Convolutional Neural Network Approach for Steady-State Somatosensory Evoked Potential-based Robotic Exoskeleton Control

Abstract

Steady-state somatosensory evoked potentials (SSSEP)-based brain-computer interface (BCI) has been developed to control real-life assisting robots, such as an exoskeleton for body-disabled patients. Although the SSSEP-BCI system has been advanced by various research groups, improving the accuracy is required for applying stable control of the appliance. Moreover, the SSSEP-BCI study on body-disabled patients is still required to confirm applicability. In this study, the convolutional neural network (CNN) was applied to the SSSEP-BCI and compared with the regularized linear discriminant analysis (RLDA) as the widely used method. For extracting the spatial feature, the common spatial pattern (CSP) was used in both methods, and the efficiency of the CNN was validated via the SSSEP data from five healthy subjects and two spinal cord injury (SCI) patients. As a result of 9-fold cross-validation, the CSP-CNN method showed significantly higher accuracy than CSP-RLDA by 7.8% (Healthy subject) and 7.2% (SCI patients). We can therefore conjecture that CNN can improve the performance of SSSEP-BCI.