Data-driven subclassification of type 2 diabetes in Mexico: a conceptual framework to improve diabetes care

Summary

Type 2 diabetes (T2D) is a heterogeneous disease and a major public health concern in low- and middle-income countries (LMICs). To address this heterogeneity, several subgroup classification frameworks have been proposed. Among them, the most widely used is the data-driven classification proposed by Ahlqvist et al., which includes severe autoimmune diabetes (SAID), severe insulin-deficient diabetes (SIDD), severe insulin-resistant diabetes (SIRD), mild obesity-related diabetes (MOD), and mild age-related diabetes (MARD). However, reproducing this framework is limited by the need for specialized measurements of beta-cell function and insulin resistance, which are not routinely available in most clinical or epidemiological settings in LMICs. To overcome these barriers, we derived a classification algorithm to reproduce these diabetes subgroups. Here, we present a four-step roadmap in which this framework may help characterize T2D heterogeneity across Mexico, potentially guide pathophysiology-oriented treatment strategies, support public health monitoring, and enhance our understanding of T2D heterogeneity to improve diabetes care in our country.

Type 2 diabetes (T2D) is the most prevalent form of diabetes, accounting for over 90% of all cases globally. According to the International Diabetes Federation (IDF), in 2025 T2D ranked as the eighth leading cause of mortality worldwide, affecting approximately 589 million adults, of whom 252 million remain undiagnosed.¹ Moreover, the IDF has estimated that four out of five adults living with T2D reside in low- and middle-income countries (LMICs).¹ In Mexico, T2D is the second leading cause of death and had an estimated prevalence of 18.3% in 2022.² Although the national prevalence of diabetes-related complications has not been reported, a study conducted by Ovalle-Luna et al. using a large sample of patients treated at the Mexican Social Security Institute (≈297,000 patients) showed that 34.9% had at least one chronic complication, with neuropathy (13.9%), retinopathy (12.8%), and kidney disease (12.3%) being the most frequent complications.³ In a simulation study conducted by Reynoso-Noverón et al., the authors applied the UKPDS outcome model to ENSANUT 2006 data and estimated that over 20 years, per 1000 people with diabetes, there would be 112 new cases of ischemic heart disease, 260 myocardial infarctions, 113 cases of heart failure, 101 strokes, 62 amputations, and 539 premature diabetes-related deaths. Furthermore, the authors predicted that 54% of patients living with diabetes would die within 20 years, with an average remaining life expectancy of 10.9 years.⁴ Therefore, T2D represents a substantial and growing public health challenge in Mexico. However, a key barrier to improving outcomes is that T2D is often treated as a single homogeneous disease, while some individuals may differ substantially in underlying pathophysiology, treatment response, and risk of complications. One potential explanation for why T2D carries a high burden in countries such as Mexico lies in the inherent heterogeneity of this condition. It is now well established that T2D is a polygenic and heterogeneous condition which stems from the interplay of environmental factors, lifestyle, ethnicity, and genetic susceptibility, further modulated by aging-related physiological decline.⁵ This interplay leads to immunological and metabolic heterogeneity across diverse pathophysiological pathways including inflammation, whole-body and tissue-specific insulin response and insulin resistance, and adipose tissue distribution, all of which may influence treatment response and risk of long-term complications.⁵ However, current clinical classifications do not fully capture this heterogeneity, contributing to delayed diagnosis, suboptimal glycemic control, and persistent complication burden.⁶

