| This dissertation documents methods for automatic detection and classification of epileptiform transients, which are important clinical issues. There are two main topics: (1) Detection of paroxysmal activities in EEG; and (2) Classification of paroxysmal activities. This machine learning algorithms were trained on expert opinion which was provided as annotations in clinical EEG recordings, which are called "yellow boxes" (YBs).;The dissertation describes improved wavelet-based features which are used in machine learning algorithms to detect events in clinical EEG. It also reveals the influence of electrode positions and cardinality of datasets on the outcome. Furthermore, it studies the utility of using fuzzy strategies to obtain better performance than using crisp decision strategies.;In the yellow-box detection study, this dissertation makes use of threshold strategies and implementation of ANNs. It develops two types of features, wavelet and morphology, for comparison. It also explores the possibility to reduce input vector dimension by pruning. A full-scale real-time simulation of YB detection is performed. The simulation results are demonstrated using a web-based EEG viewing system designed in the School of Computing at Clemson, called EEGnet. Results are compared to expert marked YBs. |
