Title:
Probabilistic Enhancement of EEG Component Using Prior Distribution of Correlations Between Channels

Authors:
Hayato Maki, Tomoki Toda, Sakriani Sakti, Graham Neubig, Satoshi Nakamura

Institution:
Nara Institute of Science and Technology

abstract:
EEG is increasingly used for study on cognitive process or creating brain machine interfaces (BMI) but its signal-noise ratio is very low, which presents serious problems for EEG interpretation and analysis from EEG recordings.

Synchronous addition of trials to cancel out background noise or rejecting trials contaminated by eye blinks cause substantial data loss. Therefore, a technique to separate each component of EEG observation is seriously required.

Blind source separation using independent component analysis(ICA) assumes that the number of independent signal sources is the same as the number of sensors. The assumption is questionable in the context of EEG signal separation because we do not know the effective number of statically independent brain signals contributing to the EEG.

In this research, we employ probabilistic model to separate signals. It assumes that each EEG component's amplitude follows Gaussian distribution in every slot of time-frequency domain. We introduce a prior distribution of correlations between EEG channels that follows Inverse Wishart distribution. Finally, the pattern recognition was carried out, which demonstrated the effectiveness of the proposed approach.