| The electroencephalogram (EEG) is an important diagnostic tool in the field of neurology, particularly for the diagnosis of seizure disorders. The EEG can record seizure events to help characterize their location, frequency and intensity. However, because seizures and interictal evidence of seizures occurs infrequently in some types of seizure disorders, it is often necessary to conduct long-term EEG recordings.; To help reduce the difficulty of reviewing and interpreting and interpreting long-term EEG recordings, much work has been done on the field of computer-assisted EEG interpretation. The earliest automated EEG interpretations were based upon analysis of the waveform morphology and/or frequency components of the recording. More recently, attempts have been made to add higher-order processing stages. The most common approaches have used rule-based expert systems and artificial neural networks.; In contrast to rule-based expert systems and artificial neural networks, hidden Markov models and Bayesian networks have received little attention from automated EEG interpretation researchers. This dissertation investigates the hypothesis that an automated EEG interpretation system that combines (1) the ability of hidden Markov models to classify noisy time-series data with (2) the ability of Bayesian networks to integrate uncertain evidence, will perform particularly well at seizure detection.; In particular, I will describe a system using hidden Markov models and Bayesian networks to perform automated seizure recognition in long-term EEG recordings using temporal and spatial context. At the lowest level, the system uses a Fourier transform to extract frequency components from the time series data, and transforms those frequency components into model parameters. At a higher level, the system utilizes a hidden Markov model to represent temporal context, and a Bayesian network to represent spatial context.; While there is considerable variation among cases, it will be shown that in the average case both the hidden Markov layer and Bayesian network layer individually and in combination improve the performance of the system, even if the input is not highly correlated with seizure. In addition, in the best case in which the input is well correlated with seizure, the output of the hidden Markov and Bayesian layers can be even more predictive. |
