GluFormer

GluFormer treats a full day's worth of continuous glucose monitor readings as one unbroken line and learns the overall shape of that line to identify blood glucose patterns.

Think of it like scanning an ECG waveform from start to finish with your eyes, looking for unusual sections. A glucose graph mixes together slow, large-scale trends that rise and fall over hours with short ripples caused by meals or brief sensor noise. GluFormer doesn't separate these two — it learns them together as one combined signal.

This made it decent at capturing the big-picture trend, but limited when fine-grained distinction of short ripples was needed for prediction. Later models tried to address this by explicitly processing the slow trend and the short-term fluctuations separately.

The article introduces GluFormer alongside CGMformer and CGM-JEPA as earlier models that "processed glucose signals as a single representation stream, failing to distinguish slow baseline patterns from short-term fluctuations." It notes that the newly released GlucoFM scored 5.8 points higher in average PR-AUC than GluFormer when compared on the same pretraining data — worth noting that GluFormer is a separate, earlier model distinct from GlucoFM.