GlucoCast

GlucoCast is a conditional diffusion framework for continuous glucose monitoring (CGM) data. It generates synthetic CGM trajectories conditioned on contextual signals (meals, insulin, activity) and uses the generated data to train downstream forecasters that outperform LSTM and CNN baselines on the OhioT1DM benchmark by roughly 18% RMSE.

The motivating constraint is data sharing. Real CGM traces are sensitive health data, difficult to release at scale. A generator that produces plausible, privacy-preserving traces unlocks training pipelines that would otherwise be blocked on data access.

Blood glucose forecasting helps people with type 1 diabetes anticipate hyper- and hypoglycaemic events. Supervised models (LSTM, CNN, Transformer) have shown promising results on public benchmarks but remain hard to deploy broadly because patient data is protected.

Diffusion models have emerged as a strong generative family for structured continuous data. Unlike GANs they have stable training dynamics; unlike VAEs they produce high-fidelity samples. Conditioning a diffusion model on the exogenous signals that drive glucose trajectories is a natural fit for this problem.

The generator learns to reverse a forward diffusion process on CGM windows, conditioned on aligned meal, insulin, and activity signals. Downstream forecasters are trained on mixtures of real and synthetic data, then evaluated against real held-out trajectories from OhioT1DM.

Models trained with GlucoCast-augmented data outperform LSTM and CNN baselines trained on real data alone by approximately 18% RMSE on OhioT1DM. The improvement is consistent across the standard 30-minute and 60-minute forecast horizons used in the benchmark.