Artificial intelligence in neurosurgery: a systematic scoping review of machine learning, deep learning, and large language model applications

Background Artificial intelligence (AI) encompasses machine learning (ML), deep learning (DL), and large language models (LLMs), all of which are increasingly integrated into neurosurgical practice. Objectives This systematic scoping review evaluates AI applications in neurosurgery from March 2021 to March 2026, focusing on three primary hypotheses: (H1) the utility of ML/DL in neuro-oncological surgery; (H2) the role of AI in spine surgery planning and outcome prediction; and (H3) the performance of LLMs in neurosurgical knowledge-assessment and case-vignette benchmarking tasks. Methods A PRISMA 2020-compliant search was conducted in PubMed, EMBASE, Web of Science, and Scopus. After screening 1,847 records, 23 manuscripts met the eligibility criteria and were included. Descriptive statistics, frequency analyses, and subgroup comparisons were performed. Results Across included studies, CNN-based and hybrid models reported neuro-oncology accuracy values up to 87.2% with AUC values around 0.89, while AI models in spine surgery reported accuracy values ranging from 81.5% to 84.0% and AUC values from 0.79 to 0.85. In LLM benchmarking studies, GPT-4 achieved the highest reported performance across neurosurgical knowledge-assessment tasks, with accuracy ranging from 74.0% to 83.4%; however, these results represent descriptive, study-level comparisons rather than pooled meta-analytic estimates. Conclusions AI demonstrates strong diagnostic and predictive potential across all three neurosurgical domains. GPT-4 currently leads among evaluated LLMs. Data scarcity and prospective validation remain the primary barriers to clinical adoption.