Changes in brain rhythms and connectivity tracking fear acquisition and reversal

The experiments took place at the Centre for studies and research in Cognitive Neuroscience (CsrCN) of the University of Bologna. Twenty healthy volunteers were recruited. All participants were right-handed and had normal or corrected to normal vision and reported no medical or psychiatric illness. One participant did not complete the experimental session because of a fainting and was excluded from the dataset. Thus, nineteen participants have completed the study (8 males, mean age = 23.48, std = 1.85). The study followed the American Psychological Association Ethical Principles of Psychologists and Code of Conduct and the Declaration of Helsinki and was approved by the local Bioethics Committee of the University of Bologna (Protocol number 71559). Each participant signed an informed consent prior to the start of the experiment and all data were analyzed and reported anonymously.

The experiment consists of two phases. During the first phase, the acquisition phase, participants view two different visual stimuli (Japanese hiragana) on screen (Starita and di Pellegrino 2018; Starita et al. 2019b). One hiragana, in the following denoted as Image 1, is used as a control stimulus (CS-), while the other hiragana, denoted as Image 2, acts as a conditioned stimulus (CS +). That is, following 50% of visualizations of the CS + stimulus, an aversive shock (unconditioned stimulus, US) was administered, whereas no shock ever occurred after the CS- stimulus. During the second phase, the reversal phase, the hiragana are reversed, so that Image 2 acts as the new CS- while Image 1 acts as the new CS + . The unconditioned stimulus (US) consists of a 2 ms electrostatic stimulation (Effting and Kindt 2007; Krypotos et al. 2014, 2015; Starita et al. 2022; Stemerding et al. 2022) administered in the right wrist (dominant hand) through a Digitimer stimulator (model DS7A, Digitimer Ltd., M = 51.25 mA, SD = 1.88 mA) via Ag/AgCl pregelatinized electrodes (Friendship Medical, SEAg-S-15000/15 × 20). The intensity of the shock is calibrated for each subject, via verbal feedback and an ascending level procedure, up to a level defined as 'very annoying and unpleasant, but never painful'. Figure 1 illustrates the experimental paradigm.

The task is generated thanks to the OpenSesame software (Mathôt et al. 2012). Each phase of the task (both acquisition and reversal) included two blocks (i.e., we have 4 blocks in total: Acq1, Acq2, Rev1 and Rev2), which were interspersed with a 5-min break. For each block, 40 stimuli were shown on screen (i.e., 40 trials, 20 per CS). Each stimulus was presented for 6 s, with an interstimulus time interval ranging from 11 to 14 s. In each block, stimuli followed each other in random order, with only constraints on the first two trials (one CS- and one CS + /US) and on the number of consecutive stimuli of the same type, never more than two.

Subjects received no information regarding which visual stimulus would be associated with the shock, only the following instructions at the beginning of each block were provided "You will see two different images, which will appear one at a time on the screen. Occasionally, the image may shock you. Your task is to figure out which image will shock you. Press any key to get started." The CS-US relationship is then learned from scratch with experience. During each block, the subjects' electroencephalographic signal and skin conductance were recorded.

Skin conductance was recorded during each block at 1000 Hz (10 Hz low-pass filter, gain switch set to 5) through the BIOPAC MP-150 system (Goleta, CA). Pregelatinized electrodes (BIOPAC EL501) were connected on the palmar surface of the left non-dominant, non-shocked hand. The signal was then digitized and down sampled at 200 Hz using Autonomate software (version 2.8, Green et al. 2014) to detect trough-to-peak SCR values. SCR was considered valid if trough-to-peak deflection began between 0.5 s and 4.5 s after Image onset, lasted no longer than 5 s, and was greater than 0.02 µS. Skin conductance values from trials that did not meet all of these criteria were set as zero SCR and retained in the analyses (Starita et al. 2019a).

Note that, in the present work, SCR is used for the sole purpose of demonstrating the correct acquisition and reversal of fear in our subject group. The relationship between SCR and brain rhythms (i.e., between the central nervous system and autonomic responses) was analyzed in the previous work. Similarly, subjective ratings of CS-US contingency awareness and CS valence collected at the end of each block confirmed acquisition and reversal at an explicit level. The details of such results can be found in (Starita et al. 2023).

