The Role of Machine Learning in Predicting the Onset and Progression of Neuropathic Pain After Spinal Cord Injury: A Literature Review

Abstract

Introduction: Developing a diagnostic tool that can determine whether a patient will develop neuropathic pain following a spinal cord injury can aid clinicians in treatment procedures and improve patient outcomes. Developing new detection technology can take years, thus finding a way to use existing diagnostic tools would be optimal. Machine learning can be leveraged to incorporate existing data and classify patient outcomes when there are obvious patterns for classification. Methods: A review of full reports published in English was conducted through PubMed. The relevant keywords used in this search included “neuropathic pain”, “spinal cord injury”, machine learning, and “predict” among others. Eight relevant citations were retrieved and reviewed. Results: A decision tree regressor model using clinical measures for neuropathic pain and level of spinal cord injury found that BMI and anxiety scores were the most influential variables in predicting outcomes. A similar tree for functional magnetic resonance imaging (fMRI) data found ventral and dorsal tissue bridges to be predictors of neuropathic pain. Another fMRI study pointed to a strong correlation between changes in perioperative blood oxygen levels at the ipsilateral frontal lobe and neuropathic pain outcomes. Magnetic resonance spectroscopy (MRS) implicated a lower glutamate-glutamine/myoinositol ratio in high neuropathic pain. Various machine learning algorithms were evaluated in building an EEG classifier in two separate studies, and classification accuracies greater than 80% were reached in both. A classifier built using positron emission tomography data attained classification accuracies of 87.5%. Discussion: The most common machine learning algorithm used in building classifiers was support vector machines, linear discriminant analysis and neural net. Regression trees were also used, but they were used to elucidate the variables influencing predictions. Each study has its limitations, either due to limitations of the study method, classification method or data type. Conclusion: There exist many methods to study neuropathic pain and spinal cord injury and each method provides different information regarding the mechanism of pain, influential variables, and physiological changes that occur with pain. Classification can be done using any of these methods to achieve acceptable accuracies, but these accuracies are not enough for a clinical prognostic classifier.