Evaluation of coronary heart disease risk prediction based on simple physical examination parameters by machine learning model: a retrospective cohort model development and validation study

Background To develop and externally validate a coronary heart disease (CHD) risk model from routine clinical indicators and identify key predictors. Methods The Framingham Heart Study cohort ( n = 4,240) was used. Missing values and outliers were handled, and class imbalance was corrected with SMOTEENN/SMOTETomek. Data were split 7:3 for training and internal validation. A two-tier feature selection (chi-square, mutual information, ANOVA F -test) retained ten variables. A stacked ensemble of gradient boosting, random forest, and XGBoost with a logistic-regression meta-learner was trained. Performance was measured by AUC, accuracy, precision, recall, and F 1. External validation used a retrospective hospital cohort ( n = 200; 2024–2025). Model explanations were derived with SHAP. Results Internal validation yielded AUC 0.977 and accuracy 0.942 ( F 1: 0.944). External validation achieved AUC 0.929 and accuracy 0.885. SHAP identified systolic blood pressure, age, total cholesterol, and fasting glucose as leading contributors, with plausible nonlinear effects and interactions. Conclusion A model built from routinely available measures demonstrates strong discrimination for CHD risk and generalizes to an external cohort, offering a clinically interpretable tool for cardiovascular risk assessment.