Comparative evaluation of machine learning models for predicting postoperative infection after open pulmonary lobectomy

To characterize postoperative infection (POI) following open pulmonary lobectomy and comparatively evaluate the performance of multiple machine learning (ML) algorithms for POI prediction using clinical information available at the end of surgery. The study included 3,160 patients who underwent open pulmonary lobectomy at West China Hospital from 2017 to 2025. Fourteen prespecified clinical variables available by the end of surgery were used as predictors. Nine ML algorithms were developed and compared using cross-validation, hyperparameter tuning, and SMOTE. Model performance was assessed for discrimination, classification, calibration, and clinical utility, with SHAP analysis used to characterize feature contributions. Among the 3,160 patients, 369 (11.68%) developed POI. Respiratory tract infection and surgical site infection were the most common infection categories, accounting for 6.17% and 5.38%, respectively; organ/space infections accounted for 70.59% of surgical site infections. Among the nine ML models, SVM achieved the highest AUC (0.645, 95% CI 0.589–0.700), followed by RF (0.632, 95% CI 0.573–0.691), LR (0.624, 95% CI 0.567–0.682). KNN showed the highest accuracy (0.806) and specificity (0.887), SVM showed the highest sensitivity (0.586) and negative predictive value (0.921), and RF achieved the highest F1 score (0.276).Calibration performance varied across models, with Brier scores ranging from 0.156 to 0.241. Decision curve analysis showed positive net benefit at low threshold probabilities (0.01–0.11). SHAP analysis identified age, surgical duration, and preoperative hospital stay duration as consistently important features across SVM, MLP, and RF models. ML approaches showed potential for POI risk stratification after open pulmonary lobectomy, with SVM achieving the highest discrimination. Age, surgical duration, and preoperative hospital stay duration were consistently important features, supporting further validation of these models. Not applicable.