In medicine, AI seeks to emulate the characteristics of human intelligence, such as learning, communication, decision-making, and adaptation to changing environments in a clinical setting. Thus, it includes diverse domains, of which ML is of particular interest and has major applicability in medicine (Fig. 2). ML is a subset of AI that uses computational models to learn large complex data and generate useful predictive outputs without any explicit programing. The fundamental principles of ML are based on intricate statistical and mathematical optimization using numerical data [5]. Decision tree learning, support vector machines (SVMs), and artificial neural networks (ANNs) are three common ML models. These models are based on supervised learning methodologies in which the machine is trained to predict the outputs based on previously collected inputs for which the outputs are known; hence, a “labeled” dataset is used [6]. A decision tree is a graphical representation technique that provides a visual depiction of various decisions with the base knot and various branches attached to the knot and ending in terminal nodes (Fig. 3). Because of their peculiar structure, they are easier to interpret by clinicians. The SVM is adapted for nonlinear classification and regression, considering each input as a multidimensional vector. SVM algorithms incorporate decision boundaries called hyperplanes (Fig. 4). In spine pathologies, disk degeneration grading and identification of the type of scoliotic curves can be performed using SVM algorithms. In the ANN model, the machine integrates data using the linkage of artificial neurons. The information flows through several layers of the hidden neuronal network; finally, the processed output is generated (Fig. 5). This ANN multilayered strategy is also known as “deep learning” [7]. Other AI tools include natural language processing (NLP) and computer vision [8]. NLP involves mining and analyzing text-based data that can be incorporated into analytic algorithms [9]. In simple terms, NLP quickly extracts patient data from electronic records, including lifestyle and several other health determinants that could otherwise be cumbersome for human evaluation. Computer vision is another advanced tool that aids in the evaluation and processing of patient imaging data with higher precision and clarity. It often uses convoluted ANN pathways for image segmentation and analysis. U-Net is an example of these networks that are primarily used in neurosurgical imaging [10].

By using these specialized tools, AI facilitates and augments patient healthcare by assisting clinicians in four important aspects: diagnosis (provides inference based on patient data including signs, symptoms, and investigations), therapy (aids in surgical decision-making and enhancing postoperative patient care), prognosis (enables the surgeon to predict patient-specific outcomes and modify treatment strategies), and research (with the available patient data, AI can enhance complex statistical analysis to innovative ideas and research in spine care).

Numerous studies in AI regarding various aspects of low back pain have been published [11–14], especially disk-related pathologies. Reports have shown that factors such as disk degeneration, endplate defects, Modic changes, and vertebral osteoarthritis can be automatically detected in magnetic resonance imaging using AI algorithms such as SVMs and ANNs [15–18]. “Segmentation” of medical images to provide pixel-specific data has been an important contribution of computer vision in spinal imaging [19]. Gong et al. [20] proposed a network framework called “Axial-SpineGAN” for simultaneous segmentation and diagnosis of spinal structures using axial magnetic resonance imaging (MRI). Recently, magnetic resonance (MR) image augmentation using an ANN was approved by the Federal Drug Administration [21]. This technology facilitates MRI in a fraction of the normal study time, and the proposed benefits include improved patient satisfaction, enhanced image quality, and fewer motion artifacts. Regarding disk degeneration, the Pfirmann grading system has been validated in multiple studies; however, few authors have reported interobserver variability and heterogeneity in the results obtained using this grading system [22,23]. ML tools such as convoluted neural networks (CNNs) and deep learning can extract “radiomic data” from MR images that are quite difficult to interpret by visual inspection with the naked eye, and studies have shown an accuracy of 97% in assessing disk degeneration [24]. Salient features such as the shape and intensity of the disks can be evaluated in detail using the feature-extraction technique and critically analyzed [25]. Similarly, texture analysis of ligamentous structures in spinal MRI produced parameters for more accurate detection of lumbar canal stenosis [26]. Won et al. [27] compared the efficacy of grading lumbar canal stenosis in axial MRI between radiology experts and trained CNN, and the final agreement rates of decision-making among them were 77.9% and 74.9%, respectively, which were not significant. Other transformative innovations of AI in diagnostic imaging include MR fingerprinting and the identification of tissue properties using synthetic MRI. These techniques can be useful for the preoperative assessment of osteoporosis and early detection of spondyloarthropathies [28,29]. Apart from MRI, AI also has implications in the evaluation of plain radiographs [30,31] and computed tomography scans [32,33].

ML models, such as regression SVM and deep ANN, were adopted for calculating Cobb’s angle in patients with scoliosis and yielded satisfactory results with an error rate of <3° [34,35]. Lyu et al. [36] incorporated a neural networking algorithm in a three-dimensional (3D) ultrasound scan to detect the best quality image for detecting the deformity and reported an accuracy of 100%. Jaremko et al. [37] studied 18 testing samples using a three-layered back-propagation ANN to estimate Cobb’s angle in laser scan images of patients with deformity. They observed that the ANN of full torso imaging could distinguish Cobb’s angle >30° with acceptable sensitivity and specificity [37]. These computer-aided systems, by their automated measurements, provide reliable and objective assessments of the deformity.

