Prediction of long-term mortality in maintenance hemodialysis patients using machine learning survival models: a prospective multicenter study with SHAP-based interpretation

Background Accurate and interpretable prediction of long-term mortality in maintenance hemodialysis (MHD) patients remains challenging. This study aimed to develop and externally validate machine learning survival models integrating clinical, nutritional, inflammatory, and anthropometric predictors. Methods This prospective multicenter cohort included 1,129 MHD patients from eight dialysis centers in China. Patients were allocated to training ( n = 645), internal validation ( n = 274), and independent external test ( n = 210) cohorts. Predictors were selected using LASSO–Cox regression and Survival Boruta. Five survival machine learning models were developed and evaluated for discrimination, prediction error, calibration, and clinical utility. SHapley Additive exPlanations (SHAP) were used to interpret model predictions. Results During a median follow-up of 4.53 years, 603 patients died. 14 predictors were selected for model development. In the external test cohort, gradient boosting survival analysis (GBS) achieved a C-index of 0.748 (95% CI, 0.705–0.790), a 5-year time-dependent AUC of 0.799 (95% CI, 0.726–0.861) and the lowest integrated Brier score (0.141) among the five models. GBS also demonstrated reasonable calibration, although its discrimination was not significantly superior to that of random survival forest. SHAP analysis identified age, prealbumin and diabetes mellitus as the leading contributors to model predictions. Conclusion An interpretable GBS model integrating routinely available clinical, nutritional, inflammatory, and anthropometric predictors demonstrated promising performance for long-term mortality risk stratification in MHD patients. Further external validation and prospective clinical evaluation are warranted.