D134. Application of Machine Learning to Predict Breast Cancer Related-lymphedema Development

PURPOSE: Breast cancer-related lymphedema (BCRL) is a significant long-term complication affecting 1 in 5 patients treated for breast cancer. This study aims to evaluate the role of the superficial lymphatic anatomy in predicting lymphedema development by utilizing machine learning algorithms. METHODS: A retrospective chart review was conducted to identify breast cancer patients who underwent pre-operative indocyanine green (ICG) mapping of their superficial lymphatic anatomy. The charts were evaluated to assess for lymphedema development post surgical treatment. Thirty-one features were used to describe patients' superficial lymphatic anatomy. The Random Forest machine learning algorithm was utilized for missing data imputation and to construct a predictive model for postoperative lymphedema development based on patient demographics and superficial lymphatic anatomy. RESULTS: Among the 124 included patients, 31 (25%) developed lymphedema. The model achieved an overall accuracy rate of 80%. In instances where lymphedema was absent, the model correctly predicted negatives at an 82% rate, with a 95% true negative identification rate. Conversely, for cases of lymphedema development, the model accurately classified positives at a rate of 67%, with a 33% true positive identification rate. CONCLUSION: This pilot study highlights 1) the promising predictive capabilities of employing machine learning models to forecast postoperative outcomes in the field of plastic and reconstructive surgery; 2) the utility of mapping patients' superficial lymphatic anatomy prior to treatment, to better predict and ultimately prevent lymphedema development. Future studies incorporating larger sample sizes from multiple institutions will fortify the model's predictive capacities.