Beyond Mean Absolute Error: Ancestry-Stratified Calibration and Explainability for Warfarin Dosing Models

Background: Warfarin's narrow therapeutic index requires precise dose individualization, yet published machine-learning studies report accuracy that is inversely correlated with evaluation rigor: the field's best-reported mean absolute error (MAE) comes from an unreliably small test set. Methods: We present a reproducible, nine-phase ML pipeline on the public International Warfarin Pharmacogenetics Consortium (IWPC) dataset (n=6037 and n=1780) with fixed patient-level splits, strict leakage controls, and a nested clinical/genetic ablation across nine architectures. Results: AutoGluon with combined features achieves MAE =8.55 mg/week, R2=0.475, PW20 (percentage of predictions within 20% of the true dose) =0.466 on the larger cohort (MAE =7.96, R2=0.480, PW20 =0.497 on the smaller), exceeding prior rigorous benchmarks. A leave-one-ancestry-group-out evaluation, cross-checked across two model families, reproduces only part of an earlier analysis's known error concentration in Black/African-American patients: the Black label shows no penalty under either model, while Han Chinese and Caucasian show model-dependent penalties. A deep ensemble is badly miscalibrated (24.9%/21.6% coverage of a nominal 90% interval), while MAPIE (Model Agnostic Prediction Interval Estimator) conformal prediction on the same model achieves near-nominal coverage (89.8%/93.3%); per-ancestry coverage shows every Black/African-American label under-covered, a fairness gap the aggregate number hides. On IWPC-1780, SHAP (SHapley Additive exPlanations) attributes 64% of feature importance to genetics, confirming VKORC1 as the strongest predictor. Conclusions: To our knowledge, no previously published pipeline evaluated on the IWPC dataset jointly reports accuracy, ancestry-stratified fairness, calibration, and explainability on the same fixed splits.