From theory to practice: what is the potential of artificial intelligence in the future of neurosurgery?

KEYWORDS:

• AI
• neurosurgery
• machine learning
• data science
• robotics

1. Introduction

Artificial Intelligence (AI) is reviving and disrupting modern surgery and healthcare by employing machines that can learn, understand, and predict – whether in robotic assisted surgery or through algorithms which facilitate the design of tailored, precise treatments for surgical patients [1,2]. AI refers to algorithms which mimic and adopt human cognitive functions (i.e. decision-making, learning, and reasoning [3–5]. This functionality is achieved through training of systems on data sets where key features are labeled from known examples (supervised learning) and by training to identify patterns from unlabeled data autonomously, without specific guidance (unsupervised learning) [4]. AI has evolved to handle multidimensional data like images and language; the recent introduction of natural language processing tools and large language models like ChatGPT [6] and Bard [7] have prompted experts to compare AI developments to the Industrial Revolution and suggest we are at the dawn of a technological epoch [8]. How will this technology be advanced and applied specifically to the field of neurosurgery?

AI is already being used to leverage various classes of data in order to modernize the neurosurgical patient pathway (Figure 1). It is currently being applied in the pre-operative planning of intracranial biopsy targets [9,10]; intraoperative real-time navigation of high-risk operative corridors involving eloquent areas [9]; and post-operative predictions regarding surgical complications and optimization of patient recovery [4]. AI integration both in and outside of the operating room (OR) has resulted in enhanced surgeon performance, superior patient outcomes, and lower healthcare costs [10–12]. Advances in machine learning have allowed for the rapid creation of accurate predictive models without being hindered by some limitations of classical statistical methods. These techniques can overcome unreliability from relatively small data sets and analyze large numbers of feature variables, demonstrated by Katsuki et al. in predicting outcomes for subarachnoid hemorrhage patients [13]. Certainly, the presence of AI in neurosurgery has brought about remarkable benefits. While challenges persist, the potential for transformative advancements propels us toward an exciting future.

Figure 1. The neurosurgical patient pathway and examples of AI’s roles at each stage [5,9,14–20].

2. The future of AI in neurosurgery

As healthcare technology accelerates and becomes increasingly automated, we anticipate that neurosurgery will not lag. Many AI applications in neurosurgery are predicated on the availability of contemporary multiomic data from national or international data repositories, which has both depth in scale and richness in individual profile. When this data is coupled with state-of-the art deep-learning algorithms, predictions about neurosurgical disease are likely to be more insightful and made with greater certainty. By increasing the accuracy of predicted timescales of disease progression, and utilizing continuously adaptive analysis of patients’ healthcare records, earlier targeted surgical interventions can be offered. Large language models (LLMs) leveraging natural language processing, and deep neural networks entrained on radiomic features of interests, are a few of the ways in which AI can be utilized to identify at-risk individuals, potentially allowing more patients to receive surgery at earlier safer timepoints. These novel technologies may be capable of directing efforts toward preventative rather than reactive means of neurosurgical intervention, as is the case in other surgical oncology specialties [21].

In the neurosurgical operating theater of the future, surgeons can expect to work with augmented reality, smart wearables, and intelligent instruments – all interconnected within a continually adaptive, interactive, digital environment. Smart glasses can host a real-time HUD (Heads-up Display) alongside eye-tracking software, and highlight critical parameters including patient neurological function, anatomical landmarks, and performance levels. These visual cues would help improve awareness and prioritize relevant features of the external operative environment and the surgeon’s own internal physiological landscape. Such devices may be used to display contemporaneous electrophysiological recordings to help define eloquent areas and surgical boundaries; superimpose key anatomical information and flag up impending dangers in the observed operative field; and warn surgeons when their concentration or stamina begins to wane. Intelligent surgical instruments can autonomously respond to the local environment and operator input. The force applied to the surgical tool can be modulated based on navigational information and tissue type, with greater sensitivity and reduced force or cautery settings when approaching or manipulating delicate structures or eloquent regions. Collectively all the data sources can be used to iteratively improve the set-up of the operative environment, case planning and identify potential areas to improve workflow, interpersonal dynamics, surgical efficiency and resource allocation. While some of these features will be introduced in stages, the full digitization, monitoring, and forward prediction of the operating environment will require extensive clinical testing to ensure feasibility and safe implementation rather than representing a distraction for the surgical team.

