GlucoFM addresses a core problem in continuous glucose monitoring: treating a glucose trace as a single entangled signal mixes slow physiological baseline with fast transient events from meals, activity, stress or sensor artifacts. This entanglement forces existing foundation models to learn conflicting patterns, limits performance when labels are scarce, and inflates model size without delivering proportional gains.
The solution proposed by GlucoFM is to split the signal into two streams—a slow “state” component that captures the underlying glycemic baseline and a fast “event” component that captures short‑term deviations. By preserving the observation mask and using a causal, mask‑aware Gaussian filter, the model keeps missing data informative rather than imputing it away. Two JEMA‑style pretraining objectives then learn representations from masked patches and predict future state/event dynamics, all while operating on a modest 0.72 M‑parameter encoder that fits on a single H100 GPU.
For practitioners, the research prototype is not yet a regulated product, but the training recipe is fully reproducible: any team with a CGM corpus can train the model on a single H100 in about 120 epochs, and inference for a 24‑hour window runs comfortably on a CPU container or on‑device. This enables labs and startups to build downstream tools—risk scores, meal response forecasts, insulin‑dose advisors—without needing massive labeled datasets or huge compute budgets.
Key takeaways for the audience are: the state/event split yields a 4.1 PR‑AUC lift over the best same‑corpus baseline; performance gains are strongest on clinically central tasks such as diabetes risk, beta‑cell dysfunction and insulin resistance; and the compact model outperforms alternatives that are hundreds of times larger.
#AI #GlucoFM #CGM #Diabetes #MachineLearning #HealthTech