CGM-JEPA: Learning Transferable Continuous Glucose Monitor Representations via Self-Supervised Pretraining and Cross-View Regularization

Continuous Glucose Monitoring (CGM) offers a minimal-invasive window into glucose dynamics, yet its potential to detect early metabolic subphenotypes (such as insulin resistance (IR) and 𝛽-cell dysfunction) remains untapped. While early prediction of these markers are vital for preventing Type 2 Diabetes, gold-standard diagnostics (e.g., venous OGTT in clinical setting) are invasive and costly for population-scale screening. Conversely, self-supervised representation learning from abundant, unlabeled CGM data offers a scalable alternative, but existing methods often fail to bridge the gap between controlled clinical data and noisy, real-world wearable streams. We propose CGM-JEPA, a self-supervised predictive pretraining framework that learns embeddings directly from raw CGM data via masked representation prediction, and X-CGM-JEPA, a cross-view extension that incorporates a masked Glucodensity objective, forcing the model to capture both temporal dynamics and global glycemic distributions. We rigorously evaluate our models on two independent clinical cohorts (Initial: 𝑁=32, Validation: 𝑁=24) with gold-standard venous measurements (3-hours) paired with corresponding mean of 28 days of CGM recordings per subject, under clinically motivated settings including cohort generalization, venous-to-CGM transfer, and real-world home CGM evaluation. Our proposed models demonstrate consistently strong performance across evaluation settings. While CGM-JEPA is competitive in indomain evaluations, X-CGM-JEPA yields the most consistent gains under distribution shift, particularly in venous-supervised transfer to wearable CGM representations +6% in AUROC, highlighting the effectiveness of predictive and cross-view self-supervision for deployment-relevant metabolic risk stratification from consumergrade CGM.