EEG signals were collected from all participants through 63 wet Ag/AgCl electrodes. Each signal was referenced to the FCz and grounded to the FPz electrode. For all participants, the EEG signals were amplified by a BrainAmp DC amplifier (Brain Products, Gilching, Germany) and digitized at a sampling rate of 1000 Hz. Data corresponding to the four experimental blocks (Acq1, Acq2, Rev1 and Rev2) were then exported to MATLAB R2021a (MathWorks Inc., Natick MA, USA) and processed offline through custom-made scripts. A schematic illustration depicting the various processing steps is reported in the supplementary information (SI_1). Firstly, data were down-sampled at 500 Hz and processed with both a band-pass filter (1–60 Hz) and a notch filter (50 Hz) to remove the irrelevant EEG spectral content and electric coupling interferences. Specifically, with regard to the band-pass filter, we implemented a 5th-order elliptic filter, with a passband ripple of 0.1 dB and a stopband attenuation of 40 dB, and performed zero-phase digital filtering. With these specifications, the frequency contents in the bands of interest (from 4 Hz for the theta rhythm to 42 Hz for the gamma rhythm) are not compromised. Then, 40 stimulus-locked epochs from 0 (stimulus onset) to 6 s (stimulus duration) were extracted from EEG recordings of each block, corresponding to 20 Image 1 trials and 20 Image 2 trials. Moreover, the 10 s preceding the onset of the first stimulus were extracted from Acq1 and defined as the participant’s baseline signal. At this stage, the different epochs were visually inspected, and no trials were deemed necessary to be discarded. Once concatenated the extracted signals along the time dimension (baseline, Acq1, Acq2, Rev1 and Rev2), bad channels were identified by computing the correlation coefficient between each electrode and the others. More precisely, for each EEG electrode we calculated the mean value of the four highest (absolute) correlations and marked as bad channels those electrodes whose mean value was < 0.4 (Bigdely-Shamlo et al. 2015; da Cruz et al. 2018). Then, the remaining good channels were re-referenced to the average of all electrodes, and the reference electrode (FCz) recovered.

Subsequently, to remove the artefactual components from EEG data, we performed the Independent Component Analysis (ICA) using the EEGLAB MATLAB toolbox (https://​sccn.​ucsd.​edu/​eeglab/​index.​php). Independent Components (ICs) containing artifacts were at first identified through an EEGLAB plugin named ‘IClabel’, which defines the probability of each extracted IC to be a brain-driven (‘Brain’) or a non-brain-driven activity (‘Muscle’, ‘Eye’, ‘Heart’, ‘Line Noise’, ‘Channel Noise’ and ‘Other’). After rejecting the ICs classified as ‘Brain’ with less than 5% probability, we visually inspected all the remaining components (scalp map, time and spectral activity) and further removed only those showing clear artifactual activity. Finally, artifact-cleaned EEG signals were used to retrieve the previously identified bad channels using the spherical interpolation, and the 64 EEG signals were again re-referenced to the average of all electrodes.

Cortical source activity was reconstructed starting from the 64 artifact-cleaned EEG signals. The estimation of intracortical current densities was performed using the method eLORETA (exact Low Resolution Electromagnetic Tomography, LORETA-KEY©® software package, version: v20200414), a functional imaging technique belonging to the family of linear inverse solutions for 3D EEG source distribution modeling. Precisely, the algorithm computes the weighted minimum norm solution, so that the particular weights used in this solution endow eLORETA with the property of exact localization of test point sources under ideal (noise free) conditions (Pascual-Marqui et al. 2011).

The software employs a template three-layers head model (MNI152 template) comprising the scalp, the outer skull surface, and the inner skull surface and registered to the Talairach human brain atlas. The solution space is restricted to the grey matter of the reference brain, divided into 6239 voxels at 5 mm spatial resolution. The software LORETA-KEY©® was exploited to compute the inversion matrix starting from the Talairach coordinates of the 64 electrodes, while all subsequent processing steps were implemented in MATLAB.

