Identification and validation of an explainable machine learning model for early postoperative pulmonary complications after esophagectomy in patients with esophageal cancer.

Background: Postoperative pulmonary complications (PPCs) constitute a major adverse outcome following esophagectomy for esophageal cancer (EC). This study aimed to develop and validate an interpretable machine learning (ML) model for early prediction of PPCs within 30 days post-esophagectomy. Methods: A retrospective cohort of 975 patients undergoing esophagectomy was enrolled and divided into training and internal testing sets in a 7:3 ratio. An independent prospective cohort of 417 participants served as the external validation cohort. Feature selection was performed using the least absolute shrinkage and selection operator (LASSO) and Boruta algorithms to identify key predictors. Eight ML models were subsequently constructed using the selected features. Model performance was assessed through the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, F1-score, calibration curves, and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) method was employed to interpret the ML models. Results: Among the retrospective cohort, the incidence of PPCs was 43.2% (421/975). Feature selection identified seven independent predictors of PPCs. Among the ML models evaluated, category boosting (CatBoost) demonstrated superior performance, achieving an AUC of 0.890 [95% confidence interval (CI): 0.853-0.928] in internal testing cohort and 0.888 (95% CI: 0.856-0.920) in external validation. Calibration curves indicated strong agreement between predicted and observed outcomes, while DCA confirmed clinical utility. SHAP analysis revealed surgical duration, neoadjuvant therapy, and advanced age as the three most influential predictors of PPCs. Conclusions: The CatBoost-based model exhibited robust predictive accuracy for early PPCs following esophagectomy, achieving an AUC of 0.888 in external validation. Integration of SHAP analysis enhanced interpretability, offering actionable insights for risk stratification and guiding perioperative clinical-making. This tool may facilitate timely interventions to mitigate postoperative morbidity in EC patients.