic regression; Odds ratio; Psychiatry; Major depressive disorder; Schizophrenia (object-oriented programming); Clinical trial; Sample size determination; Clinical psychology; Internal medicine; Mood","score_opus":0.9193925220984189,"score_gpt":0.7997913039352579,"score_spread":0.11960121816316105,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3214101616","genre_codex":"review","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.09952902,0.42390716,0.282486,0.025030324,0.013088925,0.10092255,0.035580803,0.002157968,0.017297274],"genre_scores_gemma":[0.55544716,0.06627173,0.204674,0.01601611,0.004076726,0.13945691,0.009548751,0.00076877576,0.0037398722],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.20964117,0.46573427,0.2553579,0.02575519,0.041541427,0.0019701214],"domain_scores_gemma":[0.107241325,0.6828906,0.13910417,0.049076624,0.02056724,0.001120086],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.5391673,0.0031416768,0.00975218,0.0142586455,0.0018947113,0.0067500486,0.005628797,0.0049267425,0.0036278593],"category_scores_gemma":[0.74356407,0.0025714955,0.010106326,0.015441817,0.0064038998,0.005428882,0.005193271,0.0036347914,0.00077483774],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.013284461,0.00037844796,0.08698199,0.55813557,0.09597064,0.002629801,0.009284996,0.0039356397,0.0033878887,0.03008084,0.031409737,0.16452],"study_design_scores_gemma":[0.012146441,0.0024281114,0.10899358,0.38044646,0.13366105,0.0048254346,0.0035910872,0.01777237,0.018971214,0.16958235,0.14635673,0.0012251218],"about_ca_topic_score_codex":0.0022053595,"about_ca_topic_score_gemma":0.0020241598,"teacher_disagreement_score":0.46083272,"about_ca_system_score_codex":0.0054336493,"about_ca_system_score_gemma":0.009566255,"threshold_uncertainty_score":0.5682888},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","research_integrity"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","research_integrity"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4200240932","doi":"10.1093/nutrit/nuab093","title":"Are systematic reviews addressing nutrition for cancer prevention trustworthy? A systematic survey of quality and risk of bias","year":2021,"lang":"en","type":"review","venue":"Nutrition Reviews","topic":"Nutrition and Health in Aging","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"Narodowe Centrum Nauki; China Academy of Chinese Medical Sciences; Uniwersytet Jagielloński Collegium Medicum","keywords":"Observational study; Medicine; Systematic review; Cochrane Library; Data extraction; Meta-analysis; MEDLINE; Publication bias; Environmental health; Family medicine; Population; Internal medicine","score_opus":0.6607547283756908,"score_gpt":0.5532416583848381,"score_spread":0.10751306999085275,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4200240932","genre_codex":"review","genre_gemma":"review","domain_codex":"methods","domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.02046459,0.8642517,0.047369003,0.037099335,0.0039386963,0.014447922,0.00865261,0.0003810886,0.0033951],"genre_scores_gemma":[0.44176042,0.34253493,0.11726849,0.033434696,0.0031430894,0.055320162,0.0053271637,0.00044293012,0.000768136],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.2366673,0.3863958,0.294892,0.015783535,0.06441149,0.0018498003],"domain_scores_gemma":[0.064841315,0.7525485,0.114654474,0.024610681,0.04170755,0.0016374263],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.5820934,0.0023727291,0.01509125,0.023368513,0.00275765,0.01098542,0.004976255,0.0058460436,0.004668367],"category_scores_gemma":[0.85348964,0.004064519,0.013829777,0.023846159,0.0077675926,0.012935776,0.0071663805,0.0046375785,0.00085283886],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001202501,0.000057555306,0.031162867,0.80561715,0.051734097,0.0002189727,0.006856187,0.0006911951,0.0004978041,0.004171224,0.0092593,0.088531084],"study_design_scores_gemma":[0.0012379169,0.00046416037,0.026550896,0.8518622,0.05942052,0.0007150247,0.004802784,0.0019341096,0.0007756318,0.017463895,0.034400415,0.00037248738],"about_ca_topic_score_codex":0.00712802,"about_ca_topic_score_gemma":0.011332272,"teacher_disagreement_score":0.41790658,"about_ca_system_score_codex":0.011178895,"about_ca_system_score_gemma":0.025899654,"threshold_uncertainty_score":0.51535326},"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","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4200365875","doi":"10.22541/au.163827914.48781130/v1","title":"Empirical evaluation of fundamental principles of evidence-based medicine: a meta-epidemiological study","year":2021,"lang":"en","type":"preprint","venue":"","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Grading (engineering); Meta-analysis; Medicine; Odds ratio; Systematic review; Publication bias; Statistics; MEDLINE; Econometrics; Internal medicine; Mathematics; Chemistry","score_opus":0.9923385322462839,"score_gpt":0.7148133160779869,"score_spread":0.27752521616829695,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4200365875","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.019507796,0.85719913,0.09454573,0.0057517714,0.0016582788,0.015596208,0.0026967586,0.00028273958,0.0027615335],"genre_scores_gemma":[0.55831313,0.19912344,0.19303606,0.005481561,0.0015986344,0.03968038,0.0019398696,0.00019443415,0.00063249195],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.5545799,0.3524344,0.05352275,0.013373719,0.024806693,0.001282562],"domain_scores_gemma":[0.37305745,0.55591124,0.03653033,0.020862678,0.012376676,0.0012617174],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.2955369,0.0033648852,0.014874203,0.014534951,0.0011896426,0.009511434,0.0048377775,0.005858839,0.0048375893],"category_scores_gemma":[0.6038576,0.0033906726,0.027465025,0.013032181,0.0034872552,0.0069845906,0.0038393198,0.004003591,0.0006180256],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004561684,0.00024153107,0.030771863,0.3876622,0.48711154,0.0005224769,0.0007952172,0.0051429626,0.00059258257,0.0077371993,0.0023091887,0.07255162],"study_design_scores_gemma":[0.0057086847,0.003142127,0.025673172,0.21838799,0.68199944,0.0010744013,0.00062693004,0.015167904,0.0012776821,0.033201404,0.013365033,0.00037527658],"about_ca_topic_score_codex":0.0017489777,"about_ca_topic_score_gemma":0.0022143975,"teacher_disagreement_score":0.99663514,"about_ca_system_score_codex":0.0045039007,"about_ca_system_score_gemma":0.008205552,"threshold_uncertainty_score":0.8687284},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4200475773","doi":"10.1177/15586898211054243","title":"Systematic Reviews of Systematic Quantitative, Qualitative, and Mixed Studies Reviews in Healthcare Research: How to Assess the Methodological Quality of Included Reviews?","year":2021,"lang":"en","type":"article","venue":"Journal of Mixed Methods Research","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval; The Quebec Population Health Research Network; Université de Montréal; McGill University; Centre Hospitalier de l’Université de Montréal","funders":"","keywords":"Systematic review; Typology; Management science; Qualitative research; Multimethodology; Critical appraisal; Qualitative property; Health care; Quality (philosophy); Data science; Computer science; Psychology; Sociology; MEDLINE; Medicine; Social science; Alternative medicine; Epistemology; Engineering; Political science; Pathology","score_opus":0.9950623349363318,"score_gpt":0.8309246850568849,"score_spread":0.1641376498794469,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4200475773","genre_codex":"review","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0036568707,0.8312311,0.08107453,0.049451947,0.010051026,0.018627647,0.0019727505,0.00043603635,0.0034981247],"genre_scores_gemma":[0.08535467,0.37744835,0.405487,0.021150002,0.0056699123,0.10220386,0.0013377216,0.00044178966,0.0009067227],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.14190029,0.5504927,0.24030533,0.010491097,0.055074025,0.0017366282],"domain_scores_gemma":[0.07407113,0.72370696,0.09882159,0.023368632,0.0782975,0.0017341563],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.67922074,0.0043642507,0.023026658,0.04751373,0.006324611,0.022353034,0.009745764,0.010620872,0.004395],"category_scores_gemma":[0.8964524,0.0063115135,0.018660046,0.047422595,0.011082456,0.022595963,0.012933292,0.00871779,0.0013813936],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007803598,0.00006999047,0.0046089645,0.7061953,0.02762565,0.0002361012,0.007839044,0.0009393275,0.0005835909,0.017024873,0.01939277,0.21470411],"study_design_scores_gemma":[0.0011538651,0.00037225313,0.0041093603,0.8742,0.029498424,0.0004405311,0.003648761,0.002182589,0.00077396387,0.041274983,0.04190029,0.00044502737],"about_ca_topic_score_codex":0.009059548,"about_ca_topic_score_gemma":0.0161261,"teacher_disagreement_score":0.97697335,"about_ca_system_score_codex":0.021384459,"about_ca_system_score_gemma":0.05251207,"threshold_uncertainty_score":0.39557797},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4214868455","doi":"10.2196/35332","title":"Evaluation of Digital Interventions for Physical Activity Promotion: Protocol for a Scoping Review","year":2022,"lang":"en","type":"review","venue":"JMIR Research Protocols","topic":"Physical Activity and Health","field":"Medicine","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Leibniz-Gemeinschaft; Freie Hansestadt Bremen","keywords":"PsycINFO; CINAHL; Systematic review; Context (archaeology); Inclusion (mineral); Grey literature; Psychological intervention; MEDLINE; Health promotion; Promotion (chess); Medicine; Intervention (counseling); Psychology; Medical education; Nursing; Public health; Political science; Social psychology","score_opus":0.9342517862276597,"score_gpt":0.7874797437112254,"score_spread":0.14677204251643428,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4214868455","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00018647722,0.00049844006,0.0010993372,0.0001307027,0.00012554652,0.9965391,0.00095149194,0.000041148804,0.0004276264],"genre_scores_gemma":[0.00013618624,0.0004601531,0.003029901,0.00006609817,0.000014683572,0.9959258,0.00019338362,0.000004504407,0.00016931588],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9501534,0.021408355,0.01690681,0.0033431523,0.0054867757,0.0027015472],"domain_scores_gemma":[0.9291971,0.023695221,0.012000861,0.006844376,0.024940377,0.0033221394],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.103608824,0.0066380277,0.012870075,0.014927306,0.005652729,0.0076066935,0.005196756,0.008726889,0.10560052],"category_scores_gemma":[0.092239365,0.006055349,0.016143199,0.013310781,0.004887638,0.009279448,0.008527688,0.007976795,0.022231497],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.019047653,0.0024183078,0.001573427,0.7221501,0.0021771071,0.0020147962,0.00725972,0.0034809504,0.005989826,0.0107051,0.06091184,0.16227128],"study_design_scores_gemma":[0.0454615,0.0067860456,0.008237801,0.43586534,0.005041355,0.0010544483,0.0053916387,0.0021408407,0.0050513344,0.016523365,0.46761072,0.000835678],"about_ca_topic_score_codex":0.0041869134,"about_ca_topic_score_gemma":0.0088604065,"teacher_disagreement_score":0.10560052,"about_ca_system_score_codex":0.012259548,"about_ca_system_score_gemma":0.061423138,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4224314394","doi":"10.1007/s10899-022-10122-w","title":"Emerging Gambling Problems and Suggested Interventions: A Systematic Review of Empirical Research","year":2022,"lang":"en","type":"review","venue":"Journal of Gambling Studies","topic":"Gambling Behavior and Treatments","field":"Psychology","cited_by":16,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"Alberta Gambling Research Institute, University of Calgary","keywords":"Scopus; Psychology; Harm; Psychological intervention; Inclusion (mineral); Intervention (counseling); Empirical research; Systematic review; Applied psychology; MEDLINE; Social psychology; Psychiatry","score_opus":0.6723972921458732,"score_gpt":0.6382039077752859,"score_spread":0.034193384370587276,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4224314394","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00027581726,0.9992938,0.000059271944,0.000114773415,0.000038758924,0.00005673347,0.00007872457,0.0000030177607,0.00007906269],"genre_scores_gemma":[0.0029160532,0.9962949,0.00038327064,0.00020816413,0.000027762948,0.000080749225,0.000058898386,0.0000019035346,0.00002825865],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.99607277,0.0011997394,0.001433686,0.00040787866,0.0007533001,0.0001326338],"domain_scores_gemma":[0.9832895,0.013538176,0.0020833581,0.00016961919,0.0007182865,0.0002010683],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.007357786,0.001470801,0.008649087,0.0076675476,0.0005919086,0.0028944707,0.002176893,0.0023709564,0.00546174],"category_scores_gemma":[0.029201211,0.0008578247,0.0063095135,0.006719882,0.0012393326,0.0020620222,0.0017974868,0.0019576831,0.00030709946],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00041628283,0.00006395259,0.0005449424,0.8566933,0.005816448,0.000075857555,0.00013949494,0.00012267794,0.0001283734,0.0004318132,0.0019379766,0.13362898],"study_design_scores_gemma":[0.00058912346,0.00024615863,0.0034814451,0.9102514,0.05887218,0.00036303324,0.00029643564,0.00011248982,0.00015001667,0.0008762963,0.024718018,0.00004343392],"about_ca_topic_score_codex":0.006792769,"about_ca_topic_score_gemma":0.023435129,"teacher_disagreement_score":0.008649087,"about_ca_system_score_codex":0.002742818,"about_ca_system_score_gemma":0.008229109,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4224442502","doi":"10.1007/s10459-022-10103-z","title":"Exploring professional identity in rehabilitation professions: a scoping review","year":2022,"lang":"en","type":"review","venue":"Advances in Health Sciences Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":51,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; McGill University Health Centre; Centre for Interdisciplinary Research in Rehabilitation","funders":"Fonds de Recherche du Québec - Santé","keywords":"Identity (music); Rehabilitation; Medical education; Health professions; Medicine; Psychology; Physical therapy; Health care; Political science","score_opus":0.2165631896154671,"score_gpt":0.5923900897870247,"score_spread":0.37582690017155757,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4224442502","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00010424837,0.9994759,0.000050111732,0.00013692412,0.00006257243,0.000018747187,0.000020191648,0.0000012771436,0.00013008335],"genre_scores_gemma":[0.0009437874,0.99858236,0.00016991445,0.00018562107,0.000034412962,0.000023485085,0.000019543935,8.725858e-7,0.000039954433],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.9965491,0.0009817742,0.0010278274,0.00041039742,0.00084687024,0.00018386693],"domain_scores_gemma":[0.9827802,0.0140697425,0.0016078535,0.00020073583,0.0011065622,0.00023491324],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0077202828,0.001220021,0.004941051,0.009285744,0.00083995925,0.0044960272,0.0016787307,0.0030786828,0.004130697],"category_scores_gemma":[0.019704487,0.00080291025,0.0033940228,0.010338847,0.0014442942,0.00327824,0.0025042326,0.0019512511,0.0005324934],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011105258,0.0000715381,0.0007698475,0.5950298,0.0018835426,0.00012283365,0.0005734527,0.00015307099,0.00023758052,0.0016050963,0.0057652127,0.39367697],"study_design_scores_gemma":[0.000055952623,0.00008595154,0.0028534723,0.880806,0.0081758695,0.00043116705,0.0008467514,0.00007805673,0.00013590453,0.0012348514,0.105254136,0.000041933923],"about_ca_topic_score_codex":0.007448538,"about_ca_topic_score_gemma":0.022723343,"teacher_disagreement_score":0.009285744,"about_ca_system_score_codex":0.0025723677,"about_ca_system_score_gemma":0.01187086,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4231686601","doi":"10.31222/osf.io/b79mk","title":"Patient-important outcomes reported in clinical studies of pharmacologic treatments for covid-19: A protocol of a meta-epidemiological study","year":2020,"lang":"en","type":"preprint","venue":"","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Impact; McMaster University; Western University; Université de Montréal; University of Saskatchewan","funders":"","keywords":"Medicine; Epidemiology; Randomized controlled trial; Protocol (science); Pandemic; MEDLINE; Coronavirus disease 2019 (COVID-19); Clinical trial; Meta-analysis; Disease; Intensive care medicine; Family medicine; Alternative medicine; Internal medicine; Pathology","score_opus":0.9322724632072774,"score_gpt":0.6826235854805558,"score_spread":0.24964887772672162,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4231686601","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0027633032,0.0073816027,0.019310584,0.0013982181,0.0011881277,0.9580701,0.008274358,0.00033526172,0.0012784924],"genre_scores_gemma":[0.0033803927,0.0010712668,0.016925659,0.0003754099,0.0000545851,0.97702014,0.00089644163,0.000026737509,0.00024930143],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.8502342,0.09047709,0.04310063,0.006593305,0.007643584,0.0019511717],"domain_scores_gemma":[0.86596215,0.07381083,0.017894123,0.022444244,0.018020224,0.0018683975],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.16617206,0.005681848,0.013333878,0.00960399,0.0022091349,0.0064484966,0.0038502614,0.007164094,0.023727823],"category_scores_gemma":[0.25639054,0.0039593223,0.049919587,0.008977322,0.0026934901,0.004094141,0.005506393,0.008167389,0.0032034046],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.05953558,0.0013398465,0.009064446,0.6470866,0.16788223,0.0013784245,0.0026509615,0.008269325,0.003346952,0.014092874,0.03175579,0.053596973],"study_design_scores_gemma":[0.15169908,0.008501128,0.025246594,0.21548557,0.29217032,0.0009937782,0.0017074092,0.014284119,0.006515684,0.054079045,0.22820528,0.001112087],"about_ca_topic_score_codex":0.00214288,"about_ca_topic_score_gemma":0.003874502,"teacher_disagreement_score":0.9943181,"about_ca_system_score_codex":0.006364002,"about_ca_system_score_gemma":0.019062791,"threshold_uncertainty_score":0.8788123},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"reporting","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4240239701","doi":"10.31234/osf.io/c8zk2","title":"A Methodological Review of Exploratory Factor Analysis in Sexuality Research: Used Practices, Best Practices, and Data Analysis Resources","year":2017,"lang":"en","type":"review","venue":"","topic":"Cognitive and psychological constructs research","field":"Psychology","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Human sexuality; Exploratory factor analysis; Psychology; Exploratory research; Best practice; Order (exchange); Sample (material); Set (abstract data type); Exploratory analysis; Social psychology; Sociology; Data science; Computer science; Psychometrics; Political science; Social science; Clinical psychology; Business; Gender studies","score_opus":0.9750119899436716,"score_gpt":0.7506804499057284,"score_spread":0.2243315400379432,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4240239701","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005577304,0.9870046,0.006677139,0.0026256137,0.0004501729,0.00047815478,0.00026213503,0.000040840863,0.0019036705],"genre_scores_gemma":[0.0043746633,0.96877635,0.023588158,0.00095805636,0.00023689734,0.0014474264,0.0002635899,0.000049866136,0.00030505663],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.921449,0.044599414,0.01768707,0.002938309,0.012696242,0.00062996236],"domain_scores_gemma":[0.74449986,0.1966282,0.01343638,0.006772185,0.037642222,0.0010211077],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.09988348,0.0017461001,0.003370748,0.024242567,0.0020181185,0.0059587327,0.0030311157,0.0019945635,0.003724595],"category_scores_gemma":[0.21749915,0.0014903673,0.00340816,0.027631244,0.0032448466,0.006410247,0.0037460523,0.0029444557,0.0015059693],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007987099,0.00005838602,0.0020940606,0.1578227,0.00072032196,0.00012744192,0.003649913,0.00031610657,0.00040962666,0.00960618,0.021173567,0.80394185],"study_design_scores_gemma":[0.000061995874,0.00015633993,0.0067879893,0.47875604,0.002567328,0.00080841244,0.0034725913,0.0003338173,0.0008000411,0.012784907,0.49331927,0.00015129911],"about_ca_topic_score_codex":0.0080424305,"about_ca_topic_score_gemma":0.017579494,"teacher_disagreement_score":0.9001165,"about_ca_system_score_codex":0.0070589646,"about_ca_system_score_gemma":0.026928963,"threshold_uncertainty_score":0.5282407},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4248363397","doi":"10.21203/rs.3.rs-71534/v1","title":"Bias Estimation In Study Design: A Meta-Epidemiological Analysis of Transcatheter Versus Surgical Aortic Valve Replacement","year":2020,"lang":"en","type":"preprint","venue":"Research Square (Research Square)","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Women's College Hospital; University of Toronto; Western University","funders":"University of Toronto; Canadian Institutes of Health Research; University of Ottawa","keywords":"Meta-analysis; Estimation; Epidemiology; Medicine; Aortic valve replacement; Cardiology; Internal medicine; Economics; Stenosis","score_opus":0.8706539368585138,"score_gpt":0.6069626284785796,"score_spread":0.26369130837993415,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4248363397","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.021338463,0.89422554,0.069095545,0.002251208,0.0024574155,0.007100901,0.0018184903,0.00024298643,0.0014693623],"genre_scores_gemma":[0.67093986,0.16159412,0.12297357,0.0058389814,0.0022148932,0.032387964,0.002477135,0.00032230024,0.0012512524],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.553023,0.33561054,0.07220057,0.017253563,0.02043182,0.0014804042],"domain_scores_gemma":[0.44836792,0.46555448,0.05546111,0.019445108,0.010304548,0.0008668379],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.27495456,0.0040110378,0.020127146,0.013516284,0.001126451,0.0075749676,0.004265252,0.0060623786,0.0030829299],"category_scores_gemma":[0.4646929,0.0026151203,0.04905964,0.009935109,0.0023921635,0.006060819,0.004581439,0.0037026608,0.00033371372],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004635835,0.00007099158,0.02072775,0.14700358,0.7954223,0.00019933733,0.00038632232,0.0021740438,0.00026073278,0.0019396825,0.0011055003,0.02607389],"study_design_scores_gemma":[0.005392054,0.0010104747,0.011948753,0.05559899,0.9041945,0.0004405214,0.00020769483,0.0048972038,0.0007913012,0.01014595,0.0051896656,0.000182942],"about_ca_topic_score_codex":0.0019396481,"about_ca_topic_score_gemma":0.0023521876,"teacher_disagreement_score":0.99598897,"about_ca_system_score_codex":0.0036308235,"about_ca_system_score_gemma":0.004035911,"threshold_uncertainty_score":0.8941101},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaepi_narrow","metaepi_broad","metaresearch"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4253731119","doi":"10.32920/14636286","title":"The effectiveness of knowledge translation interventions for promoting evidence-informed decision-making among nurses in tertiary care: a systematic review and meta-analysis","year":2021,"lang":"en","type":"review","venue":"","topic":"Health Sciences Research and Education","field":"Health Professions","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"St. Peter's Hospital; McMaster Children's Hospital; Hamilton General Hospital; McMaster University","funders":"McMaster University; London Health Sciences Centre; Canadian Institutes of Health Research; Hamilton Health Sciences","keywords":"CINAHL; Psychological intervention; Data extraction; Knowledge translation; MEDLINE; Systematic review; Meta-analysis; Health care; Psychology; Medicine; Nursing; Knowledge management; Computer science","score_opus":0.34128447801581957,"score_gpt":0.6168342622352857,"score_spread":0.27554978421946613,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4253731119","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0068086796,0.9862578,0.0019208518,0.00038431672,0.00035299442,0.0034525841,0.00032082415,0.000057111356,0.00044471858],"genre_scores_gemma":[0.18130231,0.7906888,0.011247377,0.0012251998,0.00030817796,0.014125472,0.0006147958,0.000060973893,0.00042685238],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.94697547,0.02747312,0.015896823,0.002651137,0.0062243557,0.00077909994],"domain_scores_gemma":[0.9019574,0.07427503,0.011864295,0.003078886,0.0078714155,0.0009529183],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.06501529,0.0044431034,0.028100574,0.016166557,0.0015268205,0.0059654415,0.0036766117,0.0035414284,0.0037351223],"category_scores_gemma":[0.13979575,0.0027847665,0.05410175,0.012070781,0.0017712285,0.0046571586,0.0037766295,0.0030195753,0.00035995347],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019707151,0.00010038212,0.0011311729,0.5674237,0.40782312,0.000084802334,0.00029670185,0.00042129259,0.00014991837,0.00017272621,0.0003085373,0.020117018],"study_design_scores_gemma":[0.0012379972,0.000571755,0.0014480937,0.0911136,0.9031933,0.00006402143,0.00014081754,0.00027866056,0.00026227615,0.00027094435,0.001373406,0.000045151664],"about_ca_topic_score_codex":0.009994497,"about_ca_topic_score_gemma":0.018608604,"teacher_disagreement_score":0.06501529,"about_ca_system_score_codex":0.010185226,"about_ca_system_score_gemma":0.01557537,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4281479985","doi":"10.1136/bmjopen-2021-053417","title":"Reporting quality of clinical trial protocols: a repeated cross-sectional study about the Adherence to SPIrit Recommendations in Switzerland, CAnada and GErmany (ASPIRE-SCAGE)","year":2022,"lang":"en","type":"article","venue":"BMJ Open","topic":"Ethics in Clinical Research","field":"Medicine","cited_by":21,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Ottawa; Centre for Disability Prevention and Rehabilitation; Canadian Memorial Chiropractic College; McMaster University; Impact","funders":"University of Oxford; Bundesamt für Gesundheit; Alexander von Humboldt-Stiftung; Albert-Ludwigs-Universität Freiburg; Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung; National Science Foundation","keywords":"Medicine; Checklist; Protocol (science); Randomized controlled trial; Clinical trial; Family medicine; Consolidated Standards of Reporting Trials; Cross-sectional study; Guideline; Trial registration; Alternative medicine; Internal medicine; Pathology","score_opus":0.8625137923322216,"score_gpt":0.7433109279867433,"score_spread":0.11920286434547833,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4281479985","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8838392,0.026871562,0.02265599,0.018345892,0.0008298145,0.030580353,0.009180477,0.0004270037,0.0072696446],"genre_scores_gemma":[0.95901775,0.0016657958,0.014650541,0.0038519474,0.00017122166,0.01699854,0.0029491081,0.00013021086,0.0005647995],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.43652588,0.3189395,0.121161364,0.021479916,0.096436985,0.005456448],"domain_scores_gemma":[0.15266177,0.38684413,0.24393515,0.05326044,0.15549926,0.0077994238],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.5260737,0.0007002555,0.0020166738,0.006857666,0.003690445,0.005894035,0.003152073,0.0029396287,0.0022067775],"category_scores_gemma":[0.65267444,0.0014488547,0.002886828,0.008001488,0.0058727083,0.0037982285,0.005052541,0.0032093478,0.00038246252],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0027338276,0.0006365044,0.8903811,0.007455029,0.0042071515,0.00015825295,0.019812452,0.00061967166,0.00072994345,0.0017282631,0.011772077,0.05976571],"study_design_scores_gemma":[0.0012217655,0.0016098332,0.9677538,0.005769525,0.0012603285,0.00026469567,0.0041769976,0.0016923044,0.0009582626,0.00060945156,0.014380617,0.00030243385],"about_ca_topic_score_codex":0.12682098,"about_ca_topic_score_gemma":0.14410223,"teacher_disagreement_score":0.9690451,"about_ca_system_score_codex":0.030954909,"about_ca_system_score_gemma":0.065116905,"threshold_uncertainty_score":0.5844356},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":true,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4285090825","doi":"10.1101/2022.07.11.22277498","title":"Assessment of concordance between related systematic reviews and between related guideline recommendations: protocol for a methodological survey","year":2022,"lang":"en","type":"preprint","venue":"medRxiv","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; St. Michael's Hospital; Toronto Rehabilitation Institute; University Health Network; University of Toronto; McMaster University; Impact","funders":"","keywords":"Concordance; Guideline; Protocol (science); Systematic review; Clinical Practice; Medicine; MEDLINE; Psychology; Management science; Family medicine; Alternative medicine; Pathology; Political science; Engineering","score_opus":0.928813207935435,"score_gpt":0.6732457410896231,"score_spread":0.2555674668458119,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4285090825","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0010838353,0.00020328966,0.009493393,0.00061209773,0.00030872115,0.9836187,0.0039388007,0.00014616929,0.0005949352],"genre_scores_gemma":[0.0004884574,0.00005439572,0.007183681,0.00011813384,0.000015397825,0.9917755,0.0002585007,0.000009678732,0.00009629223],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.7032181,0.16872875,0.09265013,0.008601418,0.02261454,0.0041870424],"domain_scores_gemma":[0.53875434,0.23467876,0.055165436,0.05492447,0.11043088,0.006046134],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.3477611,0.0038996763,0.00582241,0.011017743,0.0042024893,0.005984002,0.00299486,0.0069987266,0.05386314],"category_scores_gemma":[0.4877772,0.0056406623,0.010774191,0.011038686,0.004951438,0.0057042115,0.007511242,0.008438833,0.014236063],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.052571006,0.004475956,0.008668228,0.30009893,0.008298935,0.0014236685,0.02327963,0.007723313,0.0074537857,0.039561477,0.18755588,0.35888916],"study_design_scores_gemma":[0.1274102,0.012816907,0.034615945,0.2044403,0.009419608,0.00096081797,0.007909196,0.015041638,0.015408254,0.08471702,0.48494568,0.002314315],"about_ca_topic_score_codex":0.0029920307,"about_ca_topic_score_gemma":0.0040520607,"teacher_disagreement_score":0.6522389,"about_ca_system_score_codex":0.012102289,"about_ca_system_score_gemma":0.05271058,"threshold_uncertainty_score":0.8043267},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"evaluation","study_design":"observational","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4289132654","doi":"10.1111/jebm.12483","title":"Clinical manifestations of COVID‐19: An overview of 102 systematic reviews with evidence mapping","year":2022,"lang":"en","type":"review","venue":"Journal of Evidence-Based Medicine","topic":"COVID-19 Clinical Research Studies","field":"Medicine","cited_by":37,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Medicine; Cochrane Library; Meta-analysis; MEDLINE; Systematic review; Coronavirus disease 2019 (COVID-19); Internal medicine; Web of science; Disease; Infectious disease (medical specialty)","score_opus":0.8614589479472008,"score_gpt":0.6529562840453669,"score_spread":0.20850266390183392,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4289132654","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0017351091,0.994031,0.00048056984,0.00056990486,0.00014551873,0.0010854179,0.0015596374,0.00003294218,0.00035985708],"genre_scores_gemma":[0.024792517,0.96628344,0.0029177573,0.00097052177,0.000111277404,0.0035789823,0.0011442355,0.00002847483,0.0001727795],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.96313596,0.01078774,0.01924582,0.0019369227,0.0041428385,0.00075074215],"domain_scores_gemma":[0.904565,0.06856245,0.017726729,0.0016697984,0.0068137874,0.00066226034],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.028363481,0.0030947884,0.0120685585,0.031625587,0.0012479561,0.0042780214,0.0032955268,0.0025149407,0.0051579783],"category_scores_gemma":[0.104604565,0.0022989758,0.014463647,0.028613457,0.0015279759,0.003719105,0.0037581243,0.0018338392,0.00052055845],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001769263,0.000010146141,0.00093950785,0.96999884,0.0116418945,0.000101452955,0.00022793545,0.000086134445,0.00014727331,0.00019428994,0.0011020573,0.015373588],"study_design_scores_gemma":[0.00020551588,0.00012891027,0.0041445526,0.8946982,0.08868588,0.0003052675,0.00034834078,0.00008508584,0.0001703632,0.00029943278,0.010889133,0.000039308095],"about_ca_topic_score_codex":0.010219863,"about_ca_topic_score_gemma":0.029340103,"teacher_disagreement_score":0.031625587,"about_ca_system_score_codex":0.008015777,"about_ca_system_score_gemma":0.016576977,"threshold_uncertainty_score":0.15000224},"labels":[{"model":"gemma","categories":["metaepi_broad","metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4293483968","doi":"10.1136/bmjopen-2021-060237","title":"Factors Associated with the Magnitude Of acUpuncture treatment effectS (FAMOUS): a meta-epidemiological study of acupuncture randomised controlled trials","year":2022,"lang":"en","type":"article","venue":"BMJ Open","topic":"Acupuncture Treatment Research Studies","field":"Medicine","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; Western University; York University; Hamilton Health Sciences; University of Toronto; McMaster University; Impact","funders":"China Academy of Chinese Medical Sciences; National Natural Science Foundation of China","keywords":"Medicine; Acupuncture; Meta-analysis; Meta-regression; Randomized controlled trial; MEDLINE; Sample size determination; Internal medicine; Clinical trial; Physical therapy; Alternative medicine; Pathology; Statistics","score_opus":0.3686658038902662,"score_gpt":0.4846991083205923,"score_spread":0.11603330443032611,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4293483968","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.016317954,0.9727296,0.006672631,0.00080020336,0.0006207579,0.001327996,0.00090663077,0.00007642369,0.0005478065],"genre_scores_gemma":[0.68962866,0.27328402,0.023602419,0.0027230103,0.0013860445,0.006798961,0.0017922515,0.00015927538,0.0006252641],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.87152594,0.077020206,0.032448348,0.008856333,0.008941423,0.0012077439],"domain_scores_gemma":[0.7562602,0.19877212,0.029722022,0.007867597,0.0064182905,0.0009597564],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.102775425,0.003960193,0.019177293,0.00894196,0.00082466565,0.0060300743,0.003091102,0.004198546,0.0039544194],"category_scores_gemma":[0.21283299,0.0024646313,0.06587533,0.009538807,0.0018219538,0.004067805,0.0027858168,0.0037249434,0.0003758918],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0045918045,0.00003650635,0.01346164,0.09305917,0.8817271,0.00008988079,0.000085847256,0.00044380015,0.00018955136,0.00025801314,0.0002569171,0.00579971],"study_design_scores_gemma":[0.0015431445,0.0003416139,0.0072470997,0.010073707,0.97868663,0.00010090092,0.00004783886,0.0003772361,0.00012585832,0.00064906257,0.0007764118,0.000030446387],"about_ca_topic_score_codex":0.002151804,"about_ca_topic_score_gemma":0.0029867424,"teacher_disagreement_score":0.9960398,"about_ca_system_score_codex":0.0021511794,"about_ca_system_score_gemma":0.003133684,"threshold_uncertainty_score":0.5435349},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4293918656","doi":"10.21203/rs.3.rs-2010358/v1","title":"Methodological