Performance of GPT-4 for planning acupuncture treatment: comparison with human clinician performance.

Abstract

Background: The medical knowledge of GPT-4 has been evaluated on patient data, providing diagnostic and treatment suggestions. However, few studies have directly compared the clinical suggestions of GPT-4 with those of groups of practitioners.

Methods: This study assessed the ability of GPT-4 to make medical decisions regarding acupuncture treatment by comparing its selection of acupoints with those made by human clinicians. Ten case reports published in Korean medical journals were selected and put in a standardized format. The standardized patient information was given to 80 Korean Medicine doctors and GPT-4 to diagnose and prescribe three to five acupoints per case. To evaluate the performance of GPT-4, the similarities in acupoint selection between the doctors and GPT-4 were quantified based on the percentage overlap and correlations of the selection probabilities of acupoints in each case.

Results: The average percentage overlap for acupoints among cases at the 10% cutoff was 51.3%, i.e., more than half of the GPT-4 acupoint suggestions overlapped the acupoints selected by the doctors. In half of the cases, significant correlations were observed in the acupoint selection probabilities, implying that GPT-4 acupoint suggestions are similar to those of doctors.

Conclusions: GPT-4 made reasonable acupoint suggestions, with notable overlap observed with the prescriptions of doctors. This shows its promise for supporting medical decisions, education, and personalized medicine for patients undergoing acupuncture treatment. Future studies and validation are necessary to ensure the reliability and efficacy of applying GPT-4 in real-world settings.