AI-Powered Precision: Transforming Neurosurgical Practice through Intelligent Technologies

3.1 Molecular tumour subgroup differentiation

Brain tumours are lethal among all cancer types and have sustained a mortality rate of 16,606 deaths per year in the United States in 2018. Glioblastoma, which is a diffuse glioma, is the commonest primary brain tumour with less than 2 years of median survival [6, 9, 10]. Brain tumours differ from other neoplastic tumours of the body in several aspects; mainly, definite tumour histology can only be obtained intra- or post-operatively due to the unavailability of reliable pre-surgical histology. Mostly, during pre-surgical assessments, it depends on the serial magnetic resonance imaging (MRI) interpretation of the radiologists regarding the tumour site and probable type. Intraoperative histology samples are transported to laboratories, processed and interpreted by pathologists, and usually take 20 to 30 minutes, which is a potential barrier to providing timely and effective surgical care for patients [11].

However, several brain tumours carry different molecules, which can be used as molecular markers of diagnosis, thus enabling early diagnosis, surgical decision-making, and selection of appropriate chemoradiotherapy. However, it demands immunohistochemistry, cytogenetic tests, and sometimes next-generation sequencing, which are costly and unavailable in many centres. Expert knowledge in interpreting these patterns is also challenging due to declining human resources [12]. However, some kind of predetermination of tumour type is essential before proceeding to the surgery, and AI-based molecular classification of brain tumours emerged as a more reliable modality [12].

In most of these AI systems, optical imaging methods are used, and the most common is simulated Raman histology (SRH). It produced label-free images of unprocessed biological tissues rapidly in sub-micron resolution, which is far more reliable and faster than conventional methods [11]. The biological properties of lipids, proteins and nucleic acids are used to generate images with optimised diagnostic microscopic features of the tumours while eliminating artefacts of conventional slide preparation [11].

Either way, the generated histological images should be interpreted by experts in histopathology, which takes time and may not be readily available during neurosurgeries. So, AI-based machine learning (ML) and deep learning (DL) come into play. Deep learning simulates complex neural networks and contributes to image analysis, which is usually more accurate and timelier than humans. DL consists of systems such as convolutional neural networks (CNN), recurrent neural networks (RNN), and long short-term memory (LSTM) [6].

AI-based molecular diagnosis of gliomas has been employed by combining an optical imaging method with a deep learning-based imaging classification system. The optical imaging method stimulated Raman histology (SRH), and the combined system was called DeepGlioma. The system was pre-embedded with model data on SRH image features of adult-type diffuse gliomas, and it had an average molecular genetic classification accuracy of 93.2% and can identify sub-types of diffuse gliomas with an accuracy of 91.5% [12, 13].

Similarly, combined systems have been employed to obtain near real-time bedside diagnosis of brain tumours during neurosurgical procedures. Here, SRH has been combined with deep convolutional neural networks (CNN). The CNN system has been trained to interpret more than 2.5 million SRH images, and it classifies tissues into the 13 most common brain tumours in less than 150 seconds during the surgical procedure with an accuracy of 94.6% [11, 13]. This revolution enables intraoperative decision-making of tumour resection (gross or subtotal resection) and enhances patient outcomes.

Despite several points lacking in conventional MRI imaging of brain tumours, the combined MRI radiomic with genetic expression may predict more accurate details of biophysical characteristics of MRI images, which is capable of embracing the heterogeneity of tumour types. The system is embedded with data, including pre-surgical MRI images of brain tumours, genetic details of definite post-surgical tumour diagnosis, tumour location and size, and data on patient outcomes and survival [6]. As mentioned, gliomas are the most aggressive primary CNS tumours and the clinical outcome highly depends on the extent of initial resection of the tumour and the clinical response to chemo-radiation. So, combined radiomic assessment of MRI features of AI has successfully identified several brain tumour sub-types, enabling pre-surgical planning of extension of tumour extraction and suitable chemo-radiation [6, 9, 14].

