Predicting pathologic ≥N2 disease in women with breast cancer.

The distinction between pN1 and ≥pN2 breast cancer impacts treatment decisions. Using data from a single institution on women with cN0 invasive breast cancer who were treated with upfront surgery, had 1-3 positive SLNs, and underwent completion ALND, we used gradient boosted trees (XGBoost) to develop a model for predicting ≥pN2 disease using clinicopathologic variables. Model performance was tested in a held-out subsample (20%) and validated using data from the National Cancer Database (NCDB). Of 3574 patients with cN0 breast cancer, 587 underwent upfront surgery and had 1-3 positive SLNs. Of these, 415 (70.7%) underwent completion ALND, with 64 (15.4%) having ≥pN2 disease. The trained algorithm had an AUC of 0.87 (95% CI: 0.74, 0.97) in the held-out test data, and 0.78 (95% CI: 0.76, 0.79) in recent NCDB data where completion ALND was much less commonly performed. The number of positive SLNs and the total number of SLNs removed had the greatest influence on model predictions in the held-out test data. The developed model effectively estimates the probability of ≥pN2 disease in cN0 patients with positive SLNs, providing guidance for the management of patients with breast cancer.