Use this URL to cite or link to this record in EThOS:
Title: Signal processing methods for EEG data classification
Author: Varnavas, Andreas Soteriou
ISNI:       0000 0004 2671 7755
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2008
Availability of Full Text:
Access from EThOS:
Access from Institution:
The scope of this thesis is to determine appropriate features of a person's electroencephalo-graphic (EEG) data and the way in which they can be used to predict their performance in an "oddball" experiment. We classify a person's performance in one of the following classes: "success" or "failure", depending on the reaction time related with it. Predicting a person's performance means finding the correct class where the latter belongs to, using the person's EEG data corresponding to a time period before the reaction takes place. The problem is addressed in various ways as far as the feature construction process is concerned, whereas a Gaussian classifier is used in all cases. First, the raw time signals and the magnitude of their Fourier Transform are used as features. Then the number of these features is reduced, using various feature selection methods in combination with Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA) and Non Neg-ative Matrix Factorization (NMF). Subspace methods, using PCA and NMF to construct different spaces for the two classes, are also used to perform the desired classification. Features are also constructed using a Time-Frequency representation of the EEG signals. In this case we propose two novel algorithms which analyze the magnitude of the Time-Frequency representation using NMF, in a single or multi-trial basis, and the coefficients of selected NMF components are used as features. Finally, a novel algorithm performing the desired classification based on the construction of signals characterising each of the classes is proposed. These characteristic signals are constructed linearly combining the EEG signals of the various channels, minimising the variance of the time samples over the trials belonging to the same class. A novel algorithm is also proposed for selecting the appropriate channels to be used in the construction of the characteristic signals. This algorithm is based on the identification of the channels showing the least interference from background brain activity. The maximum classification rate produced for one of the 11 subjects in our study is 97.22%. However the rates usually vary between 70% and 80%. Considering the difficulty of the problem, this is encouraging. However, with these classification rates, real applications should only consider the generation of notification signals to increase the attention of operators and not involve any critical, automatic decision making process. Moreover, the variability in the methods and channels being optimal across the various subjects, implies that in a real case, a "tailor made" system should be designed for each user.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available