The real-world evidence in electronic health records has the opportunity to catalyze retrospective research, but much of the data is found not in structured fields, but trapped within clinical notes. Existing techniques in clinical information extraction have often relied on having a large amount of labeled data. In this talk, I will describe how large language models can be used to parse clinical text with only a minimal amount of supervision.

I will briefly explain recent advances in large language models as well as challenges to deployment in the clinical domain. I will proceed by describing our recent work demonstrating how these models can be used for clinical information extraction, sometimes with no labels required. This covers a variety of tasks in clinical NLP, including concept disambiguation, biomedical evidence abstraction, and medication relation extraction.