Abstract
Continuous glucose monitoring (CGM) combined with AI offers new opportunitiesfor proactive diabetes management through real-time glucose forecasting.However, most existing models are task-specific and lack generalization acrosspatient populations. Inspired by the autoregressive paradigm of large languagemodels, we introduce CGM-LSM, a Transformer decoder-based Large Sensor Model(LSM) pretrained on 1.6 million CGM records from patients with differentdiabetes types, ages, and genders. We model patients as sequences of glucosetime steps to learn latent knowledge embedded in CGM data and apply it to theprediction of glucose readings for a 2-hour horizon. Compared with priormethods, CGM-LSM significantly improves prediction accuracy and robustness: a48.51% reduction in root mean square error in one-hour horizon forecasting andconsistent zero-shot prediction performance across held-out patient groups. Weanalyze model performance variations across patient subgroups and predictionscenarios and outline key opportunities and challenges for advancing CGMfoundation models.