Calssification of Mental Tasks using Eeg Signals / Fatma ElZahraa Abd El Rahman ElChihaby ; Supervised Mohamed Fathy AbuElyazeed , Mohamed Emad Mousa Rasmy , Mohamed Waleed Fakhr

Thesis (M.Sc.) - Cairo University - Faculty Of Engineering - Department Of Electronics and Communications

EEG is an acronym for ElectroencephalographThis is a recording (graph) of electrical signals (electro) from the brain (encephalo) by using electrodes appropriately placed on the scalpThe electrodes only receive electrical signals naturally generated by the brainThe EEG has many applications ; it can be used in the classification of brain diseases and psychiatric disorder such as obsessive compulsive disorder and schizophrenia casesMoreover , EEG helps in the classification of various mental tasks which in its turn may allow disabled persons to communicate with the outside world based on a set of normal mental activitiesFor example , it can give a physically disabled person the chance to control a device like a wheelchair In this work , we deal with the classification of five mental tasks : baseline task , multiplication task , rotation task , letter task , and count taskInitially , the EEG signal is divided into time intervals (windows) The rectangular window is used for its simplicity with 50percent overlappingThen a feature vector is associated to each of these intervalsThe features could be Linear Predictive Coefficients (LPC) , Power spectral Density (PSD) , or Mel Frequency Cepstrum Coefficients (MFCC) There are two methods to calculate the PSD ; the classical periodogram method and the averaged periodogram method (Welch's method) , which gives in its turn a 10percent higher accuracy than the classical methodThe PSD using asymmetric power ratio (APR) or the MFCC features give better results than the LPC coefficients ; the amount of increase is about 4percent - 6percentIn the purpose of increasing the classification accuracy , the PSD with APR and MFCC are concatenated together to form a new feature vectorThe Gaussian Mixture Model (GMM) is utilized as the core statistical classifier ; with diagonal covariance matrixFinally , a postprocessor is added after the classifier to cancel the classification errors of frames corrupted by artifactsIt uses the majority rule to all trial frames to get one decision each trialIn this case , the implemented system classifies all possible pairs of the five mental tasks with an average classification accuracy of 9631percentIt also achieves an average classification accuracy of 9231percent for five mental tasks classification The performance of the system is characterized by the classification accuracy , the identification time , and the classifier complexityIn order to enhance this performance , different component analysis techniques are used such as Principal Component Analysis (PCA) and Independent Component Analysis (ICA) When the ICA is used as a preprocessor ; the classification accuracy is increased by 2 - 4percentTo achieve less identification time and to reduce the complexity of the classifier , the PCA is used to determine the best directions (principal components) by which the data is effectively represented ; the identification time is decreased by 50percent - 70percentWhen both ICA and PCA are used for the concatenated set of feature , the system achieves an average classification accuracy of 9762 percent and 9631percent for all possible task pairs and for the five mental tasks respectively

Issued also as CD