Designing Retrieval-Augmented Language Models for Clinical Decision Support

Ever-increasing demands for physician expertise drive the need for trustworthy point-of-care tools that can help aid decision-making in all clinical settings. Retrieval-augmented language models carry potential to relieve the information burden on clinicians in the next generation of clinical decision support systems by unifying non-parametric knowledge representations with parametric reasoning systems. Numerous designs for these semi-parametric language models have been proposed and tested in the general domain. In this paper, we assess their design choices through the lens of clinical decision-making, categorizing methods of knowledge injection and proposing future research directions that will advance retrieval-augmented language models towards integration in clinical workflows.