Evaluation of the performance of large language models in endoscopic lumbar surgery: a comparative analysis

This study aimed to evaluate and compare the performance of three large language models (LLMs)-ChatGPT o1-preview, Claude 3.5 Sonnet, and Gemini 1.5 Pro-in providing information on endoscopic lumbar surgery based on 10 frequently asked patient questions.

The 10 high-frequently asked patient questions about endoscopic lumbar surgery were selected through discussion among authors. These questions were then submitted to the three LLMs. Responses were evaluated by five spine surgeons using a 5-point Likert scale for overall quality, text readability, content relevance, and humanistic care. Additionally, five non-medical volunteers assessed the understandability and satisfaction of the responses.

The intraclass correlation coefficients of ChatGPT o1-preview, Claude 3.5 Sonnet, and Gemini 1.5 Pro of the five evaluators were 0.522, 0.686, and 0.512, respectively. Claude 3.5 Sonnet received the highest scores for overall quality (4.86 ± 0.35, P <0.001), text readability (4.91 ± 0.32, P <0.001), and content relevance (4.78 ± 0.42, P <0.001). ChatGPT o1-preview was the most approved by non-medical background volunteers (49%), followed by Gemini 1.5 Pro (29%) and Claude 3.5 Sonnet (22%).

From the perspective of professional surgeons, Claude 3.5 Sonnet provided the highest quality and most relevant information. However, ChatGPT o1-preview was more understandable and satisfactory for non-professional users. This study not only highlights the potential of LLMs in patient education but also emphasizes the need for careful consideration of their role in medical practice, including technical limitations and ethical issues.