| Deep neural networks can be used for abstraction and as a preprocessing step for other machine learning classifiers. Our goal was to develop methods for a more accurate automated seizure detection. Deep architectures have been used for classification of events, and shown in this research to be an effective way of classifying multichannel high resolution medical data. The medical data used in this thesis was gathered from an electroencephalograph (EEG) used in a hospital setting on seizure patients.;To demonstrate the ability of deep architectures to learn and abstract from input data, the signals from the EEG that contained both seizure and non seizure data were given both as featurized data and raw data to the deep architecture. In addition to the multiple types of data preparation, a patients EEG data was tested not only against their own EEG signal training data but other patients as well. This study supports the effectiveness of deep feed forward neural networks for usage in the seizure classification scenario, as well as highlights some of the difficulties associated with training deep neural networks, as shown through experimental results. |