3.2 Intracranial aneurysm detection

Cerebral aneurysms are dilatations of cerebral arteries commonly found in the anterior circulation of the brain. However, the majority is silent, which is usually found accidentally on neuroimaging or autopsy, and some may result in subarachnoid haemorrhages (SAH) following rupture. The prevalence of unruptured cerebral aneurysms (UCAs) was 7.0%, with female predominance in China [15, 16], and the overall global prevalence is around 3.2%, according to data from 2023 [17]. Aneurysms are accountable for 85% of SAH and carry high mortality and morbidity, and surgical and endovascular managements also carry a risk of 3–10% of stroke or death [7]. The management of an aneurysm includes detection, estimation of risk of rupture, prediction of complications, decision of treatment strategies, and evaluation of the risk of recurrence. So, with the evolution of neurosurgery along with radiology, early identification and intervention have been assessed in several aspects, and AI technology has been incorporated into all of these aspects.

High-risk individuals may undergo screening either with magnetic resonance imaging (MRI), computerised tomographic angiography (CTA), or conventional angiography for early identification of cerebral aneurysms [17]. In the case of SAH, computed tomography (CT) is considered to have 100% sensitivity when performed within 6 hours of the onset of the symptoms; however, it has declined drastically over time [17]. The bleeding point of SAH can be identified using CTA, magnetic resonance angiography (MRA) or digital subtraction angiography (DSA). However, small points can be easily missed [7]. Either way, radiologists clinically evaluate these studies, which always carries an added human error.

Computer-assisted diagnosis (CAD) of aneurysms has been implemented using AI, conventional-style CAD, and deep-learning CAD systems. Conventional CAD is based on pre-fed clinical features of neuroimaging, such as curvatures of the cerebral vessels and thresholding. It has a sensitivity of around 84–100%; however, remarkably higher false positives (2.4 FPs) were noted due to the bending or branching points of the blood vessels being the major issue. Identifying cerebral aneurysms smaller than 5 mm is also challenging and shows a lower sensitivity. In contrast, deep-learning CAD systems are based on MRA and have achieved a sensitivity of 90–93%. The identification of aneurysms on MRA takes less time and is more accurate than conventional interpretation by radiologists. However, the major barrier here is the higher rate of FPs [7, 18, 19].

The HeadZNet model was developed using CTA images and a neural network segmentation model. It achieved greater sensitivity and accuracy than the above two systems in identifying aneurysms greater than 3 mm in size [7]. During their post-stroke work-up, CTA is performed on all patients who have developed acute ischaemic strokes. Thus, CAD systems associated with CTA are more practical and beneficial in clinical settings. Further validation and generalisation are needed.

Following the identification of aneurysms, stratification of the risk of rupture is needed in deciding further management, neurosurgical, or conservative. Treatment-related fatality is higher in some scenarios compared to conservative management, while the risk of rupture is low [20, 21]. When calculating the rupture risk, the size and shape of the aneurysm and the hemodynamics of the blood flowing through the aneurysm are important. Conventionally, radiologists make manual measurements of the size and shape of aneurysms, which are crucial in risk stratification; however, there are inherited intra- and interobserver variations and bias. Computer-assisted 3D recreation of aneurysms and convolutional neural network systems of DL can be employed for these risk stratifications. The AI-based automated morphologic analysis is a system that calculates the risk of rupture of cerebral aneurysms. It identifies and isolates the aneurysm sac, enabling easy identification of the neck of the aneurysm by analysing the vascular skeleton and overcoming human errors and bias [7, 22, 23]. Rapid assessment kits such as ‘A View’ can be used. However, more trials and modifications are needed before application in routine clinical practices.

The hemodynamics can be calculated by computational fluid dynamics modelling, which is used in modern practice; however, it requires human interpretation, which takes time and makes it challenging to use in real time. Automated systems using 3D angiograms have been proposed. However, further evaluation is needed before use in the clinical setting [7]. Similarly, AI neural network systems and multi-task AI have been studied in the aspect of early prediction of complications of aneurysms, such as SAH, and have proven to perform excellently compared to conventional methods.