review of NMA bias concepts provides groundwork for the development of a list of concepts for potential inclusion in a new risk of bias tool for network meta-analysis (RoB NMA Tool)","year":2022,"lang":"en","type":"preprint","venue":"Research Square","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Innovative Research Group (Canada); Ottawa Hospital; University of British Columbia","funders":"National Institute for Health Research Applied Research Collaboration West; Medical Research Council; Department of Health and Social Care; University of Bristol; Canadian Institutes of Health Research; National Institute for Health and Care Research; NIHR Bristol Biomedical Research Centre; University Hospitals Bristol NHS Foundation Trust","keywords":"Popularity; Publication bias; Quality (philosophy); Systematic review; MEDLINE; Quality of evidence; Meta-analysis; Computer science; Inclusion (mineral); Psychology; Medicine; Social psychology; Epistemology","score_opus":0.9320928261565841,"score_gpt":0.6570412207702426,"score_spread":0.27505160538634144,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4293918656","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00052296417,0.0072570145,0.9840762,0.0046439976,0.0005546919,0.00069619546,0.00044171498,0.00044720786,0.0013600942],"genre_scores_gemma":[0.014659056,0.0038405152,0.9742546,0.0014070626,0.0006173865,0.004282715,0.00028022198,0.00034051307,0.00031798976],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"systematic_review","domain_scores_codex":[0.5904572,0.2934299,0.07095916,0.008751505,0.035349898,0.0010522829],"domain_scores_gemma":[0.2421836,0.66881335,0.026360098,0.033352744,0.02807644,0.0012138139],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.36476198,0.0039723115,0.0076046386,0.023070455,0.0031414723,0.015355072,0.008176282,0.0077589313,0.013207152],"category_scores_gemma":[0.6944471,0.0035195334,0.016678477,0.013773124,0.0070796194,0.021981586,0.012643604,0.019560408,0.0034368774],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003677875,0.000108751825,0.0026768888,0.026900608,0.0058635115,0.00022879442,0.002920873,0.0060575744,0.0009037003,0.5935317,0.017581917,0.34285775],"study_design_scores_gemma":[0.00026943453,0.00013211192,0.0008040409,0.015108993,0.0028247717,0.0005288802,0.00030536114,0.018286988,0.0015512622,0.93017715,0.029710937,0.00030011137],"about_ca_topic_score_codex":0.0019390816,"about_ca_topic_score_gemma":0.0028595154,"teacher_disagreement_score":0.63523805,"about_ca_system_score_codex":0.00579202,"about_ca_system_score_gemma":0.01152797,"threshold_uncertainty_score":0.78336155},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"methods","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":"W4307503711","doi":"10.23736/s2724-5683.22.06230-5","title":"Therapeutic vs. prophylactic anticoagulation in COVID-19 patients: a systematic review and meta-analysis of real-world studies","year":2022,"lang":"en","type":"review","venue":"Minerva Cardiology and Angiology","topic":"COVID-19 Clinical Research Studies","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University Health Centre","funders":"","keywords":"Meta-analysis; Coronavirus disease 2019 (COVID-19); Medicine; 2019-20 coronavirus outbreak; Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); Internal medicine; Intensive care medicine; Virology; Disease","score_opus":0.30400328313915576,"score_gpt":0.5143024761269444,"score_spread":0.2102991929877886,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4307503711","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0052119037,0.99295604,0.0004944393,0.00020136227,0.00010023723,0.00026587854,0.0005661705,0.000017807031,0.0001860846],"genre_scores_gemma":[0.20063978,0.792439,0.002596012,0.0010506893,0.0002725408,0.0014280253,0.0012485011,0.00003336424,0.00029203383],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.98930883,0.00499205,0.0031714702,0.0010434907,0.0011291371,0.00035500326],"domain_scores_gemma":[0.9758252,0.018117901,0.0038525804,0.0006248309,0.0013266354,0.00025292175],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.015251809,0.0024280408,0.01828915,0.006159997,0.0005907867,0.0035951617,0.0019275527,0.002068472,0.0032331746],"category_scores_gemma":[0.033522587,0.0012505567,0.035466507,0.0072987033,0.0006995048,0.0015973002,0.0013762631,0.0017609458,0.00023481925],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0020385552,0.000024614092,0.004467318,0.38016823,0.60268927,0.00011623654,0.000075692624,0.0005115798,0.00016895148,0.0001308047,0.00052358897,0.009085134],"study_design_scores_gemma":[0.00044775222,0.0001451889,0.0033819396,0.027984366,0.9663812,0.00008265997,0.000038455673,0.00017716723,0.00007851978,0.00016351245,0.001100081,0.00001918577],"about_ca_topic_score_codex":0.005547957,"about_ca_topic_score_gemma":0.0102627445,"teacher_disagreement_score":0.01828915,"about_ca_system_score_codex":0.0027173671,"about_ca_system_score_gemma":0.003370554,"threshold_uncertainty_score":0.08066028},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4308446649","doi":"10.7717/peerj.14320","title":"Assessment of the association between periodontal disease and total cancer incidence and mortality: a meta-analysis","year":2022,"lang":"en","type":"review","venue":"PeerJ","topic":"Oral microbiology and periodontitis research","field":"Dentistry","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Meta-analysis; Internal medicine; Breast cancer; Lung cancer; Incidence (geometry); Subgroup analysis; Publication bias; Cancer; Population; Prostate cancer; Oncology; Environmental health","score_opus":0.15065719619435253,"score_gpt":0.44812607121056885,"score_spread":0.2974688750162163,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4308446649","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015706716,0.977998,0.0034522002,0.0004341164,0.0003398502,0.00046132383,0.0011115599,0.00008776972,0.00040841574],"genre_scores_gemma":[0.49773625,0.4838454,0.010410055,0.0015272317,0.0007801021,0.002371111,0.0023515518,0.00012526338,0.0008529701],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.98027724,0.011308224,0.004034014,0.0020905805,0.0018136796,0.0004762209],"domain_scores_gemma":[0.9706106,0.022021325,0.0034979284,0.0016174959,0.0018488262,0.00040385735],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.030301685,0.004619134,0.019535847,0.0100502195,0.0009797347,0.004687005,0.0032738894,0.0030867087,0.003962748],"category_scores_gemma":[0.046270724,0.0022555226,0.08584678,0.008481601,0.0008356861,0.0020876573,0.002105567,0.003004674,0.00038835197],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016956518,0.000022897719,0.00579556,0.036215674,0.9514205,0.00007909456,0.00003141295,0.0007601613,0.00013806655,0.00006315573,0.00021595934,0.003561826],"study_design_scores_gemma":[0.000295854,0.00009525556,0.0022870463,0.0022006056,0.9942159,0.000049237526,0.000013068911,0.00028574583,0.00006735949,0.00011862662,0.00035875454,0.000012499686],"about_ca_topic_score_codex":0.0056873388,"about_ca_topic_score_gemma":0.009119331,"teacher_disagreement_score":0.030301685,"about_ca_system_score_codex":0.0028414472,"about_ca_system_score_gemma":0.002809743,"threshold_uncertainty_score":0.16025257},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4310773520","doi":"10.1016/j.jclinepi.2022.11.020","title":"Indicators of questionable research practices were identified in 163,129 randomized controlled trials","year":2022,"lang":"en","type":"article","venue":"Journal of Clinical Epidemiology","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Public Health; Ottawa Hospital; University of Ottawa","funders":"ZonMw","keywords":"Blinding; Randomized controlled trial; Checklist; Proxy (statistics); Sample size determination; Publication bias; Medicine; Logistic regression; MEDLINE; Impact factor; Consolidated Standards of Reporting Trials; Family medicine; Meta-analysis; Statistics; Psychology; Surgery; Internal medicine","score_opus":0.9659537253604227,"score_gpt":0.7731422889942364,"score_spread":0.19281143636618625,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4310773520","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5170551,0.2657498,0.10191469,0.009412263,0.0011776256,0.039665446,0.051207554,0.0012738806,0.012543648],"genre_scores_gemma":[0.84573364,0.020227462,0.093150206,0.003114801,0.0003166733,0.025978064,0.010241758,0.0003811638,0.0008561361],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.43181184,0.17717539,0.29469952,0.018295094,0.07438092,0.003637305],"domain_scores_gemma":[0.09640565,0.6853954,0.15923296,0.034880385,0.02303758,0.001047916],"candidate_categories":["metaresearch","metaepi_broad","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.3683602,0.0016530291,0.0045755687,0.024197012,0.0026025786,0.005628525,0.002396447,0.003896984,0.0021930442],"category_scores_gemma":[0.7493345,0.0020586974,0.0059350375,0.034212478,0.0040528267,0.0052808076,0.006822972,0.0023986425,0.00090234645],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0064546843,0.0004400144,0.5498558,0.16260733,0.019390294,0.0040234425,0.017212605,0.002142551,0.0029612295,0.009356082,0.012340475,0.21321546],"study_design_scores_gemma":[0.0035063007,0.0016549944,0.5272685,0.22832178,0.05973985,0.010193075,0.00912589,0.00794605,0.013700198,0.031628847,0.106135696,0.0007787567],"about_ca_topic_score_codex":0.0016989941,"about_ca_topic_score_gemma":0.0033764506,"teacher_disagreement_score":0.996103,"about_ca_system_score_codex":0.005337709,"about_ca_system_score_gemma":0.013809378,"threshold_uncertainty_score":0.77892435},"labels":[{"model":"gemma","categories":["metaresearch","research_integrity"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad","research_integrity"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4311338564","doi":"10.1016/j.jval.2022.09.2165","title":"PCR232 The Use of EuroQol EQ-5D in the Middle East and North Africa Region: A Systematic Review","year":2022,"lang":"en","type":"review","venue":"Value in Health","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"CINAHL; PsycINFO; Data extraction; Systematic review; Medicine; EQ-5D; MEDLINE; Population; Inclusion (mineral); Family medicine; Psychology; Health related quality of life; Environmental health; Psychiatry; Psychological intervention; Political science; Pathology","score_opus":0.9323004053396625,"score_gpt":0.4378622853480493,"score_spread":0.49443811999161325,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4311338564","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00045748087,0.99854726,0.00010362592,0.0001452362,0.00007458287,0.000120695324,0.0004026905,0.0000036851527,0.00014466915],"genre_scores_gemma":[0.009161806,0.9893818,0.00054747076,0.0003732456,0.000048217873,0.00020522978,0.00020370532,0.0000048439633,0.00007370579],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9925759,0.0024616169,0.0027728947,0.0005850991,0.0013825733,0.00022184657],"domain_scores_gemma":[0.9833872,0.012010954,0.003038548,0.00021762366,0.0011846926,0.00016098004],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00946134,0.0015181683,0.00790602,0.009458727,0.00048447074,0.003081582,0.0016186968,0.0017744385,0.005907758],"category_scores_gemma":[0.041716125,0.0009346646,0.008844224,0.010904266,0.00091646734,0.0023931614,0.002115944,0.001537639,0.00032796586],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00026441956,0.000017912067,0.0015534131,0.90167636,0.017659556,0.00006249923,0.00011999903,0.0001289734,0.00010997184,0.00040423273,0.001965245,0.07603732],"study_design_scores_gemma":[0.0005163278,0.00016825733,0.009390666,0.8470467,0.11377988,0.00030728988,0.0003162285,0.00012134295,0.00016815965,0.0005465883,0.02756966,0.00006890394],"about_ca_topic_score_codex":0.013898052,"about_ca_topic_score_gemma":0.03359387,"teacher_disagreement_score":0.013898052,"about_ca_system_score_codex":0.0033605527,"about_ca_system_score_gemma":0.01007316,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"grok","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"opus","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4311586875","doi":"10.1186/s13104-022-06263-1","title":"Competition for recruitment in SARS-CoV-2 Trials in the United States: a longitudinal cohort analysis","year":2022,"lang":"en","type":"article","venue":"BMC Research Notes","topic":"Ethics in Clinical Research","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Ethica (Canada)","funders":"Narodowe Centrum Nauki","keywords":"Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); Coronavirus disease 2019 (COVID-19); Cohort; 2019-20 coronavirus outbreak; Competition (biology); Medicine; Cohort study; Sars virus; Demography; Virology; Biology; Outbreak; Disease; Infectious disease (medical specialty); Internal medicine; Ecology","score_opus":0.9400359417216635,"score_gpt":0.7114902230524065,"score_spread":0.228545718669257,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4311586875","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9939361,0.0012204554,0.0019951642,0.0008471158,0.00005740261,0.00021302207,0.0008404478,0.000019253926,0.00087106915],"genre_scores_gemma":[0.99801755,0.00009462123,0.000632074,0.00023931694,0.000030631338,0.00024192466,0.00035085413,0.000009580459,0.00038339238],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.96028244,0.0275611,0.0035320814,0.00350653,0.0029350482,0.002182826],"domain_scores_gemma":[0.887462,0.0476706,0.044132814,0.012447858,0.004202414,0.004084248],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.07881692,0.00035673607,0.00068684656,0.00081175874,0.0014463402,0.0019426903,0.001321408,0.0014886758,0.003882746],"category_scores_gemma":[0.09454417,0.0006584046,0.0024516142,0.0013767536,0.0006444326,0.0023243476,0.0016319528,0.0020989699,0.00039460216],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001041244,0.00014566886,0.99406356,0.000034858986,0.0007837036,0.000045722267,0.00026883464,0.00017957331,0.00007335288,0.0002519903,0.0006364454,0.002475052],"study_design_scores_gemma":[0.00021164316,0.0012674236,0.99085313,0.000105629755,0.0008077205,0.00019426248,0.0006272242,0.0041414155,0.0001234028,0.00035766678,0.0012770997,0.00003338541],"about_ca_topic_score_codex":0.0072130687,"about_ca_topic_score_gemma":0.009226454,"teacher_disagreement_score":0.9211831,"about_ca_system_score_codex":0.0010481597,"about_ca_system_score_gemma":0.00302793,"threshold_uncertainty_score":0.4168287},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4311692627","doi":"10.1053/j.ajkd.2022.10.015","title":"Sex and Gender in Randomized Controlled Trials of Adults Receiving Maintenance Dialysis: A Meta-epidemiologic Study","year":2022,"lang":"en","type":"article","venue":"American Journal of Kidney Diseases","topic":"Sex and Gender in Healthcare","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Impact; McMaster University; Population Health Research Institute; University of Calgary; University of Alberta","funders":"Medical Research Council","keywords":"Medicine; Randomized controlled trial; Meta-analysis; Dialysis; MEDLINE; Gerontology; Intensive care medicine; Internal medicine","score_opus":0.08894198904737506,"score_gpt":0.3763907996304896,"score_spread":0.28744881058311456,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4311692627","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.21149547,0.78020424,0.0031517115,0.00081458694,0.0011929737,0.0006269642,0.0011644796,0.00013351913,0.0012160448],"genre_scores_gemma":[0.9324796,0.060260434,0.0025880416,0.0022636638,0.0006063476,0.0005441779,0.0006752315,0.000056410943,0.0005260995],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.98148715,0.01124457,0.003737161,0.0021087872,0.00074793777,0.00067436637],"domain_scores_gemma":[0.97215265,0.019332785,0.004943959,0.0021411132,0.00090583425,0.0005237337],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.016979363,0.0027161387,0.012934918,0.0016731452,0.00083870674,0.003645159,0.0017054473,0.0036176215,0.0034780542],"category_scores_gemma":[0.042084355,0.0018920449,0.034760773,0.0027524277,0.0011190552,0.0023618483,0.0013565937,0.0033342026,0.00029110073],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.046383355,0.000098841425,0.020855298,0.019507132,0.90842944,0.00011789854,0.00007326396,0.00032910495,0.00027530442,0.00012021162,0.0003203452,0.003489814],"study_design_scores_gemma":[0.017687533,0.001195062,0.014135198,0.001654615,0.9637384,0.000075739306,0.0000730159,0.00030056867,0.00010424013,0.00037116688,0.00062814355,0.000036328824],"about_ca_topic_score_codex":0.0034007488,"about_ca_topic_score_gemma":0.005244759,"teacher_disagreement_score":0.9972839,"about_ca_system_score_codex":0.0012355443,"about_ca_system_score_gemma":0.001214925,"threshold_uncertainty_score":0.08979654},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4313824971","doi":"10.2196/43722","title":"Accumulation of Biological and Behavioral Data of Female Sex Workers Using Respondent-Driven Sampling: Protocol for a Systematic Review","year":2023,"lang":"en","type":"review","venue":"JMIR Research Protocols","topic":"HIV, Drug Use, Sexual Risk","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Respondent; Observational study; Protocol (science); Population; Sampling (signal processing); Data collection; Data quality; Scale (ratio); Data extraction; Systematic review; Observational methods in psychology; Scopus; Psychology; Medicine; MEDLINE; Environmental health; Computer science; Statistics; Alternative medicine; Geography; Biology; Political science; Engineering","score_opus":0.954106471357703,"score_gpt":0.7654041862458333,"score_spread":0.1887022851118697,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4313824971","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00028533285,0.0015001546,0.0014173383,0.0003927569,0.00031427707,0.99281025,0.0028935727,0.00007752161,0.00030879126],"genre_scores_gemma":[0.00028344063,0.00067791017,0.0022918736,0.00016357716,0.000023770472,0.9960955,0.00032173368,0.0000058509604,0.00013640012],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.93066156,0.030280143,0.022720119,0.005425519,0.008009875,0.002902775],"domain_scores_gemma":[0.9132982,0.03634909,0.018134603,0.008016657,0.021390343,0.0028111353],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.1005793,0.0061327997,0.017030885,0.014016393,0.0041817813,0.007467479,0.0054348195,0.0098407725,0.049746957],"category_scores_gemma":[0.15669064,0.005899592,0.016273804,0.0119569115,0.0062281387,0.009110154,0.005649704,0.0077508157,0.008405507],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0033537124,0.00021828435,0.00073965726,0.93302774,0.0028226739,0.00059399556,0.0022412348,0.0010090758,0.0015207822,0.0048508397,0.02619441,0.023427594],"study_design_scores_gemma":[0.026257712,0.0020216666,0.005673983,0.7555578,0.0136273345,0.00084241334,0.0024803237,0.0018902834,0.0028779558,0.015046117,0.17304216,0.00068230735],"about_ca_topic_score_codex":0.005286492,"about_ca_topic_score_gemma":0.010505104,"teacher_disagreement_score":0.1005793,"about_ca_system_score_codex":0.017111259,"about_ca_system_score_gemma":0.07145812,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":[],"domain":null,"study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4318973081","doi":"10.1136/bmjebm-2022-112053","title":"Different meta-analysis methods can change judgements about imprecision of effect estimates: a meta-epidemiological study","year":2023,"lang":"en","type":"review","venue":"BMJ evidence-based medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Meta-analysis; Statistics; Random effects model; Mathematics; Medicine; Restricted maximum likelihood; Maximum likelihood; Econometrics; Internal medicine","score_opus":0.9820049695252123,"score_gpt":0.7362051535903333,"score_spread":0.24579981593487898,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4318973081","genre_codex":"review","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.07259309,0.7599143,0.13563804,0.014048181,0.004762994,0.0060245413,0.0013071193,0.00047067902,0.0052410904],"genre_scores_gemma":[0.67572343,0.09723771,0.20566167,0.008544596,0.0018030117,0.009344206,0.00074471725,0.0003418459,0.00059885555],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.23611733,0.66761416,0.055996377,0.0096991295,0.029705776,0.0008672691],"domain_scores_gemma":[0.07915136,0.85974246,0.031956133,0.019031867,0.009654375,0.00046374576],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.52488333,0.0047944183,0.010054894,0.013212399,0.0017836511,0.013354538,0.0053790645,0.010224596,0.0024265368],"category_scores_gemma":[0.8112393,0.003440132,0.0466065,0.0103593785,0.0044246293,0.009911896,0.0052897,0.00863645,0.0003976717],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.018402154,0.00035817298,0.043099377,0.09420867,0.67202526,0.00079606264,0.0035138705,0.011045581,0.0012114411,0.009483041,0.0038987657,0.14195758],"study_design_scores_gemma":[0.016322209,0.005534968,0.03397552,0.096296035,0.7514195,0.0019067223,0.0017108571,0.023487408,0.0048072264,0.050429583,0.012790962,0.0013189776],"about_ca_topic_score_codex":0.0016326802,"about_ca_topic_score_gemma":0.002557819,"teacher_disagreement_score":0.9899451,"about_ca_system_score_codex":0.0050377194,"about_ca_system_score_gemma":0.003658023,"threshold_uncertainty_score":0.5859035},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4320007456","doi":"10.2196/preprints.45426","title":"Clinical Effectiveness, Feasibility, Acceptability, and Usability in Mobile Health Applications for Epilepsy: A Systematic Review (Preprint)","year":2022,"lang":"en","type":"review","venue":"","topic":"Mobile Health and mHealth Applications","field":"Health Professions","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Usability; Systematic review; Medicine; MEDLINE; Randomized controlled trial; Consolidated Standards of Reporting Trials; Meta-analysis; Family medicine; Medical physics; Computer science; Internal medicine","score_opus":0.2069863324394874,"score_gpt":0.5815197134930307,"score_spread":0.3745333810535433,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4320007456","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0037489715,0.98907477,0.0005149849,0.0007020167,0.0003256703,0.0039926153,0.0010609622,0.000020593558,0.00055946835],"genre_scores_gemma":[0.060977433,0.92105746,0.003755223,0.0014793276,0.00034263424,0.011245119,0.0007737263,0.000019504612,0.00034952475],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9653518,0.012877273,0.013030054,0.0016140539,0.006583365,0.00054349034],"domain_scores_gemma":[0.913401,0.061833832,0.014999536,0.0012613848,0.008009173,0.00049499434],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.029334713,0.0017745722,0.008579136,0.0074650156,0.000855211,0.004900932,0.0016927634,0.0023582815,0.0056604184],"category_scores_gemma":[0.12643708,0.0012451848,0.014781219,0.008475728,0.0011906497,0.0036798094,0.002274733,0.0017149494,0.0003734996],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002826009,0.000021258125,0.00094043935,0.9631542,0.011994261,0.000041007264,0.00020236292,0.000045714944,0.000098582146,0.00013298853,0.00097465835,0.022111896],"study_design_scores_gemma":[0.0006202513,0.00035017388,0.0063505103,0.87624764,0.10707137,0.0002265892,0.00038400147,0.00013486843,0.0002621636,0.00026655023,0.008033242,0.000052672534],"about_ca_topic_score_codex":0.0056485976,"about_ca_topic_score_gemma":0.016630918,"teacher_disagreement_score":0.029334713,"about_ca_system_score_codex":0.006248871,"about_ca_system_score_gemma":0.015061801,"threshold_uncertainty_score":0.15513861},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4323536717","doi":"10.1101/2023.03.05.23286821","title":"The impact of blinding on trial results: A systematic review and meta-analysis","year":2023,"lang":"en","type":"review","venue":"medRxiv","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Impact; Dalhousie University; McMaster University","funders":"","keywords":"Blinding; Meta-analysis; Publication bias; Medicine; Randomized controlled trial; Systematic review; MEDLINE; Odds ratio; Strictly standardized mean difference; Clinical trial; Internal medicine","score_opus":0.9072607679519703,"score_gpt":0.6315100895610637,"score_spread":0.2757506783909066,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4323536717","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0041060876,0.974442,0.011863859,0.00144757,0.0014019154,0.0050886455,0.0007842194,0.00014800324,0.00071768067],"genre_scores_gemma":[0.33384162,0.5626088,0.053132135,0.006472404,0.0021797612,0.03911738,0.0015076542,0.00036178215,0.00077857426],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.5087068,0.3403097,0.10207275,0.016181313,0.0304803,0.0022491757],"domain_scores_gemma":[0.40500537,0.49809808,0.061215937,0.018244071,0.016219433,0.0012171996],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.28436247,0.00606214,0.029588504,0.014707067,0.0018993374,0.009593459,0.0054420307,0.007467702,0.0045041186],"category_scores_gemma":[0.49657547,0.004064127,0.05436026,0.013128436,0.005688641,0.009669267,0.004743444,0.0063236495,0.0005759949],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0032949718,0.000055478544,0.003380594,0.5568624,0.40832195,0.00012953821,0.00026954658,0.001169468,0.0001769676,0.0011112833,0.0012005712,0.02402715],"study_design_scores_gemma":[0.005555801,0.00062474544,0.0039886436,0.21770234,0.75939244,0.00022793993,0.00013672908,0.0020315123,0.0007126069,0.0056357575,0.0038229316,0.00016852403],"about_ca_topic_score_codex":0.005304501,"about_ca_topic_score_gemma":0.008023398,"teacher_disagreement_score":0.9704115,"about_ca_system_score_codex":0.00980934,"about_ca_system_score_gemma":0.013890129,"threshold_uncertainty_score":0.8825085},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4323537830","doi":"10.1117/1.nph.10.2.023518","title":"Reproducibility of infant fNIRS studies: a meta-analytic approach","year":2023,"lang":"en","type":"article","venue":"Neurophotonics","topic":"Child and Animal Learning Development","field":"Psychology","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"Ministerio de Ciencia e Innovación; Ikerbasque, Basque Foundation for Science","keywords":"Neuroimaging; Sample size determination; Reproducibility; Psychology; Cognitive psychology; Contrast (vision); Meta-analysis; Repetition (rhetorical device); Multiple comparisons problem; Audiology; Developmental psychology; Computer science; Artificial intelligence; Statistics; Neuroscience; Medicine; Mathematics; Linguistics","score_opus":0.18972864079193918,"score_gpt":0.37961874807244356,"score_spread":0.18989010728050437,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4323537830","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"reproducibility","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reproducibility","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.061206892,0.83672297,0.089364216,0.001564045,0.001132615,0.0031096637,0.004552944,0.0007769371,0.0015696388],"genre_scores_gemma":[0.8882796,0.05024938,0.048911784,0.0015046274,0.0003934098,0.007883677,0.001996278,0.00031058036,0.00047069593],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.8525095,0.08477295,0.033728026,0.017317962,0.010380298,0.0012912977],"domain_scores_gemma":[0.64689636,0.3031132,0.021903355,0.021156043,0.0063021854,0.00062882964],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.12784085,0.0032965152,0.012508356,0.009857146,0.0009839016,0.004802033,0.004505892,0.0032464247,0.0026124278],"category_scores_gemma":[0.23968747,0.0017967905,0.047090147,0.009020765,0.0019545504,0.0032282404,0.0033586207,0.0025496716,0.00035871542],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014542372,0.000023325078,0.023998857,0.04937871,0.91232693,0.0001848245,0.00025248833,0.00091910956,0.0005709252,0.00056721934,0.0004414746,0.009881873],"study_design_scores_gemma":[0.0006538806,0.0004687406,0.020935316,0.008087133,0.9617485,0.00027110972,0.00024571424,0.0018479652,0.0008027784,0.002176126,0.0026818386,0.00008097003],"about_ca_topic_score_codex":0.0025387458,"about_ca_topic_score_gemma":0.0035867724,"teacher_disagreement_score":0.8721591,"about_ca_system_score_codex":0.0022316936,"about_ca_system_score_gemma":0.00207984,"threshold_uncertainty_score":0.6760951},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reproducibility","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reproducibility","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4366597924","doi":"10.1136/bmjopen-2022-069906","title":"Strategies used to manage overlap of primary study data by exercise-related overviews: protocol for a systematic methodological review","year":2023,"lang":"en","type":"article","venue":"BMJ Open","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Public Health Ontario; University of Toronto; Institute of Health Services and Policy Research","funders":"Universitat Autònoma de Barcelona; Universitat de Barcelona","keywords":"Medicine; Systematic review; MEDLINE; Psychological intervention; Cochrane Library; Protocol (science); Grey literature; Alternative medicine; Pathology; Nursing","score_opus":0.9715586565184136,"score_gpt":0.7256562755993499,"score_spread":0.2459023809190637,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4366597924","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00022477684,0.0007729512,0.0031868434,0.00035415366,0.00038864338,0.99308383,0.0015458753,0.00012865424,0.00031431214],"genre_scores_gemma":[0.00019289386,0.0002578612,0.0048239683,0.000096549404,0.000020389756,0.9943551,0.00014197,0.000009994731,0.000101360325],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.778077,0.12411627,0.065600164,0.009740683,0.017442787,0.005023048],"domain_scores_gemma":[0.7288084,0.12314541,0.047161717,0.028227506,0.06645375,0.006203144],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.2366192,0.008536994,0.017911715,0.018842842,0.0049337437,0.009411323,0.00566124,0.013325154,0.05836253],"category_scores_gemma":[0.29086673,0.0085379705,0.01872624,0.018015219,0.008549313,0.011480387,0.0073433477,0.011433037,0.01879867],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.012628871,0.0006552265,0.0012200737,0.8089997,0.003893204,0.001029781,0.008598322,0.002544993,0.0039267787,0.013421995,0.064560734,0.078520335],"study_design_scores_gemma":[0.05031373,0.003702352,0.007183424,0.5252861,0.0076309,0.0009918472,0.0053411704,0.004711778,0.0066325758,0.047557924,0.33940884,0.001239422],"about_ca_topic_score_codex":0.0030372692,"about_ca_topic_score_gemma":0.0057465867,"teacher_disagreement_score":0.76338077,"about_ca_system_score_codex":0.014832938,"about_ca_system_score_gemma":0.062870786,"threshold_uncertainty_score":0.94138443},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4367053085","doi":"10.3389/fnagi.2023.1051809","title":"Gout and risk of dementia, Alzheimer's disease or vascular dementia: a meta-epidemiology study","year":2023,"lang":"en","type":"review","venue":"Frontiers in Aging Neuroscience","topic":"Gout, Hyperuricemia, Uric Acid","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"Natural Science Foundation of Zhejiang Province; National Natural Science Foundation of China","keywords":"Medicine; Dementia; Internal medicine; Vascular dementia; Meta-analysis; Publication bias; Relative risk; Gout; Cohort study; Confidence interval; Disease","score_opus":0.16801500793577467,"score_gpt":0.38459283011815293,"score_spread":0.21657782218237825,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4367053085","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.012633327,0.98503995,0.0009852601,0.00023874421,0.0001941334,0.00021785205,0.00042063367,0.000034186643,0.00023599606],"genre_scores_gemma":[0.361708,0.62871015,0.004971197,0.0011403838,0.0006094518,0.0012650054,0.0009933437,0.000050472954,0.0005519965],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.98892844,0.00613767,0.0022122443,0.0013436431,0.0010175066,0.00036056835],"domain_scores_gemma":[0.9850492,0.010285038,0.0026646429,0.0008043236,0.0009650898,0.00023164277],"candidate_categories":["metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.018275782,0.003778532,0.016644204,0.006469338,0.000762655,0.0047760312,0.0020172237,0.0024253617,0.002497443],"category_scores_gemma":[0.030852688,0.0017840808,0.05846332,0.007221265,0.0007069949,0.0016430715,0.0018283649,0.002054113,0.00024337342],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019152091,0.000021294118,0.0068327417,0.049984377,0.9365778,0.000095894524,0.000029365627,0.0004957159,0.00017320756,0.000060030496,0.00015842692,0.0036558716],"study_design_scores_gemma":[0.00040863297,0.00016462075,0.0041407174,0.0039598513,0.9903502,0.000094318166,0.000016775422,0.00020502199,0.00008820358,0.00012985505,0.00042851063,0.000013348257],"about_ca_topic_score_codex":0.0060066185,"about_ca_topic_score_gemma":0.0086999405,"teacher_disagreement_score":0.9833558,"about_ca_system_score_codex":0.0020485509,"about_ca_system_score_gemma":0.002230579,"threshold_uncertainty_score":0.096652746},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4376270084","doi":"10.1007/s12630-023-02462-x","title":"Black representation in critical care randomized controlled trials: a meta-epidemiological study","year":2023,"lang":"en","type":"review","venue":"Canadian Journal of Anesthesia/Journal canadien d anesthésie","topic":"Ethics in Clinical Research","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Juravinski Hospital; Impact; University of Ottawa; McMaster University; University of Toronto","funders":"","keywords":"Randomized controlled trial; Medicine; Generalizability theory; Psychological intervention; Demographics; Population; Demography; Confidence interval; Epidemiology; Family medicine; Environmental health; Psychology; Nursing; Internal