Heart disease prediction through feature optimization, ensembles, and threshold calibration

Abstract

Early detection of cardiovascular disease is critical for reducing mortality and enabling timely clinical intervention. While machine learning models have been widely applied to heart disease prediction, many studies prioritize overall accuracy rather than sensitivity, which is essential in medical screening tasks. This study proposes a structured multi-stage optimization framework that integrates feature optimization, ensemble model tuning, and recall-oriented decision-threshold calibration within a unified pipeline. Feature selection was performed using RFECV and importance-based methods to identify an optimal subset of 14 predictors, and ensemble classifiers including Random Forest, Extra Trees, and XGBoost were refined through hyperparameter tuning using RandomizedSearchCV. To prioritize sensitivity, the classification threshold was calibrated from 0.5 to 0.3. In the calibrated evaluation split, recall increased to 0.961 while maintaining strong discriminative performance (AUC ≈ 0.93). To assess robustness, stratified cross-validation and bootstrap resampling were conducted; cross-validation demonstrated stable recall performance around 0.88–0.89 across models, and bootstrap analysis yielded a mean recall of 0.902 (95% CI: 0.842–0.959), confirming that the recall-oriented optimization remains statistically reliable. McNemar’s test further indicated no statistically significant performance differences among the ensemble models. These findings demonstrate that integrating feature selection, ensemble learning, and threshold calibration can improve heart disease detection while maintaining predictive stability, providing a clinically aligned machine learning framework for cardiovascular risk assessment.