Electroencephalography (EEG) Based Analysis of Human Brain Activities for Audio Stimuli

Abstract

A brain-computer interface (BCI) uses signals originated from the brain to direct some peripheral devices. This thesis deals with multichannel electroencephalogra- phy (EEG) based BCI implementation. Synchronization of neural activity between di erent parts of human brain has great signi cance in coordination of cognitive activities. The time-frequency (TF) representation can be used to measure the synchronization between di erent channels of EEG. The technique, called syn- chrosqueezing transform (SST) is one of the techniques that operates with the continuous wavelet transform (CWT) and produces impressively localized time- frequency representations of nonlinear and nonstationary signals. The SST based method is proposed to e ectively measure the synchronization in TF domain. Due to its data adaptability and frequency reassignment properties, the SST produces a well-de ned TF representation. The marginal time coherence for di erent channel pairs is used to quantify synchronization. The experiment is performed with both synthetic and real EEG data. The results show that the marginal time coherence based on the proposed SST exhibits very clear discrimination between two types of motor imagery (MI) movement. This research also presents a novel method for the selection of e ective spatial lter pair and discriminative features in EEG based MI classi cation. Usually, the spatial lter pair is selected manually. However, the manual selection of CSP lters does not con rm that the approach will achieve the best accuracy. In the proposed method, the analyzing EEG data is divided into training and test sets. The training set is used to select appropriate spatial lters with dominant features. To accomplish such features, the EEG of training set is segmented again into two subsets termed as training subset and test subset. The features of both subsets are extracted using common spatial pattern (CSP). The mutual information between the features of training subset and class levels of the training subset is calculated. Then features of training subset are ranked on the basis of values of the mutual information. Besides, the features of test subset are also ranked according to the order of the training subset features. The initial classi cation performance using training and test subsets is obtained by using linear discriminant analysis (LDA). Then a grid search method selects the e ective number of spatial lter pairs as well as the discriminative features by which the maximum accuracy score is yielded. Thus the selected spatial lter and features are used in actual classi cation accu- racy of the test set of EEG. In this research, the binary classi cation performance of the proposed approach is evaluated to classify MI data where the datasets are widely used as the publicly available dataset from BCI competition III. The ap- proach achieves more increased mean accuracies than di erent existing methods of MI tasks. Finally, the method is veri ed to classify audio stimuli based EEG. In auditory EEG, the proposed approach produces superior classi cation accuracy compared to prevailing methods.