To address this gap, several subgroup classification frameworks have been proposed to stratify T2D into clinically meaningful phenotypes. Among the most widely studied is the data-driven classification introduced by Ahlqvist et al. in 2018.⁷ In their initial publication, they characterized five distinct data-driven subgroups based on six variables including hemoglobin A1c (HbA1c), glutamic acid decarboxylase autoantibodies (GAD65-Ab), age at diagnosis, body-mass index (BMI), and measures of insulin sensitivity (HOMA2-IR) and insulin secretion (HOMA2-B). With these variables, diabetes subgroups were classified as follows: severe autoimmune (SAID), severe insulin-deficient (SIDD), severe insulin-resistant (SIRD), mild obesity-related (MOD), and mild age-related diabetes (MARD).⁷ These subgroups have since been replicated in Germany,⁸ China,⁹ India,¹⁰ Japan,¹¹ Ukraine,¹² and the Netherlands¹³ and linked with trajectories of glycemic control, diabetes-related complications, and therapeutic responses.^14,15 Furthermore, the clinical and economic utility of this framework has been evaluated in prior studies. In a study performed by Li et al., the authors used European descent cohorts and found that applying a diabetes subgroups-based treatment intensification, such as targeting weight and lipid control in addition to glycemia either based on the data-driven subgroups proposed by Ahlqvist or risk-driven subgroups, could deliver up to a 10-fold increase in quality-adjusted life years.¹⁶ In another study by Dennis et al., the authors performed a post-hoc analysis of the ADOPT and RECORD trials, demonstrating that implementing data-driven diabetes subgroup frameworks led to a better characterization of the pharmacological response. In this study, patients classified as SIRD showed better response to thiazolidinediones, while those in the MARD subgroup responded better to sulfonylureas.¹⁴ Finally, in another study conducted by our group, we used data from the Mexico City Prospective Study and applied the data-driven diabetes subgroup framework, finding that individuals classified as SIDD and MARD had a higher risk of CVD mortality. Furthermore, we found that integrating these subgroups with recalibrated clinical scores, such as SCORE2-Diabetes, significantly improved the prediction of fatal CVD.¹⁷ Although various alternative approaches have been proposed to delineate more specific, pathophysiology-oriented subgroups—such as genetic or multi-omics stratification and complex longitudinal phenotyping—these methods often require specialized biomarkers or computational tools not routinely available in clinical practice, show limited reproducibility across populations, and have not yet demonstrated clear improvements in clinical outcomes compared with simpler classification strategies.¹⁸ Hence, most pathophysiology and omics-driven diabetes classifications are difficult to replicate and implement in care settings within LMICs. In contrast, the data-driven subgroup classification proposed by Ahlqvist et al. remains a practical alternative to characterize diabetes heterogeneity, improve prediction of clinical outcomes, and potentially implement a transition towards a precision medicine approach, particularly in health systems such as Mexico's, where economic resources and personnel are constrained.

One limitation of the Ahlqvist et al. framework is its reliance on biochemical variables such as GAD65-Ab and on C-peptide to estimate beta-cell function and insulin resistance, which are not routinely available in most clinical settings or in many epidemiological studies conducted in LMICs. This limitation challenges the implementation of the Ahlqvist et al. approach in clinical practice and its potential translation to primary care evidence.¹⁹ To overcome these barriers, Bello-Chavolla et al. adapted the Ahlqvist et al. classification and developed four algorithms based on self-normalizing neural networks (SNNN), using NHANES as the discovery and validation cohort.²⁰ In this work, our group first replicated the five diabetes subgroups defined using the original clustering variables proposed by Ahlqvist et al. (i.e., GAD65-Ab, age at diagnosis, BMI, HbA1c, HOMA2-IR, and HOMA2-B) in NHANES-III. Once these subgroups were identified, they developed four SNNN algorithms to estimate the probability of belonging to each subgroup based on specific variables. These included: a model with HOMA2 estimations using C-peptide (C-peptide model), another with HOMA2 estimations using insulin (insulin model with HbA1c), a third model that does not require HbA1c but uses HOMA2 estimations (insulin model without HbA1c), and a fourth model using only fasting glucose, triglycerides, HDL-C, waist circumference, BMI, and age at diabetes diagnosis (non-insulin model). These algorithms provide an alternative approach in terms of the required variables for estimating data-driven diabetes subgroup classifications in settings where insulin or C-peptide measurements required for HOMA-based estimation are unavailable.^21,22 To assess the clinical feasibility of this framework in Mexico, we explored the cost of implementing these required measurements in ambulatory laboratory settings. The estimated costs to measure the required variables for the fourth model (non-insulin model) were on average ∼120 USD. Given the balanced classification performance and the significantly lower implementation costs compared to insulin or C-peptide testing, these algorithms represent a pragmatic, feasible, and potentially cost-effective approach for implementing diabetes subgroup stratification in primary care and other resource-limited settings. However, these algorithms do not incorporate the SAID subgroup due to the limited availability of anti-GAD65-Ab testing in most primary care settings, making reliable identification of this subgroup impractical in routine clinical contexts. These algorithms were subsequently implemented into a web-based application with potential implementation for clinical practice and clinical research (https://uiem.shinyapps.io/diabetes_clusters_app/). In the same study, our group validated and applied these algorithms to different Mexican cohorts, demonstrating transitions between groups over time and clinical and genetic factors associated with classification into each subgroup. Specifically, in a cohort of newly diagnosed patients with diabetes receiving multidisciplinary interventions, only 10.7% remained in the SIDD subgroup after three months of multidisciplinary management, with most reclassified to MOD or MARD, whereas MARD, SIRD and MOD subgroups were comparatively more stable despite the multidisciplinary intervention. Importantly, patients who persisted as SIDD had higher risks of retinopathy and kidney disease.²⁰ These findings highlight the heterogeneous and dynamic nature of T2D, supporting the notion that diabetes subgroups could be repeatedly implemented to characterize diabetes heterogeneity at a population level, guide treatment allocation over time and inform precision-based therapeutic strategies.

