Omkar Nayak
Research Mentor(s): Xiang Zhou
Mentor Department: Biostatistics
Authors: Omkar Nayak, Robert Langefeld , Matt Zawistowski, Xiang Zhou
Session: Session 6 (3:00pm – 3:50pm)
Presentation Type: Poster 101
Abstract
Electronic health records serve as a rich source of medical data that can provide valuable insights into the presence, progression, and characteristics of key diseases. While certain components of electronic health records are well-structured, a significant portion of the data consists of unstructured text data in the form of clinical notes. In recent years, Large Language Models (LLMs), such as ChatGPT, have been shown to excel in the analysis and extraction of information from unstructured text data. In this project, we leveraged modern LLM methods to design a framework that systematically extracts critical disease-related information from de-identified clinical notes efficiently at scale. We also developed statistical methods for analyzing the reasoning of LLMs that ensure logical decision-making. We applied our methods to clinical notes for >48K Michigan Medicine patients enrolled in the Michigan Genomics Initiative (MGI) biobank to identify various conditions difficult to ascertain from structured data, including family histories of cancer, stroke, and type 2 diabetes. We highlight the benefit of LLM extracted patient family histories by performing a genome-wide association study (GWAS) of type 2 diabetes. Our analysis that incorporated family history showed increased power at genetic sites known to be associated with type 2 diabetes (TCF7L2, FTO, IGF2BP2) compared to a standard GWAS analysis. Our results demonstrate the potential applications for LLMs in the analysis of clinical data and the benefits it can provide to the research community.
