ObjectiveTo investigate the value of a predictive model for sentinel lymph node (SLN) metastasis after neoadjuvant therapy (NAT) based on the radiomic features from multi-modality MRI in combination with clinicopathologic data. MethodsThe clinical data and MRI images of breast cancer patients (initially diagnosed with cN0, all underwent NAT and surgical treatment) from two hospitals (Affiliated Hospital of Southwest Medical University and Suining Central Hospital) from January 2018 to September 2024, were retrospectively collected. The radiomic features from the multi-modality images, including T2-weighted short tau inversion recovery (T2STIR), diffusion-weighted imaging (DWI), dynamic contrast-enhanced (DCE), were extracted and selected. The predictive models for SLN metastasis after NAT were constructed using four algorithms: LightGBM, XGBoost, support vector machine (SVM), and logistic regression (LR), in combination with clinicopathologic data. The models were evaluated for performance and interpretability using receiver operating characteristic (ROC) curves, calibration curves, decision curve analysis, and Shapley additive explanation (SHAP) analysis. ResultsA total of 236 breast cancer patients were enrolled in this study. Among them, 216 patients from the Southwest Medical University were subdivided in an 8:2 ratio into the training set (173) and internal validation set (43), while 20 patients from the Suining Central Hospital served as the external validation set. Among the clinical and pathological features, lymphovascular invasion (LVI) (P<0.001), perineural invasion (PNI) (P=0.002), and Ki-67(P<0.001) were identified as the risk factors for SLN metastasis after NAT. The predictive models utilizing multi-modality MRI and clinicopathologic data yielded area under the ROC curve (AUC) values for the internal and external validation sets of 0.750 [95%CI=(0.395, 1.000)]/0.625 [95%CI=(0.321, 0.926)] for LightGBM, 0.878 [95%CI=(0.707, 1.000)]/0.778 [95%CI=(0.525, 0.986)] for XGBoost, 0.641 [95%CI=(0.488, 0.795)]/0.681 [95%CI=(0.345, 1.000)] for SVM, and 0.667 [95%CI=(0.357, 0.945)]/0.583 [95%CI=(0.196, 0.969)] for LR. XGBoost demonstrated the best predictive performance. Further SHAP analysis revealed that LVI, the minimum value of first-order features from T2STIR-MRI, and platelet count were the key features influencing the predictions of the models. ConclusionThe XGBoost prediction model based on radiomic features derived from multiparametric MRI (T2STIR, DWI, and DCE) combined with clinicopathological data was able to predict SLN metastasis after NAT in breast cancer patients.