Background: Abdominal obesity is a major driver of metabolic dysfunction, yet conventional screening methods fail to capture dynamic glycemic abnormalities. CGM-based phenotyping in non-diabetic, centrally obese Southeast Asian populations remains unexplored.
Objective: To identify glucose phenotypes (glucotypes) using machine learning analysis of Continuous Glucose Monitoring (CGM) data in Thai adults with central obesity and to examine cross-dataset reproducibility.
Methods: Two independent cohorts (discovery n = 104; validation n = 148) of participants enrolled in a lifestyle modification program underwent 14-day CGM monitoring. K-means clustering was performed on four standardized metrics: time above range (TAR%), time below range (TBR%), standard deviation of daily averages, and mean daily range. Prediabetes was defined according to ADA 2024 criteria (HbA1c 5.7 - 6.4% and/or fasting blood sugar 100 - 125 mg/dL). A sensitivity analysis using TAR% and TBR% only (n = 193) was conducted. Cluster stability was assessed via bootstrap resampling (1,000 iterations) and hierarchical clustering cross-validation.
Results: Three reproducible glucotypes were identified across both cohorts: (1) Postprandial Spikers (17 -26%), characterized by elevated TAR (12.7 - 14.2%), highest HbA1c (5.7 - 5.8%, p < 0.01), and prediabetes prevalence of 35 - 36% (p = 0.010); (2) Stable Glucose (62 - 76%), well-controlled across all metrics; and (3) Hypoglycemia-Prone (7 - 14%), with elevated TBR (9.1 - 13.2%), extreme variability, and zero prediabetes. The Hypoglycemia-Prone phenotype exhibited a glycemic pattern consistent with reactive hypoglycemia, although definitive mechanistic classification requires further investigation. Linear Discriminant Analysis (LDA) achieved 92.6% cross-validated accuracy. The relative risk of prediabetes in Spikers versus Stable was 2.52 (NNS = 4.6). Sensitivity analysis (n = 193, 2-feature) confirmed the identical glucotype structure (Silhouette = 0.544, ARI = 0.791).
Conclusion: Three clinically meaningful glucotypes were identified and replicated across independent cohorts and analytical approaches, confirming CGM-based phenotyping as a robust tool for precision metabolic screening in centrally obese populations. The consistent identification of hidden prediabetes risk and a phenotype suggestive of reactive hypoglycemia supports CGM deployment in diabetes prevention programs across Southeast Asia. Future studies incorporating insulin measurements and meal timing data are needed to confirm the mechanistic basis of the Hypoglycemia-Prone phenotype.