JMIR Publications

ABSTRACT

Background:

Large Language Models (LLMs) have shown broad applicability in medicine, including the generation of clinical documents. Beyond content creation, LLMs can also be used to evaluate the quality of medical documents. Because of LLMs' ability to simulate (or impersonate) specific personas, they can offer diverse perspectives (such as those of healthcare professionals versus patients with lower health literacy) on the clarity of medical texts.

Objective:

The primary objective of this research was to evaluate the ability of LLMs to simulate diverse user personas, varying by demographic profiles including educational background, gender, visit frequency, for the task of interpreting ICU discharge summaries. The study aimed to benchmark the clarity assessments generated by these LLM personas against a baseline established by human participants with corresponding backgrounds, in order to highlight the potential and limitations of using current LLMs to create personalized health information.

Methods:

We evaluated the ability of LLMs to simulate diverse user personas for the task of interpreting ICU discharge summaries. LLMs were prompted to adopt personas with varied demographic profiles, including different educational backgrounds. The resulting LLM-generated assessments of the summaries’ clarity were then benchmarked against a baseline established by human participants with corresponding backgrounds.

Results:

LLMs demonstrated a strong ability to simulate personas based on educational attainment, accurately interpreting key medical information in 88% of cases. However, the models’ performance varied widely when other demographic variables were introduced. For instance, persona performance was highly erratic based on gender, with simulated male personas achieving 97% accuracy while female personas achieved only 44%. The inclusion of additional details, such as the frequency of prior emergency room visits, further degraded the models' performance.

Conclusions:

This research highlights both the potential and the significant limitations of using LLMs to create personalized health information. While LLMs are promising for simulating user perspectives based on education, the current models exhibit unpredictable performance when tasked with incorporating other fundamental demographic traits like gender.