GlucoFM two-stream pretraining lifts CGM prediction by 4.1 points

Researchers have built GlucoFM, a self-supervised foundation model for continuous glucose monitoring that splits every recording into two streams — a slow-moving baseline and the short deviations layered on top of it — rather than pushing the signal through the single representation stream that earlier CGM models use. Across four cohorts and seven clinical tasks, its frozen representations averaged 58.8 PR-AUC, against 54.7 for the strongest specialized CGM model retrained on the same data. Pretraining used 109,066 hours of unlabeled CGM data drawn from 477 participant and session records.

The problem GlucoFM is aimed at is not sensing. A CGM sensor sitting under the skin already reports interstitial glucose every few minutes, fasting, overnight and after meals — a far more direct read on metabolic regulation than the motion and heart-rate proxies consumer wearables collect. The problem is that almost none of those traces come with a clinical label attached, and labels are expensive.

That makes CGM an obvious target for self-supervised pretraining, and several groups have already tried: CGMformer, GluFormer and CGM-JEPA all learn from unlabeled traces. GlucoFM's argument is about what those models throw away. A CGM recording is a relatively smooth curve with short excursions written over it, and those excursions come from different sources — food, exercise, or the sensor misbehaving. Collapsing both into one representation stream asks the encoder to carry two kinds of information in one channel.

GlucoFM's two-stream encoder separates them explicitly: a low-frequency state component for slow glucose trends, and a residual event component for the short deviations, whether physiological, behavioral, or sensor error. Each recording is resampled onto a 24-hour grid at five-minute steps with an observation mask kept alongside it, so the model can tell a measured value from a missing one. Rather than reconstructing raw glucose values — which carry measurement noise and artifacts the model has no reason to memorize — pretraining predicts latent representations through two complementary objectives, which also teach it time of day and temporal structure.

The augmentations are the part worth noting for anyone who has handled sensor data. The team injected CGM-specific corruptions: baseline drift, simulated compression dropouts, sparser sampling, and short connectivity gaps. The model is deliberately trained on the failure modes of the hardware, not on an idealized trace.

Evaluation ran on four cohorts — CGMacros, Stanford, Hall and ShanghaiT2DM — against seven tasks: diabetes risk, insulin resistance, beta-cell dysfunction, hyperlipidemia, hypoglycemia, obesity and glucotype, for 14 cohort-task pairs in total. The first protocol was window-level linear probing with participant splits: freeze the encoder, train a linear classifier on individual 24-hour representations, and never let the same participant appear in both train and test. It is a narrow question deliberately posed — does one day of a stranger's glucose say anything about their phenotype.

GlucoFM posted the best task-averaged PR-AUC of the methods tested, moving the mean across those 14 evaluations from 54.7 to 58.8 — 4.1 points absolute, about 7.5 percent relative. It took the top PR-AUC in every diabetes risk and beta-cell dysfunction evaluation, and in three of four insulin resistance evaluations. Separately, the paper reports a 5.8-point average PR-AUC margin over the best tested GluFormer variant, pretrained on the same dataset. Two headline margins against two different reference points, and the write-up does not reconcile them.

The second test is dynamic rather than static: predict the full two-hour postprandial glucose trajectory, measured relative to the value at the start of the meal, using only what was known beforehand. The analysis covered 874 linked meal episodes across 34 participants, with participant-split cross-validation, and Dexcom and Libre data processed separately under identical splits. Context was added to each frozen representation in stages — one hour of pre-meal CGM, then meal composition (energy, carbohydrate, fat, protein, fiber), then participant attributes including fasting glucose, BMI and diabetes status.

With the full input set, GlucoFM reached the lowest mean absolute error at 21.88 mg/dL. The best competing foundation model came in at 22.90, and a baseline that simply predicts the training-set mean scored 27.69.

One day is a thin basis for reading an individual, so the team also tested aggregation: encode each day separately, then average the representations over windows of up to seven days, weighting every participant equally. Adding days raised PR-AUC in most settings, by 9.6 points for beta-cell dysfunction in Stanford and 14.0 points for diabetes prediction in Hall. In CGMacros the gains concentrated in the Dexcom, Libre and combined-sensor data. The exception was insulin resistance in ShanghaiT2DM under simple averaging, which suggests the right aggregation scheme is task-dependent. The practical point is that the gains come without retraining the encoder.

