GlucoFM: 0.72M Parameters Beats 385M MOMENT on Glucose Monitoring

Google Research and UNSW Sydney have published GlucoFM, a self-supervised foundation model for continuous glucose monitoring that achieves 58.8 task-averaged PR-AUC across 14 cohort–task evaluations using only 0.72M trainable parameters. The core architectural bet: rather than treating a glucose trace as one entangled time series — the approach taken by CGMformer, GluFormer, and CGM-JEPA — GlucoFM decomposes each recording into a slow physiological state stream and a transient event stream, then pretrains both branches with JEPA-style latent objectives. The result beats the strongest CGM-specific baseline retrained on the same corpus by 4.1 PR-AUC points and outperforms GluFormer (base) at 135M parameters and MOMENT (large) at 385M parameters — a result that echoes the broader pattern of small-model efficiency competing with raw scale.

The research carries an explicit caveat: GlucoFM is a prototype with no regulatory clearance, no public checkpoint as of 26 August 2026, and no intended clinical use. Every evaluation is retrospective, and the largest pretraining cohort is non-public. Code and reproducibility scripts are committed but not yet released.

Architecture

Each CGM recording is aligned to a fixed 24-hour grid at Δt = 5 minutes, yielding L = 288 positions per window. An observation mask M propagates through the entire pipeline — missing positions fill a tensor slot but are never treated as measurements. An ablation confirms that dense interpolation underperforms the mask-aware formulation.

A causal, mask-aware learnable Gaussian filter splits the signal: the filtered output is the state stream; the masked residual is the event stream. Filter bandwidth σ is learnable within 2–12 grid steps (roughly 10–60 minutes on the 5-minute grid), initialized at 6.0. The one-sided kernel enforces causality: no future glucose value leaks into the current state estimate.

Both streams are tokenized into 24 one-hour patches, fused into 128-dimensional tokens, and equipped with circular time-of-day positional features. Pretraining uses two objectives: masked contextual latent prediction over 50–60% of patches against an EMA teacher (m = 0.997), and next-patch state/event dynamics prediction via residual transition heads. CGM-aware augmentations include baseline wander, compression-like amplitude drops, decimation to 15-minute sampling, and disconnection blocks. The encoder is a 3-layer Transformer with hidden dimension 128, 4 heads, and feed-forward width 256. Total trainable parameters: 0.72M; total parameters including the EMA target branch during pretraining: 1.18M. Training ran for 120 epochs on a single NVIDIA H100 at batch size 128, over 109,066 hours of unlabeled CGM from 477 subjects spanning the Wear-CGM, ShanghaiT2DM, Stanford, BIG IDEAs, and Colas cohorts.

Benchmark Results

Evaluation used subject-disjoint 5-fold cross-validation with 10 repeats, identical splits across every method, under a frozen encoder with a logistic regression probe.

GlucoFM led PR-AUC on every diabetes-risk and beta-cell-dysfunction evaluation, won 3 of 4 insulin-resistance evaluations, and ranked first on 21 of 24 cross-dataset transfer evaluations. On two-hour postprandial glycemic response forecasting across 874 meal events from 34 participants — mixing Dexcom and Libre sensors — it achieved 21.88 mg/dL MAE against 22.90 for the best baseline and 27.69 for a train-fold mean predictor. Against a seven-day GMI threshold rule, GlucoFM improved macro-F1 by 7.4 points on Stanford and 17.4 points on CGMacros-Dexcom. Training on 20% of the pretraining corpus matched CGM-specific baselines trained on the full dataset, suggesting strong data efficiency. The weakest cohort was ShanghaiT2DM, a Libre 15-minute dataset under-represented during pretraining.

Practical Constraints

At 0.72M parameters with 24-hour-window inference, GlucoFM is CPU-deployable and on-device feasible. No checkpoint is public yet; the paper commits to releasing code and reproducibility scripts. The architecture is fully specifiable from the paper — encoder depth, filter bandwidth range, patch count, JEPA objectives, and augmentation schedule are all documented — placing it in the same tier as parameter-efficient domain-specific models where pretraining recipe matters more than scale.

GlucoFM is a systems design argument: signal decomposition before encoding, JEPA-style self-supervision without reconstruction loss, and mask-aware handling of missing data collectively produce a model that outperforms architectures nearly 500× its size on clinically relevant classification tasks. Whether this transfers to prospective CGM settings — different hardware, faster sensor iteration, non-research populations — remains the question only a clinical deployment can answer.