Large Language Models and the Degradation of the Medical Record

Abstract

P erhaps no artifact of modern medicine has redefined medical practice more than the electronic health record (EHR). Physicians now spend the majority of their time writing and reading notes on the computer, tasks that bleed into "pajama time" and invade what would be time off. Many clinicians blame computerization for widespread burnout, alienation, and the decimation of primary care. It is therefore not surprising that many physicians and hospital administrators see EHR documentation generated by large language models (LLMs) as a potential path to salvation. LLMs (such as OpenAI's GPT-4 or Google's Gemini) are trained on large amounts of text and have demonstrated impressive abilities in processing and generating hu-manlike text in many domains, including medicine. The range of proposed applications is broad, from informa-tional tasks such as reviewing charts and summarizing clinical encounters to complex reasoning tasks such as making diagnoses and recommending treatments. Whereas many physicians consider using the current technology for the latter set of tasks to be unacceptably high risk, the former set is generally viewed as presenting a low-risk opportunity for physicians to reclaim time from computer work. That possibility has generated substantial enthusiasm , with numerous start-up companies attempting to tackle medical scribe tasks and EHR vendors partnering with artificial intelligence (AI) firms. We fear, however, that instead of facilitating communication and transparency, the insertion of LLM-generated text directly into the medical record risks diminishing the quality, efficiency, and humanity of health care. Such text may include structured notes about clinical encounters, prepopulated responses to patient-portal messages , or summaries of clinical information intended for physicians. We are particularly concerned that LLM-generated text will reduce the overall informa-tional quality of the chart, rendering this critical resource less useful for both physicians and future AI models. To understand the current challenges, it's helpful to revisit the debates over the creation of modern EHRs. By the mid-20th century, medical charts had grown bloated with information from many different silos of the health care system. This "source-based medical rec-ord" was unwieldy and difficult to parse for anyone who didn't