Outcome-aware and interpretable subtyping of chronic kidney disease: an analysis of the FROM-J study

Chronic kidney disease (CKD) affects approximately 10% of the global population and exhibits substantial heterogeneity in disease progression and clinical outcomes. Despite ongoing efforts to develop new therapeutic strategies, the number of patients progressing to end-stage kidney disease (ESKD) and the incidence of cardiovascular disease (CVD) continue to rise. Although severity classification systems for CKD are well established and refined, they remain insufficient to capture prognostically relevant patient subtypes. In this study, we developed an outcome-aware and interpretable clustering framework for CKD subtyping using data from the FROM-J cohort with prognostic follow-up. A supervised XGBoost model was first trained to predict a ≥ 30% decline in estimated glomerular filtration rate (eGFR), a surrogate marker of CKD progression. SHAP (SHapley Additive exPlanations) values derived from this model were then used to quantify outcome-relevant feature contributions. Based on these feature attributions, a similarity graph was constructed, and spectral clustering was performed to identify patient subtypes driven by prognostic relevance. The proposed framework identified four CKD subtypes with distinct baseline clinical characteristics and significantly different risks of renal replacement therapy (RRT) and cardiovascular disease (CVD) events. Serum albumin, blood urea nitrogen (BUN), and smoking status consistently emerged as key features defining subtype structure and prognosis. Robust risk stratification was preserved even when clustering was restricted to these three routinely measured variables. Overall, our findings demonstrate that integrating outcome-driven feature attribution into clustering enables interpretable and clinically relevant CKD subtyping, providing a practical approach for characterizing disease heterogeneity and supporting risk stratification and personalized management.