Can Time Series Foundation Models Forecast Nocturnal Hypoglycemia? A Benchmarking Study in Type 1 Diabetes

Abstract

Nocturnal hypoglycemia in type 1 diabetes (T1D) accounts for 5–6\% of T1D mortality and imposes significant psychological burden on patients and caregivers. Continuous glucose monitors (CGMs) have improved real-time glycemic awareness, yet clinically actionable long-horizon forecasting remains unsolved. Recent time series foundation models (TSFMs) have shown strong generalization across temporal domains, but their application to blood glucose forecasting is critically underexplored. Here we benchmark six state-of-the-art TSFMs—\texttt{Chronos2}, \texttt{Sundial}, \texttt{TiDE}, \texttt{TimesFM}, \texttt{TimeGrad}, and \texttt{TinyTimeMixer}—in zero-shot and fine-tuned settings across four public T1D CGM datasets, targeting 8-hour nocturnal forecasting horizons. Attention-based architectures consistently outperform MLP-based counterparts in capturing blood glucose trajectory morphology, with \texttt{TimesFM} achieving best-in-class performance. These findings challenge the assumed competitiveness of MLP-based TSFMs for physiologically complex time series and establish a reproducible benchmark advancing the feasibility of deploying TSFMs for nocturnal hypoglycemia prevention. Code is available at \url{https://github.com/Blood-Glucose-Control/nocturnal-hypo-gly-prob-forecast/}.