Google unveils GlucoFM, a dual-stream glucose prediction model

Why CGM data has been hard to interpret

Wearables like smartwatches infer activity and sleep from indirect signals such as movement and heart rate, but they don't directly reveal how well someone's blood sugar is regulated. CGM devices, by contrast, use a small under-skin sensor to measure interstitial glucose every few minutes, capturing fasting, overnight, and post-meal patterns. The catch is that high-quality clinical labels needed to interpret these signals are rare and expensive to collect. Existing CGM foundation models like CGMformer, GluFormer, and CGM-JEPA also process glucose signals as a single representation stream, which means they can't separate slow baseline patterns from short-term fluctuations caused by meals, activity, or sensor noise.

Splitting slow trends from short-term noise

GlucoFM targets exactly this gap. It aligns 24-hour glucose records to a 5-minute grid and keeps a separate mask distinguishing observed points from gaps. The encoder splits into two branches — one handling the slow underlying glucose trend, the other handling residual fluctuations from meals, activity, and sensor noise. Rather than reconstructing raw glucose values directly, the model was pretrained with two objectives that predict latent representations, and it was deliberately exposed during training to the kinds of missing-data situations real CGM records run into: baseline drift, compression loss, sparse sampling, and brief disconnections. The model was pretrained on 109,066 hours of unlabeled data from 477 participants and sessions, combining two Wear-CGM datasets with four existing public datasets.

How much of a lead did it show?

Google Research evaluated seven clinical tasks — diabetes risk, insulin resistance, beta-cell dysfunction, dyslipidemia, hypoglycemia, obesity, and glucotype — across four cohorts: CGMacros, Stanford, Hall, and ShanghaiT2DM. The encoder was frozen and only a new linear classifier was trained on top, with training and evaluation sets split so no participant appeared in both.

Comparison	Strongest existing model	GlucoFM
Average PR-AUC (14 combinations)	54.7	58.8
vs. GluFormer (pretrained on same corpus)	baseline	+5.8 points
Post-meal 2-hour glucose prediction MAE	22.90 mg/dL	21.88 mg/dL
Cross-dataset transfer (12 evaluations)	baseline	led in 11, by 0.5–8.6 points

GlucoFM scored highest on every diabetes risk and beta-cell dysfunction evaluation, and on three of four insulin resistance evaluations. When predicting post-meal glucose change using 874 meal events (from 34 participants) captured on Dexcom and Libre CGM devices, the model combined one hour of pre-meal data, nutritional information, fasting glucose, and BMI to achieve a mean absolute error of 21.88 mg/dL — beating the best baseline of 22.90 mg/dL and a simple-average baseline of 27.69 mg/dL.

The gap widens as labels get scarcer

Averaging daily records over up to seven days improved predictive power in most settings — by 9.6 points on the Stanford beta-cell dysfunction evaluation and 14.0 points on the Hall diabetes prediction task. In cross-dataset transfer tests, where a classifier trained on one cohort was applied to a completely different cohort, GlucoFM beat the strongest existing model in 11 of 12 evaluations by margins of 0.5 to 8.6 points, with absolute PR-AUC ranging from 61.6% (Stanford→Hall) to 90.0% (Hall→CGMacros insulin resistance). GlucoFM's representations held the top score even under extremely sparse label conditions — one labeled participant per class and just 1% of observations — a result particularly valuable for real-world research settings where clinical labels are hard to come by. When the research team compared the dual-stream architecture against versions using raw input only, baseline-trend-only, or fluctuation-only representations, the fluctuation-only version performed worst, while the full model combining both streams consistently came out ahead.

Google Research's earlier work adding audio-visual clinical consultation abilities to its medical AI "AMIE" sits on the same organizational thread of healthcare AI research. Where AMIE addressed the clinical consultation touchpoint, GlucoFM sits a step earlier, in the interpretation of biosignals themselves.

In the paper, the research team acknowledged that the current pretraining population is still relatively small as a limitation, and said future work will expand to larger and more diverse populations while moving beyond independent 24-hour processing toward multi-day modeling spanning weeks or months, as well as real-time responsiveness.

Editor's take

The CGM market started out as a medical device category for diabetes patients and has, in recent years, broadened into consumer metabolic health tracking. The bottleneck along that path has stayed the same: sensors pour out data every five minutes, but the labeled clinical cases needed to translate that data into something like "this person is at high diabetes risk" are hard to accumulate even a few hundred of at a single hospital. That's precisely the gap GlucoFM is aimed at. Learning representations first from unlabeled raw signals, then boosting performance with only a small number of labels on top, is a recipe that's repeated itself in medical AI for years — but getting this level of transfer performance in a domain like glucose signals, where individual variation is high and readings differ across devices (Dexcom vs. Libre), is a notable result.

When self-supervised foundation models of similar scale get applied to real cohort data, the same pattern tends to show up: lumping everything into a single representation stream looks fine at first, then falls apart the moment the cohort changes. The fact that GlucoFM structurally separates baseline trends from short-term fluctuations and still won 11 of 12 cross-cohort transfer comparisons suggests that separation wasn't just a design preference — it actually filtered out noise.

The parties most likely to watch this technology domestically are medical device makers working with CGM hardware, digital health startups, and hospital research teams that have been accumulating cohort data. For anyone looking to cut labeling costs while keeping predictive power, it's worth considering borrowing pretrained representations like this and layering on a small amount of labels, rather than labeling an entire dataset from scratch. That said, this evaluation was limited to four cohorts of a few hundred participants each, so whether the same transfer performance holds up on a cohort reflecting Korean metabolic characteristics remains an open question.

Over the coming months, a key thing to watch will be whether this performance gap holds up as other research teams try to reproduce the paper on their own CGM cohorts. Given that Google Research has already flagged multi-day modeling and real-time responsiveness as the next steps, there's a good chance these representations will find their way into consumer metabolic health services that combine wearables with CGM.