{"id":"W4361228074","doi":"10.1101/2023.03.26.23286718","title":"Diagnosing and remediating harmful data shifts for the responsible deployment of clinical AI models","year":2023,"lang":"en","type":"preprint","venue":"medRxiv","topic":"Machine Learning in Healthcare","field":"Computer Science","cited_by":8,"is_retracted":false,"has_abstract":true,"ca_institutions":"St. Michael's Hospital; Vector Institute; Hospital for Sick Children; University of Toronto","funders":"Vector Institute; University of Toronto; Canadian Institute for Advanced Research","keywords":"Software deployment; Harm; Leverage (statistics); Demographics; Community hospital; Warning system; Health care; Transfer of learning; Medical diagnosis; Medicine; Machine learning; Artificial intelligence; Medical emergency; Emergency medicine; Computer science; Psychology; Demography; Nursing","routes":{"ca_aff":true,"ca_fund":true,"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","open_science"],"consensus_categories":[],"category_scores_codex":[0.01043781,0.0002309421,0.0005540827,0.0001264736,0.0002164985,0.0001804488,0.003633491,0.0002323829,0.000001459942],"category_scores_gemma":[0.008845568,0.0001762431,0.0001104404,0.0002171161,0.0001024784,0.0002264877,0.009218053,0.001229399,0.000003346244],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002424352,"about_ca_system_score_gemma":0.0006068271,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0004400852,"about_ca_topic_score_gemma":0.0002215842,"domain_scores_codex":[0.9959546,0.0007619286,0.001141946,0.001226227,0.0005437611,0.0003714857],"domain_scores_gemma":[0.9843671,0.01110278,0.0006894374,0.003493817,0.0002054952,0.0001413846],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0001316997,0.0001380498,0.6877843,0.002703762,0.0002778534,0.00004669025,0.004658047,0.05379662,0.0000272976,0.01976602,0.006887004,0.2237826],"study_design_scores_gemma":[0.0002034719,0.00005974581,0.05391413,0.0003655276,0.00002937556,0.00000139529,0.00001960966,0.9273497,0.00003019765,0.01661896,0.001247584,0.000160272],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1692572,0.001690298,0.7634326,0.05875107,0.004572871,0.001718657,0.0001332643,0.0004146571,0.00002937474],"genre_scores_gemma":[0.8316346,0.001162815,0.1647898,0.001111484,0.0008095327,0.0002483834,0.00006130195,0.00008265024,0.0000993799],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8735531,"threshold_uncertainty_score":0.9995034,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.3898331686800192,"score_gpt":0.4786072066113161,"score_spread":0.0887740379312969,"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."}}