Novel phenotypic clusters of adult-onset diabetic kidney disease based on static and dynamic clustering algorithms - a data-driven multi-center study.

OBJECTIVE

This study aimed to identify clinically interpretable DKD sub-phenotypes for personalized prognosis and intervention.

METHODS

We enrolled participants with type 2 diabetes mellitus (T2DM) from three hospitals in China. Clinically interpretable sub-phenotypes were identified using two complementary clustering frameworks: a static model (k-means clustering) capturing the metabolic-renal profile at DKD diagnosis, and a dynamic model (group-based multi-trajectory model, GBMTM) reconstructing longitudinal trajectories preceding diagnosis. Associations between DKD clusters and post-diagnosis DKD progression were estimated using cox regression.

RESULTS

A total of 1520 participants were included for the analysis of static clustering, 32.53% were females with a mean age of 54.4 ± 12.1 years. The static clustering method classified DKD patients into four clusters: BALANCE, YOUTH, TYPICAL and NATURAL. Compared with the participants in YOUTH, the hazard ratios (HRs) of progressing to higher stage of CKD (G3b-G5) in BALANCE, TYPICAL and NATURAL are 1.68 (95%CI:0.81-3.49), 7.85 (95%CI:4.25-14.51), 6.58 (95%CI:3.54-12.24), respectively. GBMTM identified three longitudinal trajectories. The HRs (95%CI) of progressing to higher stage of CKD (G3b-G5) was 2.48 (95%CI:1.50-4.09) in Trajectory 1 compared with Trajectory 2. In addition, the clustering results obtained from the validation cohort were consistent with the derivation cohort.

CONCLUSIONS

By integrating static and dynamic clustering approaches, this study identifies novel DKD sub-phenotypes, and evaluated the risk of DKD progression in each cluster.