ABSTRACT Aims Data‐driven clustering studies of type 2 diabetes (T2D) based on limited routine clinical variables have reproducibly described major metabolic subgroups, including severe insulin‐deficient diabetes (SIDD), severe insulin‐resistant diabetes (SIRD), mild obesity‐related diabetes and mild age‐related diabetes. We examined whether deeper metabolic phenotyping can improve phenotypic resolution beyond these low‐dimensional frameworks. Methods We performed an exploratory cross‐sectional clustering analysis in 94 adults with newly diagnosed T2D and complete data, drawn from an initial 97 screened participants who underwent deep metabolic phenotyping. Principal component analysis (PCA) guided selection of 13 relatively independent variables spanning six physiological domains: glycaemic exposure, insulin sensitivity and secretion, visceral adiposity, lipid metabolism, skeletal muscle status, hepatic injury and haematological indices. K‐means clustering was applied to standardised variables. Robustness was assessed across prespecified 5‐, 9‐ and 13‐variable panels, exclusion of age, alternative muscle and adiposity representations and comparison with hierarchical and Gaussian mixture clustering. Autoimmune diabetes was not examined due to insufficient case numbers. Results Five phenotypic clusters were observed, including preserved insulin sensitivity with compensatory hyperinsulinaemia (IS‐HI). Two clusters aligned with canonical SIDD‐like and SIRD‐like phenotypes. Within the enriched model, two additional physiologically coherent patterns were resolved: a myogenic–anaemia phenotype (MA), characterised by reduced skeletal muscle percentage and altered haematological indices and a hepato‐lipotoxic phenotype (HL), marked by dyslipidaemia and elevated hepatic enzymes. In an unconstrained sensitivity rerun of the same 13‐variable panel, however, a distinct SIDD‐like cluster did not re‐emerge, underscoring that higher‐dimensional clustering depends in part on variable selection and biologically anchored interpretation. Internal robustness analyses supported preservation of the overall cluster architecture across variable panels, although alternative clustering methods yielded non‐identical partitions. Conclusions Phenotypic depth may improve the mechanistic resolution of type 2 diabetes beyond routine‐variable frameworks. However, the present single‐centre findings are hypothesis‐generating only, do not support clinical decision‐making and all treatment‐related inferences remain speculative without longitudinal outcome data and external validation.
Wang et al. (Sun,) studied this question.