Investigating the neural correlates of imagined speech: An EEG-based connectivity analysis

The main objectives of this work are to design a framework for imagined speech recognition based on EEG signals and to represent a new EEG-based feature extraction. In this paper, after recording signals from eight subjects during imagined speech of four vowels (/æ/, /o/, /a/ and /u/), a partial functional connectivity measure, based on the spectral density of correntropy has been set up, and the brain connectivity has been analyzed. Then, the inter-regional connectivity features are defined and calculated based on statistically significant connections. Finally, selected features have been classified by SVM method. Results show a significant difference (p<0.05) between the connectivity patterns of imagined speech and the baseline in some frequency bands. The average classification accuracy for eight subjects is 81.1%. Among other findings of this study are inter-regional connectivity patterns and frequency bands during imagined speech. The proposed method outperforms the accuracy of the competing methods.