The performance of ChatGPT versus neurosurgery residents in neurosurgical board examination-like questions: a systematic review and meta-analysis

Objective Large language models and ChatGPT have been used in different fields of medical education. This study aimed to review the literature on the performance of ChatGPT in neurosurgery board examination-like questions compared to neurosurgery residents. Methods A literature search was performed following PRISMA guidelines, covering the time period of ChatGPT's inception (November 2022) until October 25, 2024. Two reviewers screened for eligible studies, selecting those that used ChatGPT to answer neurosurgery board examination-like questions and compared the results with neurosurgery residents' scores. Risk of bias was assessed using JBI critical appraisal tool. Overall effect sizes and 95% confidence intervals were determined using a fixed-effects model with alpha at 0.05. Results After screening, six studies were selected for qualitative and quantitative analysis. Accuracy of ChatGPT ranged from 50.4 to 78.8%, compared to residents' accuracy of 58.3 to 73.7%. Risk of bias was low in 4 out of 6 studies reviewed; the rest had moderate risk. There was an overall trend favoring neurosurgery residents versus ChatGPT (p < 0.00001), with high heterogeneity (I2 = 96). These findings were similar on sub-group analysis of studies that used the Self-assessment in Neurosurgery (SANS) examination questions. However, on sensitivity analysis, removal of the highest weighted study skewed the results toward better performance of ChatGPT. Conclusion Our meta-analysis showed that neurosurgery residents performed better than ChatGPT in answering neurosurgery board examination-like questions, although reviewed studies had high heterogeneity. Further improvement is necessary before it can become a useful and reliable supplementary tool in the delivery of neurosurgical education.