Machine learning-based prediction of 1-year mortality risk after off-pump coronary artery bypass grafting.

Background: Coronary heart disease (CHD) has gradually become one of the main causes of death among patients worldwide. Off-pump coronary artery bypass grafting (OPCABG) has been increasingly applied due to its avoidance of cardiopulmonary bypass. However, there is currently no study that predicts the postoperative mortality risk for patients undergoing OPCABG. To fill this gap, we identified the independent risk factors associated with poor 1-year survival outcomes in patients undergoing OPCABG and developed an effective machine learning (ML) model for prediction. Methods: Patient data were extracted from the Medical Information Mart for Intensive Care (MIMIC)-IV database. Multivariate Cox regression analysis was performed to identify independent risk factors for adverse postoperative survival outcomes in patients undergoing OPCABG. Based on these features, five survival ML models were developed, including Gradient Boosting Machine (GBM), least absolute shrinkage and selection operator-Cox regression (Lasso-Cox), Cox Boosting (CoxBoost), eXtreme Gradient Boosting (XGBoost), and partial least squares regression-Cox (PLSRCox). Model performance was assessed at 3 months, 6 months, and 1 year after surgery across the training, testing, and validation cohorts, respectively. The optimal model was further interpreted using Shapley Additive Explanations (SHAP) visualization. Results: A total of 2,280 patients who underwent OPCABG were identified from the MIMIC-IV database and randomly divided into training and testing sets in a 7:3 ratio. In the training cohort, multivariate Cox regression analysis identified creatine kinase (CK), red cell distribution width (RDW), total bilirubin (TBIL), alanine aminotransferase (ALT), chronic kidney disease (CKD), anion gap, and aspartate aminotransferase (AST) as independent risk factors for adverse postoperative survival outcomes. Among the five developed survival ML models, the CoxBoost model achieved areas under the receiver operating characteristic curve (AUCs) of 0.955, 0.958, and 0.961 at 3, 6, and 12 months, respectively, in the training set. The time-dependent concordance index (C-index) and AUC indicated strong model performance. In the testing and validation cohorts, CoxBoost also demonstrated excellent predictive capability across all time points. Conclusions: The CoxBoost model, constructed using CK, RDW, TBIL, ALT, CKD, anion gap, and AST as key predictors, effectively predicts the risk of adverse 1-year survival outcomes in patients undergoing OPCABG. However, this study has an imbalance in the sample size. Although the survival Synthetic Minority Oversampling Technique (SMOTE) was adopted to address this issue, it may still have an impact on the model's performance. We look forward to more research in the future to further explore this problem.