A Foundation Model for Glucose Time

Continuous glucose traces look like one noisy curve. They are not. They contain a slow baseline and short-term deviations from meals, activity, and sensor artifacts. GlucoFM is built around that split.

The model is a lightweight, self-supervised foundation model for continuous glucose monitoring. One stream tracks slower metabolic state. The other tracks transient events. Time of day is preserved. Missing readings are masked rather than invented. Pretraining predicts latent structure instead of reconstructing raw noise, and uses augmentations that mimic real device failures: drift, sparse sampling, brief disconnects.

Trained on more than 100,000 hours of unlabeled CGM, the resulting representations transfer to diabetes risk, insulin resistance, beta-cell dysfunction, hypoglycemia, obesity, glucotype, and post-prandial glucose response. Frozen features with simple linear probes beat strong baselines, work across datasets, and remain useful when labeled examples are scarce. Multi-day averaging further improves subject-level prediction.

The current design still works in 24-hour windows, not weeks or months. The larger point stands. Unlabeled wearable traces become reusable clinical representations when the model respects the timescales already present in the signal.

This Ledger Entry expands how the reader thinks by treating CGM not as a single curve but as a dual-timescale digital biomarker that can be pretrained once and reused across metabolic tasks.