Abstract
OBJECTIVE: To develop and externally validate an interpretable multiparametric MRI-based radiomic-clinical model using Shapley Additive Explanations (SHAP) methodology for early prediction of breast cancer sensitivity to neoadjuvant chemotherapy (NAC). METHODS: This retrospective multicentric study enrolled 223 breast cancer patients from three medical centers. Patients underwent pretreatment multiparametric MRI (DCE-MRI and DWI sequences) with Miller-Payne grades 4-5 defining NAC-sensitive. Manual tumor segmentation generated regions of interest for extracting 2,396 radiomic features per patient. Feature selection integrated reproducibility analysis (ICC > 0.7), univariable significance testing (p < 0.01), LASSO regression, and hierarchical clustering. A support vector machine (SVM) model incorporated optimized radiomic signatures and clinical variables. SHAP methodology provided global feature importance interpretation and individualized prediction explanations. RESULTS: The integrated radiomic-clinical model demonstrated superior performance to standalone clinical (AUC 0.720) and radiomic (AUC 0.833) models in the internal validation set, achieving an AUC of 0.904 (95% CI: 0.816-0.991). This advantage persisted in external validation (AUC 0.928, 95% CI: 0.874-0.982). SHAP analysis identified wavelet_HHL_glcm_Correlation_DCE as the predominant predictive feature, with high values significantly correlating to NAC-sensitive. A clinical nomogram translated model outputs into quantifiable risk probabilities, where total scores ≥130 indicated > 95% sensitivity likelihood. CONCLUSION: The SHAP-explainable radiomic-clinical model provides a clinically applicable, noninvasive tool for pretreatment stratification of NAC sensitivity. This approach enhances personalized therapeutic decision-making while minimizing unnecessary treatment toxicity.