Coaching GlucoFM to know metabolism
We pre-trained GlucoFM on 109,066 hours of unlabeled CGM information from Put on-CGM
CGM recordings can include gaps, totally different sampling intervals, and sensor artifacts. GlucoFM aligns every recording to a 24-hour, five-minute grid and retains an statement masks, retaining measured and unobserved positions distinct. Its dual-stream encoder separates a lower-frequency state part, representing slower glycemic developments, from a residual occasion part capturing short-term deviations that will come up from physiology, conduct, or sensing artifacts.
Slightly than reconstructing actual uncooked glucose readings, which may be affected by measurement noise and sensor artifacts, GlucoFM makes use of latent predictive pre-training with two complementary duties:
- Contextual prediction: We masks (i.e., disguise) components of a every day glucose sequence and ask the mannequin to foretell their latent representations from the encompassing context. By predicting in latent area, the mannequin captures broader every day glucose patterns with out having to reconstruct each sensor studying.
- Temporal dynamics: We additionally practice the mannequin to foretell how an individual’s regular baseline and short-term deviations will shift from one hour to the following. This encourages it to seize the continual nature of glucose dynamics fairly than treating readings as remoted snapshots in time.
Lastly, CGM-aware augmentations introduce baseline drift, compression-like drops, sparser sampling, and brief disconnections, exposing the mannequin to variation and missingness encountered in actual CGM recordings.
