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

Area of Science:

• Nephrology
• Data Science in Healthcare
• Predictive Modeling

Background:

• Accurate survival prediction is crucial for personalized treatment in peritoneal dialysis (PD).
• Existing methods may lack dynamic risk assessment capabilities for PD patients.

Purpose of the Study:

• To develop and validate PD-PREDICT, an XGBoost-based model for dynamic mortality risk prediction in incident PD patients.
• To assess the model's performance against traditional methods and across different validation cohorts.

Main Methods:

• Retrospective cohort study using UK Renal Registry (UKRR) data (22,711 incident PD patients).
• Development cohort (n=14,650) for training/internal testing; temporal validation (n=8,061) and external validation (n=2,180, Norwegian Renal Registry).
• Performance evaluated using Harrell's C index and Integrated Brier Score (IBS).

Main Results:

• PD-PREDICT achieved high performance: training C index 0.83, test C index 0.81 (IBS: 0.09).
• Outperformed decision tree baseline (test C index 0.78, IBS: 0.13).
• Maintained robust accuracy in temporal (C index 0.80) and external (C index 0.77) validations.

Conclusions:

• PD-PREDICT offers reliable, dynamic mortality risk predictions for PD patients.
• The model demonstrates superior performance compared to traditional approaches.
• Validated accuracy across diverse temporal and geographical settings supports clinical utility.