Applied machine learning to predict 1-year major adverse cardiovascular events in elderly patients after percutaneous coronary intervention

Cardiovascular diseases remain the leading cause of mortality worldwide, with elderly patients experiencing the worst prognosis following ST-elevation myocardial infarction (STEMI). Traditional risk stratification models demonstrate suboptimal performance in geriatric patients due to complex risk profiles involving frailty and multiple comorbidities. To develop machine learning-based predictive models for one-year major adverse cardiovascular events (MACE) in elderly patients (≥65 years) undergoing percutaneous coronary intervention (PCI) for STEMI. This retrospective cohort study analyzed 1,358 elderly patients who underwent PCI between 2015 and 2021. MACE included cardiovascular death, myocardial infarction, stroke, and revascularization within one year. Eight machine learning algorithms were evaluated: XGradient Boosting (XGB), Random Forest (RF), Logistic Regression, Neural Networks, Support Vector Machines, K-Nearest Neighbors, Decision Trees, and Naive Bayes. The Synthetic Minority Oversampling Technique (SMOTE) was applied to address class imbalance. Model performance was evaluated using various metrics, and SHAP (Shapley Additive Explanations) was used to enhance model interpretability and support clinical decision-making by identifying key risk factors driving predictions. Among the patients (mean age 74.1 ± 6.7 years, 31.8% female), 152 (11.2%) experienced MACE within one year. XGB and RF emerged as the most robust models, achieving area under the receiver operating characteristic curve (AUC) values of 94% and 95%, respectively, with RF demonstrating higher sensitivity (79%) and specificity (96%). SHAP analysis revealed pre-PCI ejection fraction, age, creatinine levels, fasting blood sugar, BMI, and LDL/HDL ratio as the most influential predictors. Machine learning models demonstrated strong predictive performance in predicting one-year MACE in elderly post-PCI patients. By combining high-performance models with SHAP-based explanations, this approach supports transparent, clinically actionable risk stratification and personalized care.