Artificial Intelligence and Machine Learning Applications in Spine Surgery

WHAT IS ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING?

AI is the broader concept of applying systems to simulate human learning and thinking. One of the main applications of AI is ML, which utilizes various computational techniques to continuously learn and self-adjust from past data in order to determine mathematical relationships inherent in the data. The majority of prior outcomes research in spine have involved statistical analyses to characterize relationships between independent and dependent variables. By doing so, the focus of these statistical analyses has been to identify the parameters of a model and understand how each impacts the prediction. Although these analyses are valuable and often offer the researcher an ease of interpretability, they are static in nature and are often subject to selection bias and limited external validity. In contrast, the focus of ML is less about the parameters of a model and more about the prediction. This often leads to “black box” algorithms, which may be difficult to conceptualize. However, ML has the potential to process an inordinate amount of data, learn and adapt to varying patient populations, and provide more powerful predictive capabilities.

In general, there are 3 main paradigms within ML: supervised learning, unsupervised learning, and reinforcement learning. Supervised learning is a more “hands-on” approach than other forms of ML. This approach relies on labeled datasets to “supervise” algorithms to classify and predict outcomes accurately. Tasks suited for this style of learning involve classification and regression. One potential drawback is that it may be time-consuming to train the data as it requires labels for each input and output, which may be extensive for large datasets. Common models used in supervised learning include decision trees (DTs), support vector machines (SVMs), and LR.

The main difference between supervised learning and unsupervised learning is the use of labeled datasets. Unsupervised learning trains algorithms based on unlabeled data without any guidance. This style of learning attempts to discover the inherent patterns of unlabeled data on its own. Tasks suited for this approach involve associations and clustering. Unsupervised learning may be used to cluster large unlabeled datasets based on their similarities and/or differences as well as process anomalies in visual or medical imaging data.

Similar to unsupervised learning, reinforcement learning does not involve external supervision. However, the goal of this approach involves a “trial-and-error” method of exploring new, unlabeled data to minimize cost or maximize reward parameters without any guidance. Over time, these algorithms will continue to explore new data and develop their own rules to maximize an outcome. Artificial neural networks (ANNs) often serve as the learning component in this approach.

As with any statistical analysis, it is important to make the distinction between correlation and causation in ML. A correlation is an association between variables. For instance, variable A may be associated with variable B. It is possible that variable A caused variable B or vice versa. However, there could be a third factor, variable C, which changes or confounds both variables A and B independently. Only after controlling for all confounders and random chance can it be assumed that a causal relationship exists between variables A and B. Although ML enables users to identify nuanced, complex relationships, these relationships are still correlations, not causations. For instance, ML may be trained on hundreds of x-ray images to detect a spinal deformity pattern; however, changes in pixel gradients or variances in patient positioning in x-ray images may introduce new variables that can influence the model. The assumption made with ML is that with a large enough training dataset, the ML will be able to encode all possible relationships into the model. One of the key factors behind the incremental gains in the accuracy of ML is the growing availability of data and stronger computer processing power. However, there may be practical limits to this. Currently, AI researchers argue for integrating a causal understanding into ML, which may require fewer training samples; however, the implementation of “causal ML” remains at the conceptual level.4,5

COMMON MACHINE LEARNING MODELS USED IN SPINE RESEARCH

DT and random forests (RFs) are examples of supervised learning, which uses a set of inputs and outputs to predict classification and regression problems such as readmission rates, patient-reported outcomes, and surgical complications. In short, a DT iteratively asks questions to partition data to reach the eventual outcome. An RF is an ensemble of many DTs. In comparison with a single DT, RF reduces overfitting and ultimately may improve accuracy.

SVMs are supervised learning algorithms that use decision boundaries or hyperplanes to help categorize 2 classes of data points and maximize the “distance” between the data of those 2 classes. SVMs may be useful in classifying images and conducting image segmentation analyses.

ANN is a commonly used ML tool that emulates the framework of the nervous system. ANN involves multiple layers with at least 1 hidden layer between the input and output. These hidden layers are interconnected by weighted linkages.

