SHAP enhanced predictive modeling for clinical machine learning

Clinical prediction is challenged by heterogeneity and noise in healthcare data. We evaluated SHAP-Enhanced Predictive Modeling (SEPM), a data-centric framework that ranks training observations using aggregated absolute Shapley Additive Explanations (SHAP) and optimizes subset retention through cross-validation. Across four clinical datasets and five classifiers, SEPM was compared with full-data training (TML) and randomized subset selection (SEPM-RAND) using nested cross-validation. A separate common-protocol XGBoost benchmark compared SEPM with Truncated Monte Carlo (TMC) Data Shapley estimation. Twenty simulation scenarios varied training sample size and outcome prevalence, with evaluation on full, high-SHAP, and low-SHAP test groups. SEPM achieved higher mean F1 scores than TML in 18 of 20 dataset-classifier combinations and SEPM-RAND in 16, retaining 42.1%-94.4% of training observations. TMC achieved higher mean F1 scores and balanced accuracy in all four datasets, whereas SEPM had lower probability-error measures and pipeline execution times (TMC-to-SEPM runtime ratios: approximately 65-1,216). The dataset-level F1 comparison was not significant (exact two-sided Wilcoxon p = 0.125). SEPM achieved the highest mean full-test F1 in all 20 simulated scenarios, but higher mean expected calibration error and Brier scores. SEPM may improve threshold-based classification with fewer training observations, but its comparison with randomized selection does not isolate ranking at identical final subset sizes, and probability quality can deteriorate. External validation, fixed-size comparisons, subgroup assessment, and calibration-aware optimization are needed before clinical use. Clinical trial registration Not applicable.

The online version contains supplementary material available at https://doi.org/10.1007/s44250-026-00440-w.