3.3 Neurological condition classification

Extensive clinical assessment and radiological imaging are essential to make a diagnosis of a neurological condition. The identification of neural structures and the function of the brain has been comprehensively analysed using advanced technology and several conditions affecting the neuronal system. With the revolution of technology, several AI-related automated diagnostic systems have been developed to diagnose several neurological conditions. These conditions can be broadly categorised into seven major aspects: tumours, seizure disorders, disorders of development, neurodegenerative disorders, headaches, and facial pain, cerebrovascular accidents and neurological infection [24, 25].

Convolutional neural networks (CNN) are the second-generation deep neural networks inspired by the architecture of the human brain, which transform visual inputs into complex systems, ultimately resulting in object recognition. This technology can be employed to recognise brain changes related to several neurological conditions, and such systems governed by AI have been introduced. Such as artificial neural networks (ANN), multilayer perceptron (MLP), recurrent neural networks (RNN), convolutional neural networks, reinforcement learning (RL), deep reinforcement learning, and spiking neural networks (SNN). These systems have a higher capacity for analysing unknown data and come to certain diagnoses in contrast to conventional methods. According to the latest WHO data, AI-based applications have a higher chance of predicting certain conditions correctly than specialists [25, 26].

MRI is the standard imaging modality for brain tumour recognition. Then the specialists, including radiologists and surgeons, came up with a possible diagnosis using their knowledge and experience. However, definitive identification of tumours and boundaries may be difficult in some instances. That being said, rather than conventional analysis by radiologists, deep and machine learning algorithms, such as CNN-based multigrading brain tumour classification systems, can be used. Based on pre-fed data, this system analyses the MRI images, detects brain tumours and performs segmentation and tumour grade classification. According to the latest available data, a 23-later CNN system can classify brain tumours with an accuracy of 97.8–100%. In the diagnosis of chondrosarcoma, AI-based algorithms have an accuracy of 90% (correctly disposing nine out of 10 cases), and neurosurgeons have between 40% and 85% accuracy. However, the image rotation affects the analyses, and a large amount of data is required for the system training. These two limitations have been overcome by vision transformer (ViT) neural networks [25, 26, 27].

Epileptic seizures develop due to sudden abnormal electrical waveforms of the brain, which can be detected using electroencephalography (EEG). AI-based systems which extract definite features of single-channel EEG analyse the waveforms using MLP systems or ANN and recognise the seizure activity. Similarly, two-dimensional deep convolutional autoencoders can also detect abnormal electrical activities using neural networks. Out of which, temporal lobe epilepsy (TLE) is the commonest, and AI have a proven accuracy of 95.8% in diagnosing TLE, compared to physicians (66.7%) based on conventional MRI data [25].

Genetic conditions like cerebral palsy, Down syndrome, autism spectrum disorder (ASD), fragile X syndrome, and attention deficit hyperactivity disorder (ADHD) lead to intellectual and developmental disabilities in children and usually can be identified before 18 years of age. However, early identification may aid in the proper management of these conditions with more promising results than late interventions. It may help in improving the quality of life of these children and high parental satisfaction. However, early identification is difficult, and it highly depends on the parents’ alertness regarding their child. Deep neural network (DNN)-based AI algorithms can utilise large amounts of data, including genetic data, health, and clinical records, neuroimaging and other investigation data, and behavioural patterns of already diagnosed children with these conditions. With critical analysis, AI can identify children with intellectual disability (ID), developmental disability (DD), or both. ASD can be identified by screening several data sets, such as fMRI images using the DNN system, detecting toddler’s eye movements, and by detecting maternal auto-antibody-based biomarkers. Similarly, ADHD can be identified by deep analysis of EEG data using support vector machine (SVM) algorithms [25].

Alzheimer’s disease (AD), Parkinson’s disease (PD), motor neuron disease (MND), and frontotemporal lobar degeneration are the major neurodegenerative disorders of concern, and early recognition is crucial but challenging. The symptoms emerge, and MRI features are visualised only after a considerable number of neurons are damaged. Thus, ML algorithms can analyse whole-brain anatomical MRI images with the support of a support vector machine (SVM). For the identification of AD, 3D neural networks are used to distinguish AD from FTLD and AD patients from healthy individuals with greater accuracy than conventional methods [25]. The mild cognitive impairment (MCI) can later be transformed into AD. Thus, the AI-based random forest system can predict this transformation 3 years before the clinical diagnosis by analysing brain MRIs. This is a great leap in neurology as early detection of these conditions determines the treatment progress and the prognosis. Then, conditions like motor neuron disease (MND), Huntington’s disease (HD), and Parkinson’s disease (PD) are associated with motor neuronal dysfunction. So, a combination of simple motor functions, such as drawing with machine learning, has been used to identify PD. Then, machine learning systems like SVM, Bayesian, and nearest neighbour have employed the gait disturbance, which can be seen prominently in patients with PD, HD and amyotrophic lateral sclerosis (ALS) to distinguish these three conditions [25, 28, 29].

