EEG-based Gait State and Gait Intention Recognition Using Spatio-Spectral Convolutional Neural Network

Abstract

EEG-based BCI was recently applied to lower limb exoskeleton robots. Various machine learning decoders have shown high accuracy performance on classifying the gait state whether the subject is walking or standing. However, there is a trade-off between the accuracy and the responsiveness due to the delay time. The delay time is critical when controlling the exoskeleton robots with EEG decoders online (real-time). In this research, we propose spatio-spectral convolutional neural networks with relatively short segment of EEG data (0.2s) having 83.4% accuracy on gait state recognition. The gait intention recognition that detects the subject's gait intention prior to the actual gait had 77.3% accuracy. We were able to classify EEG data of both healthy subjects and stroke patients at subacute and chronic phases.