Objective To construct a combined model for predicting 2-year disease progression after neoadjuvant chemotherapy (NAC) in breast cancer based on multiparametric MRI radiomics, clinicopathological features, and pectoralis muscle index (PMI), and to evaluate its predictive performance. Methods The clinical and imaging data of 215 breast cancer patients receiving NAC were retrospectively analyzed, and they were divided into a disease-progression group (n=44) or a progression-free survival group (n=171) according to whether disease progression occurred within 2 years after NAC. PMI was measured based on chest CT; independent-sample t-test and χ2-test were adopted to compare baseline characteristics between the two groups; region of interest delineation and multiparametric MRI radiomic feature extraction were performed using ITK-SNAP software and Deepwise Multimodal Research Platform, and the final radiomic feature set for modeling was determined through step-wise screening using intra-class correlation coefficient, Pearson correlation coefficient between features, and L1-regularized logistic regression. On the basis of the feature set, three progressive binary prediction models were constructed: pure radiomic model (19 radiomic features only), radiomic-clinical combined model (+clinicopathological features), and radiomic-clinical-PMI comprehensive model (+PMI). The decision-tree algorithm (Gini coefficient for splitting, maximum tree depth=5, minimum split sample size=10, minimum leaf-node sample size=5) and 10-fold stratified cross-validation with the fixed random seed set to 1 were used to assess model performance; evaluation metrics included area under the receiver operating characteristic (ROC) curve (AUC), accuracy, sensitivity, specificity, and Brier score. SHAP algorithm was applied for feature contribution analysis of the optimal-performance model. Results Statistically significant differences were observed between the two groups in estrogen receptor expression status, progesterone receptor expression status, Ki-67 proliferation index, molecular subtype, and PMI (all P<0.05). A total of 9056 original radiomic features were extracted, and 19 optimal features were retained for model construction after step-wise screening. The ROC AUCs of the pure radiomic model, radiomic-clinical combined model, and radiomic-clinical-PMI comprehensive model were 0.702, 0.723, and 0.759, respectively; their Brier scores were 0.067, 0.074, and 0.069, respectively. SHAP analysis demonstrated that the top 3 features contributing most to the radiomic-clinical-PMI comprehensive model were sequentially: apparent diffusion coefficient sequence texture feature wavelet_glrlm_wavelet-HLL-ShortRunLowGrayLevelEmphasis, Ki-67 proliferation index, and T2WI-FS sequence texture feature wavelet_glszm_wavelet-HHL-SmallAreaLowGrayLevelEmphasis.Conclusion The combined model constructed from multiparametric MRI radiomics, clinicopathological features, and PMI can effectively predict 2-year disease progression after NAC in breast cancer patients, with better predictive performance than the pure radiomic model and radiomic-clinical combined model. It provides a convenient and non-invasive auxiliary tool for prognostic evaluation after breast cancer NAC.

无