EEG-Based Intersubject Correlations Reflect Selective Attention in a Competing Speaker Scenario

Abstract

Several solutions have been proposed to study the relationship between ongoing brain activity and natural sensory stimuli, such as running speech. Computing the intersubject correlation (ISC) has been proposed as one possible approach. Previous evidence suggests that ISCs between the participants' electroencephalogram (EEG) may be modulated by attention. The current study addressed this question in a competing-speaker paradigm, where participants (N = 41) had to attend to one of two concurrently presented speech streams. ISCs between participants' EEG were higher for participants attending to the same story compared to participants attending to different stories. Furthermore, we found that ISCs between individual and group data predicted whether an individual attended to the left or right speech stream. Interestingly, the magnitude of the shared neural response with others attending to the same story was related to the individual neural representation of the attended and ignored speech envelope. Overall, our findings indicate that ISC differences reflect the magnitude of selective attentional engagement to speech.

Keywords: EEG; attended speaker paradigm; correlated component analysis; intersubject correlation; naturalistic stimuli; selective auditory attention; speech envelope tracking.

FIGURE 1

Participants were instructed to attend to one of two simultaneously presented stories while EEG was recorded. Experimental conditions differed in allocation of attention to the story presented from the front left (yellow) or front right (purple). A 49-channel configuration was used for further analysis. (Intersubject correlation) Correlations between participants’ EEG signals were maximized using Correlated Component Analysis. The matrix on the right illustrates that correlations between participants that also attended to the left or right story are categorized as ISC_same whereas correlations across conditions are categorized as ISC_other. As the correlation matrix is symmetrical one half can be neglected. (Speech envelope tracking) Speech envelope tracking was performed by segmenting the signals into 30 s epochs. For each epoch, the EEG signal of every channel was cross-correlated with the attended (red) and ignored (blue) speech envelope. The figure shows an example when the left stream is attended to, and the right stream ignored.

FIGURE 2

Cross-correlation functions of each channel averaged over epochs and participants. Topographies below show the channel weights, indicating channels that contributed most to the correlations at time lags from −50 to 300 ms in steps of 50 ms. (A) Cross-correlation functions of individual EEG channels with the attended speech envelope. The red line marks the global field power of the cross-correlation (GFP_crosscorr). (B) Cross-correlation functions of individual EEG channels with the ignored speech envelope. The blue line marks GFP_crosscorr of the ignored speech envelope.

FIGURE 3

(A) Individual ISC scores between participants within the same attentional condition (red), and other attentional condition (blue). Participants from the freefield and headphone study are marked with solid gray and dashed black lines, respectively. The horizontal black lines represent 95th percentile of chance-level ISC scores for the same and other group correlation, respectively. (B) Mean and standard error of the mean (SEM) of intersubject correlations within a condition (same) and across conditions (other) for each component. 95th percentile of chance-level correlations is marked in gray. *** indicates significance at p < 0.001. (C) Forward model of the three most correlating components. The rows represent each component starting from the strongest in descending order. The first column contains data from all participants combined. The second and third column exclusively contain data from participants that attended to the left or right story, respectively. Channel weights which indicate channels that contribute most to the correlations between participants are indicated by color.

FIGURE 4

(A) Grand average global field power (GFP) of cross-correlation between the EEG signal and the attended (red) and ignored (blue) speech envelope after averaging over all epochs and participants. Red and blue shaded areas show the confidence interval of the respective function. *** indicates a difference (p < 0.001) between the two functions in the time window of P2_crosscorr, i.e., from 130 to 184 ms time lag. (Bottom) Correlation between ISC and selective attention effect on speech envelope tracking. Attentional modulation of the speech envelope tracking is defined as the difference between the attended and ignored GFP_crosscorr in the P2_crosscorr time window. Each datapoint represents one participant. (B) Correlation with participant’s ISC_same score. (C) Correlation with participant’s ISC_other score.

FIGURE 5

(A) Prediction accuracies assessed by AUC indicate the performance of ISC to correctly classify whether or not a participant attended to the left story (ISC_left) or the right story (ISC_right). Gray-shaded area indicates significance level at p < 0.001. (B) Each datapoint represents the relation of a participant’s ISC_left and ISC_right score. The colors indicate the condition of the participant. The diagonal line illustrates how informative each datapoint was for the models.