Abstract:

BACKGROUND: Feature extraction of electroencephalogram (EEG) signals is an important step in the brain-computer interfaces (BCI) system. Effective and rapid EEG signals feature extraction is important for classification and correct understanding. Currently, power spectrum density estimation, autoregression model and wavelet transform have been used to extract EEG signals features. However, they cannot well reflect nonlinear dynamics of the brain.
OBJECTIVE: To explore the effect of kurtosis of power spectral (KPS) in the recognition of hand imagery.
METHODS: The data gained from BCI competition in 2003 provided by Graz University of Technology. The EEG signals ranging from 8 to 24 Hz were decomposed by wavelet packet. The KPS of C3 and C4 were calculated respectively. The KPS was defined as the feature vector. The left and right hand motor imaginary tasks were distinguished by the time-variable linear classifier.
RESULTS AND CONCLUSION: The proposed method was applied to the test data set with 140 trails. The satisfactory results were obtained with the highest classification accuracy of 89.29%. The maximum mutual information was 0.6269 bit. The Signal-to-Noise Ratio was 1.3848. The KPS on channels C3 and C4 between 8 and 24Hz was coincident with event-related desynchronization and event-related synchronization. The method is simple and quick and it is a promising method for on-line BCI system.