An Expert System to Detect and Classify CNS Disorders Based on Eye Test Data Using SVM and Nature-Inspired Algorithms

Abstract

AbstractChanges in eye movements have a strong relationship with the changes in the brain. Several medical studies have revealed that in most CNS disorders, ocular manifestations are often associated with brain symptoms. To date, computational intelligence has not been used to study the relationship between eye movements and brain disorders. We propose a support vector machine (SVM) based machine learning solution to identify, five disorders related to the central nervous system; Amyotrophic Lateral Sclerosis (ALS), Multiple Sclerosis (MS), and Alzheimer’s Disease (AD), Parkinson’s Disease (PD), and Schizophrenia. Apart from the SVM, the proposed solution handles two major problems which occur in the data preprocessing stage; insufficiency of real eye test data and finding optimal features set for a particular disorder. An algorithm is developed to generate synthetic data and to find the optimal features set for a particular disorder, a solution based on particle swarm optimization is proposed. We trained the SVM models using the generated synthetic data and tested with the real data. The proposed system based on SVMs with linear, polynomial, and RBF kernels were able to identify the stages of the disorders, as diagnosed in medical studies. The SVM with the RBF kernel worked with an accuracy of 97% in identifying the existence of a CNS disorder. In classifying the stages of ALS, the linear kernel worked with an accuracy of 77% while the polynomial kernel worked with an accuracy of 100%, 90%, and 64% in classifying stages of MS, AD, and Schizophrenia. For PD, SVMs with all kernels gave an accuracy of 96%. The results are encouraging, giving sufficient evidence that the proposed system works better. We further illustrate the viability of our method by comparing the results with those obtained in previous medical studies.KeywordsSVMLinear KernelRBF KernelPolynomial KernelPSOGACNS disordersSynthetic dataFeature selection