Google research unveils glucofm: a dual-stream foundation model for enhanced continuous glucose monitoring

Google Research, in collaboration with UNSW Sydney, has introduced GlucoFM, a self-supervised foundation model designed for continuous glucose monitoring (CGM). This model represents a significant advancement by splitting the glucose trace into two distinct streams: a slow physiological 'state' stream and a transient 'event' stream. Unlike existing models such as CGMformer, GluFormer, and CGM-JEPA, which encode glucose data as a single entangled sequence, GlucoFM maintains an observation mask and employs two JEPA-style latent objectives during pretraining. The result is a 0.72M-parameter encoder that achieved a task-averaged PR-AUC of 58.8 across 14 cohort-task evaluations, surpassing the strongest CGM-specific baseline retrained on the same corpus with a score of 54.7. The model was pretrained on 109,066 hours of unlabeled CGM data from 477 subjects using a single NVIDIA H100 GPU. The research team emphasized that GlucoFM is a research prototype and has not been cleared or approved by any regulatory authority. It is not intended for clinical or consumer use and is not designed to diagnose, treat, cure, or prevent disease. All evaluations are retrospective, and the largest pretraining cohort remains non-public. As of August 26, 2026, no checkpoint has been released, though the paper promises to provide code and reproducibility scripts. The model's architecture aligns each recording to a fixed 24-hour grid with a 5-minute interval, resulting in 288 positions. It preserves the absolute circadian start index and uses an observation mask to retain missing positions without counting them as measurements. A causal, mask-aware learnable Gaussian filter splits the signal into the state and event streams. The bandwidth σ is learnable within 2–12 grid steps, roughly 10–60 minutes, initialized at 6.0. A one-sided kernel ensures causality, preventing future glucose data from influencing the current state estimate. Both streams are tokenized into 24 one-hour patches, fused into 128-dimensional tokens, and given circular time-of-day features. Pretraining involves two JEPA-style objectives: masked contextual latent prediction over 50–60% of patches against an EMA teacher and next-patch state/event dynamics prediction via residual transition heads. CGM-aware augmentations include baseline wander, compression-like drops, decimation to 15-minute sampling, and disconnection blocks. The encoder consists of a 3-layer Transformer with a hidden dimension of 128, 4 heads, and a feed-forward size of 256, totaling 1.18M parameters. Pretraining utilized data from Wear-CGM, ShanghaiT2DM, Stanford, BIG IDEAs, and Colas. Under subject-disjoint linear probing across four cohorts and seven tasks, GlucoFM achieved a 58.8 task-averaged PR-AUC, outperforming the strongest CGM-specific baseline by 4.1 points and surpassing the best GluFormer variant by 5.8 points. It led in PR-AUC for all diabetes-risk and beta-cell-dysfunction evaluations and three of four insulin-resistance evaluations, ranking first in 21 of 24 cross-dataset transfer evaluations. For two-hour postprandial glycemic response forecasting, GlucoFM achieved a 21.88 mg/dL MAE with full context, outperforming the best baseline with 22.90 and a train-fold mean of 27.69 over 874 meal events from 34 participants using Dexcom and Libre sensors. It also outperformed a seven-day GMI threshold rule on macro-F1 by +7.4 points on Stanford and +17.4 on CGMacros-Dexcom. Trained on 20% of the corpus, it matched CGM baselines trained on the entire dataset. The research team highlighted that the model's recipe is deployable today, with 0.72M trainable parameters and 120 epochs on a single H100, allowing any team with a CGM corpus to reproduce it, with 24-hour-window inference running on a CPU container or on-device.