Abstract: FR-PO0004
Long-Term GFR Trajectory Prediction in Patients with CKD Using Neural Network-Based Latent-Space Neighbor Simulation
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
- Sakaguchi, Yusuke, Department of Nephrology, The University of Osaka Graduate School of Medicine, Suita, Osaka, Japan
- Kadono, Hikaru, Department of Nephrology, The University of Osaka Graduate School of Medicine, Suita, Osaka, Japan
- Yoshino, Mitsuaki, MBX Co., Ltd., Ise, Japan
- Isaka, Yoshitaka, Department of Nephrology, The University of Osaka Graduate School of Medicine, Suita, Osaka, Japan
- Kimura, Tomonori, Department of Nephrology, The University of Osaka Graduate School of Medicine, Suita, Osaka, Japan
Background
Accurate prediction of long-term GFR trajectories is essential for optimizing care in CKD. Although linear regression models have been widely used to estimate GFR slopes, GFR trajectories are often non-linear, causing discrepancies between predicted and observed values. Here, we developed an AI-based model to improve the GFR trajectory prediction. We then compared its performance with that of conventional linear models.
Methods
A total of 22,489 patients with CKD who had longitudinal eGFR data for more than 12 months were identified from the Osaka Consortium for Kidney Disease Research (OCKR) database. We categorized 30 blood and urine test items into kidney-related and other domains, training separate autoencoders to maintain domain independence. This yielded low-dimensional integrated feature vectors capturing multivariate and temporal relationships. Patients in close proximity within this space are supposed to share similar inter-domain and temporal patterns. Based on this similarity, we used Monte Carlo simulations to forecast long-term trajectories (50 months) from short-term sequences (10 to 12 months).
Results
By applying encoder-based transfer learning, we successfully generated low-dimensional feature vectors from short-term data (10–12 months). These vectors were nearly identical to those obtained from long-term data (50 months). Based on the first 12 months of test data for 65 randomly-selected cases, the simulation predicted eGFR values measured at 45–54 months with a mean absolute error (MAE) of 10.8 mL/min/1.73m2 (SD: 8.5). This significantly outperformed simple linear regression (MAE: 36.0, SD: 35.8), with consistent accuracy observed using data from 10 months onward.
Conclusion
Predicting long-term GFR trajectories by Monte-Carlo simulating patterns from integrated latent features exhibited smaller errors relative to the actual measured values than extrapolations by simple linear regression lines. This method offers a practical tool for the long-term CKD management.