Interpretable machine learning for predicting residual cancer burden following neoadjuvant therapy in HER2 positive breast cancer: A real-world study
Accurate early prediction of response to neoadjuvant therapy (NAT) in HER2-positive breast cancer could support pre-treatment risk stratification. We conducted a retrospective real-world study of 692 patients treated between November 2018 and May 2025 at a tertiary center to develop interpretable machine-learning models for pre-treatment prediction of residual cancer burden (RCB; favorable: RCB 0–I vs. poor: RCB II–III). Least absolute shrinkage and selection operator identified nine routinely available predictors, and ten algorithms were trained with 10-fold cross-validation and grid search. Among them, a gradient boosting machine (GBM) achieved the best discrimination in the independent test set (AUC 0.796) with superior calibration and net clinical benefit. SHAP highlighted HER2 status, hormone-receptor status and clinical nodal stage as the strongest contributors, consistent with biological expectations. We implemented a Shiny web application for research and educational use. In a propensity score–matched comparison of trastuzumab–pertuzumab versus trastuzumab– pyrotinib regimens, no significant difference in achieving favorable RCB was observed. This interpretable GBM model shows potential for pre-treatment risk stratification in patients with HER2-positive breast cancer; however, its current predictive performance does not support stand-alone clinical decision-making.
Authors
- 刘守贵
- 林舜国
- 郑彩红
- 韩忠华
- Chengbin Fu
- Yu Chen (ORCID: https://orcid.org/0000-0003-0747-3479)
- Jinhua Huang
- Yubin Chen
- Liantao Guo
- Junchao Lin
- Honglin Zhang
- Zhinan Chen
- Chenyan Li
- Chunsen Xu
- Hui Han
- Jingwen Zhuang
Institutions
- Fujian Medical University (CN)
- Union Hospital (CN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-19
- DOI
- https://doi.org/10.1038/s41598-026-79913-0
- Primary Topic
- Breast Cancer Treatment Studies
- Type
- article
- Field-Weighted Citation Impact
- 0.00