A stable and interpretable radiomic model for predicting neoadjuvant chemotherapy response in HR+HER2− breast cancer

To identify stable and interpretable MRI radiomic features for developing a predictive model of neoadjuvant chemotherapy (NAC) response in HR+/HER2− breast cancer. This retrospective multicenter study included four analytical cohorts representing 686 unique patients with HR+/HER2− breast cancer, including an imaging-development cohort, an external imaging-validation cohort, an imaging–transcriptomic cohort, and an independent transcriptomic cohort. NAC response was evaluated using the Residual Cancer Burden or Miller-Payne grading system. Multiregional radiomics features were extracted from pre-treatment DCE-MRI, including intratumoral subregions, peritumoral regions, kinetic features, and multiregional spatial interaction (MSI). Five feature selection strategies (stepwise regression, LASSO, RFE, GBM-based embedded selection, and a proposed stability-driven strategy) were combined with four classifiers (LR, SVM, RF, XGBoost) to build predictive models. Stable features were defined as those selected in at least 50% of the repeated partitions by at least two of four conventional methods. Biological interpretability was explored by associating radiomic predictions with transcriptomic data and SNF molecular subtypes. The stable feature-XGBoost model outperformed all other combinations, with mean AUCs of 0.902 and 0.855 in the training and testing sets, respectively, and an AUC of 0.783 in the external validation cohort. Stable features mainly derived from intratumoral subregions and MSI, underscoring the importance of spatial heterogeneity in predicting treatment response. Radiogenomic analysis revealed that MRI predicted non-responders were predominantly of the aggressive SNF3 or SNF4 subtypes and showed activation of cell-cycle, RTK, and PI3K-Akt/mTOR pathways, whereas predicted responders were mainly SNF1 or SNF2 and exhibited a more immune-active profile. A stability-driven, multiregional MRI radiomics model enables improved prediction of NAC response, offering an interpretable imaging-based approach for individualized decision-making in HR+/HER2− breast cancer.