Personalized survival prediction in elderly dialysis patients: integrating machine learning with traditional survival...

Elderly patients undergoing kidney replacement therapy (KRT) face high mortality rates. Traditional statistical models describe overall survival patterns but may fail to capture non-linear relationships among clinical factors. This study integrates classical survival analysis and machine learning (ML) to enhance individualized survival prediction in elderly patients treated with hemodialysis (HD) or peritoneal dialysis (PD).

We analyzed a retrospective cohort of 2,501 patients aged ≥ 75 years initiating KRT in Colombia (2009-2013), with follow-up through 2015. Survival was assessed using Kaplan-Meier and Cox proportional hazards (Cox PH) models. Five ML survival methods including Penalized Cox, Random Survival Forest (RSF), gradient boosting accelerated failure time (XGB-AFT), DeepSurv, and DeepHit were evaluated on an independent test set using time-dependent concordance (C-index) and the integrated Brier score (IBS). Uncertainty was quantified using non-parametric bootstrap 95% confidence intervals.

HD was associated with improved survival compared with PD (HR = 1.36, 95% CI: 1.20-1.53; p < 0.005). Predictors of reduced survival included advanced age, female sex, diabetes, hypoalbuminemia, and anemia. In head-to-head model comparison, DeepSurv achieved the best overall performance (C-index = 0.692, 95% CI: 0.649-0.734; IBS = 0.145, 95% CI: 0.129-0.164), while Cox-based models showed comparable performance. RSF and XGB-AFT demonstrated intermediate performance, and DeepHit showed lower overall accuracy in this cohort. Feature importance analysis highlighted vascular access (arteriovenous fistula), albumin, and age as the strongest predictors.

Hemodialysis is associated with a survival advantage in elderly dialysis patients. Integrating ML with classical survival analysis improves individualized prognostic assessment and highlights prognostic factors such as vascular access and nutritional/inflammatory status. These findings may inform precision medicine approaches to dialysis planning and patient-centered management.