Recently, Zhu et al. [38] collected data from 775 patients who underwent cervical spine surgery and screened 84 patient variables to identify differences between patients with positive and negative cervical ossification of the posterior longitudinal ligament (OPLL). They proposed an ML-driven nomogram to predict patients with cervical OPLL and could identify the risk factors and other associated characteristics of cervical OPLL. A few researchers have also developed computer-aided detection systems for the automated detection of spinal metastatic lesions in the thoracolumbar spine [39,40]. With more improvements in AI technology, the future could be automated reporting of radiological investigations, saving time and providing optimal results comparable with human reporting. Existing studies on the diagnosis of lumbar disk degeneration are shown in Table 1 [41–78].

AI allows for the comprehensive study of patient-specific anatomy and anatomical variations and the implementation of individualized surgeries. Lateral access lumbar spine surgeries have been on a rising trend, particularly in adult spinal deformities, owing to the magnitude of deformity correction they offer in both coronal and sagittal planes. However, complications such as lumbar plexus injury and ureteric and visceral injuries are a concern because of the narrow safety corridor. AI-enhanced ultrasound imaging was developed by Carson et al. [79] for the identification of internal and adjacent neural structures in lateral lumbar fusion surgeries, and the reported accuracy of nerve detection was >95%. Other recent advances such as robotic and navigation-guided procedures have revolutionized spine and spinal cord surgeries with their three-dimensional, real-time, and haptic feedback mechanisms. The Da Vinci Surgical System (Intuitive Surgical, Sunnyvale, CA, USA), SpineAssist (MAZOR Robotics Inc., Caesarea, Israel), ROSA (Medtech SA, Montpellier, France), and Excelsius GPS Robot (Globus Medical Inc., Audubon, PA, USA) are the commonly available surgical robots (Fig. 3). Studies have reported excellent outcomes with robotics in pedicle screw insertion, tumor excision, and spinal deformity correction surgeries [80,81]. In addition to their role in complex and challenging spine procedures, robots have been reported to be very useful in targeted procedures such as radiofrequency ablation, biopsy, and vertebral augmentation (kyphoplasty/vertebroplasty) [82]. AI-powered robots would be augmented with various DL and computer vision sensors (vision devices such as two-dimensional/3D cameras, fine tactile/vibration, proximity sensors, accelerometers, and other environmental sensors) that feed them with sensing data that they could analyze and act upon in real-time. These systems provide enhanced dexterity, precision, and stability during surgeries, allowing surgeons to perform complex operations with greater accuracy. However, the integration of AI algorithms in robotic navigation is still in its early stages and will gain much more prominence in the years to come.

Opioids are important rescue analgesics in spine surgery. More importantly, opioid therapy requires vigilant supervision, particularly in the older population, as it can induce dependence and cause adverse events following long-term use. Karhade et al. [83] proposed an AI-based model to stratify patients at risk of opioid dependence. Using this model, surgeons can identify at risk patients and adapt alternate pain management strategies, thereby mitigating the risks associated with opioid use [83]. Recently, Ayling et al. [84] examined the adverse events following lumbar spine surgeries for degenerative pathologies and reported that approximately 2.4% and 19.2% of patients had one of the major or minor perioperative adverse events, respectively. These adverse events not only increase hospitalization costs but also significantly affect surgical outcomes. Studies using the random forest approach to predict perioperative complications in adult spinal deformity surgery showed an accuracy of prediction of 87.6% [85]. Similarly, Wang et al. [86] employed a risk-stratification tool using ANNs in 12,492 patients to identify candidates who might be safe for ambulatory anterior cervical discectomy and fusion surgery, thereby reducing the chances of prolonged hospitalization. In addition, AI has been successfully evaluated to plan the discharge of patients following elective spine surgeries. The application of such algorithms would not only guide surgeons but also help patients reduce their healthcare costs and improve patient satisfaction.

Many prognostic tools are available in spine literature and are widely used in practice for the grading and prognosis of spinal disorders. Evaluating and analyzing these data and scoring systems for a larger population can be challenging for clinicians who use traditional statistical methods. ML algorithms are superior to conventional statistical models because they can analyze larger datasets and interpret nonlinear relationships in the given data. In patients with cervical spondylotic myelopathy, Khan et al. [87] used polynomial SVM learning to identify patients at risk of functional deterioration following surgery. The study included 757 patients, and the reported accuracy of detection was approximately 74.3%. McGrit et al. [88] examined data from 750 to 1,200 patients undergoing low back surgery and used regression analysis to predict the Oswestry Disability Index 1 year after surgery with an accuracy of up to 84%. Kim et al. [89] observed the incidence of postoperative C5 palsy in patients with OPLL and reported superior efficacy of ML algorithms with logistic regression models in the prediction of C5 palsy. Several similar studies using ML algorithms have reported outcome prediction following cervical and lumbar spine surgeries [90–92]. Hopkins et al. [93] performed a retrospective study of 4,046 patients undergoing posterior spinal fusion surgeries and identified patients at risk of developing surgical site infection using AI protocols. They reported positive and negative predictive values of 92.56% and 98.45%, respectively [93]. Hemiplegia/paraplegia, multilevel fusion, congestive cardiac failure, chronic pulmonary failure, and cerebrovascular disease were the risk factors with the highest significance in their study.

