{"id":"W4282836129","doi":"10.1097/aog.0000000000004841","title":"Accuracy of Surgeon Self-Reflection on Hysterectomy Quality Metrics","year":2022,"lang":"en","type":"article","venue":"Obstetrics and Gynecology","topic":"Surgical Simulation and Training","field":"Medicine","cited_by":3,"is_retracted":false,"has_abstract":true,"ca_institutions":"University Health Network; Mount Sinai Hospital","funders":"","keywords":"Medicine; Wilcoxon signed-rank test; Complication; Cohort; Surgery; Hysterectomy; Mann–Whitney U test; Internal medicine","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":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0161141,0.0002765415,0.0003592053,0.001659408,0.0003291618,0.001317146,0.000641857,0.0003495182,0.0007393647],"category_scores_gemma":[0.09559943,0.0002399759,0.0005844485,0.001632325,0.0004948727,0.000712074,0.001213769,0.0004110293,0.0002291732],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001573627,"about_ca_system_score_gemma":0.00207332,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.04504208,"about_ca_topic_score_gemma":0.04484982,"domain_scores_codex":[0.9885427,0.004196909,0.001707803,0.001240416,0.00374449,0.000567736],"domain_scores_gemma":[0.888772,0.03915395,0.04170764,0.007425662,0.02149942,0.001441337],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00005490051,0.000007990991,0.9902954,0.00003757955,0.0001052978,0.000007636438,0.0005073174,0.000356215,0.00008006243,0.00003220486,0.0002607235,0.008254703],"study_design_scores_gemma":[0.000003995423,0.0000749666,0.9959824,0.0000578534,0.00002631937,0.00005223166,0.0006701106,0.002034277,0.0002843232,0.00006591226,0.0007310179,0.00001655458],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9911062,0.0006126311,0.003642652,0.0004168518,0.00004198096,0.00004573975,0.001613035,0.00007759911,0.002443378],"genre_scores_gemma":[0.9980434,0.0001096498,0.0009781142,0.00005041084,0.00001092697,0.00001590096,0.0006070151,0.000006254747,0.0001782353],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.04504208,"threshold_uncertainty_score":0.08955979,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.0719261144888636,"score_gpt":0.3558064418682317,"score_spread":0.2838803273793681,"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."}}