From an epidemiological perspective, applying the diabetes subgroup classification framework has helped to partially explain diabetes trends and identify vulnerable groups. In two studies performed by our group, we applied the data-driven diabetes subgroups framework to national health surveys in two countries. In the United States, we found that the weighted prevalence of T2D increased from 7.5% to 13.9% (1988–2018), and this increase was primarily driven by a growth in milder diabetes phenotypes, specifically MOD and MARD. MOD increased from 2.6% to a final prevalence of 6.4%, while MARD increased from 1.2% to 3.9%. Conversely, the SIDD and SIRD subgroups exhibited more stable trajectories, with prevalences between 2.0–2.9% and 1.5–2.1%, respectively. However, in the study performed in Mexico, we found that our country experienced an even more pronounced increase in T2D prevalence (2016–2022) from 13.4% to 18.2%. In contrast to what we observed in the US, the increase in diabetes prevalence in Mexico has been primarily driven by severe phenotypes and specifically by SIDD, which remains the most prevalent subgroup (6.62%), followed by MOD (5.25%).^23,24 Notably, SIDD was disproportionately concentrated in southern regions, which are characterized by the highest rate of sociodemographic inequalities in Mexico. States with greater social lag had higher rates of the SIDD subgroup and lower rates of the MOD subgroup. These findings suggest that diabetes subgrouping could help identify vulnerable groups who may require earlier, more intensive, or tailored interventions.

