Abstract: SA-PO0551
Clustering Trajectories of eGFR Using Latent Process Mixed Models and Unsupervised Learning
Session Information
- Cystic Kidney Diseases: Clinical Research
November 08, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Genetic Diseases of the Kidneys
- 1201 Genetic Diseases of the Kidneys: Monogenic Kidney Diseases
Authors
- Zhang, Zhen, Otsuka America Pharmaceutical Inc, Princeton, New Jersey, United States
- Yu, Kunbo, Pharmapace, Inc., San Diego, California, United States
- Jiang, Huan, Otsuka America Pharmaceutical Inc, Princeton, New Jersey, United States
Background
Characterizing long-term kidney function trajectories is essential to identify distinct progression patterns in patients with autosomal dominant polycystic kidney disease (ADPKD). While Mayo Imaging Classification provides early stratification, trajectory-based approaches may uncover additional patient heterogeneity. This study aimed to apply latent process mixed models and unsupervised learning to identify distinct estimated Glomerular Filtration Rate (eGFR) trajectory clusters using real-world longitudinal data.
Methods
Using longitudinal eGFR data from a subgroup of patients in the OVERTURE study (NCT01430494) with baseline Mayo Imaging Classification, latent class mixed models (LCMM) with linear, quadratic, and cubic trends were fitted to identify trajectory clusters, considering duration as the true time scale. The optimal number of clusters was selected based on AIC. Additionally, a two-stage unsupervised learning approach was conducted: local regression models were fitted per patient, and K-means clustering was applied to the extracted coefficients. Clustering solutions from LCMM and unsupervised learning were compared by cross-tabulation and adjusted Rand Index.
Results
Among 2990 patients with sufficient longitudinal eGFR data, LCMM identified six distinct trajectory classes ranging from early rapid decliners, moderate progressors, to slow progressors, which aligned partially but not completely with Mayo Classes. K-means clustering on local regression model coefficients suggested similar patterns but with moderate concordance (adjusted Rand Index = 0.3) to LCMM. Patients in the slow-declining class had lower baseline height-adjusted TKV, younger age and favored vital signs at baseline. LCMM and unsupervised learning results reduced within-cluster sum of squares (WCSS) under Mayo classification by 61% and 37%, respectively. The clustering approaches provided complementary perspectives for patient stratification.
Conclusion
Trajectory-based clustering using LCMM and unsupervised learning methods offers a data-driven approach to enrich ADPKD patient characterization and may inform future risk stratification strategies beyond traditional imaging classifications.
Funding
- Commercial Support – Otsuka Pharmaceutical Development & Commercialization, Inc., Princeton, NJ