{"id":"W3161782857","doi":"10.1186/s12909-021-02653-4","title":"Assessing the validity of an OSCE developed to assess rare, emergent or complex clinical conditions in endocrinology & metabolism","year":2021,"lang":"en","type":"article","venue":"BMC Medical Education","topic":"Innovations in Medical Education","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"ca_institutions":"Medical Council of Canada; Ottawa Hospital; University of Ottawa","funders":"","keywords":"Generalizability theory; Formative assessment; Objective structured clinical examination; Medicine; Medical education; Delphi method; Psychology; Family medicine; Statistics; Mathematics","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":[0.002632923,0.0001656681,0.0005117685,0.0002522593,0.0001432686,0.00003254121,0.0003006717,0.0002348861,0.004808859],"category_scores_gemma":[0.05505511,0.0001209761,0.00008592534,0.00158156,0.0002997218,0.0001870226,0.0001031954,0.0006536724,0.00005585367],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000155198,"about_ca_system_score_gemma":0.01649757,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00009964522,"about_ca_topic_score_gemma":0.0008405022,"domain_scores_codex":[0.9957166,0.001184189,0.001398384,0.0004558106,0.0009046973,0.0003403238],"domain_scores_gemma":[0.9963514,0.001122035,0.0002856882,0.000778385,0.001135451,0.0003270102],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0002896093,0.02488052,0.4042577,0.001115176,0.0002426864,0.00006591438,0.00425277,0.00003535326,0.004757941,0.03701307,0.2153156,0.3077738],"study_design_scores_gemma":[0.000919911,0.00005853653,0.9121448,0.000272703,0.0001006313,0.0001953927,0.003198262,0.0003347954,0.0004920675,0.0004176566,0.08172358,0.0001416669],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9644886,0.00005187923,0.0114592,0.01900444,0.003493827,0.000556972,0.000005493172,0.00003851459,0.0009011164],"genre_scores_gemma":[0.8784574,0.00009501368,0.1021681,0.01590191,0.001175744,0.000285633,0.001222603,0.00002707984,0.0006664507],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5078871,"threshold_uncertainty_score":0.9961009,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.3838439118483818,"score_gpt":0.5621679485589601,"score_spread":0.1783240367105783,"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."}}