Abstract
BACKGROUND: Early detection of metabolic dysfunction before diabetes onset remains a critical challenge in preventive medicine. Although glucose dynamics provide high-dimensional insights into metabolic states, it remains unclear which combination of glucose dynamics-derived measures most effectively summarises interindividual differences in glucose regulation. METHODS: We analysed continuous glucose monitoring (CGM) data from 8025 adults without diagnosed diabetes, using all recordings of at least seven days' duration, to derive a low-dimensional representation of glucose dynamics. Using exploratory factor analysis, we identified a small set of latent features that accounted for variation between individuals in CGM-derived metrics. We then trained and validated machine-learning models to predict postprandial glucose trajectories from these features in 863 participants who underwent standardised meal tests. We further assessed the generalisability of this representation in an independent oral glucose tolerance test dataset. Finally, we examined associations between the derived features and markers of vascular and liver health in 1784 non-diabetic adults. RESULTS: Three features, "mean", "variance", and "autocorrelation", explain more than 80% of the interindividual differences in CGM-derived measures. A three-dimensional representation based on these features reconstructs postprandial glucose trajectories with high accuracy and outperforms fasting, mean, and two-hour postprandial glucose values. Each feature shows independent associations with carotid artery intima-media thickness and with hepatic steatosis and stiffness. CONCLUSIONS: By compressing high-dimensional glucose dynamics into three interpretable features with minimal loss of information, this framework provides a simple yet physiologically meaningful representation of glucose regulation that may facilitate a more precise and interpretable assessment of diabetes-related risk.