Performance of ChatGPT on the Peruvian National Licensing Medical Examination: Cross-Sectional Study (Preprint)

Abstract

BACKGROUND The Peruvian National Licensing Medical Examination (ENAM) is an important milestone for Peruvian medical doctors. However, the failure rate almost reaches 60%. Access to high-quality medical education is inequitable. The recent growth of Artificial Intelligence (AI) poses an opportunity to close this breach and improve Peruvian medical education. OBJECTIVE To evaluate ChatGPT 3.5 performance in the Peruvian National Licensing Medical Examination and assess the usefulness of the explanation provided for each correct answer. METHODS The dataset consisted of 180 multiple-choice questions from the 2022 Peruvian National Licensing Medical Examination (ENAM). ChatGPT performance on these questions was evaluated with three prompts: open-ended (OE), multiple-choice questions without justification (MCQ-NJ), and multiple-choice questions with justification (MCQ-J). The quality of the explanations was evaluated by two independent raters. The performance of ChatGPT was compared with the score of 1025 Peruvian junior doctors who took the ENAM in 2023 as a progress test and with the historical mean of Peruvian medical doctors from 2009-2019. RESULTS ChatGPT passed the ENAM on the three prompts with the highest accuracy on the MCQ-J, scoring 77% (139/180). Surpassing the mean score of historical junior doctors (55%) and from junior doctors of 2023 (54%). Among correct answers on MCQ-J, 64% (89/139) provided explanations of good quality. CONCLUSIONS ChatGPT not only passed the ENAM but also outperformed the mean score of Peruvian examinees, raising concerns about the current state of Peruvian medical education. The variable performance across different prompts emphasizes the need for further research on prompt engineering. Although ChatGPT provided good-quality explanations in 64% of correct answers, its use to aid medical education still requires a review process. We anticipate continuous improvement in AI performance, potentially closing the barrier to access to high-quality medical education in the future. CLINICALTRIAL None