Development and validation of a machine learning-based predictive model for early outcomes following combined suction-assisted lipectomy and lymphovenous anastomosis in breast cancer-related lymphedema: a retrospective cohort study.

Background: Breast cancer-related lymphedema (BCRL) significantly compromises quality of life. Although combined suction-assisted lipectomy (SAL) and lymphovenous anastomosis (LVA) is effective, outcomes vary considerably among patients. Currently, tools for early postoperative risk stratification are lacking. Methods: We retrospectively reviewed data from BCRL patients who underwent combined SAL and LVA at Beijing Shijitan Hospitalfrom June 2018 to June 2025. Predictive variables were selected using the Least Absolute Shrinkage and Selection Operator (LASSO) regression combined with bootstrap resampling (B = 1,000). Seven algorithms-including logistic regression (LR), decision tree (DT), random forest (RF), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), support vector machine (SVM), and artificial neural network (ANN) were compared. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration plots, decision curve analysis (DCA), and Brier score. SHapley Additive exPlanations (SHAP) analysis was conducted for model interpretation, and a web-based prediction tool was developed. Results: A total of 300 patients were enrolled (training set: n=211; validation set: n=89). The rate of satisfactory outcomes at 6 months was 72.3%. LASSO and bootstrap validation identified three stable predictors: postoperative excess limb volume (selection frequency: 100%), disease duration (83.6%), and disease severity grade (83.4%). Among the seven models, SVM exhibited the optimal balance of discrimination and clinical utility in the validation set: AUC 0.891 (95% CI: 0.812-0.970), sensitivity 90.8%, specificity 62.5%, F1-score 0.887, and Brier score 0.119. DCA indicated net clinical benefit within the threshold range of 0.1-0.6. Although ANN achieved a higher AUC than SVM (0.903 vs. 0.891) (DeLong test, P = 0.532), SVM demonstrated superior sensitivity (90.8% vs. 89.2%), F1-score (0.887 vs. 0.879), and Cohen's kappa (0.555 vs. 0.531).Furthermore, SVM's structural risk minimization principle conferred superior generalization stability compared with ANN's empirical risk minimization, making it more suitable for small-sample clinical settings. SHAP analysis revealed that postoperative excess volume was the strongest predictor. Conclusion: The 3-variable SVM model effectively predicts 6-month outcomes following combined surgery for BCRL. Integrated with SHAP analysis and a web-based tool, this model enables early postoperative risk stratification to identify high-risk patients requiring closer monitoring, providing a reference for future standardized rehabilitation protocols.