A Practical Machine Learning Model for Predicting Neoadjuvant Response in HER2-Positive Breast Cancer

Area of Science:

• Oncology
• Radiology
• Machine Learning

Background:

• Pathological complete response (pCR) is a key prognostic marker for HER2-positive breast cancer (BC) after neoadjuvant chemotherapy (NAC).
• Predicting NAC response using pre-treatment data is challenging due to limited reliable predictors.
• This study explores machine learning (ML) for predicting NAC response based on routinely available clinical and MRI data.

Purpose of the Study:

• To evaluate the efficacy of clinicopathologic and baseline MRI variables in predicting NAC response in HER2-positive BC.
• To compare seventy classical ML classifiers for their ability to predict pCR.

Main Methods:

• Retrospective analysis of 112 HER2-positive BC patients treated with NAC.
• Evaluation of 14 pre-treatment variables including hormone receptor status, tumor grade, Ki67, nodal status, and baseline MRI features.
• Comparison of 70 ML classifiers using leave-one-out cross-validation (LOOCV) and performance metrics like accuracy and AUROC.

Main Results:

• A support vector machine (SVM) with a radial basis function (RBF) kernel achieved the highest accuracy (82.1%) and AUROC (83.3%).
• Key predictors included hormone receptor (PR, ER) status, Ki67, and baseline MRI characteristics.
• The ML model demonstrated meaningful predictive performance using routinely available pre-treatment data.

Conclusions:

• Machine learning shows promise in predicting pCR for HER2-positive BC before treatment initiation.
• The approach can potentially enhance pre-treatment clinical risk assessment using existing data.
• Further validation through prospective multicentre studies is necessary for clinical implementation.