GlucoFM

A continuous glucose monitor writes a reading every five minutes, and two very different processes move that number: the slow drift of a person's glycemic regime, and the sharp excursion that follows a meal, a walk, or a sensor pressed against a mattress overnight. Encode the trace as a single sequence and a model must disentangle those timescales implicitly, from data in which they arrive already summed. GlucoFM, from Google Research and the University of New South Wales, separates them before representation learning begins: a learnable causal Gaussian filter estimates the slow trend, and the residual left behind becomes an event stream, the two encoded in parallel and fused into hourly tokens.

The same design treats sensing artifacts as structure rather than noise to be smoothed away. Irregular recordings are aligned to a 24-hour, five-minute grid of 288 positions, and an observation mask travels with the tensor so a filled position is never mistaken for a measured one. Pretraining predicts latent representations rather than reconstructing raw glucose values, keeping the objective from chasing compression drops and dropouts.

GlucoFM joins a small family of continuous glucose monitoring (CGM) foundation models — GluFormer, CGM-LSM, CGMformer, CGM-JEPA — all of which encode glucose as one representation stream. It was first posted as an arXiv preprint in May 2026 and described in a Google Research announcement in August 2026.

#Key Features

• State–event decomposition: A learnable causal Gaussian filter, bandwidth constrained to roughly 10–60 minutes, defines the slow state component; the residual becomes the event stream. Causality is enforced, so future readings never leak backward.
• Mask-aware daily grid: Recordings at 5- and 15-minute sampling share one 288-position daily grid with the observation mask preserved throughout, so gaps and device differences are modeled rather than interpolated over.
• Latent predictive pretraining: Two JEPA-style objectives run together — masked contextual prediction of hidden patches against an EMA target branch, and prediction of the next hour's state and event tokens by residual transition heads.
• CGM-aware augmentation: Baseline wander, compression-like drops, decimation to 15-minute sampling and short disconnection blocks are injected during pretraining to match what real sensors produce.
• Frozen-encoder transfer: Every downstream result uses a linear probe on a frozen encoder, including cross-cohort transfer and few-shot settings, so the backbone is never re-fit per dataset.

#Technical Details

GlucoFM is compact: a 3-layer transformer context encoder with hidden dimension 128, 4 attention heads and feed-forward dimension 256, paired with an EMA target encoder and a 1-layer predictor, totaling 0.72M trainable and 1.18M total parameters. Pretraining ran 120 epochs at batch size 128 on a single NVIDIA H100, over 109,066 hours of unlabeled CGM from 477 subject records: Google's internal Wear-CGM study (192 subjects, 75,330 hours) plus the ShanghaiT2DM, Stanford, BIG IDEAs Lab and Colas cohorts. Evaluation spans four downstream cohorts (CGMacros, ShanghaiT2DM, Stanford, Hall) and seven phenotype-classification tasks — diabetes risk, insulin resistance, β-cell dysfunction, hyperlipidemia, hypoglycemia, obesity and glucotype — for 14 cohort–task evaluations. Under subject-disjoint linear probing, average PR-AUC rises from 54.7 for the strongest CGM-specific baseline retrained on the same corpus to 58.8; GlucoFM leads 11 of 12 cross-dataset transfer evaluations. On two-hour postprandial glycemic response prediction over 874 meal events from 34 participants, the frozen embedding combined with pre-meal CGM, meal nutrition and subject context reaches 21.88 mg/dL mean absolute error against 22.90 for the best baseline. The work is an unrefereed arXiv preprint. The authors state that they will release code and reproducibility scripts; no official repository or checkpoint has been published, and the reimplementations that exist on GitHub and HuggingFace are community reconstructions rather than the authors' model.

#Applications

GlucoFM is aimed at research groups that hold CGM recordings but few clinical labels. Because the encoder stays frozen, a lab can embed each 24-hour window once and fit small task heads for diabetes risk, insulin resistance or β-cell dysfunction, averaging embeddings over up to seven days when subject-level rather than day-level prediction is wanted. The postprandial results extend this to event-conditioned prediction, the daily embedding supplying metabolic history alongside meal macronutrients and participant characteristics. The authors state plainly that GlucoFM is a research prototype, not cleared by any regulator.

#Impact

The result worth carrying forward is that a 0.72M-parameter model outperforms larger single-stream CGM foundation models trained on the same corpus, arguing that physiologically motivated structure matters more than capacity at the data scales CGM research currently reaches. That structure — decomposing by timescale, preserving the observation mask, predicting in latent space — generalizes beyond glucose to other sparsely labeled wearable biosignals. The constraints are equally clear: 477 pretraining records is a modest population for claims about demographic and device heterogeneity, the encoder still processes 24-hour windows independently, and with no public checkpoint the reported gains cannot yet be independently reproduced.