Kegler, M. , Etard, O. , Forte, A. & Reichenbach, T.
Department of Bioengineering, Imperial College London, Kensington, London SW7 2AZ, UK
We recently introduced a method for measuring the auditory brainstem responses to natural speech with a few scalp electrodes in a bipolar montage, and used the method to show a significant attentional modulation of the recorded response (Forte et al., eLife 2017). Here we sought to develop a statistical model to detect the auditory brainstem response to continuous speech from high-density EEG recordings and use it to decode attention to one of two speakers in an 'on-line' mode from only short time segments.
We employed regularized linear regression to map a fundamental waveform, that oscillates at the fundamental frequency of the voiced parts of speech, to the multichannel EEG signal at different delays. To decode attention to one of two competing speakers through the detected brainstem activity we employed linear models to reconstruct the fundamental waveform of an attended and ignored speaker from the neural recordings. With linear discriminant analysis, we classified the reconstruction accuracies according to the attentional focus.
The estimated latency of 9 ms and the topography of the detected brainstem responses were consistent with previous studies that employed short, repeated speech stimuli. Testing the developed classification method on short segments of EEG data we obtained a high accuracy of 90% using several recording channels only.
We showed that it is possible to detect the auditory brainstem response to natural speech from high-density EEG recordings. This allows for measuring the response of the auditory brainstem as well as of cortical responses to continuous speech simultaneously. The decoding of a subject's attentional focus from short segments of EEG data demonstrates that attention can be assessed fast and effectively from the auditory brainstem response to continuous speech. The accurate classification obtained with only few EEG channels makes the method computationally inexpensive and applicable in mobile devices such as hearing aids.