medicine","score_opus":0.611815265220914,"score_gpt":0.5717538600750817,"score_spread":0.04006140514583234,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4376270084","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.019265797,0.9733891,0.004702465,0.0004589574,0.00056526775,0.0004957227,0.0004913853,0.00005381801,0.00057751546],"genre_scores_gemma":[0.68396616,0.2955903,0.011982698,0.0036931299,0.00082104106,0.0018598216,0.0009830564,0.000156774,0.00094709976],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.9099742,0.060754966,0.016754808,0.0071788942,0.0038937726,0.0014433432],"domain_scores_gemma":[0.82910055,0.14172909,0.018009068,0.0072104074,0.0032447455,0.0007061595],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.06438215,0.0029692734,0.020009711,0.0031515665,0.00094382826,0.0051953755,0.0029283199,0.0052147997,0.006395953],"category_scores_gemma":[0.1427238,0.0022689006,0.031833272,0.0052960166,0.0018436866,0.0037500404,0.0022060396,0.0045222216,0.00048316052],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.012631709,0.00006265511,0.011385905,0.0976196,0.8667419,0.00012922657,0.00014031303,0.00030195317,0.00015208963,0.00056054286,0.00071387744,0.009560149],"study_design_scores_gemma":[0.004748146,0.00030400613,0.0059006875,0.0077624535,0.9785902,0.000074494215,0.00007103652,0.00038116676,0.00007719602,0.00094664074,0.001108927,0.000035144705],"about_ca_topic_score_codex":0.002967107,"about_ca_topic_score_gemma":0.0057976977,"teacher_disagreement_score":0.99703074,"about_ca_system_score_codex":0.002075331,"about_ca_system_score_gemma":0.0020548217,"threshold_uncertainty_score":0.34048945},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad","metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4377010956","doi":"10.1093/fampra/cmad052","title":"The quality of paediatric asthma guidelines: evidence underpinning diagnostic test recommendations from a meta-epidemiological study","year":2023,"lang":"en","type":"article","venue":"Family Practice","topic":"Asthma and respiratory diseases","field":"Medicine","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hospital for Sick Children; University of Toronto","funders":"U.S. Department of Health and Human Services; National Institutes of Health; Canadian Institutes of Health Research; National Institute for Health and Care Research; Canadian Thoracic Society; National Institute for Health and Care Excellence","keywords":"Medicine; Asthma; Spirometry; Epidemiology; Family medicine; Test (biology); Diagnostic test; Pediatrics; MEDLINE; Primary care; Intensive care medicine; Pathology","score_opus":0.41210880923528265,"score_gpt":0.4964183921697496,"score_spread":0.08430958293446694,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4377010956","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015372933,0.97251713,0.0045521483,0.0031517602,0.00041174385,0.0009671359,0.0017584934,0.000047779362,0.0012209531],"genre_scores_gemma":[0.6456736,0.32883728,0.014907386,0.0042645917,0.00065893016,0.0029092613,0.0024095979,0.000087355525,0.00025198382],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.77036715,0.14781995,0.05167039,0.010810576,0.017302176,0.0020297524],"domain_scores_gemma":[0.37370515,0.5303789,0.066870466,0.010627828,0.017194953,0.0012227212],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.15812522,0.0025587901,0.012064669,0.010856081,0.0010112259,0.008040133,0.004364097,0.004993646,0.0034787797],"category_scores_gemma":[0.47788,0.002267059,0.033550087,0.012917451,0.0028042358,0.0057236217,0.0036702328,0.0041466514,0.00029323384],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0025196362,0.00006426034,0.052766293,0.3231842,0.5769976,0.00023563887,0.0008219713,0.0016020916,0.00012678839,0.0010557873,0.002044254,0.03858149],"study_design_scores_gemma":[0.0015580946,0.00034249554,0.034989774,0.14063177,0.8131594,0.00029649594,0.00035351785,0.0010605237,0.00026009732,0.0028038814,0.0044566463,0.000087211716],"about_ca_topic_score_codex":0.013346776,"about_ca_topic_score_gemma":0.015390922,"teacher_disagreement_score":0.99744123,"about_ca_system_score_codex":0.006077747,"about_ca_system_score_gemma":0.007092623,"threshold_uncertainty_score":0.8362561},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"evaluation","study_design":"systematic_review","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4378903516","doi":"10.3389/ijph.2023.1605763","title":"Certainty of the Global Burden of Disease 2019 Modelled Prevalence Estimates for Musculoskeletal Conditions: A Meta-Epidemiological Study","year":2023,"lang":"en","type":"review","venue":"International Journal of Public Health","topic":"Musculoskeletal pain and rehabilitation","field":"Medicine","cited_by":31,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"European Cooperation in Science and Technology; Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung; European Centre for Chiropractic Research Excellence; National Science Foundation","keywords":"Epidemiology; Medicine; Representativeness heuristic; Disease burden; Public health; Meta-analysis; Osteoarthritis; Physical therapy; Environmental health; Demography; Population; Statistics; Alternative medicine; Internal medicine; Pathology","score_opus":0.19646743052588023,"score_gpt":0.49710522179827116,"score_spread":0.30063779127239093,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4378903516","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004034378,0.97725165,0.012069206,0.0019447174,0.00056537264,0.0010683871,0.0016688774,0.00007086146,0.0013266437],"genre_scores_gemma":[0.3085766,0.653791,0.025733326,0.0031121431,0.000970079,0.003874867,0.003290136,0.00012937532,0.0005224536],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.85685956,0.08949765,0.032882124,0.0056987386,0.014078287,0.0009836612],"domain_scores_gemma":[0.65736735,0.29134962,0.030098578,0.009221018,0.011098813,0.0008646266],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.1693926,0.003008529,0.015252812,0.011927339,0.00063334085,0.00836895,0.0037159987,0.0036956486,0.004043994],"category_scores_gemma":[0.347253,0.002772998,0.032953665,0.0067822705,0.0015387008,0.004488376,0.004213405,0.0040184245,0.0004233725],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018662752,0.00004541045,0.012458713,0.41637865,0.46631554,0.00019233991,0.00041030752,0.004610023,0.00023672737,0.0032059085,0.0022166972,0.0920635],"study_design_scores_gemma":[0.0013182112,0.00052366476,0.011879718,0.2279425,0.71746606,0.00075615523,0.00029070608,0.0028828497,0.0006651285,0.011231824,0.024852488,0.00019072743],"about_ca_topic_score_codex":0.0041863727,"about_ca_topic_score_gemma":0.0043123243,"teacher_disagreement_score":0.99699146,"about_ca_system_score_codex":0.005181278,"about_ca_system_score_gemma":0.0066392133,"threshold_uncertainty_score":0.8958444},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4379135156","doi":"10.1016/j.mayocpiqo.2023.04.010","title":"Inadequate Reporting of Cointerventions, Other Methodological Factors, and Treatment Estimates in Cardiovascular Trials: A Meta-Epidemiological Study","year":2023,"lang":"en","type":"article","venue":"Mayo Clinic Proceedings Innovations Quality & Outcomes","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; St. Michael's Hospital","funders":"Arts and Humanities Research Council; University of California, San Francisco; University of Bern; University of Toronto; Inselspital, Universitätsspital Bern; Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung; National Science Foundation","keywords":"Blinding; Medicine; Meta-analysis; Publication bias; Psychological intervention; Clinical trial; Odds ratio; MEDLINE; Confidence interval; Relative risk; Physical therapy; Internal medicine","score_opus":0.9823887722089711,"score_gpt":0.7142536049889211,"score_spread":0.26813516722005004,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4379135156","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.011878993,0.96866745,0.014050637,0.0013281718,0.0009907848,0.001570396,0.00074184674,0.00009669652,0.0006750238],"genre_scores_gemma":[0.6566767,0.2760178,0.045389272,0.0051116184,0.0026218025,0.010778326,0.0023403496,0.00031926448,0.0007448529],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.41758454,0.38532194,0.14894286,0.017215705,0.028688092,0.0022467808],"domain_scores_gemma":[0.26451632,0.5788407,0.110947505,0.03001425,0.014833561,0.0008476362],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.3791918,0.0054949545,0.022186328,0.013356335,0.0016057411,0.012312471,0.004421219,0.0063828933,0.0031413557],"category_scores_gemma":[0.5903142,0.0038382078,0.056641012,0.0149877835,0.0038007281,0.009537014,0.0055533685,0.004323559,0.00043725336],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004311949,0.000059222646,0.031520184,0.13932931,0.80524915,0.0001586586,0.0003004879,0.00086848123,0.00021187784,0.00077160867,0.00079786475,0.016421158],"study_design_scores_gemma":[0.0032658442,0.00074173306,0.020827973,0.055238366,0.9093675,0.0005021474,0.00018394724,0.0020174377,0.0006620203,0.0033861012,0.0036313864,0.00017553908],"about_ca_topic_score_codex":0.002448054,"about_ca_topic_score_gemma":0.0041812216,"teacher_disagreement_score":0.99450505,"about_ca_system_score_codex":0.004941489,"about_ca_system_score_gemma":0.00492107,"threshold_uncertainty_score":0.76556706},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4380149107","doi":"10.1016/j.jval.2023.03.1885","title":"PCR107 A Systematic Review and Quality Assessment of Cardiovascular Disease-Specific Patient Reported Outcome Measures Part I: Instrument Development and Content Validity","year":2023,"lang":"en","type":"review","venue":"Value in Health","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Content validity; Medicine; Patient-reported outcome; Quality of life (healthcare); Health related quality of life; Disease; Content (measure theory); Physical therapy; Intensive care medicine; Psychometrics; Clinical psychology; Pathology; Nursing","score_opus":0.9256591917202315,"score_gpt":0.49881490904595216,"score_spread":0.4268442826742793,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4380149107","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.001768558,0.99208504,0.0008404532,0.0005168003,0.00030327064,0.0020793008,0.0017490158,0.000024989233,0.0006324586],"genre_scores_gemma":[0.045119826,0.93924916,0.006617227,0.0015244709,0.00020514731,0.0054011196,0.0013898944,0.00003624589,0.0004569441],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9593253,0.015051093,0.016018055,0.0021457393,0.006995462,0.0004643953],"domain_scores_gemma":[0.9430874,0.03491357,0.014346153,0.001555878,0.0057114568,0.00038550442],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.028394805,0.0018606466,0.0109047135,0.0100275725,0.00093436637,0.0040264614,0.0021782937,0.0020575204,0.006441226],"category_scores_gemma":[0.13141318,0.0012132522,0.014440909,0.0108118225,0.0016638055,0.0027863637,0.0030683824,0.0020185048,0.00056550116],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00039219565,0.000023013145,0.0021172003,0.9056973,0.036300488,0.000048349102,0.00018207075,0.0001110651,0.0001638325,0.00053556607,0.0025207824,0.051908184],"study_design_scores_gemma":[0.0014147824,0.0002855374,0.013121504,0.7525245,0.20429616,0.00029171247,0.00020652456,0.00022091527,0.00029687962,0.0009144954,0.026328655,0.00009832952],"about_ca_topic_score_codex":0.010047093,"about_ca_topic_score_gemma":0.02572284,"teacher_disagreement_score":0.028394805,"about_ca_system_score_codex":0.005728427,"about_ca_system_score_gemma":0.015993817,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"opus","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4381434989","doi":"10.3390/curroncol30060443","title":"Characteristics of Phase IV Clinical Trials in Oncology: An Analysis Using the ClinicalTrials.gov Registry Data","year":2023,"lang":"en","type":"article","venue":"Current Oncology","topic":"Statistical Methods in Clinical Trials","field":"Mathematics","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Medicine; Clinical trial; Adverse effect; Psychological intervention; Internal medicine; Clinical Oncology; Sample size determination; Oncology; Cancer; Intensive care medicine","score_opus":0.9697628496947025,"score_gpt":0.7983529685784945,"score_spread":0.17140988111620803,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4381434989","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.34819633,0.2109808,0.07815821,0.005076692,0.0011126512,0.0733498,0.2594657,0.0010675504,0.022592213],"genre_scores_gemma":[0.7168412,0.021930411,0.05863027,0.0032279855,0.00044261347,0.12034496,0.07598297,0.0008104404,0.0017890495],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.6085949,0.20611095,0.12537783,0.013670379,0.043561187,0.0026847776],"domain_scores_gemma":[0.30397618,0.50566566,0.1458199,0.023779899,0.01839318,0.002365247],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.24595156,0.0011337552,0.0057966895,0.015383373,0.0008632442,0.004159405,0.001974322,0.00166156,0.0050410856],"category_scores_gemma":[0.44361567,0.0010882834,0.0076132943,0.0325989,0.0013919174,0.0044493363,0.0027949973,0.003731651,0.0008447241],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.049168877,0.00108063,0.4968349,0.082229584,0.064471826,0.0010942909,0.0018564113,0.011899067,0.0019088985,0.013116015,0.042038504,0.234301],"study_design_scores_gemma":[0.017352773,0.010066589,0.56632894,0.03146166,0.05512635,0.0049113114,0.0016526651,0.016585656,0.0039133807,0.024629429,0.26729137,0.0006798536],"about_ca_topic_score_codex":0.0010044135,"about_ca_topic_score_gemma":0.0018174207,"teacher_disagreement_score":0.9942033,"about_ca_system_score_codex":0.0026971446,"about_ca_system_score_gemma":0.008561038,"threshold_uncertainty_score":0.92987597},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4382395105","doi":"10.1177/14733250231185961","title":"Arts-based research with immigrant and racialized older adults: A scoping review","year":2023,"lang":"en","type":"review","venue":"Qualitative Social Work","topic":"Participatory Visual Research Methods","field":"Social Sciences","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"Immigration; Sociology; The arts; Gender studies; Gerontology; Psychology; Political science; Medicine; Visual arts; Art","score_opus":0.9258133317648053,"score_gpt":0.7970408334346758,"score_spread":0.1287724983301295,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4382395105","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00068971986,0.99517125,0.0005880582,0.0006948305,0.00039041543,0.000499514,0.00009745332,0.000010261537,0.001858447],"genre_scores_gemma":[0.0034464411,0.9936615,0.0013561051,0.0003574543,0.00008499025,0.0007565572,0.000099044555,0.000005316756,0.00023249017],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.98968595,0.004296087,0.0032631317,0.000545715,0.0019001967,0.00030897037],"domain_scores_gemma":[0.9533447,0.036750972,0.0030714802,0.00087251264,0.005521962,0.00043846967],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.017199256,0.0014435517,0.003485464,0.022129875,0.0026097007,0.0057174135,0.0018084885,0.003191066,0.0033266624],"category_scores_gemma":[0.070231214,0.0011825394,0.003545459,0.019413015,0.0015543033,0.0046921438,0.0036430173,0.0019542333,0.00081557967],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006832865,0.00006547182,0.0009741132,0.586418,0.0009951772,0.00034099922,0.0042466563,0.00021977545,0.00044444687,0.0044656633,0.009971423,0.39178994],"study_design_scores_gemma":[0.000012391487,0.000055801826,0.0012036356,0.91015553,0.001657188,0.00029035614,0.002199437,0.00006258587,0.00015900738,0.0010027332,0.083181515,0.00001979312],"about_ca_topic_score_codex":0.009587159,"about_ca_topic_score_gemma":0.025610214,"teacher_disagreement_score":0.022129875,"about_ca_system_score_codex":0.0041184453,"about_ca_system_score_gemma":0.024697045,"threshold_uncertainty_score":0.09095949},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4383340853","doi":"10.1016/j.jclinepi.2023.06.022","title":"Noninferiority margins exceed superiority effect estimates for mortality in cardiovascular trials in high-impact journals","year":2023,"lang":"en","type":"review","venue":"Journal of Clinical Epidemiology","topic":"Statistical Methods in Clinical Trials","field":"Mathematics","cited_by":8,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Impact; McMaster University","funders":"","keywords":"Interquartile range; Medicine; Clinical trial; MEDLINE; Sample size determination; Guideline; Relative risk; Emergency medicine; Confidence interval; Internal medicine; Statistics","score_opus":0.9328234368721308,"score_gpt":0.7624042600171351,"score_spread":0.17041917685499564,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4383340853","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0050900383,0.97047865,0.013323594,0.0050903508,0.0015551589,0.00035477086,0.00042370567,0.00009281732,0.003590889],"genre_scores_gemma":[0.44287097,0.4987005,0.033771176,0.0107352,0.009654931,0.0012648492,0.00131153,0.0002069462,0.0014839348],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.840839,0.110295326,0.026996806,0.005884681,0.014924751,0.0010594476],"domain_scores_gemma":[0.27450678,0.6793154,0.02988574,0.007074184,0.008263735,0.0009542937],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.15567182,0.0013992445,0.007322362,0.004897791,0.00048648883,0.005275303,0.0023869006,0.0031405273,0.009249546],"category_scores_gemma":[0.40156803,0.0011678966,0.0100284,0.004138358,0.002548573,0.008092084,0.00326311,0.00648315,0.0007069615],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.009918555,0.00015418159,0.011884157,0.25793016,0.1046564,0.00019497125,0.0007146574,0.0024162189,0.00076501,0.023933811,0.013571939,0.57386],"study_design_scores_gemma":[0.015882453,0.0035154766,0.10181166,0.1529927,0.27452245,0.0019348558,0.00115703,0.009089677,0.0027883986,0.31066668,0.12530276,0.00033592497],"about_ca_topic_score_codex":0.0008588048,"about_ca_topic_score_gemma":0.0016173759,"teacher_disagreement_score":0.84432817,"about_ca_system_score_codex":0.0015235484,"about_ca_system_score_gemma":0.0033592577,"threshold_uncertainty_score":0.8232811},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4383342582","doi":"10.1186/s12913-023-09726-8","title":"Overlapping research efforts in a global pandemic: a rapid systematic review of COVID-19-related individual participant data meta-analyses","year":2023,"lang":"en","type":"review","venue":"BMC Health Services Research","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; McGill University; Jewish General Hospital","funders":"Institute of Genetics; Horizon 2020 Framework Programme; Universitätsklinikum Heidelberg","keywords":"Pandemic; Coronavirus disease 2019 (COVID-19); Nursing research; Health informatics; Medicine; Meta-analysis; 2019-20 coronavirus outbreak; Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2); Public health; Systematic review; Health administration; MEDLINE; Virology; Nursing; Outbreak; Pathology; Disease; Biology","score_opus":0.9838711068747567,"score_gpt":0.7652878711180463,"score_spread":0.21858323575671035,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4383342582","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0027759357,0.9631142,0.010119946,0.0052211913,0.001299055,0.011997133,0.003917797,0.00027287356,0.0012818396],"genre_scores_gemma":[0.07526816,0.7552848,0.09123595,0.0100775575,0.0010416444,0.06161807,0.004404255,0.00039961434,0.00067007425],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.73703295,0.13358039,0.09606821,0.010588934,0.020618727,0.002110835],"domain_scores_gemma":[0.42066112,0.4628313,0.056971796,0.022223506,0.03544235,0.0018699063],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.22339757,0.0033442494,0.019681988,0.034670636,0.0019684648,0.010101223,0.00592637,0.005176419,0.0073391255],"category_scores_gemma":[0.4866036,0.0041310894,0.031013919,0.02398639,0.0026104283,0.01133613,0.009406527,0.0046541095,0.00092448207],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003384253,0.000019015059,0.0010413835,0.9169063,0.039862514,0.0001158952,0.00069566735,0.00030136475,0.00023217498,0.00091511576,0.00287928,0.03669289],"study_design_scores_gemma":[0.00058079505,0.00019023323,0.0024309666,0.8424991,0.124610685,0.00019222249,0.00061218883,0.00026280893,0.00027363194,0.0027223001,0.025527427,0.00009769719],"about_ca_topic_score_codex":0.0054658283,"about_ca_topic_score_gemma":0.017514924,"teacher_disagreement_score":0.980318,"about_ca_system_score_codex":0.011405766,"about_ca_system_score_gemma":0.033867534,"threshold_uncertainty_score":0.95768905},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W4385287363","doi":"10.1136/bmjopen-2023-073232","title":"Evaluating the impact of including non-randomised studies of interventions in meta-analysis of randomised controlled trials: a protocol for a meta-epidemiological study","year":2023,"lang":"en","type":"article","venue":"BMJ Open","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Joseph’s Healthcare Hamilton; McMaster University; Impact","funders":"National Science Fund for Distinguished Young Scholars; Canadian Institutes of Health Research; National Natural Science Foundation of China","keywords":"Medicine; Meta-analysis; Randomized controlled trial; Concordance; Systematic review; Psychological intervention; MEDLINE; Publication bias; Population; Epidemiology; Meta-regression; Internal medicine; Environmental health; Psychiatry","score_opus":0.9906337942003213,"score_gpt":0.8021268322848917,"score_spread":0.1885069619154296,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4385287363","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0008823258,0.004883545,0.04720296,0.0016308131,0.0027331915,0.9349477,0.0052027423,0.0013031776,0.001213456],"genre_scores_gemma":[0.0011602184,0.00061656826,0.054788247,0.00024493109,0.0000590753,0.9424647,0.0004072095,0.000068329515,0.00019076037],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.37808928,0.42593694,0.15188168,0.0180226,0.023157645,0.0029119481],"domain_scores_gemma":[0.42448208,0.3933179,0.062103335,0.06568301,0.04983744,0.0045762463],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.41711387,0.012377353,0.026931405,0.019064613,0.004109305,0.013189422,0.009493741,0.019052958,0.032251067],"category_scores_gemma":[0.64862096,0.0099073015,0.06013873,0.022556892,0.008054997,0.010276324,0.009762772,0.02286166,0.009475463],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.029330557,0.0009065094,0.0035481348,0.70924175,0.086865894,0.0011884389,0.0044821603,0.008561665,0.0027247802,0.019827694,0.055027835,0.07829453],"study_design_scores_gemma":[0.11917233,0.006104338,0.011657629,0.3747436,0.12615697,0.0014086213,0.001665449,0.036841992,0.008750486,0.10086563,0.21056217,0.0020708404],"about_ca_topic_score_codex":0.0021833875,"about_ca_topic_score_gemma":0.0031044066,"teacher_disagreement_score":0.9876226,"about_ca_system_score_codex":0.012472298,"about_ca_system_score_gemma":0.030458145,"threshold_uncertainty_score":0.7188024},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4385997249","doi":"10.1111/edt.12872","title":"Methodological assessment and overall confidence in the results of systematic reviews with meta‐analyses focusing on traumatic dental injuries: A cross‐sectional study","year":2023,"lang":"en","type":"review","venue":"Dental Traumatology","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Checklist; Medicine; Systematic review; Meta-analysis; Confidence interval; Protocol (science); MEDLINE; Psychology; Alternative medicine; Pathology; Internal medicine","score_opus":0.9687362985305835,"score_gpt":0.7013148111192546,"score_spread":0.2674214874113289,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4385997249","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.26250032,0.6303715,0.07800371,0.0048853196,0.0015539067,0.012698015,0.006126883,0.00034438173,0.0035158792],"genre_scores_gemma":[0.9193421,0.028034844,0.03514132,0.0015145327,0.00047886776,0.0131005775,0.0020232953,0.000105234525,0.00025929534],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.36466965,0.296939,0.2738246,0.02130529,0.041711703,0.0015498421],"domain_scores_gemma":[0.096612744,0.69630456,0.15582165,0.022058072,0.02822333,0.00097964],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.44847202,0.001981604,0.0079206005,0.016696043,0.0018796945,0.007011377,0.0030957167,0.0033578072,0.0027188098],"category_scores_gemma":[0.7449917,0.002752878,0.027980363,0.017162362,0.004138804,0.006407047,0.0058562616,0.0029519643,0.00031888302],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0042031845,0.00024628575,0.44169077,0.18373995,0.30753252,0.000465985,0.008298541,0.0017519391,0.00072770973,0.0019181559,0.0019786097,0.047446314],"study_design_scores_gemma":[0.0031036222,0.004367787,0.38674492,0.13655494,0.42768255,0.0037678778,0.005905016,0.007105444,0.0023352252,0.009521662,0.012159779,0.00075123203],"about_ca_topic_score_codex":0.0016981823,"about_ca_topic_score_gemma":0.0028322507,"teacher_disagreement_score":0.9920794,"about_ca_system_score_codex":0.0038352862,"about_ca_system_score_gemma":0.005739597,"threshold_uncertainty_score":0.6801322},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4386570142","doi":"10.1016/j.cpr.2023.102336","title":"Cognitive-behavioral treatment for gambling harm: Umbrella review and meta-analysis","year":2023,"lang":"en","type":"review","venue":"Clinical Psychology Review","topic":"Gambling Behavior and Treatments","field":"Psychology","cited_by":36,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University","funders":"National Institute on Alcohol Abuse and Alcoholism; National Institutes of Health","keywords":"Meta-analysis; PsycINFO; Psychology; Harm; Cognitive behavioral therapy; MEDLINE; Systematic review; Clinical psychology; Cognition; Randomized controlled trial; Psychiatry; Medicine; Social psychology","score_opus":0.945307245757294,"score_gpt":0.7391619732899881,"score_spread":0.20614527246730585,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4386570142","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0004092867,0.9990682,0.000092529626,0.000067197434,0.00005925376,0.00007258365,0.000083705556,0.0000055500254,0.00014169939],"genre_scores_gemma":[0.007043872,0.9914397,0.00058221474,0.0005031093,0.00008260215,0.00012047955,0.00010592599,0.0000048942647,0.00011724693],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9985245,0.00049141777,0.0004435624,0.00020690043,0.00026318047,0.00007057757],"domain_scores_gemma":[0.99703145,0.002127648,0.0005206309,0.00007196869,0.00018582639,0.0000625477],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0034185478,0.00164223,0.010089224,0.0031757185,0.000344111,0.0020818303,0.0017922225,0.0016710883,0.005625678],"category_scores_gemma":[0.009048245,0.0006052075,0.009135581,0.0034431561,0.0005163825,0.0011327794,0.0010509093,0.0018520241,0.0004284814],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0023148637,0.00017677735,0.00048498358,0.7727253,0.07110833,0.00007930159,0.0000865232,0.00030876184,0.0004649515,0.00034091677,0.003292984,0.14861628],"study_design_scores_gemma":[0.0047874376,0.0013572273,0.0070665977,0.35799503,0.5869234,0.0004407869,0.00018523473,0.0003720775,0.00044324237,0.0009948431,0.039330862,0.00010312875],"about_ca_topic_score_codex":0.0041704327,"about_ca_topic_score_gemma":0.013802256,"teacher_disagreement_score":0.010089224,"about_ca_system_score_codex":0.0012929585,"about_ca_system_score_gemma":0.0021889054,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4386864680","doi":"10.1186/s12885-023-11392-2","title":"Sleep disturbances and the risk of lung cancer: a meta-epidemiological study","year":2023,"lang":"en","type":"review","venue":"BMC Cancer","topic":"Sleep and related disorders","field":"Psychology","cited_by":23,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"China Academy of Chinese Medical Sciences; National Natural Science Foundation of China","keywords":"Publication bias; Medicine; Funnel plot; Meta-analysis; Chronotype; Internal medicine; Odds ratio; Cochrane Library; Lung cancer; Confidence interval; Oncology; Circadian rhythm","score_opus":0.15490475250364255,"score_gpt":0.44885807022148977,"score_spread":0.2939533177178472,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4386864680","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015537953,0.98100305,0.0016968542,0.0003088283,0.00028807909,0.0002767205,0.00059624744,0.00003545375,0.0002568195],"genre_scores_gemma":[0.38540298,0.60333014,0.006438129,0.0010511455,0.00060764665,0.0013627836,0.0013048232,0.000049848764,0.00045247338],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9901759,0.0052072294,0.0022609595,0.001162561,0.0008668792,0.00032657027],"domain_scores_gemma":[0.9809533,0.01373425,0.0026227455,0.0010400737,0.0013555969,0.0002941121],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.015111803,0.0032187812,0.013578063,0.007336507,0.0007740544,0.0038668865,0.0021723423,0.0025187656,0.002661267],"category_scores_gemma":[0.031007377,0.0018873528,0.051045537,0.0075822095,0.00063195283,0.0016497861,0.0017313582,0.0025452285,0.00025681857],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019996234,0.000058815036,0.0139566865,0.084705465,0.8919112,0.00017531986,0.000060551607,0.00075414707,0.00026378434,0.00011388975,0.00028474053,0.0057157646],"study_design_scores_gemma":[0.00032881857,0.00024794764,0.007920779,0.007051428,0.98274404,0.0001290566,0.00003422328,0.0004235919,0.00012154945,0.00022300256,0.00075281423,0.000022815237],"about_ca_topic_score_codex":0.0050012,"about_ca_topic_score_gemma":0.007941581,"teacher_disagreement_score":0.99678123,"about_ca_system_score_codex":0.0017331633,"about_ca_system_score_gemma":0.0024553067,"threshold_uncertainty_score":0.079919815},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4387222396","doi":"10.1007/s40279-023-01941-x","title":"Certainty of Evidence Assessment in Systematic Reviews Published by High-Impact Sports Science Journals: A Meta-epidemiological Study","year":2023,"lang":"en","type":"article","venue":"Sports Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":3,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact; University of Saskatchewan","funders":"","keywords":"Certainty; Systematic review; Grading (engineering); Meta-analysis; Publication bias; Quality of evidence; Medicine; Citation; Evidence-based medicine; Sports medicine; MEDLINE; Bibliometrics; Psychology; Alternative medicine; Computer science; Library science; Political science; Pathology; Mathematics; Engineering","score_opus":0.9041380651400333,"score_gpt":0.6472378102149855,"score_spread":0.25690025492504776,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4387222396","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.040349048,0.9382005,0.0152923325,0.0011222942,0.0008148116,0.0016214947,0.0016187273,0.0001370816,0.00084375724],"genre_scores_gemma":[0.85364,0.11862928,0.019104952,0.0022109165,0.00097328064,0.003037792,0.0018860559,0.00015796813,0.00035972946],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.51311594,0.28649786,0.14310528,0.027082074,0.027135538,0.00306335],"domain_scores_gemma":[0.20788993,0.6847789,0.06414094,0.026864307,0.014500353,0.0018256423],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.28884494,0.0061667296,0.03756827,0.023920422,0.002090979,0.013258383,0.0062117004,0.009286065,0.0041216663],"category_scores_gemma":[0.6354684,0.0072768633,0.0877481,0.016900321,0.0041078962,0.012286399,0.011003157,0.007926853,0.0004538727],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0042372956,0.000051887695,0.021316394,0.054808125,0.91325194,0.00012983638,0.00030705947,0.00043826678,0.00010113732,0.00025979776,0.00023305202,0.0048652813],"study_design_scores_gemma":[0.0025255845,0.00056494283,0.012331576,0.011174296,0.9689242,0.00037549727,0.0002046222,0.0007742796,0.00015639319,0.001984097,0.00084442215,0.00014001531],"about_ca_topic_score_codex":0.0055780048,"about_ca_topic_score_gemma":0.006120303,"teacher_disagreement_score":0.99383324,"about_ca_system_score_codex":0.0053193304,"about_ca_system_score_gemma":0.005125978,"threshold_uncertainty_score":0.8769808},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"evaluation","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4387381774","doi":"10.48083/wura1857","title":"Methodological Quality of Systematic Reviews for Questions of Therapy and Prevention Published in the Urological Literature (2016–2021) Fails to Improve","year":2023,"lang":"en","type":"article","venue":"Société Internationale d’Urologie Journal","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Systematic review; Medicine; Weakness; Meta-analysis; Protocol (science); MEDLINE; Confidence interval; Medical physics; Alternative medicine; Surgery; Internal medicine; Pathology","score_opus":0.9138691369725518,"score_gpt":0.6449601844319653,"score_spread":0.26890895254058644,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4387381774","genre_codex":"review","genre_gemma":"empirical","domain_codex":"reporting","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.043712858,0.82567865,0.060936533,0.028956844,0.005805013,0.010579803,0.016253028,0.00054042833,0.007536819],"genre_scores_gemma":[0.7376438,0.124010265,0.07695157,0.017269615,0.0027999692,0.030090487,0.009597933,0.00053303264,0.0011034154],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.21463767,0.32141602,0.356437,0.030880323,0.073357135,0.0032719134],"domain_scores_gemma":[0.090871505,0.6585327,0.16350041,0.03417147,0.051317368,0.0016065366],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.5493863,0.0035196068,0.0126224235,0.033710435,0.003487301,0.016634732,0.00596958,0.004853632,0.005968957],"category_scores_gemma":[0.81796956,0.0043928265,0.02301178,0.031451963,0.007202583,0.01277634,0.010679466,0.0036980195,0.0011274709],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014394436,0.00008495123,0.055666916,0.73790103,0.116062626,0.00039672302,0.00690641,0.0014408135,0.0007822639,0.004504407,0.006233992,0.06858039],"study_design_scores_gemma":[0.0021460138,0.0011080195,0.09943025,0.6349268,0.16958386,0.0014692143,0.0057761515,0.003626842,0.0021823775,0.0163051,0.06272386,0.00072149077],"about_ca_topic_score_codex":0.009151488,"about_ca_topic_score_gemma":0.020864397,"teacher_disagreement_score":0.9873776,"about_ca_system_score_codex":0.012380379,"about_ca_system_score_gemma":0.024135686,"threshold_uncertainty_score":0.55568695},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W4387695005","doi":"10.3346/jkms.2023.38.e333","title":"Causes for Retraction in the Biomedical Literature: A Systematic Review of Studies of Retraction Notices","year":2023,"lang":"en","type":"review","venue":"Journal of Korean Medical Science","topic":"Academic integrity and plagiarism","field":"Social Sciences","cited_by":47,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital; University of Ottawa","funders":"H. Lundbeck