Yikuan Li, M.S., Hanyin Wang, Ph.D., Halid Z. Yerebakan, Ph.D., Yoshihisa Shinagawa, Ph.D., and Yuan Luo, Ph.D.
Advancing health data interoperability can significantly benefit research, including phenotyping, clinical trial support, and public health surveillance. Federal agencies such as the Office of the National Coordinator of Health Information Technology, the Centers for Disease Control and Prevention, and the Centers for Medicare & Medicaid Services are collectively promoting interoperability by adopting the Fast Healthcare Interoperability Resources (FHIR) standard. However, the heterogeneous structures and formats of health data present challenges when transforming electronic health record data into FHIR resources. This challenge is exacerbated when critical health information is embedded in unstructured rather than structured data formats. Previous studies relied on separate rule-based or deep learning–based natural language processing (NLP) tools to complete the FHIR transformation, leading to high development costs, the need for extensive training data, and the complex integration of various NLP tools. In this study, we assessed the ability of large language models (LLMs) to convert clinical narratives into FHIR resources. The FHIR–generative pretrained transformer (GPT) was developed specifically for the transformation of clinical texts into FHIR medication statements. In experiments involving 3671 snippets of clinical texts, FHIR-GPT achieved an exact match rate of more than 90%, surpassing the performance of existing methods. FHIR-GPT improved the exact match rates of existing NLP pipelines by 3% for routes, 12% for dose quantities, 35% for reasons, 42% for forms, and more than 50% for timing schedules. These findings provide confirmation of the potential for leveraging LLMs to enhance health data interoperability. (Funded by the National Institutes of Health and by an American Heart Association Predoctoral Fellowship.)