Since each cortical source is described by a three-dimensional current density vector, by right multiplying the inversion matrix by the 64 EEG signals, we can extract the three scalar components of the current density vector for the 6239 voxels and at each time instant. Then, as the choice of constrained dipole orientations was made, the 3D current densities were projected on the voxels’ normal versor obtaining one time series for voxel.

Finally, according to the atlas used by LORETA-KEY©® (76 ROIs, see Table 1) voxels were grouped in functionally significant Regions of Interest (ROIs), and the signal representing each ROI was obtained as the mean activity of the voxels belonging to the ROI. The relationship between the ROIs’ names, MNI coordinates, and Brodmann areas is provided in the supplementary information (SI_2).

The power spectral density (PSD) was evaluated on the 76 reconstructed cortical ROIs and during each trial using the Welch’s periodogram method (Hamming window of 2 s, 50% overlap, 10 s zero padding).

Power analysis in theta (4–8 Hz), alpha (8–14 Hz) and gamma (30–42 Hz) bands was then performed separately in each experimental block (Acq1, Acq2, Rev1 and Rev2) and for each stimulus (Image 1 and Image 2). In particular, for each participant, the following analyses were performed in each ROI. The power was computed in each frequency band and each trial, and normalized to the baseline condition. That is, for each participant the power in the alpha, theta and gamma bands, for each trial and each phase, was divided by the power of the respective band computed in the 10 s participant's baseline. This trial-by-trial power signal was then separated between Image 1 and Image 2 (20 trials per image and per block) and used for two computations. 1) For each frequency band and each block, a mean power was computed by averaging the trial-by-trial power over the 20 trials of each image, to obtain one power value for block and stimulus. 2) For each frequency band and each block, the trial-by-trial power was used to compute a moving average signal for each image, considering a window of 3 trials (sliding one trial at a time), in order to evaluate power temporal evolution for each stimulus over the four experimental blocks. Since the chosen window is odd, it is centered on the trial in the current position. When there are not enough trials to fill the window (that is, regarding the first and last trials of each block), the window size is automatically truncated and the average is taken over only the two trials that fill the window.

Importantly, the data resulting from mean power computation (computation 1) were used to identify, without any a priori assumption, the ROIs and rhythms implicated in fear conditioning, and thus deserving an in-depth inspection in this work. To this end, for each frequency band we followed two steps. Step 1: we selected all areas that exhibit a statistical significant difference (corrected) in power between Image 1 and Image 2 (i.e., between CS + and CS-) in at least one of the four experimental blocks, to detect a possible involvement either in acquisition or reversal. We anticipate here that all these corrected significances occur only in the acquisition phase. Step 2: since the focus in this work is on reversal, we further restricted our analysis to those ROIs (among those selected in step 1) that exhibit a statistically significant difference (although not corrected) between Image 1 and Image 2 both during acquisition and reversal. As shown in “Results”, this resulted in two regions for the theta rhythm, eleven regions for the alpha rhythm, and no region for the gamma rhythm.

To further investigate the possible neural mechanisms underlying fear acquisition and reversal, we evaluated the functional connectivity in each considered frequency band among the 76 reconstructed cortical ROIs by using the Spectral Granger Causality estimator, which provides weighted and directional metrics of the causal interactions between ROIs. The connectivity analysis was limited to the theta and alpha bands, since the power analysis in the gamma band provided inconclusive results (see “Results”). The Granger Causality is based on the autoregressive (AR) modeling framework and estimates the functional connectivity between ROIs by comparing the prediction ability of two AR models (of a certain order \(p\)) on the same process \({x}_{k,j}\). Specifically, given two time series \({x}_{k,i}[n]\) and \({x}_{k,j}[n]\), where n is the discrete time (\(n=0, 1,\dots ,N-1\)), representing the activity at two distinct cortical ROIs (\({ROI}_{i}\) and \({ROI}_{j}\)) for each participant \(k\) (\(k=1,\dots ,19\)). The Granger Causality estimator quantifies the causal interaction from \({ROI}_{i}\) to \({ROI}_{j}\) as the improvement in predictability of \({x}_{k,j}[n]\) when using a bivariate AR model, based on both past values of \({x}_{k,j}\) and past values of \({x}_{k,i}\), compared to a univariate AR model, based only on past values of \({x}_{k,j}\).

