Machine learning-based prediction of 30-day outcomes after surgery for left-sided infective endocarditis

Surgery is often essential for patients with left-sided infective endocarditis (IE). Accurate preoperative risk assessment remains challenging. This study aimed to develop and internally validate machine learning (ML) models to predict 30-day postoperative mortality or major complications in patients undergoing surgery for left-sided IE. A retrospective analysis of adult patients between May 2009 and May 2025 was performed. The primary outcome was a composite of all-cause death or major complications within 30 days, including postoperative stroke, malignant arrhythmia, prolonged ventilation, acute kidney injury requiring dialysis, shock, and multiple organ dysfunction syndrome. An ML-based pipeline was used to select variables. Eight ML algorithms were evaluated with 6 strategies to address class imbalance. Model performance was assessed in terms of discrimination, calibration, and overall prediction error using out-of-fold (OOF) predictions in the training set, and the selected final model was subsequently evaluated in the test set. Model interpretation was performed using SHapley Additive exPlanations (SHAP) analysis, and an online tool was developed for clinical implementation. Among the 992 included patients, 100 (10.1%) experienced the primary outcome. Ten predictors were selected for model development. Of the 56 candidate models evaluated, 13 met the screening criteria based on OOF predictions. The categorical boosting (CatBoost) model trained with random over-sampling (RO) showed the lowest OOF Brier score and was selected as the final model. In the test set, CatBoost-RO achieved an AUC of 0.814 (95% CI: 0.724–0.904), a calibration intercept of -0.036 (95% CI: -0.474–0.401), a calibration slope of 0.995 (95% CI: 0.658–1.331), and a Brier score of 0.065 (95% CI: 0.053–0.078). SHAP analysis identified total bilirubin and N-terminal pro-B-type natriuretic peptide as the most influential predictors. The CatBoost model trained with RO showed favorable discrimination, calibration, and low prediction error for estimating the risk of 30-day death or major complications after surgery for left-sided IE, which may support preoperative risk stratification in this population, although external validation is warranted.