Short-term postprandial glucose monitoring reveals stable traits from noisy free-living meals

Technical Summary: Short-term postprandial glucose monitoring reveals stable traits from noisy free-living meals

Problem Statement
Postprandial glucose responses (PPGR) exhibit significant heterogeneity, both between individuals and within the same individual across different meals. While large glucose excursions are recognized as early markers of impaired control, the "noise" introduced by free-living conditions—such as varying meal compositions, physical activity, and diurnal physiology—obscures the identification of stable, individual-specific metabolic traits. Current precision nutrition frameworks often focus on predicting single meal responses using population-level covariates (e.g., microbiome, anthropometrics) or identifying phenotypes from standardized challenges. However, these approaches struggle to separate context-dependent variations from persistent physiological differences in routine, uncontrolled settings. The core challenge is determining whether reproducible individual glucose response traits can be isolated from the high within-individual variability inherent in free-living data without relying on standardized meal challenges.

Methodology
The study analyzed data from the "Food & You" cohort, comprising 992 adults without diagnosed diabetes. The dataset included 50,463 free-living meals and 4,524 standardized meal challenges (glucose drink, white bread, white bread with butter). Continuous glucose monitoring (CGM) data was recorded every 15 minutes, and dietary intake was logged via a smartphone application.

• Outcome Measures: Four complementary PPGR outcomes were derived from the 2-hour postprandial window: Incremental Area Under the Curve (iAUC), Maximum Glucose (MaxGlu), Peak Duration (time above baseline), and Glucose at 120 minutes (Glu120).
• Modeling Framework: The authors employed a Multivariate Mixed-Effects Machine-Learning (MMER) framework.
  • Fixed Effects: An XGBoost model estimated the non-linear, population-shared response surface based on meal features (nutrient content, recent intake, pre-meal glucose, and glycemic trends). Demographic and anthropometric data were excluded from the model training to ensure the model relied solely on meal-onset information.
  • Random Effects: Participant-specific deviations were modeled as random intercepts (overall glycemic elevation) and random slopes (macronutrient responsiveness) for each of the four outcomes.
• Dimensionality Reduction: The covariance matrix of the random effects was eigendecomposed to identify the principal axes of between-individual variation ($PC_{Full}$).
• Validation Strategy:
  • Cross-Validation: A temporally blocked, participant-wise 5-fold cross-validation was used to prevent data leakage, with a 6-hour embargo around test blocks.
  • Transfer Learning: Models trained exclusively on free-living meals were tested on the held-out standardized meal challenges to evaluate the generalizability of the estimated traits.
  • Reliability: Split-half analyses (odd/even meals and temporal splits) and monitoring duration analysis (1 to 14 days) were conducted to assess the stability and convergence of the estimated traits.

Key Results

1. Variance Decomposition: Within-individual variability dominated free-living responses, accounting for 82% to 99% of the total variance across outcomes. However, between-individual differences were systematic and became more apparent under standardized conditions (Intraclass Correlation Coefficients increased from ~33% in free-living to ~60-70% in standardized challenges).
2. Predictive Performance: The MMER-XGBoost model (incorporating both random intercepts and slopes) achieved the highest predictive performance.
   • On free-living meals, it reached a multivariate $R^2$ of 0.51, outperforming non-individualized XGBoost ($R^2$ = 0.45) and linear models.
   • Crucially, on held-out standardized meals, the MMER-XGBoost model achieved a multivariate $R^2$ of 0.31, significantly outperforming the non-individualized XGBoost ($R^2$ = 0.18). This demonstrates that traits learned from noisy free-living data successfully transfer to controlled challenges.
3. Low-Dimensional Trait Space: Between-individual variation was concentrated in a low-dimensional space.
   • $PC_{Full}1$ (69.3% of variance): Represented an overall tendency toward larger or smaller glucose excursions (positive loadings on all intercepts).
   • $PC_{Full}2$ (11.8% of variance): Reflected carbohydrate responsiveness combined with lower late glucose elevations (positive carbohydrate slopes, negative late-time intercepts).
   • The effective dimensionality of the trait space was estimated at approximately 2.0.
4. Stability and Convergence:
   • $PC_{Full}1$ scores were highly reproducible across non-overlapping meal subsets and converged rapidly. Estimates based on just 3 days of monitoring achieved an ICC of ~0.75 with the 14-day reference, approximating the full-duration estimate.
   • $PC_{Full}2$ scores were less stable, requiring longer monitoring (8–11 days) to reach similar agreement levels.
   • The dominant axis ($PC_{Full}1$) remained robust even when standardized meals were excluded from training, confirming its utility as a generalizable phenotype.

Significance and Claims
The paper claims that short-term continuous glucose monitoring is sufficient to measure individual glucose response traits, shifting the focus from single-meal prediction to metabolic phenotyping. The authors demonstrate that despite high noise in free-living data, a low-dimensional structure of individuality exists and can be recovered using mixed-effects machine learning.

The primary contribution is the identification of a candidate coordinate ($PC_{Full}1$) for stratification that represents an individual's overall glycemic elevation tendency. This trait is reproducible from as little as three days of monitoring and generalizes to standardized challenges. The study suggests that precision nutrition can move beyond context-specific predictions to account for persistent physiological differences between individuals, provided that modeling approaches explicitly separate population-shared response surfaces from participant-specific deviations. The findings support the feasibility of using routine self-monitoring to identify metabolic phenotypes without the need for invasive or controlled clinical challenges.