Large Language Model GPT-4 Compared to Endocrinologist Responses on Initial Choice of Glucose-Lowering Medication Under Conditions of Clinical Uncertainty.

Abstract

Objective: To explore how the commercially available large language model (LLM) GPT-4 compares to endocrinologists when addressing medical questions when there is uncertainty regarding the best answer.

Research design and methods: This study compared responses from GPT-4 to responses from 31 endocrinologists using hypothetical clinical vignettes focused on diabetes, specifically examining the prescription of metformin versus alternative treatments. The primary outcome was the choice between metformin and other treatments.

Results: With a simple prompt, GPT-4 chose metformin in 12% (95% CI 7.9-17%) of responses, compared with 31% (95% CI 23-39%) of endocrinologist responses. After modifying the prompt to encourage metformin use, the selection of metformin by GPT-4 increased to 25% (95% CI 22-28%). GPT-4 rarely selected metformin in patients with impaired kidney function, or a history of gastrointestinal distress (2.9% of responses, 95% CI 1.4-5.5%). In contrast, endocrinologists often prescribed metformin even in patients with a history of gastrointestinal distress (21% of responses, 95% CI 12-36%). GPT-4 responses showed low variability on repeated runs except at intermediate levels of kidney function.

Conclusions: In clinical scenarios with no single right answer, GPT-4's responses were reasonable, but differed from endocrinologists' responses in clinically important ways. Value judgments are needed to determine when these differences should be addressed by adjusting the model. We recommend against reliance on LLM output until it is shown to align not just with clinical guidelines but also with patient and clinician preferences, or it demonstrates improvement in clinical outcomes over standard of care.