|

Google Research Introduces GlucoFM: A 0.72M-Parameter Dual-Stream Foundation Model for Continuous Glucose Monitoring

ByRicardo

Google Research and UNSW Sydney have launched GlucoFM, a self-supervised basis mannequin for continuous glucose monitoring. Its core transfer is a cut up. Existing CGM fashions — CGMformer, GluFormer, CGM-JEPA — encode a glucose hint as one entangled sequence. GlucoFM decomposes it right into a sluggish physiological “state” stream and a transient “occasion” stream, retains the remark masks intact, and pretrains with two JEPA-style latent goals. The result’s a 0.72M-parameter encoder that reached 58.8 task-averaged PR-AUC throughout 14 cohort–job evaluations, towards 54.7 for the strongest CGM-specific baseline retrained on the identical corpus. It was pretrained on 109,066 hours of unlabeled CGM from 477 topics, on a single H100.

Is it deployable?

As analysis infrastructure, sure. As a medical or shopper product, not but.

The analysis group state it instantly: GlucoFM is a analysis prototype, has not been cleared or permitted by any regulatory authority, and isn’t supposed to diagnose, deal with, treatment or stop illness. Every analysis is retrospective, the most important pretraining cohort is private, and no checkpoint has shipped as of 26 August 2026 — the paper commits to releasing code and reproducibility scripts.

What is deployable at this time is the recipe. At 0.72M trainable parameters and 120 epochs on a single NVIDIA H100, any group with a CGM corpus can reproduce it, and 24-hour-window inference runs on a CPU container or on-device.

The drawback with treating CGM as one sign

Existing CGM basis fashions like CGMformer, GluFormer and CGM-JEPA encode a glucose hint as a single entangled sequence. But CGM carries two issues without delay: a sluggish regulatory baseline, and brief transient deviations from meals, exercise, stress or sensor artifacts. Clinical labels are additionally costly and cohort-specific, which caps supervised coaching.

Architecture

GlucoFM aligns every recording to a set 24-hour grid at Δt = 5 minutes, giving L = 288 positions, and preserves absolutely the circadian begin index. An remark masks M is retained finish to finish — lacking positions are crammed solely to construct a tensor and by no means counted as measurements. An ablation exhibits dense interpolation underperforms this mask-aware default.

A causal, mask-aware learnable Gaussian filter then splits the sign: the filtered pattern turns into the state stream, the masked residual the occasion stream. Bandwidth σ is learnable inside 2–12 grid steps, roughly 10–60 minutes, initialized at 6.0. A one-sided kernel enforces causality, so future glucose by no means leaks into the present state estimate.

Both streams are tokenized into 24 one-hour patches, fused into 128-dimensional tokens, and given round time-of-day options. Pretraining makes use of two JEPA-style goals: masked contextual latent prediction over 50–60% of patches towards an EMA instructor (m = 0.997), and next-patch state/occasion dynamics prediction through 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 whole parameters. Pretraining used 109,066 hours of unlabeled CGM from 477 topics throughout Wear-CGM, ShanghaiT2DM, Stanford, BIG IDEAs and Colas. <!– INTERACTIVE EXPLAINER EMBED GOES HERE –>

Results

Under subject-disjoint linear probing throughout 4 cohorts and 7 duties (14 cohort–job evaluations), GlucoFM reached 58.8 task-averaged PR-AUC towards 54.7 for the strongest CGM-specific baseline retrained on the identical corpus — +4.1 factors, about 7.5% relative — and 5.8 above one of the best GluFormer variant. It led PR-AUC on each diabetes-risk and beta-cell-dysfunction analysis and three of 4 insulin-resistance evaluations, and ranked first on 21 of 24 cross-dataset switch evaluations.

For two-hour postprandial glycemic response forecasting it reached 21.88 mg/dL MAE with full context, towards 22.90 for one of the best baseline and 27.69 for a train-fold imply, over 874 meal occasions from 34 members throughout Dexcom and Libre sensors. It additionally beat a seven-day GMI threshold rule on macro-F1 by +7.4 factors on Stanford and +17.4 on CGMacros-Dexcom. Trained on 20% of the corpus, it already matched CGM baselines skilled on all of it.

Key Takeaways

  • GlucoFM splits CGM right into a sluggish “state” stream and a transient “occasion” stream as a substitute of 1 entangled sequence.
  • 0.72M trainable parameters beat a 135M GluFormer and a 385M MOMENT on task-averaged PR-AUC.
  • 58.8 vs 54.7 PR-AUC over one of the best same-corpus CGM baseline throughout 14 cohort–job evaluations.
  • Strongest positive factors are on diabetes danger, beta-cell dysfunction and insulin resistance — the clinically central duties.
  • It is a analysis prototype with no regulatory clearance and no public checkpoint but.

Check out the Paper and Technical Details. Also, be happy to comply with us on Twitter and don’t overlook to affix our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

Need to associate with us for selling your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar and many others.? Connect with us

The put up Google Research Introduces GlucoFM: A 0.72M-Parameter Dual-Stream Foundation Model for Continuous Glucose Monitoring appeared first on MarkTechPost.

Similar Posts