Using Hand-Crafted and Learned EEG Features for the Detection of Epileptic Seizures

Electroencephalogram (EEG) signals have been used extensively to detect epilepsy seizures. The classic flow includes the extraction of meaningful features, followed by the classification of the resulting feature vectors. The relevance of features has a huge impact on detection accuracy. To capture potential brain interactions related to epileptic episodes, this paper introduces feature extractors that fuse different types of features gathered from all electrodes. The features are defined in the frequency and time domains or learned via autoencoders. The effectiveness of unsupervised feature learning is analyzed for autoencoders applied on data clusters separately, or on all data in one step. The experimental results show improved performance when features are extracted from all electrodes, by fusing learned and hand-crafted attributes. Compact relevant encoding is produced using autoencoders or the histogram of variations between successive optima.