A/S","keywords":"Misconduct; Subspecialty; Medicine; Meta-analysis; MEDLINE; Systematic review; Confidence interval; Scientific misconduct; Publication bias; Family medicine; Internal medicine; Alternative medicine; Pathology; Political science","score_opus":0.1576615077346172,"score_gpt":0.4965828687945747,"score_spread":0.3389213610599575,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4387695005","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"evaluation","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"evaluation","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0037009616,0.99447495,0.0002402892,0.00046263935,0.0001522479,0.00037435317,0.00037496522,0.000009072372,0.00021052589],"genre_scores_gemma":[0.06625231,0.9288743,0.001710923,0.00091659714,0.00022867619,0.0012705926,0.0006160724,0.000015861275,0.00011459907],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.91117454,0.035587683,0.03513481,0.0041290675,0.012929242,0.001044685],"domain_scores_gemma":[0.6571104,0.26367167,0.05895739,0.004328691,0.014722438,0.0012095545],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.05767701,0.0018909727,0.007193488,0.025243575,0.0014450724,0.004009945,0.0023572978,0.0022275385,0.003155138],"category_scores_gemma":[0.22322404,0.0015002784,0.010730505,0.024704989,0.0025331925,0.0048286263,0.0030200477,0.0017923462,0.00025872054],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00025947284,0.000024708039,0.007584814,0.93214285,0.0154335145,0.0001939594,0.0009482373,0.000056998615,0.000109030494,0.00021959246,0.0013839668,0.041642837],"study_design_scores_gemma":[0.00023871867,0.00023409337,0.020314476,0.89106923,0.07062565,0.0008983011,0.0017891765,0.000084105326,0.00028245992,0.0005009179,0.013883861,0.00007896746],"about_ca_topic_score_codex":0.0033774984,"about_ca_topic_score_gemma":0.010302746,"teacher_disagreement_score":0.99777246,"about_ca_system_score_codex":0.0045544617,"about_ca_system_score_gemma":0.01566845,"threshold_uncertainty_score":0.30502886},"labels":[{"model":"gemma","categories":["metaresearch","research_integrity"],"domain":"evaluation","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad","research_integrity"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4388033764","doi":"10.1111/iwj.14351","title":"Deficiencies in reporting inclusion/exclusion criteria and characteristics of patients in randomized controlled trials of therapeutic interventions in pressure injuries: a systematic methodological review","year":2023,"lang":"en","type":"review","venue":"International Wound Journal","topic":"Pressure Ulcer Prevention and Management","field":"Health Professions","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Medicine; Inclusion and exclusion criteria; Psychological intervention; Generalizability theory; Randomized controlled trial; Inclusion (mineral); Quartile; Systematic review; Adverse effect; Randomization; Sample size determination; MEDLINE; Alternative medicine; Intensive care medicine; Internal medicine; Confidence interval; Nursing; Pathology; Statistics","score_opus":0.43995663140885366,"score_gpt":0.5881098872714793,"score_spread":0.14815325586262562,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4388033764","genre_codex":"review","genre_gemma":"review","domain_codex":"reporting","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"reporting","domain_consensus":"reporting","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.012066504,0.8041192,0.041417483,0.017032072,0.0062328256,0.1057439,0.009836313,0.00033342783,0.003218236],"genre_scores_gemma":[0.19827464,0.26488054,0.15106556,0.018835116,0.0018941149,0.3580203,0.0057681417,0.00027971988,0.0009818642],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.18509115,0.3424365,0.41260728,0.012382656,0.045606196,0.0018762093],"domain_scores_gemma":[0.15994127,0.617759,0.13322411,0.035251483,0.051972628,0.0018515065],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.52527547,0.0029423693,0.018788135,0.016842032,0.0028638327,0.009032111,0.0060826195,0.007600807,0.0033902344],"category_scores_gemma":[0.7844401,0.0041619833,0.018821722,0.022156514,0.0057787285,0.009469338,0.005446887,0.00454329,0.00081918703],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014509491,0.000094777315,0.0072566555,0.8941602,0.030726628,0.00017538051,0.002300834,0.0005139047,0.00032651835,0.003294159,0.0060000503,0.05369991],"study_design_scores_gemma":[0.0036452897,0.0006276158,0.0077557247,0.8752599,0.07119502,0.00031960208,0.00109753,0.0011990509,0.0011412497,0.006888976,0.030607117,0.0002629321],"about_ca_topic_score_codex":0.0058811996,"about_ca_topic_score_gemma":0.010031555,"teacher_disagreement_score":0.47472453,"about_ca_system_score_codex":0.0148504805,"about_ca_system_score_gemma":0.046470135,"threshold_uncertainty_score":0.5854199},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4388422401","doi":"10.1186/s12969-023-00917-5","title":"Reporting of determinants of health inequities and participant characteristics in randomized controlled trials of juvenile idiopathic arthritis in Canada: a scoping review","year":2023,"lang":"en","type":"review","venue":"Pediatric Rheumatology","topic":"Autoimmune and Inflammatory Disorders Research","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Arthritis Research Centre of Canada; Centre for Advancing Health Outcomes; University of British Columbia Hospital; University of Ottawa; Research Canada; University of British Columbia","funders":"Canadian Institutes of Health Research","keywords":"Medicine; Socioeconomic status; Randomized controlled trial; Arthritis; Psychological intervention; Health equity; Gerontology; Health care; Family medicine; Physical therapy; Population; Environmental health; Public health; Internal medicine; Psychiatry; Nursing","score_opus":0.16991437890352346,"score_gpt":0.4286133349586202,"score_spread":0.25869895605509674,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4388422401","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.002508148,0.98046297,0.0017232825,0.0019683123,0.00052424043,0.008086475,0.0032090659,0.000048029877,0.0014694604],"genre_scores_gemma":[0.10780044,0.8262207,0.014419777,0.004131353,0.00041686604,0.04290018,0.003425319,0.00008723932,0.0005981738],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.6539363,0.13806912,0.15976048,0.009648117,0.03560662,0.0029793605],"domain_scores_gemma":[0.34306252,0.5207745,0.07570031,0.009513443,0.049077794,0.0018715473],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.2207693,0.0024854986,0.017323343,0.026633475,0.0032566013,0.01157643,0.0055138585,0.0055240258,0.0046195025],"category_scores_gemma":[0.59625304,0.0025929182,0.014069328,0.034103252,0.004505543,0.0053194184,0.004287756,0.003472058,0.00035200443],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000539471,0.000023001638,0.0023696646,0.9442651,0.019492723,0.00008445718,0.00067217177,0.00023653373,0.00008155092,0.00089314504,0.0017576563,0.029584467],"study_design_scores_gemma":[0.000724641,0.00013532351,0.0042521073,0.9219714,0.062035277,0.00008841769,0.00036156355,0.00021213882,0.00018941991,0.00074947294,0.0092233885,0.000056894787],"about_ca_topic_score_codex":0.14504091,"about_ca_topic_score_gemma":0.2421657,"teacher_disagreement_score":0.9588437,"about_ca_system_score_codex":0.04115632,"about_ca_system_score_gemma":0.12344238,"threshold_uncertainty_score":0.96093017},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":true,"about_ca_topic":true,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":true,"confidence":"high"}],"label_agreement":"split"},{"id":"W4389131041","doi":"10.1057/s41599-023-02403-5","title":"Evaluations of training programs to improve capacity in K*: a systematic scoping review of methods applied and outcomes assessed","year":2023,"lang":"en","type":"article","venue":"Humanities and Social Sciences Communications","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"Bill and Melinda Gates Foundation","keywords":"Inclusion (mineral); Categorization; Sample (material); Medical education; Curriculum; Sample size determination; Psychology; Systematic review; Computer science; MEDLINE; Knowledge management; Medicine; Political science; Artificial intelligence; Statistics; Pedagogy","score_opus":0.9420600838320179,"score_gpt":0.7521992021815397,"score_spread":0.1898608816504782,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4389131041","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0065057976,0.9796173,0.003103596,0.0016472245,0.0005925689,0.0047650686,0.00092776417,0.00004370144,0.0027969428],"genre_scores_gemma":[0.07803144,0.89688647,0.0109338695,0.0010633499,0.00022881398,0.011705348,0.0007434864,0.00004776209,0.0003595008],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.86572707,0.06902649,0.03870891,0.0041617537,0.020647945,0.0017278595],"domain_scores_gemma":[0.6463897,0.2749715,0.03495049,0.0059581436,0.036486533,0.0012435915],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.13496183,0.0021417253,0.007217721,0.029411728,0.0024513938,0.008232765,0.0032945795,0.0025145605,0.0040255664],"category_scores_gemma":[0.3467923,0.0019574594,0.007972811,0.028930591,0.0034940357,0.0073109027,0.0048869057,0.002514002,0.0005008782],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00034927228,0.00009905762,0.0025164804,0.7849718,0.0042271176,0.00009010187,0.0022268388,0.0004380657,0.00018452121,0.0017192636,0.0020982313,0.20107937],"study_design_scores_gemma":[0.00006264149,0.0001440656,0.0022210148,0.9798962,0.00607851,0.0000641533,0.0011036247,0.00011722486,0.00026236658,0.00047547158,0.00954392,0.000030677405],"about_ca_topic_score_codex":0.011579104,"about_ca_topic_score_gemma":0.022909349,"teacher_disagreement_score":0.86503816,"about_ca_system_score_codex":0.014940228,"about_ca_system_score_gemma":0.049856618,"threshold_uncertainty_score":0.7137549},"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":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4389687393","doi":"10.23736/s1973-9087.23.08310-7","title":"The influence of bias in randomized controlled trials on rehabilitation intervention effect estimates: what we have learned from meta-epidemiological studies","year":2023,"lang":"en","type":"article","venue":"European Journal of Physical and Rehabilitation Medicine","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"Ministero della Salute","keywords":"Blinding; Rehabilitation; Randomized controlled trial; Medicine; Meta-analysis; Publication bias; Physical therapy; Physical medicine and rehabilitation; Epidemiology; Intervention (counseling); Reporting bias; Selection bias; MEDLINE; Clinical psychology; Psychiatry; Surgery; Pathology","score_opus":0.48881372665186773,"score_gpt":0.5026068912095041,"score_spread":0.01379316455763635,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4389687393","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0013464926,0.9661383,0.015191668,0.013603154,0.0023669212,0.00035735883,0.00026182484,0.000074180294,0.0006600926],"genre_scores_gemma":[0.12867478,0.7631916,0.05415621,0.038517877,0.010537201,0.0035098875,0.00068108324,0.00034802253,0.00038331363],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.4396459,0.36619708,0.13187957,0.023365352,0.03622218,0.0026899665],"domain_scores_gemma":[0.05708273,0.8809866,0.029904675,0.016700879,0.014757334,0.00056770333],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.4975832,0.0043045618,0.023345536,0.016711723,0.0021751048,0.017718904,0.0069426317,0.011261236,0.0045587444],"category_scores_gemma":[0.79682046,0.004190436,0.022453295,0.01520794,0.008804685,0.026259039,0.008790448,0.013006049,0.000825276],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0018440018,0.00009315094,0.012682342,0.55251366,0.12616828,0.0004730189,0.0035475066,0.0019932871,0.00039922964,0.017113954,0.010099961,0.27307165],"study_design_scores_gemma":[0.0011299222,0.0006974183,0.008539202,0.7005482,0.1506387,0.0010477845,0.0014555969,0.0021837214,0.0012228354,0.07410577,0.057921138,0.0005097453],"about_ca_topic_score_codex":0.005435969,"about_ca_topic_score_gemma":0.0066984487,"teacher_disagreement_score":0.97665447,"about_ca_system_score_codex":0.008190667,"about_ca_system_score_gemma":0.013719797,"threshold_uncertainty_score":0.6195694},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W4390574318","doi":"10.1007/s00125-023-06052-w","title":"Twenty years of participation of racialised groups in type 2 diabetes randomised clinical trials: a meta-epidemiological review","year":2024,"lang":"en","type":"review","venue":"Diabetologia","topic":"Ethics in Clinical Research","field":"Medicine","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact; McMaster University Medical Centre","funders":"McMaster University","keywords":"Medicine; Type 2 diabetes; Randomized controlled trial; Ethnic group; MEDLINE; Diabetes mellitus; Meta-analysis; Clinical trial; Epidemiology; Demography; Research design; Gerontology; Family medicine; Internal medicine","score_opus":0.8713873905798888,"score_gpt":0.7038143442371966,"score_spread":0.1675730463426922,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4390574318","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0018427863,0.99664813,0.00045468926,0.0002687295,0.00028295844,0.0001919095,0.000117510004,0.000012725026,0.0001805223],"genre_scores_gemma":[0.109617166,0.88194907,0.0026584414,0.0021811496,0.00068339135,0.0021126973,0.00045694792,0.00004394704,0.00029718253],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.908772,0.057910576,0.020294553,0.004876035,0.006927613,0.0012192128],"domain_scores_gemma":[0.7660628,0.18630663,0.03433816,0.005470883,0.0065940805,0.0012274912],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.08854829,0.0025951713,0.020718018,0.007833784,0.0009867958,0.00669922,0.002934112,0.004728221,0.0054660775],"category_scores_gemma":[0.17902967,0.0023277858,0.047962554,0.007834292,0.0016139915,0.0057311594,0.0033725344,0.0040994016,0.00040347344],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0058024973,0.00006887814,0.0048224344,0.38517973,0.5716394,0.000121085424,0.00021901903,0.0005935311,0.00011931224,0.00046555945,0.00088460423,0.03008401],"study_design_scores_gemma":[0.0024560094,0.00070722785,0.0065648044,0.10598738,0.8775528,0.00021134391,0.0001321205,0.00027163816,0.00013320995,0.0008944256,0.005037658,0.0000513574],"about_ca_topic_score_codex":0.0035562566,"about_ca_topic_score_gemma":0.0058460543,"teacher_disagreement_score":0.97928196,"about_ca_system_score_codex":0.0050264173,"about_ca_system_score_gemma":0.005812413,"threshold_uncertainty_score":0.46829373},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4390765507","doi":"10.1007/s11655-023-3617-0","title":"Factors and Their Impact on Treatment Effect of Acupuncture in Different Outcomes: A Meta-Regression of Acupuncture Randomized Controlled Trials","year":2024,"lang":"en","type":"review","venue":"Chinese Journal of Integrative Medicine","topic":"Acupuncture Treatment Research Studies","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; Impact","funders":"","keywords":"Acupuncture; Medicine; Randomized controlled trial; Meta-analysis; Confidence interval; Strictly standardized mean difference; Physical therapy; Meta-regression; MEDLINE; Cochrane Library; Internal medicine; Alternative medicine; Pathology","score_opus":0.1108893715685012,"score_gpt":0.4947666850293225,"score_spread":0.38387731346082127,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4390765507","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.012056513,0.9844742,0.0019186804,0.00022990735,0.0003221477,0.00033316683,0.00033447018,0.000037849954,0.0002930163],"genre_scores_gemma":[0.43332613,0.5545474,0.0070212786,0.0013757547,0.0006732862,0.0012133926,0.0010334331,0.00012005623,0.0006892077],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9650111,0.018810207,0.009352707,0.0034310843,0.0026827406,0.00071214576],"domain_scores_gemma":[0.91401327,0.074073255,0.0065915943,0.0025593885,0.002264315,0.0004982541],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.038781364,0.004138564,0.020136163,0.0040106284,0.00082343206,0.00467941,0.0029645928,0.0033525468,0.005007969],"category_scores_gemma":[0.086621016,0.0021337406,0.066799216,0.0059162634,0.0015023183,0.0027490621,0.0025427432,0.005171557,0.0004340109],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.005493706,0.000033249144,0.006030771,0.062388193,0.91756064,0.000061885316,0.000065629705,0.00032089325,0.00012597856,0.000118089534,0.00016823747,0.0076327063],"study_design_scores_gemma":[0.0012373503,0.0001734588,0.002804571,0.0032105187,0.9916951,0.00003579719,0.000029267794,0.00014329287,0.00005706818,0.00021348253,0.00038794652,0.000012190409],"about_ca_topic_score_codex":0.0059921816,"about_ca_topic_score_gemma":0.007821163,"teacher_disagreement_score":0.9958614,"about_ca_system_score_codex":0.0020153392,"about_ca_system_score_gemma":0.0032777213,"threshold_uncertainty_score":0.20509791},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_narrow","metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4391147924","doi":"10.2139/ssrn.4696882","title":"Hierarchy of Patient-Reported Outcome Measures in Randomized Controlled Trials for Chronic Non-Specific Low Back Pain: A Meta-Epidemiological Study – Study Protocol","year":2024,"lang":"en","type":"preprint","venue":"SSRN Electronic Journal","topic":"Musculoskeletal pain and rehabilitation","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"","keywords":"Epidemiology; Randomized controlled trial; Medicine; Protocol (science); Meta-analysis; Outcome (game theory); Physical therapy; Chronic pain; Hierarchy; Low back pain; Alternative medicine; Internal medicine; Pathology","score_opus":0.11050970728391607,"score_gpt":0.4062562167360266,"score_spread":0.2957465094521105,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4391147924","genre_codex":"protocol","genre_gemma":"protocol","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":"protocol","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.011218678,0.019293943,0.016062507,0.000602471,0.0010297749,0.93509954,0.013625321,0.00081651605,0.002251288],"genre_scores_gemma":[0.023798626,0.0021349888,0.012956651,0.00043266627,0.00011076535,0.95788896,0.0017355623,0.00010389022,0.00083791796],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.89261127,0.057195198,0.031802032,0.007914811,0.0073972903,0.0030794467],"domain_scores_gemma":[0.91957533,0.042668253,0.020025637,0.00916768,0.0067877555,0.0017752909],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.07974789,0.007954568,0.02432784,0.0074175056,0.0031884657,0.0064005386,0.0037010482,0.007001667,0.043112885],"category_scores_gemma":[0.12714563,0.0063902945,0.038106777,0.005746009,0.0045332643,0.0042468244,0.0051153377,0.009909093,0.0043949024],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.3405632,0.001591327,0.00625427,0.3862136,0.20599835,0.00046384113,0.0010599229,0.0028781195,0.0023564762,0.0060948823,0.013690769,0.03283527],"study_design_scores_gemma":[0.6732851,0.009576448,0.017272972,0.042667743,0.20336848,0.0003216508,0.00043093358,0.0046224385,0.0022101928,0.00863122,0.03702763,0.00058524834],"about_ca_topic_score_codex":0.0029327075,"about_ca_topic_score_gemma":0.006433535,"teacher_disagreement_score":0.9756722,"about_ca_system_score_codex":0.006216636,"about_ca_system_score_gemma":0.011388161,"threshold_uncertainty_score":0.4217522},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4391656148","doi":"10.1016/j.radonc.2024.110148","title":"Applying the fragility index to randomized controlled trials evaluating total neoadjuvant therapy for rectal cancer: A methodological survey","year":2024,"lang":"en","type":"review","venue":"Radiotherapy and Oncology","topic":"Colorectal Cancer Surgical Treatments","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Kingston Health Sciences Centre; Queen's University; McMaster University","funders":"","keywords":"Colorectal cancer; Medicine; Fragility; Index (typography); Randomized controlled trial; Neoadjuvant therapy; Oncology; Internal medicine; Cancer; Computer science; Breast cancer","score_opus":0.42799547503695806,"score_gpt":0.5724683855566545,"score_spread":0.1444729105196964,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4391656148","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.012406195,0.9040336,0.063645504,0.0023225427,0.0009784552,0.0122550465,0.001839818,0.00013404485,0.0023848077],"genre_scores_gemma":[0.37124693,0.42094624,0.14580324,0.008002581,0.0010971982,0.04773792,0.003568812,0.00025375868,0.0013432875],"study_design_codex":"meta_analysis","study_design_gemma":"observational","domain_scores_codex":[0.5755789,0.25975576,0.12827621,0.00948883,0.025667675,0.0012326139],"domain_scores_gemma":[0.31049064,0.61357254,0.048115805,0.014214689,0.012878979,0.0007273699],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.28176463,0.0024721858,0.012034743,0.007328212,0.0008203775,0.0054075127,0.0034269649,0.0031535805,0.0044938675],"category_scores_gemma":[0.5921878,0.0015947577,0.026121523,0.008186768,0.0023021563,0.0034435142,0.0037771105,0.0042356965,0.0003869879],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008149158,0.000113535665,0.022853421,0.341795,0.461487,0.0002013544,0.00089395646,0.0026484213,0.00062202004,0.0055597215,0.0029687597,0.15270767],"study_design_scores_gemma":[0.010935176,0.003642501,0.027229525,0.16859812,0.71773046,0.0008224907,0.00077016605,0.0067911427,0.0010333863,0.028125951,0.03401094,0.0003101504],"about_ca_topic_score_codex":0.0017422994,"about_ca_topic_score_gemma":0.0037789014,"teacher_disagreement_score":0.7182354,"about_ca_system_score_codex":0.005252432,"about_ca_system_score_gemma":0.005062928,"threshold_uncertainty_score":0.8857121},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4392298597","doi":"10.1371/journal.pone.0298411","title":"Validity and reliability International Classification of Diseases-10 codes for all forms of injury: A systematic review","year":2024,"lang":"en","type":"review","venue":"PLoS ONE","topic":"Medical Coding and Health Information","field":"Health Professions","cited_by":24,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Michael's Hospital","funders":"Canadian Institutes of Health Research","keywords":"Medicine; Kappa; Cohen's kappa; Systematic review; MEDLINE; Validity; Poison control; Emergency medicine; Psychometrics; Clinical psychology; Statistics","score_opus":0.4717224928504905,"score_gpt":0.5194885194684242,"score_spread":0.04776602661793372,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4392298597","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0031900369,0.992622,0.00033645387,0.0003998874,0.00017250078,0.0010882304,0.0017200476,0.000011996432,0.00045877844],"genre_scores_gemma":[0.06034324,0.9301139,0.0025954973,0.00081677607,0.00019738036,0.0040518735,0.0016969658,0.000015020338,0.00016933534],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9624162,0.012238391,0.013993018,0.0025067197,0.008214395,0.0006313264],"domain_scores_gemma":[0.88646346,0.077633575,0.02243426,0.0015737562,0.01126851,0.0006264353],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.023765918,0.0017193665,0.009718998,0.018667793,0.000889267,0.0030606,0.0022599073,0.0017996861,0.0027175534],"category_scores_gemma":[0.13219835,0.0012617938,0.009423497,0.016175356,0.0015923892,0.0028888956,0.001971025,0.0013801911,0.0002518328],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021454619,0.000029933655,0.008899577,0.93204457,0.015865121,0.00007486523,0.00038418427,0.00013715865,0.000075074146,0.0002472621,0.0021644996,0.039863274],"study_design_scores_gemma":[0.00035050785,0.00014120524,0.028050024,0.88934577,0.069345534,0.00037655054,0.00048689728,0.00018599335,0.00013454072,0.00039296463,0.011122049,0.00006802015],"about_ca_topic_score_codex":0.017869847,"about_ca_topic_score_gemma":0.039403703,"teacher_disagreement_score":0.9762341,"about_ca_system_score_codex":0.005513038,"about_ca_system_score_gemma":0.018960906,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4392604238","doi":"10.1101/2024.03.06.24303892","title":"Publication bias in pharmacogenetics of statin-associated muscle symptoms, an umbrella review with a meta-epidemiological study","year":2024,"lang":"en","type":"review","venue":"medRxiv","topic":"Lipoproteins and Cardiovascular Health","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"Centre National de la Recherche Scientifique","keywords":"Funnel plot; SLCO1B1; Publication bias; Statin; Meta-analysis; Medicine; Odds ratio; Internal medicine; Pharmacogenetics; Pharmacology; Genotype; Genetics; Biology","score_opus":0.28214524288905907,"score_gpt":0.446024021149999,"score_spread":0.16387877826093994,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4392604238","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0031317226,0.9907031,0.0028920171,0.00087652006,0.00054286275,0.0005371992,0.00078309607,0.00003380203,0.0004996185],"genre_scores_gemma":[0.26363838,0.70556736,0.0129103875,0.0065139807,0.0020239176,0.0060910536,0.0020936343,0.00015073446,0.00101054],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.9216534,0.04039764,0.02465328,0.0056376834,0.006738787,0.0009191879],"domain_scores_gemma":[0.81492037,0.14436112,0.025841683,0.00695714,0.007332887,0.00058672804],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.07509449,0.003234221,0.014811011,0.012337285,0.001212254,0.0064318376,0.0024964735,0.0030810505,0.004838774],"category_scores_gemma":[0.17705043,0.001865583,0.030773658,0.012247705,0.0014056247,0.003387347,0.0029234914,0.0018964643,0.0004991324],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013616813,0.000029746532,0.010154209,0.27647513,0.6879085,0.00028869393,0.00017068797,0.00060780894,0.0002830372,0.00058718614,0.0016782759,0.02045506],"study_design_scores_gemma":[0.00066313724,0.00016255729,0.0076331836,0.052965034,0.93121374,0.00024120558,0.000081650585,0.00056043715,0.00022133421,0.0016254559,0.004588398,0.00004378605],"about_ca_topic_score_codex":0.0036626712,"about_ca_topic_score_gemma":0.0060367775,"teacher_disagreement_score":0.9967658,"about_ca_system_score_codex":0.0031596557,"about_ca_system_score_gemma":0.0045548445,"threshold_uncertainty_score":0.39714235},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_narrow","metaepi_broad","research_integrity"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4392947957","doi":"10.1371/journal.pone.0298526","title":"Do statistical heterogeneity methods impact the results of meta- analyses? A meta epidemiological study","year":2024,"lang":"en","type":"article","venue":"PLoS ONE","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Meta-analysis; Random effects model; Statistics; Confidence interval; Study heterogeneity; Pooling; Medicine; Restricted maximum likelihood; Interquartile range; Mathematics; Sample size determination; Maximum likelihood; Internal medicine; Computer science; Artificial intelligence","score_opus":0.9732907961212317,"score_gpt":0.7002165803757118,"score_spread":0.27307421574551993,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4392947957","genre_codex":"review","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0077699586,0.8838202,0.082606524,0.014266513,0.004679023,0.002636837,0.0016015104,0.00038215905,0.0022373127],"genre_scores_gemma":[0.50518066,0.30415305,0.15000956,0.015478463,0.0068015666,0.014916812,0.001668962,0.0008772412,0.00091364305],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.41211212,0.5115421,0.04076893,0.015350517,0.018689,0.0015373022],"domain_scores_gemma":[0.24629404,0.6945866,0.030533124,0.019738339,0.008163311,0.0006846809],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.36324233,0.003999351,0.017195128,0.013471672,0.0014116028,0.011256437,0.0057895244,0.006624019,0.005096776],"category_scores_gemma":[0.6638061,0.003327116,0.058099296,0.015878776,0.004035089,0.0100772735,0.00521838,0.0071461033,0.000831424],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002756306,0.000074971336,0.021218697,0.109864965,0.7623198,0.0003593557,0.0008442846,0.00471823,0.0002682058,0.0076593934,0.0064860727,0.083429754],"study_design_scores_gemma":[0.004995996,0.0010493206,0.014616816,0.08304467,0.8079356,0.00054251193,0.0005757509,0.010135212,0.0009075606,0.058225896,0.017555444,0.0004152837],"about_ca_topic_score_codex":0.0034753643,"about_ca_topic_score_gemma":0.0044005243,"teacher_disagreement_score":0.99600065,"about_ca_system_score_codex":0.0053309435,"about_ca_system_score_gemma":0.007016801,"threshold_uncertainty_score":0.7852356},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4394792109","doi":"","title":"Meta-Analyses Proved Inconsistent in How Missing Data Were Handled Across Their Included Primary Trials: A Methodological Survey","year":2020,"lang":"en","type":"article","venue":"DOAJ (DOAJ: Directory of Open Access Journals)","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Missing data; Primary (astronomy); Statistics; Computer science; Data mining; Psychology; Data science; Mathematics","score_opus":0.9953610906384595,"score_gpt":0.7960798083345333,"score_spread":0.19928128230392617,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4394792109","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.011282867,0.8622278,0.09091233,0.011896313,0.0038325046,0.011290487,0.0050231144,0.0005740162,0.0029605334],"genre_scores_gemma":[0.43017066,0.3035894,0.16991797,0.02891246,0.0037154672,0.05343484,0.006968465,0.0020404719,0.001250291],"study_design_codex":"meta_analysis","study_design_gemma":"observational","domain_scores_codex":[0.3578797,0.35519966,0.23884259,0.019364927,0.026805308,0.0019077578],"domain_scores_gemma":[0.2009488,0.68488854,0.04903254,0.04398908,0.020117257,0.0010238454],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.47185382,0.0044371192,0.01750084,0.014872323,0.0020448032,0.013287603,0.00825508,0.0057768,0.0041681635],"category_scores_gemma":[0.8076729,0.005204048,0.039049365,0.016113153,0.005116981,0.010331239,0.0075731226,0.006249323,0.00080953795],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0032906944,0.00008774306,0.0267469,0.3367147,0.53202426,0.0003967443,0.0026714213,0.0020109636,0.0005219929,0.008600954,0.0066327513,0.08030086],"study_design_scores_gemma":[0.0051301727,0.00070722314,0.018025449,0.23326391,0.667227,0.0009828336,0.0012963014,0.0022114695,0.0015289721,0.02951627,0.03969879,0.00041145564],"about_ca_topic_score_codex":0.004654045,"about_ca_topic_score_gemma":0.0076466226,"teacher_disagreement_score":0.9824992,"about_ca_system_score_codex":0.005714088,"about_ca_system_score_gemma":0.009765143,"threshold_uncertainty_score":0.6512983},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W4395045665","doi":"10.1101/2024.04.22.24305806","title":"Cracking the Code: A Scoping Review to Unite Disciplines in Tackling Legal Issues in Health Artificial Intelligence","year":2024,"lang":"en","type":"review","venue":"medRxiv","topic":"Artificial Intelligence in Healthcare and Education","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University; University of Toronto; Library and Archives Canada; Royal College of Physicians and Surgeons of Canada; Queen's University; Hospital for Sick Children; University of Ottawa","funders":"Canadian Institutes of Health Research; Hospital for Sick Children","keywords":"Code (set theory); Engineering; Engineering ethics; Computer science; Programming language","score_opus":0.3442469389447934,"score_gpt":0.5690300085512622,"score_spread":0.22478306960646877,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4395045665","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0012284337,0.98441535,0.0020726714,0.007118489,0.0013666059,0.0013174885,0.00030031428,0.00002439207,0.0021563116],"genre_scores_gemma":[0.010999113,0.9756783,0.005476016,0.0036383388,0.00038524083,0.002959829,0.0003901451,0.00003373893,0.00043924007],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.9465789,0.02178301,0.017671762,0.0024452268,0.010553466,0.00096753344],"domain_scores_gemma":[0.69288975,0.23880382,0.025122311,0.0052272785,0.036371116,0.0015856999],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.05928201,0.001815375,0.004993075,0.05818823,0.0034664827,0.009428355,0.0038768454,0.0058872807,0.0041817874],"category_scores_gemma":[0.2569166,0.0016184034,0.005007466,0.043944508,0.0059432182,0.012826055,0.0066245776,0.0041531525,0.00075316115],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010676972,0.000048411523,0.0016501797,0.67301726,0.0013384002,0.00037908248,0.00569112,0.00037679012,0.00043043302,0.009605991,0.017224424,0.2901313],"study_design_scores_gemma":[0.000011790862,0.000026148047,0.0007070276,0.95573026,0.0011086403,0.00011524501,0.0012305586,0.00009362761,0.0000829509,0.0013819046,0.03949472,0.000017063252],"about_ca_topic_score_codex":0.0121947015,"about_ca_topic_score_gemma":0.0372904,"teacher_disagreement_score":0.05928201,"about_ca_system_score_codex":0.010877197,"about_ca_system_score_gemma":0.05954945,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["bibliometrics"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4396607175","doi":"10.1016/j.jscai.2024.101629","title":"B-56 | Impact of Persistent Iatrogenic Atrial Septal Defect Following Transcatheter Mitral Valve Repair: A Meta-Analysis","year":2024,"lang":"en","type":"article","venue":"Journal of the Society for Cardiovascular Angiography & Interventions","topic":"Cardiac Valve Diseases and Treatments","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Medicine; Cardiology; Internal medicine; Mitral valve; Meta-analysis","score_opus":0.04445524211100815,"score_gpt":0.3645031608368662,"score_spread":0.320047918725858,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4396607175","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.027374322,0.96196234,0.004177973,0.00081081816,0.0010665252,0.00044509364,0.0028456736,0.00017217656,0.0011450715],"genre_scores_gemma":[0.7170178,0.2679539,0.0053286823,0.0021452042,0.000889587,0.0016269893,0.0028994794,0.00015483846,0.0019833408],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.98975134,0.005894266,0.0019761166,0.0012734067,0.0007376987,0.0003671558],"domain_scores_gemma":[0.98888946,0.0084195305,0.0012093447,0.0006004163,0.0006797158,0.0002015288],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.014780706,0.003813061,0.014620217,0.0051823375,0.00066089013,0.004091542,0.0018143593,0.003210454,0.0074588],"category_scores_gemma":[0.021443585,0.0017018522,0.09767576,0.0053767688,0.00064899086,0.001743934,0.0015553611,0.0032136845,0.0007064332],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0021372144,0.0000144538435,0.0024083538,0.025899401,0.9668811,0.00005848188,0.000016297763,0.0004947035,0.00014625455,0.000053994743,0.000266745,0.0016231409],"study_design_scores_gemma":[0.0005771555,0.00015594639,0.002765025,0.0016638993,0.9934371,0.00006028471,0.000016127186,0.00030104205,0.00006993449,0.00017002356,0.00077078893,0.000012560189],"about_ca_topic_score_codex":0.0073895613,"about_ca_topic_score_gemma":0.0094238315,"teacher_disagreement_score":0.014780706,"about_ca_system_score_codex":0.0016835949,"about_ca_system_score_gemma":0.0016005094,"threshold_uncertainty_score":0.07816875},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"empirical","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":"low"}],"label_agreement":"split"},{"id":"W4396989130","doi":"10.1681/asn.20223311s1266b","title":"Sex