Comparing MEG and EEG measurement set-ups for a brain-computer interface based on selective auditory attention

Abstract

Auditory attention modulates auditory evoked responses to target vs. non-target sounds in electro- and magnetoencephalographic (EEG/MEG) recordings. Employing whole-scalp MEG recordings and offline classification algorithms has been shown to enable high accuracy in tracking the target of auditory attention. Here, we investigated the decrease in accuracy when moving from the whole-scalp MEG to lower channel count EEG recordings and when training the classifier only from the initial or middle part of the recording instead of extracting training trials throughout the recording. To this end, we recorded simultaneous MEG (306 channels) and EEG (64 channels) in 18 healthy volunteers while presented with concurrent streams of spoken "Yes"/"No" words and instructed to attend to one of them. We then trained support vector machine classifiers to predict the target of attention from unaveraged trials of MEG/EEG. Classifiers were trained on 204 MEG gradiometers or on EEG with 64, 30, nine or three channels with trials extracted randomly across or only from the beginning of the recording. The highest classification accuracy, 73.2% on average across the participants for one-second trials, was obtained with MEG when the training trials were randomly extracted throughout the recording. With EEG, the accuracy was 69%, 69%, 66%, and 61% when using 64, 30, nine, and three channels, respectively. When training the classifiers with the same amount of data but extracted only from the beginning of the recording, the accuracy dropped by 11%-units on average, causing the result from the three-channel EEG to fall below the chance level. The combination of five consecutive trials partially compensated for this drop such that it was one to 5%-units. Although moving from whole-scalp MEG to EEG reduces classification accuracy, usable auditory-attention-based brain-computer interfaces can be implemented with a small set of optimally placed EEG channels.

Fig 1. Experimental design and conditions.

(A) Spoken-word stimuli recorded with a dummy head were presented to the participant. In the acoustic scene, the two speakers appeared at  ± 40 degrees of the participant’s head mid-line. (B) The stimulus onset asynchrony (SOA) was 1 s, leaving about 400-ms silent interval between consecutive stimuli. The black triangles indicate stimulus onsets. (C) The high- and low-pitch version of the words alternated (standard) but occasionally two same-pitch words were presented consecutively (deviant). (D) Classifiers were trained with trials extracted randomly across the entire recording (random training), or with trials only from the beginning of the recording (early training), or with trials extracted randomly from the latter half of the measurement (random late training), or with trials from the middle of recording (middle training).

Fig 2. Layout of the EEG channel selections in the measurement.

The 64-channel EEG set (all dots), and the subsets of 30 (blue dots and all circles), nine (all blue circles) and three (thick blue circles) channels.

Fig 3. Classification accuracy and information transfer rate (ITR) as a function of the measurement set-up and the number of combined trials.

(A) Classification accuracy when classifiers were trained with 25% of the data extracted randomly across the entire recording (grey) or with only the first quarter of the data (blue, green). (B) Classification accuracy when classifiers were trained with the middle blocks B9–B12 and tested with last four blocks B13–B16. The dots in panels A and B represent the mean accuracy across the 11 participants, and the shading the standard error of mean. (C) Information transfer rate as bits/min with training as in panel A. (D) Same as in Panel C but training as in Panel B. The dots in Panels C and D represent the mean ITR across the 11 participants, and the shading the standard error of mean.

Fig 4. Classification accuracy as a function of the amount of training data.

The dots in the learning curves represent the mean classification accuracy across the 11 participants. The shading indicates the standard error of mean calculated over 20 iterations to each training set size.