Then, cerebrospinal accidents (strokes) are the most common cause of neurological deficits in patients, which result from interrupted cerebral perfusion, mainly due to atherosclerosis. Early treatment is paramount within 3 to 4.5 hours of the onset of symptoms. Otherwise, prolonged hypoxia to brain matter results in permanent neurological deficits or even death. AI-based algorithms such as SVM, random forest, and ANN can be used to detect sub-types of strokes, namely ischaemic strokes and haemorrhagic strokes. MRI image analysis using the SMV method can achieve nearly 95% accuracy in diagnosing acute ischaemic strokes. Similarly, stratification of stroke severity and risk of after-effects, such as cerebral oedema, can be done by AI-based algorithms [25, 30, 31].

Lastly, neurological infections are difficult to acquire, but once acquired, they are difficult to treat. The causative agents may be bacterial, viral, or rarely fungal. Most of the infections, regardless of the cause, show similar symptoms, making them difficult to diagnose. Bacterial and viral meningitis show symptoms like headache, neck stiffness, and fever and constitutional symptoms like nausea and vomiting. Differentiating between these two is crucial in deciding the management. Radiological investigations such as computed tomography (CT), magnetic resonance imaging (MRI), blood investigations and cerebrospinal fluid (CSF) analysis using lumbar puncture aid in arriving at a diagnosis. However, these conventional methods may take time; some are invasive and may stress out the patients. So, naval AI-based algorithms fed with laboratory parameters such as neutrophil and lymphocyte count of cerebrospinal fluid (CSF), neutrophil-to-lymphocyte ratio (NLR), serum albumin, c-reactive protein (CRP), glucose and CSF lymphocytes-to-blood CRP ratio (LCR) can predict the type of diagnoses with much higher accuracy than conventional invasive methods. Similarly, neonatal sepsis with or without meningoencephalitis also has high mortality rates, whereas early diagnosis with clear differentiation is important. AI-based machine learning systems have achieved this task with high accuracy using metabolomics data of sepsis. AI-based algorithms with the capability to predict the risk of prognosis of several CNS infections are in use. Maternal infection may adversely affect fetal brain development and may result in autism spectrum disorder (ASDs). Conditions like autoimmune encephalitis ultimately resulted in cognitive dysfunction, reduction in consciousness and speech disorders. So, early risk prediction may aid in avoiding these consequences [25, 32, 33, 34, 35].

4.1 Virtual surgical simulations: Creating patient-specific anatomical reconstructions

With the advancing neuroimaging techniques, diagnosis and determination of treatment options have evolved, and patient prognosis has improved. Even though we made a tentative diagnosis, some complex neurosurgical conditions remained inoperable due to a poor understanding of the surrounding anatomy of the pathology until the patient opened up, and some are located in hard-to-access regions, making neurosurgeons hesitate [36, 37].

AI-based, several 3D models have been employed, such as immersive surgical simulations through augmented reality (AR), virtual reality (VR), Mixed Reality (MxR), extended reality (XR), and 3D printing applications. These facilitated the reconstruction of 3D models of neuroanatomy by sophisticated algorithms using 2D radiological images. Magnetic resonance imaging (MRI) and computerised tomography (CT) are used as baseline imaging. MRI T1-weighted images are used to map the external anatomy of the skull, and both T1W and T2W MRI sequences are used to define the grey and white matter of the brain. Then, the vascular network is reconstructed using CT angiograms, contrast-enhanced MR venograms, and time-of-flight (TOF) MR angiograms by blood flow dynamics. These baseline data undergo segmentation, 3D reconstruction, model refinement, and ultimately integration into AI systems, giving rise to 3D models. Besides this visual simulation, AI-based automated modelling and reconstruction of full anatomical structures based on patient-specific data is immensely beneficial in deciding surgical options [38, 39].