Transfer was tested by training a risk classifier on one cohort and evaluating it on participants from another. GlucoFM beat the strongest competing method in 11 of 12 evaluations, by margins from 0.5 to 8.6 PR-AUC points, and lost one by 0.6. Absolute PR-AUC ran from 61.6 percent on both Stanford to Hall transfer tasks up to 90.0 percent for insulin resistance on Hall to CGMacros. The researchers read this as evidence that the frozen representations can ignore cohort-specific noise and pick up shared metabolic structure.

Few-shot behavior was probed two ways: varying the number of labeled participants per class, and varying the fraction of observations available per participant. GlucoFM sat above every other model at every data volume tested, including the hardest settings — one participant per class, and 1 percent of observations available. The margin was widest when labeled participants were scarce.

Here is where I would temper the enthusiasm. The relative gains are consistent, which matters more than their size — a method that wins 11 of 12 transfers is doing something structural, not getting lucky. But the absolute numbers are modest. An average of 58.8 PR-AUC across 14 clinical tasks is not a model anyone should hand a diagnosis to, and the postprandial result buys 1.02 mg/dL over the next-best model while a training-set mean already gets within 5.81 mg/dL of it. The headline improvement is real and small, and the gap between "best available representation" and "clinically useful" is not measured here at all.

The ablation is the most convincing part of the paper, because it tests the actual claim. The full two-stream architecture was compared against three simpler variants: raw glucose values fed directly, a state-only version, and an event-only version. Event-only was the worst by a clear margin — short fluctuations alone do not describe metabolic state. Raw input and state-only landed in comparable territory. The full two-stream version was consistently first, which supports the decomposition rather than merely asserting it.

What the write-up is quiet about is the four tasks nobody reports wins on. GlucoFM's leadership is itemized for diabetes risk, beta-cell dysfunction and insulin resistance — the three tasks where glucose dynamics are most directly the thing being measured. Hyperlipidemia, hypoglycemia, obesity and glucotype are in the task list and then vanish from the results narrative. Also absent: any comparison against the metrics clinicians already use. Every baseline here is another learned model or a training-set mean. Nobody reports what time-in-range or mean glucose would have scored on the same 14 evaluations, and that is the number a skeptical endocrinologist would ask for first.

The pretraining corpus is the other constraint, and the authors say so. Those 109,066 hours come from 477 participant and session records — roughly nine and a half days apiece, which is a lot of signal from very few people. Metabolic responses vary across individuals, cohorts and CGM hardware, and a model trained on hundreds of records has not seen that variation. The stated plan is larger and more diverse pretraining groups, plus a move from independently encoded 24-hour windows to native multi-day modeling, which would let the model see trends unfolding over weeks and months.

The work is credited to Zechen Li, Keerthana Natarajan, Weizhi Zhang, Simon A. Lee, Yuwei Zhang, Maxwell A Xu, Menglian Zhou, Zeinab Esmaeilpour and Flora D. Salim of the University of New South Wales, with Mark Malhotra, Lindsey Sunden, Shwetak Patel, Yuzhe Yang and Ahmed A. Metwally. Bobak J. Mortazavi and Ricardo Gutierrez-Osuna of Texas A&M University supplied the CGMacros dataset. The two Wear-CGM studies were approved by Advarra under IRB numbers Pro00059582 and Pro00069880, with written participant consent for de-identified secondary research and algorithm development; the published datasets carry their own approvals.

The architectural bet here is narrow and testable: that glucose has at least two timescales worth modeling separately, and that a model told this explicitly learns more from the same data than one left to discover it. The ablation says yes. What the rest of the paper says is that the resulting advantage is a few points wide — which is the position every foundation model occupies right before somebody scales the pretraining set by two orders of magnitude and finds out whether the inductive bias was worth building in or merely worth having until the data arrived.