Back to 2025 Abstracts
Using Machine Learning to Predict Breast Cancer-Related Lymphedema Following Axillary Lymph Node Dissection
Benjamin D. Wagner
*1, Jonlin Chen
1, Francis D. Graziano
1, Jonathan Rubin
1, Arielle N. Roberts
1, Michelle Coriddi
1, Andrea V. Barrio
2, Babak J. Mehrara
1, Danielle Rochlin
11Plastic and Reconstructive Surgery Service, Department of Surgery, Memorial Sloan Kettering Cancer Center, New York, NY; 2Breast Service, Department of Surgery, Memorial Sloan Kettering Cancer Center, New York, NY
Background: Breast cancer-related lymphedema (BCRL) is a common complication following axillary lymph node dissection (ALND), significantly impacting quality of life, physical function, and psychological well-being. Early identification of high-risk patients can facilitate timely referral for immediate lymphatic reconstruction or early decongestive therapy. Existing BCRL prediction models have limited accuracy and generalizability, and few have utilized machine learning (ML) techniques for predictive modeling. This feasibility study aimed to develop ML models to predict BCRL in patients undergoing ALND.
Methods: Demographic and clinical data were prospectively collected from female breast cancer patients undergoing ALND at Memorial Sloan Kettering Cancer Center between 2016 and 2024. BCRL was defined as relative volume change ≥10% at ≥6 months post-ALND. Data trained both preoperative and postoperative BCRL prediction models using six ML algorithms of varying explainability and complexity. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and Brier scores.
Results: A total of 479 patients (mean age: 48.3±11.2 years; BMI: 26.3±5.6 kg/m2), with a mean follow-up time of 24.4±8.8 months were included. 23.8% of patients developed BCRL an average 16.6±7.5 months post-ALND. The top-performing preoperative model achieved an AUC of 0.787 (95% CI: 0.555-0.817, Brier score 0.175), while the top postoperative model achieved an AUC of 0.710 (95% CI: 0.470-0.744, Brier score 0.210). Key predictors included age, race, and BMI in the preoperative model, and adjuvant chemotherapy and number of positive lymph nodes in the postoperative model.
Conclusion: Prediction models for BCRL following ALND were developed using ML. These models demonstrate the potential of ML for early BCRL identification and risk stratification, though also highlight the difficulties in accurately predicting BCRL development. Future studies with larger, multi-institutional datasets are warranted to enhance model predictive performance.
