Google Research and UNSW Sydney has released GlucoFMSelf-supervised Foundation Model for continuous glucose monitoring. Its main move is to split. Existing CGM models — CGMformer, GluFormer, CGM-JEPA — encode a glucose trace as one entangled sequence. GlucoFM breaks it down into a physiologically slow sequence. “state” The transient and the stream “event” The stream keeps the mask of observation intact and it pretrains using two latent JEPA objectives. The result is a 0.72M-parameter encoder that reached 58.8 task-averaged PR-AUC across 14 cohort–task evaluations, against 54.7 for the strongest CGM-specific baseline retrained on the same corpus. The encoder was pretrained using 109,066 unlabeled hours of CGM data from 477 different subjects on one H100.
Is it deployable?
Yes, as research infrastructure. Not yet.
Researchers state that GlucoFM, a prototype for research, is neither cleared nor approved by regulatory authorities, nor is it intended to cure, diagnose or prevent any disease. Every evaluation is retrospective, the largest pretraining cohort is non-public, and no checkpoint has shipped as of 26 August 2026 — the paper commits to releasing code and reproducibility scripts.
It is today’s recipe that is deployedable. With 0.72M parameters trainable and 120 epochs per single NVIDIA H100Any team with CGM corpus is able to reproduce this, and the 24-hour-window-inference can be run either on a computer container or device.
CGM is not a signal.
Existing CGM foundation models like CGMformer, GluFormer You can also find out more about the following: CGM-JEPA It is possible to encode the glucose trace using a multi-dimensional sequence. CGM has two aspects: A slow regulatory baseline and transient short deviations caused by meals, activity or sensor artifacts. Also, clinical labels can be expensive and are specific to a cohort. This limits the amount of supervised learning.
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. Dense interpolation is inferior to this default mask-aware.
The signal is split by a Gaussian learnable filter that has a causal mask. State of the art It is the residual that has been masked. You can also find out more about the following: stream. Bandwidth σ is learnable within 2–12 grid steps, roughly 10–60 minutes, initialized at 6.0. One-sided kernels enforce causality so that glucose estimates for the future are never influenced by current values.
The time of day features are circular and combine both streams into 24 patches. 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 enhancements are baseline wandering, drops that look like compression, decimation for 15-minute sample, and blocks of disconnection.
The encoder is a 3-layer Transformer, hidden dimension 128, 4 heads, feed-forward 256 — 0.72M trainable and 1.18M total parameters. Wear-CGM was used to pretrain 109 066 hours of CGM. ShanghaiT2DM, Stanford, BIG IDEAs Colas.
The results of the study are:
Under subject-disjoint linear probing across four cohorts and seven tasks (14 cohort–task evaluations), 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.
Over 874 meals from 34 participants, the two-hour glycemic postprandial response was forecasted to be 21.88mg/dL with context. This compares to 22.90mg/dL for baseline, and 27.69mg/dL for train-fold average. Dexcom You can also find out more about the following: Libre sensors. The macro-F1 also surpassed a GMI threshold of seven days on Stanford by 7,4 points and +17,4 on CGMacros Dexcom. The baselines for CGM were already the same as those trained across all corpora.
What you need to know
- GlucoFM separates CGM from a rapid “state” The transient and the stream “event” The stream is better than a single, entangled series.
- On PR-AUC averaged by task, 0.72M trainable parameter beats a 135M GLUFORMER and a MOMENT of 385M.
- 58.8 vs 54.7 PR-AUC over the best same-corpus CGM baseline across 14 cohort–task evaluations.
- Strongest gains are on diabetes risk, beta-cell dysfunction and insulin resistance — the clinically central tasks.
- This is still a prototype for research, and there has been no official approval or public test.
Click here to find out more Paper You can also find out more about the following: Technical Details. Also, feel free to follow us on Twitter Don’t forget about our 150k+ML SubReddit Subscribe Now our Newsletter. Wait! Are you using Telegram? now you can join us on telegram as well.
You can partner with us to promote your GitHub Repository OR Hugging Page OR New Product Launch OR Webinar, etc.? Connect with us

