A comparative study of machine learning and traditional survival models in predicting 5-year post‑ coronary artery bypass grafting mortality

Introduction: Coronary artery disease (CAD) is a major public health issue worldwide and in Iran, contributing substantially to morbidity and mortality. For patients with symptomatic and advanced forms of the disease, coronary artery bypass grafting (CABG) remains a widely utilized and effective treatment strategy. This study aimed to investigate five-year post-CABG survival and identify key predictors of mortality using traditional and machine learning (ML) survival models. Methods This retrospective cohort study included 2,860 patients who underwent isolated CABG between 2016 and 2021 at Farshchian Cardiovascular Hospital in Hamadan, Iran. Kaplan–Meier analysis and log-rank tests were used for univariate survival analysis. To assess predictive performance, several statistical and ML models were applied, including Cox regression, Bayesian Neural Network (BNN) survival, random forest, bagging survival, support vector machines (SVMs), and parametric Weibull regression. Model performance was evaluated using accuracy, area under the curve (AUC), Brier score, and F1-score. Result A total of 2,860 participants were included in the study, of whom 4.9% died during the follow‑up period. The mean age at death was significantly higher compared to that of survivors (68 vs. 63 years, p < 0.001), with a greater proportion aged ≥ 76 years among the deceased (26.6% vs. 8.5%). Among all models, the BNN Survival model demonstrated the best overall predictive performance. This model showed the highest sensitivity at 0.90 (95% CI: 0.70–0.99), with a specificity of 0.61 (95% CI: 0.47–0.83), and an AUC of 0.77 (0.68–0.85). Variable importance analysis showed age as the most influential predictor (normalized importance close to 1.0), followed by residential area (~ 0.65). Conclusion Advanced ML techniques, especially BNN survival models, can greatly improve the accuracy of long-term mortality predictions after CABG compared to traditional methods. Using these models in clinical practice could help tailor risk assessments to each patient and improve post-surgery care.