The evidence derived from Mexico and other countries may support the feasibility for implementation studies, such as randomized clinical trials. While cluster-based diabetes classifications are not yet incorporated into clinical guidelines, these findings support the development of a pragmatic roadmap for potential implementation. Here, we propose a conceptual four-step approach through which data-driven diabetes subgroups could be implemented (Fig. 1). Step 1 would be classification implementation using the algorithm developed by Bello-Chavolla et al., systematically measuring the variables required for the non-insulin-based algorithm, including biochemical (fasting glucose, HDL-C, triglycerides), anthropometric (BMI, waist circumference, height), and clinical variables (age at diabetes diagnosis). Ideally, to improve the performance of the algorithm, HbA1c and insulin could be added to estimate HOMA2-B and HOMA2-IR. At this stage, subgroup classification could support early risk stratification, rather than direct treatment decisions. From a public health perspective, diabetes subgroup mapping could help identify regions with high-risk subgroups and severe phenotypes (such as SIDD and SIRD) and support stratified resource planning, community interventions, and surveillance. Furthermore, studies which could help inform their applicability include prospective studies in patients with long-standing diabetes to assess the prognostic implications of the diabetes subgroup framework to predict diabetes outcomes, as well as emulated target trials to evaluate the potential effect of pharmacological interventions across subgroups. Step 2 would involve a personalized treatment plan, in which therapies and monitoring could be aligned with subgroup profiles. Evidence from the ADOPT trial suggests that prescription of pharmacological interventions based on diabetes subgroups may be beneficial.¹⁴ For example, patients classified as SIRD demonstrated a better glycemic response to thiazolidinediones, while MARD to sulfonylureas. Conversely, SIDD was characterized by a faster increase in glycemia and an increased risk of diabetes-related complications, which suggest that this subgroup could benefit from earlier and more intensive glucose management, potentially including insulin, SGLT2 inhibitors, or DPP-4 inhibitor initiation. Regarding lifestyle interventions, it is important to note that while the Look AHEAD trial did not originally used the Ahlqvist et al. clustering framework, a subsequent re-analysis characterized four distinct subgroups based on the cohort's composition.²⁵ This evidence highlights that intensive lifestyle interventions (ILI) for weight loss should be implemented with caution: while ILI was associated with reduced CVD risk in several subgroups, it was also associated with an increased risk of CVD events specifically among individuals with poor glucose control (HbA1c >9%), a profile characterized by severe hyperglycemia similar to the SIDD cluster. Consequently, for patients presenting with this clinical profile, achieving metabolic stability may be necessary before initiating intensive weight-reduction programs to optimize outcomes and avoid potential harm. Furthermore, non-pharmacological strategies should be tailored to the predominant phenotype. For example, individuals in the MOD subgroup may specifically benefit from hypo-caloric diets for weight loss, whereas management for the MARD subgroup should prioritize protein-balanced nutrition and resistance training to mitigate age-related sarcopenia. Recognizing autoimmune diabetes also remains important for clinical decision-making. In resource-limited settings where GAD65-Ab testing is often unavailable, pragmatic indicators such as early insulin requirement, poor response to oral agents, lower BMI, ketosis at presentation, or a personal or family history of autoimmune disease may help clinicians identify individuals who warrant early insulin therapy or referral for confirmatory testing. Nevertheless, at this stage, there is a clear gap in pragmatic and randomized clinical trials, as well as implementation studies, to evaluate the effectiveness of individualized therapies based on diabetes subgroups. Step 3 would focus on measuring clinical and economic benefit, based on the proportion of eligible patient subgroups classified and the time to treatment decision. These benefits would require the evaluation of these subgroups registered using information from electronic-health records (EHRs), diabetes surveillance systems and national health surveys. At the individual level, treatment goals would prioritize HbA1c control, reduction of therapeutic inertia, and early complication rates. At the health-care system level, key indicators would include hospitalizations, emergency visits, and medication expenditure patterns. From an epidemiological perspective, these indicators would support evaluation of subgroup-informed care and guide resource allocation to areas that could benefit most from community interventions, novel therapies, and multidisciplinary programs. These approaches can inform tailored treatment strategies, improve metabolic control, and reduce complications by aligning therapy with each patient's profile. In this regard, some studies that could be implemented to evaluate the impact at this stage include cost-effectiveness studies to assess the benefits of targeted versus universal interventions, as well as ecological studies to evaluate the prevalence of diabetes subgroups and the rates of subgroup-related complications in national health surveys. Finally, Step 4 would consist of nationwide integration and policy translation, requiring the adoption of standardized datasets that allow the integration of diabetes subgroup classification into interoperable EHRs and clinical registries. This would be accompanied by standardized diabetes subgroup care pathways and training programs to support individualized management at the primary care level, ultimately leading to incorporation of diabetes subgroups into institutional clinical guidelines and protocols. However, extensive prospective validation, pragmatic clinical trials, and implementation studies are still needed to evaluate the prognostic value and the pharmacological and clinical benefits of adopting the data-driven subgroup approach in the Mexican context. Ultimately, this framework aims to standardize diabetes subgroup classification and treatment pathways across institutions and to enable routine monitoring of quality indicators, coverage of subgroup classification, time to treatment adjustment, HbA1c control, and subgroup-specific complication rates, at clinic, regional, and national levels.

