Artificial Intelligence and Robotics in Neurosurgery: A Systematic Review of Performance, Validation Quality, and Translational Readiness

Technical Summary: Artificial Intelligence and Robotics in Neurosurgery

Problem Statement
Neurosurgery demands extreme precision, where minor gains in intraoperative decision-making significantly impact patient safety and long-term neurological outcomes. Despite advancements in navigation and imaging, persistent challenges remain, including brain shift during operations, variable targeting precision in stereotactic interventions, steep learning curves, and time-consuming planning. While Artificial Intelligence (AI), Machine Learning (ML), Deep Learning (DL), and robotics have advanced individually to address these issues, existing literature has largely examined these domains in isolation. There is a lack of comprehensive reviews that integrate these technologies while critically appraising the quality of validation and translational readiness. Furthermore, few reviews apply a transparent, domain-specific risk of bias (RoB) assessment across both clinical interventional studies and algorithmic development studies, making it difficult to identify clear translational gaps.

Methodology
This systematic review adhered to PRISMA 2020 guidelines. A systematic search was conducted across PubMed/MEDLINE, Embase, Scopus, and Web of Science for studies published between January 2015 and June 2026. The search targeted quantitative outcomes related to AI, ML, DL, robotics, and adaptive stimulation in neurosurgery.

• Selection: From an initial pool of over 1,400 records, 56 original studies reporting quantitative outcomes were included.
• Classification: Studies were categorized into five technology domains: Deep Learning (DL) for brain tumor detection, Machine Learning (ML) for outcome prediction, Robotic Stereoelectroencephalography (SEEG), Deep Brain Stimulation (DBS) innovations, and other applications (e.g., simulations, documentation).
• Risk of Bias Framework: A novel, customized five-domain RoB framework was developed and applied to all 56 studies. The domains assessed were: (1) Data source and selection, (2) Validation method, (3) Risk of overfitting and data leakage, (4) Outcome measurement, and (5) Reporting quality.
• Analysis: Due to high heterogeneity in study designs and technologies, a meta-analysis was not performed. Instead, a structured narrative synthesis supported by quantitative summaries was utilized.

Key Contributions

1. Integrated Synthesis: This is one of the first reviews to conjointly examine AI/ML/DL and robotic technologies in neurosurgery, moving beyond isolated domain assessments.
2. Customized RoB Framework: The application of a tailored five-domain RoB tool allowed for a granular critical appraisal of both algorithmic development and clinical interventional studies, highlighting specific weaknesses such as overfitting and lack of external validation.
3. Quantification of the Translational Gap: The review explicitly quantifies the disparity between high technical performance metrics and the low level of prospective, externally validated clinical evidence.
4. Convergence Pathways: The paper proposes specific pathways for future integration between intelligent algorithms and robotic platforms, moving beyond siloed development.

Results

• Study Distribution: The largest domain was DL for brain tumor detection (11 studies, 19.6%), followed by ML for outcome prediction (10 studies), Robotic SEEG (8 studies), and DBS innovations (8 studies). Publication trends accelerated significantly after 2020.
• Risk of Bias: Overall, 50% of the included studies were rated as high risk of bias, and 57.1% relied exclusively on internal validation. Only 12.5% were prospective clinical trials, and merely 7.1% conducted external validation on independent cohorts.
• Performance Metrics:
  • DL Tumor Detection: Reported a median accuracy of 98.6% (range 95.86%–99.7%). However, 8 of the 11 studies were rated high risk of bias due to internal validation only. One study (Cepeda et al.) achieved low risk of bias with external validation, reporting a Dice score of 0.81 for contrast-enhancing tumors.
  • Robotic Applications: Robotic SEEG and DBS lead implantation demonstrated a median targeting error of approximately 1.5 mm (ranging from 1.14 mm to 2.74 mm). These studies generally showed better RoB profiles (mostly moderate risk) compared to pure algorithm development.
  • Adaptive DBS: Adaptive systems showed a statistically significant broadening of the therapeutic window (p = 0.005 in Parkinson's disease) and a reduction in total electrical energy delivered by approximately 22 µW compared to continuous stimulation.
  • Human-AI Collaboration: In glioblastoma resection estimation, combining explainable AI with human expertise improved accuracy from 0.53 (experts alone) to 0.94.
  • Documentation: NLP tools (ChatGPT) reduced operative report generation time from 15–21 minutes to 2.3–4.6 minutes, though factual accuracy varied (71–85%).
• Translational Bottleneck: A major bottleneck was identified where 88% of studies lacked prospective validation. The field is characterized by a high volume of retrospective and algorithmic development studies but a scarcity of prospective, multi-center clinical trials.

Significance and Claims
The authors claim this review provides the first integrated synthesis of AI and robotic technologies in neurosurgery using a rigorous, customized RoB framework. The significance lies in its ability to:

• Expose Methodological Fragility: It highlights that high reported performance (e.g., 98.6% accuracy in tumor detection) does not equate to robustness when external validation is absent.
• Identify Clinical Maturity: It establishes that adaptive neuromodulation systems and robotics currently demonstrate stronger clinical translation and more favorable risk of bias profiles than pure deep learning applications, which remain largely in the algorithmic development stage.
• Propose Actionable Pathways: By mapping the translational gap, the review proposes concrete pathways for convergence, such as closed-loop adaptive DBS using multimodal biomarkers and AI-enhanced robotic navigation.

The paper concludes that realizing the full potential of these technologies requires a deliberate shift from siloed, retrospective algorithmic improvements toward prospective, externally validated, multi-center studies and intentional convergence between robotic platforms and intelligent algorithms.