The guidance provided by these 3D, 360-degree imaging continuum from the initial patient consultation, pre-operative planning of neurosurgical treatment, intraoperative navigation throughout the surgery, and, ultimately, post-operative consultations. This is especially useful in complex vascular pathologies, as it defines their correlation with surrounding structures, which is challenging to achieve via 2D images [37, 40, 41]. Similarly, lesions of the skull base are hard to reach and challenging with complex neuroanatomy and critical structures [38, 42].

VR is one of the systems that can provide a three-dimensional (3D) 360-degree imaging guide of neurosurgical pathology [38, 43]. The surgical theatre visualisation platform provides personalised, 3D, and 360-degree VR of the lesion, giving a better understanding of the site of the lesion and surrounding neuronal anatomy. This model can be used to provide an understanding of their condition and possible surgical options to the patient and their families, which is hardly achieved by 2D images. Proper informed consent can be obtained, and higher patient satisfaction has been reported [44]. The surgical rehearsal platform (SRP) analysed eight imaging modalities and reproduced a live 3D model accurately, displaying even surgical tools and anatomy around the pathology, providing a better chance of panning complex neurosurgeries and rehearsal of several possible approaches. Similarly, pre-operative modifications can be employed in pre-planned management. SRP utilises individualised volumetric scans of patients in the format of Digital Imaging and Communications in Medicine (DICOM), and images are processed by incorporating multiple imaging modalities. Diffusion tensor imaging (DTI) is performed only in areas of interest. This holistic approach enables neurosurgeons to have a pre-surgical idea about the per-patient brain anatomy and anticipate possible barriers and shortcomings during surgery [38, 39].

Complex pathologies like meningiomas, chondrosarcomas, chordomas and sinonasal malignancies extend into the anterior skull base and can be approached by advanced AI-based endoscopic endonasal approach (EEA) with 3D modelling. Reconstruction provides a real-time guide with manipulation, and incorporating the AR system enables precise tumour extraction. This combined approach has reported a reduction of cerebrospinal fluid (CSF) leakage rates from 40% to 2.9%, which is a great achievement. The reduction of injuries to the internal carotid artery (ICA) is from 0.9% to 0.3%. A significant reduction of cranial nerve (CN) dysfunction has also been achieved, remarking the benefit of advanced technology in patient safety with surgical precision [38, 45].

4.2 Predictive modelling: Analysing complex medical datasets to anticipate surgical challenges

Especially in neurosurgery, accurate surgical decision-making by analysing risk-benefit profiles is crucial to deciding whether to proceed with neurosurgery or manage conservatively. Unnecessary surgical risk may increase neurosurgical fatalities and lifelong disability. Similarly, one surgical curative treatment may lead to another unfavourable consequence. Ideally, before deciding on definitive management, prediction of risk of surgery, survival following surgery, improvement of symptomatology, and possible adverse effects should be taken into consideration. Clinicians can make predictions based on research data. However, individualised prediction is challenging due to the heterogeneity of these neurological conditions. So, AI-based machine learning (ML) systems come into play for these predictions in a more customised manner [46, 47].

ML is increasingly used in clinical research. This algorithm can be fed with a large set of data from different population groups, which is comprehensively analysed and provides a clinically meaningful relationship with the provided input and the output data. This patient data includes genetic details, medical history and imaging study results. As they learn from experience, previously unknown relationships can also be predicted using this data [46]. It has proven beneficial in understanding complex neurological disorders, similarly predicting long-term outcomes and disease progression, and establishing personalised patient management plans [48]. Also, high patient satisfaction can be obtained through this integrated approach. ML has predicted the outcome of several neurosurgical conditions, including brain tumours, traumatic brain injury, spinal lesions, hydrocephalus, and neurovascular disease, enhancing clinical decision-making better than some established prognostic parameters and enabling informed best practices [39, 46, 48].

