MACHINE LEARNING-GUIDED CLUSTERING FOR PHENOTYPIC SUB CLASSIFICATION OF TYPE 2 DIABETES MELLITUS: A CROSSSECTIONAL STUDY EMPLOYING STATISTICAL DATA AUGMENTATION
Abstract
Background: Type 2 diabetes mellitus (T2DM) is not a single
uniform disorder but a heterogeneous metabolic condition, with wide
variation among affected individuals in clinical presentation and
cardiovascular (CV) hazard. Existing diagnostic frameworks fail to
adequately capture this diversity, which may hinder efforts toward
individualized, mechanism-based treatment strategies. Methods:
Clinical and biochemical data were obtained from 127 participants
(108 patients with T2DM and 19 metabolically healthy controls).
To increase statistical power without compromising biological
realism, a statistically grounded data-augmentation procedure was
applied, expanding the cohort to 600 subjects (500 diabetic, 100
non-diabetic) through Gaussian Mixture Modeling combined with
skew-normal distribution sampling. K-means clustering, principal
component analysis (PCA), and multivariate regression modeling
were subsequently used to delineate phenotypic subgroups and
identify determinants of glycemic status and cardiovascular risk.
Results: Two distinguishable diabetic phenotypes emerged from
cluster analysis: Cluster 1 (Metabolically Compensated n=351,
58.5%), comprising comparatively younger individuals (43.9±6.8
years) with better-controlled HbA1c (6.94±1.35%) and intermediate
cardiovascular prevalence (45.6%); and Cluster 2 (Metabolically
Decompensated, n=249, 41.5%), consisting of older individuals
(59.5±7.1 years) with markedly deranged glycemic control (HbA1c
10.43±1.13%) and a substantially higher cardiovascular burden
(85.1%). Compared with controls, diabetic participants showed
Volume 3, Number 3, 2026 Pages 268-281
Bioscience Insights Volume 3, Number 3, 2026
269 | P a g e
markedly higher HbA1c (9.00±1.79% vs 5.33±0.37%,
p<1×10⁻⁷⁰, Cohen's d=2.24) and random blood glucose
(280.9±79.4 vs 105.6±16.1 mg/dL, p<1×10⁻⁷⁹, Cohen's
d=2.41). A multivariable regression model predicted HbA1c with
outstanding accuracy (R²=0.988), with systolic blood pressure the
dominant contributing variable (β=8.72). A logistic model predicting
cardiovascular disease achieved an area under the curve (AUC) of
0.786. On PCA, the leading two components together accounted for
94.5% of total variance. Conclusions: Application of unsupervised
machine learning to phenotypic stratification uncovers biologically
and clinically meaningful diabetic subgroups that differ substantially
in cardiovascular risk. The decompensated subgroup calls for
intensive, multi-pronged management, whereas the compensated
subgroup may respond well to lifestyle-centered approaches. The tight
linkage between blood pressure and glycemic status points to
overlapping underlying pathophysiology and reinforces the case for a
precision-medicine framework in diabetes care.
Views & Downloads Statistics (Last 12 Months)
Monthly performance and trends of article views and downloads
