Google Research publishes GlucoFM, a foundation model for continuous glucose data
Trained on 109,000 hours of unlabelled CGM data, the model splits the glucose curve into a slow baseline and short-term deviations. It is a research contribution, not offered as a product or clinical tool.
Google Research has published GlucoFM, a foundation model built on continuous glucose monitoring (CGM) data. It learns patterns from unlabelled data to produce representations reusable across metabolic health predictions.
What distinguishes the approach is that it does not treat the glucose curve as a single signal. A slower "state" stream carries a person's baseline glycemic trend; an "event" stream carries short-term deviations from meals, activity or sensor artefacts. Both streams preserve time of day and track where readings are missing.
Pre-training uses latent prediction objectives rather than reconstructing raw readings: predicting masked segments from surrounding context, and forecasting how the baseline and deviations shift hour to hour. The training data is 109,066 hours of unlabelled CGM records drawn from Wear-CGM studies and four published datasets, totalling 477 participant/session records.
Evaluation covers seven clinical prediction tasks across four cohorts: diabetes risk, insulin resistance, beta-cell dysfunction, hyperlipidemia, hypoglycemia, obesity and glucotype classification. Postprandial glycemic response forecasting was also tested on 874 meal events.
By the figures the company reports, the model improved PR-AUC by 5.8 percentage points on average over GluFormer, the strongest baseline retrained on the same data, and reached a mean absolute error of 21.88 mg/dL on postprandial prediction against the baseline's 22.90. It led in 11 of 12 cross-dataset transfer evaluations. These are Google's own measurements.
The work is presented as a research contribution; the page makes no statement about a product release or clinical availability. The authors also note limitations: the pre-training population is modest, and the model currently processes 24-hour windows independently.
For details see Google Research's post and the arXiv paper.
