Subtypes of newly diagnosed type 2 diabetes and risk of complications: analysis of electronic health records in the USA.

Subtypes of newly diagnosed type 2 diabetes and risk of complications: analysis of electronic health records in the USA - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Diabetologia . Author manuscript; available in PMC: 2026 May 31. Published in final edited form as: Diabetologia. 2026 Feb 21;69(6):1545–1556. doi: 10.1007/s00125-026-06687-5 Subtypes of newly diagnosed type 2 diabetes and risk of complications: analysis of electronic health records in the USA Zhongyu Li 1 , Star Liu 2 , Joyce C Ho 3, 4 , K M Venkat Narayan 4, 5 , Mohammed K Ali 4, 5, 6 , Jithin Sam Varghese 4, 5 Author information Article notes Copyright and License information 1 Nutrition and Health Sciences Doctoral Program, Laney Graduate School, Emory University, Atlanta, GA, USA 2 Biomedical Informatics and Data Science Doctoral Program, School of Medicine, Johns Hopkins University, Baltimore, MD, USA 3 Department of Computer Science, College of Arts and Sciences, Emory University, Atlanta, GA, USA 4 Emory Global Diabetes Research Center of the Woodruff Health Sciences Center and Emory University, Atlanta, GA, USA 5 Hubert Department of Global Health, Rollins School of Public Health, Emory University, Atlanta, GA, USA 6 Department of Family and Preventive Medicine, School of Medicine, Emory University, Atlanta, GA, USA ✉ Jithin Sam Varghese, jvargh7@emory.edu Contribution statement JSV and ZL conceptualised the study. JSV, ZL and SL developed the analytic plan with inputs from JCH. ZL and JSV led the data extraction and analysis, and wrote the first draft. All authors contributed to interpretation of the data. ZL and JSV are the guarantors of this work, and, as such, had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. All authors critically reviewed and approved the final version of the manuscript. Issue date 2026 Jun. PMC Copyright notice PMCID: PMC13221843  NIHMSID: NIHMS2173820  PMID: 41723302 The publisher's version of this article is available at Diabetologia Abstract Aims/hypothesis Data-driven subtyping of type 2 diabetes has not been translated into clinical practice due to the lack of routine fasting glucose and insulin measurements. We aimed to identify type 2 diabetes subtypes in clinical settings using electronic health records and study their epidemiology. Methods We identified 727,076 adults (≥18 years) with newly diagnosed type 2 diabetes from Epic Cosmos research platform data across all 50 states and the District of Columbia between 2012 and 2023. Classification models developed in cohort studies were applied to study the sociodemographic distribution of subtypes. Cox proportional hazards regression models, adjusted for age and sex, were used to assess the rates of microvascular complications (retinopathy, neuropathy and nephropathy) and macrovascular complications (severe atherosclerotic cardiovascular disease [ASCVD], other ASCVD and heart failure). Results Among newly diagnosed individuals (mean age 64.4 years [SD 13.3], 52% female), 21.6% were classified as having severe insulin-deficient diabetes (SIDD), 23.8% were classified as having mild obesity-related diabetes (MOD), 40.9% were classified as having mild age-related diabetes and 13.7% were classified as having the mixed subtype. Compared with those classified as having MOD, individuals classified as having SIDD had higher HRs for retinopathy (HR 2.83; 95% CI 2.73, 2.93), neuropathy (HR 1.57; 95% CI 1.54, 1.60), nephropathy (HR 1.34; 95% CI 1.32, 1.37), severe ASCVD (HR 1.49; 95% CI 1.46, 1.53), other ASCVD (HR 1.23; 95% CI 1.21, 1.25) and heart failure (HR 1.17; 95% CI 1.15, 1.20). SIDD and MOD were more prevalent among Hispanics (28.4% and 30.1%, respectively) and non-Hispanic Black people (25.5% and 30.0%, respectively) compared with non-Hispanic White people (20.1% and 21.6%, respectively), and were also more prevalent in the District of Columbia and Utah, respectively, compared with the rest of the country. Conclusions/interpretation Individuals with different type 2 diabetes subtypes, identified through electronic health records, differ in terms of their risk of vascular complications. These findings support leveraging routine electronic health record data to improve the characterisation of patient heterogeneity at the time of diabetes diagnosis. Keywords: Clinical translation, Clusters, Electronic health records, Health disparities, Machine learning, Macrovascular complications, Microvascular complications, Real-world evidence, Subphenotypes Introduction Type 2 diabetes affects over 34 million adults in the USA, with 1.2 million individuals being newly diagnosed annually [ 1 , 2 ]. Recent studies using unsupervised machine learning techniques based on clinical features, biomarkers or omics data have identified distinct subtypes of type 2 diabetes [ 3 - 5 ]. Among the various approaches employed, the most widely replicated method is that from the All New Diabetics in Scania (ANDIS) cohort, which identified five subtypes of diabetes – one autoimmune (GADA-positive) and four type 2 diabetes subtypes – using age, BMI, HbA 1c and homeostatic indices derived from fasting glucose and insulin (or C-peptide) [ 6 ]. These subtypes, namely severe insulin-deficient diabetes (SIDD), severe insulin-resistant diabetes (SIRD), mild obesity-related diabetes (MOD) and mild age-related diabetes (MARD), are associated with differences in clinical characteristics, time to the need for pharmacological treatments, and risks of vascular complications [ 3 , 4 ]. Although the four type 2 diabetes subtypes were replicated in diverse global populations, the observed clinical heterogeneity has not yet been translated into health gains from better risk stratification [ 7 - 12 ]. Most studies that identified the four subtypes were conducted using surveys and observational cohort studies, in which the fasting C-peptide, glucose or insulin data necessary to calculate homeostatic indices are collected [ 13 ]. However, these biomarkers are not routinely collected in clinical practice in the USA, where treatment decisions primarily rely on HbA 1c levels. As a result, it remains unknown whether these subtypes can be replicated for use in routine care using metabolic panels and vital signs available in electronic health records (EHRs). Identifying type 2 subtypes in data from routine clinical settings could address significant knowledge gaps and enable a deeper understanding of the clinical and public health implications of heterogeneity in type 2 diabetes. First, the geographic and sociodemographic variability in type 2 subtypes within the USA remains poorly understood, as findings from ethnically homogeneous European cohorts cannot be generalised to the US population [ 7 , 14 , 15 ]. Understanding these variations could inform public health strategies tailored to vulnerable populations affected by type 2 diabetes. Second, while SIDD and SIRD are linked to increased microvascular