Prediction of Multiple Movement Intentions from CNV Signal for Multi-Dimensional BCI

TLDR: In this paper, a BCI was created that used electroencephalography (EEG) to allow multiple dimensions of control, achieved through the movement or motor imagery of the right hand, left hand, tongue, and right foot.

Abstract: Patients that suffer from loss of motor control would benefit from a brain-computer interface (BCI) that would, optimally, be noninvasive, allow multiple dimensions of control, and be controlled with quick and simple means. Ideally, the control mechanism would be natural to the patient so that little training would be required; and the device would respond to these control signals in a predictable way and on a predictable time scale. It would also be important for such a device to be usable by patients capable and incapable of making physical movements. A BCI was created that used electroencephalography (EEG). Multiple dimensions of control were achieved through the movement or motor imagery of the right hand, left hand, tongue, and right foot. The movements were non-sustained to be convenient for the user. The BCI used the 1.5 seconds of the Bereitschaftspotential prior to movement or motor imagery for classification. This could allow the BCI to execute an action on a time scale anticipated by the user. To test this BCI, eight healthy participants were fitted with 29 EEG electrodes over their sensorimotor cortex and one bipolar electrooculography electrode to detect eye movement. Each participant completed six blocks of 100 trials. A trial included visual presentation of three stimuli: a cross, an arrow, and a diamond. Participants rested during the presentation of the cross. The arrow indicated the action that the participant should perform: right hand squeeze, left hand squeeze, press of the tongue against the roof of the mouth, or right foot toe curl. The diamond indicated that the participant should execute the movement during the first three blocks; and that the participant should imagine executing the movement during the last three blocks. Trials affected by motion artifacts, in particular face muscle activity, were removed. Of the remaining data, about 80% were used to train a Bayesian classification and about 20% were used to test this classification. Prediction of the four movements reached accuracies above 150% that of random classification for both real and imagined movements. This suggests a promising future for this BCI.