Abstract
Cardiovascular diseases are among the leading causes of global mortality and morbidity, underscoring the need for accurate and explainable predictive models. This study aimed to develop and evaluate explainable AI models for predicting major cardiovascular events (MCVE) over 60 months. We conducted a secondary analysis of two large cohort studies nested within the Golestan (northeast Iran) and Pars (south Iran) Cohorts, which followed identical protocols. After applying exclusion criteria, 9,769 participants. Predictors included demographic, clinical, and biochemical variables, with class imbalance addressed via SMOTE. Machine learning models (XGBoost, logistic regression, decision tree, SVM, random forest, KNN) and a multilayer perceptron were trained and evaluated. Model performance was assessed using accuracy, sensitivity, specificity, positive predictive value, negative predictive value, AUC, and F1 score. Interpretability was examined using SHapley Additive exPlanations. Machine-learning models showed variable performance across cohorts, with XGBoost demonstrating the best and most stable discrimination. In the pooled cohort, XGBoost achieved the highest AUC (0.84), with good accuracy (0.81) and specificity (0.85). Its performance remained consistent in the PolyIran and PolyPars cohorts, with AUC values of 0.83 and 0.85, respectively, suggesting good generalizability. Random Forest was the second-best model. The MLP deep-learning model showed only modest performance. SHAP analysis identified age, creatinine, and systolic blood pressure as the most important predictors across cohorts, with higher values associated with increased MCVE risk. Higher FBS, BMI, LDL/HDL ratio, male sex, and smoking were also associated with increased risk, whereas PolyPill use and greater adherence showed protective effects. Explainable AI models, particularly XGBoost, demonstrated high predictive accuracy for MCVE and revealed creatinine as a novel risk indicator not included in traditional CVD scores. These findings support incorporating kidney function screening into CVD risk stratification and highlight the potential of AI-driven tools to enhance prevention strategies.