An improved EEG pattern classification system based on dimensionality reduction and classifier fusion
- Publication Type:
- Issue Date:
Analysis of brain electrical activities (Electroencephalography, EEG) presents a rich source of information that helps in the advancement of affordable and effective biomedical applications such as psychotropic drug research, sleep studies, seizure detection and brain computer interface (BCI). Interpretation and understanding of EEG signal will provide clinicians and physicians with useful information for disease diagnosis and monitoring biological activities. It will also help in creating a new way of communication through brain waves. This thesis aims to investigate new algorithms for improving pattern recognition systems in two main EEG-based applications. The first application represents a simple Brain Computer Interface (BCI) based on imagined motor tasks, whilst the second one represents an automatic sleep scoring system in intensive care unit. BCI system in general aims to create a lion-muscular link between brain and external devices, thus providing a new control scheme that can most benefit the extremely immobilised persons. This link is created by utilizing pattern recognition approach to interpret EEG into device commands. The commands can then be used to control wheelchairs, computers or any other equipment. The second application relates to creating an automatic scoring system through interpreting certain properties of several biomedical signals. Traditionally, sleep specialists record and analyse brain signal using electroencephalogram (EEG), muscle tone (EMG), eye movement (EOG), and other biomedical signals to detect five sleep stages: Rapid Eye Movement (REM), stage 1,... to stage 4. Acquired signals are then scored based on 30 seconds intervals that require manually inspecting one segment at a time for certain properties to interpret sleep stages. The process is time consuming and demands competence. It is thought that an automatic scoring system mimicking sleep expert rules will speed up the process and reduce the cost. Practicality of any EEG-based system depends upon accuracy and speed. The more accurate and faster classification systems are, the better will be the chance to integrate them in wider range of applications. Thus, the performance of the previous systems is further enhanced using improved feature selection, projection and classification algorithms. As processing EEG signals requires dealing with multi-dimensional data, there is a need to minimize the dimensionality in order to achieve acceptable performance with less computational cost. The first possible candidate for dimensionality reduction is employed using channel feature selection approach. Four novel feature selection methods are developed utilizing genetic algorithms, ant colony, particle swarm and differential evolution optimization. The methods provide fast and accurate implementation in selecting the most informative features/channels that best represent mental tasks. Thus, computational burden of the classifier is kept as light as possible by removing irrelevant and highly redundant features. As an alternative to dimensionality reduction approach, a novel feature projection method is also introduced. The method maps the original feature set into a small informative subset of features that can best discriminate between the different class. Unlike most existing methods based on discriminant analysis, the proposed method considers fuzzy nature of input measurements in discovering the local manifold structure. It is able to find a projection that can maximize the margin between data points from different classes at each local area while considering the fuzzy nature. In classification phase, a number of improvements to traditional nearest neighbour classifier (kNN) are introduced. The improvements address kNN weighting scheme limitations. The traditional kNN does not take into account class distribution, importance of each feature, contribution of each neighbour, and the number of instances for each class. The proposed kNN variants are based on improved distance measure and weight optimization using differential evolution. Differential evolution optimizer is utilized to enhance kNN performance through optimizing the metric weights of features, neighbours and classes. Additionally, a Fuzzy kNN variant has also been developed to favour classification of certain classes. This variant may find use in medical examination. An alternative classifier fusion method is introduced that aims to create a set of diverse neural network ensemble. The diversity is enhanced by altering the target output of each network to create a certain amount of bias towards each class. This enables the construction of a set of neural network classifiers that complement each other.
Please use this identifier to cite or link to this item: