On August 26, Google Research released GlucoFM, a self-supervised foundation model built purely on continuous glucose monitoring (CGM) traces. The paper is posted to arXiv as 2605.30865, authored by Google Research's Ahmed A. Metwally and Zechen Li. The problem it targets is narrow but concrete: an arm-worn sensor spits out a reading every five minutes, and a few days of that becomes a long, noisy curve. Extracting clinically useful signals from it — insulin resistance, beta-cell function — has traditionally required manual feature engineering.
The model splits the curve into two streams. One handles slow variables — metabolic trends over hours to a day. The other handles fast variables — post-meal spikes, overnight hypoglycemia, short-term deviations. Each stream is encoded separately, then merged into a joint representation. The team's ablation study found that a version relying solely on the "event stream" performed worst — momentary fluctuations alone can't support a stable metabolic picture.
The numbers, laid out
Pretraining consumed 109,000 hours of unlabeled CGM data, drawn from 477 participant/session records across five datasets. Evaluation spans four cohorts — CGMacros, a Stanford cohort, the Hall cohort, and ShanghaiT2DM — across seven clinical prediction tasks, for 14 cohort-task combinations in total. Against the best variant of GluFormer, GlucoFM's PR-AUC came out 5.8 percentage points higher on average, a relative gain of roughly 7.5%. It leads across the board on diabetes risk and beta-cell function, and wins three of four evaluations on insulin resistance. For postprandial glucose response prediction, its MAE is 21.88 mg/dL versus a baseline of 22.90. On cross-dataset transfer, it wins 11 of 12 evaluations, by margins of 0.5 to 8.6 points.
The few-shot results are closer to how this would actually get used. Given just one labeled participant per category, or only 1% of observations, GlucoFM still holds its lead. In clinical research, labeling costs are often an order of magnitude higher than data collection costs — a more practical finding than any leaderboard score.
A contrast in data scale
Do the rough math: 109,000 hours at five-minute intervals works out to roughly 1.3 million data points. Language models these days routinely start in the trillions of tokens; by that measure, this dataset barely registers. At 477 records across five datasets, the scale is far from massive either way. Google is candid about this in its limitations section: the pretraining population is relatively small, and the model currently processes independent 24-hour windows only — it can't yet see week- or month-level trends.
Signal-based foundation models don't follow the same scaling logic as language models. CGM curves have low degrees of freedom and strong internal structure, so data from a few hundred people is enough for a model to learn what counts as normal variation. In its roadmap, Google says it plans to expand to larger, more diverse populations, build native multi-day modeling, support real-time adaptation, and better understand how noise patterns differ across cohorts.
Getting onto devices
The evaluation was run separately on both Dexcom and Libre sensor data, showing the model isn't locked to one hardware platform's sampling characteristics. The postprandial-response data — 874 paired meal events from 34 participants — isn't large, but it's clean. CGM has spent the past couple of years expanding beyond diabetes management into the general health-conscious population, and once prescription-free versions arrive, the number of wearers — and the pool of unlabeled curves available for self-supervised training — will only grow. A lightweight model that transfers well from few examples sits squarely in that gap.
Sources: Google Research official blog, arXiv paper 2605.30865, CocoLoop; model architecture, pretraining data volume, and the 14 cohort-task evaluation setup follow the official blog post.