Evaluating the limits of AI in medical specialisation: ChatGPT's performance on the UK Neurology Specialty Certificate Examination.

BACKGROUND

Large language models such as ChatGPT have demonstrated potential as innovative tools for medical education and practice, with studies showing their ability to perform at or near the passing threshold in general medical examinations and standardised admission tests. However, no studies have assessed their performance in the UK medical education context, particularly at a specialty level, and specifically in the field of neurology and neuroscience.

METHODS

We evaluated the performance of ChatGPT in higher specialty training for neurology and neuroscience using 69 questions from the Pool-Specialty Certificate Examination (SCE) Neurology Web Questions bank. The dataset primarily focused on neurology (80%). The questions spanned subtopics such as symptoms and signs, diagnosis, interpretation and management with some questions addressing specific patient populations. The performance of ChatGPT 3.5 Legacy, ChatGPT 3.5 Default and ChatGPT-4 models was evaluated and compared.

RESULTS

ChatGPT 3.5 Legacy and ChatGPT 3.5 Default displayed overall accuracies of 42% and 57%, respectively, falling short of the passing threshold of 58% for the 2022 SCE neurology examination. ChatGPT-4, on the other hand, achieved the highest accuracy of 64%, surpassing the passing threshold and outperforming its predecessors across disciplines and subtopics.

CONCLUSIONS

The advancements in ChatGPT-4's performance compared with its predecessors demonstrate the potential for artificial intelligence (AI) models in specialised medical education and practice. However, our findings also highlight the need for ongoing development and collaboration between AI developers and medical experts to ensure the models' relevance and reliability in the rapidly evolving field of medicine.