CGM-JEPA

CGM-JEPA is an AI model pretrained only on raw data from a continuous glucose monitor (CGM) — a small sensor worn under the skin that measures blood glucose every few minutes. Instead of relying on doctor-assigned diagnostic labels, it looks for patterns directly in the curve that glucose readings trace over time.

Think of it like filming a stream all day long with a single camera. That camera doesn't distinguish between the stream's slow, sweeping bends and the small droplets splashing off a rock — it just captures everything as one continuous video. In terms of blood glucose, this means the model lumps together the slow rise-and-fall that happens over a day with the short spikes triggered by meals or exercise, learning them as one blended signal.

Because of this approach, CGM-JEPA and other models of its generation had a key limitation: they couldn't separate slow underlying trends from brief fluctuations. This limitation resurfaced as a point of comparison when Google Research later introduced GlucoFM, a model that learns to split the two apart from the start.

In the article, CGM-JEPA is cited alongside CGMformer and Glucoformer as an example of earlier models that "processed glucose signals as a single representational stream, failing to distinguish slow baseline patterns from short-term fluctuations." It's worth noting that CGM-JEPA is not made by the same developer as the newly introduced GlucoFM, nor is it an earlier version of GlucoFM — it's simply one of the prior works cited for comparison.