Google Research and UNSW Sydney have launched GlucoFM, a new foundation model for continuous glucose monitoring that separates blood sugar data into two distinct streams. The system processes a glucose trace by splitting it into a slow physiological state and a transient event stream, rather than encoding the data as a single entangled sequence. This approach uses a 0.72M-parameter encoder to reach a task-averaged PR-AUC of 58.8 across 14 cohort-task evaluations. That score beats the 54.7 achieved by the strongest CGM-specific baseline retrained on the same data. The model was pretrained on 109,066 hours of unlabeled CGM from 477 subjects using a single NVIDIA H100.
Is it ready for clinical use?
The team describes GlucoFM as a research prototype. It has not received clearance from any regulatory authority and is not intended to diagnose, treat, cure or prevent disease. All evaluations are retrospective, the largest pretraining cohort remains private, and no checkpoint has shipped as of 26 August 2026. The paper commits to releasing code and reproducibility scripts later.
What is available now is the recipe. At 0.72M trainable parameters and 120 epochs on a single NVIDIA H100, any team with a CGM corpus can reproduce the model. Inference for a 24-hour window runs on a CPU container or on-device hardware.
The problem with treating CGM as one signal
Existing foundation models like CGMformer, GluFormer and CGM-JEPA encode a glucose trace as a single entangled sequence. But CGM carries two things at once: a slow regulatory baseline and short transient deviations from meals, activity, stress or sensor artifacts. Clinical labels are also expensive and cohort-specific, which limits supervised training.
Architecture
GlucoFM aligns each recording to a fixed 24-hour grid at Δt = 5 minutes, giving L = 288 positions, and preserves the absolute circadian start index. An observation mask M is retained end to end — missing positions are filled only to build a tensor and never counted as measurements. An ablation shows dense interpolation underperforms this mask-aware default.
A causal, mask-aware learnable Gaussian filter then splits the signal: the filtered trend becomes the state stream, the masked residual the event stream. Bandwidth σ is learnable within 2–12 grid steps, roughly 10–60 minutes, initialized at 6.0. A one-sided kernel enforces causality, so future glucose never leaks into 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 uses two JEPA-style objectives: masked contextual latent prediction over 50–60% of patches against an EMA teacher (m = 0.997), and next-patch state/event dynamics prediction via residual transition heads. CGM-aware augmentations add baseline wander, compression-like drops, decimation to 15-minute sampling, and disconnection blocks.
The encoder is a 3-layer Transformer, hidden dimension 128, 4 heads, feed-forward 256 — 0.72M trainable and 1.18M total parameters. Pretraining used 109,066 hours of unlabeled CGM from 477 subjects across Wear-CGM, ShanghaiT2DM, Stanford, BIG IDEAs and Colas.
Results
Under subject-disjoint linear probing across four cohorts and seven tasks, GlucoFM reached 58.8 task-averaged PR-AUC against 54.7 for the strongest CGM-specific baseline retrained on the same corpus — +4.1 points, about 7.5% relative — and 5.8 above the best GluFormer variant. It led PR-AUC on every diabetes-risk and beta-cell-dysfunction evaluation and 3 of 4 insulin-resistance evaluations, and ranked first on 21 of 24 cross-dataset transfer evaluations.
For two-hour postprandial glycemic response forecasting it reached 21.88 mg/dL MAE with full context, against 22.90 for the best baseline and 27.69 for a train-fold mean, over 874 meal events from 34 participants across Dexcom and Libre sensors. It also beat 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 already matched CGM baselines trained on all of it.
What it means
For people managing diabetes, this work offers a more efficient way to process glucose data. The dual-stream design means the model can distinguish between the body’s long-term regulation and sudden spikes caused by food or activity. This separation could lead to better forecasts for post-meal responses and more accurate identification of diabetes risk or insulin resistance, all while requiring far less computing power than previous models.