Frequency-domain Granger Causality can be formalized starting from the spectral derivation of the bivariate representation of the activity of the two ROIs, \({x}_{k,j}[n]\) and \({x}_{k,i}[n]\) via the Fourier Transformation. According to Geweke (1982, 1984) the power spectrum of a time series \({x}_{k,j}[n]\) can be decomposed into an ‘intrinsic’ and a ‘causal’ part, considering the latter predicted by the other time series \({x}_{k,i}[n]\).

The GC spectrum from \(i\) to \(j\) \({(GC}_{i\to j}(f))\) is defined as the logarithm of the ratio between the total power spectrum of \({x}_{k,j}[n]\) at frequency \(f\) and the difference between the total power spectrum and the ‘causal’ power predicted by \({x}_{k,i}[n]\) at the same frequency. Accordingly, at a given frequency f, the estimated quantity \({GC}_{i\to j}\left(f\right)\) is zero when the causal power of \({x}_{k,i}[n]\) onto \({x}_{k,j}[n]\) is zero and increases (> 0) as the causal power increases. For each participant \(k\), a GC spectrum was computed in each block and for each stimulus by linking 20 trials for each experimental block and stimulus. In all cases, the order \(p\) of the AR models was set equal to 30 on the basis of a previous analysis (Magosso et al. 2021; Tarasi et al. 2021) which showed that for \(p\ge 30\) the estimated values of GC do not change substantially. It is worth noting that spectral GC provides a connectivity matrix (\(GC\left(f\right):\) 76 × 75, discarding auto connectivity) for each frequency sample (\(n\,\, sample = 2501, frequency\,\, resolution=0.1 Hz\)). To obtain a single connectivity value representative of the rhythms under analysis, we computed the mean value of GC(f) in the given frequency band. Additionally, for each participant, the theta and alpha connectivity matrices were normalized so that the sum of all connections in each matrix is equal to 100. Then, from these normalized matrices (named complete connectivity matrices), sparse normalized connectivity matrices were obtained by performing a statistical analysis between the two types of stimuli (Image 1 and Image 2) independently for each experimental block, using the non-parametric permutation t-test (see details below in the “Statistical analysis”). Hence, for each block and for each frequency band, only connections significantly different (p value < 0.05) between Image 1 and Image 2 among all the 76 × 75 possible connections were retained in each subject connectivity matrix, while the others were set to zero.

Finally, both the complete and the resulting sparse connectivity matrices were averaged across participants to characterize each block and each stimulus (4 blocks × 2 stimuli) with a connectivity matrix.

To understand how power changes are transmitted from a generative node to others nodes in the network, we focused on the set of connections exiting from the areas identified in the previous frequency domain analysis, following the two-step method described in ‘Cortical Power computation’. The resulting functional network is represented through a graph where the involved cortical ROIs are the nodes and the connectivity values are displayed by weighted and directed arrows.

For each experimental block, a two-tailed permutation-based t-test for dependent samples between Image 1 and Image 2 was performed on SCR first, and then on the normalized mean power and connectivity data in each considered frequency band. The test was performed through a custom-made MATLAB script from the functions implemented in the FieldTrip toolbox (Oostenveld et al. 2010). It should be noted that the two-tailed test was used as we had no a priori hypothesis about how the power and connectivity were varying (increase/decrease) between the two stimuli. The distribution of the t-statistic for each cortical ROI under the null hypothesis was empirically realized by generating 5000 random permutations of the observed values between the 2 stimulus conditions (Monte Carlo method). The uncorrected p-value was the proportion of the permutation distribution greater than or at most equal to the observed t-statistic computed on the non-permuted values. Then, for cortical mean power analysis (i.e. average power over the 20 trials) a correction for multiple comparisons (76 comparisons, one per ROI) was achieved, separately for each block, using the false discovery rate correction (Benjamini–Hochberg procedure) (Benjamini and Hochberg 1995). Also, the SCR statistics largely survive correction (4 comparisons, one per block).