The spine is the most common site of skeletal metastasis, and approximately 70% of patients with malignancy could develop spinal metastasis [94]. Postoperative outcomes and overall survival rates in these patients have improved over the years. Karhade et al. [95] analyzed 1,790 patients with spinal metastasis and deployed Bayesian algorithm (ML) to identify 30-day mortality rates following spinal metastasis surgery. An open-access web application was also developed to identify these high-risk patients using ML algorithms. The authors concluded that as the volume of data in oncology increases, the creation of learning models and use of these systems as accessible tools may significantly enhance prognosis and appropriate management. A similar study included 1,053 patients with spinal epidural abscesses and showed promising results on the internal validation of ML algorithms to predict in-hospital and 90-day postdischarge mortality in these patients [96]. AI has thus emerged as a decision support tool, enabling surgeons’ clinical decision-making process to be augmented by their predictive power. In addition to these clinical utilities, computer vision-based algorithms record the rehabilitation and functional assessment of patients with spinal cord injury. Likitlersuang et al. [97] designed a tool using an egocentric camera system to detect the functional interactions of the hand with objects during activities of daily living in patients with cervical spinal cord injury during rehabilitation. This device can directly collect quantitative information on hand functions in home and community settings.

With the introduction of AI systems, the term “smart gait” is becoming popular, where an integrated human gait data analysis such as human activity recognition, gait phase detection, gait event prediction, fall detection, recognition of a person’s age and sex, and abnormal gait detection using AI tools can be performed. One of the important clinical signs in patients with radiculopathy caused by lumbar degenerative conditions is the listing of their trunk. This could alter the weight-bearing areas of the feet and cause variations in the gait pattern. Hayashi et al. [98] used SVM algorithms to analyze gait alteration in patients with L4 and L5 radiculopathy caused by lumbar canal stenosis and reported an accuracy of 80.4%. Similar published studies in the AI literature have focused on the identification of loads and stress patterns in various ligaments and joints of the foot, ankle, and knee, ensuring their use in orthopedic disorders such as adult-acquired flatfoot and osteoarthritis.

Adult spinal deformity is a complex pathology with heterogeneous clinical presentation and management options. Ames et al. [99] applied hierarchical clustering using AI to propose a classification model that would guide surgeons in deciding the appropriate surgical treatment. They are mainly useful in identifying patients at low risk and those likely to improve with surgical procedures. Regarding spinal fixation in these patients, pedicle screw instrumentation is preferred for deformity correction to regain sagittal and coronal balance. Several finite-element studies on the biomechanical properties and pullout strength of pedicle screws have been performed. However, these studies could not be individualized to patient-specific factors such as osteoporosis and sarcopenia, which are more prevalent in this population. Practically, the surgeon appraises the screw hold only by subjectively assessing the insertional torque of the screws. An experimental model to assess the pullout strength of pedicle screws using ML was developed by Khatri et al. [100], and they studied various parameters such as pedicle screw insertion angle, depth, density, and reinsertion. They observed that the model would also have a good clinical application with the inclusion of variables from the patient database, such as age, bone mineral density, and levels of activity. In addition, the authors suggested that failure mechanisms such as toggling and cyclical loading of pedicle screws can be investigated by incorporating the AI database.

Although the effects and contributions of AI and ML could be largely rewarding, the validity of the input data must be thoroughly investigated along with the integration of unstructured, scattered data to avoid erroneous predictions. External validation studies are needed to confirm the clinical efficacy in patients [8]. Selection biases could be a problem with these algorithms because the population data used for machine training may not necessarily be representative of the overall patient population. Another medicolegal consideration is the security and privacy of the patient data necessary for training these tools because it could result in breaches in the confidentiality of patient databases and cyber theft. Thus, data collection must be regulated in AI depository software to avoid privacy disclosure. During the early phase of their use, the need for data scientists to interpret health data could be one of the barriers to the widespread use of AI.

Many institutions, medical device companies, and pharmaceuticals have received significant incentives and funding to promote newer innovations such as AI and ML. These AI algorithms must not be programed to recommend specific pathways to increase the prioritization of designers and their funding companies. An ethical and regulatory framework must be developed for these algorithms so that AI can operate within well-defined norms.

In a recent systematic review, Liawrungrueang et al. [101] observed that AI and its integration with augmented reality (AR) and virtual reality (VR) appears promising for improving overall surgical safety. They also stated that VR surgical simulators create a secure environment for targeted surgical scenarios and foster self-guided learning. These extensive datasets can be processed by AI programs and aid in a deeper understanding of specific performance metrics during simulated operative tasks. In minimally invasive spine procedures, combining AR in the surgical workflow would be a major improvement because it would provide a 3D visualization of the anatomical structures. Considering all the effects of AI, in the coming years, it will serve as an invaluable additive tool in the surgical armamentarium; however, it could never replace the vast clinical knowledge and skills of surgeons.