Development and validation of an interpretable stacking-based risk model for breast cancer- related lymphoedema: a cross-sectional study.

Background: Breast Cancer-Related Lymphedema (BCRL) is one of the common complications after breast cancer treatment. It is characterized by irreversibility and high disability. Currently, early identification of high-risk patients remains a clinical pain point. This study focuses on BCRL risk prediction, integrating machine learning techniques with multi-dimensional clinical data to construct a practical prevention and treatment tool, providing a reference for precise prevention and treatment of BCRL. Methods: Patients who underwent breast cancer surgery and treatment from January 2021 to May 2025 were included. The diagnostic criterion for BCRL was "the difference in upper limb circumference ≥2 cm". Using LASSO regression, Support Vector Machine Recursive Feature Elimination (SVM-REF), Random Forest (RF) and Boruta, 34 variables were subjected to feature selection, and features selected by all four algorithms were used as modeling variables. The SHAP value analysis was performed to construct and validate an interpretable BCRL risk prediction model using a Stacking ensemble learning approach, and a visual Web-based risk prediction tool was developed. Results: A total of 570 eligible patients were included. They were stratified into the training set (420 cases) and the temporal validation set (150 cases). Seven core characteristics were selected, including clinical characteristics (BMI, tumor clinical stage, surgical method, axillary lymph node handling method and diabetes) and modifiable characteristics (patients' awareness of BCRL and the time spent holding their mobile phones daily after surgery). The Stacking model performed the best, featuring high accuracy, high precision and interpretability. The ROC-AUC values for the training and validation sets were 0.911 and 0.868, respectively, and the PR-AUC values were 0.857 and 0.789, respectively. The Web tool, StackBCRL, can be accessed online. Conclusions: Through the construction of an interpretable BCRL risk prediction model based on Stacking ensemble learning by integrating postoperative behavioral characteristics (the time spent holding their mobile phones daily after surgery and patients' awareness of BCRL), the developed BCRL risk assessment system achieved real-time individualized risk prediction and visualization. It provides a simple and practical method for large-scale, low-cost early screening of BCRL, is suitable for promoting clinical application and provides an innovative solution for BCRL prevention and control.