Effect of a machine learning prediction model on the false-negative rate of sentinel lymph node biopsy for clinically node-positive breast cancer after neoadjuvant chemotherapy

Machine learning (ML) models can be used to predict axillary pathological complete responses (pCRs) in clinically node-positive (cN+) breast cancer after neoadjuvant chemotherapy (NAC). We developed an ML model combining clinicopathological characteristics and axillary ultrasound features before and after NAC to predict the possibility of axillary pCR in NAC-treated cN + breast cancer.

Patients with cN + breast cancer who received NAC were categorized into training and verification cohorts (7:3 ratio). Independent predictors of axillary pCR were selected using univariate and multivariate logistic regression analyses; six ML models were developed to predict pCRs. Another independent prospective cohort of 126 patients was enrolled to evaluate false-negative cases when the best-performing model was used to guide patient selection for sentinel lymph node biopsy (SLNB).

Overall, 614 patients with breast cancer were included. Age, menstrual status, cN staging before NAC, molecular subtype, histological grade, tumor shrinkage percentage after NAC, and lymph node cortical thickening ≥3 mm before and after NAC were independent predictors of axillary pCR. A multilayer perceptron model had the best stability and predictive performance, yielding the highest area under the receiver operating characteristic curve of 0.801 (training) and 0.774 (validation). When applying this model to the independent test cohort to guide patient selection for SLNB, the false-negative rate was reduced from 22.2 % to 1.4 %.

We established an ML model with excellent performance to predict pCR in cN + breast cancer after NAC. The ML model demonstrated potential to reduce the false-negative rate of single-tracer SLNB when used as an adjunct to clinical judgment.