Convolutional neural network approach for motor imagery and steady-state somatosensory evoked potential-based hybrid brain-computer interface using dry electrodes

Abstract

Recently, brain-computer interfaces (BCIs) based on steady-state somatosensory evoked potential (SSSEP) using selective attention (SA) have been developed to overcome class number limitations and non-natural gaze. In this study, we compared the performance of motor imagery (MI), SA using vibration stimuli, and a hybrid (HY) BCI combining MI and SA using dry electrodes. Eight healthy subjects participated in the experiments, during which they performed MI, SA, and HY tasks while wearing an electroencephalogram (EEG) head cap with 23 dry electrodes. Based on the displayed instructions, the subjects imagined the hand-grasping actions (left or right) during the MI tasks and focused on one of the vibration stimuli during the SA task. In the HY task, the subjects performed both tasks simultaneously. The acquired EEG data from each task were then classified using a convolutional neural network-based method. The results of the 10-fold cross-validation, considering the left-hand, right-hand, and idle classes revealed that the HY task exhibited an accuracy of 63.9 f 10.4 % (Mean f SD), which was significantly higher than that of the MI (52.1 f 11.7 %) and SA (62.9 f 10.4 %) tasks. Furthermore, there was no statistical difference in accuracy when comparing the use of 5 channels, 9 channels, and all 23 channels for the MI, SA, and HY tasks. These experimental results support the notion that the HY-BCI can achieve higher accuracy even in environments using dry electrodes.