complication risks, evidence regarding subtype-specific differences in macrovascular outcomes, such as atherosclerotic cardiovascular disease (ASCVD), remains inconclusive [ 6 , 16 - 21 ]. Small sample size and short follow-up durations limit the ability to detect clinically meaningful differences between subtypes [ 22 , 23 ]. Operationalising type 2 diabetes subtypes in EHRs through application of validated classification models using routinely collected nonfasting clinical measures may enable inexpensive and long-term patient follow-up, and offer opportunities for clinical risk prediction and public health surveillance. Our objectives were to identify type 2 diabetes subtypes in a large EHR database from the USA, and to examine how individuals with these subtypes differ in terms of their sociodemographic distribution and progression to vascular complications. To overcome the barrier of identifying subtypes in the absence of fasting glucose and insulin data, we developed classification models using pooled cohorts from the USA and applied them to EHR data. The findings of this study lay the groundwork for risk stratification and for conducting pragmatic clinical trials of tailored management strategies for type 2 diabetes subtypes in routine care. Methods This study was exempt from the need to receive ethical approval for the analysis of secondary datasets. All participants in the cohort studies and trials gave written informed consent before participation. Study population and data sources Pooled cohort studies Data for developing subtype classification models for newly diagnosed type 2 diabetes came from six NIH-funded longitudinal cohorts: Atherosclerosis Risk in Communities (ARIC), Coronary Artery Risk Development in Young Adults (CARDIA), the non-intervention arm of the Diabetes Prevention Program (DPP), the DPP Outcomes Study (DPPOS – the long-term follow-up of DPP), the Jackson Heart Study (JHS) and the Multi-Ethnic Study of Atherosclerosis (MESA) [ 24 - 28 ]. These cohorts employed internally standardised protocols (see electronic supplementary material [ESM] Tables 1 and 2 ), and enrolled participants with varying clinical and sociodemographic profiles, enabling us to include a broad spectrum of individuals with diabetes for developing and validating subtype classification models. EHRs on the Epic Cosmos research platform Epic Cosmos research platform data were used to operationalise the validated subtype classification algorithm in routine care, and to assess its epidemiological and clinical implications at the population level. The Epic Cosmos research platform data includes de-identified longitudinal EHRs for over 290 million individuals from all 50 states of the USA and the District of Columbia that are compliant with the Health Insurance Portability and Accountability Act (HIPAA) [ 29 ]. Individuals were linked longitudinally across health systems by Epic Cosmos through an internal privacy-preserving process, and are broadly representative of the US population [ 30 ]. Data collection We obtained and harmonised variables relevant to clinical phenotyping at or within 1 year of type 2 diabetes diagnosis, as well as additional demographic information from cohort studies and EHRs. Variables used in phenotyping included age at diagnosis, BMI (kg/m 2 ), fasting blood glucose (mg/dl) and fasting insulin (μIU/ml), HbA 1c (mmol/mol), systolic BP (mmHg), diastolic BP (mmHg), LDL-cholesterol (mg/dl), HDL-cholesterol (mg/dl) and triglycerides (mg/dl) and the triglyceride:HDL-cholesterol ratio (an indirect marker of insulin resistance) ( ESM Tables 2 and 3 ) [ 31 ]. We calculated HOMA2-B and HOMA2-IR based on fasting insulin and fasting blood glucose, using the calculator published by the University of Oxford [ 32 ]. We imputed missing values of cardiometabolic biomarkers using k -nearest neighbour imputation (where k =5). We additionally extracted other clinical and sociodemographic characteristics recorded prior to the diagnosis visit: year of birth, biological sex (male/female), self-reported race/ethnicity (Hispanic, non-Hispanic White, non-Hispanic Black, non-Hispanic other), insurance status (Medicare, Medicaid, private/unspecified, self-pay) for the month of diagnosis, history of comorbidities based on ICD-10-CM codes, and prescriptions for the year prior to diagnosis. The percentile ranking for socioeconomic position for the zip code of residence was available through linked Social Vulnerability Index 2020 data from the Centers for Disease Control and Prevention [ 33 ]. Analytic sample: newly diagnosed type 2 diabetes Pooled cohort studies The definition of a new diagnosis of diabetes was harmonised across the pooled cohorts based on the combination of cohort-specific questionnaire and biomarker data. Participants were classified as having newly diagnosed type 2 diabetes if they did not have a diabetes diagnosis at the time of enrolment to the cohort study, and either self-reported a new diagnosis, reported a diagnosis by a physician or health provider, or had used diabetes medications in the last year. For those who did not self-report a diagnosis of type 2 diabetes, we relied on the 2024 ADA diagnostic criterion to identify individuals with diabetes ( ESM Table 4 ). Using the harmonised definition, we identified 7623 participants with newly diagnosed diabetes in the pooled cohorts. We excluded all individuals for whom age, BMI and HbA 1c were missing at diagnosis of type 2 diabetes ( n =4233) or for whom HOMA2-B and HOMA2-IR values were implausible ( n =13; acceptable input range for HOMA2-B and HOMA2-IR calculation: insulin 20–400 pmol/l; glucose 3.0–25.0 mmol/l]). The excluded individuals were similar to the final analytic sample in terms of demographic and metabolic characteristics ( ESM Table 5 ). The final analytic sample of newly diagnosed type 2 diabetes for the pooled cohort dataset consisted of 3377 individuals ( ESM Figs 1 and 2 ). Epic Cosmos records Adults (aged 18–99 years) with type 2 diabetes that was newly diagnosed between January 2012 and December 2023 were identified using the SUPREME-DM computable phenotype based on inpatient diagnosis codes or any combination of lab results, outpatient diagnosis codes (ICD-10-CM) and diabetes medications occurring within 2 years of each other ( ESM Table 4 ) [ 34 , 35 ]. We defined the date of diagnosis as the date that the second SUPREME-DM criterion was met, and restricted diagnoses to individuals who had in-person encounters in each of the 2 years before detection. The final analytic sample for Epic Cosmos records comprised 727,076 individuals. Data analysis Classification model for subtypes in pooled cohorts We performed de novo hierarchical and k -means clustering (with k values ranging from 2 to 10) using five key variables: age, BMI, HbA 1c , HOMA2-B and HOMA2-IR. Both of these unsupervised approaches identified an optimal solution of four clusters ( ESM Fig. 3 ) [ 36 ]. We used the clusters from the k -means clustering and labelled them based on their similarity