Convolutional neural networks (CNNs) are a subset of ANN that leverage a mathematical operation called convolution and are best suited for image, speech, and audio signal inputs. In short, CNNs take an input picture, assign relevance (eg, weights) to different aspects in the image through multiple layers, and distinguish them.

APPLICATIONS OF MACHINE LEARNING IN THE CERVICAL SPINE

A number of studies have sought to develop ML algorithms to predict outcomes and complications after cervical spine surgery.6–8 One of the earliest applications of ML in cervical spine outcomes was introduced by Arvind et al in 2018, who examined more than 20,000 patients who underwent anterior cervical discectomy and fusion (ACDF) from the National Surgical Quality Improvement Program (NSQIP) database and compared the performance of various ML models (ANN, LR, SVM, and RF) to predict surgical complications.9 These authors found that ANN and LR outperformed the American Society of Anesthesiologists (ASA) physical status classification for every complication (area under curve [AUC] for cardiac, venous thromboembolism, wound complication, and mortality: ANN 0.772, 0.656, 0.518, 0.979; LR 0.759, 0.639, 0.501, 0.974; ASA 0.566, 0.397, 0.455, 0.346). Furthermore, ANN outperformed LR in predicting venous thromboembolism, wound complication, and mortality rate. In comparison, both the SVM and RF models were unable to perform better than random chance, which suggests that selecting the appropriate ML algorithm is an important factor to consider. As with any ML model, the performance hinges on the quality of the training data. Although NSQIP includes a large national sample of surgical patients, it lacks highly granular data pertinent to spine surgery. As a result, many surgery-specific variables, which may serve as stronger inputs, are not available to include in these models.

In 2019, one of the first studies to use ML to predict patient-reported outcomes after surgery for degenerative cervical myelopathy was performed by Merali et al.10 These authors used a prospective, multicentered database to review 757 patients and compared multiple ML models (DT, RF, SVM, LR, and ANN) to predict patient-reported outcomes (Short Form-6D and modified Japanese Orthopaedic Association [mJOA]) after surgery. As measured by the AUC, the RF and SVM models outperformed LR and DT in up to 2-year follow-ups, which was attributed to the ability of these models to process more complex nonlinear and conditional relationships than either LR or DT models. The RF model found that the longer duration of preoperative symptoms, worse preoperative disease severity, older age, greater body weight, and current smoking status were associated with worse postoperative outcomes. Of note, this study did not include radiographic parameters in their models, which may have improved the performance. Furthermore, ANN performed poorly compared with the other models. This was likely due to the relatively limited number of training samples since ANN models are known to generally require larger training samples than SVM or RF for adequate training.

Wong et al used SVM to identify demographic, radiographic, and paraspinal muscle parameters that would predict proximal/distal adjacent segment disease after 2-level ACDF.11 The SVM model achieved high accuracy (96.7%) and an AUC (0.97) for predicting adjacent segment disease. This study found that preoperative total cross-sectional area of cervical paraspinal muscle, relative fat composition, and asymmetry at C5 to C7 were predictive for early onset of adjacent segment disease.

Another application of ML in cervical spine has involved medical image analysis.12 Computer vision techniques have shown the potential for a wide spectrum of medical applications and recently shown adoption within spine surgery. Huang et al reviewed over 300 images of 9 different ACDF systems from 5 different companies to identify the manufacturer and model of anterior cervical spinal hardware.13 Knowing the manufacturer and model of prior cervical implants can facilitate faster and safer revision surgeries. These authors used a computer vision algorithm “bag of visual worlds” instead of other well-known techniques such as CNN because the primary task was the classification of a relatively sparse dataset. Images were subsequently classified by an SVM classifier. These authors achieved an accuracy of greater than 90% with a relatively small sample size, and this performance persisted for 1-level, 2-level, and 3-level plates.

Recently, ML applications in predicting the diagnosis and severity of cervical spondylotic myelopathy have shown promise. Hopkins et al used ANN on clinical and radiographic factors on 18 images to predict cervical spondylotic myelopathy as well as mJOA scores.14 These authors achieved a median accuracy of 90% as well as mJOA scores within 0.4 points. This study was limited by the small sample size and lack of comparative partition analyses for training, testing, and validating, which are commonly used to optimize ML models.