Adult spinal deformity (ASD) is a condition with higher rates of post-operative complications. No guarantee can be given to ensure patients’ intact baseline neural functions following surgery. So, in such conditions, shared decision-making by the patient and the clinician, depending on the predictive data provided by the AI system, is beneficial. Such systems have shown 87% accuracy in predicting major intra- and perioperative complications of this surgery, enabling point-of-care decision-making [49].

5.1 Navigation tools: Real-time AI-assisted guidance systems

As discussed, the pre-surgical simulation of neuroanatomy via AI-based tools is trending now. Complex AI-based tools also achieve AI-guided real-time navigation throughout the surgery. Several AI-based navigation tools are in use, and they have improved surgical precision with a drastic reduction of intraoperative complications. The surgical navigation advance platform (SNAP) is one of a kind, and it provides 360-degree navigation of surgical tools and the real-time manipulation of underlying pathology throughout the surgery. Collectively, this approach is beneficial in achieving a safe resection of pathology with minimal neurological damage [37].

According to global statistics, skull surgeries are more difficult to perform than surgeries elsewhere. To overcome this barrier, an AI-based Markov model is used. This system recognises when surgeons have completed the manual manipulations of the surgery and implements the next task in order. Then, surgeons can safely move on to the next step of surgery with greater efficiency while minimising intraoperative errors [26, 38, 39].

Real-time tissue diagnosis by AI-based biopsy analysis during surgeries is immensely important in managing neurosurgeries. For conventional frozen section analysis, biopsy samples should be sent to laboratories and should be analysed by experts, requiring time. However, AI can accomplish this task in less than 3 minutes intraoperatively. Surprisingly, the accuracy of AI diagnosis is almost 100%, while conventional methods carry a margin of error. This is a greater leap in the surgical aspect, which may entirely change pre-determined treatment plans in some patients [26].

5.2 AI-assisted cellular identification with unprecedented accuracy

Cells are the basic unit of life, with which all living beings continue. Still, scientists are trying to figure out the dynamic function and behaviours of cells of viruses, bacteria, and human beings for future predictions. Virtual cell models are continuously on their way to development; however, most of them are based on assumptions made from observations. The biological system is multiscale, comprised of atoms, molecules, and cells, and building up the histological appearance, giving rise to various functional properties which is non-linear in one scale than another. Thus, the transformation through time and space is unpredictable. Similarly, one biological process is a result of multiple interactions such as genetic regulation, metabolic pathways and dynamic signalling and response cascades. A minute change in the input process may result in a greater dynamic change in the output, non-linear dynamics [50].

The virtual cell is an AI-generated dynamic 3D model of cells which is simulated by biology, physiology, several image-based models, and evidence-based graphs. This may function as a real cell while facilitating new research approaches, as cell-based medicine is the core of regenerative medicine. Bioengineers can produce microorganisms that can produce clinically important substances such as new drugs. Similarly, by analysing individual genomic data, creating personalised cell models may enable the creation of tailor-made treatment strategies for each patient [50, 51].

In this context, innovative cell-based approaches are being explored instead of traditional methods such as radiotherapy, chemotherapy, and surgery for neuronal tumours. Numerous phase 2 clinical trials investigate effective and safe treatments for malignant brain tumours through genetic therapies and immunotherapies. Gliomas are becoming more and more resistant to conventional therapies and have a high recurrence rate, most probably due to the presence of glioma stem cells (GSC). So, stem cell targeted therapy is emerging. Neural precursor cells (NPC) can migrate throughout the brain and differentiate into neurons, glial cells and oligodendroglia cells. This ability has been combined with novel technology and selectively delivers drugs to the tumour site [52, 53, 54]. Retrovirus-guided transfer of specific important genes into the NPCs has been tested. Then, NPCs with the ability to produce interleukins, lysosomes, and cytotoxic mediators such as tumour necrosis factor (TNF) aid in cancer cell apoptosis and selective antagonism of glioblastoma cell invasion [52, 55]. Similarly, novel treatments based on non-technologies and translational medicine are in the way of approaching the conditions which have been considered as incurable, spinal cord injuries [50, 52].