to the original ANDIS study subtypes as MARD, MOD, SIRD and SIDD. Our sensitivity analyses, which included leaving one cohort out at a time, comparisons with the original ANDIS centroids, and a complete-case analysis without imputation, suggested the robustness of the identified subtypes ( ESM Methods ). We subsequently constructed one-vs-all logistic regression models to predict the probability of being classified into each of the four subtypes as a function of nine routine clinical variables (i.e. age at diagnosis, BMI, fasting blood glucose and fasting insulin, HbA 1c , systolic and diastolic BP, LDL-cholesterol, HDL-cholesterol, triglycerides and the triglyceride-to-HDL-cholesterol ratio). Each model was fitted on a 70% training subset ( n =2363) of the pooled cohort dataset, and subsequently validated on a 30% holdout test dataset ( n =1014). To determine the optimal predicted probability cut-off for subtype membership classification, we employed a fivefold cross-validation approach. We selected the threshold that maximises the F1 score (a combination of sensitivity and positive predictive value) and summarised key indicators of discrimination (i.e. sensitivity, specificity, AUC, positive predictive value, negative predictive value and F1 score) for each model. De novo clustering using the nine variables did not identify the four ANDIS study subtypes. Classification of subtypes in EHRs We applied the classification models developed in the pooled cohorts to newly diagnosed individuals on the Epic Cosmos research platform. For each participant, predicted probabilities were generated from the one-vs-all logistic regression models. Individuals were assigned to subtypes based on whether they exceeded the predicted probability threshold for SIDD, followed by MARD and MOD, based on the a priori risk of complications [ 8 ]. Given the poor discrimination of the SIRD model (AUC=0.62, F1=0.28), individuals who did not exceed the probability thresholds for the subtypes SIDD, MARD and MOD were categorised as ‘mixed’. Time to pharmacological treatment and risk of complications Cox proportional hazards regression models, adjusted for sex and age, were used to estimate cumulative incidence curves. We estimated five-year cumulative incidence for the prescription of glucose-lowering medication classes (insulin, metformin and incretin mimetics [glucagon-like peptide-1 receptor agonists [GLP-1RA] or GLP-1/glucose-dependent insulinotropic polypeptide receptor agonists [GLP-1/GIP-RA]). We estimated the ten-year cumulative incidence of new-onset microvascular complications (retinopathy, neuropathy and nephropathy) and macrovascular complications (severe ASCVD, heart failure [preserved and reduced ejection fraction] and other ASCVDs) based on billed diagnostic codes ( ESM Table 4 ), among those who were free of each complication at diagnosis. Model validation and evaluation To validate the sequential classification models, we compared the predicted subtypes with those predicted using cluster centroids in the National Health and Nutrition Examination Surveys (NHANES) from 1999 to 2018 ( ESM Methods ). To assess the clinical utility of subtypes, we compared three models: a model with subtypes only, a model with known clinical risk factors for microvascular complications (age, sex, HbA 1c , BMI and systolic BP) and risk scores for macrovascular complications (SCORE-2 linear predictor), and a model that included both subtypes and clinical variables [ 37 ]. Model discrimination and calibration were evaluated using Harrell’s C-index and bootstrap-corrected calibration with 1000 bootstrap resamples, respectively. To assess residual confounding, we examined the association of subtypes with pharyngitis and allergic rhinitis as negative controls [ 38 ]. All analyses were performed using Python version 3.12.1 (Python Software Foundation, Wilmington, DE, USA) and R version 4.2.3 (R Foundation for Statistical Computing, Vienna, Austria). Results The individuals with newly diagnosed diabetes ( n =727,076) identified on the Epic Cosmos research platform were 64.4±13.3 years old (mean ± SD), 51.5% female, 67.6% non-Hispanic White, 16.8% non-Hispanic Black and 7.0% Hispanic. Their cardiometabolic profiles (median HbA 1c 53 mmol/mol; IQR 49–63 mmol/mol [7.0%, IQR 6.6–7.9%]; systolic BP 131.2±15.5 mmHg; median triglyceride:HDL-cholesterol ratio 3.3; IQR 2.1–5.3) were worse compared with the pooled cohort data ( Table 1 and ESM Table 5 ). Table 1. Descriptive characteristics of individuals with newly detected type 2 diabetes by subtype on the Epic Cosmos research platform Characteristic Overall SIDD MARD MOD Mixed N 727,076 156,951 297,568 172,922 99,635 Age at detection of SUPREME-DM phenotype (years) 64.4±13.3 59.9±13.5 74.8±7.9 51.9±9.4 62.9±5.7 Female 51.5 50.0 49.8 57.6 48.6 Race/ethnicity Non-Hispanic White 67.6 62.9 73.3 61.5 69.0 Non-Hispanic Black 16.8 19.8 13.1 21.2 15.3 Hispanic 7.0 9.3 4.8 8.9 6.8 Non-Hispanic other 8.6 8.0 8.8 8.5 8.8 Social vulnerability index a 60.4 (33.5–82.1) 64.1 (37.0–84.6) 57.4 (31.2–79.7) 62.7 (35.5–83.8) 59.6 (32.8–81.4) Insurance Medicare 39.6 30.8 61.7 14.2 31.6 Medicaid 13.5 16.3 9.8 17.8 12.9 Key biomarkers HbA 1c (mmol/mol) 53 (49–63) 80 (72–97) 50 (48–53) 51 (48–55) 54 (48–60) HbA 1c (%) 7.0 (6.6–7.9) 9.5 (8.7–11.0) 6.7 (6.5–7.0) 6.8 (6.5–7.2) 7.1 (6.8–7.6) BMI (kg/m 2 ) 33.2±6.8 34.0±6.9 29.5±5.1 38.8±6.1 33.8±5.3 Systolic BP (mmHg) 131.2±15.5 132.5±16.3 131.1±15.7 130.9±14.8 129.9±15.0 Diastolic BP (mmHg) 74.7±10.2 76.2±10.4 71.4±9.5 78.7±9.7 74.9±9.3 LDL-cholesterol (mmol/l) 2.31±0.98 2.49±1.08 2.10±0.90 2.53±0.96 2.19±0.92 HDL-cholesterol (mmol/l) 1.16±0.38 1.11±0.39 1.24±0.39 1.13±0.34 1.10±0.36 Triglycerides (mmol/l) 1.86±1.04 2.10±1.2 1.60±0.83 1.96±1.06 2.08±1.15 Triglyceride:HDL-cholesterol ratio 3.3 (2.1–5.3) 3.9 (2.4–6.4) 2.8 (1.8–4.2) 3.6 (2.3–5.5) 3.9 (2.5–6.4) Follow-up duration (months) 39.9 (19.9–72.9) 40.9 (21.0–74.8) 35.0 (16.0–65.9) 45.9 (23.0–79.8) 44.9 (23.0–79.8) Open in a new tab Values are mean ± SD or median (IQR) for continuous variables, and percentages for categorical variables The number of observations for which values are missing is presented in ESM Table 3 a For the social vulnerability index, 0 represents low vulnerability and 100 represents high vulnerability Classification models for incident diabetes in the pooled cohorts In the analytic pooled cohort sample ( n =3377), the mean age of participants at diagnosis of type 2 diabetes was 63.4 years (SD 12.4). The overall pooled cohort was 59.8% female, 49.9% White, 35.5% Black and 14.6% from other racial groups, and had a median HbA 1c of 45 mmol/mol with IQR 40–50 mmol/mol (6.3%; IQR 5.8–6.7%), mean BMI of 33.2 kg/m 2 (SD 7.0), median HOMA2-B of 108 (IQR 73.3–156.4) and median HOMA2-IR of 2.8 (IQR 1.4–4.7). Descriptive characteristics of the pooled cohort sample, EHR sample and each cohort study are presented in ESM Tables 5 - 7 . The individuals classified as having MARD were older (mean ± SD 72.5±7.6 years) with lower BMI (mean ± SD 