Feature extraction in development of brain-computer interface: a case study

Abstract

In this study we compared two spectral estimation methods for feature extraction in development of computer access for communication and control based on EEG signals. We compared power spectrum calculated by fast Fourier transform (FFT) and autoregressive (AR) coefficients calculated by Burg's algorithm. We analyzed four surface EEG signals recorded above sensory-motor areas, while the subject was attempting to use only mental activities to modulate his EEG signals resulting in desired simple movements of animated object on the feedback computer screen. To establish communication channel based on EEG signals, we asked our subject to determine which mental activity produced reproducible control actions. During the on-line training we used AR coefficients of recorded EEG signals for the feedback generating variables, i.e. the object's movement direction and speed were dependent on the EEG pattern. Our subject was able to determine which mental activity resulted in desired movements and to learn to control object movements with more than 90% accuracy. We performed off-line analysis to determine if FFT or AR can be more successful in extracting relevant information from the EEG signals before performing pattern recognition by a machine learning classifier. Probability of misclassification was used as the outcome measure. Off-line analysis showed that the overall best classification accuracy achieved with AR exceeds that achieved with FFT.