Feature space embeddings for extracting structure from single channel wake EEG using RBF networks

This paper concerns structure extraction for the analysis of real-world single channel wake EEG signals. The long term objective is a study of vigilance in wake EEG. The authors compare the use of PCA and ICA-motivated embeddings based on single channel EEG data, to increase signal-to-noise ratio and to obtain a robust feature space which is more amenable to characterisation by neural networks. The work is novel in several respects: the use of unsupervised data (most work in this area is through evoked potential response experiments), the treatment of EEG by short window embeddings directly from the time domain (most work averages 30 second segments of power spectra), the comparison of ICA and PCA embeddings on single channel data (most ICA work so far reported has employed multiple sensor channels) and the use of a `NeuroScale´ architecture for topographic feature extraction