EEG classification of imagined syllable rhythm using Hilbert spectrum methods

We conducted an experiment to determine whether the rhythm with which imagined syllables are produced may be decoded from EEG recordings. High density EEG data were recorded for seven subjects while they produced in imagination one of two syllables in one of three different rhythms. We used a modified second-order blind identification (SOBI) algorithm to remove artefact signals and reduce data dimensionality. The algorithm uses the consistent temporal structure along multi-trial EEG data to blindly decompose the original recordings. For the four primary SOBI components, joint temporal and spectral features were extracted from the Hilbert spectra (HS) obtained by a Hilbert-Huang transformation (HHT). The HS provide more accurate time-spectral representations of non-stationary data than do conventional techniques like short-time Fourier spectrograms and wavelet scalograms. Classification of the three rhythms yields promising results for inter-trial transfer, with performance for all subjects significantly greater than chance. For comparison, we tested classification performance of three averaging-based methods, using features in the temporal, spectral and time-frequency domains, respectively, and the results are inferior to those of the SOBI-HHT-based method. The results suggest that the rhythmic structure of imagined syllable production can be detected in non-invasive brain recordings and provide a step towards the development of an EEG-based system for communicating imagined speech.