Clustering-based subgroups of type 2 diabetes mellitus and their associations with diabetes remission after bariatric surgery

To the Editor: Obesity is a pivotal risk factor for type 2 diabetes mellitus (T2DM), and its frequent co-occurrence has contributed to increased morbidity and mortality worldwide. Bariatric surgery represents an effective intervention for obese patients with diabetes. However, the extent of postoperative diabetes remission varies considerably across individuals.[1] Current predictive scoring models are derived from variables identified through univariate analyses, which may insufficiently capture the heterogeneity of T2DM and are limited in their ability to discriminate among distinct diabetes subgroups. In contrast, clustering-based approaches offer a more integrative, phenotype-driven classification system that may better inform surgical outcomes. A well-established clustering framework classifies T2DM into four subtypes: severe insulin-resistant diabetes (SIRD), mild age-related diabetes (MARD), severe insulin-deficient diabetes (SIDD), and mild obesity-related diabetes (MOD). These subtypes differ in clinical presentation, complication risks, and response to pharmacologic treatments, suggesting their potential for guiding individualized treatment.[2] While this classification has been initially investigated in European populations undergoing bariatric surgery,[3] its relevance and applicability in Chinese diabetic patients undergoing surgery remain unexplored. Ethnic differences are known to influence diabetes phenotypes and treatment response. In the Chinese population, T2DM develops at a lower body mass index (BMI), often accompanied by intrinsic β-cell dysfunction, increased visceral adiposity, and heightened insulin resistance.[4] Furthermore, East Asian patients have been shown to exhibit more favorable glycemic responses following bariatric surgery.[5] In this multicenter study, we applied the clustering framework to two independent cohorts of Chinese T2DM patients undergoing bariatric surgery to evaluate its association with diabetes remission, and to explore its clinical utility in surgical diabetes management. The discovery cohort comprised 207 patients with T2DM who underwent sleeve gastrectomy (SG) or Roux-en-Y gastric bypass (RYGB) between 2011 and 2020 at five centers: Zhongshan Hospital, Drum Tower Hospital, Second Xiangya Hospital, Daping Hospital, and Shanghai Tenth People’s Hospital. The study inclusion criteria were a diagnosis of T2DM at baseline and availability of 1-year postoperative follow-up data. Patients were excluded if they lacked glycemic outcomes or had missing baseline clustering variables [Supplementary Figure 1 and Supplementary Table 1, https://links.lww.com/CM9/C666]. An independent validation cohort comprised 57 T2DM patients who underwent SG at Beijing Friendship Hospital between 2021 and 2023, with 1-year diabetes assessment. The study protocol was approved by the Ethics Committee of Zhongshan Hospital, Fudan University (No. B2021-157R), and local approvals were obtained at all participating centers [Supplementary Table 2, https://links.lww.com/CM9/C666]. Written informed consent was obtained from all participants. Clinical data were anonymized by removing personal identifiers, and data management procedures adhered to institutional and national data protection regulations. Demographics characteristics, medical history, medication use, anthropometric and laboratory data were collected at baseline and 1-year follow-up. Homeostasis Model Assessment version 2-insuin resistance (HOMA2-IR) and HOMA2-β were calculated using the HOMA2 Calculator (University of Oxford, available at: https://www.dtu.ox.ac.uk/homacalculator/). T2DM was diagnosed with the American Diabetes Association criteria. Early-onset diabetes was defined as age of diagnosis <40 years, while late-onset was defined as >40 years old. Diabetic retinopathy was diagnosed via fundus photography, and non-alcoholic fatty liver disease (NAFLD) was diagnosed using ultrasonography or a previous diagnosis. The primary outcome was complete remission of T2DM at 1-year post-surgery. According to the American Society for Metabolic and Bariatric Surgery criteria: complete remission was defined as fasting plasma glucose (FPG) <5.6 mmol/L and HbA1c <6.0%, without anti-diabetic medications use. Partial remission was defined as FPG 5.6–6.9 mmol/L and HbA1c 6.0–6.4%. K-means clustering was performed using five standardized baseline variables: body mass index (BMI), age at diabetes diagnosis, HbA1c, HOMA2-IR, and HOMA2-β, with the k value set to 4. Two patients