Finally, for scalp-level and moving average analyses, a cluster-based permutation t-test (Maris and Oostenveld 2007) was performed, to account for spatial dependence between electrodes and temporal dependence between trials, respectively.

First, a statistical analysis was performed between the SCR during the presentation of Image 1 and Image 2 in each block, to assess the correct acquisition and reversal of fear.

For each block (Acq1, Acq2, Rev1 and Rev2), we found a statistically significant difference between images in SCR values. In detail: p_Acq1 = \(3.9*{10}^{-4}\), p_Acq2 = \(3.9*{ 10}^{-4}\), p_Rev1 = \(6.0*{ 10}^{-3}\), and p_Rev2= \(3.9*{ 10}^{-4}\) (uncorrected). Moreover, in all blocks SCR was higher during CS + than CS- (i.e., during the presentation of Image 2 in Acq1 and Acq2, and Image 1 in Rev1 and Rev2). This analysis confirms the successful acquisition and reversal of fear for the subject group. A figure showing the group means SCR values for the two images, in the four blocks, is provided in the supplementary information (SI_3).

This work aims to investigate the role of brain rhythms at the cortical level. However, to ensure the reliability of our findings, strengthen their value, and enable comparisons with previous studies (e.g., Mueller et al. 2014; Sperl et al. 2019; Bierwirth et al. 2021), scalp-level power was also computed for a preliminary analysis. Differential scalp-level maps for the theta, alpha, and gamma bands normalized mean power are reported in Fig. 2. Theta band shows greater power during CS + (i.e., Image 2 in Acq1 and Acq2; Image 1 in Rev1 and Rev2) than CS- in fronto-medial channels, for all the four blocks; alpha band shows greater power during CS- (i.e., Image 1 in Acq1 and Acq2; Image 2 in Rev1 and Rev2) than CS + in left parietal channels, for all the four blocks; while gamma band power shows less consistency among all phases of the experiment. Cluster-based analysis shows statistically significant (corrected) differences especially in the acquisition phase.

Data resulting from the mean power analysis (computation 1, see “Cortical Power Computation”) were used to identify, without any a priori assumptions, the ROIs and rhythms most involved in fear acquisition and reversal. For these regions, we displayed the power mean value across the four blocks and the moving average of the trials. Finally, we further performed functional connectivity analysis.

The previous analysis revealed the presence of several regions that exhibit significant differences between Image 1 and Image 2 power. In this section we investigate, through Granger connectivity, how information is transferred from these regions toward other regions in the brain (i.e., how this increased or decreased power is transferred). Connectivity in the theta band is illustrated considering the connections that emerge from the two regions, CG l and CG m, since both display significant connectivity differences between Image 1 and Image 2. Regarding the alpha rhythm, only connectivity from two areas (CG l and PG l) is shown. In fact, connectivity was evaluated also from the remaining areas but statistical differences between Image 1 and Image 2 were unclear and did not provide any evident network topology.

In each of the following Figs. 9, 10, 11, and 12, the first column displays the connections that are stronger during the processing of Image 1 (Image1 > Image2: blue arrows), whereas the second column shows the connections that are stronger during the processing of Image 2 (Image2 > Image1: red arrows). In the first set of plots, we show the connections, which exhibit the largest differences (in absolute value) between Image 1 and Image 2 in each block, independently of the statistical difference (Figs. 9, 11), to show a general trend. Since all connectivity matrices are normalized to 100, and we have a total of 76 × 75 connections, the mean value of the connections is 0.0175. We chose to plot all connection differences that overcome 1/4 of the mean value for theta and 1/6 for alpha band. These different thresholds were chosen because the differences turned out higher in the theta than in the alpha range.

In the second set of plots, only connections that exhibit a significant statistical difference between Image 1 and Image 2 (p < 0.05, uncorrected) are displayed (Figs. 10, 12).

