A time‑dependent risk prediction model for distant metastasis in early‑stage breast cancer based on explainable ensemble learning: a retrospective cohort study

Background: Distant metastasis is a leading cause of death in early-stage breast cancer, but current tools are imprecise or costly. This study aimed to develop an SHapley Additive exPlanations (SHAP)‑enhanced ensemble learning model using routine clinicopathological features to predict metastasis risk. Methods: This retrospective cohort study enrolled 351 patients with stage I-III breast cancer diagnosed between 2016 and 2024. The primary endpoint was distant metastasis‑free survival (DMFS). Comprehensive clinicopathological variables were refined using recursive feature elimination with cross‑validation. A stacking ensemble framework was constructed incorporating CoxNet, random survival forest (RSF), gradient boosting survival trees (GBST), and DeepSurv as base learners, with LightGBM as the meta‑learner. Model performance was assessed using the concordance index (C‑index), time‑dependent area under the curve (AUC), integrated Brier score, and calibration curves. SHAP was applied for global and local interpretability. Results: During a median follow‑up of 42 months, 89 distant metastasis events occurred. Eight core predictors were identified. The Stack‑LightGBM model achieved a global C‑index of 0.82 [95% confidence interval (CI): 0.77-0.87] and time‑dependent AUCs of 0.85 (95% CI: 0.80-0.90), 0.82 (95% CI: 0.77-0.87), and 0.79 (95% CI: 0.74-0.84) for 1‑, 3‑, and 5‑year DMFS, respectively, outperforming all single models. SHAP analysis revealed N stage and Ki‑67 as dominant risk drivers, with non‑linear effects and clinically meaningful feature interactions. Kaplan-Meier (KM) analysis yielded 5‑year distant metastasis‑free survival rates of 95.2%, 78.5%, and 51.3% for low‑, intermediate‑, and high‑risk groups, respectively (log‑rank P