Clinical Trajectories for Prognostic Subgrouping in IgA Nephropathy: A Deep Learning Approach

IgA nephropathy (IgAN) shows wide heterogeneity from spontaneous remission to progressive kidney failure, challenging risk assessment based on single time-point markers. We posited that integrating the early dynamics of hematuria, proteinuria, and eGFR could reveal reproducible subgroups that capture treatment response and disease activity beyond static predictors. We therefore applied unsupervised deep learning to first-year clinical trajectories in a large, prospective, nationwide cohort and evaluated the association between trajectory-defined subgroups and long-term outcomes. 

Data were drawn from the Japan IgA Nephropathy Prospective Cohort Study (J-IGACS; 33 university and 11 regional core hospitals), including biopsy-proven IgAN enrolled 2005–2015 with follow-up through 2021. Inclusion required baseline eGFR ≥30 mL/min/1.73 m², ≥12 months of follow-up, and ≤1 missing value among urinary RBC, proteinuria, and eGFR at 0, 6, and 12 months; 873/1,130 patients met criteria. The primary endpoint was ≥30% eGFR decline from baseline. Median follow-up was 78 months (IQR 42–96).
Urinary RBC was ordinally encoded (five grades <5 to >50/HPF); proteinuria and eGFR were standardized with fold-wise parameters; remaining baseline missingness was handled by chained equations. An LSTM autoencoder embedded the three-marker 0–12-month trajectories; we clustered latent vectors with k-means, selecting K by silhouette and Calinski–Harabasz indices. Cox models tested associations with the primary outcome (Model 1 univariable; Model 2 + age and MAP; Model 3 + Oxford scores). Incremental value was assessed by C-statistic and IDI comparing a baseline predictor model vs. baseline+cluster model. 

K=3 was optimal, yielding Cluster 1 (n=284), Cluster 2 (n=215), and Cluster 3 (n=374). Cluster 1 exhibited marked improvement: severe hematuria (>50/HPF) fell from 46.8% to 0.4%, proteinuria decreased from 1.15 to 0.23 g/day, and eGFR remained ~90 mL/min/1.73 m² (Figure 1). Cluster 2 showed persistent severe hematuria (26.5% at 12 months) with modest proteinuria reduction (0.90→0.46 g/day) and mild eGFR decline (76.6→74.6). Cluster 3 began with the lowest eGFR and had incomplete proteinuria improvement (1.05→0.53 g/day) despite milder hematuria. These patterns aligned with baseline differences: Cluster 1 younger with more crescents and more immunosuppression/tonsillectomy; Cluster 3 older with more T-lesions and greater RAS inhibitor use.
Event-free survival differed across clusters (log-rank p<0.001). Versus Cluster 1, Clusters 2/3 were independently associated with the primary outcome (Model 3 HR 2.12, 95%CI 1.35–3.34). Adding cluster membership to static baseline predictors significantly improved C-statistic and IDI. Findings were robust after adjusting for initial treatments (adjusted HR 1.90, 95%CI 1.19–3.02), for more severe endpoints (40%/50% eGFR decline, kidney failure), for alternative clustering (Gaussian mixture), and in complete-case analyses.

An unsupervised deep-learning approach to first-year trajectories identified three clinically interpretable IgAN subgroups with distinct long-term risks, capturing information not contained in baseline clinicopathological features alone. Trajectory-based classification improved model performance and may support dynamic risk stratification and personalized follow-up strategies.