Summary

A medRxiv preprint analysing 50,463 free-living meals from 992 adults without diagnosed diabetes found that model-derived glucose-response scores estimated from three days of monitoring approximated scores from 14 days. The analysis also identified two major dimensions of individual meal responses and tested them against 4,524 held-out standardized meals.

A preprint from researchers at EPFL suggests that three days of post-meal glucose monitoring may be enough to estimate a person’s recurring glucose-response pattern, even though responses vary substantially from meal to meal.

The analysis examined 50,463 meals recorded under free-living conditions from 992 adults without diagnosed diabetes. The researchers used a multivariate mixed-effects machine-learning framework to separate patterns shared across the study population from patterns specific to individual participants. They then evaluated the model against 4,524 standardized meals that had been excluded from training.

The work was posted on medRxiv on September 15, 2026. It is a preprint, and its immediate contribution is a model of individual glucose-response patterns rather than a treatment or clinical-outcome result.

Contents

How the analysis separated meal noise from personal patterns

Postprandial glucose refers to the blood-glucose response following a meal. In everyday life, that response is affected by the meal itself and by the person eating it, making it difficult to identify a stable individual pattern from a small number of observations.

The researchers analysed four glucose-response outcomes jointly rather than treating each response in isolation. The mixed-effects model was designed to distinguish population-level structure from participant-specific variation. This allowed the analysis to ask whether a person’s response profile could be reproduced across different groups of meals.

The comparison with standardized meals provided a separate test of the model. Because those 4,524 meals were held out during training, their results could be used to evaluate whether patterns derived from free-living meals transferred to a more consistent meal setting.

Two dimensions described individual responses

The model found that variation within the same person generally dominated the variation between participants in the free-living data. In other words, the same participant could respond differently across everyday meals. Despite that noise, the differences between people were concentrated mainly in two dimensions.

The dominant dimension represented an overall tendency towards larger glucose excursions. A glucose excursion is the rise and subsequent response pattern following a meal. The second dimension reflected carbohydrate responsiveness, with lower late glucose elevations forming part of that pattern.

Scores along the dominant dimension were reproducible when the researchers compared non-overlapping subsets of meals. They were also associated with responses to the held-out standardized meals after the analysis considered age, sex, body-mass index, and average glucose level and variability. This indicates that the model-derived score captured information beyond those basic measurements in this dataset.

Why the three-day result matters

The researchers estimated the dominant-axis scores using different monitoring periods. Scores calculated from three days of monitoring approximated scores calculated from 14 days. The finding suggests that a relatively short observation period may capture a repeatable component of a person’s meal-related glucose response without requiring two weeks of data.

That could make short-term monitoring useful for research in which participants need to be grouped according to their glucose-response patterns. The authors describe the dominant axis as a candidate coordinate for stratification: a way to organise participants by a measurable response characteristic before comparing interventions or studying metabolic variation.

The evidence currently concerns adults without diagnosed diabetes and modelled glucose-response patterns. Application to people with diagnosed diabetes and effects on patient outcomes were outside the reported analysis. The study therefore provides a method for measuring response individuality, while the clinical usefulness of that measurement remains a future research question.

Sources