Code, data and results for: Same prediction, different explanations — agreement, faithfulness, stability and error-signal value of post-hoc explanations for clinical prediction models

This archive contains the analysis code, input data, per-patient outputs and figures for a study of post-hoc explanations of clinical prediction models. Five public clinical datasets were used: breast cancer (WDBC), Cleveland heart disease, South African heart disease, Pima diabetes and the Diabetes 130-US Hospitals electronic health record cohort (30-day readmission; 69,970 patients). On each, we trained logistic regression, an explainable boosting machine, random forest and XGBoost. Held-out predictions for 120 patients per dataset were explained with KernelSHAP, LIME and occlusion (7200 explanations). We evaluated the explanations' top-3 feature agreement, deletion-based faithfulness and stability under reruns and input perturbation. We also tested whether disagreement between explanations identifies misclassified predictions, compared with model confidence, using patient-level (cluster-bootstrap) inference. Running code/run_all.sh reproduces all results, tables and figures from the raw data; see README.md. Version 1.1 adds the electronic health record cohort; results for the 4 original datasets are identical to version 1.0. The benchmark datasets come from the Penn Machine Learning Benchmarks repository (commit 7c1f4bd); the Diabetes 130-US Hospitals dataset comes from the UCI Machine Learning Repository (doi:10.24432/C5230J; CC BY 4.0). This record accompanies a manuscript submitted to the Journal of the American Medical Informatics Association (JAMIA).