Abstract
BACKGROUND: Large language models (LLMs) are increasingly evaluated in medical education and clinical decision support, but their performance in highly specialized fields, such as nephrology, is not well established. We compared two advanced LLMs, GPT-4 and the newly released o1 pro, on comprehensive nephrology board renewal examinations. METHODS: We administered 209 Japanese Self-Assessment Questions for Nephrology Board Renewal from 2014 to 2023 to o1 pro and GPT-4 using ChatGPT pro. Each question, including images, was presented in separate chat sessions to prevent contextual carryover. Questions were classified by taxonomy (recall/interpretation/problem-solving), question type (general/clinical), image inclusion, and nephrology subspecialty. We calculated the proportion of correct answers and compared performances using chi-square or Fisher's exact tests. RESULTS: Overall, o1 pro scored 81.3% (170/209), significantly higher than GPT-4's 51.2% (107/209; p < 0.001). o1 pro exceeded the 60% passing criterion every year, while GPT-4 achieved this in only two out of the 10 years. Across taxonomy levels, question types, and the presence of images, o1 pro consistently outperformed GPT-4 (p < 0.05 for multiple comparisons). Performance differences were also significant in several nephrology subspecialties, such as chronic kidney disease, confirming o1 pro's broad superiority. CONCLUSION: o1 pro significantly outperformed GPT-4 in a comprehensive nephrology board renewal examination, demonstrating advanced reasoning and integration of specialized knowledge. These findings highlight the potential of next-generation LLMs as valuable tools in nephrology, warranting further and careful validation.