Efficient channel reduction and feature extraction schemes for analyzing mental and cognitive task based EEG signals

Abstract

In view of recent increase in awareness of workplace safety, improved productivity, and social security, the importance of efficient channel reduction and feature extraction scheme for analyzing and classifying alcoholics from non-alcoholics has increased prodigiously nowadays. Most of the reported approaches use complex algorithm for channel reduction and feature extraction technique with a large feature dimension. Moreover, the channel reduction techniques have no physiological or neurological correspondence, hence, differ from person to person and trial to trial. Therefore, all the channels are to be used for the collection of EEG data which, when fed to a network or algorithm, result in a reduced set of channels. In this thesis, some channel reduction techniques are proposed based on extensive analysis on neurological behavior of different location of human brain, such as: lobe-based scheme, cortical function-based scheme, hemispheric lateralization based-scheme, Brodmann’s localization theory-based scheme, and weighted score-based scheme. Through extensive study of different brain functions, it is shown that different regions of the brain as well as various channels can be discarded which results in a significant reduction in number of channels to be used. The proposed methods in this study even allow to use only one sixth of the total channels to precisely capture mental and cognitive tasks. In the process of selection, apart from biological concept, EEG signal characteristics are also taken into consideration. It is shown that the proposed weighted scoring method offers the highest number of channel reduction where both quantitative and qualitative weighting of various brain functions and scoring of channels are proposed based on similarity of the brain functions with the mental and cognitive task and proximity of the spatial areas with respect to the active functional zone. In order to classify alcoholic and nonalcoholic, two different feature sets are proposed. Reflection coefficient of gamma band visual evoked potential is proposed as a feature extraction scheme and further reduction of channels is achieved by analyzing the feature. Next, variance of spatially distributed EEG is proposed as another feature extraction technique with the lowest feature dimension. Extensive simulation is carried out on publicly available datasets where different types of cross validation techniques are used by different classifiers such as k-nearest neighbor and support vector machine. It is found that both the feature extraction schemes can offer very high classification accuracy even with the proposed reduced set of channels. The proposed method offers EEG-based classification of alcoholic from healthy persons with very small number of channels which is expected to reduce time and complexity of computation.