and Gender Considerations in Randomized Controlled Trials in Adults Receiving Chronic Dialysis: A Meta-Epidemiologic Study","year":2022,"lang":"en","type":"article","venue":"Journal of the American Society of Nephrology","topic":"Dialysis and Renal Disease Management","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; University of Calgary; University of Alberta","funders":"","keywords":"Randomized controlled trial; Medicine; Meta-analysis; Dialysis; Internal medicine; Kidney disease; Intensive care medicine","score_opus":0.06676825503888333,"score_gpt":0.34291590974973396,"score_spread":0.2761476547108506,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4396989130","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.01800082,0.9711944,0.006409085,0.001614403,0.00086356694,0.0005916915,0.00046352044,0.000061275,0.00080113055],"genre_scores_gemma":[0.7312861,0.23621422,0.020108342,0.0058707576,0.0018349864,0.003072464,0.00088229566,0.00012751167,0.0006033035],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.78498244,0.16346163,0.032710288,0.009019053,0.008513364,0.0013132929],"domain_scores_gemma":[0.64329225,0.30117345,0.036533326,0.011202088,0.006783131,0.0010157435],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.15398596,0.0024878453,0.011582717,0.0060582627,0.0008749586,0.005735715,0.0021736543,0.0034572128,0.0029097258],"category_scores_gemma":[0.24379915,0.0015678147,0.04987084,0.0068223393,0.0017023376,0.004015229,0.0019885018,0.0027689596,0.00023453152],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.007153704,0.000053177802,0.038933825,0.06944802,0.8620835,0.0002155454,0.0001828438,0.0009231947,0.00022857153,0.0008785399,0.00079623784,0.019102858],"study_design_scores_gemma":[0.0027063948,0.00090362557,0.02044583,0.022851832,0.94394714,0.00031138177,0.00015160853,0.0012240643,0.00032257242,0.0033290908,0.0037345418,0.00007190215],"about_ca_topic_score_codex":0.0020578357,"about_ca_topic_score_gemma":0.003139942,"teacher_disagreement_score":0.99751216,"about_ca_system_score_codex":0.0023370874,"about_ca_system_score_gemma":0.0030016177,"threshold_uncertainty_score":0.8143653},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4398792297","doi":"10.1016/j.jss.2024.04.054","title":"Factors Associated With Loss to Follow-Up in Surgical Trials: A Systematic Review and Meta-Analysis","year":2024,"lang":"en","type":"review","venue":"Journal of Surgical Research","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":3,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Health Sciences Centre; Sunnybrook Health Science Centre; St. Michael's Hospital; Public Health Ontario; University of Toronto","funders":"","keywords":"Meta-analysis; Medicine; Systematic review; MEDLINE; Internal medicine; Biology","score_opus":0.9649304470218802,"score_gpt":0.6943321596978526,"score_spread":0.2705982873240277,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398792297","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.014935725,0.98217946,0.0015831423,0.00019881119,0.00028147275,0.00036591914,0.00027889904,0.000038922313,0.000137726],"genre_scores_gemma":[0.4941973,0.49371973,0.0069805663,0.0015373555,0.0005592919,0.0017476982,0.0007579096,0.00007546726,0.00042468376],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.96049356,0.021198483,0.011295479,0.0032529333,0.002861585,0.0008978639],"domain_scores_gemma":[0.92688566,0.05774627,0.009824004,0.002778267,0.0021258353,0.000640022],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.03668081,0.0033217648,0.022833044,0.0051618954,0.0007923963,0.005114816,0.0025198967,0.004272717,0.0031306616],"category_scores_gemma":[0.08330769,0.0030236132,0.05721668,0.0057105813,0.0014508435,0.003606472,0.0029003944,0.00429326,0.00032079822],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0033917676,0.00003851224,0.0063005975,0.09631326,0.88792956,0.000075478274,0.000078122364,0.00031974193,0.000095127194,0.00006570174,0.00019484444,0.005197301],"study_design_scores_gemma":[0.00090662064,0.00018101477,0.002993765,0.005994689,0.9890805,0.00005444143,0.00004593198,0.0002070047,0.000053572505,0.00016067745,0.00029678366,0.000025028123],"about_ca_topic_score_codex":0.004526543,"about_ca_topic_score_gemma":0.010514695,"teacher_disagreement_score":0.97716695,"about_ca_system_score_codex":0.0028701636,"about_ca_system_score_gemma":0.0039200243,"threshold_uncertainty_score":0.19398892},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4399706167","doi":"10.1016/j.jval.2024.03.2436","title":"PT22 Consistency Between Comparative Efficacy Estimates from Target Trial Emulations and Their Origin Target Trials: A Meta-Epidemiological Study","year":2024,"lang":"en","type":"article","venue":"Value in Health","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Cascades (Canada); University of Toronto; McMaster University","funders":"","keywords":"Consistency (knowledge bases); Epidemiology; Meta-analysis; Medicine; Computer science; Artificial intelligence; Internal medicine","score_opus":0.8587779294536831,"score_gpt":0.5553820052404167,"score_spread":0.30339592421326644,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4399706167","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.19711217,0.48143962,0.27005613,0.008663202,0.006207007,0.0051598367,0.016329465,0.0017756716,0.01325684],"genre_scores_gemma":[0.9605279,0.009975905,0.017671471,0.0027702628,0.0011600798,0.0023504067,0.0031809856,0.00055061525,0.001812521],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.59184194,0.33182365,0.03965371,0.023093605,0.01165155,0.0019355805],"domain_scores_gemma":[0.20108715,0.6966075,0.03998591,0.053102806,0.008077019,0.001139559],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.2922475,0.0035498568,0.010885947,0.0052148565,0.0010254034,0.009133868,0.0067835487,0.006981043,0.014735965],"category_scores_gemma":[0.60529476,0.0037349872,0.038313344,0.008908444,0.0048186076,0.0078084157,0.003970312,0.008774421,0.001911259],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.026530823,0.00011522886,0.046106618,0.021821963,0.8685399,0.00042307007,0.0004741879,0.0025217934,0.00034169783,0.003156449,0.0033508064,0.026617495],"study_design_scores_gemma":[0.02008975,0.002709709,0.047043454,0.0069673103,0.887504,0.0014174333,0.00028995582,0.008535485,0.0011042325,0.016932061,0.0071320958,0.0002744838],"about_ca_topic_score_codex":0.0014231218,"about_ca_topic_score_gemma":0.0007975648,"teacher_disagreement_score":0.9964501,"about_ca_system_score_codex":0.0016619603,"about_ca_system_score_gemma":0.0013922964,"threshold_uncertainty_score":0.87278485},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4400573007","doi":"10.1093/ndt/gfae156","title":"Heart failure events in randomized controlled trials for adults receiving maintenance dialysis: a meta-epidemiologic study","year":2024,"lang":"en","type":"review","venue":"Nephrology Dialysis Transplantation","topic":"Heart Failure Treatment and Management","field":"Medicine","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Memorial University of Newfoundland; Impact; University of Alberta; Population Health Research Institute; Canadian VIGOUR Centre; McMaster University","funders":"British Heart Foundation","keywords":"Medicine; Heart failure; Dialysis; Randomized controlled trial; Internal medicine; Intensive care medicine; Cardiology","score_opus":0.09110686111055549,"score_gpt":0.3835865122031745,"score_spread":0.29247965109261903,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4400573007","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.03099202,0.96323484,0.0020756866,0.0005943302,0.00046574834,0.0009333868,0.0010107729,0.000063272244,0.0006299015],"genre_scores_gemma":[0.72084755,0.26307437,0.007159271,0.0025172106,0.0005643158,0.0037295173,0.001635648,0.000065176624,0.00040695476],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9464848,0.03242342,0.013182103,0.0041392036,0.0029393015,0.0008310946],"domain_scores_gemma":[0.9098453,0.067578815,0.01554814,0.0032104973,0.0031322222,0.00068500434],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.04209663,0.0022631935,0.016028523,0.005247638,0.00070347346,0.0039265724,0.0019889115,0.0027115713,0.0036044817],"category_scores_gemma":[0.09354974,0.0015895232,0.040762592,0.005932883,0.00092854013,0.002718686,0.0018341509,0.0027480153,0.00025992337],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004737719,0.00004867616,0.013886657,0.10187128,0.8739904,0.00010695197,0.00007765398,0.0003378666,0.000109083456,0.00017577018,0.00038852438,0.004269418],"study_design_scores_gemma":[0.0028968274,0.0003329678,0.009026475,0.011318378,0.9745847,0.000102701495,0.000051437084,0.00034534495,0.00008653317,0.00033983786,0.00089366786,0.000021182781],"about_ca_topic_score_codex":0.0033250833,"about_ca_topic_score_gemma":0.005261199,"teacher_disagreement_score":0.9977368,"about_ca_system_score_codex":0.0021757341,"about_ca_system_score_gemma":0.002157579,"threshold_uncertainty_score":0.22263092},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4400747098","doi":"10.1111/dom.15771","title":"Bariatric surgery and all‐cause mortality: A methodological review of studies using a non‐surgical comparator","year":2024,"lang":"en","type":"review","venue":"Diabetes Obesity and Metabolism","topic":"Bariatric Surgery and Outcomes","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University; Jewish General Hospital","funders":"Atara Biotherapeutics; Kowa Company; Amarin Corporation; Novo Nordisk; Merck; Novartis; Pfizer; Seqirus; AstraZeneca; National Institute of Diabetes and Digestive and Kidney Diseases; Boehringer Ingelheim","keywords":"Medicine; Surgical procedures; Surgery; General surgery; Intensive care medicine","score_opus":0.36883434458384223,"score_gpt":0.46638567469115594,"score_spread":0.0975513301073137,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4400747098","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0012933015,0.99633944,0.00068270916,0.00046530928,0.00018874463,0.00058047264,0.00025048776,0.000007590429,0.00019190331],"genre_scores_gemma":[0.0607488,0.9249321,0.006550255,0.0019082893,0.0003445468,0.004577169,0.0007343627,0.000031333566,0.00017321714],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.86019045,0.055821486,0.06357086,0.005488045,0.014045611,0.0008836004],"domain_scores_gemma":[0.5796017,0.3340613,0.06373059,0.0068877395,0.014893661,0.00082500326],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.10561095,0.0020707764,0.011256937,0.014250981,0.001177801,0.005153706,0.004476358,0.0038375456,0.003150479],"category_scores_gemma":[0.3272092,0.0020885656,0.019776514,0.014387123,0.002061462,0.003955005,0.0031306276,0.0020558694,0.0002991274],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008077614,0.00003033479,0.0039872145,0.91523165,0.04441147,0.000070623115,0.00026002293,0.00018138511,0.00015097864,0.00049032347,0.0009543645,0.033423968],"study_design_scores_gemma":[0.0009198011,0.00042225517,0.01117663,0.73338133,0.23655117,0.00037130117,0.00038653106,0.00024400293,0.00044524047,0.0014855518,0.014544659,0.00007151221],"about_ca_topic_score_codex":0.005148088,"about_ca_topic_score_gemma":0.011166858,"teacher_disagreement_score":0.89438903,"about_ca_system_score_codex":0.0064270976,"about_ca_system_score_gemma":0.012087755,"threshold_uncertainty_score":0.5585308},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4400853486","doi":"10.1038/s41433-024-03255-2","title":"Exploring the fragility of meta-analyses in ophthalmology: a systematic review","year":2024,"lang":"en","type":"review","venue":"Eye","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":3,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; McMaster University; Impact","funders":"","keywords":"Medicine; Fragility; Optometry; Ophthalmology; MEDLINE; Neuro-ophthalmology; Meta-analysis; Glaucoma; Pathology; Chemistry","score_opus":0.9808495744060876,"score_gpt":0.6805419536740732,"score_spread":0.30030762073201434,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4400853486","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"reproducibility","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"reproducibility","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0029916728,0.98191226,0.011284319,0.0016652896,0.000689713,0.0006551701,0.0003481071,0.00009689215,0.00035649413],"genre_scores_gemma":[0.33017537,0.60657156,0.044371244,0.0106808925,0.0026385363,0.0037254118,0.0009372444,0.00028482528,0.00061484776],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.5665536,0.33680305,0.061481066,0.0144698545,0.018955491,0.001736999],"domain_scores_gemma":[0.18346791,0.763391,0.028476529,0.016126413,0.007817522,0.0007205888],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.30990574,0.0054465258,0.028556822,0.0151600465,0.0014728388,0.009573102,0.0058525265,0.007980418,0.0043414356],"category_scores_gemma":[0.669813,0.0048126387,0.04652988,0.015002926,0.004285969,0.008599731,0.0068078735,0.0074165305,0.0005393079],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011105754,0.000024739422,0.0070619634,0.17549501,0.7841215,0.00029318532,0.00026825749,0.0013056818,0.000113796035,0.001536393,0.0013803715,0.027288508],"study_design_scores_gemma":[0.0008504512,0.00026602377,0.0038090562,0.05008883,0.93130386,0.0002679888,0.00019738311,0.001788109,0.00016675562,0.008293307,0.0028611312,0.00010704692],"about_ca_topic_score_codex":0.006157452,"about_ca_topic_score_gemma":0.014243741,"teacher_disagreement_score":0.99455345,"about_ca_system_score_codex":0.0053960998,"about_ca_system_score_gemma":0.008045432,"threshold_uncertainty_score":0.8510091},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4401337768","doi":"10.5430/jnep.v14n12p10","title":"Adrift down under: An integrative review of international students’ experiences in Australian higher education including shaping factors","year":2024,"lang":"en","type":"article","venue":"Journal of Nursing Education and Practice","topic":"International Student and Expatriate Challenges","field":"Social Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Popularity; Acculturation; Psychology; Social connectedness; Qualitative research; Learning community; Public relations; Pedagogy; Sociology; Political science; Social psychology; Social science; Immigration","score_opus":0.22535668201679224,"score_gpt":0.5472963545256717,"score_spread":0.32193967250887945,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4401337768","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0022914242,0.99448836,0.00016756136,0.0011652937,0.00038167863,0.0000604168,0.00006463388,0.0000054654142,0.0013751015],"genre_scores_gemma":[0.01510819,0.9817976,0.0005160416,0.0014190131,0.00015567652,0.00014919683,0.000099428034,0.000008205473,0.00074668525],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9958311,0.0018892207,0.00088070374,0.00030874435,0.00092087756,0.00016940558],"domain_scores_gemma":[0.9871383,0.009153998,0.0012631421,0.00020865392,0.0018869546,0.0003490783],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0058764345,0.0006997358,0.0017758345,0.00869714,0.0012192305,0.003856666,0.0009182373,0.0012193408,0.0036670764],"category_scores_gemma":[0.0140449945,0.00061400444,0.0016170541,0.011154838,0.0011136105,0.004429873,0.0024110393,0.0015842204,0.00037598785],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001189561,0.000082931256,0.0027824175,0.4490044,0.00076545787,0.000543507,0.02628489,0.00018792317,0.0009620987,0.0039818715,0.022032622,0.49325296],"study_design_scores_gemma":[0.00002207675,0.0003003542,0.024147337,0.5660825,0.0021204234,0.0016591296,0.020541025,0.0000936286,0.00036384648,0.0011815437,0.38341603,0.00007219013],"about_ca_topic_score_codex":0.011200285,"about_ca_topic_score_gemma":0.029352194,"teacher_disagreement_score":0.011200285,"about_ca_system_score_codex":0.002291193,"about_ca_system_score_gemma":0.0094255665,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4402643319","doi":"10.1016/j.sapharm.2024.09.004","title":"Development and evaluation of a model to identify publications on the clinical impact of pharmacist interventions","year":2024,"lang":"en","type":"article","venue":"Research in Social and Administrative Pharmacy","topic":"Pharmaceutical Practices and Patient Outcomes","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre Hospitalier Universitaire Sainte-Justine","funders":"Hospices Civils de Lyon; Centre Hospitalier Universitaire de Québec","keywords":"Psychological intervention; Pharmacist; Medicine; Psychology; Pharmacy; Family medicine; Nursing","score_opus":0.9444726263386489,"score_gpt":0.7745015023084805,"score_spread":0.16997112403016845,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402643319","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.28115347,0.0021156077,0.5256223,0.015262105,0.000927877,0.09418402,0.044608086,0.012770679,0.023355769],"genre_scores_gemma":[0.2396802,0.00043331983,0.7217193,0.0005706887,0.00008079534,0.02681139,0.008844173,0.00012240639,0.001737624],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9630701,0.019777533,0.0067396113,0.0034308285,0.0059598223,0.0010220881],"domain_scores_gemma":[0.6355732,0.29980558,0.019340092,0.010465756,0.031459358,0.0033560037],"candidate_categories":["metaresearch","bibliometrics"],"consensus_categories":[],"category_scores_codex":[0.060636453,0.0021040107,0.0023457946,0.016287467,0.0018028633,0.009042527,0.0039423946,0.0029145677,0.010703377],"category_scores_gemma":[0.1891181,0.0009183647,0.0041632345,0.0097247595,0.0011573336,0.0068614865,0.0041096834,0.0018103542,0.0020433287],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.010080469,0.009815457,0.33616516,0.012934545,0.00416551,0.0004746261,0.0039223824,0.051574897,0.0031990234,0.025255904,0.023160184,0.5192518],"study_design_scores_gemma":[0.008731909,0.00703506,0.08903793,0.0043182326,0.006442087,0.00068745407,0.0071072085,0.78025126,0.014146103,0.040201012,0.041422784,0.0006188907],"about_ca_topic_score_codex":0.029534696,"about_ca_topic_score_gemma":0.026613439,"teacher_disagreement_score":0.98371255,"about_ca_system_score_codex":0.007158714,"about_ca_system_score_gemma":0.024227818,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"simulation_or_modeling","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":[],"domain":null,"study_design":"design_other","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"simulation_or_modeling","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4403458056","doi":"10.1016/j.jclinepi.2024.111568","title":"Many randomized trials in a large systematic review were not registered and had evidence of selective outcome reporting: a metaepidemiological study","year":2024,"lang":"en","type":"review","venue":"Journal of Clinical Epidemiology","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":6,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Dalhousie University","funders":"Canadian Institutes of Health Research","keywords":"Epidemiology; Medicine; Meta-analysis; Randomized controlled trial; Systematic review; Outcome (game theory); MEDLINE; Internal medicine","score_opus":0.9846993043956908,"score_gpt":0.7822405296101224,"score_spread":0.2024587747855684,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403458056","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.027925167,0.9270465,0.025527207,0.005193013,0.0021775288,0.005962615,0.0033218013,0.00023504754,0.0026111486],"genre_scores_gemma":[0.6217243,0.29142213,0.048883785,0.019359395,0.001893573,0.011611709,0.0030277383,0.0005867034,0.0014906857],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.64249796,0.20085247,0.116918184,0.017935906,0.019385688,0.002409806],"domain_scores_gemma":[0.31692028,0.56563956,0.07156383,0.028339226,0.016243666,0.0012934152],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.24893424,0.0035771043,0.015398229,0.009021695,0.0018234121,0.008236538,0.003568503,0.005507441,0.0060817865],"category_scores_gemma":[0.5606114,0.004288137,0.02826128,0.0113530625,0.003794892,0.008354558,0.0044250283,0.0055215564,0.0007765094],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0061399117,0.000106903586,0.023908323,0.24513568,0.68711776,0.0005637973,0.00065726583,0.00027090262,0.00035579587,0.001587346,0.0028116773,0.031344574],"study_design_scores_gemma":[0.003572405,0.00034006886,0.009118485,0.04477581,0.93356335,0.00059033907,0.00020200822,0.000354802,0.0004331933,0.0027510412,0.004194887,0.00010356254],"about_ca_topic_score_codex":0.00434385,"about_ca_topic_score_gemma":0.0103903245,"teacher_disagreement_score":0.9964229,"about_ca_system_score_codex":0.004353872,"about_ca_system_score_gemma":0.00868436,"threshold_uncertainty_score":0.92619777},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","research_integrity"],"domain":"reporting","study_design":"observational","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4403591944","doi":"10.1016/j.atherosclerosis.2024.118624","title":"Publication bias in pharmacogenetics of statin-associated muscle symptoms: A meta-epidemiological study","year":2024,"lang":"en","type":"review","venue":"Atherosclerosis","topic":"Lipoproteins and Cardiovascular Health","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"Association Nationale de la Recherche et de la Technologie","keywords":"Pharmacogenetics; Medicine; Statin; Epidemiology; Meta-analysis; Hydroxymethylglutaryl-CoA Reductase Inhibitors; Publication bias; Internal medicine; Pharmacology; Genetics; Genotype; Biology","score_opus":0.3292323609253614,"score_gpt":0.4284963298699932,"score_spread":0.0992639689446318,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403591944","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.028608017,0.94139004,0.016748248,0.004297728,0.0021629573,0.0013717005,0.003453839,0.0002108397,0.0017566466],"genre_scores_gemma":[0.787738,0.17525195,0.01753287,0.007847581,0.0024356656,0.0046635442,0.0028906139,0.00020996989,0.0014298541],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.9485379,0.029161207,0.012387746,0.0045926506,0.0042632613,0.0010572418],"domain_scores_gemma":[0.83938354,0.13186064,0.015811782,0.0068131853,0.005527969,0.00060293986],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.07409009,0.0025505668,0.00939078,0.004972439,0.0008850941,0.0047840616,0.002193951,0.00310311,0.0056257234],"category_scores_gemma":[0.1305673,0.0013342362,0.037150137,0.0065810345,0.0010998521,0.002728856,0.001968239,0.0024614388,0.0003932656],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004457353,0.000044598746,0.046586215,0.11310744,0.8104276,0.0005090673,0.00022213835,0.001166936,0.0005217412,0.00091487257,0.003157334,0.018884731],"study_design_scores_gemma":[0.0013335733,0.00026742805,0.020289952,0.012793039,0.95713305,0.00030679585,0.00008848327,0.0013997583,0.0004747186,0.0022154714,0.0036380396,0.000059654594],"about_ca_topic_score_codex":0.002950201,"about_ca_topic_score_gemma":0.0045898054,"teacher_disagreement_score":0.99744946,"about_ca_system_score_codex":0.0017685033,"about_ca_system_score_gemma":0.0033150446,"threshold_uncertainty_score":0.39183056},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4404923291","doi":"10.1186/s12916-024-03778-1","title":"Integration of non-randomized studies with randomized controlled trials in meta-analyses of clinical studies: a meta-epidemiological study on effect estimation of interventions","year":2024,"lang":"en","type":"review","venue":"BMC Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University; St. Joseph’s Healthcare Hamilton; Impact","funders":"West China Hospital, Sichuan University; Sichuan Provincial Administration of Traditional Chinese Medicine; National Science Fund for Distinguished Young Scholars; Sichuan University; National Natural Science Foundation of China","keywords":"Medicine; Randomized controlled trial; Meta-analysis; Systematic review; MEDLINE; Concordance; Psychological intervention; Internal medicine; Psychiatry","score_opus":0.9831690436353597,"score_gpt":0.7817473869103255,"score_spread":0.2014216567250342,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4404923291","genre_codex":"review","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.007116962,0.7206984,0.2509719,0.0041366196,0.0037219955,0.00933565,0.0012922605,0.0006347924,0.0020914334],"genre_scores_gemma":[0.2791634,0.18671246,0.48384386,0.005019785,0.0031201073,0.039277986,0.0015124561,0.00073820545,0.0006118444],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.11992395,0.75267434,0.08593093,0.014524214,0.026108973,0.0008376972],"domain_scores_gemma":[0.083321616,0.8276784,0.042498842,0.03177748,0.014156571,0.0005671473],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.59441227,0.007595165,0.03328933,0.03957043,0.0023319144,0.018214336,0.008183176,0.008986355,0.0054124263],"category_scores_gemma":[0.86905533,0.007350706,0.07542479,0.039855853,0.005168329,0.019378548,0.012136445,0.009238432,0.0009082945],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0025530553,0.00013203264,0.015462685,0.27747372,0.5891928,0.000369486,0.0017075419,0.004945986,0.0004068451,0.0076215286,0.0022766725,0.09785772],"study_design_scores_gemma":[0.00324755,0.0012823911,0.008778485,0.13955443,0.79348385,0.00057869626,0.00049724156,0.014202765,0.0010633824,0.026337285,0.010503058,0.00047076578],"about_ca_topic_score_codex":0.0028153558,"about_ca_topic_score_gemma":0.0035519165,"teacher_disagreement_score":0.9924048,"about_ca_system_score_codex":0.008176958,"about_ca_system_score_gemma":0.011184864,"threshold_uncertainty_score":0.5001619},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4405184040","doi":"10.1007/s11739-024-03811-4","title":"Interventions to reduce burnout in emergency medicine: a national inventory of the Canadian experience to support global implementation of wellness initiatives","year":2024,"lang":"en","type":"review","venue":"Internal and Emergency Medicine","topic":"Healthcare professionals’ stress and burnout","field":"Health Professions","cited_by":3,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Newfoundland and Labrador Centre for Applied Health Research; Memorial University of Newfoundland; Western University; University of Manitoba","funders":"","keywords":"Burnout; Psychological intervention; Debriefing; Medicine; Mindfulness; Curriculum; Nursing; Grey literature; Intervention (counseling); Medical education; MEDLINE; Psychology; Clinical psychology","score_opus":0.22688376493843343,"score_gpt":0.5841488141598157,"score_spread":0.35726504922138225,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4405184040","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0039265547,0.9935364,0.00011113238,0.00083727634,0.00017446585,0.00021154102,0.0003350931,0.000008066733,0.00085952244],"genre_scores_gemma":[0.018199835,0.979387,0.0011745358,0.00045496225,0.000050874016,0.00017579788,0.00036091098,0.000004119304,0.0001919784],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9981306,0.00032991543,0.0002589546,0.00014943327,0.0008808907,0.00025022143],"domain_scores_gemma":[0.9940247,0.0021183277,0.0007509222,0.000104080216,0.002341143,0.00066089595],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0064232433,0.0010907057,0.003350349,0.0035275947,0.0014622296,0.0020564299,0.0018387135,0.00089137896,0.00211802],"category_scores_gemma":[0.009584597,0.00040639,0.0030357519,0.0053843274,0.0005968013,0.00075597095,0.0015072167,0.0021500723,0.00015840601],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00046908215,0.00050248584,0.008035465,0.15187347,0.004569637,0.000065232736,0.0012776209,0.00044281883,0.00022370006,0.00092222873,0.012238637,0.81937957],"study_design_scores_gemma":[0.0009122522,0.000933099,0.24378078,0.49368688,0.024253309,0.00028757897,0.0039088083,0.0005488247,0.00054125336,0.00085495465,0.23006746,0.00022484029],"about_ca_topic_score_codex":0.5575426,"about_ca_topic_score_gemma":0.7885465,"teacher_disagreement_score":0.44245738,"about_ca_system_score_codex":0.014108254,"about_ca_system_score_gemma":0.057719756,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":true,"confidence":"high"},{"model":"grok","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":true,"confidence":"high"},{"model":"opus","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":true,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4405223579","doi":"10.1093/jnci/djae318","title":"Justification, margin values, and analysis populations for oncologic noninferiority and equivalence trials: a meta-epidemiological study","year":2024,"lang":"en","type":"article","venue":"JNCI Journal of the National Cancer Institute","topic":"Statistical Methods in Clinical Trials","field":"Mathematics","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"London Health Sciences Centre; Western University","funders":"National Cancer Institute; University of Texas MD Anderson Cancer Center; National Institutes of Health; Andrew Sabin Family Foundation","keywords":"Margin (machine learning); Equivalence (formal languages); Clinical endpoint; Meta-analysis; Population; Medicine; Statistics; Clinical trial; Mathematics; Computer science; Internal medicine","score_opus":0.9378516184211011,"score_gpt":0.7058903326026658,"score_spread":0.23196128581843534,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4405223579","genre_codex":"review","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.030023178,0.7932782,0.14433591,0.008749559,0.00401087,0.013163911,0.001725378,0.00043960876,0.0042733946],"genre_scores_gemma":[0.76350445,0.0684234,0.11885251,0.009603154,0.0022482486,0.034775276,0.0014988008,0.00025465657,0.0008396173],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.38582116,0.53888655,0.04982474,0.008688702,0.015815465,0.00096333],"domain_scores_gemma":[0.2791421,0.6478822,0.044613127,0.020415992,0.0072134756,0.00073311495],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.36804473,0.0032836916,0.014442075,0.0072657415,0.0008886708,0.0067263558,0.0042719063,0.0074701584,0.004143189],"category_scores_gemma":[0.5986807,0.002335557,0.034162354,0.0073304987,0.0027302308,0.007033708,0.003051706,0.0060672853,0.0005118194],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.03729749,0.00021910928,0.020185905,0.16174252,0.6798503,0.00038412635,0.0006895188,0.005188847,0.00077615044,0.010875052,0.004522747,0.078268245],"study_design_scores_gemma":[0.032039296,0.003102367,0.011492654,0.0545486,0.82813114,0.00052928476,0.0003007086,0.014518269,0.00162498,0.04062272,0.012757381,0.00033272192],"about_ca_topic_score_codex":0.0006443718,"about_ca_topic_score_gemma":0.0008115879,"teacher_disagreement_score":0.9967163,"about_ca_system_score_codex":0.0024873635,"about_ca_system_score_gemma":0.00242627,"threshold_uncertainty_score":0.7793133},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4405596178","doi":"10.1016/j.jclinepi.2024.111648","title":"Meta-epidemiology and reporting characteristics of mapping reviews: a scoping review","year":2024,"lang":"en","type":"review","venue":"Journal of Clinical Epidemiology","topic":"Data-Driven Disease Surveillance","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University Medical Centre; Bruyère","funders":"National Office for Philosophy and Social Sciences","keywords":"Epidemiology; MEDLINE; Systematic review; Meta-analysis; Medicine; Data science; Environmental health; Computer science; Political science; Pathology","score_opus":0.8119543235295741,"score_gpt":0.6447508381367261,"score_spread":0.16720348539284802,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4405596178","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00094930583,0.9964185,0.001093216,0.00032535358,0.000115687886,0.00032537096,0.00053363526,0.000019291521,0.00021978316],"genre_scores_gemma":[0.03417908,0.9572816,0.0041998387,0.0012413599,0.00027479054,0.0014921057,0.0010803846,0.00006885479,0.00018204648],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.891501,0.041168593,0.047905758,0.006939203,0.011327197,0.0011582125],"domain_scores_gemma":[0.5958525,0.33452728,0.04604494,0.0078124623,0.014822515,0.0009404003],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.10266451,0.0026037795,0.017455483,0.018342761,0.0009545635,0.008741024,0.003841591,0.003783942,0.005108979],"category_scores_gemma":[0.38830978,0.0029168634,0.024564961,0.018836,0.0022047404,0.0064791283,0.004473223,0.0032877298,0.0006366206],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00066869886,0.000028123353,0.006946373,0.7626815,0.16055046,0.00010869856,0.00043019123,0.00036139952,0.00015983597,0.0008228401,0.0022784958,0.06496347],"study_design_scores_gemma":[0.0003222161,0.00014830762,0.007955372,0.5187978,0.4585471,0.00038019288,0.00027200027,0.00025902363,0.00023901706,0.001687539,0.011316721,0.0000747405],"about_ca_topic_score_codex":0.0073409188,"about_ca_topic_score_gemma":0.014894737,"teacher_disagreement_score":0.99739623,"about_ca_system_score_codex":0.0060117966,"about_ca_system_score_gemma":0.012347353,"threshold_uncertainty_score":0.5429483},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4406118033","doi":"10.1093/ehjqcco/qcae114","title":"Atrial fibrillation screening and clinical outcomes: a meta-analysis of randomized controlled trials","year":2025,"lang":"en","type":"review","venue":"European Heart Journal - Quality of Care and Clinical Outcomes","topic":"Atrial Fibrillation Management and Outcomes","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Fraser Health; University of British Columbia","funders":"","keywords":"Randomized controlled trial; Atrial fibrillation; Meta-analysis; Medicine; Internal medicine; Intensive care