An exploratory machine learning analysis of 10-year mortality after isolated CABG

Official URL: https://dx.doi.org/10.1155/jocs/3088313

Background: Conventional risk scoring systems often lack optimization for long-term prognosis, necessitating the exploration of machine learning (ML) methods in predicting mortality after coronary artery bypass grafting (CABG). This study aimed to identify risk factors for long-term mortality in isolated CABG patients and compare the performance of ML models with EuroSCORE II. Methods: This single-center, retrospective study analyzed the 10-year follow-up data of 350 patients who underwent isolated CABG between January 2011 and January 2012. The dataset comprised 52 variables, including clinical and procedural data. After preprocessing and SMOTE resampling, ensemble ML models were trained using Lasso-based feature selection; performance was compared with the full EuroSCORE II model. Result: The best-performing model utilized the top 8 features selected by Lasso regression, achieving a higher recall (0.768) and F1 score (0.462) compared to the model using the full feature set. The eight most predictive features were X-clamp time, cardiopulmonary bypass (CPB) time, first-year blood glucose levels, MACCE, elevated pulmonary artery pressure (PAH), 10th-year blood glucose levels, intervention for peripheral arterial disease (PAD), and first-year LDL levels. The ML pipeline (AUC: 0.559) achieved discrimination comparable to EuroSCORE II (AUC: 0.604) (p = 0.403). Conclusion: ML-based analysis identified procedural duration, long-term metabolic control, and PAD-related variables as key determinants of long-term mortality after CABG. Notably, MACCE history and elevated PAH further contributed to risk stratification. ML models achieved discrimination comparable to traditional risk scores while offering improved interpretability for long-term outcomes, suggesting that integration with established clinical scoring systems may enhance patient-specific mortality prediction.