To develop an interpretable MRI-based machine learning model combining radiomic and clinicoradiological features for predicting human epidermal growth factor receptor 2 (HER2) status, and to explain predictions using Shapley additive explanations (SHAP) analysis. This retrospective study included 322 breast cancer patients from two centers. Patients were divided into training (n = 153), internal validation (n = 39), and external validation (n = 130) sets. Radiomic features were extracted from T2-weighted fat-suppressed (T2WI-FS) and dynamic contrast-enhanced (DCE) MRI. After normalization and feature selection with LASSO, logistic regression was used to build radiomics, clinicoradiological, and combined models. Performance was assessed with the area under the curve (AUC), calibration, and decision curve analysis (DCA). SHAP analysis was applied for interpretability. Among 3290 extracted features, 24 were retained for modeling. The combined clinicoradiological–radiomics model achieved the best performance (AUCs: 0.899 in training, 0.891 in internal validation, and 0.840 in external validation), outperforming single-modality models. Calibration and DCA confirmed good agreement and a higher clinical benefit. SHAP analysis showed DCE-derived features contributed most to the prediction. The interpretable clinicoradiological–radiomics model demonstrated robust performance in predicting HER2 expression in breast cancer and may assist individualized treatment planning.
Xia et al. (Sun,) studied this question.