Kidney Week

Abstract: FR-PO0252

High-Dimensional Generative Latent Representations Capture the Pathophysiological Structure of CKD and Improve Survival Prediction

Session Information

• CKD: Omics, Systemic Stressors, and Targeted Pharmacotherapy
  October 23, 2026 | Location: Exhibit Hall A, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: CKD (Non-Dialysis)

• 2201 CKD (Non-Dialysis): Epidemiology, Risk Factors, and Prevention

Authors

• Kanda, Eiichiro, Kawasaki Ika Daigaku, Kurashiki, Okayama Prefecture, Japan

Eiichiro Kanda, MD, PhD, MPH

• Epureanu, Bogdan I., University of Michigan, Ann Arbor, Michigan, United States

Bogdan I. Epureanu, PhD

• Pennathur, Subramaniam, University of Michigan, Ann Arbor, Michigan, United States

Subramaniam Pennathur, MD

• Adachi, Taiji, Kyoto Daigaku, Kyoto, Kyoto Prefecture, Japan

Taiji Adachi, PhD

Background

Chronic kidney disease (CKD) progression reflects interacting metabolic, inflammatory, and structural pathways that are not fully represented by conventional low dimensional clinical parameters. We hypothesized that high dimensional generative latent spaces more effectively capture this complex pathophysiology and enhance survival prediction.

Methods

We analyzed 33 clinical variables from 3,129 patients over three years. Ten generative architectures were combined with Cox regression to evaluate survival prediction accuracy using Harrell’s C index. Latent representations ranging from 2 to 1,024 dimensions were extracted to assess how latent space dimensionality influences prognostic performance.

Results

Patients had a mean age of 62.0 years and an average eGFR of 50.6 mL/min/1.73m². The baseline Cox model achieved a C index of 0.85. Generative models paired with Cox regression—particularly conditional variational autoencoders (CVAEs)—demonstrated dimension dependent improvements in predictive accuracy (Figure 1). Although some architectures showed nonmonotonic variability, likely reflecting model specific sensitivity, CVAE performance improved consistently with increasing dimensionality, reaching a C index of 0.94 at 1,024 dimensions.

Conclusion

CKD progression appears inherently high dimensional, and generative latent spaces preserve clinically meaningful structure beyond traditional approaches. CVAE derived representations provide robust prognostic signals and may enable future applications such as individualized risk trajectories, high resolution disease mapping, and counterfactual simulations.