29.4±4.6 kg/m 2 ) and HOMA2-IR (median 1.9; IQR 1.3–2.9). MOD was characterised by a younger age (mean ± SD 52±7.9 years) and higher BMI (mean ± SD 37.9±7.0 kg/m 2 ) ( ESM Table 7 ). SIRD was characterised by high BMI (mean ± SD 32.4±6.0 kg/m 2 ), HOMA2-B (median 300; IQR 249.8–367.4) and HOMA2-IR (median 9.0; IQR 7.4–11.8), and later age at diagnosis (mean ± SD 65.7±9.4 years) than individuals classified as having MOD and SIDD but earlier than those classified as having MARD. SIDD was characterised by an early age at onset (mean ± SD 61.1±13.7 years), high HbA 1c (median 79 mmol/mol [IQR 72–90 mmol/mol] or 9.4% [IQR 8.7–10.4%]) and poor beta cell function (median HOMA2-B 41.9; IQR 21.9–72.8). The distribution of the five key variables and nine clinical variables used in developing the logistic regression models for subsequent EHR application are shown in ESM Fig. 4 and ESM Fig. 5 , respectively. The one-vs-all logistic regression models using nine clinical variables demonstrated strong discrimination, with minimal loss of performance between the training and test datasets ( ESM Table 8 ). The optimal predicted probability thresholds for classification that maximised the F1 score were 0.16, 0.38 and 0.32, respectively, for SIDD, MOD and MARD. In particular, the model for classifying SIDD exhibited high discrimination, with an AUC of 0.99 (95% CI 0.99, 1.00) and an F1 score of 0.90. Similarly, the models for MOD and MARD had high AUCs of 0.96 (95% CI 0.95, 0.97) and 0.93 (95% CI 0.92, 0.95), and F1 scores of 0.88 and 0.87, respectively. Coefficients and standard errors for the models are presented in ESM Table 9 . Epidemiology of diabetes subtypes in EHRs The classification models used to predict subtype membership based on EHR data from the Epic Cosmos platform identified 21.6% individuals as having SIDD, 40.9% as having MARD, 23.8% as having MOD, and 13.7% as mixed subtype, with median follow-up durations ranging from 35 to 46 months. The distribution of the nine clinical variables used in subtyping type 2 diabetes is shown in Fig. 1 . Descriptive characteristics for subtypes were similar between cohort studies and EHRs ( Table 1 ). For instance, individuals classified as having MOD had the highest BMI (mean ± SD 38.8±6.1 kg/m 2 ), while those classified as having MARD had the lowest mean BMI (mean ± SD 29.5±5.1 kg/m 2 ). Individuals classified as having SIDD had the highest mean HbA 1c at diagnosis (median 80 mmol/mol [IQR 72–97 mmol/mol] or 9.5% [IQR 8.7–11.0%]). Those classified as mixed subtype (i.e. participants who were not classified into any of the other three subtypes) had a cardiometabolic profile resembling that of individuals classified as having MOD. Fig. 1. Open in a new tab Distribution of clinical characteristics of individuals classified as having the various subtypes of type 2 diabetes on the Epic Cosmos research platform. TG:HDL, triglyceride:HDL-cholesterol ratio The social vulnerability index was higher for individuals classified as having SIDD (median 64.1; IQR 37.0–84.6) or MOD (median 62.7; IQR 35.5–83.8) compared with those classified as having MARD (median 57.4; IQR 31.2–79.7). In addition, SIDD and MOD were more common among Hispanics (28.4% and 30.1%, respectively) and non-Hispanic Black adults (25.5% and 30.0%, respectively) compared with non-Hispanic White adults (20.1% and 21.6%, respectively). The crude distribution of SIDD, MOD and MARD among individuals newly diagnosed with type 2 diabetes in EHRs also showed regional variations ( Fig. 2 ). For instance, the District of Columbia (27.8%) and South Carolina (26.1%) had the highest proportions of cases classified as SIDD, while Idaho (18.7%) and Kansas (18.9%) had the lowest proportions ( Fig. 2a ). MARD was predominant across the USA ( Fig. 2b ). Utah had the highest proportion of MOD (29.2%), followed by Colorado (28.7%) and Alaska (27.4%) ( Fig. 2c ). We observed additional geographic variations in subtypes when further stratifying by race and ethnicity. Relatively higher proportions of non-Hispanic Black and Hispanic adults were classified as having SIDD (≥20%) or MOD (≥30%), while the proportion classified as having MARD was highest among non-Hispanic White people across the majority of states ( ESM Fig. 6 ). Fig. 2. Open in a new tab Geographic distribution of type 2 diabetes subtypes for individuals on the Epic Cosmos research platform. Estimates are based on a sample of 727,076 observations. ( a ) SIDD, ( b ) MARD, ( c ) MOD Time to pharmacological treatments The time to prescription of insulin (HR 1.65; 95% CI 1.62, 1.67) and incretin mimetics (HR 1.22; 95% CI 1.20, 1.24) was shorter among individuals classified as having SIDD ( Fig. 3a - c ) in the first 5 years after diagnosis compared with individuals classified as having MOD. Individuals classified as having SIDD were also less likely to receive metformin (HR 0.92; 95% CI 0.91, 0.94). Compared with individuals classified as having MOD, initiation of prescription of insulin (HR 0.87; 95% CI 0.86, 0.89), metformin (HR 0.81; 95% CI 0.80, 0.83) and incretin mimetics (HR 0.42; 95% CI 0.41, 0.43) was later among individuals classified as having MARD. Fig. 3. Open in a new tab Time to pharmacological treatments and microvascular and macrovascular complications after diagnosis of type 2 diabetes for individuals on the Epic Cosmos research platform. Numbers at risk indicate the number of individuals in each subtype who had not yet experienced the event or been prescribed the medication and remained under follow-up at each time point. Sex and age-adjusted cumulative incidence curves for ( a ) time to insulin prescription, ( b ) time to metformin prescription, ( c ) time to prescription of incretin mimetics (GLP-1RA or GLP-1/GIP-RA), ( d ) time to retinopathy diagnosis, ( e ) time to neuropathy diagnosis, ( f ) time to nephropathy diagnosis, ( g ) time to any severe ASCVD event (non-ST-elevation myocardial infarction, ST-elevation myocardial infarction, other acute coronary syndromes or stroke), ( h ) time to other ASCVD events (peripheral artery disease, transient ischaemic attack, coronary heart disease, atrial fibrillation, stable angina or cardiomyopathy), and ( i ) time to first heart failure event (preserved or reduced ejection fraction). HRs are provided in ESM Tables 10 and 11 . Unadjusted Kaplan–Meier cumulative incidence curves are shown in ESM Figs 11 and 12 Time to microvascular complications Compared with individuals classified as having MOD, individuals classified as having SIDD were more likely to develop retinopathy (HR 2.83; 95% CI 2.73, 2.93), neuropathy (HR 1.57; 95% CI 1.54, 1.60) or nephropathy (HR 1.34; 95% CI 1.32, 1.37) in the first 10 years after diagnosis ( Fig. 3d - f ). Similarly, individuals classified as having MARD were more likely to develop retinopathy (HR 1.21; 95% CI 1.16, 1.26) but less likely to develop neuropathy (HR 0.90; 95% CI 0.88, 0.92) or nephropathy (HR 0.95; 95% CI 0.93, 0.97) compared with individuals classified as having MOD. Individuals classified