Development of Efficient Algorithms for the Classification of Imagined Speech

Abstract

newline The completely paralyzed and quadriplegic patients cannot communicate with others. newlineHowever, the imagined thoughts of these patients can be used to drive assistive devices by newlinebrain-computer interfacing (BCI), the success of which relies on better classification newlineaccuracies. This work examines the possibility of better feature extraction techniques from newlineelectroencephalography (EEG) signals of imagined speech, from which a robust model can be newlinebuilt with the help of a classifier. The feature extraction techniques like spatial features derived newlinefrom common spatial patterns (CSP), hybrid time-spatial, and hybrid time-frequency-spatial newlinefeatures are used to discriminate between imagined vowels /a/, /u/, and their corresponding newline/rest/ intervals. The random forest classifier, combined with hybrid time-spatial features, newlinedecoded the tasks with the highest accuracy reaching 91%. Moving on, we have performed an newlineexperiment in which a 32-channel industry-standard EEG device was used to record 26 newlineimagined English alphabets from 13 subjects. We denoised the imagined signals by discrete newlinewavelet transform and extracted the spatial filters by the CSP method and time-domain newlinefeatures. Spatial features, when classified with linear support vector machine and time-domain newlinefeatures classified by random forest, gave the best results. Alpha, beta, and theta bands could newlineclassify imagined alphabets better than other bands and had average classification accuracies newlineof 88.59%, 87.39%, and 88.97%, respectively, by using spatial features and 81.88%, 76.72%, newlineand 79.25%, respectively by time-domain features. The grand average accuracies of all the 26 newlinealphabets in six EEG frequency bands was 77.97% in a subject-independent binary newlineclassification framework.