Pretrained time‑series models need task-specific tuning, and dietary data helps predict short‑term glucose
Continuous glucose monitors (CGMs) record glucose many times per hour and can be used to forecast short‑term glucose trends. This paper tests whether large pretrained time‑series models — so‑called foundation models — can predict CGM values out of the box, and whether adding dietary information (food images and nutrient records) improves those forecasts.
The researchers ran a controlled comparison on eight public CGM datasets that cover people with Type 1 diabetes, Type 2 diabetes, and people without diabetes. They evaluated classical methods (for example, Elastic Net), recent specialized time‑series models (for example, PatchTST), and several foundation models. They followed a unified protocol that varied how much past data the model saw and how far ahead it tried to predict.
They found that foundation models used zero‑shot — that is, without additional training on CGM data — did not consistently beat strong task‑specific baselines. However, a small amount of further training on CGM data (lightweight fine‑tuning) substantially improved results. For example, when fine‑tuned the Chronos‑Bolt foundation model reduced root mean square error (RMSE, a standard measure of prediction error) by about 6.5% to 18.4% in the Type 1 diabetes cohort and by about 8.6% to 18.2% in the non‑diabetes and Type 2 diabetes cohorts. These gains appeared both for in‑distribution tests and for some out‑of‑distribution settings.
To test dietary context, the team used the CGMacros dataset, which pairs CGM traces with food images and macronutrient records. They built a residual‑based fusion system that first predicts glucose from CGM alone and then adds a meal‑driven correction. Adding dietary information cut overall RMSE by roughly 3% and reduced error after meals (postprandial periods) by roughly 15% compared with CGM alone. The study also found that Chronos‑based CGM representations were more closely linked to observed meal‑related glucose rises than representations from LSTM or CatBoost models, even when those other models also used dietary inputs.