Google Research Introduces GlucoFM: A Dual-Stream Foundation Model for Continuous Glucose Monitoring

Google Research and UNSW Sydney have developed GlucoFM, a self-supervised foundation model for continuous glucose monitoring. Unlike existing models like CGMformer, GluFormer, and CGM-JEPA, which encode glucose traces as one entangled sequence, GlucoFM splits the signal into two streams: a slow 'state' stream and a transient 'event' stream.

The researchers argue that this approach is more accurate, as continuous glucose monitoring (CGM) carries both a slow regulatory baseline and short transient deviations from meals, activity, stress, or sensor artifacts. This dual-stream architecture allows for more precise predictions of glucose levels.

GlucoFM was pre-trained on 109,066 hours of unlabeled CGM data from 477 subjects across five cohorts. The model reached a task-averaged PR-AUC of 58.8 against the strongest CGM-specific baseline retrained on the same corpus, with +4.1 points and 7.5% relative improvement.