Kidney Week

Abstract: TH-PO1195

Improving Kidney Failure Risk Predictions for Clinical Trials Across CKD Stages 1-4

Session Information

• Late-Breaking Research Posters
  November 06, 2025 | Location: Exhibit Hall, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: CKD (Non-Dialysis)

• 2302 CKD (Non-Dialysis): Clinical, Outcomes, and Trials

Authors

• Shpaner, Leonid, University of California Los Angeles David Geffen School of Medicine, Los Angeles, California, United States

Leonid Shpaner, MS

• Petousis, Panayiotis, University of California Los Angeles David Geffen School of Medicine, Los Angeles, California, United States

Panayiotis Petousis

• Nicholas, Susanne B., University of California Los Angeles David Geffen School of Medicine, Los Angeles, California, United States

Susanne B. Nicholas, MD, PhD, MPH

• Bui, Alex, University of California Los Angeles David Geffen School of Medicine, Los Angeles, California, United States

Alex Bui, PhD

• Duru, Obidiugwu, University of California Los Angeles David Geffen School of Medicine, Los Angeles, California, United States

Obidiugwu Duru, MD, MS

• Kornowske, Lindsey M., Providence Health and Services, Renton, Washington, United States

Lindsey M. Kornowske, MS

• Jones, Cami R., Providence Health and Services, Renton, Washington, United States

Cami R. Jones, PhD

• Daratha, Kenn B., Providence Health and Services, Renton, Washington, United States

Kenn B. Daratha, PhD

• Norris, Keith C., University of California Los Angeles David Geffen School of Medicine, Los Angeles, California, United States

Keith C. Norris, MD, PhD

• Tuttle, Katherine R., Providence Health and Services, Renton, Washington, United States

Katherine R. Tuttle, MD, FASN

Group or Team Name

• Center for Kidney Disease Research, Education, and Hope (CURE-CKD).

Background

Recruitment prescreening and reliable outcome assessment are persistent bottlenecks in kidney trials. The Kidney Failure Risk Equation (KFRE) is limited to CKD stages 3-4. However, considering individuals in CKD 1 and 2 supports early prediction and detection, enabling prescreening upstream of traditional practice. We evaluated KFRE and compared with machine learning (ML) models developed across CKD 1-4 with and without missing data.

Methods

Registry-based EHR data from UCLA and Providence Health were split into train/validation/test sets (60/20/20). We defined four cohorts: CKD 3-4 with (n=97,386) and without (n=4,190) missing data; and CKD 1-4 with (n=637,684) and without (n=21,246) missing data. In addition to using the KFRE 4-variable equation, we developed four ML models (logistic regression, random forest, XGBoost, CatBoost) to predict 2- and 5-year kidney failure. Performance metrics were AUROC and average precision.

Results

As KFRE can only be used on CKD 3-4, all models were tested on the CKD 3-4 test holdout (n=4,190). In CKD Stages 3-4 without missing data, KFRE 4-variable achieved AUROC 0.871 and average precision 0.483 at 2 years, and AUROC 0.831 and average precision 0.458 at 5 years. Extending prediction to our best model (XGBoost), developed from data of CKD Stages 1-4 (including missing data), achieved an AUROC of 0.887 and average precision of 0.533 at 2 years; and an AUROC of 0.868 and average precision of 0.562 at 5 years. For all other cohorts, models, and metrics, see Table 1.

Conclusion

XGBoost outperformed KFRE on AUROC and average precision in CKD 3-4. The best performing model used the full CKD spectrum for training and included records with missing data, which enables earlier prediction by incorporating more information. These results support a single model for prescreening and improved outcome assessment in kidney trials.

Model evaluation results on test holdout of CKD 3-4. AUROC and AP at 2 and 5 years. 95% CIs. CKD Stages: development cohort by stage with or without missing data. * with missing data.

Funding

• Other NIH Support

Digital Object Identifier (DOI)

• doi: 10.1681/ASN.2025k91143bs