Continuous Glucose Monitoring Reduces Glucose Dynamics to 3 Features

Key Takeaways

• Three CGM-derived features captured most interindividual variation in glucose dynamics among adults without diagnosed diabetes.
• The three-feature representation reproduced standardized postprandial glucose trajectories with high accuracy and also generalized to external validation datasets, with lower accuracy in real-world meal analyses.
• The features were associated with carotid intima-media thickness and liver ultrasound measures, and clustering based on mean, standard deviation, and AC_Var yielded six groups with lower confidence bounds in high-value clusters.

A primary research Communications Medicine analysis found that at least seven days of continuous glucose monitoring from 8025 adults without diagnosed diabetes could be summarized by three features: mean, variance, and autocorrelation. Exploratory factor analysis identified those axes, which accounted for 82% of total between-person variation in glucose dynamics. The investigators used this as a compact, non-diagnostic summary of routine CGM patterns in adults without diagnosed diabetes.

The analysis drew on five observational cohorts, with feature discovery performed in the 10 K Project and postprandial reconstruction tested in 863 PREDICT1 participants. In the 10 K Project, participants were aged 40 to 70 years, and all included CGM recordings lasted at least seven days. Exploratory factor analysis identified mean, variance, and autocorrelation as the core axes, and principal component analysis supported the same structure with 85% of variance explained. The authors used these three axes as the minimal representation of glucose dynamics in people without diagnosed diabetes.

In PREDICT1 meal-test modeling, a two-dimensional encoding reconstructed standardized postprandial glucose trajectories with R=0.88, while the three-dimensional model reached R=0.92. In the San Francisco Bay Area Study1, the three-feature model achieved R=0.91 during oral glucose tolerance testing. In real-world 10 K Project meal analyses, the three-dimensional model reached R=0.506 at the 20 g carbohydrate threshold, with similar values at 40 g and 60 g. Across the datasets tested, the compact feature set preserved much of the shape information in postprandial trajectories.

Among 1784 non-diabetic adults with simultaneous ultrasound and CGM assessments, the same features were significantly associated with carotid intima-media thickness and liver measures of attenuation, elasticity, and dispersion. After adjustment for blood pressure, lipid profiles, mean, and standard deviation, AC_Var remained associated with IMT at β=0.011 and hepatic attenuation at β=0.010. Mean and AC_Var were also independently associated with IMT two years after CGM measurement. In the Kobe University analysis, the model using mean, standard deviation, and AC_Var showed superior predictive performance for disposition index compared with fasting blood glucose, HbA1c, and 120-minute plasma glucose. These associations were modest in a largely healthy population.

K-means clustering based on mean, standard deviation, and AC_Var identified six groups across the study population. In high-value clusters, the lower 95% confidence bounds were 111 mg/dL for mean, 17 mg/dL for standard deviation, and 0.075 for AC_Var. The investigators presented these values as feature-based reference thresholds within the study population rather than diagnostic cutoffs.

The analysis was observational and model-based, and validation was limited to the datasets and CGM windows described. The three-feature framework organized glucose-dynamics variation, external reconstruction, and organ-health associations within the populations studied.