Comparative Analysis of Large Language Models in Hemodialysis Vascular Access: ChatGPT-5, Gemini-2.5, and DeepSeek-V3

Abstract

Objectives: Vascular access is a key component of effective hemodialysis treatment. Patient education regarding selection, complications, and daily care remains challenging. Recently, large language models (LLMs) have been explored as supportive tools in this field. This study compared three widely used LLMs (ChatGPT-5, Gemini-2.5, and DeepSeek-V3) in addressing patient-centered questions on hemodialysis vascular access. Methods: Twenty-five frequently asked patient questions were compiled from literature, educational materials, and expert input. Each question was submitted to the three LLMs in standardized sessions, and answers were anonymized. Four cardiovascular surgeons independently evaluated responses in a blinded manner using 5-point Likert scales for accuracy, clarity, and scientificity. Statistical analyses compared model performances. Results: DeepSeek-V3 provided significantly responses with a higher word count and achieved higher scientific depth scores compared to ChatGPT-5 and Gemini-2.5 (P<0.01). Accuracy scores showed ChatGPT-5 demonstrated significantly lower than both Gemini-2.5 and DeepSeek-V3 (P<0.001). No significant differences were observed among models regarding clarity. Overall, DeepSeek achieved the highest mean scores across all criteria. Conclusions: DeepSeek-V3 demonstrated superior scientific depth and overall reliability, whereas ChatGPT-5 showed relative weaknesses in accuracy. The comparable clarity across models highlights LLMs’ potential as supportive tools for patient education. Future studies should further validate these tools in real-world clinical settings and define appropriate safeguards for their use.

Ähnliche Arbeiten