Kidney Week

Abstract: FR-PO0005

Predicting Two-Year Risk of Progression to ESKD in Intensive Care Unit Survivors of AKI: A Competing Risk-Random Survival Forest Approach

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

• Ekrikpo, Udeme E., Division of Nephrology, Department of Medicine, University of Alabama at Birmingham, Birmingham, Alabama, United States

Udeme E. Ekrikpo, MBBS, MSc, PhD

• Acharya, Dilaram, Division of Nephrology, Department of Medicine, University of Alabama at Birmingham, Birmingham, Alabama, United States

Dilaram Acharya, PhD

• Ghazi, Lama, Department of Epidemiology, School of Public Health, University of Alabama at Birmingham, Birmingham, Alabama, United States

Lama Ghazi, MD, PhD

• Babroudi, Seda, Division of Nephrology, Tufts Medical Center, Boston, Massachusetts, United States

Seda Babroudi, MD, MS, FASN

• Chen, Jin, Division of Nephrology, Department of Medicine, University of Alabama at Birmingham, Birmingham, Alabama, United States

Jin Chen, PhD

• Takeuchi, Tomonori, Department of Internal Medicine, University of New Mexico, Albuquerque, New Mexico, United States

Tomonori Takeuchi, MD

• Neyra, Javier A., Division of Nephrology, Department of Medicine, University of Alabama at Birmingham, Birmingham, Alabama, United States

Javier A. Neyra, MD, MS, FASN

Background

Survivors of acute kidney injury (AKI), particularly after intensive care unit (ICU) admission, face elevated long-term risk of progression to end-stage kidney disease (ESKD). High competing mortality and complex nonlinear clinical relationships complicate accurate risk prediction. We developed and internally validated a competing-risk random survival forest (CR-RSF) model to predict two-year ESKD among ICU survivors with AKI.

Methods

We conducted a retrospective cohort study of 34,209 adult ICU survivors with AKI (2012–2022) within a large academic health system linked to the United States Renal Data System. The primary outcome was incident ESKD within 2 years; death was treated as a competing event. An 85-variable CR-RSF model (1,500 trees) incorporating demographic, comorbidity, physiological, laboratory, and medication data available at discharge was trained. Variable reduction using permutation importance, minimal depth, and consensus rank-sum scoring yielded a parsimonious 15-variable model. Performance was assessed using out-of-bag predictions with time-dependent area under the curve (AUC), Brier score, integrated prediction accuracy (IPA), calibration intercept and slope, and decision curve analysis.

Results

During follow-up, 362 patients (1.1%) developed ESKD and 2,951 (8.6%) died. The 15-variable CR-RSF demonstrated strong discrimination (AUC 0.89 [95% CI 0.87–0.91] at 1 year; 0.82 [0.79–0.84] at 2 years) and excellent calibration (slope 0.905 at 1 year; 0.916 at 2 years; intercepts near zero). The Brier scores were low (0.0059 at 1 year; 0.0089 at 2 years), and the IPA was 14.8% at 2 years. Decision curve analysis showed consistent net benefit across 1–20% thresholds; at a 2% threshold, the model identified 3 and 5.5 additional true-positive patients per 1,000 at 1 and 2 years, respectively.

Conclusion

A machine learning–based competing-risk approach enables accurate, clinically meaningful prediction of long-term ESKD after ICU-associated AKI and supports precision post-discharge risk stratification. External validation is warranted.

Funding

• NIDDK Support