medicine; Cardiology","score_opus":0.5744384648020419,"score_gpt":0.5930892405730938,"score_spread":0.018650775771051897,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406118033","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.006874075,0.9890771,0.001867387,0.00040403492,0.0003270352,0.00055367727,0.0005793635,0.000062316554,0.00025502098],"genre_scores_gemma":[0.26928297,0.7125646,0.010113087,0.0016062054,0.00071120873,0.0037971593,0.0014859583,0.00009725725,0.00034162833],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.954384,0.030082079,0.009001916,0.002919327,0.0029567943,0.00065577205],"domain_scores_gemma":[0.90961057,0.07485135,0.010308922,0.0023928536,0.002309284,0.00052693055],"candidate_categories":["metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.043912355,0.004244468,0.027990805,0.008826748,0.0006629065,0.0047028186,0.0028803167,0.0032210678,0.0036800297],"category_scores_gemma":[0.09863027,0.002092892,0.05980059,0.008655533,0.0010916537,0.0025212418,0.0018428497,0.0031459075,0.00034909663],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0031715275,0.000029199093,0.0031117673,0.11313769,0.87361413,0.00006478769,0.000036585367,0.00074505457,0.000079897975,0.0001126033,0.0003502687,0.0055465396],"study_design_scores_gemma":[0.0014478058,0.00021849721,0.0018128424,0.009177117,0.9860293,0.000055212775,0.000016784208,0.00030723197,0.000050322742,0.0003410306,0.0005264268,0.000017482565],"about_ca_topic_score_codex":0.0046562348,"about_ca_topic_score_gemma":0.0063204532,"teacher_disagreement_score":0.99575555,"about_ca_system_score_codex":0.0031633708,"about_ca_system_score_gemma":0.0030642191,"threshold_uncertainty_score":0.23223346},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4407264846","doi":"10.1007/s10926-025-10271-w","title":"A Systematic Review and Evaluation of the Tools Measuring Work-Related Psychosocial Factors in Prospective Research on Sickness Absence of Health Care Workers","year":2025,"lang":"en","type":"review","venue":"Journal of Occupational Rehabilitation","topic":"Workplace Health and Well-being","field":"Health Professions","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Institut Universitaire en Santé Mentale de Québec; Institut de recherche Robert-Sauvé en santé et en sécurité du travail","funders":"Institut de Recherche Robert-Sauvé en Santé et en Sécurité du Travail","keywords":"Psychosocial; PsycINFO; Medicine; MEDLINE; Health care; Context (archaeology); Health psychology; Systematic review; Nursing; Public health; Psychiatry","score_opus":0.19633202670793937,"score_gpt":0.5526199643596708,"score_spread":0.3562879376517314,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4407264846","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0032956183,0.9929183,0.0005161754,0.00022645539,0.0001942475,0.0017842214,0.0008113528,0.000019170535,0.00023442546],"genre_scores_gemma":[0.0283998,0.96582764,0.0025027622,0.00050672,0.000065805754,0.002075716,0.00042455862,0.000011180776,0.00018586387],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.98364925,0.0044502323,0.00775505,0.0011067668,0.0026488383,0.00039000844],"domain_scores_gemma":[0.96486205,0.024445143,0.0070579364,0.00061382283,0.002586744,0.00043438518],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.016918687,0.0021292537,0.016457906,0.01130663,0.0010838427,0.0045957873,0.003162165,0.0027881186,0.0039713113],"category_scores_gemma":[0.060524553,0.0015091405,0.014115922,0.0119971605,0.001522286,0.0030725757,0.0030554975,0.0019521983,0.00027199584],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00030069862,0.000020915042,0.00089520076,0.9599494,0.020315994,0.00006074392,0.00022762781,0.00006429747,0.000103840204,0.0001258462,0.00046955515,0.017465867],"study_design_scores_gemma":[0.0005999532,0.0003020839,0.005713958,0.74214274,0.24116316,0.00022181867,0.0004070117,0.0001083471,0.00016222904,0.00021265468,0.008896938,0.0000691087],"about_ca_topic_score_codex":0.017711379,"about_ca_topic_score_gemma":0.05404069,"teacher_disagreement_score":0.98308134,"about_ca_system_score_codex":0.005588486,"about_ca_system_score_gemma":0.019653147,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4407554448","doi":"10.1016/j.amjsurg.2025.116250","title":"Inconsistency between abstract and main text reporting is common in pilot randomized controlled trials of surgical interventions: A methodological survey","year":2025,"lang":"en","type":"review","venue":"The American Journal of Surgery","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Joseph’s Healthcare Hamilton; Impact; McMaster University","funders":"","keywords":"Medicine; Randomized controlled trial; Confidence interval; Psychological intervention; MEDLINE; Clinical trial; Intervention (counseling); Family medicine; Surgery; Internal medicine; Psychiatry","score_opus":0.944213899314882,"score_gpt":0.6576468836200432,"score_spread":0.2865670156948388,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4407554448","genre_codex":"review","genre_gemma":"empirical","domain_codex":"reporting","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":"reporting","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.033049673,0.83902407,0.06645368,0.016143871,0.0066607883,0.021963967,0.009961974,0.00056205573,0.0061799344],"genre_scores_gemma":[0.6173455,0.17137718,0.12029445,0.025883192,0.004617668,0.0481815,0.009313574,0.001062358,0.001924635],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.1947581,0.36654302,0.3689254,0.017613977,0.050658707,0.0015007153],"domain_scores_gemma":[0.050498262,0.7869313,0.10387959,0.02652222,0.03131888,0.00084986375],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.59340686,0.0024814692,0.010425926,0.015168839,0.0019562924,0.009247357,0.004343945,0.005121792,0.005658508],"category_scores_gemma":[0.8681292,0.004389341,0.028062003,0.012260276,0.0065501193,0.007314333,0.0067164684,0.005234053,0.00071256],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0078049074,0.00021947866,0.047422986,0.42379448,0.3564295,0.0003690217,0.0064851893,0.0009881421,0.0007434926,0.0053879726,0.0121704815,0.13818443],"study_design_scores_gemma":[0.012487264,0.0029930396,0.07166486,0.31172675,0.5125212,0.0018360345,0.0025978386,0.002949368,0.0024325142,0.02538446,0.052565236,0.0008413959],"about_ca_topic_score_codex":0.0035562958,"about_ca_topic_score_gemma":0.006631051,"teacher_disagreement_score":0.40659314,"about_ca_system_score_codex":0.006896327,"about_ca_system_score_gemma":0.012687316,"threshold_uncertainty_score":0.5014018},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad","metaresearch"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4408378962","doi":"10.1136/bmjopen-2024-091342","title":"Access and use of general and mental health services before and during the COVID-19 pandemic: a systematic review and meta-analysis","year":2025,"lang":"en","type":"review","venue":"BMJ Open","topic":"COVID-19 and Mental Health","field":"Psychology","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Mental health; Meta-analysis; PsycINFO; Pandemic; Population; MEDLINE; Observational study; Cochrane Library; Family medicine; Psychiatry; Environmental health; Coronavirus disease 2019 (COVID-19); Disease; Pathology","score_opus":0.40098588280857894,"score_gpt":0.5946177684024331,"score_spread":0.19363188559385414,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4408378962","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00394016,0.9936231,0.00047165653,0.00020671863,0.000124772,0.00056179083,0.0008254905,0.000025738806,0.00022059103],"genre_scores_gemma":[0.12015981,0.8711289,0.0024103555,0.0010485025,0.00028089396,0.0033293692,0.001258206,0.000039712424,0.00034428682],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9874281,0.004757177,0.0048537436,0.0012003573,0.0013285249,0.00043199965],"domain_scores_gemma":[0.96533227,0.02550828,0.0054972917,0.0010656692,0.0022299807,0.00036647965],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01862907,0.0027004736,0.023258878,0.011013945,0.0007295422,0.003787879,0.0026090231,0.0025275538,0.0045391],"category_scores_gemma":[0.04826759,0.0014863238,0.03769695,0.011744473,0.0009917062,0.0024230857,0.0021062784,0.0017663455,0.00032785672],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009309405,0.00002279718,0.0023888177,0.6428246,0.34437284,0.00008714052,0.000111644214,0.00025365906,0.00009728675,0.000110584864,0.00050013163,0.008299519],"study_design_scores_gemma":[0.00039071238,0.00020918554,0.0037194036,0.09135599,0.9016466,0.000102773665,0.000096807846,0.00011908601,0.00008896864,0.00019102587,0.0020480163,0.000031406642],"about_ca_topic_score_codex":0.007920439,"about_ca_topic_score_gemma":0.016611403,"teacher_disagreement_score":0.023258878,"about_ca_system_score_codex":0.0035300427,"about_ca_system_score_gemma":0.005666592,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W4409203502","doi":"10.1016/j.drugalcdep.2025.112665","title":"Covariates in studies examining longitudinal relationships between substance use and mental health problems among youth: A meta-epidemiologic review","year":2025,"lang":"en","type":"review","venue":"Drug and Alcohol Dependence","topic":"Substance Abuse Treatment and Outcomes","field":"Medicine","cited_by":7,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre Hospitalier Universitaire Sainte-Justine; McMaster University; Université de Montréal; St. Joseph’s Healthcare Hamilton","funders":"Canadian Institutes of Health Research","keywords":"Mental health; Substance use; Meta-analysis; Psychology; Longitudinal study; Covariate; Clinical psychology; Psychiatry; Medicine","score_opus":0.517409130212774,"score_gpt":0.4333312985156157,"score_spread":0.08407783169715832,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4409203502","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0014540433,0.99782723,0.0001909563,0.00007745226,0.000086804655,0.00004710247,0.00024823943,0.000008150938,0.00005998023],"genre_scores_gemma":[0.039683674,0.95734465,0.0013579101,0.000671732,0.00014208563,0.00017293228,0.0004669733,0.000014937174,0.00014516103],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9943997,0.0018269297,0.00219987,0.00080378004,0.0005818878,0.00018768552],"domain_scores_gemma":[0.97752017,0.016823975,0.00365042,0.0007321052,0.0010406817,0.00023262341],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.011478402,0.0020540631,0.010890472,0.005376828,0.0005020737,0.003461234,0.0019838165,0.0025627874,0.0049199234],"category_scores_gemma":[0.033718884,0.0020857798,0.021855058,0.007285721,0.0007585496,0.0022140604,0.0022639225,0.002163487,0.00039258986],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011069869,0.00008554033,0.022883285,0.4750574,0.44712242,0.00015906396,0.00018089556,0.0003131177,0.0003174371,0.00024819613,0.0014345865,0.051091075],"study_design_scores_gemma":[0.0004175429,0.00024301329,0.023423456,0.08764132,0.8800185,0.00018096418,0.00016356815,0.00012585272,0.0001504379,0.00044097955,0.007137134,0.000057294255],"about_ca_topic_score_codex":0.009507512,"about_ca_topic_score_gemma":0.025308894,"teacher_disagreement_score":0.9891095,"about_ca_system_score_codex":0.0014304237,"about_ca_system_score_gemma":0.004276312,"threshold_uncertainty_score":0.06070435},"labels":[{"model":"gemma","categories":["metaepi_broad","metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4410804962","doi":"10.1200/jco.2025.43.16_suppl.e18109","title":"Comparison of survival outcomes of transoral surgery with or without adjuvant chemoradiotherapy vs chemoradiotherapy alone for oropharyngeal cancer: A systematic review and meta-analysis.","year":2025,"lang":"en","type":"review","venue":"Journal of Clinical Oncology","topic":"Head and Neck Cancer Studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Medicine; Chemoradiotherapy; Oncology; Meta-analysis; Adjuvant; Internal medicine; Head and neck cancer; Surgery; Cancer; Overall survival","score_opus":0.42385987848991047,"score_gpt":0.593154952517222,"score_spread":0.16929507402731148,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4410804962","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004001947,0.9941246,0.00042882183,0.00018759085,0.0001632468,0.00039132294,0.00046706205,0.000016453996,0.00021895667],"genre_scores_gemma":[0.14713117,0.8454244,0.0025295732,0.0010991943,0.00030517858,0.0020841225,0.0009772911,0.000030620933,0.0004184643],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.99230677,0.0037803063,0.0020272157,0.00077176845,0.00085364433,0.00026043016],"domain_scores_gemma":[0.98752636,0.009186463,0.0020160994,0.00029074377,0.00079972425,0.00018064641],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.011789092,0.0025088643,0.01739536,0.005308203,0.00057621126,0.0030677726,0.0019195992,0.0026091125,0.0037633516],"category_scores_gemma":[0.027243244,0.0012111766,0.045523416,0.005784517,0.000721505,0.0019969398,0.0014125031,0.0021618106,0.00032868734],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0029474979,0.000041041334,0.0017793267,0.49877217,0.4837503,0.000090184956,0.00008184071,0.00046268167,0.00020906013,0.00016362437,0.0004924012,0.011209788],"study_design_scores_gemma":[0.000910629,0.0003086959,0.0019060607,0.03564184,0.9591332,0.0000652696,0.00004285644,0.0001717745,0.0001083329,0.00018267645,0.0015064873,0.000022196786],"about_ca_topic_score_codex":0.0066779996,"about_ca_topic_score_gemma":0.015648872,"teacher_disagreement_score":0.01739536,"about_ca_system_score_codex":0.0029511086,"about_ca_system_score_gemma":0.0045722034,"threshold_uncertainty_score":0.06234747},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4410896045","doi":"10.23736/s0026-4806.25.09678-8","title":"Definitions for hypertensive response to exercise: a systematic review","year":2025,"lang":"en","type":"review","venue":"Minerva Medica","topic":"Cardiovascular and exercise physiology","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Medicine; Physical medicine and rehabilitation; Physical therapy","score_opus":0.06265724139647665,"score_gpt":0.35302126853064536,"score_spread":0.2903640271341687,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4410896045","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0011069021,0.9949545,0.00047988002,0.00048208312,0.00013202969,0.0013965465,0.0010638677,0.000021729633,0.0003624107],"genre_scores_gemma":[0.021284636,0.9718087,0.0019403483,0.00089183985,0.000098173005,0.0030746358,0.00073993404,0.000013097824,0.00014872136],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9836352,0.00542755,0.0067138723,0.0010584313,0.0027948879,0.00037012444],"domain_scores_gemma":[0.9579982,0.029059274,0.008789622,0.0006183528,0.0032305038,0.00030406116],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.015788324,0.0022038103,0.010430228,0.014229119,0.0007739301,0.003306867,0.003245038,0.0020339177,0.0050900364],"category_scores_gemma":[0.065754965,0.001277411,0.011570157,0.013719249,0.0015896533,0.003352698,0.0022883257,0.0015416427,0.00043076585],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009825337,0.00000706054,0.00055907934,0.97763526,0.006098251,0.00005248967,0.00013442265,0.000089212874,0.00005071396,0.00019543203,0.0010019034,0.014077972],"study_design_scores_gemma":[0.00020676838,0.00009298625,0.0029402738,0.9337318,0.052807376,0.00023128952,0.00020375755,0.000105541,0.00009373441,0.00036570668,0.009182562,0.00003815151],"about_ca_topic_score_codex":0.010086801,"about_ca_topic_score_gemma":0.02457568,"teacher_disagreement_score":0.9842117,"about_ca_system_score_codex":0.0061857104,"about_ca_system_score_gemma":0.014976216,"threshold_uncertainty_score":0.0834977},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4411507891","doi":"10.1016/j.jclinepi.2025.111872","title":"Target trial emulation of cardiovascular outcome trials of medications used to treat type 2 diabetes using real-world data: a systematic review of observational studies","year":2025,"lang":"en","type":"review","venue":"Journal of Clinical Epidemiology","topic":"Diabetes Treatment and Management","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Jewish General Hospital","funders":"Canadian Institutes of Health Research; McGill University","keywords":"Observational study; Medicine; Type 2 diabetes; Emulation; Clinical trial; Diabetes mellitus; Meta-analysis; Systematic review; MEDLINE; Outcome (game theory); Randomized controlled trial; Intensive care medicine; Internal medicine; Psychology; Endocrinology","score_opus":0.8697186499513354,"score_gpt":0.6643221108102215,"score_spread":0.2053965391411139,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411507891","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004896615,0.9835613,0.002637609,0.0009745449,0.00044241027,0.005159688,0.001717498,0.000050159604,0.00056012743],"genre_scores_gemma":[0.15271556,0.7941401,0.021150772,0.004254785,0.0005107161,0.024024455,0.0027511043,0.00009291815,0.00035963743],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.6862658,0.12766108,0.15111226,0.010670524,0.023053667,0.0012366573],"domain_scores_gemma":[0.37378904,0.4894499,0.09565377,0.013037039,0.026828514,0.0012417801],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.1671208,0.0023796118,0.017025044,0.021059157,0.0012821265,0.005633694,0.004431009,0.0037742048,0.0025005487],"category_scores_gemma":[0.4806101,0.0026177568,0.015773216,0.017655963,0.003164112,0.006262622,0.0035672027,0.0020990267,0.0003444246],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00052394177,0.000018160319,0.0029864893,0.9355398,0.038276177,0.000098928365,0.00038808447,0.00014131145,0.00016833388,0.00030545954,0.0007324966,0.020820856],"study_design_scores_gemma":[0.0013192454,0.00034388178,0.0068165236,0.7895412,0.18899292,0.0003200665,0.00048216246,0.00031794087,0.00054661336,0.0007587235,0.01047323,0.00008751583],"about_ca_topic_score_codex":0.0047812406,"about_ca_topic_score_gemma":0.010901274,"teacher_disagreement_score":0.98297495,"about_ca_system_score_codex":0.008760211,"about_ca_system_score_gemma":0.018909277,"threshold_uncertainty_score":0.88382983},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4411555570","doi":"10.1016/j.jclinepi.2025.111876","title":"Adherence to RECORD reporting guidelines among observational studies using routinely collected health data published in general medical journals: a meta-epidemiologic study","year":2025,"lang":"en","type":"article","venue":"Journal of Clinical Epidemiology","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"SickKids Foundation; Hospital for Sick Children; Queen's University; McMaster University; Women's College Hospital","funders":"","keywords":"Observational study; Medicine; Family medicine; MEDLINE; Epidemiology; Medical record; Environmental health; Pathology; Internal medicine","score_opus":0.9959672742694313,"score_gpt":0.8133929164219663,"score_spread":0.182574357847465,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411555570","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.37798628,0.51280403,0.07240283,0.005301099,0.0026219306,0.008757146,0.015410888,0.000597527,0.0041182954],"genre_scores_gemma":[0.91038036,0.022430276,0.050688546,0.0030159412,0.00041072635,0.005763775,0.0065416573,0.00025055415,0.00051805977],"study_design_codex":"meta_analysis","study_design_gemma":"observational","domain_scores_codex":[0.36550432,0.3491744,0.22354464,0.025514584,0.03443036,0.001831571],"domain_scores_gemma":[0.17652218,0.5671472,0.14888108,0.077267155,0.029018097,0.0011644539],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.33044782,0.0019850782,0.008755824,0.007372791,0.0022799096,0.0073504145,0.005444594,0.0060463473,0.0015910137],"category_scores_gemma":[0.6240453,0.004027754,0.025024908,0.014942979,0.003101946,0.0071434923,0.004122543,0.004773124,0.00032311902],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.006565352,0.00045050465,0.4015298,0.045598302,0.49846953,0.00022225565,0.004127055,0.00093097804,0.00044648245,0.0012916268,0.003758268,0.0366099],"study_design_scores_gemma":[0.0059895325,0.0021087527,0.40741292,0.03030569,0.52672666,0.0011603866,0.0026176376,0.006823978,0.0015574553,0.004360277,0.010207921,0.0007287679],"about_ca_topic_score_codex":0.005788413,"about_ca_topic_score_gemma":0.006666095,"teacher_disagreement_score":0.9980149,"about_ca_system_score_codex":0.0025573801,"about_ca_system_score_gemma":0.0041094082,"threshold_uncertainty_score":0.82567704},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4411977700","doi":"10.1186/s12916-025-04202-y","title":"Sex and gender considerations in randomized controlled trials in critical care nephrology: a meta-epidemiologic study","year":2025,"lang":"en","type":"review","venue":"BMC Medicine","topic":"Sex and Gender in Healthcare","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Alberta Health Services; University of Alberta","funders":"University of Alberta; King Chulalongkorn Memorial Hospital","keywords":"Medicine; Randomized controlled trial; Nephrology; Meta-analysis; Internal medicine; MEDLINE; Intensive care medicine","score_opus":0.6544556132467321,"score_gpt":0.5824175902738793,"score_spread":0.07203802297285278,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411977700","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0136157265,0.96404594,0.015250067,0.0023946657,0.0014047492,0.001319823,0.00087339856,0.00010035537,0.0009953871],"genre_scores_gemma":[0.738074,0.1946546,0.043255676,0.009130982,0.0027130789,0.0098103015,0.0013523208,0.00024061052,0.00076841575],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.55382687,0.34082156,0.07183878,0.015489805,0.016035236,0.0019878384],"domain_scores_gemma":[0.30286002,0.61201,0.0573642,0.018670985,0.008064168,0.0010305712],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.27638158,0.0029287308,0.015259961,0.008906798,0.0012175523,0.007023036,0.0033870651,0.004776938,0.0039419285],"category_scores_gemma":[0.42517027,0.0018894821,0.05346739,0.010652134,0.0026074986,0.0061200014,0.003007488,0.0036647078,0.0003567228],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0056775864,0.00006342297,0.050281428,0.13321348,0.7818112,0.0002775538,0.00037588374,0.0010754808,0.0002260024,0.002043163,0.0012246763,0.023730125],"study_design_scores_gemma":[0.0032026432,0.0011583806,0.026331743,0.056452163,0.8932977,0.00048083262,0.0003116492,0.0021489302,0.0005023124,0.008970038,0.0070165494,0.00012710273],"about_ca_topic_score_codex":0.0017453266,"about_ca_topic_score_gemma":0.0024224678,"teacher_disagreement_score":0.99707127,"about_ca_system_score_codex":0.0030554961,"about_ca_system_score_gemma":0.0035912849,"threshold_uncertainty_score":0.8923503},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaepi_broad","metaepi_narrow"],"domain":null,"study_design":"design_other","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4412582784","doi":"10.1136/bmjopen-2024-097148","title":"Data sharing statements for clinical trials: a cross-sectional survey of cardiology journals","year":2025,"lang":"en","type":"article","venue":"BMJ Open","topic":"Research Data Management Practices","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"St. Joseph’s Healthcare Hamilton","funders":"National Natural Science Foundation of China","keywords":"Medicine; Clinical trial; Publication; Data sharing; Odds; Odds ratio; MEDLINE; Family medicine; Logistic regression; Alternative medicine; Internal medicine; Pathology","score_opus":0.9414220175727007,"score_gpt":0.759484280899632,"score_spread":0.18193773667306867,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4412582784","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"reproducibility","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":"reproducibility","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9933832,0.0019859704,0.00030138853,0.0015631061,0.000030555642,0.00012387423,0.0012895601,0.000022423868,0.0012998341],"genre_scores_gemma":[0.99745125,0.0008403053,0.00047677418,0.00035150113,0.000064775144,0.000118107775,0.00053459074,0.000009083433,0.0001535564],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.95495504,0.0175498,0.015681233,0.0026193864,0.007936804,0.0012576483],"domain_scores_gemma":[0.55697423,0.15157631,0.25338078,0.010171101,0.019503873,0.008393717],"candidate_categories":["metaresearch","open_science"],"consensus_categories":[],"category_scores_codex":[0.031023854,0.00027062645,0.00070915057,0.006974768,0.0009948986,0.0024792608,0.00096955843,0.0011789347,0.0027147557],"category_scores_gemma":[0.15640461,0.0006081492,0.00064761983,0.010115131,0.0011619925,0.0042289062,0.0026819408,0.0010111477,0.0008027228],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010526132,0.00007315804,0.99291635,0.00023723698,0.00007181183,0.000056694993,0.0016399251,0.00003347979,0.00016295108,0.000054917564,0.0006262058,0.00402207],"study_design_scores_gemma":[0.000021132635,0.00025665856,0.9914124,0.00020692048,0.00004802636,0.0005587298,0.005150985,0.00020459933,0.00016188719,0.00009179992,0.0018566852,0.000030140665],"about_ca_topic_score_codex":0.0014663173,"about_ca_topic_score_gemma":0.0014071991,"teacher_disagreement_score":0.9990305,"about_ca_system_score_codex":0.0010567261,"about_ca_system_score_gemma":0.0018777184,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["metaresearch","scholarly_communication","open_science"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaresearch","metaepi_broad","open_science","scholarly_communication","research_integrity"],"domain":"reproducibility","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaresearch","open_science"],"domain":"reproducibility","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4413118160","doi":"10.1186/s13098-025-01909-z","title":"Comparative effectiveness of high-intensity interval training and moderate-intensity continuous training on cardiometabolic health in patients with diabesity: a systematic review and meta-analysis of randomized controlled trials","year":2025,"lang":"en","type":"review","venue":"Diabetology & Metabolic Syndrome","topic":"Cardiovascular and exercise physiology","field":"Medicine","cited_by":22,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia, Okanagan Campus; University of British Columbia; McMaster University","funders":"","keywords":"Medicine; High-intensity interval training; Glycated hemoglobin; Internal medicine; Interval training; Cardiorespiratory fitness; Cochrane Library; Insulin resistance; Overweight; Randomized controlled trial; Meta-analysis; Confidence interval; Lipid profile; Physical therapy; Obesity; Type 2 diabetes; Diabetes mellitus; Endocrinology; Cholesterol","score_opus":0.06320879528680888,"score_gpt":0.3466840780606362,"score_spread":0.28347528277382733,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413118160","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0052200225,0.9934596,0.0002839023,0.00011089734,0.00012955967,0.00042264504,0.00020601905,0.000013316128,0.00015393838],"genre_scores_gemma":[0.14816192,0.84559757,0.0024557356,0.0007499877,0.00031077975,0.0019910932,0.00049816014,0.000018695055,0.00021605013],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9902171,0.0042560278,0.0031884233,0.00092344766,0.0011818045,0.00023315648],"domain_scores_gemma":[0.9831782,0.01164953,0.0035401413,0.0003622203,0.0010360387,0.00023385766],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01299966,0.002406866,0.021785257,0.0049989917,0.0005966409,0.0029660014,0.0020499083,0.0024447225,0.0030106225],"category_scores_gemma":[0.029215567,0.0013383427,0.030360166,0.0055446904,0.0007610193,0.0016044871,0.001219601,0.0017561277,0.00019347994],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.004786977,0.000080613805,0.002223486,0.5574704,0.41724676,0.00006992357,0.00007781554,0.00038878116,0.00022022045,0.000103643826,0.00033660728,0.016994856],"study_design_scores_gemma":[0.0024758608,0.0006773202,0.0037310815,0.055964887,0.93531364,0.00008759767,0.00005968796,0.00021438386,0.0001647192,0.00013236117,0.0011523846,0.000026117774],"about_ca_topic_score_codex":0.0049304487,"about_ca_topic_score_gemma":0.009590026,"teacher_disagreement_score":0.021785257,"about_ca_system_score_codex":0.002628768,"about_ca_system_score_gemma":0.003677857,"threshold_uncertainty_score":0.06874955},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4413340902","doi":"10.3390/curroncol32080462","title":"High Outcome-Reporting Bias in Randomized-Controlled Trials of Acupuncture for Cancer Chemotherapy-Induced Nausea and Vomiting: A Systematic Review and Meta-Epidemiological Study","year":2025,"lang":"en","type":"review","venue":"Current Oncology","topic":"Acupuncture Treatment Research Studies","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Universitätsspital Zürich","keywords":"Medicine; Nausea; Acupuncture; Meta-analysis; Vomiting; Clinical trial; Reporting bias; Randomized controlled trial; MEDLINE; Publication bias; Metric (unit); Outcome (game theory); Prospective cohort study; Systematic review; Epidemiology; Physical therapy; Alternative medicine; Intensive care medicine; Internal medicine; Pathology","score_opus":0.6576657626141557,"score_gpt":0.6228960202366118,"score_spread":0.03476974237754382,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413340902","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004759124,0.9738273,0.0119324215,0.0019039145,0.001257728,0.003946792,0.0012117423,0.00010107656,0.001059838],"genre_scores_gemma":[0.3246876,0.5664148,0.05468537,0.011789525,0.0018481498,0.03621179,0.0029297732,0.00027046417,0.0011624785],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.5718687,0.25876173,0.12569055,0.015884982,0.025984723,0.0018093504],"domain_scores_gemma":[0.41884273,0.48163855,0.06267885,0.019254757,0.01663449,0.0009506881],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.27275917,0.002953997,0.018527037,0.008154755,0.0016656932,0.007666437,0.0036917077,0.0045274603,0.0046691876],"category_scores_gemma":[0.46814972,0.0027175273,0.027865857,0.011436196,0.003399065,0.0059678312,0.0034539911,0.0036791223,0.0005166873],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017144344,0.00006039441,0.009915896,0.6528687,0.2881564,0.00023302856,0.00054777804,0.00066973513,0.00024513516,0.002404852,0.0024590492,0.04072457],"study_design_scores_gemma":[0.0040540993,0.00082296703,0.015795516,0.31742373,0.6238986,0.0006040369,0.0004729572,0.0018033322,0.0011223359,0.012603743,0.021147914,0.00025064405],"about_ca_topic_score_codex":0.00502745,"about_ca_topic_score_gemma":0.0083276,"teacher_disagreement_score":0.997046,"about_ca_system_score_codex":0.007156826,"about_ca_system_score_gemma":0.011932893,"threshold_uncertainty_score":0.89681745},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"reporting","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"reporting","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W4413408738","doi":"10.1177/01410768251366880","title":"Quantifying the impact of immortal time bias: empirical evidence from meta-analyses","year":2025,"lang":"en","type":"article","venue":"Journal of the Royal Society of Medicine","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ottawa Hospital; Mental Health Research Canada; University of Ottawa","funders":"","keywords":"Interquartile range; Meta-analysis; Subgroup analysis; MEDLINE; Study heterogeneity; Medicine; Systematic review; Confidence interval; Publication bias; Random effects model; Statistics; Surgery; Internal medicine; Mathematics","score_opus":0.9106758573085312,"score_gpt":0.6310800049927433,"score_spread":0.2795958523157879,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413408738","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0066176234,0.92181075,0.05989953,0.0021968784,0.001531181,0.0035880331,0.0025423078,0.00027816682,0.0015355117],"genre_scores_gemma":[0.5548051,0.29499057,0.10782666,0.0075626373,0.002622866,0.025711702,0.00462058,0.0006368828,0.0012228686],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.5466023,0.32941648,0.08022847,0.017746601,0.024526339,0.0014798709],"domain_scores_gemma":[0.2216016,0.7013493,0.04309535,0.024897235,0.008431614,0.000624763],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.34069481,0.006109349,0.024570493,0.019512208,0.001075744,0.009875962,0.006438392,0.0063927076,0.0059187934],"category_scores_gemma":[0.65579885,0.003247391,0.06299295,0.01546796,0.004546705,0.010441066,0.006663999,0.0064329524,0.00069061556],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016218234,0.000037888618,0.0082120085,0.22680536,0.73018795,0.00021038405,0.00032864203,0.0030416877,0.00015567344,0.0026116197,0.0012486866,0.0255383],"study_design_scores_gemma":[0.0018793094,0.0005979677,0.0074874596,0.073529854,0.87674177,0.00044560846,0.00025515232,0.003939254,0.00068242784,0.026121505,0.008093697,0.00022599722],"about_ca_topic_score_codex":0.0026520882,"about_ca_topic_score_gemma":0.0033534549,"teacher_disagreement_score":0.97542953,"about_ca_system_score_codex":0.0041103256,"about_ca_system_score_gemma":0.00457399,"threshold_uncertainty_score":0.8130407},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_narrow","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4413788159","doi":"10.1016/j.jhepr.2025.101566","title":"Liver transplantation for HCC with macrovascular invasion: A systematic review and meta-analysis of observational studies","year":2025,"lang":"en","type":"review","venue":"JHEP Reports","topic":"Hepatocellular Carcinoma Treatment and Prognosis","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto; Princess Margaret Cancer Centre; Toronto General Hospital; University Health Network","funders":"Incyte; Roche; AstraZeneca; Servier; MSD","keywords":"Liver transplantation; Observational study; Meta-analysis; Medicine; Internal medicine; Transplantation; Oncology","score_opus":0.3662705535403707,"score_gpt":0.38299568102747755,"score_spread":0.01672512748710686,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413788159","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0031322394,0.99542564,0.00034977793,0.00018401469,0.00008981092,0.00019126161,0.00046270198,0.000011417394,0.00015308357],"genre_scores_gemma":[0.08941633,0.9063664,0.001558757,0.00068720774,0.00018963071,0.00083406566,0.0007737726,0.000015424963,0.0001584043],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.99281126,0.002925662,0.0025929422,0.00065152097,0.0007627934,0.0002558183],"domain_scores_gemma":[0.9803709,0.014287795,0.0032394254,0.00046105555,0.0014444783,0.00019645193],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01136883,0.0022292358,0.014603046,0.006408352,0.00062957464,0.0027059922,0.0017688762,0.0015697791,0.003046721],"category_scores_gemma":[0.028215077,0.0011192844,0.030058483,0.00807826,0.00057999964,0.0015843095,0.0014195978,0.0016505874,0.00022532565],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00095183915,0.000025538735,0.0047044586,0.62681633,0.35225818,0.00016092195,0.0001096548,0.00040644212,0.0001990808,0.00012920493,0.00076205656,0.013476252],"study_design_scores_gemma":[0.00030254817,0.00014136021,0.005523614,0.058018524,0.9330591,0.00014120057,0.000084616055,0.00016860166,0.00010700736,0.00021750451,0.002208205,0.000027746928],"about_ca_topic_score_codex":0.006850519,"about_ca_topic_score_gemma":0.016302655,"teacher_disagreement_score":0.014603046,"about_ca_system_score_codex":0.0023901488,"about_ca_system_score_gemma":0.0043914984,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4415273748","doi":"10.3389/fimmu.2025.1613441","title":"Air