Imagined speech EEG-based BCI using dynamic mode decomposition

Abstract

To aid individuals with speech impairments, the research aims to decode imagined speech from non-stationary EEG signals using brain-computer interfaces (BCIs). The proposed method combines multichannel EEG signals with time-frequency representations (TFRs) based on Hilbert spectral analysis of modes decomposed by dynamic mode decomposition (DMD). The imagined electroencephalogram (EEG) signals of six imagined speech commands (Up, Down, Left, Right, Upward, and Downward) from 15 subjects sampled at 1024 Hz were measured using a six-concatenated channel standard physiological signal system; the signal was filtered to remove artifacts between 2 and 40 Hz using a finite impulse response pass-band filter. The proposed method employs a convolutional neural network (CNN) model that takes TFRs as input images to decode the imagined speech commands. The parameters of the model are optimized to achieve the best performance in decoding speech commands. In a comparison with the multi-fast and adaptive empirical mode decomposition method, the proposed method utilizing DMD achieves significantly improved decoding accuracy. EEG signals are decoded with an accuracy of 70% for six classes of imagined speech commands. DMD and CNN are integrated in the proposed method to extract dynamic information from imagined speech EEG signals, represented as spatial modes and their instantaneous frequencies (alpha and beta).