AI-based optimization is anticipated in key phases of postoperative recovery through automated monitoring and response of physiological parameters. Dynamic implantable technology in the form of brain control interfaces (BCIs) has considerable potential to improve rehabilitation through automated monitor-and-response mechanisms. BCIs function by interacting directly with neuronal circuitry, algorithmically decoding real-time cortical signals, and thus facilitating enhanced modulated brain and spinal cord stimulation. Indeed, in recent pilot work, this technology has managed to demonstrate significant improvement in gait rehabilitation and restoring competent walking over complex terrain in spinal cord injury patients [22]. In the future, BCIs may be able to recruit supplementary cortical and subcortical areas and facilitate neuroplastic processes permitting maintenance or restoration of key cognitive domains as well as supporting rehabilitation.

3. Ethical considerations

Since the term ‘robotics’ was coined in the 1920’s, fiction has gradually become reality, and there has been an urgent need for ethical frameworks that ensure a harmonious relationship between autonomous agents and their human creators. However, as more complex technology becomes available in clinical practice, more nuanced socio-ethical issues are anticipated, with some of direct relevance to neurosurgery. A select few examples are discussed below whilst the remainder are summarized in Table 1.

Table 1. Summary of foreseeable ethical issues arising because of AI implementation.

Issue	Case Example(s)	Solutions
Patient harm	IBM Watson Health Cancer AI algorithm, a clinical decision aid, recommended unsafe cancer treatments to physicians following inadequate training on too few case examples. One published example of this error was the recommendation to give Bevacizumab to a patient with bleeding despite severe hemorrhage as a known risk of this drug [23,24].	IDEAL Framework – guideline for safe surgical innovation [25].
		Robust frameworks for early evaluation and reporting such as DECIDE-AI [26].
		Registry of devices and record follow up of complications to identify and correct problems once technology is integrated into clinical practice.
		Collate as much unbiased data as possible to create the most reliable AI predictions.
		Utilise simulations in the development process.
Workforce de-skilling and job loss	AI automation is predicted to take over 300 million full-time jobs worldwide [27]. However, the surgical profession is the least likely to lose jobs to automation due to complexity and skill requirements [28].	Awareness and Identification of skills being lost to automation from educational bodies. This could be supplemented with simulation training from medical school onwards. More senior surgeons could also benefit from training in digital skills to avoid becoming out of date. One example of a digital literacy program is the European Skills Agenda from the European Commission which aims to upskill and train the medical workforce in digital skills.
Over-treatment	PSA screening was associated with overdiagnosis (detection of cancer that would otherwise not have been diagnosed in the patients life time) in 15–37% [29].	Screening program ideals could be considered before mass testing of biometric markers to ensure testing is meaningful, efficient and ethical [30,31].
		Role of an independent regulator to assess the patient impact and potential benefit of screening on mass a given biomarker
Environmental damage	Large scale medical data required for AI learning needs to be stored. A single data center can typically consume the equivalent electricity of 50,000 homes. Direct damage to ecosystems occurs by drawing on local water systems for cooling, preventing the water from enriching the ground and the animals dependent on it [32].	Investment in green energy supplies and support local environments when planning new data centers.Repurposing of heat produced in data centers
Increased health inequality	High costs associated with AI implementation is likely to restrict access to more affluent areas. This will likely exacerbate existing healthcare disparities between wealthy and poor communities.	Local input when distribution of technology is decided to avoid poorer communities being left behind. In a similar fashion to managing COVID vaccination distribution globally, global health organizations should prioritize fair distribution and access.
	Algorithms trained on biased data may exacerbate existing health inequalities [33]. A widely used health prediction algorithm in the US exhibited significant racial bias leaving out large proportions of colored patients from receiving additional health support [34].	Incorporation of known healthcare disparity variables into AI models.
Accountability	In a 13-year period, one study identified more than 8000 robotic system malfunctions reported to the USFDA including broken pieces of instruments falling into patients. In such cases, the liability for error remains unclear [35]	Clear legal frameworks discussed and created by government and healthcare leaders to delineate culpability in potential scenarios.
Patient sentiment	Surveys have found that patients overall find AI use both acceptable and appropriate for aspects of neurosurgical intervention such as image interpretation and operative planning. However, fully automated operations do not appear to be in public favor with only 17.7% of participants finding it acceptable [36]	Obtain valid informed consent from patients regarding data usage and potential injury resulting from autonomous procedures.
		Build up public trust in AI technology in a gradual manner with emphasis on AI assistance [37].
Data security	The ‘WannaCry’ ransomware cyber-attack in 2017 affected more than 60 NHS trusts which resulted in difficulties in accessing patient records and subsequent cancellations of appointments [38].	Investment must be directed toward cyber security in preparation of population wide data collection and amalgamation.
		Potentially allow patient access to data tracking systems to maintain transparency.