Here, a summary of the main results is provided. The scalp-level analysis (Fig. 2) shows a greater theta power during CS + than during CS-, especially in fronto-medial channels, and an increased alpha power during CS- on a broader area, especially involving the left-parietal electrodes. These results are stressed by the study performed at the cortical source level, showing a statistically significant (corrected) increase in theta power during CS + in the left and medial cingulate gyrus (Figs. 3 and 4) and a significant (corrected) increase in alpha power during CS- in the left motor and somatosensory areas (contralateral to the shock, Figs. 6 and 7). It is worth noting that these changes remain consistent during reversal (considering the new CS + and the new CS-), although significant only in the first block. Furthermore, concerning the theta response, reversal occurs just after one trial (Fig. 5), whereas three-four trials are necessary to observe a reversal in alpha power (Fig. 8).

Connectivity analysis, limited to the outgoing connections from the most significant regions revealed by the previous study, shows stronger theta connectivity emerging from the left cingulate gyrus during CS + (Figs. 9 and 10) and stronger connectivity arising from the left precentral gyrus during CS- (Figs. 11 and 12). Both are also evident during reversal. Interestingly, these connectivity differences between CS + and CS- are higher during the last reversal block, when differences in power density become elusive, suggesting a role for synaptic plasticity in reversal.

A first significant result concerns the role of the theta rhythm in fear acquisition. Our data show an increase in theta power in the left-mid cingulate cortex in response to the CS + stimuli. This role is particularly marked during the first block of acquisition (see also Table 2) but remains evident (although less pronounced) during the second acquisition block and the first reversal block. Noticeably, in the second reversal block, although CS + theta power is still higher than the CS- theta power, the difference becomes less substantial, and the role of theta rhythm progressively attenuates. The same pattern is confirmed by the trial-by-trial analysis using the moving average signal, which provides additional attractive cues. Indeed, at the beginning of the first acquisition block, theta power in the cingulate cortex increases abruptly both during CS + and CS-, probably signaling an alert phase of the experiment. Then, after just 1–2 trials, theta power differentiated between CS + and CS-. It is worth noting that this difference becomes maximally evident during the mid-period of the acquisition phase and then progressively declines toward the end of acquisition 2. The rapidity of the theta response is confirmed by looking at the reversal phase: just 1–2 shock-associated trials are sufficient to significantly increase theta power in response to the new aversive image.

These patterns, taken together, suggest that theta oscillations in the cingulate cortex signal the presence of a new aversive event, and this pattern is already evident during the first trials of the learning phase. However, these results also underline some differences between acquisition and reversal. Although the theta power difference between CS + and CS- develops promptly, it remains weaker during the reversal phase than in the previous acquisition phase, especially during the second reversal period. Overall, it seems that the theta rhythm signals the novelty of the aversive event and then declines when the association has been established, with this decline especially evident in the second reversal period, making power difference during reversal less manifest than during acquisition. This result agrees with a recent finding by (Taub et al. 2018): the authors suggest that the increase in theta power during aversive conditioning is correlated with the magnitude of conditioned responses but declines once the association is stabilized. Similarly, (Ridderbusch et al. 2021) observed a temporary increase in neural activation in the anterior cingulate cortex after re-exposure to the US after extinction training and suggested that this is associated with exploratory behavior, signaling changes in US-expectancy and arousal ratings.

Several studies underline the implication of the anterior (mid or dorsal) cingulate cortex in fear acquisition (Milad et al. 2007a; Toyoda et al. 2011; Mueller et al. 2014; Feng et al. 2014; Bierwirth et al. 2021). These results substantially agree with ours. Our study adopted the subdivision among ROIs illustrated in Table 1, according to the atlas used by LORETA-KEY©®. Using this atlas, we found significant theta power differences in the cingulate gyrus (left and medial). According to this atlas, the cingulate gyrus includes, among the others, the posterior portions of the Brodmann regions 24 and 32, which are traditionally ascribed to the ACC. However, it is worth noting that the Atlas also includes two other “cingulate” regions named “anterior cingulate” and “posterior cingulate” (see Table 1). In particular, the region called “anterior cingulate” includes the most anterior portions of areas 24 and 32. This subdivision agrees with the functional description of the cingulate gyrus proposed by (Vogt 2009). The author states that “the greatest number of “fear” activations occur in the anterior part of the midcingulate cortex MCC and not in ACC”, The first roughly corresponds to the posterior portions of the Brodmann areas 32 and 24, i.e. to the CG region used in the present atlas.

