Kidney Week

Abstract: TH-PO0646

Real-World Prediction of CKD Progression Using an Electronic Health Record (EHR)-Based Survival Boosting Model: Benchmarking Against a Klinrisk Proxy

Session Information

• CKD: Clinical - Guidelines, Biomarkers, Progression, Procedures, Outcomes, and More
  October 22, 2026 | Location: Exhibit Hall A, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: CKD (Non-Dialysis)

• 2202 CKD (Non-Dialysis): Clinical, Outcomes, and Trials

Authors

• Balaji, Shyam Sunder, nference, Bengaluru, Karnataka, India

Shyam Sunder Balaji

• Varma, Gowtham, nference, Bengaluru, Karnataka, India

Gowtham Varma, MBBS

• Anand, Deepak, nference, Bengaluru, Karnataka, India

Deepak Anand, PhD

• Thiagarajan, Viswanathan, nference, Bengaluru, Karnataka, India

Viswanathan Thiagarajan

• Soundararajan, Venkataramanan, nference, Bengaluru, Karnataka, India

Venkataramanan Soundararajan, PhD

Background

Risk prediction for chronic kidney disease (CKD) progression enables risk-based monitoring and can improve feasibility of clinical trials through event-rate enrichment. We evaluated whether a comprehensive featured EHR model improves prediction of CKD progression at clinically relevant horizons compared with established baselines Klinrisk-style proxy, and assessed performance across baseline CKD stage and comorbidity subgroups.

Methods

Using the nference nSights federated EHR database, models were developed in a single US health system (representing catchment of 8M patient lives across 3 decades), with evaluation on a held-out validation cohort (N=26,805) requiring demographics, baseline serum creatinine, and urine albumin-to-creatinine ratio (UACR). Indexing and inclusion criteria mirrored published Klinrisk methodology. The primary endpoint was CKD progression defined as ≥40% eGFR decline, eGFR <15 mL/min/1.73m², or end-stage kidney disease, whichever occurred first. The model incorporated demographics, comorbidity burden and disease durations, medication classes, vitals, laboratory panels, and proteinuria testing, and compared against a laboratory-feature comparator using age, sex, eGFR, UACR, blood urea nitrogen, glucose, hemoglobin, and hematocrit, approximating the published Klinrisk model.

Results

In the held-out validation cohort, the EHR-enriched model outperformed the KlinRisk proxy at 2 years, with C-index 0.784 versus 0.724 and AUROC 0.888 versus 0.795 across 463 progression events. At 5 years, the EHR-enriched model achieved AUROC 0.861 across 1,246 events. At an F1-optimized threshold, PPV was 0.271 at 2 years and 0.245 at 5 years. Discrimination remained robust in earlier-stage CKD, with AUROC 0.893 at 2 years and 0.859 at 5 years in G1-G3a disease, and persisted in advanced CKD, with AUROC 0.733 at 2 years and 0.772 at 5 years in G3b-G5 disease. In early CKD stages, enrichment exceeded the proxy in G1, G2, and G3a, (baseline event rate 1.5-2.7%) with relative gains of 8.3%, 48.5%, and 16.4%, respectively.

Conclusion

An extended feature EHR machine learning model improved real-world prediction of CKD progression over clinical baselines and a Klinrisk-style proxy, while providing substantial event-rate enrichment for potential trial feasibility applications. External validation across additional health systems within the federated network is planned.