On 26 August 2026 Google Research published GlucoFM, a self-supervised foundation model for continuous glucose monitoring. The paper has been on arXiv since 29 May, with a second version on 25 August, from fourteen authors across Google Research and the University of New South Wales.
The architectural premise is that a glucose trace is not just another one-dimensional time series. It carries a slow drift of the baseline and, on top of it, short excursions that are typically postprandial, and the two run on different scales. GlucoFM separates them at the level of the representation.
How the model is built
The signal is aligned to a 24-hour chronological grid at five-minute steps, and unmeasured positions stay marked by an observation mask rather than being interpolated. On that grid the model produces two families of token: state tokens for the slow component and event tokens for the short deviations.
The encoder is a 3-layer Transformer with hidden dimension 128, 4 attention heads and feed-forward dimension 256, alongside a one-layer predictor and a target encoder updated by exponential moving average. At evaluation time the target branch is discarded and only the frozen encoder remains.
The parameter count is the figure the Google post leaves out and the paper states: 0.72 million trainable parameters, 1.18 million in total during pretraining, the difference being the target branch.
The training data
Pretraining uses 109,066 hours of unlabelled traces from 477 subjects, aggregating five cohorts at five- and fifteen-minute cadences:
Wear-CGM— 192 subjects, 75,330 hours, 5 minutesColas— 206 subjects, 9,544 hours, 5 minutesShanghaiT2DM— 44 subjects, 12,414 hours, 15 minutesStanford— 19 subjects, 8,761 hours, 5 minutesBIG IDEAs— 16 subjects, 3,017 hours, 5 minutes
Downstream evaluation uses four separate cohorts, 203 participants and 71,669 hours, across seven phenotype-classification tasks: diabetes risk, insulin resistance, beta-cell dysfunction, glucotype, hyperlipidaemia, hypoglycaemia and obesity.
Two cohorts appear on both sides, and the methodological point deserves a careful reading: separation is defined at dataset level over subject groups, so no subject used for pretraining reappears in the downstream test groups. The Wear-CGM protocols are approved by the Advarra IRB under numbers Pro00059582 and Pro00069880, with written informed consent.
The pretraining objectives
Two latent-prediction tasks, both in representation space rather than over reconstructed glucose values. The first is JEPA-style contextual prediction: portions of the sequence are masked with a ratio sampled between 0.5 and 0.6, and their representation is predicted from the surrounding context. The second predicts the hour-to-hour transition of baseline and deviation, with a residual update form that pushes the heads to model change rather than state.
The augmentations simulate what happens to a sensor worn on a person: baseline drift, compression-like drops, sparser sampling, disconnections.
Probes applied to the two streams before fusion confirm the intended separation. State tokens better preserve the hourly glucose level, event tokens better preserve short-term change.
The measured results
The main protocol is linear probing with a frozen encoder and subject-disjoint splits, over 14 cohort-task evaluations.
On that protocol GlucoFM moves average PR-AUC from 54.7 to 58.8, which is 4.1 points above the best CGM-specific foundation model retrained on the same corpus, and 5.8 points above the best GluFormer variant evaluated. It leads on every diabetes-risk and beta-cell dysfunction task, and on three of the four insulin-resistance tasks.
The paired gain across the 14 evaluations is +4.11 PR-AUC points over CGM-JEPA, with a 95% confidence interval of 2.40 to 5.81, and +4.74 over CGMformer, from 2.47 to 7.02. The six intervals reported are Nadeau-Bengio corrected, exclude zero and stay significant after Holm adjustment, at an adjusted p of 0.012 or below.
On two-hour postprandial glycaemic response prediction with complete context, mean absolute error on the trajectory is 21.88 mg/dL against 22.90 for CGM-JEPA, the strongest comparator across all four endpoints. In the equally weighted two-sensor mean the reduction against CGM-JEPA is 4.5% on the trajectory, 5.8% on positive iAUC, 2.2% on peak rise and 3.4% on peak time.
Cohort transfer is measured by training the linear probe on one dataset and testing it directly on another. Across the twelve PR-AUC evaluations the model leads in eleven, by margins between 0.5 and 8.6 points, and trails once by 0.6 points on insulin resistance from Stanford to CGMacros. Counting PR-AUC and ROC-AUC together, the paper reports first place on 21 evaluations out of 24.
Size against performance
The comparison table is the part that matters most to anyone working on models, because scale does not order the results.
| Model | Parameters | Mean PR-AUC |
|---|---|---|
| GlucoFM | 0.72 M | 58.8 |
| CGM-JEPA | 0.52 M | 54.7 |
| CGMformer | 0.85 M | 54.1 |
| GluFormer tiny | 0.65 M | 53.0 |
| GluFormer base | 135 M | 52.3 |
| Mantis | 8 M | 55.5 |
| MantisV2 | 4.2 M | 56.9 |
The 135-million-parameter model sits at the bottom, below its own tiny variant at 0.65 million and below both Mantis variants. Among the CGM-specific models the spread between first and last is 6.5 points, and all of them stay under a megaparameter except the one that does worst.
In a domain with 109,066 hours of data the advantage comes from the structure imposed on the representation, not from the number of weights. It is a result you can act on: an encoder of 0.72 million parameters runs on modest hardware and opens the way to inference near the sensor, which for a continuous five-minute stream is the natural place for it.
Limits
The paper states its own, and states them plainly. The pretraining population remains modest and may not cover demographic, device, lifestyle and disease heterogeneity. Evaluation is retrospective and confined to phenotype classification and meal-aligned postprandial prediction, so it does not touch longitudinal outcomes, treatment response or prospective clinical deployment. The encoder models 24-hour windows independently, and the multi-day dimension is left to downstream aggregation.
The authors write that GlucoFM is a research prototype, cleared by no regulatory authority and not intended to diagnose or treat any disease, and that it should be validated within the intended population and clinical workflow before real-world use. That is the distinction separating a model from a certified medical device, and the same question applies as for any clinical software.
Code and reproducibility scripts are promised in the future tense in the paper. As of 29 August 2026 no repository or downloadable checkpoint appears to have been published, so the figures reported here are not currently reproducible by third parties.
On my own work: the architecture, corpus and table figures come from the version 2 PDF on arXiv, extracted and read locally. The Google blog post summarises the work without giving the parameter count, and its description of the corpus composition does not match table 2 of the paper, which is what I used.
- https://research.google/blog/glucofm-foundation-model-for-continuous-glucose-monitoring/
- https://arxiv.org/abs/2605.30865
- https://arxiv.org/pdf/2605.30865v2
Cover image: French barograph from 1930, National Air and Space Museum, photo by Votpuske, 2023 — CC BY 4.0 — https://commons.wikimedia.org/wiki/File:Barograph_(France,_1930).jpg