Explainable Stacked Ensemble Learning Model for Clinical Heart Disease Risk Prediction

Successive feature-selection stages can simplify clinical prediction models without necessarily improving their performance. This study examines a sequential Mutual Information (MI)–correlation–LASSO procedure within a stacked ensemble using the Framingham Heart Study and UCI Heart Disease datasets. Five configurations were compared: all processed features, MI only, MI plus correlation, the complete sequence, and recursive feature elimination with cross-validation (RFECV). Model development used the training partitions, with performance reported on held-out observations. The complete procedure reduced Framingham from 15 to 14 predictors and UCI from 15 to 13. Framingham ROC-AUC ranged from 0.677 to 0.685 across configurations. On UCI, MI plus correlation achieved the highest F1-score (0.858), while the complete sequence achieved the highest ROC-AUC (0.897) and PR-AUC (0.916). TreeSHAP and full-stack KernelSHAP identified several common leading predictors, alongside retained variables with low attribution magnitudes. The findings show that the benefit of successive selection stages varied by dataset and metric, and that selection status differed from a predictor’s contribution to the fitted model.