Performance and exploration of ChatGPT in medical examination, records and education in Chinese: Pave the way for medical AI

Abstract

Background: Although chat generative pre-trained transformer (ChatGPT) has made several successful attempts in the medical field, most notably in answering medical questions in English, no studies have evaluated ChatGPT's performance in a Chinese context for a medical task.

Objective: The aim of this study was to evaluate ChatGPT's ability to understand medical knowledge in Chinese, as well as its potential to serve as an electronic health infrastructure for medical development, by evaluating its performance in medical examinations, records, and education.

Method: The Chinese (CNMLE) and English (ENMLE) datasets of the China National Medical Licensing Examination and the Chinese dataset (NEEPM) of the China National Entrance Examination for Postgraduate Clinical Medicine Comprehensive Ability were used to evaluate the performance of ChatGPT (GPT-3.5 and GPT-4). We assessed answer accuracy, verbal fluency, and the classification of incorrect responses owing to hallucinations on multiple occasions. In addition, we tested ChatGPT's performance on discharge summaries and group learning in a Chinese context on a small scale.

Results: The accuracy of GPT-3.5 in CNMLE, ENMLE, and NEEPM was 56% (56/100), 76% (76/100), and 62% (62/100), respectively, compared to that of GPT-4, which was of 84% (84/100), 86% (86/100), and 82% (82/100). The verbal fluency of all the ChatGPT responses exceeded 95%. Among the GPT-3.5 incorrect responses, the proportions of open-domain hallucinations were 66 % (29/44), 54 % (14/24), and 63 % (24/38), whereas close-domain hallucinations accounted for 34 % (15/44), 46 % (14/24), and 37 % (14/38), respectively. By contrast, GPT-4 open-domain hallucinations accounted for 56% (9/16), 43% (6/14), and 83% (15/18), while close-domain hallucinations accounted for 44% (7/16), 57% (8/14), and 17% (3/18), respectively. In the discharge summary, ChatGPT demonstrated logical coherence, however GPT-3.5 could not fulfill the quality requirements, while GPT-4 met the qualification of 60% (6/10). In group learning, the verbal fluency and interaction satisfaction with ChatGPT were 100% (10/10).

Conclusion: ChatGPT based on GPT-4 is at par with Chinese medical practitioners who passed the CNMLE and at the standard required for admission to clinical medical graduate programs in China. The GPT-4 shows promising potential for discharge summarization and group learning. Additionally, it shows high verbal fluency, resulting in a positive human-computer interaction experience. GPT-4 significantly improves multiple capabilities and reduces hallucinations compared to the previous GPT-3.5 model, with a particular leap forward in the Chinese comprehension capability of medical tasks. Artificial intelligence (AI) systems face the challenges of hallucinations, legal risks, and ethical issues. However, we discovered ChatGPT's potential to promote medical development as an electronic health infrastructure, paving the way for Medical AI to become necessary.