A machine learning approach predicts mortality in acute coronary syndrome (ACS) patients following percutaneous coronary intervention (PCI)

Background Patients with acute coronary syndrome (ACS) treated with percutaneous coronary intervention (PCI) carry a residual risk of adverse events and all-cause mortality. Traditional risk scores such as TIMI and GRACE focus on short-term outcomes and are not designed to capture the complex interplay of post-intervention clinical, lifestyle, and socioeconomic factors. Machine learning (ML) offers a novel approach to adaptive risk stratification by identifying patterns across diverse patient data. Purpose To develop, externally validate, and explain a ML-based model for predicting all-cause mortality in ACS patients following PCI. Methods Data from 1,261 ACS patients (mean age 61±11.1 years, 16.6% women) treated with PCI were obtained from the Santiago de Compostela Cardiac Rehabilitation program (2015–2020). The dataset included demographics, socioeconomic factors, medical history, laboratory results, cardiopulmonary exercise testing, and lifestyle data. The central electronic health record system was accessed to retrieve date and cause of death within 12.0 ± 20.9 months of follow-up. Seven ML algorithms were trained using 50 repeats of 4-fold stratified cross-validation. Key performance metrics, including AUC, specificity, sensitivity, accuracy, and balanced accuracy, were compared across all models. Feature selection was performed using random mutation implementation of the Genetic Algorithm and model predictions were explained using SHapley Additive exPlanation (SHAP) values. External validation was conducted on 566 ACS patients post-PCI and a broader population of 1,386 CAD patients from the European multicentre EU-CaRE cohort. Model replicability was assessed using the independent validation screening (IVS) tool. Results LightGBM, Random Forest, and SVM outperformed logistic regression (mean AUC >0.8). LightGBM achieved the highest balanced accuracy (0.87), with sensitivity of 0.86 and specificity of 0.88. External validation in ACS post-PCI patients confirmed reliable identification of true positives (sensitivity 0.78, specificity 0.73). SHAP analysis identified nine key predictors: diabetes, physical exercise capacity (peak METs), average daily sitting duration, statin prescription, family history of heart disease, number of affected vessels, revascularization history, serum creatinine, and chronic hepatopathy. Notably, patients achieving <6 METs or sitting ≥9 hours/day showed substantially increased mortality risk. The model received an overall IVS rating of “recommended.” Conclusions The ML-based model provides a transparent, high-accuracy tool for identifying ACS patients post-PCI at elevated mortality risk, integrating clinical, lifestyle, and socioeconomic factors derived from standard assessments and electronic health records. It may serve as a decision support tool to identify high-risk patients who could benefit from intensified personalized treatment and lifestyle interventions.