3D Hand Movement Velocity Reconstruction using Power Spectral Density of EEG Signals and Neural Network

Three dimensional (3D) limb motion trajectory is predictable with a non-invasive brain-computer interface (BCI). To date, most non-invasive motion trajectory prediction BCIs use potential values of electroencephalographic (EEG) signals as the input to a multiple linear regression (mLR) based kinetic data estimator. We investigated the possible improvement in accuracy of 3D hand movement prediction (i.e., the correlation of registered and reconstructed hand velocities) by replacing raw EEG potentials with spectrum power values of specific EEG bands. We also investigated if a non-linear neural network based estimator outperformed the mLR approach. The spectrum power model provided significantly higher accuracy (R~0.60) compared to the similar EEG potentials based approach (R~0.45). Additionally, when replacing the mLR based kinetic data estimation module with a feed-forward neural network (NN) we found the NN based spectrum power model provided higher accuracy (R~0.70) compared to the similar mLR based approach (R~0.60).