as having the mixed subtype were more likely to develop retinopathy (HR 1.43; 95% CI 1.37, 1.49), neuropathy (HR 1.10; 95% CI 1.08, 1.13) or nephropathy (HR 1.03; 95% CI 1.01, 1.05) compared with individuals classified as having MOD. Time to macrovascular complications Compared with individuals classified as having MOD, those classified as having SIDD had elevated rates of severe ASCVD events (HR 1.49; 95% CI 1.46, 1.53), heart failure (HR 1.17; 95% CI 1.15, 1.20) or other ASCVD events (HR 1.23; 95% CI 1.21, 1.25) ( Fig. 3g - i ). Individuals classified as having MARD had similar rates of severe ASCVD, heart failure and other ASCVD to those classified as having MOD. Individuals classified as having the mixed subtype displayed rates between those for individuals classified as having SIDD or MOD for severe ASCVD and other ASCVD. Similar results were observed when comparing SIDD with all non-SIDD subtypes ( ESM Table 12 and ESM Fig. 7 ). Consistent patterns were also observed when examining individual ASCVD endpoints and subtypes of heart failure ( ESM Table 13 ). Sensitivity analyses In external validation using NHANES between 1999 and 2018, the classification models showed high discrimination for SIDD (AUC 0.99), MOD (AUC 0.94) and MARD (AUC 0.97), indicating that generalisation beyond the pooled cohorts is possible ( ESM Fig. 8 ). Subtype-only models were non-inferior to models that included clinical risk factors and the SCORE-2 linear predictor. Addition of subtypes to the latter models did not meaningfully change the results ( ESM Table 14 ). The negative control analysis showed no evidence of spurious positive associations between subtype and unrelated health conditions: compared with MOD, the HRs for SIDD were 0.81 (95% CI 0.80, 0.83) for allergic rhinitis, and 0.88 (95% CI 0.86, 0.90) for pharyngitis, with near-null or slightly inverse associations for other subtypes ( ESM Table 15 ). Discussion By operationalising reliable classification models developed in pooled cohorts to use routinely collected clinical data in EHRs, we examined the epidemiology and prognosis of type 2 diabetes subtypes in the USA. The proportions of subtypes identified in the EHRs were consistent with previous studies, with MARD being the most common, followed by MOD and SIDD [ 13 ]. Over one in five individuals with newly diagnosed type 2 diabetes was classified as having SIDD, and these individuals had the highest rates of both microvascular and macrovascular complications; SIDD was also disproportionately present in minority racial and ethnic groups. The distinct geographic and sociodemographic distributions of the subtypes have significant implications for diabetes surveillance, clinical decision support, and health equity and policy. Consistent with previous studies, subtypes were non-inferior to established clinical risk factors in terms of predicting the onset of complications [ 4 , 17 ], and present an interpretable and reproducible summary of type 2 diabetes heterogeneity at diagnosis [ 4 ]. Our findings are consistent with previous studies that indicated elevated rates of all microvascular complications among individuals classified as having SIDD, compared with individuals classified as having MOD or MARD [ 3 , 4 , 6 , 39 ]. Some studies also suggest that individuals classified as having SIRD have similar rates of retinopathy, but higher rates of kidney-related outcomes such as nephropathy and albuminuria, compared with individuals classified as having MOD or MARD ( ESM Fig. 9 ) [ 6 , 18 , 39 - 42 ]. In our sample, individuals classified as having the mixed subtype had similar rates of retinopathy and nephropathy as those classified as having MOD and MARD. The associations of subtypes with macrovascular complications were less consistent among previous studies, probably due to a combination of low event rates, smaller sample size and shorter durations of follow-up ( ESM Fig. 10 ) [ 6 , 17 - 19 , 39 , 43 , 44 ]. In the present study, individuals classified as having SIDD showed the highest overall rates of macrovascular complications, particularly for severe ASCVD, while those classified as having the mixed subtype showed lower rates of severe ASCVD but intermediate rates for heart failure. The elevated risk of severe ASCVD among individuals classified as having SIDD is biologically plausible given their characteristic severe hyperglycaemia, which promotes formation of advanced glycation end-products, oxidative stress, endothelial dysfunction and chronic inflammation [ 22 ]. These processes, together with insulin deficiency-related proinflammatory and pro-thrombotic effects, may drive atherothrombotic complications. This framework aligns with previous observations that individuals with microvascular damage are also at increased risk for macrovascular disease [ 23 ], highlighting the need for further work to disentangle how hyperglycaemia, insulin deficiency and insulin resistance jointly shape vascular outcomes across subtypes. Beyond such pathophysiological pathways, classification of individuals as having SIDD may also capture underlying aetiological social and environmental determinants of health, such as structural disparities in healthcare access, which are associated with adverse cardiovascular outcomes [ 45 ]. Our findings highlight the overlap of clinical severity and sociodemographic vulnerability in the USA: classification of individuals as having SIDD was more prevalent among non-Hispanic Black and Hispanic adults compared with non-Hispanic White people, MARD was the most common subtype, particularly in non-Hispanic White populations, and MOD was more common among younger individuals [ 7 ]. Despite the current approaches in clinical practice that may prioritise prescription of incretin mimetics for individuals with SIDD even without awareness of their subtype membership, nearly half of the individuals classified as having SIDD were not treated with insulin or incretin mimetics in the first 5 years after diagnosis. These patterns align with previous studies that used smaller cohorts of individuals with new and previously diagnosed type 2 diabetes in national surveys and individuals within a university health system [ 46 , 47 ]. The disproportionately high prevalence of individuals classified as having SIDD among minority racial and ethnic groups suggests that environmental factors and gene–environment interactions may contribute to this burden, further emphasising the importance of equitable treatment strategies as well as opportunities to integrate social determinants of health into precision prognosis regarding diabetic complications. This study has several strengths. First, we developed reliable subtyping algorithms that use routine clinical measurements from cohort studies that have a low risk of undiagnosed diabetes or missing data due to left censoring of newly diagnosed diabetes. Second, we applied harmonised definitions across cohort studies, combining self-reported data and glycaemic biomarkers to minimise information