Back to 2026 Abstracts
Machine Learning Prediction of Breast Cancer-Related Lymphedema
Benjamin D. Wagner
*1, Ariel Gabay
1, Jonlin Chen
1, Andrea V. Barrio
2, Thomas Amburn
2, Natalia Polidorio
2, Tiana Le
2, Lillian A. Boe
3, Geoffrey Hespe
1, Jonas Nelson
1, 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; 3Biostatistics Service, Department of Epidemiology and Biostatistics, Memorial Sloan Kettering Cancer Center, New York, NY
Background: Breast cancer-related lymphedema (BCRL) is a chronic and morbid sequela of breast cancer treatment that adversely affects quality of life, physical function, and psychological well-being. Early identification of high-risk patients may enable timely referral for immediate lymphatic reconstruction or early decongestive therapy. Existing prediction models have limited accuracy and generalizability, and few incorporate machine learning (ML). This study developed and evaluated ML models to predict BCRL using routinely available demographic, clinical, and treatment variables.
Methods: Demographic and clinical data were collected from patients undergoing axillary lymph node dissection (ALND) and/or sentinel lymph node biopsy (SLNB) at Memorial Sloan Kettering Cancer Center (2010-2024). Five supervised ML algorithms (LightGBM, XGBoost, Random Forest, AdaBoost, and logistic regression) were trained to predict BCRL. Performance was assessed using area under the curve (AUC), sensitivity, specificity, F1 score, and Brier score with 95% confidence intervals. Pairwise comparisons used Wald statistics. Model interpretability was evaluated with SHAP analysis.
Results: Among 20,024 patients (median follow-up, 61.7 months), median age was 55 years and median BMI was 25.0 kg/m
2. BCRL developed in 27.8% of patients undergoing ALND and 4.1% after SLNB. Model performance was moderate to strong (AUC, 0.79-0.83), with XGBoost and AdaBoost achieving the highest AUC (0.83). Sensitivity ranged from 0.63 to 0.67, with high specificity (0.81-0.84). Models were well calibrated (Brier score, 0.17-0.21). SHAP analysis identified axillary surgery type, radiation therapy, chemotherapy timing, age, and BMI as key predictors.
Conclusion: ML models demonstrate reliable performance for identifying patients at elevated risk of BCRL using standard clinical variables. With further refinement and external validation, these models may support individualized risk stratification and guide early preventive interventions.
Back to 2026 Abstracts