with extreme outliers (>5 standard deviations from the mean) were excluded. This approach identified four subgroups, SIRD, MARD, SIDD, and MOD. Principal component analysis (PCA) was applied to visualize cluster separation. Two-step clustering with log-likelihood distance and a pre-specified number of four clusters was used for internal validation. Cluster similarity was evaluated using the Jaccard coefficient, with the value >0.75 indicating good stability. Patients in the validation cohort were assigned to clusters based on Euclidean distance to the discovery cohort cluster centroids. Continuous variables were expressed as mean ± standard deviation or median with interquartile range, and categorical variables as percentages. Group comparisons were performed using ANOVA or Student’s t-test for normally distributed variables, Kruskal–Wallis or Wilcoxon rank-sum test for skewed variables, and chi-squared or Fisher’s exact test for categorical variables. Logistic regression was employed to evaluate the association between baseline cluster assignment and diabetes remission. Generalized linear models (GLMs), adjusted for surgery type, were used to compare 1-year changes in clinical parameters. A two-sided P <0.05 was considered statistically significant. All analyses were performed using SPSS v26.0 (IBM Corp., Armonk, NY, USA), and figures were generated with GraphPad Prism v9.0 (GraphPad Software, San Diego, CA, USA) and R v4.2.3 (R Foundation for Statistical Computing, Vienna, Austria). In the discovery cohort, 205 patients (mean age 39.2 years; BMI 38.0 kg/m2; male 90/205 [43.9%]) were classified into SIRD, MARD, SIDD, and MOD. The cluster-specific characteristics differed significantly. SIRD was the most obese and insulin-resistant, with higher β-cell function than MARD and SIDD. MARD was characterized by late-onset diabetes, the lowest BMI, and reduced β-cell function. In contrast, SIDD exhibited the poorest glycemic control and the most impaired β-cell function. MOD was moderately obese, showed higher insulin resistance and had the highest β-cell function (all P <0.001) [Supplementary Figure 2A; Supplementary Tables 3 and 4, https://links.lww.com/CM9/C666]. PCA demonstrated clear separation between clusters, and two-step clustering yielded a comparable distribution (Jaccard index: 0.79) [Supplementary Figures 3 and 4, https://links.lww.com/CM9/C666]. The prevalence of NAFLD also differed among clusters (P = 0.039), with higher proportion in SIRD. Liver enzymes were significantly elevated in SIRD versus MARD and SIDD (all P <0.01), and were numerically higher than in MOD (alanine aminotransferase [ALT]: P = 0.157; aspartate aminotransferase [AST]: P = 0.085; gamma-glutamyl transferase: P = 0.101), suggesting a potential trend. MOD exhibited mild glycemic impairment, with the lowest FPG and 2h-PG levels (both P <0.01). Both MARD and SIDD had longer diabetes duration and more insulin use, MARD displayed reduced insulin secretion (C-peptide), whereas SIDD demonstrated marked insulin deficiency relative to SIRD and MOD. Diabetic retinopathy and coronary artery disease were observed only in MARD and SIDD, whereas hyperuricemia was more common in SIRD [Supplementary Tables 3 and 4, https://links.lww.com/CM9/C666]. As shown in the Sankey diagram, 131 patients (63.9%) achieved 1-year complete remission, more frequently in 92.1% (35/38) of SIRD and 85.1% (40/47) of MOD than in (44.9%) 31/69 of MARD and 49.0% (25/51) of SIDD (P <0.001) [Supplementary Figures 2B, https://links.lww.com/CM9/C666]. Logistic regression analysis, adjusted for sex, surgery type, and center, revealed significantly lower remission odds in MARD (Odds Ratio [OR] 0.08, 95% confidence interval [CI] 0.02–0.28, P <0.001) and SIDD (OR 0.08, 95% CI 0.02–0.31, P <0.001) versus SIRD, whereas MOD exhibited comparable odds (OR 0.41, 95% CI 0.09–1.80, P = 0.236). These associations remained significant after adjustment for diabetes duration and were consistent with findings from the two-step clustering analysis [Supplementary Tables 5 and 6, https://links.lww.com/CM9/C666]. Postoperative metabolic changes also varied across clusters [Supplementary Tables 7–9, https://links.lww.com/CM9/C666]. SIRD exhibited the greatest total weight loss (TWL) (mean difference vs. MOD, SIDD, MARD: 6.5%, 6.2%, 4.2%, all P <0.05), along with the largest reductions in HOMA2-IR, C-peptide, and liver enzymes. In adjusted logistic regression, complete remission was independently associated with TWL (OR 1.11, 95% CI 1.04–1.18; P = 0.001) and ALT change (OR 0.98, 95% CI 0.97–0.99; P = 0.022). MOD also demonstrated greater reductions in