bias. Third, we used a validated computable phenotype for type 2 diabetes newly diagnosed in EHRs, and characterised the epidemiology and progression of the subtypes by leveraging the largest integrated EHR database in the USA. Fourth, we externally validated the classification algorithms in the nationally representative NHANES cohort, demonstrating high discrimination and generalisability of subtype identification using routine clinical variables. Fifth, we incorporated negative control outcomes to evaluate potential residual confounding and support the specificity of subtype–outcome associations. This study also has several limitations. First, transporting classification models from cohort studies to EHR data requires the assumption of no unmeasured confounding and effect modification. Furthermore, the pooled cohorts may not fully represent the heterogeneity in diabetes within the USA, particularly representation of minority populations such as Alaska Natives, Native Americans, South Asians and Pacific Islanders. Although we conducted several sensitivity analyses to assess the robustness of the findings, there is a need for further validation. Second, this study is susceptible to biases associated with use of data from EHRs. The SUPREME-DM computable phenotype, although highly specific, does not specify the order in which the criteria (lab results, medications, diagnostic codes) for an individual to be classified as newly diagnosed with type 2 diabetes are to be met. Therefore, nearly one in three individuals and one in six individuals, respectively, had a history of insulin use or diabetic nephropathy on or before the diagnosis date in the analytic sample, suggesting that the duration of type 2 diabetes may be underestimated. Additionally, the high sensitivity and low specificity of ICD-10-CM codes may overestimate the risk of diabetes complications [ 34 ]. Third, although the translation of type 2 diabetes subtypes from cohort studies to EHRs was successful for SIDD, MOD and MARD, opportunities exist for improving the classification of individuals in EHRs as having SIRD. Finally, although the subtypes are easily interpretable, biologically plausible and widely replicated, they may be less effective than continuous risk prediction models for predicting complications [ 17 ]. Nevertheless, advancements in risk prediction models and data-driven subtyping may be mutually informative: subtypes can capture clinically meaningful heterogeneity and prediction models can help to refine and validate subtype definitions to better characterise diabetes heterogeneity. Given the current focus on translating precision medicine into public health [ 48 ], we underscore two takeaways. First, use of routine administrative data in clinical practice can enable inexpensive surveillance of high-risk subtypes of newly diagnosed type 2 diabetes and monitoring of their prognosis across diverse populations [ 49 ]. Second, the overlap of high-risk subtypes with social vulnerability emphasises the need for equitable access to high-quality diabetes care. By operationalising diabetes subtypes in EHRs, this study enables opportunities for pragmatic trials of tailored treatment protocols and targeted public health surveillance, bridging a crucial gap in addressing the burden of diabetes. Supplementary Material Supplementary File NIHMS2173820-supplement-Supplementary_File.pdf (1.7MB, pdf) Supplementary Information The online version of this article ( https://doi.org/10.1007/s00125-026-06687-5 ) contains peer-reviewed but unedited supplementary material. Research in context. What is already known about this subject? Subtyping type 2 diabetes may improve risk stratification and personalised treatment The four-subtype framework comprising severe insulin-deficient diabetes (SIDD), severe insulin-resistant diabetes (SIRD), mild obesity-related diabetes (MOD) and mild age-related diabetes (MARD) has been replicated globally Implementation and integration of type 2 diabetes subtype classification within US clinical practice is limited due to the reliance of classification methods on fasting insulin and glucose, which are infrequently measured at diagnosis in the USA What is the key question? Can type 2 diabetes subtypes be identified using routinely collected clinical data, and how do they differ in terms of their risks of vascular complications? What are the new findings? Subtypes of type 2 diabetes derived from electronic health record data have varying distributions across US geographic regions, and distinct patterns of complications Individuals classified as having SIDD had 34–183% higher HRs for microvascular complications and 17–49% higher HRs for macrovascular complications than those classified as having MOD High-risk type 2 diabetes subtypes were disproportionately prevalent among minority populations, including Hispanic and non-Hispanic Black individuals How might this impact on clinical practice in the foreseeable future? Use of routinely available clinical data can support early subtype classification of type 2 diabetes to characterise patient heterogeneity at the time of diagnosis Acknowledgements We thank A. Tobarran (Office of Information Technology, Emory Healthcare) and G. Najarro (Georgia Clinical and Translational Science Alliance, Emory School of Medicine) for facilitating access to the Epic Cosmos platform, as well as K. Yates, M. Howat and D. Sandman from the Epic Cosmos team for their support. We also thank P. Vellanki (Division of Endocrinology, Emory School of Medicine) for her thoughtful comments. A previous version of this article was submitted to the medRxiv preprint server ( https://www.medrxiv.org/content/10.1101/2024.10.08.24315128v1 ) on 8 October 2024. Funding This research received no specific grant from any funding agency in the public, commercial or not-for-profit sectors. Abbreviations ANDIS All New Diabetics in Scania ASCVD Atherosclerotic cardiovascular disease EHR Electronic health record GLP-1RA Glucagon-like peptide-1 receptor agonists GLP-1/GIP-RA Glucagon-like peptide-1/glucose-dependent insulinotropic polypeptide receptor agonists MARD Mild age-related diabetes MOD Mild obesity-related diabetes NHANES National Health and Nutrition Examination Survey SIDD Severe insulin-deficient diabetes SIRD Severe insulin-resistant diabetes Data availability Data for the NIH-funded cohort studies are available to registered users on request from the NIDDK Central Repository (NIDDK-CR; https://repository.niddk.nih.gov ) and NHLBI BioLINCC ( https://biolincc.nhlbi.nih.gov/home/ ). Epic Cosmos access is available through institutional representatives of participating institutions and the Epic Cosmos team after completing certification requirements. Code availability The code for reproducing the subtypes is available on https://github.com/chroniq-lab/diabetes_subtypes_cohorts . References 1. Centers for Disease Control & Prevention (2023) National Diabetes Statistics Report. Available from https://www.cdc.gov/diabetes/php/data-research/index.html . Accessed 20 Jan 2026 2. International Diabetes Federation (n.d.) The Diabetes Atlas. Available from https://diabetesatlas.org/ . Accessed 6 Oct 2022 3. Herder C, Roden M (2022) A novel diabetes typology: towards precision diabetology from pathogenesis to treatment. Diabetologia 65(11):1770–1781. 