Performances among various common spatial pattern methods for simultaneous MEG/EEG data

Abstract

Brain Computer Interface (BCI) is a communication pathway between devices (computers) and the human brain. It treats brain signals in a real-time basis and deciphers some of what the human brain is doing to give us certain information. In this work, we develop the BCI system based on simultaneous electroencephalograph (EEG) and magnetoencephalography (MEG) using various preprocessing and feature extraction methods along with Fisher linear discriminant analysis (FLDA) classifier. Common spatial pattern (CSP) is a spatial filter whose spatially projected signal has maximum power for one class and minimum power for the other. Each single trial is computed by the variance in the time domain. We choose a proper number of patterns in order to make a feature vector. In this work, 6 CSP patterns, the first three and the last three ones are selected. A feature vector consists of 6 variances of each extracted CSP pattern from projected data. Among various CSP methods, we used normal common spatial patterns (CSP), invariant common spatial patterns (iCSP), and common spectral spatial patterns (CSSP) methods to measure the performances. Simultaneous MEG/EEG datasets (340 channels) for four subjects from Eleckta Vectorview system were digitally acquired at a 1 KHz and 8-30Hz bandpass filtered. Total 340 channels consist of three kinds of channel types such as 102 magnetometers, 204 gradiometers and 40 EEG electrodes. Three different modalities such as EEG-only, MEG-only, and simultaneous MEG and EEG were analyzed in order to study comparative BCI performances on three variants of CSP. Particularly, for simultaneous MEG/EEG data we proposed three different combination ways for BCI and their performances were discussed.