Comparing EEG artifact detection methods for real-world BCI

Abstract

One major challenge to the real-world use of brain-computer interface (BCI) technology is the decrease in classifier performance caused by degradations in electroencephalogram (EEG) signal quality due to artifacts from non-neural electrophysiological activity and the gross movement of sensors and other EEG hardware. These artifacts can contaminate or mask the neural signal and thus cause a decrease in the performance of BCI classifiers due to the system’s diminished ability to extract relevant features. One strategy to combat this effect is to identify and remove artifact-contaminated segments of data. We compared four methods that utilize higher order statistics to detect and artifact data on their ability to improve BCI classifier performance. We evaluated these methods on two datasets: a motor movement task and a rapid serial visual presentation (RSVP) task. In addition to comparing artifact detection methods, we compared the improvement in BCI classifier performance gained by removing artifact data to the decrease in performance caused by diminishing the amount of data available for classifier training. We found that overall the use of abnormal spectra to detect artifacts resulted in the greatest improvement to BCI classifier performance.