Feature extraction from multichannel EEG transients for clustering purposes

This paper treats the problem of clustering multichannel EEG transients occurring in epilepsy, so called spikes. The authors present a new method for feature extraction which is an expansion along an orthonormal set of matrices, obtained from truncated SVD and Gram-Schmidt orthogonalization. The method is compared to a method used earlier where each channel was Hermite function expanded. The result is that using the same expansion order the new method represents a set of spikes with higher fidelity than the Hermite function method. The output from the feature extraction is a vector for each spike in a set. The vectors from a set is clustered with the standard NM algorithm. Two sets of spikes have been investigated, both being clustered in appearance.