Optimal Feature Selection for Seizure Detection: A Subspace Based Approach

Abstract

An epileptic seizure detector's performance definitely depends on features extraction and selection. In this study, we present the short-time average magnitude difference function (sAMDF) as a computationally efficient feature to distinguish seizures from EEG and it is compared with the frequently used curve length. We also suggest using a subspace based approach for feature selection that exploits divergence measure as the dissimilarity criterion. In this approach, basically features are linearly transformed into another reduced space for optimality while decreasing the computational burden. Seizure discrimination performances of transformed features and original features are compared. The obtained results demonstrate that the feature selection with a divergence-based subspace approach is quite useful to discriminate the seizure parts of the signal from the nonseizure ones.