Urology consultants versus large language models: Potentials and hazards for medical advice in urology

Johanna Eckrich; Jörg Ellinger; Alexander Cox; Johannes Stein; Manuel Ritter; Andrew Blaikie; Sebastian Kuhn; Christoph Raphael Buhr

doi:10.1002/bco2.359

Urology consultants versus large language models: Potentials and hazards for medical advice in urology

BJUI Compass. 2024 Apr 3;5(5):438-444. doi: 10.1002/bco2.359. eCollection 2024 May.

Authors

Johanna Eckrich¹, Jörg Ellinger¹, Alexander Cox¹, Johannes Stein¹, Manuel Ritter¹, Andrew Blaikie², Sebastian Kuhn³, Christoph Raphael Buhr^{2

4}

Affiliations

¹ Department of Urology University Hospital Bonn Bonn Germany.
² School of Medicine University of St Andrews St Andrews UK.
³ Institute of Digital Medicine Philipps-University Marburg and University Hospital of Giessen and Marburg Marburg Germany.
⁴ Department of Otorhinolaryngology University Medical Center of the Johannes Gutenberg-University Mainz Mainz Germany.

Abstract

Background: Current interest surrounding large language models (LLMs) will lead to an increase in their use for medical advice. Although LLMs offer huge potential, they also pose potential misinformation hazards.

Objective: This study evaluates three LLMs answering urology-themed clinical case-based questions by comparing the quality of answers to those provided by urology consultants.

Methods: Forty-five case-based questions were answered by consultants and LLMs (ChatGPT 3.5, ChatGPT 4, Bard). Answers were blindly rated using a six-step Likert scale by four consultants in the categories: 'medical adequacy', 'conciseness', 'coherence' and 'comprehensibility'. Possible misinformation hazards were identified; a modified Turing test was included, and the character count was matched.

Results: Higher ratings in every category were recorded for the consultants. LLMs' overall performance in language-focused categories (coherence and comprehensibility) was relatively high. Medical adequacy was significantly poorer compared with the consultants. Possible misinformation hazards were identified in 2.8% to 18.9% of answers generated by LLMs compared with <1% of consultant's answers. Poorer conciseness rates and a higher character count were provided by LLMs. Among individual LLMs, ChatGPT 4 performed best in medical accuracy (p < 0.0001) and coherence (p = 0.001), whereas Bard received the lowest scores. Generated responses were accurately associated with their source with 98% accuracy in LLMs and 99% with consultants.

Conclusions: The quality of consultant answers was superior to LLMs in all categories. High semantic scores for LLM answers were found; however, the lack of medical accuracy led to potential misinformation hazards from LLM 'consultations'. Further investigations are necessary for new generations.

Keywords: Bard; ChatGPT; artificial intelligence (AI); chatbots; digital health; global health; large language models (LLMs); low‐ and middle‐income countries (LMICs); telehealth; telemedicine; urology.