EEG-Based Hand Movement Recognition: Feature Domain and Level of Decomposition

Abstract

This manuscript reports feature domains for the recognition of right and left hand movements using Electroencephalogram (EEG). A 21-channel EEG dataset of seven subjects during right and left hand fist open and close movements was collected from PhysioNet of the BCI2000 Instrumentation system. Features in time, frequency, and time-frequency domains have been explored. Support vector machine with radial basis function kernel was used for the recognition of right and left hand fist open and close movements. The recognition with time, frequency, and time-frequency domain features resulted in an accuracy of 90%, 92%, 97.5%, respectively. Time-frequency domain features obtained through discrete wavelet transform (DWT) at four decomposition levels have resulted in maximum recognition rate. The highest recognition rate of 98.6 \( \pm \) 0.6 % has resulted in DWT features at level 2. This was substantiated by the fact that DWT features at level 2 establish maximum correlation with pre-processed EEG. Experimental result shows that time-frequency is the best performing feature domain among the three. Further, correlation measure of time-frequency domain features with EEG are established as a benchmark for selecting DWT decomposition level.