Predicting non-sentinel lymph node involvement in breast cancer patients

531 Background: Current convention is to perform a completion axillary lymph node dissection (ALND) for invasive breast cancer (BC) patients (pts) with positive sentinel lymph node(s) (+SLN), even though <50% will have non-sentinel lymph node metastasis (+NSLN). Our goal was to develop and compare predictive models of NSLN status among +SLN pts to identify for whom completion ALND may be omitted. Methods: We constructed 3 models by recursive partitioning with receiver-operating-characteristic curves (RP-ROC), boosted classification and regression trees (CART), and a multivariate logistic regression (MLR) informed by CART. Models were developed using a multi-institutional database of 1,040 BC pts who underwent SLN biopsy and completion ALND at academic or community hospitals as part of a prospective, consented study. Accuracies were compared to the Memorial Sloan-Kettering Breast Cancer Nomogram (Nomogram). Results: 976 BC pts had evaluable SLNs and 285 had +SLNs; 101(35.4%) +SLN pts had +NSLNs. Tumor size, lymphovascular invasion (LVI), and SLN metastasis size were the best predictors of risk (multivariate P-values<0.001) of +NSLN. 39 of 156 +SLN pts (25%) with T1 tumors had +NSLNs vs. 62 of 129 (48%) with T2/T3 tumors. 73.7% with LVI vs 19.5% without LVI had +NSLNs. 4.7% of pts with isolated tumor cells, 42% with micrometastasis and 71% with macrometastasis in SLN had +NSLNs. MLR informed by CART identified 2 highly predictive variables, the LVI × size of SLN metastasis [OR 4.73(3.11–7.20, 95%CI), P<0.001] and tumor size × size of SLN metastasis [OR 1.18(1.10–1.26, 95%CI), P<0.001]. While RP-ROC and boosted CART stratified pts into low-risk (4.3%-9.9%), moderate-risk (33.3%-42.9%), and high-risk (62.2%-93%) groups, MLR predicted NSLN status with accuracy superior to RP-ROC, boosted CART, and the Nomogram: 83.3%, 76.7%, 67.7%, and 76.7%, respectively, after 10-fold cross validation. The Nomogram’s sensitivity was significantly inferior to those of RP-ROC, boosted CART, and MLR: 53.8%, 78.8%, 78.2%, and 78.0%, respectively. Conclusions: In pts with +SLNs, interactions between clinicopathologic characteristics are highly informative in predicting risk of +NSLN. However, neither our methods nor the Nomogram achieved sufficient accuracy to recommend a change in current clinical practice. No significant financial relationships to disclose.