EEG Data Analysis, Feature Extraction and Classifiers

Epileptiform transients (ETs) are an important kind of EEG signal. They have various morphologies and can be difficult to detect. This thesis describes several approaches to detecting and classifying epileptiform transients (ETs), including Bayesian classification (with Gaussian Assumption), artificial neural networks (Backpropagation FeedForward Network) and k-NNR. Various features were extracted, including the shape, frequency domain and wavelet transform coefficients. The long term goal of this research is to determine the required size of a dataset to obtain clinically significant machine classification results. The immediate goal is to identify a reasonable feature set which can achieve acceptable classification performance with reasonable computational complexity. We have explored the effect on the results by changing window size, by filtering and by adding spatial information. Unsupervised methods, e.g. clustering, have also been explored. Presently, ANN provides the best classification method using a wavelet set for the best feature set. Future directions are indicated