Performance of ChatGPT on Clinical Medicine Entrance Examination for Chinese Postgraduate in Chinese (Preprint)

Abstract

The ChatGPT, a Large-scale language models-based Artificial intelligence （AI）, has fueled interest in medical care. However, the ability of AI to understand and generate text is constrained by the quality and quantity of training data available for that language. This study aims to provide qualitative feedback on ChatGPT's problem-solving capabilities in medical education and clinical decision-making in Chinese.

This study aims to provide qualitative feedback on ChatGPT's problem-solving capabilities in medical education and clinical decision-making in Chinese.

A dataset of Clinical Medicine Entrance Examination for Chinese Postgraduate was used to assess the effectiveness of ChatGPT3.5 in medical knowledge in Chinese language. The indictor of accuracy, concordance (explaining affirms the answer) and frequency of insights was used to assess performance of ChatGPT in original and encoding medical questions.

According to our evaluation, ChatGPT received a score of 153.5/300 for original questions in Chinese, which is slightly above the passing threshold of 129/300. Additionally, ChatGPT showed low accuracy in answering open-ended medical questions, with total accuracy of 31.5%. While ChatGPT demonstrated a commendable level of concordance (achieving 90% concordance across all questions) and generated innovative insights for most problems (at least one significant insight for 80% of all questions).

ChatGPT's performance was suboptimal for medical education and clinical decision-making in Chinese compared with in English. However, ChatGPT demonstrated high internal concordance and generated multiple insights in Chinese language. Further research should investigate language-based differences in ChatGPT's healthcare performance.

RR2-https://doi.org/10.1101/2023.04.12.23288452