Explainable Machine Learning for Cardiovascular Disease Prediction: Internal Validation, Cross-Dataset Evaluation, and a Preventive Decision-Support Framework

Background and Objective: Cardiovascular diseases (CVDs) remain the leading cause of mortality worldwide, and the early identification of individuals at risk is a cornerstone of cardiovascular disease prevention [۱]. Machine learning methods can model complex and nonlinear relationships among clinical and behavioral factors and therefore offer considerable potential for cardiovascular disease risk prediction. However, model interpretability, generalizability across different populations, and the translation of model outputs into information that can support prevention remain important challenges in the application of these methods in healthcare. This study aimed to develop and internally evaluate several machine learning models for cardiovascular disease prediction, assess model transferability to a different dataset, interpret model behavior using SHapley Additive explanations (SHAP), and develop a research-oriented framework for risk stratification and preventive decision support. Methods: The development dataset consisted of ۷۰,۰۰۰ records from the CVD-۷۰K dataset. After quality control of blood pressure and anthropometric variables, ۶۸,۶۷۵ observations remained. Logistic Regression, tuned Random Forest, and XGBoost models were evaluated using a stratified ۸۰/۲۰ train-test split and stratified five-fold cross-validation within the training set. A transferable feature set comprising age, sex, systolic blood pressure, and cholesterol was defined, and four-feature Random Forest and XGBoost models were evaluated, without retraining, on ۷۴,۲۷۶ cleaned observations from the Yemen dataset. Because of the substantial difference in age distribution, cross-dataset evaluation was performed both on the entire Yemen dataset and on the overlapping age subgroup of ۲۹-۶۴ years. The full XGBoost model was interpreted using SHAP, and its outputs were used to construct operational risk strata and overlapping phenotypes of modifiable risk factors. Results: On the CVD-۷۰K test set, XGBoost achieved an accuracy of ۰.۷۳۰۱, an F۱-score of ۰.۷۱۴۲, and an ROC-AUC of ۰.۷۹۵۲, while the tuned Random Forest achieved an ROC-AUC of ۰.۷۹۶۳. Mean five-fold cross-validation ROC-AUC values for Logistic Regression, Random Forest, and XGBoost were ۰.۷۹۱۹ ± ۰.۰۰۲۷, ۰.۸۰۱۹ ± ۰.۰۰۳۲, and ۰.۸۰۰۵ ± ۰.۰۰۳۳, respectively. ROC-AUC values for XGBoost models using ۴, ۶, ۸, and ۱۰ features were ۰.۷۹۰۲, ۰.۷۹۳۰, ۰.۷۹۵۴, and ۰.۷۹۵۲, respectively. On the full Yemen dataset, ROC-AUC values for Random Forest and XGBoost were ۰.۷۲۸۸ and ۰.۷۲۷۲, respectively; both models achieved sensitivity above ۰.۹۲ but specificity below ۰.۱۶. In the ۲۹-۶۴-year subgroup, accuracy, specificity, and F۱-score improved, whereas ROC-AUC decreased to ۰.۶۶۴۱ for Random Forest and ۰.۶۶۱۷ for XGBoost. In Yemen, Brier scores were ۰.۲۳۲۸ and ۰.۲۳۳۹, calibration intercepts were -۰.۷۱۴۸ and -۰.۶۸۸۴, and calibration slopes were ۱.۱۶۰۴ and ۱.۰۸۱۵ for Random Forest and XGBoost, respectively. SHAP analysis identified systolic blood pressure, cholesterol, and age as the most important contributors to model predictions. The observed prevalence of CVD increased from ۲۵.۱۹% in the low-risk group to ۷۹.۴۵% in the high-risk group. Conclusion: Machine learning models demonstrated moderate discrimination in CVD-۷۰K, and the four-feature set retained most of the predictive performance of the full model. However, the altered performance pattern, low specificity, and calibration measures observed in the Yemen dataset demonstrated that satisfactory internal performance does not guarantee transferability to a different dataset. Age-restricted subgroup analysis improved several threshold-dependent performance metrics but did not improve ROC-AUC. Combining cross-dataset evaluation, SHAP-based interpretation, and profiling of modifiable risk factors provided a research framework for more transparent model evaluation; however, this framework is not intended as a diagnostic tool or as a source of individualized treatment recommendations.

کلیدواژه ها:

Cardiovascular disease ، Machine learning ، Explainable artificial intelligence ، XGBoost ، SHAP ، Cross-dataset evaluation ، Transferability ، Risk stratification