5.3 Robotic surgical support: Minimising tissue damage and improving surgical outcomes

As discussed, AI may have supportive aspects in managing neurosurgical patients from the diagnosis of a condition up to discharge. Intraoperative navigation is an advanced technology developed around AI that enables safe neurosurgery-related decision-making during surgery, thus minimising intraoperative errors. With the demand for minimally invasive approaches to the brain and spine, AI-based robotic devices come into play. The first reported case of robotics was the use of the PUMA (Programmable Universal Machine for Assembly) 560 robotic system in mid-1980 for a neurosurgical biopsy, and the first approved robotic device used in neurosurgery was the NeuroMate robot (Integrated Surgical Systems). Especially in fields like neurosurgery, where surgeries are done within a limited space, this technology is used mainly for anatomical localisation of the pathology and brain structures, anatomical planning in the assessment of deep brain structures, in the stabilisation of the surgeon’s hand and spinal surgical procedures like pedicle screw placement. Some commonly used robotic systems in neurosurgical fields are the Da Vinci, Robotic Stereotactic Assistance (ROSA), SOCR ATES surgical system, and these robotics have mainly been used in surgeries involving optic chiasmata, subthalamic nuclei, cerebral vascular networks and the pituitary gland Neuromata, Pathfinder, NeuroArm, SpineAssist, and Renaissance neurosurgical system also some examples [50, 51, 52, 53].

These robotic systems can be categorised into three main types: active, semi-active, and master-slave. NeuroArm is a type of master-slave robotic, which the surgeon can remotely control. This system is dependent on the surgeon’s input and is unable to make independent decisions. It is an MRI-compatible robotic with 8 degrees of freedom (DOF) and the capability of mimicking a neurosurgeon’s arm functions but in a more precise manner. Up until now, it has been used in MRI-guided tumour biopsies, hematoma evacuation procedures, and microsurgical dissections [54].

Semi-active robotic systems integrate pre-programmed data with the surgeon’s inputs and function in a hybrid manner. In conjunction with the skills of the surgeon and AI knowledge of the robotic, more precise results can be obtained in this manner. The Steady Hand System is such an example, and it can allow surgeons to expedite fine dissections without fatigability. The ‘evolution 1’ system facilitates fine endoscopic procedures. The ROSA system can analyse MRI or CT images and pre-plan the placement of electrodes and biopsies, avoiding important vascular and neural structures of the brain. This system has only 6 degrees of freedom, and this assistant system is used for deep brain stimulation (DBS) to the bilateral subthalamic nucleus in patients with Parkinson’s disease. This system has shown excellent accuracy in this function. Then, these pre-plans are loaded onto the robot, and throughout surgery, the navigation and intraoperative images can be obtained by an optical device in a robotic arm. As mentioned, this system is commonly used in stereotactic procedures, such as the placement of microcatheters in gliomas for the introduction of chemotherapy, insertion of DBS electrodes and multiple-depth electrodes for stereoelectroencephalography (SEEG). Then, the robotic-guided endoscopic cerebral hemisphere resection can be done with a combined application of an endoscope to the robotic arm ROSA, especially in children with drug-resistant epilepsy. Some other similar applications of endoscopic robotics are endoscopic intranasal therapy, endoscopic third ventricle patency, and intraventricular tumour resections [8, 26, 55, 56, 57].

Then, the active robotic systems assist surgeons in carrying out several tasks in a more precise manner and can conduct pre-programmed tasks autonomously. PUMA, SpineAssist, and Renaissance systems are such examples and they are widely used in instrumentation processes like guided biopsies and pedicle screw implantations [52, 54].

The Da Vinci system has 7 degrees of freedom, is an HB imaging system with three operating arms, and uses the transoral robotic surgery (TORS) approach. This TORS approach is mainly used for manipulations in the area of the anterior skull base as an open surgical approach in this area carries a significant risk of mortality compared to head and neck surgeries. Surgeries in the anterior skull base and sellar region involving manipulation of optic chiasmas and pituitary glands, such as in the cases of pituitary adenomas and cystic sellar masses, are hard to perform due to limited accessibility. Whereas the TORS approach using robotics makes it easy [26, 50, 52, 55].

