Mortality prediction in patients with acute myocardial infarction and prior ischemic stroke: a prospective cohort study with explainable machine learning and SHAP analysis

Background Patients with acute myocardial infarction (AMI) and a prior history of ischemic stroke (IS) represent a high-risk population with limited prognostic tools. This study aimed to develop and validate an interpretable machine learning model for predicting all-cause mortality in AMI patients with prior IS. Methods This prospective single-center cohort included 543 consecutive patients with AMI and prior IS who underwent Percutaneous Coronary Intervention (PCI). Candidate predictors were screened using univariable Cox regression, Boruta, and LASSO Cox regression. Six survival models were compared: Cox, elastic-net Cox, random survival forest (RSF), GBM survival, XGBoost survival, and Survival SVM. Internal validation used 100 bootstrap resamples with out-of-bag (OOB) evaluation. Performance was evaluated using Harrell's C-index, IPCW time-dependent AUC, calibration measures, Brier score, and decision-curve analysis. SHAP-based and permutation analyses were used for model interpretation. Results During a median follow-up of 3.03 years, 66 patients (12.2%) died. Six predictors were selected: age, hemoglobin, prognostic nutritional index, estimated glomerular filtration rate, atherogenic index of plasma, and monocyte-to-lymphocyte ratio. Although RSF achieved the numerically highest OOB C-index, elastic-net Cox showed comparable discrimination, better calibration, a lower 3-year Brier score, and greater interpretability and was selected as the final model. Its OOB C-index was 0.714 (95% CI 0.616–0.790), and its 1-, 2-, and 3-year IPCW AUCs were 0.745, 0.704, and 0.701, respectively. Age and hemoglobin made the largest contributions to model predictions. Elastic-net Cox showed numerically better performance than GRACE, but statistical superiority in discrimination was not demonstrated. Conclusions The six-variable elastic-net Cox model provided moderate discrimination and relatively favorable calibration. Independent external validation and prospective implementation studies are required before clinical use.