pollution, emerging chemical exposures, and systemic lupus erythematosus: a meta-epidemiology study","year":2025,"lang":"en","type":"article","venue":"Frontiers in Immunology","topic":"Systemic Lupus Erythematosus Research","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"National Natural Science Foundation of China","keywords":"Systemic lupus erythematosus; Systemic lupus; Autoimmune disease; Autoimmunity; Lupus erythematosus; Autoantibody","score_opus":0.02863846204813981,"score_gpt":0.3139015928459654,"score_spread":0.2852631307978256,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4415273748","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.047232658,0.94480914,0.0043401467,0.0005428964,0.0005382792,0.00032599687,0.0015739431,0.0001088342,0.0005281993],"genre_scores_gemma":[0.6547919,0.32996047,0.009420032,0.0011017461,0.0008490934,0.0011204924,0.0019765685,0.000094238676,0.00068545487],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.98840475,0.0061399946,0.002323121,0.0017697136,0.00093532394,0.0004271213],"domain_scores_gemma":[0.9851833,0.009910577,0.0021170639,0.0013835172,0.0011071542,0.00029847398],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.015814807,0.0041069738,0.013558029,0.0076946714,0.00084120187,0.0042712837,0.002087432,0.0025140846,0.0025400573],"category_scores_gemma":[0.028403958,0.0020039026,0.061966974,0.008527305,0.00067567086,0.001611454,0.0019545122,0.0021875156,0.0002722932],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016667781,0.000030451225,0.019488553,0.021333264,0.95242953,0.00014820637,0.000029081157,0.0009282278,0.00031474818,0.000083129264,0.0001993981,0.0033486371],"study_design_scores_gemma":[0.00026575997,0.00018188986,0.0072629564,0.0017052614,0.9893021,0.00011453481,0.0000222408,0.0003803998,0.000122021695,0.00014016953,0.00048873614,0.000013937116],"about_ca_topic_score_codex":0.0048770765,"about_ca_topic_score_gemma":0.006612053,"teacher_disagreement_score":0.98644197,"about_ca_system_score_codex":0.001492897,"about_ca_system_score_gemma":0.0022204216,"threshold_uncertainty_score":0.083637714},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"observational","genre":"empirical","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":"low"}],"label_agreement":"split"},{"id":"W4415679231","doi":"10.1016/j.jpsychires.2025.10.067","title":"Gambling disorder and suicide risk: A meta-epidemiology study","year":2025,"lang":"en","type":"review","venue":"Journal of Psychiatric Research","topic":"Gambling Behavior and Treatments","field":"Psychology","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Gambling disorder; Poison control; Human factors and ergonomics; Suicide prevention; Association (psychology); Injury prevention; Occupational safety and health","score_opus":0.7015815654600905,"score_gpt":0.658004460249008,"score_spread":0.04357710521108249,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4415679231","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.021010833,0.9745317,0.0020606478,0.0004730321,0.0003495439,0.00024776603,0.0008735394,0.00006330871,0.000389548],"genre_scores_gemma":[0.44850406,0.53851205,0.007585394,0.001326038,0.0008774298,0.0009783687,0.0017154652,0.0000737094,0.00042749994],"study_design_codex":"meta_analysis","study_design_gemma":"systematic_review","domain_scores_codex":[0.99053615,0.005023058,0.0020464975,0.0012097331,0.0008866765,0.00029788134],"domain_scores_gemma":[0.9860852,0.008981362,0.0026316913,0.0009907089,0.0010861309,0.00022488344],"candidate_categories":["metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.014064176,0.0032250076,0.013666629,0.007822622,0.00062242034,0.0038136805,0.0018643033,0.0020300944,0.0023534077],"category_scores_gemma":[0.029885938,0.0017468268,0.057524655,0.0070901653,0.00048773133,0.0014586423,0.0015254841,0.0020131653,0.0002510642],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014893847,0.00003106146,0.018247413,0.04827571,0.9241071,0.00011660268,0.000036727735,0.0006274806,0.00020116649,0.0000916321,0.00029562,0.006480133],"study_design_scores_gemma":[0.00038525122,0.00018477911,0.009591674,0.0064124614,0.9817825,0.0001671341,0.000027541,0.000399953,0.0000911994,0.0001813848,0.00075654965,0.000019609934],"about_ca_topic_score_codex":0.0045990516,"about_ca_topic_score_gemma":0.007885077,"teacher_disagreement_score":0.9863334,"about_ca_system_score_codex":0.0016238294,"about_ca_system_score_gemma":0.0021316165,"threshold_uncertainty_score":0.074379385},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W4416642995","doi":"10.1016/j.jclinepi.2025.112089","title":"Methodological guidance for individual participant data meta-analyses: a systematic review","year":2025,"lang":"en","type":"review","venue":"Journal of Clinical Epidemiology","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Observational study; Checklist; Set (abstract data type); Identification (biology); Guideline; Quality (philosophy); Forcing (mathematics); Observational methods in psychology","score_opus":0.9985771036472134,"score_gpt":0.8425318825246398,"score_spread":0.15604522112257357,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4416642995","genre_codex":"methods","genre_gemma":"review","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0017619447,0.41744608,0.48338822,0.016862478,0.0073688985,0.052924376,0.013292927,0.0035437476,0.00341133],"genre_scores_gemma":[0.026196098,0.08685622,0.80295914,0.010550576,0.0018613274,0.06623444,0.003119546,0.00097365584,0.0012490221],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.46373546,0.40386587,0.09839276,0.009763714,0.023141341,0.0011008077],"domain_scores_gemma":[0.34285784,0.5757114,0.028976593,0.0192064,0.031368434,0.0018793197],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.31969762,0.0056961793,0.026912693,0.017562907,0.0019328935,0.010768601,0.010808584,0.011458141,0.015354106],"category_scores_gemma":[0.6728524,0.008769023,0.035722133,0.011349423,0.0040251096,0.006725416,0.006427521,0.0138882985,0.005039823],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00439051,0.00015382207,0.0018230911,0.6136274,0.088759564,0.0005346242,0.0016070136,0.0036969862,0.0011295582,0.010379564,0.075733125,0.19816478],"study_design_scores_gemma":[0.023461834,0.0010951196,0.005725133,0.48993865,0.2525045,0.0013686242,0.00060859194,0.018832697,0.0025949073,0.08276467,0.11980049,0.0013047578],"about_ca_topic_score_codex":0.007723104,"about_ca_topic_score_gemma":0.016349336,"teacher_disagreement_score":0.6803024,"about_ca_system_score_codex":0.005795052,"about_ca_system_score_gemma":0.024116501,"threshold_uncertainty_score":0.83893394},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W4417202017","doi":"10.14309/01.ajg.0001136700.06357.98","title":"S2310 Comparing Robotic, Laparoscopic and Open Pancreaticoduodenectomy: A Systematic Review and Network Meta-Analysis of Randomized Controlled Trials and Propensity-Matched Studies","year":2025,"lang":"en","type":"article","venue":"The American Journal of Gastroenterology","topic":"Pancreatic and Hepatic Oncology Research","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Confidence interval; Randomized controlled trial; Gastric emptying; Relative risk; Meta-analysis; Pancreatic fistula; Fistula; Laparoscopy","score_opus":0.138185232599625,"score_gpt":0.42511641634441677,"score_spread":0.28693118374479176,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4417202017","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.007013842,0.98765033,0.0017928289,0.00044499626,0.00030264113,0.00088559964,0.0014162123,0.000048894,0.0004446128],"genre_scores_gemma":[0.2711408,0.70948297,0.008629339,0.001576742,0.00075316726,0.005182946,0.0024886436,0.00007343942,0.0006719187],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9763485,0.011985531,0.0068832594,0.0019697975,0.00236373,0.00044906882],"domain_scores_gemma":[0.96442664,0.027076695,0.005630433,0.0008686032,0.0017383442,0.00025927904],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.024042942,0.0029146967,0.019975496,0.010822042,0.00066960603,0.004438179,0.0022124483,0.0026018424,0.005765919],"category_scores_gemma":[0.05752636,0.0014378245,0.043400243,0.010651085,0.0008661395,0.002311899,0.001699553,0.0016673235,0.00038216545],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016912771,0.000027274697,0.004605275,0.34675032,0.6341631,0.000101058096,0.00006483874,0.00061417633,0.00014830897,0.00021315046,0.0005597712,0.011061444],"study_design_scores_gemma":[0.000543987,0.00013808806,0.0026402948,0.024026869,0.97051907,0.00006775656,0.000036363857,0.00029463365,0.00007664312,0.00031539387,0.001318921,0.000022003902],"about_ca_topic_score_codex":0.004480157,"about_ca_topic_score_gemma":0.0076169996,"teacher_disagreement_score":0.024042942,"about_ca_system_score_codex":0.0027940592,"about_ca_system_score_gemma":0.0047401018,"threshold_uncertainty_score":0.1271528},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4417223881","doi":"10.1053/j.ajkd.2025.09.023","title":"Representation of Older Patients Receiving Maintenance Dialysis in Randomized Controlled Trials: A Meta-epidemiologic Study","year":2025,"lang":"en","type":"article","venue":"American Journal of Kidney Diseases","topic":"Dialysis and Renal Disease Management","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ontario Stroke Network; McMaster University; University of Alberta","funders":"","keywords":"Dialysis; Representation (politics); Randomized controlled trial; MEDLINE; Kidney disease","score_opus":0.030261947742329164,"score_gpt":0.3538328628151803,"score_spread":0.32357091507285113,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4417223881","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.10560492,0.8646021,0.019702971,0.0020729676,0.0014544841,0.0023901847,0.001818835,0.00020160944,0.0021519118],"genre_scores_gemma":[0.9183269,0.062902614,0.011060288,0.0022983078,0.00069313915,0.002895867,0.0012502273,0.00009819733,0.0004744106],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.80483884,0.12839521,0.045025557,0.0117378,0.0084799975,0.0015225034],"domain_scores_gemma":[0.7094052,0.23590589,0.03334476,0.013789891,0.0063939714,0.001160233],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.13142201,0.0023267681,0.013809372,0.004431253,0.0008413663,0.005772026,0.002601005,0.0037462532,0.0037464048],"category_scores_gemma":[0.24574345,0.0018116074,0.03836679,0.006386269,0.0013084468,0.003952472,0.0026309395,0.003312248,0.00031150848],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.011434934,0.00009326505,0.047953457,0.037828922,0.89170694,0.00014794628,0.00018263755,0.0005512826,0.00017274337,0.0005574849,0.0006152718,0.008755098],"study_design_scores_gemma":[0.0040957085,0.0005202295,0.01759126,0.0067680324,0.96640956,0.00017456409,0.00010423519,0.0011843344,0.0001433398,0.0014667485,0.0014926638,0.000049206188],"about_ca_topic_score_codex":0.0015944169,"about_ca_topic_score_gemma":0.002377747,"teacher_disagreement_score":0.9976732,"about_ca_system_score_codex":0.0015973846,"about_ca_system_score_gemma":0.0016737212,"threshold_uncertainty_score":0.6950344},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W6887815647","doi":"10.17605/osf.io/tsh5w","title":"Prevalence, incidence, and factors associated with pain-related disabilities among Indigenous peoples in Canada: A scoping review.","year":2022,"lang":"en","type":"other","venue":"Open Science Framework","topic":"Pain Management and Opioid Use","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Indigenous; Population; Ethnic group; Protocol (science); Work (physics); Inclusion (mineral)","score_opus":0.017826733034589835,"score_gpt":0.2906009424584869,"score_spread":0.2727742094238971,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6887815647","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.060000353,0.6950614,0.0012867707,0.0074423687,0.0006784789,0.011266889,0.20073132,0.00018283282,0.023349533],"genre_scores_gemma":[0.25663447,0.6596293,0.009981222,0.0025929066,0.00018206936,0.013788557,0.050700713,0.000068362344,0.006422406],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9939116,0.00074157346,0.0018226386,0.0003905936,0.0024429848,0.0006905432],"domain_scores_gemma":[0.9744742,0.006026046,0.003612059,0.00047614676,0.014531188,0.00088038866],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0103171775,0.0007190141,0.002203195,0.018919028,0.0042802216,0.0038237115,0.0020680223,0.0012734497,0.004789318],"category_scores_gemma":[0.039809264,0.00070207805,0.0030441773,0.035822906,0.00095943373,0.0013490494,0.0024572697,0.0011207542,0.00038492528],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00062501756,0.00015656796,0.15270528,0.44880667,0.004012903,0.0006655524,0.011668951,0.0006142605,0.00047408804,0.002963569,0.106372744,0.2709344],"study_design_scores_gemma":[0.0002026773,0.00013973648,0.4873301,0.41263065,0.010393655,0.0006184971,0.009993143,0.00038837202,0.0003416069,0.00072388496,0.07704314,0.00019453271],"about_ca_topic_score_codex":0.9790211,"about_ca_topic_score_gemma":0.9904531,"teacher_disagreement_score":0.045585636,"about_ca_system_score_codex":0.045585636,"about_ca_system_score_gemma":0.24912548,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":true,"confidence":"low"},{"model":"grok","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":true,"confidence":"medium"},{"model":"opus","categories":["metaepi_broad"],"domain":null,"study_design":"systematic_review","genre":"protocol","about_ca_system":false,"about_ca_topic":true,"confidence":"low"}],"label_agreement":"split"},{"id":"W6890027164","doi":"10.3205/25ebm077","title":"Research integrity assessment of randomized controlled trials in COVID-19 evidence syntheses: focus on ethics and trial registration","year":2025,"lang":"en","type":"article","venue":"German Medical Science (German Research Foundation)","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Cochrane","funders":"","keywords":"Randomized controlled trial; Research ethics; Focus (optics); Clinical trial; Research design; Trial registration; MEDLINE; Ethics committee","score_opus":0.8941467816641128,"score_gpt":0.7364740610737476,"score_spread":0.1576727205903652,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6890027164","genre_codex":"methods","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00668162,0.04077261,0.45412973,0.08534915,0.022208735,0.3231885,0.022863232,0.004834487,0.039971907],"genre_scores_gemma":[0.03468935,0.014535778,0.4451308,0.007649096,0.0051030708,0.48127192,0.005835025,0.001806955,0.003977904],"study_design_codex":"systematic_review","study_design_gemma":"observational","domain_scores_codex":[0.08647135,0.678232,0.18728465,0.010020932,0.035348848,0.0026422243],"domain_scores_gemma":[0.032978613,0.64129406,0.08301941,0.12677433,0.11037454,0.0055590975],"candidate_categories":["metaresearch","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.77403426,0.0040781866,0.01658704,0.024621297,0.0067935986,0.024230374,0.0109520955,0.01925473,0.04865831],"category_scores_gemma":[0.9144363,0.007279189,0.01262291,0.030704029,0.023394132,0.016512968,0.015175659,0.018109532,0.02429326],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0066880356,0.00041360443,0.0034559327,0.37880513,0.003971926,0.00052830426,0.015576746,0.0017999281,0.0023248403,0.14530936,0.17315757,0.26796868],"study_design_scores_gemma":[0.010444644,0.0014723066,0.008056112,0.344181,0.0043378733,0.0014581251,0.0026067744,0.00836494,0.005558719,0.22486675,0.3876619,0.0009908562],"about_ca_topic_score_codex":0.0016204317,"about_ca_topic_score_gemma":0.0019446171,"teacher_disagreement_score":0.98074526,"about_ca_system_score_codex":0.010931259,"about_ca_system_score_gemma":0.10041868,"threshold_uncertainty_score":0.27865607},"labels":[{"model":"gemma","categories":["metaresearch","research_integrity"],"domain":"methods","study_design":"systematic_review","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad","research_integrity"],"domain":"methods","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W6902207847","doi":"10.6084/m9.figshare.16916538.v1","title":"Additional file 1 of Patient-important outcomes reported in randomized controlled trials of pharmacologic treatments for COVID-19: a protocol of a META-epidemiological study","year":2021,"lang":"en","type":"article","venue":"Figshare","topic":"COVID-19 Clinical Research Studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; McMaster University; Université de Montréal; University of Saskatchewan","funders":"","keywords":"Randomized controlled trial; Protocol (science); MEDLINE; Clinical trial; N of 1 trial","score_opus":0.41525225793469334,"score_gpt":0.5653596647413841,"score_spread":0.15010740680669077,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6902207847","genre_codex":"dataset","genre_gemma":"protocol","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00047361534,0.000367339,0.0014068933,0.0004174938,0.00013284858,0.009451205,0.9851,0.0003694704,0.002281174],"genre_scores_gemma":[0.030626707,0.0020146456,0.036165558,0.004056684,0.0006587559,0.4003295,0.494771,0.0013849244,0.029992208],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99228525,0.0036150964,0.0021468883,0.00083024637,0.00073215796,0.00039028074],"domain_scores_gemma":[0.8343484,0.14386387,0.008603897,0.005356384,0.0067273136,0.0011001298],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.013588719,0.00189488,0.0029777905,0.0038004627,0.0010585362,0.0023358425,0.0018968568,0.0023256605,0.80234534],"category_scores_gemma":[0.14311141,0.0016905433,0.0035799043,0.0063193897,0.0006627496,0.0023494978,0.001438406,0.002092207,0.056927945],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0083632795,0.00035681145,0.002393572,0.12453847,0.0019968434,0.00011813899,0.00023251736,0.0013198978,0.00024044747,0.004509986,0.8332626,0.02266737],"study_design_scores_gemma":[0.20068982,0.002624086,0.025841922,0.056455277,0.008504492,0.0007102121,0.00057101797,0.0039080465,0.0012853366,0.03391967,0.66501427,0.0004757442],"about_ca_topic_score_codex":0.0032158743,"about_ca_topic_score_gemma":0.0061626774,"teacher_disagreement_score":0.9981051,"about_ca_system_score_codex":0.0019134744,"about_ca_system_score_gemma":0.004695314,"threshold_uncertainty_score":0.2819304},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"not_applicable","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W6920678091","doi":"10.6084/m9.figshare.16916544.v1","title":"Additional file 3 of Patient-important outcomes reported in randomized controlled trials of pharmacologic treatments for COVID-19: a protocol of a META-epidemiological study","year":2021,"lang":"en","type":"article","venue":"Figshare","topic":"COVID-19 Clinical Research Studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; McMaster University; Université de Montréal; University of Saskatchewan","funders":"","keywords":"Randomized controlled trial; Protocol (science); MEDLINE; Clinical trial; N of 1 trial","score_opus":0.410528904978086,"score_gpt":0.5647141059089654,"score_spread":0.15418520093087945,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6920678091","genre_codex":"dataset","genre_gemma":"protocol","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00044287552,0.00031769444,0.0014734744,0.00047698454,0.00015650949,0.011663609,0.9822673,0.00048225743,0.0027191606],"genre_scores_gemma":[0.025668332,0.0016523163,0.034010276,0.004512964,0.00064322853,0.42539495,0.47163874,0.0016981777,0.034781042],"study_design_codex":"not_applicable","study_design_gemma":"systematic_review","domain_scores_codex":[0.9908958,0.004106529,0.0026730835,0.00092602056,0.0009278314,0.00047069334],"domain_scores_gemma":[0.8290604,0.14628324,0.009367921,0.006179527,0.007834271,0.001274553],"candidate_categories":["metaresearch","metaepi_broad","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.015116047,0.002009704,0.0033380445,0.0040817903,0.0012037463,0.0028081576,0.002080824,0.0025634018,0.82635987],"category_scores_gemma":[0.15213202,0.0018599684,0.004236817,0.006671529,0.00074000476,0.0025823899,0.0017052299,0.0022172905,0.06520923],"study_design_candidate":"systematic_review","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.008882603,0.0003461334,0.0021789167,0.13383476,0.0020982935,0.00012996538,0.00024210071,0.0013520411,0.00027698014,0.0043954966,0.8248021,0.021460488],"study_design_scores_gemma":[0.19781025,0.00225498,0.022147378,0.05368725,0.007734853,0.0006213941,0.0005699204,0.0036300684,0.0013172877,0.035756707,0.673992,0.00047796094],"about_ca_topic_score_codex":0.003715893,"about_ca_topic_score_gemma":0.0075659435,"teacher_disagreement_score":0.99666196,"about_ca_system_score_codex":0.0022437433,"about_ca_system_score_gemma":0.0051180837,"threshold_uncertainty_score":0.24767649},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"not_applicable","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W6920752056","doi":"10.6084/m9.figshare.14913276.v1","title":"Additional file 5 of Validity of observational evidence on putative risk and protective factors: appraisal of 3744 meta-analyses on 57 topics","year":2021,"lang":"en","type":"article","venue":"Figshare","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Observational study; Credibility; Table (database); Kappa; Measure (data warehouse); Critical appraisal","score_opus":0.9569077944506005,"score_gpt":0.5921725702233445,"score_spread":0.364735224227256,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6920752056","genre_codex":"dataset","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00028409093,0.00016100057,0.0005220245,0.00024770462,0.00004796682,0.0005128532,0.9971,0.0002454097,0.0008789478],"genre_scores_gemma":[0.027305678,0.0015017372,0.020396952,0.0022705707,0.00047185918,0.034021206,0.89055,0.0017606078,0.021721374],"study_design_codex":"not_applicable","study_design_gemma":"meta_analysis","domain_scores_codex":[0.99501437,0.0014785589,0.0017987051,0.0006587882,0.00075382343,0.00029583674],"domain_scores_gemma":[0.7930912,0.18156998,0.010352058,0.004278956,0.009821464,0.00088626484],"candidate_categories":["metaresearch","metaepi_broad","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.008642506,0.0017662545,0.0027443366,0.0068086633,0.0010374784,0.002815404,0.0022643546,0.0020362253,0.90425235],"category_scores_gemma":[0.1500346,0.001289437,0.003624817,0.008632267,0.00046239857,0.0040036915,0.0021179002,0.001251458,0.07739797],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017982996,0.00016222433,0.0037557771,0.10596849,0.0009962829,0.00015146083,0.00024799726,0.0010624068,0.00020278312,0.0025610465,0.8607245,0.022368776],"study_design_scores_gemma":[0.049930964,0.0009762185,0.0419689,0.06983259,0.0053451005,0.0009981692,0.0010417256,0.0044427873,0.0013898587,0.035587773,0.7880089,0.00047698547],"about_ca_topic_score_codex":0.005338815,"about_ca_topic_score_gemma":0.012279127,"teacher_disagreement_score":0.9972557,"about_ca_system_score_codex":0.00237909,"about_ca_system_score_gemma":0.0040360163,"threshold_uncertainty_score":0.13657242},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"not_applicable","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W6920871351","doi":"10.6084/m9.figshare.16916541","title":"Additional file 2 of Patient-important outcomes reported in randomized controlled trials of pharmacologic treatments for COVID-19: a protocol of a META-epidemiological study","year":2021,"lang":"en","type":"article","venue":"Figshare","topic":"COVID-19 Clinical Research Studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; McMaster University; Université de Montréal; University of Saskatchewan","funders":"","keywords":"Randomized controlled trial; Protocol (science); MEDLINE; Clinical trial; N of 1 trial","score_opus":0.40676208952822795,"score_gpt":0.5646415612489036,"score_spread":0.15787947172067568,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6920871351","genre_codex":"dataset","genre_gemma":"protocol","domain_codex":null,"domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00049061584,0.00034489948,0.0015199143,0.00048727807,0.00016269664,0.01235978,0.9815838,0.00043145297,0.002619514],"genre_scores_gemma":[0.026899846,0.0016455483,0.03451596,0.0042895307,0.0006628569,0.43660975,0.46153942,0.0014607719,0.03237642],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9915929,0.003850082,0.0023914676,0.0009027921,0.00082662445,0.00043622692],"domain_scores_gemma":[0.8410734,0.13579398,0.008688607,0.0059428676,0.0072552427,0.0012458093],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.014373383,0.0019031587,0.003206111,0.0039185667,0.0011302674,0.00262537,0.0019774344,0.002526049,0.8097869],"category_scores_gemma":[0.14498551,0.001812268,0.0039797286,0.0064194286,0.0007022848,0.0024172247,0.0015589857,0.0021516667,0.05801323],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0088602025,0.0003586343,0.0022557788,0.13213214,0.002052586,0.00013542129,0.00024481132,0.0013649205,0.00027747743,0.0046857214,0.8256031,0.022029249],"study_design_scores_gemma":[0.19992381,0.0023423154,0.022969695,0.05236276,0.0077631995,0.0006454718,0.00054456596,0.00364414,0.0012621934,0.03416694,0.67391276,0.0004621416],"about_ca_topic_score_codex":0.0033542328,"about_ca_topic_score_gemma":0.006692949,"teacher_disagreement_score":0.9980968,"about_ca_system_score_codex":0.0020178806,"about_ca_system_score_gemma":0.0048989207,"threshold_uncertainty_score":0.2713158},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"not_applicable","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W6927340553","doi":"10.3389/fnagi.2023.1051809.s001","title":"Table_1_Gout and risk of dementia, Alzheimer's disease or vascular dementia: a meta-epidemiology study.DOCX","year":2023,"lang":"en","type":"dataset","venue":"Figshare","topic":"Bacillus and Francisella bacterial research","field":"Biochemistry, Genetics and Molecular Biology","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Gout; Cohort; Confidence interval; Cohort study; Risk assessment; Disease; Dementia","score_opus":0.10522562552551656,"score_gpt":0.3401329899959581,"score_spread":0.23490736447044153,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6927340553","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.002261192,0.014356366,0.0019455893,0.0015788608,0.001088609,0.007400568,0.96787274,0.0007559222,0.0027401892],"genre_scores_gemma":[0.08440516,0.040234067,0.04537093,0.01592745,0.00316644,0.15878683,0.5830683,0.002516858,0.06652395],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.998268,0.0004960267,0.00041132022,0.00030210003,0.00036006424,0.00016241947],"domain_scores_gemma":[0.9740646,0.020248957,0.002724202,0.0006796944,0.0020534194,0.00022916147],"candidate_categories":["metaepi_broad","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.005102771,0.003425786,0.0063635525,0.0051584584,0.00091986393,0.0025971443,0.0029814593,0.0018172598,0.49607602],"category_scores_gemma":[0.042900775,0.0020118284,0.013799276,0.0074059474,0.0004711741,0.0045906072,0.0014715708,0.0019675547,0.014650123],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.003525637,0.00012714973,0.0054979096,0.5680532,0.01869156,0.00025292783,0.0001639852,0.0017148312,0.00061077165,0.0021419001,0.38516107,0.014059046],"study_design_scores_gemma":[0.07459812,0.0018070138,0.08721468,0.21370597,0.089570545,0.0020891277,0.0008088466,0.008049321,0.002611995,0.023436137,0.49533886,0.0007693575],"about_ca_topic_score_codex":0.010445954,"about_ca_topic_score_gemma":0.017189432,"teacher_disagreement_score":0.9936364,"about_ca_system_score_codex":0.0020825705,"about_ca_system_score_gemma":0.0049021365,"threshold_uncertainty_score":0.71878636},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"not_applicable","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"not_applicable","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W6939482417","doi":"10.6084/m9.figshare.14913270","title":"Additional file 3 of Validity of observational evidence on putative risk and protective factors: appraisal of 3744 meta-analyses on 57 topics","year":2021,"lang":"en","type":"article","venue":"Figshare","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Observational study; Missing data; Data collection; Observational methods in psychology; Risk assessment","score_opus":0.9569709275465907,"score_gpt":0.5920804522400188,"score_spread":0.3648904753065719,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6939482417","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00021274229,0.00014264329,0.0004285725,0.00022818409,0.000054024204,0.00038178966,0.99743515,0.00021956496,0.0008974428],"genre_scores_gemma":[0.022968724,0.0013798239,0.015295933,0.0024593915,0.00049812417,0.026773302,0.9054222,0.0016782857,0.023524184],"study_design_codex":"not_applicable","study_design_gemma":"meta_analysis","domain_scores_codex":[0.996247,0.0010952014,0.0011803901,0.0005791899,0.0006211632,0.00027707877],"domain_scores_gemma":[0.85434705,0.12669146,0.007326185,0.003614591,0.007251516,0.0007692221],"candidate_categories":["metaresearch","metaepi_broad","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.007113554,0.0019372847,0.0027967866,0.0058728503,0.0009663113,0.0026748437,0.0026295634,0.0022631665,0.9098527],"category_scores_gemma":[0.123322725,0.0012881978,0.0036032572,0.0079911165,0.0004596482,0.0039626765,0.001914325,0.0014087819,0.09178832],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0016553532,0.00013706094,0.0028168792,0.07298951,0.0008987121,0.00013304486,0.00014727663,0.0011620802,0.00017408974,0.002408234,0.90146124,0.016016493],"study_design_scores_gemma":[0.058029972,0.0008846583,0.030622095,0.054776516,0.00504292,0.0009563796,0.0007333767,0.0056248773,0.0012306799,0.037441697,0.80423766,0.00041919842],"about_ca_topic_score_codex":0.0065481956,"about_ca_topic_score_gemma":0.011704184,"teacher_disagreement_score":0.99720323,"about_ca_system_score_codex":0.0022090047,"about_ca_system_score_gemma":0.0038608063,"threshold_uncertainty_score":0.1285842},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"not_applicable","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W6939574047","doi":"10.6084/m9.figshare.14746350","title":"Additional file 1 of Bias estimation in study design: a meta-epidemiological analysis of transcatheter versus surgical aortic valve replacement","year":2021,"lang":"en","type":"article","venue":"Open MIND","topic":"Cardiac Valve Diseases and Treatments","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Women's College Hospital; University of Toronto; Western University","funders":"","keywords":"Table (database); Chart; Aortic valve replacement; Flow chart; Valve replacement; Random error","score_opus":0.20481244366084914,"score_gpt":0.43322146448937443,"score_spread":0.2284090208285253,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6939574047","genre_codex":"dataset","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00049977563,0.00044032276,0.0016555173,0.00028857082,0.00009492037,0.0028586762,0.9928048,0.00041680533,0.00094056805],"genre_scores_gemma":[0.052505426,0.0036027823,0.07610598,0.0038687391,0.00086358393,0.2318238,0.5980187,0.0032657234,0.029945362],"study_design_codex":"not_applicable","study_design_gemma":"meta_analysis","domain_scores_codex":[0.99284667,0.0023289789,0.002504446,0.00097150105,0.00091507926,0.000433271],"domain_scores_gemma":[0.791341,0.18509185,0.011623109,0.0043238024,0.0068910993,0.00072914356],"candidate_categories":["metaresearch","metaepi_broad","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.012779776,0.0027182915,0.005098311,0.006863557,0.0010880699,0.0028902723,0.0028796552,0.0025328323,0.85100526],"category_scores_gemma":[0.15919258,0.001636222,0.0056783124,0.0094707655,0.0005811398,0.0037125722,0.0018894279,0.001681386,0.04291028],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.003988631,0.0002897191,0.005500742,0.3938678,0.0034054543,0.0002565336,0.0003729887,0.0026695929,0.0004052644,0.004888362,0.55262893,0.03172597],"study_design_scores_gemma":[0.09070505,0.0022686922,0.044496253,0.16734523,0.019922722,0.001784484,0.0011223956,0.016114054,0.0025071441,0.055981494,0.59681463,0.00093779975],"about_ca_topic_score_codex":0.0048739766,"about_ca_topic_score_gemma":0.0103675155,"teacher_disagreement_score":0.9949017,"about_ca_system_score_codex":0.0025702298,"about_ca_system_score_gemma":0.0049615414,"threshold_uncertainty_score":0.21252292},"labels":[{"model":"gemma","categories":["metaepi_broad","metaepi_narrow","metaresearch"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W6939662579","doi":"10.6084/m9.figshare.14788632","title":"Additional file 1 of Most published meta-regression analyses based on aggregate data suffer from methodological pitfalls: a meta-epidemiological study","year":2021,"lang":"en","type":"article","venue":"Open MIND","topic":"Mycorrhizal Fungi and Plant Interactions","field":"Agricultural and Biological Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Aggregate (composite); Aggregate data; Data file; Component (thermodynamics); Data collection","score_opus":0.5631933366838925,"score_gpt":0.43149222169421486,"score_spread":0.13170111498967763,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6939662579","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00067302084,0.00068576634,0.0014952181,0.0002856007,0.00009463174,0.0006812478,0.9949256,0.00033460793,0.000824199],"genre_scores_gemma":[0.09111227,0.0028725106,0.06090114,0.0065216785,0.00090150174,0.06069949,0.74487126,0.003043635,0.029076574],"study_design_codex":"not_applicable","study_design_gemma":"observational","domain_scores_codex":[0.99216485,0.0030460423,0.0024510466,0.0012120503,0.00075736316,0.00036874096],"domain_scores_gemma":[0.7272968,0.24260475,0.015678728,0.0073817144,0.006143781,0.00089423155],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.01300652,0.0021693036,0.0037733046,0.0059318678,0.0009451088,0.003039411,0.0032321527,0.0027227406,0.7853126],"category_scores_gemma":[0.19293463,0.0019680883,0.0066926833,0.008990895,0.00069125986,0.0028706098,0.0020696416,0.0016945589,0.045602366],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0036063376,0.00026382302,0.009233404,0.22755525,0.013494711,0.0003284126,0.00031589155,0.002164332,0.00028450458,0.0032038346,0.71543664,0.024112962],"study_design_scores_gemma":[0.10398889,0.0016305681,0.05856896,0.075957105,0.060352117,0.0025523994,0.001019418,0.008657867,0.001684904,0.036047168,0.6487976,0.00074295525],"about_ca_topic_score_codex":0.005440091,"about_ca_topic_score_gemma":0.011964034,"teacher_disagreement_score":0.9978307,"about_ca_system_score_codex":0.00148414,"about_ca_system_score_gemma":0.0034547613,"threshold_uncertainty_score":0.30622548},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"observational","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"not_applicable","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W6939682473","doi":"10.6084/m9.figshare.16916538","title":"Additional