Abstract: SA-PO0345
Improving Mortality Prediction in Adults with Kidney Failure: Conformal Prediction Methodology with Data from the United State Renal Data System (USRDS)
Session Information
- Dialysis: Epidemiology and Facility Management
November 08, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Dialysis
- 801 Dialysis: Hemodialysis and Frequent Dialysis
Authors
- Liu, Mei, University of Florida Department of Health Outcomes & Biomedical Informatics, Gainesville, Florida, United States
- Xu, Qi, University of Florida Department of Health Outcomes & Biomedical Informatics, Gainesville, Florida, United States
- Shukla, Ashutosh M., VA North Florida South Georgia Veterans Health System, Gainesville, Florida, United States
Background
Mortality risk prediction in kidney failure is critical for guiding clinical decision-making and resource allocation. Traditional prediction models provide point estimates without quantifying uncertainty, limiting their clinical interpretability.
Methods
In this retrospective study, we analyzed USRDS data from 838,307 patients with kidney failure aged 18–90 who initiated chronic dialysis between 2011 and 2018. We constructed a comprehensive dataset by integrating demographic, clinical, and claims-based features. We developed CatBoost mortality prediction models for multiple horizons. To quantify prediction uncertainty, we implemented a conformal prediction (CP) framework, including a class-conditional variant to account for class imbalance. Model performance was assessed via AUROC and AUPRC, while CP validity and efficiency were evaluated by coverage and prediction set sizes. Feature importance was assessed using SHAP values, and feature interactions were explored to understand how key predictors contributed to different CP prediction sets.
Results
Validation AUROC ranged from 0.84 at 90 days to 0.80 at 2 years, while AUPRC rose from 0.27 to 0.59 over the same horizons. CP maintained overall accuracy near the nominal 95% confidence level, with class-conditional CP showing improved sensitivity for the deceased compared to the standard CP. Although CP produced a higher proportion of uncertain predictions—rising from 51% at 90 days to 62% at 2 years—both CP approaches demonstrated enhanced precision and negative predictive value than traditional binary classification, with average improvements of 0.18 and 0.07, respectively, at 2 years. Subgroup analyses revealed variations in CP performance by demographics and dialysis/access type. Feature analysis identified albumin, age at dialysis initiation, non-Hispanic White race, and a claim-based frailty index as consistently influential predictors, along with a novel predictor—distance to dialysis facility.
Conclusion
Our study demonstrates that integrating conformal prediction with advanced machine learning techniques provides a robust and clinically interpretable framework for mortality risk stratification in patients with kidney failure. By quantifying predictive uncertainty, this approach offers nuanced risk assessments to inform individualized treatment strategies and resource allocation in dialysis care.
Funding
- Veterans Affairs Support