CGMformer

CGMformer is an AI model that learns patterns on its own, without labeled answers, using nothing but the glucose readings recorded every few minutes by a small sensor worn under the skin.

A glucose graph actually blends two different signals: a slow trend that rises and falls over the course of a day, and short, sharp spikes triggered by eating or moving around. CGMformer treated this graph as a single continuous line, reading it all at once. It's a bit like looking at a weather record and lumping the slow seasonal shift together with a sudden rain shower into one undifferentiated line.

This approach mattered as an early way of handling glucose signals, but it had a limitation: it couldn't separate the slow trend from the short-term fluctuations. Later models tried to improve on this by learning the two apart.

The article explains that "existing CGM foundation models such as CGMformer, GluFormer, and CGM-GPT processed glucose signals as a single stream of representations, failing to distinguish slow baseline patterns from short-term fluctuations." A common misunderstanding here is that CGMformer is not the newly released Google model, but a prior-generation model mentioned for comparison.