One of the largest real-world studies of its kind has used an unsupervised machine learning algorithm to sort nine thousand people with type 2 diabetes into three clinically distinct groups, and the results reveal striking differences in how doctors prescribe the modern generation of heart- and kidney-protecting glucose-lowering drugs. The analysis, drawn from a registry spanning 61 medical centers in China between 2023 and 2024, suggests that the choice between two celebrated drug classes is far from uniform: it shifts systematically depending on the patient’s age, body weight, metabolic profile, and history of cardiovascular disease.
The drugs in question are sodium–glucose cotransporter 2 inhibitors, known as SGLT2 inhibitors, and glucagon-like peptide-1 receptor agonists, or GLP-1 receptor agonists. Together they are described in the study as cardiorenal protective treatments because large outcome trials have shown that, beyond lowering blood glucose, they reduce the risk of cardiovascular events and slow the progression of kidney disease in people with type 2 diabetes. International and Chinese guidelines now recommend their use for patients at elevated cardiovascular or renal risk, making them a cornerstone of modern diabetes management rather than optional add-ons.
Yet a persistent question in clinical practice has been who actually receives these treatments, and whether prescribing reflects the specific risk profile of each patient. Generic averages can obscure the answer. A clinic population may show an overall utilization rate that looks reasonable while concealing large pockets of under-treatment among the very patients who stand to benefit most. This is precisely the kind of problem that phenotype clustering, a technique that groups patients by multidimensional similarity rather than by any single diagnostic label, is designed to address.
The research team, led by investigators at Peking University People’s Hospital and collaborating institutions across China, extracted baseline data from the iCaReMe China observational registry cohort, a multicenter effort to assess the management and quality of care for patients with type 2 diabetes, hypertension, heart failure, and chronic kidney disease. From this dataset of 9000 patients, they applied the Partitioning Around Medoids algorithm, a robust clustering method that builds groups around actual patient cases, called medoids, rather than around mathematical averages that can be distorted by outliers. The technique has gained traction in diabetes research in recent years, following influential work that identified novel subgroups of adult-onset diabetes in European, and subsequently Chinese and American, populations.
The resulting three clusters were clinically coherent and immediately recognizable to any practicing endocrinologist. Cluster 1 consisted of older patients carrying a high risk of atherosclerotic cardiovascular disease, the buildup of cholesterol-laden plaques in arteries that underlies heart attacks and strokes. Cluster 2 captured metabolically stable patients without cardiorenal risk, the comparatively healthy core of the diabetes population. Cluster 3 contained younger patients with obesity, high metabolic risk, and a short duration of diabetes, a profile increasingly common as the global burden of type 2 diabetes shifts toward earlier onset. The enrolled cohort had a median age of 55 years, was 60.8 percent male, and had a median diabetes duration of 5.37 years.
When the researchers compared prescribing patterns across the clusters, the differences were statistically significant. Overall utilization of cardiorenal protective treatments ranged from 44.1 percent in the metabolically stable Cluster 2, to 49.0 percent in the high cardiovascular risk Cluster 1, to 51.1 percent in the young, obese, high-risk Cluster 3, which had the highest rate of all three groups. Even so, the numbers mean that nearly half of patients in every cluster, including those with established risk factors, were not receiving a drug class that guidelines prioritize for exactly their situation.
The composition of treatment also diverged by phenotype. In the metabolically stable Cluster 2, prescribing was dominated by SGLT2 inhibitor monotherapy, which accounted for 73.1 percent of cardiorenal protective treatment in that group. Cluster 3, the younger patients with obesity and high metabolic risk, showed a marked preference for GLP-1 receptor agonist-based regimens, which made up 43.6 percent of use in that cluster. Cluster 1, the older patients at high atherosclerotic risk, fell in between, with 31.5 percent of their cardiorenal protective treatment involving GLP-1 receptor agonists. These divergent patterns plausibly reflect the differing secondary benefits of the two drug classes: GLP-1 receptor agonists are associated with substantial weight reduction, which is particularly relevant for the obese younger phenotype, while SGLT2 inhibitors have established benefits for heart failure and kidney outcomes, relevant to older patients with cardiovascular burden.
The technical significance of the study lies in its multilevel correlation framework, which links patient-level phenotypes to treatment-level decisions within a standardized analytical structure. Rather than treating the diabetes population as a monolith, the framework allows clinicians and health system planners to ask how a specific, well-defined subgroup is being treated and whether that treatment matches the evidence. Partitioning Around Medoids is particularly suited to this task because it works with mixed clinical variables and produces interpretable representative cases for each cluster, making the groups easier to communicate than abstract statistical dimensions.
The findings arrive at a moment of mounting global urgency. The Global Burden of Disease Study 2021 estimated more than 800 million people living with diabetes worldwide, with prevalence projected to continue rising through 2050, and cardiovascular disease remains the leading cause of death in people with type 2 diabetes. Identifying the patients most likely to benefit from cardiorenal protective therapy, and the subgroups in which those drugs are being underused, is therefore a matter of population-scale consequence. The observation that the metabolically stable cluster still received cardiorenal protective agents at a rate of 44 percent, while remaining under-treated overall, illustrates both the diffusion of these therapies into routine practice and the gap that persists.
For clinical practice, the study’s message is that phenotype-aware prescribing could sharpen the use of SGLT2 inhibitors and GLP-1 receptor agonists, ensuring that older patients with atherosclerotic risk and younger patients with obesity-driven metabolic risk each receive the drug class best matched to their profile. The authors note that three clinically distinct phenotypes were reliably identified and that prescribing patterns already differ meaningfully across them, providing a foundation for standardized application of cardiorenal protective treatments. As registries like iCaReMe China continue to accumulate longitudinal data, cluster-based frameworks of this kind may become a routine instrument for auditing and improving the quality of diabetes care, turning the promise of precision medicine into an operational reality for hundreds of millions of patients.
Subject of Research: Phenotype cluster analysis of cardiorenal protective glucose-lowering treatment patterns in type 2 diabetes
Article Title: Cardiorenal Protective Glucose-Lowering Treatment Patterns in Type 2 Diabetes Mellitus Based on a Real-World Study: A Clinical Phenotype Cluster Analysis
Article References: Cardiorenal Protective Glucose-Lowering Treatment Patterns in Type 2 Diabetes Mellitus Based on a Real-World Study: A Clinical Phenotype Cluster Analysis. (n.d.). https://doi.org/10.1007/s12325-026-03826-7
Image Credits: AI Generated
DOI: 10.1007/s12325-026-03826-7
Keywords: type 2 diabetes, SGLT2 inhibitors, GLP-1 receptor agonists, cardiorenal protection, cluster analysis, clinical phenotypes, real-world study, precision medicine, obesity, cardiovascular risk, diabetes guidelines, registry study
News Source: Teresa Odom. (October 12, 2026). Machine Learning Reveals Three Diabetes Patient Types Shaping Heart and Kidney Drug Choices. Scienmag.
