Foundation models in healthcare show promise by extracting meaningful representations from structured electronic health record (EHR) data.
State-of-the-art foundation models have demonstrated high performance in various clinical applications, even with limited labeled data.
There is a lack of consensus on the clinical utility of these models due to insufficient comprehensive and meaningful tasks for evaluation.
A new study proposes clinically relevant tasks and evaluates state-of-the-art foundation models on 14 tasks using EHR data from Columbia University Irving Medical Center to guide the development of future healthcare models.