10.1007/s00125-021-05625-x [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Stefan N, Schulze MB (2023) Metabolic health and cardiometabolic risk clusters: implications for prediction, prevention, and treatment. Lancet Diabetes Endocrinol 11(6):426–440. 10.1016/S2213-8587(23)00086-4 [ DOI ] [ PubMed ] [ Google Scholar ] 5. Deutsch AJ, Ahlqvist E, Udler MS (2022) Phenotypic and genetic classification of diabetes. Diabetologia 65(11):1758–1769. 10.1007/s00125-022-05769-4 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Ahlqvist E, Storm P, Käräjämäki A et al. (2018) Novel subgroups of adult-onset diabetes and their association with outcomes: a data-driven cluster analysis of six variables. Lancet Diabetes Endocrinol 6:361–369. 10.1016/S2213-8587(18)30051-2 [ DOI ] [ PubMed ] [ Google Scholar ] 7. Varghese JS, Narayan KMV (2022) Ethnic differences between Asians and non-Asians in clustering-based phenotype classification of adult-onset diabetes mellitus: a systematic narrative review. Primary Care Diabetes 16(6):853–856. 10.1016/j.pcd.2022.09.007 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Misra S, Wagner R, Ozkan B et al. (2023) Precision subclassification of type 2 diabetes: a systematic review. Commun Med 3:138. 10.1038/s43856-023-00360-3 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Chung WK, Erion K, Florez JC et al. (2020) Precision medicine in diabetes: a consensus report from the American Diabetes Association (ADA) and the European Association for the Study of Diabetes (EASD). Diabetes Care 43:1617–1635. 10.2337/dci20-0022 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Li X, van Giessen A, Altunkaya J et al. (2023) Potential value of identifying type 2 diabetes subgroups for guiding intensive treatment: a comparison of novel data-driven clustering with risk-driven subgroups. Diabetes Care 46(7):1395–1403. 10.2337/dc22-2170 [ DOI ] [ PubMed ] [ Google Scholar ] 11. Dwibedi C, Ekström O, Brandt J et al. (2024) Randomized open-label trial of semaglutide and dapagliflozin in patients with type 2 diabetes of different pathophysiology. Nat Metab 6:50–60. 10.1038/s42255-023-00943-3 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Shields BM, Dennis JM, Angwin CD et al. (2023) Patient stratification for determining optimal second-line and third-line therapy for type 2 diabetes: the TriMaster study. Nat Med 29:376–383. 10.1038/s41591-022-02120-7 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Varghese JS, Carrillo-Larco RM, Narayan KV (2023) Achieving replicable subphenotypes of adult-onset diabetes. Lancet Diabetes Endocrinol 11(9):635–636. 10.1016/S2213-8587(23)00195-X [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Anjana RM, Baskar V, Nair ATN et al. (2020) Novel subgroups of type 2 diabetes and their association with microvascular outcomes in an Asian Indian population: a data-driven cluster analysis: the INSPIRED study. BMJ Open Diab Res Care 8:e001506. 10.1136/bmjdrc-2020-001506 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Ke C, Narayan KMV, Chan JCN, Jha P, Shah BR (2022) Pathophysiology, phenotypes and management of type 2 diabetes mellitus in Indian and Chinese populations. Nat Rev Endocrinol 18(7):413–432. 10.1038/s41574-022-00669-4 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Aoki Y, Hamrén B, Clegg LE et al. (2021) Assessing reproducibility and utility of clustering of patients with type 2 diabetes and established CV disease (SAVOR-TIMI 53 trial). PLoS ONE 16:e0259372. 10.1371/journal.pone.0259372 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Dennis JM, Shields BM, Henley WE, Jones AG, Hattersley AT (2019) Disease progression and treatment response in data-driven subgroups of type 2 diabetes compared with models based on simple clinical features: an analysis using clinical trial data. Lancet Diabetes Endocrinol 7:442–451. 10.1016/S2213-8587(19)30087-7 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Pigeyre M, Hess S, Gomez MF et al. (2022) Validation of the classification for type 2 diabetes into five subgroups: a report from the ORIGIN trial. Diabetologia 65:206–215. 10.1007/s00125-021-05567-4 [ DOI ] [ PubMed ] [ Google Scholar ] 19. Tanabe H, Saito H, Kudo A et al. (2020) Factors associated with risk of diabetic complications in novel cluster-based diabetes subgroups: a Japanese retrospective cohort study. J Clin Med 9:2083. 10.3390/jcm9072083 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Xiong X, Yang Y, Wei L, Xiao Y, Li L, Sun L (2021) Identification of two novel subgroups in patients with diabetes mellitus and their association with clinical outcomes: a two-step cluster analysis. J Diabetes Investig 12:1346–1358. 10.1111/jdi.13494 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Zou X, Huang Q, Luo Y et al. (2022) The efficacy of canagliflozin in diabetes subgroups stratified by data-driven clustering or a supervised machine learning method: a post hoc analysis of canagliflozin clinical trial data. Diabetologia 65:1424–1435. 10.1007/s00125-022-05748-9 [ DOI ] [ PubMed ] [ Google Scholar ] 22. Li Y, Liu Y, Liu S et al. (2023) Diabetic vascular diseases: molecular mechanisms and therapeutic strategies. Sig Transduct Target Ther 8:152. 10.1038/s41392-023-01400-z [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Brownrigg JRW, Hughes CO, Burleigh D et al. (2016) Microvascular disease and risk of cardiovascular events among individuals with type 2 diabetes: a population-level cohort study. Lancet Diabetes Endocrinol 4:588–597. 10.1016/S2213-8587(16)30057-2 [ DOI ] [ PubMed ] [ Google Scholar ] 24. Rosamond WD, Folsom AR, Chambless LE, Wang C-H (2001) Coronary heart disease trends in four United States communities. The Atherosclerosis Risk in Communities (ARIC) Study 1987–1996. Int J Epidemiol 30:S17–S22. 10.1093/ije/30.suppl_1.s17 [ DOI ] [ PubMed ] [ Google Scholar ] 25. Liu K, Daviglus ML, Loria CM et al. (2012) Healthy lifestyle through young adulthood and the presence of low cardiovascular disease risk profile in middle age: the Coronary Artery Risk Development in (Young) Adults (CARDIA) Study. Circulation 125:996–1004. 10.1161/CIRCULATIONAHA.111.060681 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. White NH, Pan Q, Knowler WC et al. (2022) The effect of interventions to prevent type 2 diabetes on the development of diabetic retinopathy: the DPP/DPPOS Experience. Diabetes Care 45:1640–1646. 10.2337/dc21-2417 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Echouffo-Tcheugui JB, Musani SK, Bertoni AG, Correa A, Fox ER, Mentz RJ (2022) Patients phenotypes and cardiovascular risk in type 2 diabetes: the Jackson Heart Study. Cardiovasc Diabetol 21:89. 