ChatGPT in Iranian medical licensing examination: evaluating the diagnostic accuracy and decision-making capabilities of an AI-based model

Abstract

Introduction Large language models such as ChatGPT have gained popularity for their ability to generate comprehensive responses to human queries. In the field of medicine, ChatGPT has shown promise in applications ranging from diagnostics to decision-making. However, its performance in medical examinations and its comparison to random guessing have not been extensively studied. Methods This study aimed to evaluate the performance of ChatGPT in the preinternship examination, a comprehensive medical assessment for students in Iran. The examination consisted of 200 multiple-choice questions categorised into basic science evaluation, diagnosis and decision-making. GPT-4 was used, and the questions were translated to English. A statistical analysis was conducted to assess the performance of ChatGPT and also compare it with a random test group. Results The results showed that ChatGPT performed exceptionally well, with 68.5% of the questions answered correctly, significantly surpassing the pass mark of 45%. It exhibited superior performance in decision-making and successfully passed all specialties. Comparing ChatGPT to the random test group, ChatGPT’s performance was significantly higher, demonstrating its ability to provide more accurate responses and reasoning. Conclusion This study highlights the potential of ChatGPT in medical licensing examinations and its advantage over random guessing. However, it is important to note that ChatGPT still falls short of human physicians in terms of diagnostic accuracy and decision-making capabilities. Caution should be exercised when using ChatGPT, and its results should be verified by human experts to ensure patient safety and avoid potential errors in the medical field. Data are available on reasonable request.