Abstract

Foundation time-series models have recently shown potential for forecasting complex temporal signals, but their behavior in patient-specific continuous glucose monitoring (CGM) forecasting remains insufficiently understood, particularly when only glucose history is available. This study provides a patient-level benchmark of foundation models for 30 min ahead glucose prediction in adults with type 1 diabetes mellitus (T1DM) under a strictly univariate CGM-only setting. Using the HUPA–UCM dataset from 25 individuals, we evaluated TimeGPT, Chronos, and Sundial against representative statistical, machine learning, and deep learning forecasters, including ARIMA, ETS, gradient-boosting models, recurrent networks, and neural forecasting architectures. Models were assessed using a local walk-forward validation strategy over the final 24 h of CGM data for each patient. Foundation models achieved the strongest global performance, with Sundial obtaining the lowest overall MAE (6.06mg/dL), while TimeGPT and Chronos remained among the most competitive approaches. However, patient-level analyses showed that this advantage was not uniform: ARIMA remained highly competitive in selected individuals, and no single model consistently dominated across the cohort. These findings suggest that foundation time-series models are promising tools for short-horizon CGM forecasting, but their use should be framed within patient-specific model selection rather than as universal replacements for classical forecasting methods.

IPC Classification

Keywords