Machine learning models for classifying all-cause mortality and deaths from cardiovascular disease and stroke

All-cause, cardiovascular, and stroke mortality are major public-health burdens, but mortality-status classifiers require rigorous validation and cautious calibration interpretation. We analyzed 4,137 sampled participants with linked binary mortality labels: 1,473 all-cause deaths, including 821 cardiovascular deaths and 223 stroke deaths. Single-task Elastic Net, Random Forest, XGBoost, and support vector machine models, together with an exploratory joint multinomial Elastic Net, were evaluated using 5-fold outer and 3-fold inner nested stratified cross-validation. Imputation, encoding, normalization, and class up-sampling were restricted to training folds, and performance was calculated from pooled out-of-fold predictions. For all-cause mortality, Elastic Net and the support vector machine achieved similar AUROC values of 0.925 after rounding, whereas Elastic Net achieved the highest AUPRC (0.886). Elastic Net performed best for cardiovascular mortality (AUROC 0.873; AUPRC 0.550). For stroke mortality, the support vector machine achieved both the highest AUROC (0.837) and AUPRC (0.189). The joint model did not improve discrimination and performed worse than single-task Elastic Net for cardiovascular mortality. Calibration curves and statistics showed meaningful deviations from ideal calibration; therefore, class-rebalanced probabilities should not be interpreted as validated absolute-risk estimates. SHAP beeswarm and dependence plots identified age and lactate dehydrogenase as prominent nonlinear model-attribution features. These models provide internally validated mortality-status classification, but clinical risk prediction requires outcome dates, fold-contained recalibration, and external validation.