Development and validation of an explainable machine learning model based on CT-derived body composition for predicting anastomotic leakage after rectal cancer surgery

Body composition plays a crucial role in the development of postoperative complications, particularly anastomotic leakage (AL), after rectal cancer (RC) surgery. However, few studies have integrated body composition into predictive models for AL risk. This study aimed to develop and validate an explainable machine learning (ML) model based on preoperative CT-derived body composition indices to predict AL in patients undergoing laparoscopic anterior resection (LAR) for RC. This retrospective study included patients who underwent LAR for RC from two independent medical centers. The derivation cohort was enrolled from Xuzhou Central Hospital between January 2019 and March 2025, and was randomly split into a training set (70%) and an internal validation set (30%). Four feature selection methods were applied to screen key preoperative predictive factors for AL. Seven ML models were constructed using the selected variables. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis. The Shapley Additive exPlanations (SHAP) method was used to interpret feature importance. Furthermore, patients from General Hospital of Ningxia Medical University, included between January 2023 and January 2026, served as an independent external validation cohort to assess the generalizability of the optimal model. A web-based risk calculator was also developed for clinical application. A total of 745 patients were included in the derivation cohort, with 521 in the training set and 224 in the internal validation set. The overall incidence of AL was 12.8%. Six independent predictors of AL were identified: male sex, preoperative serum albumin < 35 g/L, sarcopenia, elevated visceral adipose tissue index, neoadjuvant chemoradiotherapy, and tumor location. The eXtreme Gradient Boosting (XGBoost) model showed optimal performance, with AUC values of 0.885 and 0.823 in the training and internal validation sets, respectively. The external validation cohort consisted of 271 patients, among whom 24 developed AL. The XGBoost model yielded an AUC of 0.757 in the external cohort. The combined model incorporating body composition significantly outperformed the clinical-only model across all three datasets (DeLong test p < 0.001, p = 0.007, and p = 0.011, respectively). The explainable XGBoost model integrating CT-derived body composition indices with conventional clinical variables provides acceptable preoperative discrimination for AL after LAR, with acceptable cross-center generalizability. This model may assist clinicians in identifying high-risk patients before surgery to optimize perioperative management.