![]() | Only 14 pages are availabe for public view |
Abstract Computer Interface (BCI) processes the brain activity signals or the Electroencephalogram (EEG) and translates it into system commands using pattern recognition algorithms. In the present work, our aim is to make a start in the field of BCI and to investigate the possibility of enhancing the performance of Mental Task BCI systems. The presence of artifacts, such as eye blinks, in EEG recordings obscures the underlying processes and makes analysis difficult. Independent component Analysis Algorithm (ICA) has already shown its value in processing EEG and removing eye blink artifacts, but basically it is an offline method inappropriate for online implementation. In this thesis, we present a modification for the Independent Component Analysis (ICA) technique for preprocessing EEG segments commonly handled by rejecting the contaminated EEG segments. The modified ICA-based preprocessing technique automatically identifies and removes eye blink artifacts from EEG segments efficiently which prevents considerable data loss due to artifact rejection. The results show that our ICA-based preprocessing enhances the classification performance significantly over classification of raw unprocessed EEG segments. In addition, the system usability was improved by avoiding artifact rejection commonly employed in previous studies which lowers the practicality for the user having to repeat the rejected trials to avoid considerable loss of data. In addition, a comparative study is performed in this thesis in order to compensate the lack of published objective comparisons between different pattern recognition algorithms for classification of mental tasks from EEG signals by testing them within the same context and using the same protocol. For this purpose, Feed-Forward Neural Network (ANN), Radial Basis function Support Vector Machine (RBF-SVM), Linear and Quadratic Discriminant classifiers (LDC and QDC) are applied to a known EEG dataset in the BCI field. Features are extracted using Parametric Autoregressive modeling (AR), Autoregressive spectral analysis (PSD) and power differences between four frequency bands (Pdiff). Training was performed using cross-validation to control overfitting. In general, the best classification results for all feature extraction techniques were achieved using neural network classifier; the maximum accuracies for AR, PSD, and Pdiff are 71%, 73%, and 72% respectively averaged over all subjects. Features based on the frequency domain properties are more discriminative and achieve better results. |