生成AI（Perplexity）による肝臓専門医認定試験問題への挑戦

抄録

We evaluated Perplexity AI's performance on the Japanese Hepatology Board Certification Exam. Perplexity achieved a 75.2% overall accuracy rate on 105 questions, but performance varied across categories. Compared with human test-takers, Perplexity demonstrated lower accuracy on image-based (63.6% vs. 76.9%) and X2-type questions (71.4% vs. 77.1%), although the differences were statistically nonsignificant. Results suggest that Perplexity AI can mimic the knowledge of hepatology specialists but has limitations including potential gaps relating to recent research data, integration of clinical guidelines, and language comprehension. Although AIs remain promising for medical support, human ethics and expertise remain crucial. This first assessment of a generative AI's performance on the Hepatology Board Certification Exam indicates the potential for AI to enhance diagnostic accuracy and the quality of patient care in hepatology.