"PD-PREDICT": A Machine Learning Model for Patient Survival in Peritoneal Dialysis

Accurate survival prediction in peritoneal dialysis (PD) patients is essential for personalized treatment planning and shared decision-making. We developed and validated PD-PREDICT, an XGBoost-based model to generate dynamic mortality risk estimates in incident PD patients. We conducted a retrospective cohort study using data from the UK Renal Registry (UKRR), comprising 22,711 incident PD patients treated between January 1, 2007, and September 1, 2022. The development cohort (n = 14,650; January 2007-December 2016) was split into training and internal test sets. Temporal validation employed an independent UKRR cohort (n = 8,061; January 2017-December 2021). External validation used 2,180 patients from the Norwegian Renal Registry. Model performance was assessed by Harrell's concordance index ( C index), Integrated Brier Score (IBS), decision curve analysis, and 50 iteration bootstrap for C index stability. In the development cohort, PD-PREDICT achieved a training C index of 0.83 and test C index of 0.81 (IBS: 0.09). The decision tree baseline model yielded a test C index of 0.78 (IBS: 0.13). Bootstrap analysis confirmed C index stability (0.81; 95% confidence interval [CI], 0.79-0.83). Temporal validation produced a C index of 0.80, and external validation in Norway yielded 0.77. PD-PREDICT provides robust, dynamic mortality risk predictions for PD patients, outperforming traditional methods and maintaining accuracy across temporal and geographic validations.