Evaluation of Large language model performance on the Multi-Specialty Recruitment Assessment (MSRA) exam.

Introduction: AI-powered platforms have gained prominence in medical education and training, offering diverse applications from surgical performance assessment to exam preparation. This research paper examines the capabilities of Large Language Models (LLMs), including Llama 2, Google Bard, Bing Chat, and ChatGPT-3.5, in answering multiple-choice questions of the Clinical Problem Solving (CPS) paper of the Multi-Specialty Recruitment Assessment (MSRA) exam.

Methods: Using a dataset of 100 CPS questions from ten subject categories, we assessed the LLMs' performance against medical doctors preparing for the exam.

Results: Results showed that Bing Chat outperformed all other LLMs and even surpassed human users from the Qbank question bank. Conversely, Llama 2's performance was inferior to human users. Google Bard and ChatGPT 3.5 did not exhibit statistically significant differences in correct response rates compared to human candidates. Pairwise comparisons demonstrated Bing Chat's significant superiority over Llama 2, Google Bard, and ChatGPT 3.5. However, no significant differences were found between Llama 2 and Google Bard, Llama 2, and ChatGPT-3.5, and Google Bard and ChatGPT-3.5.

Discussion: Freely available LLMs have already demonstrated that they can perform as well or even outperform human users in answering MSRA exam questions. Bing Chat emerged as a particularly strong performer. The study also highlights the potential for enhancing LLMs' medical knowledge acquisition through tailored fine-tuning. Medical knowledge tailored LLMs such as Med-PaLM, have already shown promising results.

Conclusion: We provided valuable insights into LLMs' competence in answering medical MCQs and their potential integration into medical education and assessment processes.