{"id":"W4396894318","doi":"10.1038/s41746-024-01118-4","title":"Shortcut learning in medical AI hinders generalization: method for estimating AI model generalization without external data","year":2024,"lang":"en","type":"article","venue":"npj Digital Medicine","topic":"Machine Learning in Healthcare","field":"Computer Science","cited_by":65,"is_retracted":false,"has_abstract":true,"ca_institutions":"Princess Margaret Cancer Centre; Vector Institute; University Health Network; Toronto General Hospital; University of Toronto; Ted Rogers Centre for Heart Research","funders":"University of Toronto; Natural Sciences and Engineering Research Council of Canada; University Health Network; Canadian Institute for Advanced Research","keywords":"Generalization; Computer science; Artificial intelligence; Machine learning; Mathematics","routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001818252,0.0002599707,0.0003732166,0.0003269904,0.0001429219,0.0004042335,0.001289044,0.0001655225,0.00003688415],"category_scores_gemma":[0.002687702,0.0002197525,0.00004163275,0.0006734377,0.00007371791,0.001734694,0.0005240092,0.0007046069,0.000006809141],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001226002,"about_ca_system_score_gemma":0.0004948213,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001709406,"about_ca_topic_score_gemma":0.0000594872,"domain_scores_codex":[0.9966655,0.0001615446,0.0007118508,0.0009497579,0.001065805,0.000445495],"domain_scores_gemma":[0.9983618,0.0004074646,0.0001113694,0.0006964497,0.0001527368,0.0002701778],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00001249564,0.00002706496,0.009062691,0.0004263517,0.00002318654,0.00007537079,0.001818788,0.510054,0.00005595134,0.03671651,0.004030154,0.4376974],"study_design_scores_gemma":[0.0004596931,0.0001213281,0.0001205619,0.0008872121,0.00001284822,0.0001097969,0.00002285703,0.9887929,0.00000737044,0.006492275,0.002759875,0.0002132883],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.0007516966,0.000641217,0.9713503,0.02494612,0.0009446021,0.0003479409,0.00001046973,0.0004851406,0.0005225294],"genre_scores_gemma":[0.4063966,0.00002516098,0.5853858,0.005390911,0.001410436,0.0000582414,0.0005656239,0.0000691035,0.0006981218],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.4787389,"threshold_uncertainty_score":0.8961245,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.06358721130492692,"score_gpt":0.4232754399995051,"score_spread":0.3596882286945782,"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."}}