The Role of Machine Learning in Spine Surgery: The Future Is Now

Abstract

The recent influx of machine learning centered investigations in the spine surgery literature has led to increased enthusiasm as to the prospect of using artificial intelligence to create clinical decision support tools, optimize postoperative outcomes, and improve technologies used in the operating room. However, the methodology underlying machine learning in spine research is often overlooked as the subject matter is quite novel and may be foreign to practicing spine surgeons. Improper application of machine learning is a significant bioethics challenge, given the potential consequences of over- or underestimating the results of such studies for clinical decision-making processes. Proper peer review of these publications requires a baseline familiarity of the language associated with machine learning, and how it differs from classical statistical analyses. This narrative review first introduces the overall field of machine learning and its role in artificial intelligence, and defines basic terminology. In addition, common modalities for applying machine learning, including classification and regression decision trees, support vector machines, and artificial neural networks are examined in the context of examples gathered from the spine literature. Lastly, the ethical challenges associated with adapting machine learning for research related to patient care, as well as future perspectives on the potential use of machine learning in spine surgery, are discussed specifically.

Keywords: artificial intelligence; deep learning; machine learning; orthopedic surgery; spine surgery.

Figure 1

Schematic of an artificial neuron with inputs (x_{1, 2…n}), weights (w_{1, 2…n}), bias (b), transfer function (∑), activation function (φ) and output (y).

Figure 2

Diagram of a single Classification and Regression Tree (CART) with five terminal nodes or leaves.

Figure 3

A decision tree analysis to stratify spinal cord injury cases and to identify clusters of homogeneous patients that would respond similarly to treatment. The root node was based on the American Spinal Injury Association Impairment Scale (AIS), which ranged from grade A through D. The subsequent internal node was based on AOSpine injury classification (class A/B or C). Each branch then underwent another node based on anatomical region (cervical or thoracic). Class A/B cervical injuries were divided further based on age. Six unique terminal nodes or clusters were identified. Reproduced with permission by Tee et al. (32).

Figure 4

Biomechanical model for testing pedicle screw pullout-strength. (A) Schematic of rigid polyurethane foam to mimic normal, osteoporotic, and extremely osteoporotic bone. (B) Apparatus to test pedicle screw pull-out. (C) Force vs. displacement graph from pullout studies. (D) Anatomy of pedicle screw instrumentation. (E) Decision tree learning to predict pedicle screw pullout success vs. failure in relation to foam density, screw depth and insertion angle. Reproduced with permission by Varghese et al. (34).

Figure 5

Support vector machines (SVMs). (A) Two-dimensional representation of a binary classification problem in an SVM represented by two features (x, y) with a maximum-margin hyperplane. (B) The same binary classification problem as before, but the outcomes are linearly inseparable in two-dimensions. (C) A linear kernel is used to transform the previous plot, allowing for a hyperplane to be constructed in three-dimensions (x, y, z). (D) A two-dimensional representation of the same hyperplane created in higher dimensions. (E) An overfitting SVM that is being influenced by outlier data or noise. (F) An underfitting SVM that fails to maximize the distance between two outcomes in a binary classification task.

Figure 6

Schematic of a deep artificial neural network with multiple hidden layers.

Figure 7

A convolutional neural network schematic for image classification.

Figure 8

Masking in a convolutional neural network. This automated vertebral column segmentation model incorporates masking to generate two redundant classifiers for a traditionally binary classification task. Class 1 represents the background. Class 2 and 3 are unique redundant classifiers. Class 4 is the spine. By doing this, the machine becomes more adept at identifying varieties of vertebra, instead of the background. Reproduced with permission by Vania et al. (33).

Figure 9

Visual representation of oversampling low incidence complications via adaptive synthetic sampling approach to imbalanced learning or ADASYN. Oversampled synthetic neural networks are created and then compared to subsets of the “no complication” cohort (68).

Figure 10

Augmented reality system that superimposes pedicle screw trajectories from computer assisted navigation onto the operating field. By minimizing the need to memorize trajectories from a separate screen, the surgeon is more readily able to identify safe zones. The blue, red, pink, yellow, and green lines represent correct, medial, lateral, superior and inferior breaches, respectively. Reproduced with permission by Nguyen et al. (69).

Figure 11

A proof of concept application of Microsoft Hololens for reverse total shoulder arthroplasty. (A) The surgeon is able to view in real-time and place in space a 3D hologram from a CT of the patient's scapula. (B) A 3D hologram of the patient's scapula is superimposed intraoperatively in order to fully visualize the glenoid and other relevant anatomy. Reproduced with permission by Gregory et al. (74).