Machine learning-based hybrid risk estimation system (ERES) in cardiac surgery: Supplementary insights from the ASA score analysis

Accurate prediction of postoperative mortality risk after cardiac surgery is essential to improve patient outcomes. Traditional models, such as EuroSCORE I, often struggle to capture the complex interactions among clinical variables, leading to suboptimal performance in specific populations. In this study, we developed and validated the Ensemble-Based Risk Estimation System (ERES), a machine learning model designed to enhance mortality prediction in patients undergoing coronary artery bypass grafting and/or valve surgery. A retrospective analysis of 543 patients was performed using six machine learning algorithms applied to preoperative clinical data to assess predictive accuracy and clinical outcomes. Feature selection techniques, including Gini importance, Recursive Feature Elimination, and Adaptive Synthetic Sampling, were employed to improve accuracy and address class imbalance. ERES, which utilizes 15 key features, demonstrated superior predictive performance compared to EuroSCORE I. Calibration plots indicated more accurate probability estimates, whereas SHAP analysis identified creatinine, age, and left ventricular ejection fraction as the most significant predictors. The decision curve analysis further confirmed the superior clinical utility of ERES across a range of decision thresholds. Additionally, although the American Society of Anesthesiologists (ASA PS) score had limited predictive power independently, its combination with EuroSCORE I enhanced the predictive performance. Integrating machine learning models like ERES into clinical practice can improve decision making and patient outcomes although external validation is warranted for broader implementation.Author summary: Accurately predicting mortality risk in patients undergoing cardiac surgery is critical for improving clinical outcomes and guiding preoperative decision-making. Traditional scoring systems such as EuroSCORE I, while widely used, often fall short in capturing the complex, non-linear relationships among patient variables. In this study, we developed the Ensemble-Based Risk Estimation System (ERES), a machine learning-based model that leverages preoperative clinical and laboratory data to more precisely estimate short-term mortality risk after coronary artery bypass and/or valve surgery. Using advanced techniques such as feature selection, data balancing, and model calibration, ERES outperformed both traditional models and other machine learning algorithms across multiple performance metrics. The model identified creatinine, age, and ejection fraction as key predictors of postoperative outcomes. Additionally, although the ASA score alone showed limited predictive ability, its integration with EuroSCORE I improved the overall model performance. Our findings suggest that machine learning models like ERES offer a more personalized and accurate approach to risk stratification, especially in complex or high-risk cases. We strongly believe that supporting studies like this will contribute to the development of more effective health policies and evidence-based clinical practices.