{"id":"W2757874438","doi":"10.1097/acm.0000000000001938","title":"In Search of Black Swans: Identifying Students at Risk of Failing Licensing Examinations","year":2017,"lang":"en","type":"article","venue":"Academic Medicine","topic":"Medical Education and Admissions","field":"Medicine","cited_by":15,"is_retracted":false,"has_abstract":true,"ca_institutions":"Impact; Western University","funders":"","keywords":"Summative assessment; Predictive validity; Medicine; Receiver operating characteristic; Predictive modelling; Risk assessment; Cohort; Framingham Risk Score; Internal medicine; Psychology; Clinical psychology; Computer science; Machine learning; Mathematics education; Formative assessment","routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001954686,0.0001194719,0.0004858636,0.0003132702,0.0001595153,0.000003660645,0.0003766932,0.0001836183,0.001387289],"category_scores_gemma":[0.01891592,0.00009214112,0.00004706898,0.0001990624,0.0006408258,0.0001232107,0.000188116,0.0008086623,0.00002198438],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00008990501,"about_ca_system_score_gemma":0.0002569562,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0004208203,"about_ca_topic_score_gemma":0.00001692923,"domain_scores_codex":[0.9977061,0.0001032914,0.0007738167,0.0002430697,0.0009344443,0.000239343],"domain_scores_gemma":[0.9978538,0.0004309891,0.0004693835,0.0005434906,0.0001829062,0.0005193963],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00009979447,0.0001436313,0.8571089,0.000746407,0.000111082,0.00002811333,0.0375574,0.00000800209,0.09276665,0.0006158169,0.006632156,0.004182075],"study_design_scores_gemma":[0.005687724,0.0002127269,0.943653,0.007490909,0.0003328267,0.00003069056,0.01616002,0.001185181,0.02132561,0.0005802637,0.003179719,0.0001612702],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9876072,0.0005185108,0.0002373145,0.008587819,0.000332767,0.0003206139,0.00000316408,0.00001092177,0.002381636],"genre_scores_gemma":[0.9902952,0.002209641,0.0001785062,0.0005676642,0.0002980902,0.000004030169,0.00001116399,0.00001475218,0.006420917],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.08654419,"threshold_uncertainty_score":0.9995256,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.1283881165028875,"score_gpt":0.4837211549512799,"score_spread":0.3553330384483925,"validation_status":"score_only:v0-immature-baseline","note":"Baseline scores from an immature model (maturity gate not passed). Scores rank; they never assert a category."}}