Which prediction models best identify additional axillary disease after a positive sentinel node biopsy for breast cancer?

Abstract

To determine which web-based model best identifies women at low risk of further axillary disease after a positive sentinel lymph node (SLN+) biopsy. 673 women with T1-2cN0M0 SNB+ breast cancer who underwent completion axillary dissection (AxD) were identified. A subgroup not eligible to avoid AxD as part of the Z0011 study was defined (Z0011 exclusion group). Predicted risk of further axillary disease was generated using seven web-based models. "Low risk" was defined as a ≤10% risk of further axillary disease. False negative ("low risk" prediction but AxD+) rates (FNRs), area under the receiver operating characteristic curve (AUC), and Brier score were determined for each model. 6 of 7 models identified "low risk" patients but FNRs ranged from 14 to 30%. The Stanford and Memorial Sloan-Kettering (MSKCC) models had the best FNRs. FNRs were lower with SLN micrometastasis (7-15%) and higher in the Z0011 exclusion group (21-41%). All models under-predicted further nodal disease in low risk patients and over-predicted in higher-risk patients. The Stanford and MSKCC models were able to identify women with SLN micrometastasis with a ≤10% FNR. Models were not able to accurately identify low risk women from a cohort that would have been excluded from Z0011.