GlucoFM: Foundation model for continuous glucose monitoring

Details

Consumer CGM devices track interstitial glucose every few minutes, but extracting clinical meaning from those traces requires labeled data that is expensive to obtain. GlucoFM, published by Google Research, is a lightweight self-supervised foundation model built specifically for CGM data that produces transferable representations without requiring large amounts of clinical labels. The model's core innovation is a dual-stream architecture that explicitly separates slow-moving baseline glucose trends from short-term deviations — such as post-meal spikes, activity-driven dips, or sensor artifacts. Prior CGM foundation models like CGMformer, GluFormer, and CGM-JEPA process glucose as a single undifferentiated stream; GlucoFM's latent-prediction objectives instead learn the daily context and temporal evolution of each stream separately, preserving time-of-day and missingness signals. Evaluated across four diverse cohorts and seven clinical prediction tasks — diabetes risk, insulin resistance, beta-cell dysfunction, hyperlipidemia, hypoglycemia, obesity, and glucotype — GlucoFM achieved a PR-AUC 5.8 percentage points higher than the best GluFormer variant on matching data. It also recorded the lowest mean absolute error on post-prandial glycemic response (PPGR) forecasting, validated across both Dexcom and Libre CGM devices. The model shows strong cross-dataset transfer and few-shot adaptation, meaning it can be applied to new clinical populations with very limited labeled data — a practical requirement for real-world metabolic health deployments where annotation is costly.