Figure 4 | Scientific Reports

Figure 4

From: Preparatory activity of anterior insula predicts conflict errors: integrating convolutional neural networks and neural mass models

Figure 4The alternative text for this image may have been generated using AI.

The spatio-spectral preparatory neural signature of conflict errors. (a) The spectral density of the convolutional layer for each ROI indicates the involvement of broad-band arrhythmic activity of cognitive control network in conflict processing. (b) Bar plots illustrate the mean absolute error (MAE) of prediction, in the occlusion experiment, arranged in a descending order across the ROIs. The error bars represent the standard error of the mean (s.e.m.). A higher MAE for a given ROI indicates a more pronounced adverse effect of that specific ROI’s removal on the prediction of the CNN, and consequently, on cognitive control performance. The point at which the MAE begins to decrease significantly in relation to ROIs is denoted by an asterisk on the third ROI, suggesting that preparatory activity of the initial three ROIs (i.e., the right anterior insula, right precentral gyrus, and the right pars opercularis) exerted the most substantial influence on the CNN’s prediction of cognitive control performance. The inner violin plot shows the distribution of the MAEs, and the scatter plot shows individual MAE values. Interquartile ranges are shown by the black box, and the white dot within the black box indicates the median of MAE. (c) The Jansen-Rit neural mass model is comprised of three interconnected neural populations, namely pyramidal projection neurons, excitatory and inhibitory inter-neurons forming feedback loops. The core conflict processing network was modeled as a set of coupled Jansen-Rit models, (d) through a connectivity network built from functional connectivity analysis of the simultaneous fMRI dataset. (e) The time delays of these connections were estimated based on inter-regional Euclidean distances and a conduction speed of 3 m/s. (f) Through a grid-search framework spanning the interval of 0.0001 to 0.05, a global coupling scale factor "c" was systematically fine-tuned to uniformly adjust all connection weights. For every "c" value, a set of simulated EEG signals was produced, which was subsequently fed into the pre-trained CNN to calculate the Hit score. The optimal parameter "c" was defined as the value that yielded a Hit score significantly above the chance level, determined through a 5000-resampling bootstrap test. The distribution of the accuracies associated with tested "c" values is shown in the violin plot in the right panel. The scatter plot within the violin plot shows individual accuracies. The black box represents interquartile ranges, and the white dot within the black box represents the median of the accuracy. (g) The Hit score was monitored as a sine wave stimulation was applied to each node of the neural mass model. The stimulation was characterized by a normalized amplitude "a" spanning the range of 0.5 to 5 and a frequency "f" in the range of 2 to 200 Hz. The largest change in the Hit score was found for the left anterior insula (left heatmap). A 5000-randomization test was carried out to create the statistical map (right heatmap) and to assess the statistical significance of improvement in the Hit score followed by each stimulation. Applying stimulation with a frequency of approximately 130 Hz and a normalized amplitude of around 1, targeted at the left anterior insula, led to the most significant improvement in the Hit score. The histogram indicates the distribution of the enhanced Hit scores (i.e., cognitive enhancement) for different stimulation protocols. The warm-colored cluster around 0.60 in the histogram is related to 130 Hz stimulation of left anterior insula.

Back to article page