Evaluating Zero-Shot Foundation Models for Time Series Forecasting in Clinical Settings: A Simulation Study with Electronic Health Records
Abstract
Longitudinal healthcare data offer significant potential for advancing clinical decision-making through time series forecasting. Despite the development of high-performing taskspecific models, their clinical implementation is often limited by challenges in generalizability, data sharing, and resource constraints. Foundation models, which demonstrate zero-shot capabilities and reduced dependency on task-specific data, present a promising alternative. This study evaluates the zero-shot forecasting performance of three foundation models—Chronos, Time-LLM, and Time-MoE—compared to optimized task-specific models, using electronic health records from a German university hospital for training and two external validation datasets. Three clinical use cases with diverse temporal and predictive properties were analyzed. In this study, task-specific models, particularly deep learning models, outperformed zero-shot models in accuracy across most scenarios. However, zero-shot models demonstrated competitive performance, particularly in external validation datasets, underscoring their strong generalization potential. These findings suggest that the ease of implementation and transferability of zero-shot foundation models make them a viable option for clinical scenarios where retraining is impractical.
Amin Dada
Team Lead NLP
