Machine learning-based model to predict liposuction outcomes in unilateral breast cancer-related lymphedema
TL;DR
Although not possessing the highest AUC, the SVM demonstrated exceptional resistance to overfitting, evidenced by a minimal AUC decrement of merely 0.0298 from the training to the validation set, underscoring its tangible clinical utility.
Abstract
Background Breast cancer-related lymphedema (BCRL) is a common and disabling complication after breast cancer surgery, with substantial effects on limb function and quality of life. Liposuction is an established option for selected patients with chronic BCRL. However, postoperative response is heterogeneous. This study aimed to develop a machine learning model to predict liposuction efficacy in patients with unilateral BCRL. Methods We analyzed 623 unilateral BCRL cases undergoing liposuction at Beijing Shijitan Hospital, randomly splitting them 7:3 into training (n=437) and validation (n=186) cohorts. Least absolute shrinkage and selection operator (LASSO) regression with cross-validation guided feature selection. Seven algorithms—logistic regression (LR), support vector machines (SVM), decision trees (DT), artificial neural networks (ANN), LightGBM, XGBoost, and random forests (RF)—were trained and benchmarked. Performance via AUC, calibration, DCA, and Brier score identified SVM as optimal. SHAP interpretation facilitated deployment of a web-based calculator. Results The overall rate of favorable outcomes following liposuction was 64.7%. Cross-validated Lasso analysis, factoring in clinical validity and predictive importance, yielded four key predictors: history of erysipelas, preoperative affected-to-unaffected limb volume difference (Preoperative_difference), extracellular water ratio of the affected limb (r-ECW%), and body fat percentage. Among the seven models evaluated, the SVM exhibited the most balanced overall performance, achieving an accuracy of 77.42%, precision of 75.00%, specificity of 90.00%, F1-score of 0.632, Brier score of 0.163, and an AUC of 0.818 (95% CI: 0.757–0.876). Although not possessing the highest AUC, the SVM demonstrated exceptional resistance to overfitting, evidenced by a minimal AUC decrement of merely 0.0298 from the training to the validation set. Both calibration and decision curve analyses corroborated its robust generalizability, underscoring its tangible clinical utility. Conclusions An SVM model predicting surgical outcomes in BCRL was created and integrated into a user-friendly online tool. This calculator guides surgical choices based on predictive outputs, offering a valuable reference for refining patient management and treatment plans.