Kernel SVM Classifiers based on Fractal Analysis for Estimation of Hearing Loss

Abstract

Hearing screening consists of analyzing the hearing capacity of an individual, regardless of age. It identifies serious hearing problems, degree, type and cause of the hearing loss and the needs of the person to propose a solution. Auditory evoked potentials (AEPs) which are detected on the EEG auditory cortex area are very small signals in response to a sound stimulus (or electric) from the inner ear to the primary auditory areas of the brain. AEPs are noninvasive methods used to detect hearing disorders and to estimate hearing thresholds level. In this paper, due to the nonlinear characteristics of EEG, Detrented Fluctuation Analysis (DFA) is used to characterize the irregularity or complexity of EEG signals by calculating the Fractal Dimension (FD) from the recorded AEP signals of the impaired hearing and the normal subjects. This is to estimate their hearing threshold. In order to classify both groups, hearing impaired and normal persons, support vector machine (SVM) is used. For comparably evaluating the performance of SVM classifier, three kernel functions: linear, radial basis function (RBF) and polynomial are employed to distinguish normal and the abnormal hearing subjects. Grid search technique is selected to estimate the optimal kernel parameters. Our results indicate that the RBF kernel SVM classifier is promising; it is able to obtain a high training as well as testing classification accuracy.