file 1 of Patient-important outcomes reported in randomized controlled trials of pharmacologic treatments for COVID-19: a protocol of a META-epidemiological study","year":2021,"lang":"en","type":"article","venue":"Figshare","topic":"Mycorrhizal Fungi and Plant Interactions","field":"Agricultural and Biological Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; McMaster University; Université de Montréal; University of Saskatchewan","funders":"","keywords":"Randomized controlled trial; Protocol (science); MEDLINE; Clinical trial; N of 1 trial","score_opus":0.1812265184946321,"score_gpt":0.391598901425066,"score_spread":0.2103723829304339,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6939682473","genre_codex":"dataset","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0009750393,0.0005751092,0.0020013968,0.0004518001,0.00013942685,0.011182694,0.9820671,0.0003734189,0.002234034],"genre_scores_gemma":[0.051038094,0.0020632125,0.034404915,0.00349652,0.00051551056,0.46347266,0.41744873,0.0010443943,0.026515896],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99444675,0.0026062902,0.0015314212,0.0006442244,0.0004832296,0.00028806346],"domain_scores_gemma":[0.8885286,0.09557397,0.0071659666,0.0037268433,0.0042468607,0.00075772585],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.011440062,0.0016355966,0.0028236532,0.0027026013,0.00084837055,0.0018370968,0.0015476297,0.0019862046,0.7143981],"category_scores_gemma":[0.10415587,0.0012542497,0.0035313948,0.004579871,0.000512666,0.0017891744,0.0011537343,0.0018236553,0.032458305],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.015942572,0.00049604574,0.00404303,0.18886004,0.004763677,0.000171196,0.00029982862,0.002311338,0.0003921592,0.006026959,0.74582815,0.0308651],"study_design_scores_gemma":[0.30123258,0.003796177,0.03734268,0.0651675,0.015934935,0.0007166662,0.0006176792,0.0063349893,0.0015362381,0.03585039,0.5309594,0.0005106695],"about_ca_topic_score_codex":0.0029109325,"about_ca_topic_score_gemma":0.0056512766,"teacher_disagreement_score":0.2856019,"about_ca_system_score_codex":0.0016647681,"about_ca_system_score_gemma":0.003761815,"threshold_uncertainty_score":0.40737647},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"not_applicable","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W6958558395","doi":"10.6084/m9.figshare.14746350.v1","title":"Additional file 1 of Bias estimation in study design: a meta-epidemiological analysis of transcatheter versus surgical aortic valve replacement","year":2021,"lang":"en","type":"article","venue":"Figshare","topic":"Prenatal Screening and Diagnostics","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Women's College Hospital; University of Toronto; Western University","funders":"","keywords":"Table (database); Chart; Aortic valve replacement; Flow chart; Valve replacement; Random error","score_opus":0.2528468836712392,"score_gpt":0.36584832081618757,"score_spread":0.11300143714494837,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6958558395","genre_codex":"dataset","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0004548268,0.00044980025,0.001534284,0.00026981413,0.0000887148,0.0025420815,0.99329454,0.00041779788,0.0009481684],"genre_scores_gemma":[0.05103615,0.003636853,0.07334825,0.0039028495,0.0008658693,0.21167904,0.6197294,0.0036612498,0.03214033],"study_design_codex":"not_applicable","study_design_gemma":"meta_analysis","domain_scores_codex":[0.99307656,0.0023129885,0.002322783,0.0009598162,0.00089856033,0.00042925205],"domain_scores_gemma":[0.79362196,0.18383706,0.011128418,0.0041467114,0.006522828,0.0007429888],"candidate_categories":["metaresearch","metaepi_broad","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.01197555,0.0026934107,0.0049664048,0.006972843,0.0010388829,0.00285342,0.0028837807,0.0025027927,0.8585819],"category_scores_gemma":[0.154848,0.0016262842,0.0057490394,0.009478409,0.00058798527,0.0037524037,0.0018923222,0.0016747309,0.04795489],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0038542172,0.00027310703,0.005223229,0.37701762,0.0034510947,0.00024373415,0.000351206,0.0026476108,0.00038555957,0.0047178697,0.5709903,0.03084451],"study_design_scores_gemma":[0.08731697,0.0022126308,0.043297697,0.16077043,0.01921097,0.0017981828,0.0010535108,0.01538966,0.0023431063,0.055508893,0.6101963,0.00090172596],"about_ca_topic_score_codex":0.004759241,"about_ca_topic_score_gemma":0.010077316,"teacher_disagreement_score":0.9950336,"about_ca_system_score_codex":0.0024884937,"about_ca_system_score_gemma":0.0046902713,"threshold_uncertainty_score":0.20171571},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"observational","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"design_other","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W6958604890","doi":"10.6084/m9.figshare.18739501.v1","title":"Additional file 2 of Omega-6 sparing effects of parenteral lipid emulsions—an updated systematic review and meta-analysis on clinical outcomes in critically ill patients","year":2022,"lang":"en","type":"article","venue":"Figshare","topic":"Clinical Nutrition and Gastroenterology","field":"Nursing","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Kingston General Hospital; Queen's University","funders":"","keywords":"Critically ill; Critical illness; MEDLINE; Clinical trial; Sepsis","score_opus":0.07060435332087839,"score_gpt":0.3580250775955709,"score_spread":0.2874207242746925,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6958604890","genre_codex":"dataset","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"review","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000670979,0.0013707723,0.00026632633,0.00037974995,0.00008869178,0.0005459749,0.99542147,0.00016790663,0.0010881154],"genre_scores_gemma":[0.04573514,0.008778902,0.013817133,0.0048516374,0.0009088972,0.03131365,0.85359025,0.0010503012,0.03995405],"study_design_codex":"not_applicable","study_design_gemma":"meta_analysis","domain_scores_codex":[0.99844116,0.00030582753,0.00057470886,0.0002572935,0.00028239307,0.00013846626],"domain_scores_gemma":[0.9628628,0.028932426,0.004365081,0.00073512073,0.0027828016,0.00032178784],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002715276,0.0014069973,0.0033187873,0.004091148,0.0005533047,0.0018664402,0.0016335758,0.0013391299,0.74586433],"category_scores_gemma":[0.046960123,0.00080629374,0.004276405,0.0074465713,0.00027879333,0.0029855769,0.0012019628,0.0009797409,0.028369568],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0034776472,0.00011436723,0.0042812475,0.46801126,0.0036412042,0.00021840392,0.00015745933,0.0007054089,0.00045601625,0.0019309652,0.49721313,0.019792855],"study_design_scores_gemma":[0.06234089,0.0012906445,0.066521935,0.19599555,0.026656765,0.0014644556,0.00071731536,0.0028954856,0.0014578269,0.017054344,0.623162,0.00044270093],"about_ca_topic_score_codex":0.0043919026,"about_ca_topic_score_gemma":0.010723851,"teacher_disagreement_score":0.74586433,"about_ca_system_score_codex":0.0012578334,"about_ca_system_score_gemma":0.0023787427,"threshold_uncertainty_score":0},"labels":[{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"not_applicable","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W6959115587","doi":"10.6084/m9.figshare.c.6842755","title":"Sleep disturbances and the risk of lung cancer: a meta-epidemiological study","year":2023,"lang":"en","type":"other","venue":"Figshare","topic":"","field":"","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Sleep (system call); Funnel plot; Chronotype; Odds ratio; Confidence interval; Lung cancer; Incidence (geometry); Observational study; Meta-analysis","score_opus":0.10701209725830152,"score_gpt":0.35819529942424977,"score_spread":0.25118320216594825,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6959115587","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.014021232,0.9825843,0.0016402943,0.00030774076,0.00028150206,0.00027580993,0.00059648254,0.000034662495,0.00025793555],"genre_scores_gemma":[0.3717545,0.6168752,0.0063516917,0.0011212798,0.00062575366,0.0014061705,0.0013674002,0.00005175788,0.00044616274],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9898957,0.0052843713,0.0023656837,0.0012099842,0.00090079237,0.00034344062],"domain_scores_gemma":[0.9806368,0.014003221,0.0026096958,0.0010481143,0.0014069944,0.00029524145],"candidate_categories":["metaresearch","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.015502002,0.0032591685,0.013689716,0.0073096706,0.00079920737,0.0039013044,0.0022048801,0.002516341,0.002734824],"category_scores_gemma":[0.032138553,0.0018846212,0.05166515,0.007601084,0.00065595296,0.0016984215,0.0017526271,0.0026210297,0.00026421066],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019082507,0.00005732368,0.013696456,0.08986158,0.8870114,0.00016985256,0.00006235234,0.0007232677,0.00024650764,0.00011834606,0.00029868959,0.005845987],"study_design_scores_gemma":[0.00032682455,0.0002264329,0.007527374,0.0075524976,0.98267835,0.00012364407,0.000033838893,0.00039254417,0.00011420059,0.00022612237,0.00077560666,0.000022542807],"about_ca_topic_score_codex":0.0052371244,"about_ca_topic_score_gemma":0.008325265,"teacher_disagreement_score":0.9863103,"about_ca_system_score_codex":0.0017825637,"about_ca_system_score_gemma":0.0025636666,"threshold_uncertainty_score":0.08198345},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"observational","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"design_other","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W6976945731","doi":"10.6084/m9.figshare.14788632.v1","title":"Additional file 1 of Most published meta-regression analyses based on aggregate data suffer from methodological pitfalls: a meta-epidemiological study","year":2021,"lang":"en","type":"article","venue":"Figshare","topic":"Geotechnical and construction materials studies","field":"Engineering","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Aggregate (composite); Aggregate data; Data file; Component (thermodynamics); Data collection","score_opus":0.5199242974097767,"score_gpt":0.3867582678563971,"score_spread":0.13316602955337964,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6976945731","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00041390004,0.000464354,0.0010422494,0.0002474314,0.00009345594,0.00062611286,0.99586064,0.00029679682,0.0009551557],"genre_scores_gemma":[0.061587945,0.00239529,0.044906355,0.0066816565,0.00087205885,0.059297595,0.78685564,0.003506117,0.033897355],"study_design_codex":"not_applicable","study_design_gemma":"meta_analysis","domain_scores_codex":[0.99294734,0.002627247,0.0022421898,0.0010808452,0.0007132181,0.0003891945],"domain_scores_gemma":[0.7384537,0.23177427,0.013624614,0.008070537,0.007140592,0.0009362937],"candidate_categories":["metaresearch","metaepi_broad","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.012264007,0.0024740866,0.0036755616,0.0065608174,0.0010583159,0.0031559048,0.0033834423,0.0029998173,0.86248505],"category_scores_gemma":[0.19023702,0.0022114832,0.007032104,0.009811962,0.0007917868,0.003148771,0.0021207095,0.0017742492,0.077285506],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0029716515,0.0002196376,0.0060875267,0.1404572,0.0072780997,0.0002697107,0.00023049192,0.0016334774,0.00019075297,0.0027452516,0.81829304,0.019623214],"study_design_scores_gemma":[0.10625547,0.0013908796,0.046007015,0.06710187,0.035957888,0.0023336844,0.0009685481,0.006052773,0.0013286335,0.036459286,0.6954559,0.00068798603],"about_ca_topic_score_codex":0.0056563835,"about_ca_topic_score_gemma":0.011948887,"teacher_disagreement_score":0.9963244,"about_ca_system_score_codex":0.0016032569,"about_ca_system_score_gemma":0.0035342434,"threshold_uncertainty_score":0.19614834},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"not_applicable","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"not_applicable","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W6977116624","doi":"10.6084/m9.figshare.16916541.v1","title":"Additional file 2 of Patient-important outcomes reported in randomized controlled trials of pharmacologic treatments for COVID-19: a protocol of a META-epidemiological study","year":2021,"lang":"en","type":"article","venue":"Figshare","topic":"COVID-19 Clinical Research Studies","field":"Medicine","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; McMaster University; Université de Montréal; University of Saskatchewan","funders":"","keywords":"Randomized controlled trial; Protocol (science); MEDLINE; Clinical trial; N of 1 trial","score_opus":0.40676208952822795,"score_gpt":0.5646415612489036,"score_spread":0.15787947172067568,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6977116624","genre_codex":"dataset","genre_gemma":"protocol","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"protocol","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00049061584,0.00034489948,0.0015199143,0.00048727807,0.00016269664,0.01235978,0.9815838,0.00043145297,0.002619514],"genre_scores_gemma":[0.026899846,0.0016455483,0.03451596,0.0042895307,0.0006628569,0.43660975,0.46153942,0.0014607719,0.03237642],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9915929,0.003850082,0.0023914676,0.0009027921,0.00082662445,0.00043622692],"domain_scores_gemma":[0.8410734,0.13579398,0.008688607,0.0059428676,0.0072552427,0.0012458093],"candidate_categories":["metaresearch","metaepi_broad","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.014373383,0.0019031587,0.003206111,0.0039185667,0.0011302674,0.00262537,0.0019774344,0.002526049,0.8097869],"category_scores_gemma":[0.14498551,0.001812268,0.0039797286,0.0064194286,0.0007022848,0.0024172247,0.0015589857,0.0021516667,0.05801323],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0088602025,0.0003586343,0.0022557788,0.13213214,0.002052586,0.00013542129,0.00024481132,0.0013649205,0.00027747743,0.0046857214,0.8256031,0.022029249],"study_design_scores_gemma":[0.19992381,0.0023423154,0.022969695,0.05236276,0.0077631995,0.0006454718,0.00054456596,0.00364414,0.0012621934,0.03416694,0.67391276,0.0004621416],"about_ca_topic_score_codex":0.0033542328,"about_ca_topic_score_gemma":0.006692949,"teacher_disagreement_score":0.99679387,"about_ca_system_score_codex":0.0020178806,"about_ca_system_score_gemma":0.0048989207,"threshold_uncertainty_score":0.2713158},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"not_applicable","genre":"protocol","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"not_applicable","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W6977629057","doi":"10.6084/m9.figshare.c.6842755.v1","title":"Sleep disturbances and the risk of lung cancer: a meta-epidemiological study","year":2023,"lang":"en","type":"other","venue":"Figshare","topic":"Computational Physics and Python Applications","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Sleep (system call); Funnel plot; Chronotype; Odds ratio; Confidence interval; Lung cancer; Incidence (geometry); Observational study; Meta-analysis","score_opus":0.060412125168337064,"score_gpt":0.32732866330819943,"score_spread":0.2669165381398624,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6977629057","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.014021232,0.9825843,0.0016402943,0.00030774076,0.00028150206,0.00027580993,0.00059648254,0.000034662495,0.00025793555],"genre_scores_gemma":[0.3717545,0.6168752,0.0063516917,0.0011212798,0.00062575366,0.0014061705,0.0013674002,0.00005175788,0.00044616274],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.9898957,0.0052843713,0.0023656837,0.0012099842,0.00090079237,0.00034344062],"domain_scores_gemma":[0.9806368,0.014003221,0.0026096958,0.0010481143,0.0014069944,0.00029524145],"candidate_categories":["metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.015502002,0.0032591685,0.013689716,0.0073096706,0.00079920737,0.0039013044,0.0022048801,0.002516341,0.002734824],"category_scores_gemma":[0.032138553,0.0018846212,0.05166515,0.007601084,0.00065595296,0.0016984215,0.0017526271,0.0026210297,0.00026421066],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019082507,0.00005732368,0.013696456,0.08986158,0.8870114,0.00016985256,0.00006235234,0.0007232677,0.00024650764,0.00011834606,0.00029868959,0.005845987],"study_design_scores_gemma":[0.00032682455,0.0002264329,0.007527374,0.0075524976,0.98267835,0.00012364407,0.000033838893,0.00039254417,0.00011420059,0.00022612237,0.00077560666,0.000022542807],"about_ca_topic_score_codex":0.0052371244,"about_ca_topic_score_gemma":0.008325265,"teacher_disagreement_score":0.9863103,"about_ca_system_score_codex":0.0017825637,"about_ca_system_score_gemma":0.0025636666,"threshold_uncertainty_score":0.08198345},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"meta_analysis","genre":"other","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W6977951747","doi":"10.7939/dvn/10468","title":"PEDro or Cochrane to Assess the Quality of Clinical Trials? A meta-epidemiological study","year":2015,"lang":"en","type":"dataset","venue":"Borealis","topic":"","field":"","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Quality (philosophy); Clinical trial; Scale (ratio); MEDLINE; Meta-analysis; Quality of evidence","score_opus":0.8582710749683431,"score_gpt":0.6583148020165382,"score_spread":0.19995627295180485,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6977951747","genre_codex":"dataset","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015712366,0.29838595,0.028630715,0.005491802,0.0029968384,0.030664409,0.6059511,0.0022571394,0.009909738],"genre_scores_gemma":[0.3128027,0.09650423,0.11482588,0.011160962,0.0016665508,0.26843795,0.1865525,0.0023917493,0.0056574326],"study_design_codex":"systematic_review","study_design_gemma":"meta_analysis","domain_scores_codex":[0.89915866,0.04798321,0.03732089,0.007801958,0.0068202796,0.00091507623],"domain_scores_gemma":[0.82759964,0.10128712,0.041952245,0.018556194,0.009007803,0.0015969978],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.0520113,0.0022006973,0.016571848,0.012345189,0.00068701844,0.004756868,0.0029933113,0.002439018,0.032607596],"category_scores_gemma":[0.25518882,0.0017560909,0.028264174,0.013902942,0.0011571242,0.0035584997,0.0046765236,0.0029150473,0.0025896525],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.01251152,0.0000499177,0.009345303,0.5998769,0.30236006,0.00009939625,0.00023058851,0.0008114486,0.00042025518,0.0027307447,0.049690068,0.021873882],"study_design_scores_gemma":[0.064778134,0.0009530353,0.033857454,0.12798068,0.61008376,0.0006265783,0.00025154446,0.0018330197,0.0007999126,0.012765023,0.14565687,0.00041398636],"about_ca_topic_score_codex":0.0043052463,"about_ca_topic_score_gemma":0.008384814,"teacher_disagreement_score":0.9977993,"about_ca_system_score_codex":0.00213463,"about_ca_system_score_gemma":0.0036611594,"threshold_uncertainty_score":0.2750653},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"evaluation","study_design":"not_applicable","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"not_applicable","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W6996152782","doi":"","title":"Relationship between the EQ-5D index and measures of clinical outcomes in selected studies of cardiovascular interventions","year":2009,"lang":"en","type":"article","venue":"DOAJ (DOAJ: Directory of Open Access Journals)","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Angina; Observational study; Psychological intervention; Disease; Index (typography); Ceiling effect; Canadian Cardiovascular Society; Linear regression","score_opus":0.9223982087177003,"score_gpt":0.6821050020118999,"score_spread":0.24029320670580034,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6996152782","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.78654534,0.18908724,0.008010426,0.002129775,0.00028342655,0.0004987149,0.01030906,0.000037629667,0.003098439],"genre_scores_gemma":[0.99198735,0.0032294472,0.0018277455,0.00030303636,0.00016802714,0.0004431337,0.0019349841,0.000009136273,0.000097074946],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9033887,0.068927966,0.014077737,0.005441421,0.007310674,0.0008534415],"domain_scores_gemma":[0.692152,0.2387344,0.052847866,0.0073602013,0.007630728,0.0012748172],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.06523141,0.00060374907,0.0017023998,0.0055500274,0.0005163035,0.0022193887,0.0012183777,0.0010954577,0.002650933],"category_scores_gemma":[0.13118355,0.0004948494,0.00491261,0.00963337,0.000997733,0.001419981,0.0020029624,0.0015464482,0.00017424286],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00095093716,0.000042029256,0.97576255,0.0020131366,0.014033154,0.000040898867,0.00011333925,0.00046673598,0.00006770132,0.00027224663,0.00034557274,0.0058917124],"study_design_scores_gemma":[0.00027747347,0.0005802238,0.98802537,0.00088247686,0.005809644,0.00025703668,0.00017130059,0.0015856016,0.00015638172,0.0008223722,0.001399348,0.000032794836],"about_ca_topic_score_codex":0.001580912,"about_ca_topic_score_gemma":0.0022245115,"teacher_disagreement_score":0.06523141,"about_ca_system_score_codex":0.001055573,"about_ca_system_score_gemma":0.00090539304,"threshold_uncertainty_score":0.34498084},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_broad"],"domain":null,"study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W7075305078","doi":"","title":"Awareness of Implicit Attitudes Revisited: A Meta-Analysis on Replications Across Samples and Settings","year":2024,"lang":"en","type":"other","venue":"OSF Preprints (OSF Preprints)","topic":"Diverse Scientific and Economic Studies","field":"Economics, Econometrics and Finance","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Implicit-association test; Consistency (knowledge bases); Replication (statistics); Correlation; Cognition; Association (psychology); Social cognition","score_opus":0.08092403851394613,"score_gpt":0.30419126661776064,"score_spread":0.2232672281038145,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7075305078","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"reproducibility","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reproducibility","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.18983333,0.7844564,0.017703114,0.001051658,0.0010600572,0.0014725067,0.001804177,0.00024503213,0.0023737827],"genre_scores_gemma":[0.9510337,0.038396273,0.00677237,0.000769265,0.00026128656,0.0011246656,0.0012310152,0.00012470176,0.00028675565],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.92333895,0.045152225,0.016083201,0.009040885,0.005821198,0.00056355086],"domain_scores_gemma":[0.75452,0.1882752,0.025629148,0.020986192,0.009773054,0.0008163885],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.08909145,0.002401564,0.008282949,0.00483012,0.000939217,0.005019719,0.0036369108,0.0022126492,0.0024775914],"category_scores_gemma":[0.14514738,0.0016389561,0.028475726,0.0056005083,0.0027884236,0.0028960346,0.002274931,0.0021458268,0.00033794547],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019861436,0.00008844469,0.042376194,0.05529887,0.88412917,0.0001765509,0.00052726583,0.00035299774,0.0008349657,0.00033405394,0.00034017235,0.0135551505],"study_design_scores_gemma":[0.0007644021,0.0008013987,0.041678265,0.010664242,0.93998474,0.0002767132,0.00038988056,0.00070660457,0.0012818209,0.0010202948,0.0023668683,0.00006469007],"about_ca_topic_score_codex":0.0019928445,"about_ca_topic_score_gemma":0.0034136125,"teacher_disagreement_score":0.9109086,"about_ca_system_score_codex":0.0012868097,"about_ca_system_score_gemma":0.0015797527,"threshold_uncertainty_score":0.47116625},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"reproducibility","study_design":"meta_analysis","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reproducibility","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W7110863974","doi":"10.3389/fimmu.2025.1613441.s001","title":"Table 1_Air pollution, emerging chemical exposures, and systemic lupus erythematosus: a meta-epidemiology study.docx","year":2025,"lang":"","type":"dataset","venue":"Figshare","topic":"","field":"","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Incidence (geometry); Air pollutants; Observational study; Funnel plot; Systemic lupus erythematosus; Statistical analysis; Meta-analysis","score_opus":0.0734439943976077,"score_gpt":0.3210556761698789,"score_spread":0.24761168177227122,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7110863974","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004111883,0.0153315095,0.0029789181,0.0011482436,0.00087086717,0.003752636,0.9693945,0.0007222387,0.0016891541],"genre_scores_gemma":[0.15616554,0.052180007,0.06156409,0.009332454,0.0032327839,0.08811837,0.5775085,0.0024200648,0.049478192],"study_design_codex":"systematic_review","study_design_gemma":"not_applicable","domain_scores_codex":[0.9978536,0.0006884006,0.0005578274,0.00035799688,0.00035973822,0.00018250807],"domain_scores_gemma":[0.97675747,0.019225044,0.0019025012,0.00059362006,0.0013127471,0.00020865089],"candidate_categories":["metaepi_broad","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0053442977,0.0035294322,0.0059589525,0.0068707275,0.00075887324,0.0022538719,0.0030388176,0.0014866939,0.33577114],"category_scores_gemma":[0.039111592,0.0018099999,0.013761308,0.009252435,0.00038460802,0.0031552028,0.0014576794,0.0017727646,0.009517382],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.005931149,0.00019847724,0.021117035,0.544574,0.050120037,0.00060323265,0.00018249633,0.002991152,0.0009115954,0.001819833,0.34868583,0.022865197],"study_design_scores_gemma":[0.054133646,0.0028241845,0.15833434,0.112774394,0.226434,0.0030956413,0.0009098847,0.011462179,0.0039058728,0.017851565,0.40761995,0.00065441354],"about_ca_topic_score_codex":0.007631369,"about_ca_topic_score_gemma":0.0136181,"teacher_disagreement_score":0.994041,"about_ca_system_score_codex":0.0014458114,"about_ca_system_score_gemma":0.0034190514,"threshold_uncertainty_score":0.9474419},"labels":[{"model":"gemma","categories":["metaepi_broad"],"domain":null,"study_design":"observational","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"},{"model":"gpt","categories":["insufficient_payload"],"domain":null,"study_design":"not_applicable","genre":"dataset","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W7117368100","doi":"10.1016/j.jclinepi.2025.112121","title":"Systematic reviews of quasi-experimental studies: challenges and considerations","year":2025,"lang":"en","type":"article","venue":"Journal of Clinical Epidemiology","topic":"Meta-analysis and systematic reviews","field":"Decision Sciences","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University Health Centre; McGill University","funders":"Fonds de Recherche du Québec - Santé; Natural Sciences and Engineering Research Council of Canada; Canadian Institutes of Health Research","keywords":"Systematic review; Preprint; Consistency (knowledge bases); Variety (cybernetics); Transparency (behavior); Grey literature; MEDLINE; Protocol (science); Identification (biology); Comparability","score_opus":0.9664743422323613,"score_gpt":0.7342298440974422,"score_spread":0.2322444981349191,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7117368100","genre_codex":"review","genre_gemma":"methods","domain_codex":"methods","domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":null,"domain_candidate":"methods","domain_consensus":"methods","prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0090421215,0.44376454,0.1921505,0.3119713,0.023241287,0.010520594,0.0022379402,0.00070005393,0.006371663],"genre_scores_gemma":[0.23171571,0.11599259,0.48752382,0.103738025,0.022549648,0.035166997,0.0012564712,0.00066206895,0.0013947184],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.09591648,0.749909,0.093772635,0.013696892,0.045018647,0.0016863982],"domain_scores_gemma":[0.015428485,0.9365114,0.014213015,0.011768314,0.020325234,0.0017535046],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.860829,0.006355839,0.034061465,0.015614877,0.0074458453,0.019225365,0.022827491,0.016657704,0.0048103663],"category_scores_gemma":[0.92579305,0.008331708,0.017052788,0.014294916,0.032542124,0.03300768,0.0167603,0.018448722,0.0016283736],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.009977908,0.0011585228,0.02469566,0.3115184,0.10829981,0.0014559567,0.016131738,0.009788056,0.0011842648,0.10778783,0.072412886,0.33558893],"study_design_scores_gemma":[0.008468584,0.0037227513,0.020260004,0.25744238,0.044711594,0.0024135774,0.012083643,0.020977266,0.0014121747,0.5195443,0.10726163,0.0017021103],"about_ca_topic_score_codex":0.025885371,"about_ca_topic_score_gemma":0.07762841,"teacher_disagreement_score":0.139171,"about_ca_system_score_codex":0.023064485,"about_ca_system_score_gemma":0.06328265,"threshold_uncertainty_score":0.1716227},"labels":[{"model":"gemma","categories":["metaresearch"],"domain":"methods","study_design":"theoretical_or_conceptual","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"methods","study_design":"systematic_review","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W7117383184","doi":"10.1016/j.msksp.2025.103480","title":"Latent classes of trial reporting and publication practices in spinal manipulation research: a meta-epidemiological study","year":2025,"lang":"en","type":"article","venue":"Musculoskeletal Science and Practice","topic":"Musculoskeletal pain and rehabilitation","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta; Université du Québec à Trois-Rivières","funders":"","keywords":"Interpretability; Credibility; Spinal manipulation; Transparency (behavior); Alternative medicine; Clinical trial; MEDLINE","score_opus":0.3890855268524256,"score_gpt":0.5452740837269713,"score_spread":0.1561885568745457,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7117383184","genre_codex":"methods","genre_gemma":"empirical","domain_codex":"methods","domain_gemma":"reporting","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"reporting","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.12050151,0.2175855,0.6140179,0.015275286,0.0021230232,0.018679887,0.004817811,0.0014487662,0.005550263],"genre_scores_gemma":[0.78859246,0.015356683,0.1659306,0.0034978692,0.0007422895,0.02177964,0.0026466404,0.00057424925,0.00087960705],"study_design_codex":"observational","study_design_gemma":"meta_analysis","domain_scores_codex":[0.17228515,0.60772467,0.14331773,0.029666103,0.0444308,0.0025756015],"domain_scores_gemma":[0.06055522,0.6697038,0.15735786,0.09139374,0.020173496,0.00081593177],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.68642205,0.0027819287,0.008133524,0.014559778,0.0032657774,0.013821761,0.0050201663,0.0069310726,0.0036109693],"category_scores_gemma":[0.83449537,0.0033872584,0.027180286,0.017651739,0.008215783,0.012572925,0.00882833,0.0062040137,0.0006364468],"study_design_candidate":"meta_analysis","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.009010007,0.00053128373,0.44005743,0.05766682,0.24180704,0.00040397435,0.009590393,0.008238672,0.0012250728,0.023991304,0.0072083795,0.20026964],"study_design_scores_gemma":[0.0071295365,0.0043695993,0.32150468,0.07052691,0.33379242,0.0017444644,0.0042537507,0.07869981,0.00671905,0.12627432,0.043166563,0.0018189023],"about_ca_topic_score_codex":0.0034868554,"about_ca_topic_score_gemma":0.0032258942,"teacher_disagreement_score":0.9972181,"about_ca_system_score_codex":0.006179598,"about_ca_system_score_gemma":0.009166335,"threshold_uncertainty_score":0.38669747},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":["metaresearch","metaepi_broad"],"domain":"reporting","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"agree"},{"id":"W7117489679","doi":"10.1016/j.semarthrit.2025.152907","title":"Determining the best discriminatory physical functioning outcome measurement instrument for psoriatic arthritis trials: A meta-epidemiological study","year":2025,"lang":"en","type":"article","venue":"Seminars in Arthritis and Rheumatism","topic":"Spondyloarthritis Studies and Treatments","field":"Medicine","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Krembil Foundation; University Health Network; Ottawa Hospital; University of Ottawa","funders":"Parker Institute for Cancer Immunotherapy; Oak Foundation","keywords":"Psoriatic arthritis; Outcome (game theory); Patient-reported outcome; Arthritis; MEDLINE; Psychometrics","score_opus":0.16281216313373342,"score_gpt":0.37394785276592624,"score_spread":0.2111356896321928,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7117489679","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":"methods","model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":"methods","domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.29857033,0.6606126,0.031316776,0.0030056674,0.0018752426,0.0015996598,0.0015434906,0.0002932296,0.0011830503],"genre_scores_gemma":[0.95878243,0.021657996,0.014105449,0.0021048079,0.0005222751,0.0012394029,0.001131292,0.00012486264,0.00033147287],"study_design_codex":"meta_analysis","study_design_gemma":"meta_analysis","domain_scores_codex":[0.87399024,0.09852011,0.014422267,0.008112062,0.0033407507,0.0016145104],"domain_scores_gemma":[0.8403716,0.13283725,0.012382389,0.0102861505,0.0029454047,0.0011772168],"candidate_categories":["metaresearch","metaepi_narrow","metaepi_broad"],"consensus_categories":[],"category_scores_codex":[0.11819529,0.0037497217,0.020910768,0.0036153449,0.0012544906,0.007073642,0.0032008612,0.004834094,0.0029304035],"category_scores_gemma":[0.13013807,0.0025971346,0.041775513,0.0039767623,0.0020281824,0.0046861973,0.0035174098,0.0045203334,0.00038421262],"study_design_candidate":"meta_analysis","study_design_consensus":"meta_analysis","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.022823567,0.0001269102,0.03397788,0.013343654,0.91887236,0.000064233376,0.000120202014,0.0008641222,0.00032697644,0.00027228892,0.00043831818,0.008769483],"study_design_scores_gemma":[0.013621344,0.0011934313,0.01911963,0.0016548639,0.96032333,0.000079464,0.00011228279,0.0019453733,0.00022115327,0.0009951959,0.00066436175,0.00006951587],"about_ca_topic_score_codex":0.001695256,"about_ca_topic_score_gemma":0.0024503784,"teacher_disagreement_score":0.9962503,"about_ca_system_score_codex":0.0014721362,"about_ca_system_score_gemma":0.002091353,"threshold_uncertainty_score":0.6250839},"labels":[{"model":"gemma","categories":["metaresearch","metaepi_broad","metaepi_narrow"],"domain":"methods","study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":["metaepi_narrow","metaepi_broad","metaresearch"],"domain":"methods","study_design":"meta_analysis","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"}]}