APPLICATIONS OF MACHINE LEARNING IN THE LUMBAR SPINE

Several studies have used ML algorithms to predict complications after lumbar surgery. Kim et al used the NSQIP database to study more than 22,000 patients who underwent lumbar fusion surgery.15 They trained and validated both ANN and LR models to predict postoperative complications (cardiac, wound, venous thromboembolism, and mortality). On the basis of AUC values, ANN and LR had comparable AUC values for predicting all complication types; however, ANN had greater sensitivity for detecting wound complications and mortality. The main difference between ANN and LR models is that ANN allows the characterization of nonlinear relationships in the data and has the potential for slightly better accuracy for relatively rare complications, which is important for medical prognostication.16,17

Hopkins et al examined the efficacy of deep neural network to predict 30-day hospital readmission after posterior lumbar fusion based on more than 20,000 patients in the NSQIP database.18 This model took advantage of the 177 unique input variables of the NSQIP database and fed data through a series of 7 layers, each with varying degrees of forward and backward communicating neurons and sporadic dropout layers to avoid overfitting. Ultimately, this model outperformed LR by achieving an AUC of 0.81 (vs 0.72).

Ogink et al used the NSQIP database to study patients with degenerative spondylolisthesis and compared 4 ML algorithms (DT, SVM, Bayes point machine, and ANN) in their ability to predict discharge placement after surgery.19 The AUC was fairly similar among all 4 models (DT AUC = 0.733, SVM AUC = 0.742, ANN AUC = 0.755, and Bayes point machine AUC = 0.753). This study demonstrated the predictive capabilities of ML on discharge placement with good discrimination and performance. These authors have subsequently developed a web application for further clinical utility.

Kuris et al examined the records of more than 60,000 patients who underwent lumbar fusion from the NSQIP database and ANN to predict 30-day readmission.20 The ANN algorithm was able to achieve an accuracy of 94.6% for anterior lumbar interbody fusion, 94% for posterior lumbar interbody fusion, and 92.6% for posterior spinal fusion with AUC values 0.64 to 0.65.

Harada et al performed a retrospective, single-center review of 2630 patients who underwent lumbar microdiscectomy to examine the utility of ML to predict the risk of recurrent herniated nucleus pulposus.21 Input variables included patient demographics/comorbidities, clinical factors (eg, number of herniated levels, type of herniation, duration of symptoms, motor examination, surgical approach), patient-reported outcomes, and radiographic factors. These authors used an ML algorithm called “extreme gradient boosting classification (XGBoost classification),” which involves both linear model and tree learning algorithms. This model demonstrated excellent model discrimination with an accuracy 0.7, an AUC 0.72, and a Brier score of 0.21.

To the authors’ knowledge, Roller et al were the first to use ML to predict the level of lumbar spinal decompression. These authors used CNN to predict the lumbar decompression level based on magnetic resonance images of 141 patients.22 This algorithm was able to predict with an accuracy of 65%, with higher mean scores for L3-S1 surgical levels.

MACHINE LEARNING APPLICATIONS IN ADULT SPINAL DEFORMITY

FUTURE DIRECTIONS OF ARTIFICIAL INTELLIGENCE IN SPINE SURGERY

The current literature on ML in spine surgery is promising as it has already demonstrated wide-ranging applications for the cervical spine, the lumbar spine, and ASD. Other possible ML applications include integration in the preoperative planning phase and with other existing technologies, such as robot-assisted surgery and/or augmented reality systems. A variety of surgical options exist for ASD (eg, 3-column osteotomies, level selection, and interbody fusion), each with their own potential risks and benefits. The determination of the optimal surgery for each individual patient is often at the discretion of the surgeon. Furthermore, predicting postoperative alignment and specifically reducing the risk of PJK/PJF remain a significant challenge for ASD surgery. ML has the potential to augment surgeon decision-making and improve surgical outcomes. This will hinge on continuing to build an ensemble of models based on high-quality and robust databases as well as further validation and ultimate consolidation of published models to better integrate in clinical practice. It is important to keep in mind, however, that the goal of AI and ML is not to replace but to complement the surgeon in providing safer and more efficient patient-centered care.