To explore the value of interpretable machine learning (ML) models based on magnetic resonance imaging (MRI) parameters in predicting sentinel lymph node metastasis (SLNM) in elderly patients with breast cancer.

This single-center retrospective study enrolled 93 elderly (≥60 years) breast cancer patients diagnosed between January 2019 and September 2025. Clinical data and preoperative dynamic contrast-enhanced MRI (DCE-MRI) and diffusion-weighted imaging (DWI) images were collected. LASSO regression was used for feature selection. Based on the selected five features, four predictive models were constructed: multivariate logistic regression, extreme gradient boosting (XGBoost), support vector machine (SVM), and random forest (RF). The models' discrimination ability was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. Calibration and clinical net benefit were assessed via calibration curves and decision curve analysis. SHapley Additive exPlanations (SHAP) was employed for interpretability analysis of the best-performing model. T test was used to compared the measurement data, Chi-square test was used to compared the count data.

Forty-five cases were positive for sentinel lymph node biopsy (metastasis group), and 48 cases were negative (non-metastasis group). Compared with the non-metastatic group, the metastatic group exhibited higher proportion of Ki-67 high expression, larger tumor maximum diameter, higher long-to-short diameter ratio, higher proportions of annular enhancement and adjacent vascular proliferation, and higher early enhancement rate, along with a lower relative apparent diffusion coefficient (rADC), all with statistically significant differences (χ²=5.604, 3.930, 3.891; t=2.083, 4.313, 5.055, 2.408; P < 0.05 or P < 0.01). Five features were selected by LASSO for model construction: early enhancement rate, rADC, annular enhancement, long-to-short diameter ratio, and Ki-67 index. The XGBoost model demonstrated the best predictive performance (AUC=0.915, 95%CI: 0.870-0.950; accuracy=85.7%, sensitivity=83.3%, specificity=86.7%), with its calibration curve closest to the ideal diagonal and the highest clinical net benefit in the probability threshold range of 0.2-1.0. SHAP analysis revealed that early enhancement rate contributed the most to the model predictions, followed by rADC, annular enhancement, Ki-67 index, and long-to-short diameter ratio.

The XGBoost model integrating preoperative multiparametric MRI and clinical variables can effectively predict sentinel lymph node metastasis status in elderly breast cancer patients. SHAP interpretability analysis elucidates the contribution of key imaging and pathological features, enhancing model transparency and clinical credibility.