EEG Subspace Representations and Feature Selection for Brain-Computer Interfaces

Electroencephalogram (EEG) signals recorded from a persons scalp have been used to control binary cursor movements. Multiple choice paradigms will require more sophisticated protocols involving multiple mental tasks and signal representations that capture discriminatory characteristics of the EEG signals. In this study, six-channel EEG is recorded from a subject performing two mental tasks. The signals are transformed via the Karhunen-Loéve or maximum noise fraction transformations and classified by quadratic discriminant analysis. In addition, classification accuracy is tested for all subsets of the six EEG channels. Best results are approximately 90% correct when training and testing data are recorded on the same day and 75% correct when training and testing data are recorded on different days.