Many other results in rodents, primates, and humans underline the impact of theta oscillations in fear learning. Synchronization at theta frequencies is suggested to characterize activity in amygdala-hippocampal pathways associated with the consolidation of fear memory (Pape et al. 2005) and to represent a general mechanism of fear learning across species (Chen et al. 2021). A shared hypothesis is that the theta rhythm develops in the amygdala and hippocampus limbic system and is then transmitted to the ventromedial prefrontal cortex and to the anterior midcingulate cortex to synchronize ACC activity, and to transfer error signal information to support memory formation (Verbeke et al. 2021). Indeed, the anterior midcingulate cortex receives afferents from the amygdala (Vogt and Pandya 1987; Vogt 2005). Furthermore, synchronized frontomedial theta oscillations are a potential mechanism to support memory communication between brain regions (Sperl et al. 2019).

As to the last point (i.e., theta transmission and synchronization), a novel significant result in our study concerns the pattern of connections emerging from the previous two regions (CG l and CG m). We observed that Granger connectivity in the theta range is stronger during CS + than during CS-, and these differences are also evident during reversal. A possible interpretation is that the increased theta power in these regions is then transmitted to other areas of the brain, thus producing a generalized theta synchronization, subserving the retrieval of fear responses, or a general process of adaptive control of an unpleasant event (see also Shackman et al. 2011; Mueller et al. 2014)). Interestingly, if attention is focused only on connections statistically different between CS + and CS- (not only to the absolute differences), the increase in connectivity during CS + is especially evident during the first acquisition and second reversal blocks. Hence, a puzzling phenomenon is that theta power differences between CS + and CS- decline in reversal 2, whereas connectivity differences become more evident in the same block. Thus, functional connectivity does not simply reflect a change in power of the theta rhythm transmitted to other regions but may also depend on an effective alteration of synapses, especially in the last portion of the experiment. Further studies are needed to clarify this crucial point.

A significant observation emerging from our data is that alpha power changes are less localized than the changes in theta power and involve a more extensive network mainly located in portions of the posterior frontal cortex and parietal lobes, with a predominance in the left hemisphere. Some of these zones (the precentral gyrus, paracentral lobule and inferior parietal lobule) are implicated in motor and sensory innervation. As expected, alpha power is smaller during CS + than CS- in all these regions, reflecting a condition of greater arousal. However, it is worth noting that alpha power is higher than baseline during all phases of the experiment, probably since the initial period of the experiment (before any trial) was felt as the most stressful condition for the participants, possibly due to uncertainty of what will happen next. Alpha power differences are more evident in the acquisition phase than in the reversal phase, and, in the cingulate gyrus, alpha power exhibits a progressive increase during the experiment, suggesting increasing relaxation of the participants over the course of the experiment.

The pattern of alpha power changes during fear conditioning was investigated in detail by (Chien et al. 2017). The authors observed a significant alpha event-related desynchronization (ERD; i.e., a decrease in power) at parietal and occipital channels, hence over sensory structures related to (visual) CS processing. These changes were especially evident in the early phase of the stimulus train, reflecting a difference between the early and late stages of acquisition. By comparing their results with SCR data, the authors concluded that alpha power changes mainly reflect the valence and salience of the stimulus, i.e., the ability of CS to capture attention and motivate behavior. However, at odds with our results, the authors did not find significant differences in alpha power between CS + and CS-. Differences between our results and those by Chien et al. can be explained by thinking that these authors mainly focused on the magnitude of alpha ERD, which is maximal in the occipital regions, implicated in the visual processing of the external stimuli. Conversely, we focused on statistical differences between CS + and CS-, concentrated in parietal and posterior frontal regions, i.e., in the zones mainly involved in tactile and motor processing.