HOMA2-IR and C-peptide than MARD and SIDD. Although SIDD showed the most pronounced improvements in glucose levels and HOMA2-β, the remission remained limited. In the validation cohort (mean age 36.7 years; mean BMI 39.6 kg/m2; male 23/57 [40.4%]), cluster distribution and characteristics resembled those of discovery cohort. Consistent with the initial findings, 72.7% (8/11) of SIRD and 81.3% (13/16) of MOD demonstrated higher remission rates than in 33.3% (4/12) of MARD and 16.7% (3/18) of SIDD [Supplementary Table 10; Supplementary Figure 5, https://links.lww.com/CM9/C666]. Our study demonstrates that the clustering framework identified T2DM subgroups with distinct phenotypes and differential responses to bariatric surgery in Chinese population. SIRD and MOD achieved the most favorable remission outcomes, likely attributable to preserved β-cell function and greater improvements in insulin resistance. In SIRD, reductions in liver enzymes also associated with remission. In contrast, MARD and SIDD displayed more advanced β-cell dysfunction, contributing to lower remission. This clustering approach, adapted from a validated framework, incorporated five clinically accessible and biologically relevant variables that reflect key dimensions of T2DM pathophysiology, enabling reliable identification of subgroups with distinct disease trajectories and treatment responses.[2] K-means clustering was selected for its compatibility with continuous variables, relatively strong interpretability, and prior success in diabetes subtyping. The four-cluster solution yielded a stable, and clinically meaningful classification, as confirmed by both internal and external validation. We observed a higher prevalence of SIDD and MARD compared to previous study conducted in Western population,[3] likely reflecting ethnic differences-the lower BMI and limited insulin secretory capacity in East Asians. MARD patients were elder with mild obesity, while SIDD-characterized by early-onset, severe insulin deficiency, and hyperglycemia was more frequent, consistent with recognized β-cell vulnerability in this population. SIRD exhibited pronounced insulin resistance and hyperinsulinemia; its elevated liver enzymes and NAFLD prevalence suggest hepatic insulin resistance may potentially influence remission.[6] The clustering approach effectively distinguished subgroups with varying remission potential. SIRD and MOD demonstrated higher remission likelihood, associated with favorable baseline traits, particularly, elevated C-peptide reflected preserved compensatory β-cell responsiveness. Conversely, MARD and SIDD exhibited advanced β-cell dysfunction, possibly driven by aging, prolonged duration, glucolipotoxicity or genetic susceptibility, limiting remission. Postoperative weight loss, most pronounced in SIRD, and improvements in insulin resistance, especially in SIRD and MOD, were key contributors to remission. In SIRD, reductions in liver enzymes further supported hepatic steatosis as a reversible contributor to remission. Although glycemic control and β-cell function improved in MARD and SIDD, residual β-cell capacity appeared insufficient for remission. Early intervention to alleviate metabolic stress through pharmacotherapy and structured lifestyle modifications (weight management, dietary changes, and physical activity) may help preserve β-cell function. Emerging evidence also points to roles of gut microbiota shifts, gut-brain hormonal signaling, and dietary patterns in postoperative glycemic improvement, offering a broader understanding of remission mechanisms. Clinically, this clustering strategy offers a practical framework for stratifying T2DM patients and guiding expectations regarding surgical outcomes. SIRD and MOD may be considered optimal candidates for surgery for their higher remission likelihood, whereas MARD and SIDD may require tailored postoperative strategies, such as dietary management, physical activity programs, and enhanced behavioral support, to improve glycemic outcomes and sustain weight loss. This classification system may help guide surgical dicision-making and individualize postoperative management strategies. Acknowledgements We thank all the medical staff and field investigators for their work, and study participants for their cooperation. Funding This work was supported by grants from the National Key Research and Development Program of China (No. 2022YFC2505204), Noncommunicable Chronic Diseases-National Science and Technology Major Project (No. 2024ZD0531700) and the National Natural Science Foundation of China (Nos.82370871, and 32241011). Conflicts of interest None.