Aims/hypothesis <p>Available methods for predicting the onset and progression of diabetic kidney disease (DKD) and end-stage kidney disease (ESKD) are not yet ready for clinical application. We used a Japanese diabetes cohort study (J-DREAMS) to examine whether the Ahlqvist et al diabetes clustering is useful for stratifying DKD or ESKD outcomes independent of known risk factors in real-world settings.</p> Methods <p>Data-driven cluster analysis using <i>k</i>-means was performed based on GAD antibody levels, age at diagnosis, BMI, HbA_1c and HOMA2 estimates of beta cell function and insulin resistance in 12,093 individuals with type 1 or type 2 diabetes. The risk of developing DKD/ESKD was analysed using Kaplan–Meier analysis and the Cox proportional hazards model.</p> Results <p>Diabetes clustering classified individuals in the J-DREAMS cohort into five subtypes, the clinical characteristics of which were comparable to those of the previously reported five subtypes. Kaplan–Meier curve analysis showed that events for chronic kidney disease (CKD) stages 3b, 4 and 5 were highest in the severe insulin-resistant diabetes subtype. The Cox proportional hazards model showed that the severe insulin-resistant diabetes subtype had significant HRs after correction for multiple confounding factors. The Cox proportional hazards model showed that each subtype had a diverse combination of factors associated with CKD stage 3b and proteinuria events.</p> Conclusions/interpretation <p>Data-driven analysis provides diabetes subtyping, which can predict the probability of developing DKD/ESKD; each subtype has diverse combinations of factors predisposing to DKD development and progression. Data-driven diabetes subtyping to predict the likelihood of developing DKD/ESKD and mitigating predisposing factors may help personalise prevention strategies.</p> Graphical Abstract <p></p>