Loading…
Glucose forecasting from continuous glucose monitor streams over a two-hour horizon. Cuts one-hour rMSE 48.51% on OhioT1DM without training on it.
No providers recorded yet. Browse all providers
Continuous glucose monitors sample interstitial glucose every five minutes, producing one of the densest physiological time series routinely collected outside a hospital. Converting that stream into an actionable near-term forecast has been harder than it looks. Published glucose predictors are typically fit on a few dozen patients from a single cohort, and their accuracy over a clinically useful two-hour horizon degrades to the point where a patient cannot safely act on the number.
CGM-LSM, developed at Johns Hopkins University with Welldoc and Northwestern University, takes the approach large language models take to text: pretrain a single decoder autoregressively on a large corpus, then read forecasts off it directly. The authors call it a Large Sensor Model — a patient is a sequence of glucose time steps, and next-step prediction is the pretraining objective. Because that objective is exactly the downstream task, no task head or per-patient fitting stands between the pretrained checkpoint and a forecast.
It sits in the same emerging class as wearable-sensor foundation models such as NormWear, but narrows to one sensor and one physiological quantity, and spends the resulting capacity on generalizing across patient populations rather than across modalities.
CGM-LSM is a decoder-only transformer built on the GPT-2 architecture as implemented in Hugging Face Transformers: 12 layers, 12 attention heads, 768-dimensional embeddings, and a vocabulary of 400 glucose tokens plus 17 special tokens. Pretraining used de-identified CGM data from Welldoc, initially 21,215,912 records from 617 patients and filtered to 15,961,183 valid 26-hour instances from 592 patients (291 type 1, 301 type 2), then down-sampled 10% to roughly 1.6 million instances to reduce overlap between closely spaced windows. Training ran for ten epochs on a single A100 80GB with AdamW at a learning rate of 0.00005 and batch size 256; the third-epoch checkpoint was kept after validation loss began rising.
On the public OhioT1DM benchmark — 58,414 instances from 12 type 1 patients, never seen during pretraining — CGM-LSM reached an rMSE of 9.02 mg/dL at 30 minutes, 15.90 mg/dL at one hour (48.51% below a vanilla transformer trained on OhioT1DM directly), and 26.88 mg/dL at two hours, 26.67% below the strongest baseline. On held-out Welldoc patients, two-hour rMSE was 29.81 mg/dL for type 1 and 23.49 mg/dL for type 2, within 5.43% and 3.73% of the internal test sets drawn from patients the model had seen.
The intended use is micro-level diabetes self-management: showing a patient where their glucose is heading over the next two hours so they can eat, dose, or exercise in response, rather than the long-horizon risk scores that most diabetes AI produces. Because a single checkpoint serves patients it has never encountered, it fits the onboarding problem faced by CGM vendors and digital-health platforms, where a new user has no history to fit a personal model against. The subgroup analysis also gives clinical teams a concrete map of where the forecast is weaker.
CGM-LSM is evidence that the autoregressive pretraining recipe transfers to physiological sensor streams, and that scale plus diversity in the pretraining corpus buys generalization that per-cohort models cannot reach. The paper is candid about where it falls short. Accuracy declines during daytime and mealtime hours and at both hypoglycemic and hyperglycemic extremes, where training examples are sparse and behavior dominates physiology, and error is consistently higher for type 1, younger, and female patients — subgroups the authors trace to greater measured glucose variability. The model also sees only glucose: meals, insulin, and activity are absent. It remains a preprint, first posted in December 2024 and substantially revised since. The training and evaluation code was released in July 2025 under no license, covering pretraining, the OhioT1DM evaluation, and the notebooks that convert CGM records into the tokenized format the model consumes. The pretrained weights are not distributed, and the Welldoc pretraining corpus is proprietary, so reproducing the reported numbers means retraining on data a reader would have to supply.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.