A significantly greater alpha-power suppression for threat-conditioned CS + than CS- was observed by (Panitz et al. 2019) at the scalp sites corresponding to visuocortical (i.e., occipital) areas. These differences likely indicate a sustained allocation of visual attention to the conditioned threat cue. Our results replicate and extend the results by (Panitz et al. 2019), demonstrating a similar effect at the cortical source level and on a broader set of regions (not only visuocortical but also somatosensory and motor).

Since the precentral gyrus is primarily involved in motor processing, a decreased alpha during CS + in this zone may reflect greater motor activation in preparation for an escape (for instance, preparation of movement of the right arm where the shock is delivered). Indeed, alpha ERD reflects the gradual release of inhibition associated with the emergence of a task-response. In contrast, an increase in alpha oscillations (event related synchronization, ERS) is observed with the CS-. Alpha ERS is commonly ascribed to idling or suppressing activity in task-irrelevant sites (Klimesch et al. 2007). Hence, our result supports the idea that alpha power changes observed in parietal and posterior frontal zones primarily reflect a preparative response to an action (during CS +) or a partial idling (during CS-) of the same activity.

Finally, it is worth noting that the time response of this alpha pattern is slower than that of the theta response: as evident looking at the moving average, alpha ERS during CS- requires 3–4 trials to develop. Another interesting aspect is that alpha power exhibits a drastic fall (ERD) at the beginning of any new block, reflecting greater attention/arousal due to the unfamiliar new conditions. Then alpha power progressively increases (especially in CS-), reducing the response in motor areas.

The connectivity pattern in the alpha band further underlines the pivotal role played by the left precentral gyrus. Stronger outflow connectivity is evident in this area during CS- than during CS + , reflecting the higher alpha power transmitted towards other occipital, parietal and frontal regions. It is worth noting that the reverse of this connectivity pattern is relatively slow, being maximally evident during reversal 2 than during reversal 1. This phenomenon is similar to what has already been observed for the theta connectivity from the left and medial cingulate cortex. In other terms, connectivity changes mature more slowly during reversal than during acquisition, becoming fully evident in the second reversal block. This pattern probably reflects synaptic changes necessary to overcome a previous pattern of connectivity developed during the acquisition phase. Indeed, as shown in recent modeling studies (Ursino et al. 2020; Ricci et al. 2021) functional connectivity mainly reflects the amount of information transmitted from one region to another: the latter can depend both on the power in the source region and on the strength of the effective connectivity linking the two regions.

Unexpectedly, the pattern of alpha-band connectivity emerging from the left cingulate cortex apparently contradicts the pattern of alpha power: in fact, many of these connections are higher during CS + than during CS-, i.e., in conditions of ERD. We do not have a definitive explanation for this pattern. However, we suspect that these seemingly anomalous connectivity patterns reflect non-linear phenomena and are strongly affected by changes in theta power (which, as demonstrated above, are significant in the left cingulate cortex and are higher in CS + than CS-). In previous papers (Ursino et al. 2020, 2021; Ricci et al. 2021) using a neural mass model as ground truth and comparing the actual connectivity values in the model with those obtained with methods for functional connectivity assessment, we demonstrated that non-linear phenomena play a significant role in connectivity estimation, resulting in possible interference between frequency bands and alterations in the connectivity values.

There is a consensus in the literature that gamma power is implicated in inhibiting a previously acquired fear response (Stujenske et al. 2014; Courtin et al. 2014; Fenton et al. 2016). (Mueller et al. 2014) observed that, in humans, vmPFC gamma activity differs between extinguished CS + and CS-. The role of the vmPFC in extinction is further supported by neuroimaging studies (Milad et al. 2007b). However, (Schiller et al. 2008) pointed out that reversal is a more complex process than extinction. Using fMRI, these authors observed that, during reversal, the activity in the vmPFC signals the presence of a safe stimulus (hence the new CS- in reversal, previously CS + during acquisition), which can be interpreted as an unexpected reward.

According to the studies mentioned above, a significant gamma activity in the vmPFC was expected in reversal; however, in our research, we were unable to find any corrected statistical difference in gamma between CS + and CS- during any phase of the experiment. For this reason, gamma activity was not further analyzed.