3.1. Biases and accountability

A major concern is the reinforcement of preexisting biases in AI algorithms, originating from historical imbalanced data, leading to inaccurate classification and misdiagnoses. For example, if there is disproportionate representation of certain demographic groups or specific pathologies, algorithms will be weighted toward the majority class and be less accurate in minority prediction. Establishing transparent decision-making frameworks which promote inclusive processes using multi-center balanced datasets are therefore vital. In addition, involving patients in comprehending how AI generates its recommendations creates a credible and reliable system [34].

Accountability of AI usage represents another area of concern. There remains ambiguity over who holds overall responsibility in which circumstance, as discussed in the ‘Second Law: Robot Malfunctions Must Be Reported’ in the iRobotic Surgeon study [39]. If errors were to occur in the design or implementation of a clinical AI tool – does the responsibility lie with the technology developer, medical professionals, or the technology itself? A case report detailing a medical instrument malfunctioning describes how the device did not notify the surgeon of the error, nor did it report any faults [40] – fortunately, in this instance, the surgery continued uneventfully once rectified, but this could have caused prolonged operative time and unnecessary anxiety. Establishing clear lines of accountability is therefore crucial and robust regulatory frameworks are needed to govern systems outlining responsibilities of stakeholders, evaluating safety, efficacy, and reliability.

3.2. Cost and inequality

Equal access to AI-assisted neurosurgery is limited by the prohibitively high costs for some healthcare providers. While in an ideal world all patients and providers would have access to AI technology, resource allocation must be balanced with financial demands elsewhere in the health system. The global market in AI-supported healthcare is expected to have reached $6.6 billion in 2021 [41]. However, resource limited healthcare systems in the wake of the COVID pandemic may struggle to set aside the large sums required for the adoption of new technology. For example, the UK only allows for USD $5,634 (GBP £4390) per capita on health expenditure as of 2021, leaving little room to invest in AI [42].

Despite the large initial investment, AI implementation is likely to lower individual health-care costs overall by increasing efficiency and reducing the need for redundant human input. As more tasks become automated, patients are likely to experience less variability in the care they receive, reducing inequality and healthcare disparity. However, unless AI technology and infrastructure is distributed evenly, geographical disparities will manifest leading to a two-tier healthcare system. In some countries, a disconnected digital evolution has already left hospitals without access to key benefits such as electronic prescribing or note taking, leaving patients with a postcode lottery on the efficiency of care they receive.

3.3. Data

The global DataSphere is expected to reach 175 zettabytes (1.75 × 10²³ bytes) by 2025 [43]. Healthcare data is expected to contribute significantly to this growth, as sufficient data is critical for AI systems to be reliable enough for clinical use. Storage of this data in data centers requires enormous amounts of energy to maintain and significantly depletes local water supplies when drawn for cooling. The average sized data center is estimated to consume 200 terawatt hours per year – more than the total annual energy consumption of certain cities [44]. Such centers can be made more energy efficient by repurposing heat typically captured in cooling systems. Another way to reduce the environmental impact of storing so much data is to investigate more efficient storage solutions, the future of which may lie in memory encoded by quantum states instead of binary code.

3.4. Training

Alongside the rapid development of healthcare AI has been a parallel increase in programming and robotic courses to enable medical staff to implement technologies in their own clinical practice [45]. However, this educational rollout is unlikely to have kept up with the speed and breadth of AI advancement, nor is it equal among healthcare staff. Indeed, there remains a lack of integration of health informatics competencies in both undergraduate and surgical postgraduate training curricula representing a bottleneck of implementation where only the select few among surgical residents are AI or robotically trained.

Of equal importance is ensuring adequate scientific literacy in the neurosurgical community to competently critically appraise papers using AI. Methodologically flawed studies may not rigorously separate training data, test data, and validation data, potentially resulting in biased or flawed conclusions (see Table 2 for definitions of key terms). Professionals, researchers, and academics should approach these papers with critical literacy, scrutinizing methodologies, and results, ensuring the reliability of insights and maintaining high standards of healthcare and surgical practice.