6.1 Big data analytics

AI is used in almost all areas of medicine, and most of these novel interventions come into play as a result of research. With the electrification of the medical record system, a large amount of patient data has come into our access within finger distance. So big data analytics can be done using AI-based algorithms using these data to identify novel biomarkers, novel treatment strategies, and especially in diagnosing neurological conditions using patient patterns and trends.

Novel drug invention is an important aspect of medicine, but the underlying process is expensive and takes a considerable amount of time. This process starts with research, and then the development of a new drug leads to the conduct of clinical trials, and finally, the drug can be released into the market. The average estimated investment for this whole process is about $1.3 billion per drug [58], and the median time taken to develop a drug ranges from 5.9 to 7.2 years for non-oncological drugs, while it takes 13.1 years for an oncological drug. Even with this great afford, only 13.8% of clinically trailed drugs get the approval [59]. So, with the development of technology, large patient-based data collections have been created, and AI and Machine learning (ML)-based algorithms analyse these enormous amounts of data at once and understand the echo-biological system, thus enabling the development of novel drugs for new indications. AI has been used in developing new drugs for decades, and the latest was during the COVID-19 pandemic. And owing to AI’s prediction ability, research study designs and analysis can be planned, thus enabling the best possible clinical trial with high approval rates. This has made drug discovery more efficient and cost-effective by minimising cost and time and enhancing the probability of approval [60].

6.2 Machine learning in neurosurgical research

As mentioned previously, machine learning has branched out in all aspects of the neurosurgical field, from the first consultation of the patient up to post-surgical care. AI is used to diagnose clinically challenging neurosurgical conditions such as aneurysms and brain tumours and to predict the risk of several critical conditions. Similarly, pre-operative planning, intraoperative guidance, and navigation are great leaps in neurosurgery and enable safe and more accurate neurosurgical procedures. All being said, these promising inventions came through research [61].

6.3 Predictive healthcare strategies

AI is mainly used in diagnosing medical conditions but can also predict patient outcomes, risk of recurrence, probability of survival and risk of complications in several challenging conditions [8]. Several ML algorithms have been invented for such specific tasks as follows.

Traumatic brain injury (TBI) is a major neurosurgical emergency resulting in death or lifelong disability worldwide. In resource-poor settings, the situation is more pathetic due to a lack of trained, skilled staff and limited access to diagnostic technologies. So, to support timely and objective clinical decision-making, prognosis-related predictive models have been tested in low- and middle-income (LMIC) countries. For the development of this ML model, data including socio-demographic details, details of initial assessment, type of injury, vital signs, and treatment received have been used. The discharge score of the Glasgow scale has been used as the outcome variable. This prognostic tool has achieved an average of 66.2% to 86.5% accuracy in predicting the outcome of TBI patients. With further research, this ML-based predictive model can be implemented in resource-poor settings all over the world for a better triage process and to reduce TBI-related mortality [8, 62].

Then, lumbar spinal stenosis (LSS) is a common health concern among the ageing population. Usually, conservative measures are recommended, but some may need a surgical approach. Even though the indication remains the same, the surgical outcome varies greatly among individuals. The risk of adverse outcomes following surgery should be considered individually; however, it is not feasible. So, ML algorithms have been used to generate prediction models for short- and long-term outcomes following decompression surgery for LSS. The length of surgery (duration more than 45 minutes) can be predicted with an accuracy of 78%, and the accuracy of predicting re-operation varies between 63% and 69%. The extended hospital stay was predicted with an accuracy of 77%. These pre-surgical predictions give an insight into the surgical procedure and enable personalised decision-making regarding management and shared decision-making [8, 63].

Hospital re-admission following surgery significantly increases the cost of healthcare setting. Thus, a predictive model with ML has been tested to predict the re-admission following lumbar laminectomy using the database of the American College of Surgeons National Surgical Quality Improvement Program (ACS NSQIP) to develop interventions to minimise these re-admissions. Notably, the predictive model has accurately predicted over 95% of re-admissions using all available data of the patients while achieving over 79% accuracy using only pre-discharged data of the patients [8, 63].