10.1186/s12933-022-01501-z [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Bild DE (2002) Multi-ethnic study of atherosclerosis: objectives and design. Am J Epidemiol 156:871–881. 10.1093/aje/kwf113 [ DOI ] [ PubMed ] [ Google Scholar ] 29. Tarabichi Y, Frees A, Honeywell S et al. (2021) The Cosmos Collaborative: a vendor-facilitated electronic health record data aggregation platform. ACI Open 05:e36–e46. 10.1055/s-0041-1731004 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Epic Systems Corporation (n.d.) About Epic Cosmos. Available from https://cosmos.epic.com/about/ . Accessed 28 Jul 2024 31. Giannini C, Santoro N, Caprio S et al. (2011) The triglyceride-to-HDL cholesterol ratio. Diabetes Care 34:1869–1874. 10.2337/dc10-2234 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Levy JC, Matthews DR, Hermans MP (1998) Correct homeostasis model assessment (HOMA) evaluation uses the computer program. Diabetes Care 21:2191–2192. 10.2337/diacare.21.12.2191 [ DOI ] [ PubMed ] [ Google Scholar ] 33. Flanagan BE, Hallisey EJ, Adams E, Lavery A (2018) Measuring community vulnerability to natural and anthropogenic hazards: the Centers for Disease Control and Prevention’s social vulnerability index. J Environ Health 80(10):34–36 [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Richesson RL, Rusincovitch SA, Wixted D et al. (2013) A comparison of phenotype definitions for diabetes mellitus. J Am Med Inform Assoc 20:e319–e326. 10.1136/amiajnl-2013-001952 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Nichols GA, Schroeder EB, Karter AJ et al. (2015) Trends in diabetes incidence among 7 million insured adults, 2006–2011. Am J Epidemiol 181:32–39. 10.1093/aje/kwu255 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Satopaa V, Albrecht J, Irwin D, Raghavan B. finding a “kneedle” in a haystack: detecting knee points in system behavior. In Proceedings of the 31st International Conference on Distributed Computing Systems Workshops, Minneapolis, MN, USA. IEEE, New York, pp. 166–171. 10.1109/ICDCSW.2011.20 [ DOI ] [ Google Scholar ] 37. SCORE2-Diabetes Working Group and the ESC Cardiovascular Risk Collaboration (2023) SCORE2-Diabetes: 10-year cardiovascular risk estimation in type 2 diabetes in Europe. Eur Heart J 44:2544–2556. 10.1093/eurheartj/ehad260 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 38. Lipsitch M, TchetgenTchetgen E, Cohen T (2010) Negative controls: a tool for detecting confounding and bias in observational studies. Epidemiology 21:383–388. 10.1097/EDE.0b013e3181d61eeb [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 39. Bjarkø VV, Haug EB, Langhammer A et al. (2024) Clinical utility of novel diabetes subgroups in predicting vascular complications and mortality: up to 25 years of follow-up of the HUNT Study. BMJ Open Diab Res Care 12:e004493. 10.1136/bmjdrc-2024-004493 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 40. Preechasuk L, Khaedon N, Lapinee V et al. (2022) Cluster analysis of Thai patients with newly diagnosed type 2 diabetes mellitus to predict disease progression and treatment outcomes: a prospective cohort study. BMJ Open Diab Res Care 10:e003145. 10.1136/bmjdrc-2022-003145 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 41. Fedotkina O, Sulaieva O, Ozgumus T et al. (2021) Novel reclassification of adult diabetes is useful to distinguish stages of β-cell function linked to the risk of vascular complications: the DOLCE Study from Northern Ukraine. Front Genet 12:637945. 10.3389/fgene.2021.637945 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 42. Zaharia OP, Strassburger K, Strom A et al. (2019) Risk of diabetes-associated diseases in subgroups of patients with recent-onset diabetes: a 5-year follow-up study. Lancet Diabetes Endocrinol 7:684–694. 10.1016/S2213-8587(19)30187-1 [ DOI ] [ PubMed ] [ Google Scholar ] 43. Wang J, Liu J-J, Gurung RL et al. (2022) Clinical variable-based cluster analysis identifies novel subgroups with a distinct genetic signature, lipidomic pattern and cardio-renal risks in Asian patients with recent-onset type 2 diabetes. Diabetologia 65(12):2146–2156. 10.1007/s00125-022-05741-2 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 44. Song X, Lv Y, Huang N et al. (2022) Clinical characteristics of inpatients with new-onset diabetes mellitus in eastern China: based on novel clustering analysis. Front Endocrinol 13:927661. 10.3389/fendo.2022.927661 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 45. Teshale AB, Htun HL, Owen A et al. (2023) The role of social determinants of health in cardiovascular diseases: an umbrella review. J Am Heart Assoc 12:e029765. 10.1161/JAHA.123.029765 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 46. Antonio-Villa NE, Fernández-Chirino L, Vargas-Vázquez A, Fermín-Martínez CA, Aguilar-Salinas CA, Bello-Chavolla OY (2022) Prevalence trends of diabetes subgroups in the United States: a data-driven analysis spanning three decades from NHANES (1988–2018). J Clin Endocrinol Metab 107:735–742. 10.1210/clinem/dgab762 [ DOI ] [ PubMed ] [ Google Scholar ] 47. Lu B, Li P, Crouse AB et al. (2025) Data-driven cluster analysis reveals increased risk for severe insulin-deficient diabetes in Black/African Americans. J Clin Endocrinol Metab 110(2):387–395. 10.1210/clinem/dgae516 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 48. Khoury MJ, Iademarco MF, Riley WT (2016) Precision public health for the era of precision medicine. Am J Prev Med 50:398–401. 10.1016/j.amepre.2015.08.031 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 49. Roberts MC, Holt KE, Del Fiol G, Baccarelli AA, Allen CG (2024) Precision public health in the era of genomics and big data. Nat Med 30:1865–1873. 10.1038/s41591-024-03098-0 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary File NIHMS2173820-supplement-Supplementary_File.pdf (1.7MB, pdf) Data Availability Statement Data for the NIH-funded cohort studies are available to registered users on request from the NIDDK Central Repository (NIDDK-CR; https://repository.niddk.nih.gov ) and NHLBI BioLINCC ( https://biolincc.nhlbi.nih.gov/home/ ). Epic Cosmos access is available through institutional representatives of participating institutions and the Epic Cosmos team after completing certification requirements. The code for reproducing the subtypes is available on https://github.com/chroniq-lab/diabetes_subtypes_cohorts . ACTIONS View on publisher site PDF (824.4 KB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• From narratives to numbers and back: Assessing the psychosocial aspects of diabetes in the era of high technology with emerging qualitative and quantitative methodologies.#1097910
• Development and Exploration of the Effectiveness and Feasbility of a Digital Intervention for Type 2 Diabetes Mellitus#683029
• Digital Health Interventions for Diabetes Distress in Type 2 Diabetes: Protocol for a Scoping Literature Review Focused on Equity and Inclusion.#140458
• An Analytical Study on the Prevalence of Type-2 Diabetes among American Adults#11954
• An Analytical Study on the Prevalence of Type-2 Diabetes among American Adults#11953