Table 2. Glossary of terms.

Glossary Term	Definition
Algorithm	Instructions for computer to complete a task [46]
Artificial Intelligence	Providing computers with algorithms which enable them to perform tasks requiring human-like intelligence including learning and problem solving [4,9]
Artificial Neural Networks (ANNs)	Non-linear data modeling inspired by biological nervous system structures. ‘Signals’ are processed in layers of computational units called ‘neurons’. Input signals are mapped to output signals via hidden dynamic variable connections between layers [4]
Black Box Problem	The inability to understand how or why a machine learning model arrives at its decisions due to its inherent hidden layers of processing [4]
Clinical Decision Support Systems (CDSSs)	AI systems used to improve accuracy and pace in diagnoses and treatments such as in MDTs [20]. Through a classification model, utilizing data mining techniques, time spent in decision-making can be reduced as medical professionals have access to updated medical findings in a knowledge database through smart applications.
Computer Vision	Automated acquisition, processing and interpretation of images [46]
Deep Learning	Multi-layered neural network with increasingly higher order representation of objects in each layer. One approach is using unsupervised learning to first find robust features which are then refined and used as features in a final supervised model [4,46]
Feature	Variables that are of significant effect size to be included in the model, e.g. age at diagnosis, gender, IDH mutation status
Frame Problem	The difficulty in identifying and updating a set of axioms to properly describe the environment for autonomous agents [47]. In other words, algorithms may encounter difficulties in adapting to the variability in clinical presentations not encountered in the training process [24]
Machine Learning	Domain of artificial intelligence involving the process of computers learning from data [4]
Model	A mathematical representation trained from a given set of data that can be used to make predictions.
Natural Language Processing (NLP)	Domain of AI for computers to understand meaning and sentiment from human language in the format of unstructured data [4]
Reinforcement learning	Algorithms are trained using trial and error through a system of reward and punishment facilitating efficient task completion [24]
Supervised Learning	Computer learns to either classify or predict from labeled data with a known output [4]. An example of this would be a system learning to classify brain tumors from images labeled by a human.
Unsupervised learning	Computer finds naturally occurring patterns or groupings within the data which has no known output [4]. This would be of use to find patterns in data not known by the human supplying the data such as identifying genetic mutations associated with worse prognosis in glioblastoma patients.

4. Expert opinion

The integration of AI into neurosurgery signals a transformative era characterized by precision, innovation, and enhanced outcomes for patients with brain or spinal disease. AI’s ability to learn, understand, and predict is redefining surgical practices, from robotic-assisted procedures to tailored personalized treatments. Within neurosurgery, AI’s influence is already tangible, optimizing patient pathways through pre-operative planning, intraoperative navigation, and post-operative predictions. As technology continues to advance across both machine learning and robotic implementations, ethical considerations including data biases, accountability, cost, and resource management are crucial focal points for a responsible AI-driven future. Fast, accurate large scale predictive algorithms, intelligent operative environments and advanced brain-computer interfaces represent just a handful of the prominent motifs within the AI landscape, showcasing its substantive potential as AI reshapes the boundaries of neurosurgical healthcare.

In our view, artificial intelligence will continue to expand within neurosurgery, infiltrating almost every aspect of the surgical patient pathway at pace. With ever-increasing datasets that harness more manifold data types, successive iterations of AI models will become more precise and will adopt a greater role in surgical decision making. Whether neurosurgeons are forced to adapt in light of evidence demonstrating greater system efficiency and better patient outcomes, or if they choose to become partners and innovators in the adoption of these technologies remains uncertain. This represents a critical challenge for healthcare leaders and surgical educationalists. Certainly, initiatives aimed at improving digital literacy and AI awareness among surgeons will go some way toward bridging the knowledge gap and help improve the appraisal and application of these novel tools. If, however, we do not actively embrace AI implementation and digital training, patients will simply face the prospect of care in a two-tier system. Neurosurgical departments and staff will either be classed as those which are AI-capable and those which are not.

Declaration of interest

The authors have no relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock ownership or options, expert testimony, grants or patents received or pending, or royalties.

Reviewer disclosures

Peer reviewers on this manuscript have no relevant financial or other relationships to disclose.

Additional information

Funding

This paper was not funded.