Fig. 1.

Conceptual four-step roadmap for implementing data-driven diabetes subgroup classification across clinical care and public health in Mexico. Abbreviations: BMI, body mass index; HbA1c, glycated hemoglobin; METS-IR, metabolic score for insulin resistance; METS-VF, metabolic score for visceral fat; SIDD, severe insulin-deficient diabetes; SIRD, severe insulin-resistant diabetes; MOD, mild obesity-related diabetes; MARD, mild age-related diabetes; EHR, electronic health record.

However, important gaps remain to potentially translate these findings into clinical and public health settings. First, there is additional need for longitudinal validation studies to further characterize subgroup stability and the prognostic value of subgroup transitions in predicting clinical outcomes. Second, there is a need for randomized clinical trials or implementation studies that demonstrate improvements in control rates, complications, and patient-reported outcomes when using diabetes subgroup classifications to guide treatment decisions. However, caution is warranted, as these subgroups may vary over time and could reflect short-term transitions rather than stable pathophysiological states. In this context, subgroup classification could be adopted as a complementary approach to existing guideline-based treatment algorithms. Accordingly, the American Diabetes Association acknowledges T2D as a highly heterogeneous condition and emphasizes a person-centered management strategy that prioritizes treatment selection based on individual clinical characteristics and the presence of comorbidities, rather than a uniform approach.²⁶ Similarly, the 2026 American Association of Clinical Endocrinology Consensus Statement introduced a new “Diabetes Classification Algorithm” that reinforces this shift toward individualized diagnosis, urging clinicians to systematically confirm a type 2 diabetes phenotype and consider alternative etiologies in order to align management with the underlying pathophysiology.²⁷ Third, although the algorithm does not include the SAID subgroup due to the limited availability of autoimmune antibody testing in most primary care settings, recognizing adult-onset autoimmune diabetes remains essential. Fourth, there is a need for cost-effectiveness analyses focusing on evaluating targeted interventions based on personalized subgroup classification versus universal interventions. Fifth, although we have generated epidemiological evidence on diabetes subgroups in the United States and Mexico, their routine estimation using nationally representative surveys has not yet been implemented. Therefore, there is a need for public-health actions to incorporate the proposed algorithm to systematically report and monitor the prevalence diabetes subgroups at the national level. Sixth, most of the evidence derived from this data-driven diabetes subgroup framework arises from cohorts with relatively recent-onset diabetes, in which pathophysiological profiles may differ from those observed in individuals with longer-standing disease. Therefore, future studies should include and compare both patient profiles to clarify whether time since diabetes diagnosis may influence treatment response and, potentially, diabetes-related outcomes. Finally, the potential implementation of this evidence into clinical guidelines will depend on transparent, workflow-friendly tools embedded in EHRs, which may be challenging given that Mexican health care systems are fragmented, consequently limiting follow-up in secondary and tertiary care systems.²⁸ Ultimately, addressing these gaps will be essential to translate diabetes subgroups into actionable, equitable tools for improving outcomes in diverse and socially disadvantaged populations in the country.

In summary, we propose that the implementation of a data-driven diabetes subgroup classification framework could help characterize T2D heterogeneity across Mexico. Proof-of-concept for individualized therapy already exists in monogenic diabetes, where matching treatment to the genetic diagnosis, such as prescribing sulfonylureas in HNF1A-MODY, yields superior outcomes compared with standard care.²⁹ Using data-driven diabetes subgroups as an individualization framework may provide clearer pathophysiology-oriented treatment strategies, and therefore, be a promising approach to improve diabetes care, monitoring of public health interventions, and allow for better understanding of the underlying heterogeneity in T2D in Mexico, other LMICs and beyond.

Contributors

Each author contributed important intellectual content during manuscript drafting or revision and accepted accountability for the overall work by ensuring that questions pertaining to the accuracy or integrity of any portion of the work are appropriately investigated and resolved.

Declaration of interests

Nothing to disclose.