Kidney Week

Abstract: FR-PO0001

Real-Time Cumulative Sum-GFR Trajectory Monitoring Affords a 15-Month Predictive Lead Time for ESKD

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

• Capriles, Guido M., TriHealth Inc, Cincinnati, Ohio, United States

Guido M. Capriles, MD

• Dumbauld, Steven L., TriHealth Inc, Cincinnati, Ohio, United States

Steven L. Dumbauld, MD

• Dumbauld, Diane, TriHealth Inc, Cincinnati, Ohio, United States

Diane Dumbauld, APRN

• Mena, Angel, TriHealth Inc, Cincinnati, Ohio, United States

Angel Mena, MD

• Forgo, Vincent Zugudaah, TriHealth Inc, Cincinnati, Ohio, United States

Vincent Zugudaah Forgo, MS

• Abdeltawwab, Mohannad, TriHealth Inc, Cincinnati, Ohio, United States

Mohannad Abdeltawwab, MD, MBChB

Background

CKD progression is often insidious. Outpatient surveillance relies on static estimated glomerular filtration rate (eGFR) thresholds, delaying detection of rapid decline. We retrospectively validated the cumulative sum (CUSUM) algorithm—a statistical process control tool evaluating continuous eGFR trajectory velocity—to predict pathological renal decline within a community electronic health record (EHR) cohort.

Methods

A 10-year retrospective EHR analysis (TriHealth, OH) required ≥9 outpatient eGFRs (2021 CKD-EPI) to establish baseline variance. To isolate chronic decline, an acute kidney injury phenotype (all eGFRs <90 mL/min/1.73 m² clustered exclusively within 3 months) served as an exclusion criterion. Age-adjusted baselines were derived from a normative cohort (n=7,683; absent CKD, eGFR 60-90). The incident end-stage kidney disease (ESKD) cohort (n=481) excluded patients with initial eGFR <60 to eliminate lead-time bias. The algorithm was optimized via 10-fold cross-validation, locking the signal threshold (T=-2) for persistent deviation and tuning the parameter (w) to minimize false-positives.

Results

The optimal model (w=1.0, T=-2) prioritized specificity to mitigate alert fatigue, yielding cross-validated sensitivity of 96.4% and specificity of 94.1%. Longitudinally, CUSUM-GFR predicted ESKD prior to diagnosis in 125 progressors. Predictive alerts triggered while 12.0% were in CKD Stage 2 (eGFR 60-74) and 62.4% in Stage 3a (eGFR 45-59). Overall, 74.4% of progressors were identified at eGFR >45. The model afforded a 15.6-month mean predictive lead time prior to ESKD, restricting the 10-year longitudinal patient-level false-positive rate in the normative cohort to 16.8%.

Conclusion

Continuous CUSUM-GFR trajectory monitoring predicts ESKD progression, affording a 15.6-month lead time with a clinically viable false-positive rate. Evaluating longitudinal velocity rather than isolated eGFR thresholds facilitates earlier detection of pathological decline. This expands the therapeutic window for initiating guideline-directed therapies and arteriovenous access planning, potentially reducing emergent dialysis.