Kidney Week

Abstract: FR-PO0013

Competing Risks of Dialysis and Death at eGFR ≤20: Evidence for Risk-Based Kidney Transplant Referral

Session Information

• Artificial Intelligence: Clinical Prediction, Decision Support, and Digital Kidney Care
  October 23, 2026 | Location: Exhibit Hall A, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Artificial Intelligence, Digital Health, and Data Science

• 300 Artificial Intelligence, Digital Health, and Data Science

Authors

• Paulus, Amber B., Virginia Commonwealth University, Richmond, Virginia, United States

Amber B. Paulus, PhD

• Silvey, Scott, Virginia Commonwealth University, Richmond, Virginia, United States

Scott Silvey, PhD

• Patel, Nilang G., Richmond VA Medical Center, Richmond, Virginia, United States

Nilang G. Patel, MD, MS, FASN

• Gupta, Gaurav, Virginia Commonwealth University, Richmond, Virginia, United States

Gaurav Gupta, MD

Group or Team Name

• Kidney disease and Transplant Epidemiology (KiTE) Research Consortium

Background

eGFR ≤20 mL/min/1.73 m² is a policy-relevant threshold for kidney transplant waiting-time accrual and a common referral trigger. Yet it assumes that patients crossing this threshold are on the same trajectory. We tested whether EHR data could separate dialysis risk from competing mortality risk at first eGFR ≤20.

Methods

We identified adult VCU Health patients from 2022-2025 who newly reached eGFR ≤20 and had no prior kidney transplant. Patients with prior dialysis or missing baseline UACR/UPCR were excluded. Dialysis initiation and mortality were modeled as competing outcomes with censoring on January 1, 2026. Fine-Gray regression and random survival forest models were compared across 4-variable, 6-variable, and all-predictor sets. Performance was assessed using 5-fold cross-validated time-dependent AUC and Brier score.

Results

Of 12,512 screened patients, 1,127 met inclusion criteria. Median age was 66 years, 55.9% were Black, median eGFR was 18, and median UACR was 378 mg/g. During median follow-up of 693 days, 240 patients initiated dialysis and 314 died; median time to death was shorter than time to dialysis, 82 vs 225 days. Dialysis and death separated into distinct phenotypes. Patients initiating dialysis had lower eGFR and higher UACR, while patients who died had similar eGFR and UACR but greater cardiac/pulmonary disease, active malignancy, active infection, active drug abuse, and uncontrolled psychiatric disease. Active malignancy and infection were defined by diagnosis codes within 3 days of index eGFR and were prognostic predictors, not transplant exclusion criteria. For dialysis, added comorbidity variables did not improve discrimination; the strongest model used age, sex, eGFR, and UACR, with AUCs of 0.73, 0.75, and 0.71. For death, all-predictor models performed better than 4- or 6-variable models, with AUCs up to 0.77, 0.76, and 0.77.

Conclusion

At eGFR ≤20, more patients died than initiated dialysis, and mortality occurred earlier. Kidney failure risk was largely captured by kidney measures, whereas mortality risk required comorbidity context. eGFR ≤20 alone is an incomplete referral trigger. Risk-based transplant referral should pair kidney failure prediction with competing mortality estimation to prioritize timely evaluation while prompting individualized care planning. Findings are prognostic, not causal.

Acknowledgment

This work was supported by VCU Department of Internal Medicine Pilot Funding and the Gupta-Levy Foundation through the project Optimizing Kidney Transplant Referrals Through Predictive Modeling. The authors also acknowledge the Kidney and Transplant Epidemiology (KITE) Research Consortium for its scientific and analytic support.

Funding

• Private Foundation Support