Artificial Intelligence in Gastroenterology Education: DeepSeek Passes the Gastroenterology Board Examination and Outperforms Legacy ChatGPT Models
Andrew Ibrahim, Pojsakorn Danpanichkul, Alberto Hayek, Elahna Paul, Annmarie Farag, Masab Mansoor, Charat Thongprayoon, Wisit Cheungpasitporn, Mohamed O. Othman
Abstract
INTRODUCTION: DeepSeek was evaluated in gastroenterology board examination performance against legacy ChatGPT offline models, which previously showed failing performance. METHODS: The performances of the DeepSeek base R1 model and search-augmented R1 model were assessed using American College of Gastroenterology self-assessments (455 questions). RESULTS: DeepSeek exceeded the passing threshold. Search-augmented DeepSeek scored 81.5% across all questions, and the R1 base model scored 77.1%. Both search-augmented and offline DeepSeek models surpassed offline ChatGPT-3 (65.1%) and ChatGPT-4 (62.4%) ( P < 0.001). DISCUSSION: DeepSeek exhibited passing performance on the gastroenterology board examination but had gaps in niche topics and image exclusion limit utility. It may supplement education if validated by specialists.