Analysis and classification of EEG with adapted wavelets and local discriminant bases

ABSTRACTPhD THESISANALYSIS AND CLASSIFICATION OF EEG WITH ADAPTEDWAVELETS AND LOCAL DISCRIMINANT BASESNuri Fırat İNCEDEPARTMENT OF.ELECTRICAL AND ELECTRONICS ENGINEERING.INSTITUTE OF BASIC AND APPLIED SCIENCESUNIVERSITY OF ÇUKUROVASupervisor: Yrd. Doç. Dr. Sami ARICA2nd Supervisor: Prof. Dr. Pekcan UNGANYear: 2005, Pages: 99Jury: Prof. Dr. Yakup SARICADoç. Dr. Caner ÖZDEMİRDoç. Dr. Turgut İKİZYrd. Doç. Dr. Ali KOKANGÜLElectroencephalogram (EEG) can be used as a strategic tool in establishingcommunication and control between handicapped people and their environment. The socalled ?Brain Computer Interface? (BCI) is constructed by analysis and classification ofspecific patterns in the ongoing EEG which are induced without any need of muscular act.Motor Imagery has similar behavior and can be used as a strategy in the construction of BCI.Motor Imagery induced EEG patterns also have strong relationship to the real performanceof the event. Several methods such as band power and autoregressive model parameters wereused to analyze and classify the single trial EEG for a BCI task. Most of these methods usedfixed time points or frequency indexes. However the movement EEG is non-stationary andcontains subject specific patterns. Therefore it is crucial to extract local and subjectdepended information in an automated manner. In this work an adaptive time-frequencyapproach is investigated to analyze and classify real and imagery hand movement EEGs. Atfirst movement EEG is adaptively divided in time axis by using the Best Base (BB) approachwhich is obtained from the Local Cosine Packets by entropy minimization. BB has capturedtime varying properties of the signal by adjusting analysis segments where it is assumed theycorrespond to physiological states. In the latter case a modified version of Best Basealgorithm, ?Local Discriminant Bases? (LDB) were used to extract subject specific time-frequency features in an automated manner for classification of left and right handmovement imagery. Unfortunately this method suffers from the lack of translation invarianceand causes high dimensionality. Therefore several feature extraction and dimensionreduction methods such as modified mel-scale, principal component analysis and spin cycleprocedures are applied to improve the classification performance. One of the interestingresults is the difference of the adaptive segmentations and feature characteristic of bothhemispheres. The algorithm did not only adapt to time and frequency but also to space. As afurther step the number of electrodes is increased to benefit from the different cortical areas.Accordingly our algorithm has promising success for the BCI technology.Key Words: EEG, Brain Computer Interface, Movement Imagery, Adaptive Time-Frequency Analysis.I