Post-meal, or postprandial glucose, responses vary considerably between meals and individuals, complicating efforts to define stable personal glycemic traits in real-world settings. This study aimed to determine whether short-term ambulatory monitoring of free-living meals can identify reproducible, participant-specific glucose-response characteristics that generalize to standardized conditions.
The analysis included 50,463 free-living meals collected from 992 adults without diagnosed diabetes. A separate set of 4,524 standardized meals was excluded from training and used as a held-out test set to evaluate model transfer. The primary dataset derives from the Food & You study infrastructure; the authors report that the raw Food & You data will be publicly released and linked from their lab site.
To disentangle shared population patterns from participant-specific responses, the investigators applied a multivariate mixed-effects machine-learning framework that jointly modeled four glucose-response outcomes. This approach allowed separation of within-individual variability (meal-to-meal fluctuation for the same person) from between-individual structure (systematic differences across participants). The framework also enabled calculation of participant-level scores along learned axes of between-individual variation, which were then tested for reproducibility and predictive value on held-out standardized meals.
Across the free-living meals, within-individual variability was the dominant source of response variation, indicating meaningful noise in ambulatory monitoring. However, the between-individual variation was not diffusely distributed: it concentrated on two interpretable dimensions.
The first, termed the dominant axis, represented an overall propensity toward larger postprandial glucose excursions across outcome measures.
The second axis captured carbohydrate responsiveness, characterized by relatively higher early glucose rises but lower late postprandial elevations.
The concentration of between-individual differences into these two axes implies that, despite noisy free-living data, individual differences in glucose handling are structured and low-dimensional.
Dominant-axis scores demonstrated reproducibility across non-overlapping subsets of meals within participants, indicating these scores capture stable personal traits rather than idiosyncratic meal noise. Crucially, dominant-axis scores were associated with responses to the held-out standardized meals even after adjusting for age, sex, body mass index, and simple glucose summary statistics such as mean and variability. This indicates that the learned axis captures participant-specific information beyond standard covariates and summary metrics.
The authors assessed monitoring duration and report that dominant-axis scores estimated from 3 days of monitoring approximated those estimated from 14 days. This finding supports the feasibility of short-term ambulatory monitoring to recover individual glucose-response traits that are meaningful and transferable to standardized meal responses.
The results support the use of short-term, real-world glucose monitoring to quantify individual postprandial response tendencies. A reproducible dominant axis that indexes overall excursion tendency may serve as a candidate coordinate for participant stratification in research or individualized intervention strategies. Because the axis contributed predictive value beyond age, sex, BMI, and basic glucose summaries, it may capture physiology or behavioral patterns not reflected in typical clinical covariates.
The report emphasizes that free-living monitoring contains substantial within-individual variability, which complicates signal extraction but does not preclude recovery of low-dimensional participant traits. Specific limitations beyond variability (for example, sample representativeness, device adherence, or meal annotation quality) were not detailed in the source text.
Analysis code and results have been made available on GitHub (https://github.com/digitalepidemiologylab/postprandial-glucose-traits-analyses) and archived on Zenodo (DOI:10.5281/zenodo.22677786). The raw Food & You data will be released publicly and a link will be posted at digitalepidemiologylab.org according to the authors.
Ethical approval was provided by the Geneva ethics commission (ethical approval Number: 2017-02124). The authors state that participant consent and relevant reporting guidelines were followed. The study is registered under Clinical Trial NCT03848299. The authors declared no competing interests.
In this large ambulatory dataset, a multivariate mixed-effects machine-learning approach separated noisy within-individual postprandial variability from stable between-individual structure. Between-individual differences were concentrated primarily on two axes: an overall propensity for larger postprandial glucose excursions and a distinct carbohydrate-responsiveness axis with differing late glucose patterns. Dominant-axis scores were reproducible, transferred to held-out standardized meals beyond common covariates and summary measures, and could be approximated from as little as 3 days of monitoring. These findings support short-term monitoring as a practical method to quantify individual glucose-response traits and propose a candidate coordinate for stratification in research and potentially personalized care.
Competing interests: none declared by the authors.
Data and code access: GitHub and Zenodo links provided by the authors in the source article.
Posted: September 15, 2026.