Machine learning-guided prevention and management of low anterior resection syndrome: Development of an XGBoost prediction model and validation via SHAP

Objective: Colorectal cancer is among the top three cancers in terms of incidence and mortality worldwide. Although laparoscopic and robotic-assisted sphincter-preserving surgeries reduce permanent colostomy rates to under 20%, 60%–80% of patients develop postoperative low anterior resection syndrome (LARS), nearly half of whom progress to major LARS. This study aims to develop a high-precision machine learning model for predicting LARS, thereby optimizing the early identification, prevention, and management of major LARS in rectal cancer patients, providing a reliable tool for personalized clinical decision-making. Methods: This retrospective study screened 3,986 rectal cancer patients who underwent laparoscopic and robotic-assisted sphincter-preserving surgeries from January 2012 to January 2022. Key predictors were identified via LASSO regression to develop an XGBoost machine learning model for major LARS prediction, which was validated via SHapley additive exPlanations (SHAP). Results: The XGBoost model achieved 93% accuracy for major LARS prediction, with 84% precision, 74% recall, and an F1 score of 0.78, outperforming POLARS (69% accuracy, 82% precision, 36% recall, F1 score of 0.5). SHAP analysis confirmed that tumor height was the strongest predictor, followed by age at surgery, stoma status, preoperative radiotherapy, and gender. The model enabled real-time risk stratification, reducing overtreatment in non-LARS and minor LARS patients in clinical application. The model has been integrated into a user-friendly offline software (XGBoostLARS) and has been applied to the early clinical identification, prediction, and management of LARS. Conclusion: This high-precision XGBoost model optimizes the early identification, prevention, and management of major LARS, leading to new progress in personalized treatment for rectal cancer survivors.