Clinical Notes as Narratives: Implications for Large Language Models in Healthcare

Abstract

OpenAI’s ChatGPT sparked tremendous excitement regarding potential healthcare applications of large language models (LLM). LLMs trained on electronic health record (EHR) notes could enrich the feature space for many tasks including risk prediction, data classification (e.g., identifying protected health information), augmented documentation, and patient communication. Crucially, LLMs will learn not only from objective clinical data, but also from patient narratives—subjective texts authored by human clinicians, who may be sources of bias. In recognizing clinical notes as clinical narratives, and clinicians as narrators, we gain important insights into potential downstream implications of training LLMs on EHRs. Here we argue that a richer understanding of notes’ narrative elements, informed by principles from the field of narratology, could facilitate the development of LLMs that are more conscious of bias and enable the delivery of high-quality, human-centered care.