Deep Learning-Identified Clinical Trajectory Patterns and Associations with Kidney Outcomes in IgA Nephropathy

Background

The heterogeneous course of IgA nephropathy limits risk stratification based on static markers. We sought to identify clinical trajectory subgroups using unsupervised deep learning and validate their association with long-term renal outcomes.

Methods

We analyzed 873 biopsy-proven cases from the nationwide Japan IgA Nephropathy Prospective Cohort Study (J-IGACS). A long short-term memory autoencoder was used to generate low-dimensional representations of hematuria, proteinuria, and estimated glomerular filtration rate (eGFR) over the first 12 months after renal biopsy. We applied k-means clustering to these representations. The primary outcome was a 30% decline in eGFR from baseline.

Results

Three trajectory clusters were identified. Cluster 1 (n=284) showed rapid resolution of hematuria and proteinuria with stable eGFR and favorable prognosis. Cluster 2 (n=215) exhibited persistent severe hematuria, modest proteinuria reduction, and mild eGFR decline. Cluster 3 (n=374) presented with the lowest baseline eGFR and showed further decline within the first 12 months after biopsy, with incomplete proteinuria resolution despite milder hematuria. Clusters 2 and 3 had worse outcomes than Cluster 1. In Cox models adjusted for age, mean arterial pressure, and Oxford classification, cluster membership was independently associated with the primary outcome (hazard ratio 2.12; 95% CI 1.35-3.34 for Clusters 2 and 3 versus 1).

Conclusions

An unsupervised deep learning approach applied to trajectories of hematuria, proteinuria, and eGFR within the first year after renal biopsy identified three patient subgroups with distinct long-term renal risks. Trajectory-based classification may complement established baseline predictors and support more dynamic risk stratification in IgA nephropathy.

Key Points

• Deep learning on clinical trajectories revealed the heterogeneity of IgA nephropathy, identifying three distinct patient subgroups.

• These subgroups, reflecting a spectrum of progression patterns and treatment responses, had distinct long-term renal outcomes.

• This approach may provide a dynamic framework to understand clinical course, moving beyond static, single-point risk assessment.