Consumer wearables use motion and physiological sensors to estimate activity and sleep, but these signals provide only an indirect view of glucose regulation. Continuous glucose monitors (CGM) track interstitial glucose every few minutes, capturing fasting, overnight, and post-meal patterns. Yet interpreting these traces is challenging when clinical labels are sparse and costly. Many existing CGM foundation models process glucose through a single representation stream, but CGM contains slow baseline patterns punctuated by short-term deviations.

Gluco FM, a self-supervised foundation model with a dual-stream design, separates slower glycemic trends from short-term deviations while preserving time-of-day and missingness. Evaluated across four cohorts on seven clinical prediction tasks, Gluco FM’s PR-AUC was 5.8 percentage points higher on average than the best-performing Glu Former variant. It led all diabetes-risk and beta-cell-dysfunction evaluations and three of four insulin-resistance evaluations.

On postprandial glycemic response forecasting, Gluco FM achieved the lowest mean absolute error across Dexcom and Libre devices, with the best cross-dataset transfer and strong few-shot adaptation. Pre-trained on 109,066 hours of unlabeled CGM data, Gluco FM uses latent predictive pre-training with contextual prediction and temporal dynamics tasks, plus CGM-aware augmentations to handle missingness and artifacts.