Identification of Relevant Inter-channel EEG Connectivity Patterns: A Kernel-Based Supervised Approach

TLDR: A kernel-based inter-channel connectivity relevance analysis is proposed, for such a purpose, linear dependencies between channel signals are extracted using coherence measures over specific sub-frequency bands, and a similarity criterion is implemented to rank the contribution of each channel-to-channel connection for a specific task.

Abstract: Extraction of brain patterns from electroencephalography signals to discriminate brain states has been an important research field to the develop of non-invasive applications like brain-computer-interface systems or diagnosis of neurodegenerative diseases. However, most of the state-of-the-art methodologies use observations derived from each electrode independently, without considering the possible dependencies between channels. To improve understanding of brain functionality, connectivity analysis have been developed. Nevertheless in those works, where connectivity measures are included, the total number of connections is high dimensional, and the relevance of connectivity values is not considered. To cope with this issue, we propose a kernel-based inter-channel connectivity relevance analysis (termed ConnRA), for such a purpose, linear dependencies between channel signals are extracted using coherence measures over specific sub-frequency bands, and a similarity criterion is implemented to rank the contribution of each channel-to-channel connection for a specific task. Experimental validation carried out on a database of brain-computer interfaces, demonstrate very promising results, making the proposed methodology a suitable alternative to support many neurophysiological applications.