Using Brain-Computer Interface to understand and decode neural processes associated with imagined speech

Imagined speech plays a central role in human consciousness, and its understanding is of great importance for the construction of a speech-like Brain-Computer Interface (BCI). Such BCIs would considerably benefit patients suffering from severe conditions where they are unable to verbally communicate, despite being fully conscious. The goal of this study was to build a non-invasive closed-loop BCI for a vowel task. To this end, an offline vowel / syllable / word task was designed to better understand neural processes underlying imagined speech and evaluate binary classification ability. We recorded six subjects brain activities using electroencephalography (EEG) during the offline task (listening to auditory stimuli and subsequently imagining saying them) and performed a comparative analysis between different classification approaches. The best method consisted of a semi-automatic feature selection from a list of 41 EEG subband spectral features, followed by principal component analysis and nearest centroid classification. Using this method, vowel vs rest average classification accuracies were 68.9 ± 12.4%s.d. (best subject 85.5%) and 71.1 ± 12.7% (90.6%) for imagined speech and auditory stimulation conditions, respectively (p < 0.05). For the best pairs, vowel vs vowel corresponding average classification accuracies were 58.7 ± 6.0% (63.5%) and 60.4 ± 3.9% (65.9%). The online task, where users generated vowel speech imagery signals to control a ball on a screen, resulted in 67.7 ± 10.9% (83.9%) classification accuracy for the best pair. We conclude non-invasive speech imagery BCIs are within reach, but their applicability still requires efforts in signal acquisition, processing and classification improvement, and in investigating the neural principles behind imagined speech.