Survey Insights: The paper highlights the hype surrounding clinical foundation models and urges a refocus on the actual gaps in our understanding of their effectiveness. It provides a thorough review of over 80 foundation models trained on non-imaging EMR data to categorize their architectures and assess their potential use cases.

• Most existing foundation models are trained on limited and narrowly-scoped datasets, which may not adequately represent complex healthcare scenarios.
• The evaluation tasks employed often fail to give meaningful insights into the practical utility of the models for health systems.
• A new evaluation framework grounded in healthcare metrics is proposed to better measure clinical foundation models’ benefits.
• The survey underscores the importance of setting realistic expectations and addressing foundational gaps in the design and assessment of these models.

This paper lays bare the need for a paradigm shift in evaluating healthcare AI. It suggests that more realistic, diverse datasets and clinically meaningful evaluation benchmarks are needed to truly enhance patient care and hospital operations through AI.