Decoding auditory working memory content from EEG responses to auditory-cortical TMS

Abstract

Working memory (WM), short term maintenance of information for goal directed behavior, is essential to human cognition. Identifying the neural mechanisms supporting WM is a focal point of neuroscientific research. One prominent theory hypothesizes that WM content is carried in "activity-silent" brain states involving short-term synaptic changes. Information carried in such brain states could be decodable from content-specific changes in responses to unrelated "impulse stimuli". Here, we used single-pulse transcranial magnetic stimulation (spTMS) as the impulse stimulus and then decoded content maintained in WM from EEG using multivariate pattern analysis (MVPA) with robust non-parametric permutation testing. The decoding accuracy of WM content significantly enhanced after spTMS was delivered to the posterior superior temporal cortex during WM maintenance. Our results show that WM maintenance involves brain states, which are activity silent relative to other intrinsic processes visible in the EEG signal.

Keywords: Auditory working memory; MVPA; TMS-EEG; Transcranial magnetic stimulation; Working memory.

Fig. 1.

Task design. (a) Examples of the modulation patterns for ripple sounds (b) The target brain area for the TMS pulse, adapted from Uluç and colleagues [21]. (c) The auditory WM retro-cue paradigm. The timeline of events in one trial is depicted.

Fig. 2.

MVPA pipeline. Preprocessed whole head EEG data was entered into a PCA for spatial feature selection. The cut-off for the PC selection (n_PC = 8) was determined based on the elbow in the grand-average cumulative variance curve, calculated across all conditions, WM classes, and participants. Then the data was entered into a searchlight analysis with cross-participant 4-class SVM classification. Leave-one-out method was used for cross-validation. The analysis resulted with a decoding accuracy (DAc) time series where DAc is the value assigned to the centroid of searchlight sliding window. For statistical significance, we used maximum statistics with 500 permutations.

Fig. 3.

The results of E-field calculations and searchlight MVPA decoding of WM content from EEG. (a) Group median E-field maps for the Active TMS condition. (b) Group median E-field maps for the Control TMS condition. (c) Null distribution for 500 permutations, utilized to determine the statistical significance of decoding accuracies. From each permutation, the maximum cluster sum of normalized decoding accuracy was identified and added to this null distribution. The vertical dotted line illustrates the critical value for p_corrected<0.05 (cluster sum(z) = 130). (d) Decoding accuracies in the cross-participant four-class MVPA (% of correctly classified trials). The time series reflect the SVM decoding accuracies at the centroid of each sliding 300-ms searchlight. These decoding accuracies were derived from an iterative leave-one-participant out cross-validation procedure: In each searchlight time window, the data of each participant was used once as the test set and those from the rest of the remaining participants as the training set. The light red line denotes the Active TMS condition and the light blue line the Control TMS condition. The dotted horizontal line indicates the chance level of decoding-accuracy in a four-class classification (25 %). The time window when the decoding accuracy was significantly higher than chance level in the Active TMS condition is shown in dark red (p_Corrected<0.05, non-parametric cluster-based permutation test).

Fig. 4.

The results of within-participant searchlight MVPA decoding of WM content from EEG for Control and Active TMS conditions. (a) Decoding accuracies in the within-participant four-class MVPA. The time-resolved decoding reflects the accuracies at the centroid of each sliding 300-ms searchlight. The thin red line denotes the Active TMS condition and the thin blue line the Control TMS condition. The dotted horizontal line indicates the chance level of decoding-accuracy in a four-class classification (25 %). The transparent red and blue areas denote the standard error of the mean. The time window when the decoding accuracy was significantly higher than chance level in the Active TMS condition is shown in dark red (p_Corrected<0.05, non-parametric cluster-based permutation test). (b) Null distribution for 500 permutations, utilized to determine the statistical significance of decoding accuracies. From each permutation, the maximum cluster sum of decoding accuracy was identified and added to this null distribution. The vertical dotted line illustrates the critical value for p_corrected<0.05 (cluster sum of decoding accuracy = 24.5).

Fig. 5.

Topographical and butterfly plots of ERP time courses for time of interest (TOI) in the active TMS condition. Different colors in the ERP plots refer to different electrodes. The electrode map above the figures denotes the locations of electrodes. The timeline starts from t = 0 s at the visual retro-cue. (a) TOI ERP topographical plots and time courses for WM before TMS pulse. The plot depicts the results for the visual retro-cue that starts the maintenance period. (b) TOI ERP topographical plots and time courses after TMS pulse. The timeline starts from t = 0 s at the TMS pulse. (c) TMS evoked response for Active TMS session. (d) TMS evoked response for Control TMS session.