Living, Iterative Validation and Evaluation of AI in Medicine

About This Grant

Project Summary Foundation models (FMs), including large language models (LLMs), have created immense promise and excitement with their potential to change multiple aspects of how we pursue the science, practice and delivery of medicine. Given the potential, there is a need to create systematic ways to verify presumed benefits of using LLMs for medical tasks. The challenges in discerning hype from benefit are largely due to the absence of shared benchmarks that enable consistent evaluation (benchmarking), little guidance on trade-offs in performance for different post-training strategies (adaptation recipes), and a dearth of studies on the optimal mode of combining human expertise with models (human-AI teaming). We investigate methods to systematically evaluate LLMs in a setting where there is interaction between human experts to generate reusable guidance for performance benchmarking, adaptation recipes using human preferences, and implementing human-ai teaming in clinical settings. If successful, we will develop and implement concrete procedures to verify benefits of using LLMs in the context of the clinical tasks for which they are used. Overall, we seek to improve the creation and evaluation of foundation models in healthcare by rigorous benchmarking, developing systematic adaptation recipes, and guidance for optimal human-AI teaming setups.