Evaluating Zero-Shot Foundation Models for Time Series Forecasting in Clinical Settings: A Simulation Study with Electronic Health Records.

Longitudinal healthcare data enable time series forecasting to support clinical decision-making, but task-specific models face challenges in generalizability and resource constraints. This study evaluates the zero-shot performance of three foundation models-Chronos, Time-LLM, and Time-MoE-against optimized task-specific models using electronic health records from a German hospital and two external validation datasets. Across three clinical use cases, task-specific models outperformed zero-shot models in accuracy, but zero-shot models demonstrated strong generalization in external datasets. These findings suggest zero-shot models can be a viable option for scenarios where retraining is impractical, offering ease of implementation and transferability.