Machine Learning for the Prediction of Cervical Spondylotic Myelopathy: A Post Hoc Pilot Study of 28 Participants

Abstract

Background: Cervical spondylotic myelopathy (CSM) severity and presence of symptoms are often difficult to predict based simply on clinical imaging alone. Similarly, improved machine learning techniques provide new tools with immense clinical potential.

Methods: A total of 14 patients with CSM and 14 controls underwent imaging of the cervical spine. Two different artificial neural network models were trained; 1) to predict CSM diagnosis; and 2) to predict CSM severity. Model 1 consisted of 6 inputs including 3 common imaging scales for the evaluation of cord compression, alongside 3 objective magnetic resonance imaging measurements. The outcome for model 1 was binary to predict CSM diagnosis. Model 2 consisted of 23 input variables derived from probabilistic volume mapping measurements of white matter tracts in the region of compression. The outcome of model 2 was linear, to predict the modified Japanese Orthopedic Association (mJOA) score.

Results: Model 1 was used in predicting CSM. The mean cross-validated accuracy of the trained model was 86.50% (95% confidence interval, 85.16%-87.83%) with a median accuracy of 90.00%. Area under the curve (AUC) was calculated for each repetition. Average AUC for each repetition was 0.947 with a median AUC of 1.0. Average sensitivity, specificity, positive predictive value, and negative predictive value were 90.25%, 85.05%, 81.58%, and 91.94%, respectively. Model 2 was used in modeling mJOA. The mJOA model predicted scores, with a mean and median error of -0.29 mJOA points and -0.08 mJOA points, respectively, mean error per batch was 0.714 mJOA points.

Conclusions: Machine learning provides a promising method for prediction, diagnosis, and even prognosis in patients with CSM.

Keywords: Artificial intelligence; CSM; Cervical myelopathy; Cervical spondylotic myelopathy; Machine learning; Spine.

Copyright © 2019 Elsevier Inc. All rights reserved.

Figures

Figure 1-

Algorithm specifications for Deep Neural Network (DNN) used in Model 1. All variables underwent processing through a 7 layer artificial neural network with above specified neurons in each layer. Images were analyzed in groups of 4 at a time until all 18 were completed, at which point the algorithm would adjust predictive variable weights and repeat for 25 epochs/repeats. Upon completion, the final algorithm would be used to verify accuracy on 10 unseen images from the original dataset. Images were redistributed randomly into test/training groups, model weights were reinitialized and randomized and the entire process was repeated 200 times for population data collection.

Figure 2-

Algorithm specifications for Deep Neural Network (DNN) used in Model 2. All variables underwent processing through a 9 layer artificial neural network with above specified neurons in each layer. Images were analyzed in groups of 3 at a time until all 78 were completed, at which point the algorithm would adjust predictive variable weights and repeat for 1250 epochs/repeats. Upon completion, the final algorithm would be used to verify error on 26 unseen data points from the original dataset. Images were redistributed randomly into test/training groups, model weights were reinitialized and randomized and the entire process was repeated 150 times for population data collection.

Figure 3-

Distribution of cross validation accuracies of CSM predictions on 10 previously unseen patients after first training and model adjustment on random sub-sets of 18 patients.

Figure 4-

Representative receiver operator curve for Model 1. Area under the curve for this particular repetition was 0.880. Average AUC for all models was 0.947 with a median AUC being 1.0.

Figure 5-

Outcome Distribution of Model 2. (A) Distribution of mean error per batch in units of points of mJOA score. (B) Distribution of median error per batch in units of points of mJOA score.