Auditory aversive generalization learning prompts threat-specific changes in alpha-band activity

Abstract

Pairing a neutral stimulus with aversive outcomes prompts neurophysiological and autonomic changes in response to the conditioned stimulus (CS+), compared to cues that signal safety (CS−). One of these changes—selective amplitude reduction of parietal alpha-band oscillations—has been reliably linked to processing of visual CS+. It is, however, unclear to what extent auditory conditioned cues prompt similar changes, how these changes evolve as learning progresses, and how alpha reduction in the auditory domain generalizes to similar stimuli. To address these questions, 55 participants listened to three sine wave tones, with either the highest or lowest pitch (CS+) being associated with a noxious white noise burst. A threat-specific (CS+) reduction in occipital–parietal alpha-band power was observed similar to changes expected for visual stimuli. No evidence for aversive generalization to the tone most similar to the CS+ was observed in terms of alpha-band power changes, aversiveness ratings, or pupil dilation. By-trial analyses found that selective alpha-band changes continued to increase as aversive conditioning continued, beyond when participants reported awareness of the contingencies. The results support a theoretical model in which selective alpha power represents a cross-modal index of continuous aversive learning, accompanied by sustained sensory discrimination of conditioned threat from safety cues.

Introduction

It has been well established that cues predicting aversive outcomes capture and hold access to limited capacity systems in the human brain, including systems mediating sensation, perception, and attention (Li and Keil 2023). The neurophysiological processes involved in acquiring this privileged access are less understood, including how cues in different sensory modalities acquire sensitivity to conditioned threat. A growing body of work has consistently shown that pairing a visual cue with an aversive outcome over time prompts a greater power reduction in posterior alpha-band (8 to 13 Hz) oscillations compared to unpaired cues (Panitz et al. 2019; Friedl and Keil 2020). Additionally, unpaired visual cues that are perceptually similar to the conditioned stimulus also prompt alpha-band power reduction in the same parieto-occipital areas, but to a lesser degree, creating a Gaussian pattern of effects (Yin et al. 2020; Friedl and Keil 2021). These changes have been taken as evidence for widespread cortical disinhibition, consistent with concepts such as arousal or attention (Bacigalupo and Luck 2022; Li and Keil 2023). In line with this notion, human and animal work suggests that alpha power changes reflect changes in the activity of widespread cortical and thalamocortical networks involved in prioritizing stimulus processing after initial sensory evaluation (Bazanova and Vernon 2014). Evidence from monkey local field potential studies has shown that stimulus-induced reduction in alpha band power is accompanied by reduction in neural communication within thalamocortical networks, for example, between the lateral geniculate nucleus and primary visual cortex (Bollimunta et al. 2011).

Despite abundant evidence in the visual domain, it is unclear how alpha-band power is altered during auditory aversive conditioning, in which neutral tones are paired with an aversive stimulus. Tones that are paired with aversive white noise bursts modulate the late positive potential (LPP) component of the event-related potential (ERP), similar to visual cue paradigms (Pavlov and Kotchoubey 2019). The LPP, in turn, tends to mirror alpha power changes during emotional scene perception (De Cesarei and Codispoti 2011; Ferrari et al. 2020). In line with the notion that alpha power reduction may be a cross-modal index of aversive learning, Hartmann et al. (2012) found lowered alpha power in auditory cortex when participants were instructed that certain tone cues would be followed by a noxious noise. However, the extent and temporal dynamics of alpha power changes during associative learning are still to be established. Specifically, it remains unclear if auditory aversive conditioning results in selective, stimulus-specific alpha-band power reduction and, if so, which cortical areas contribute to these effects. Research in the auditory domain specifically suggests there are more subtle alpha oscillations that emanate from the auditory cortices, with alpha power modulations evident in tasks that challenge language, tinnitus, and auditory anticipation (Weisz et al. 2011). Lastly, it is unclear if auditory conditioning prompts generalization across similar tone cues—a hallmark of visual associative learning.

To address these questions, alpha-band power in human dense-array electroencephalography (EEG) was examined in a differential auditory conditioning paradigm in which the highest or lowest pitch of three tones were paired with a noxious white noise burst (Fig. 1). To establish the extent to which learning took place, self-reports of valence and arousal as well as pupil dilation were collected in addition to EEG alpha power. In differential Pavlovian conditioning, an initially neutral stimulus (the CS+) such as a soft tone or visual pattern is repeatedly paired with an unconditioned stimulus (US) such as an electric shock or noxious noise. Other stimuli (the CS−) are never paired with the US. When multiple CS− are used, it is possible to examine response tuning along a gradient that is defined by perceptual similarity with the CS+. In this type of experiment, the CS− are then referred to as generalization stimuli (GS). In the present study, three tones were initially presented to participants without any aversive US pairing (the habituation phase). Then, either the lowest or highest pitch was paired with a noxious US white noise (the acquisition phase). Self-reports for each cue were collected at early and late periods of both the habituation and acquisition phases. Additionally, pupil dilation was recorded simultaneously with the EEG. Pupil dilation is well-established index of sympathetic activity in response to cues across many different modalities including sound (Partala and Surakka 2003). Including this measure was intended to provide an additional emotion-related measure for each cue, to be used as a manipulation check and for comparison with alpha power changes and participant ratings.

To focus and simplify the analyses, the results were interpreted through two a priori models representing the most likely patterns of cue reactivity. As previously mentioned, occipital–parietal alpha power changes are known to generalize to cues that are visually similar to CS+ (Yin et al. 2020; Friedl and Keil 2021). Accordingly, it was hypothesized that threat-related changes in parietal alpha oscillations would generalize across a gradient of auditory stimuli varying in pitch. Another possibility is an all-or-nothing pattern in which there is only reactivity to the CS+. To represent these possible reactivity patterns, the models were specified by assigning a single weight-value for each cue representing the expected increase or decrease in reactivity. These weights were then used in planned comparison analyses, within the framework of the general linear model, also known as F-contrasts (Rosenthal and Rosnow 1985). Both models predict an increase for the CS+ with a positive weight. The models differ in their expectations for the next closest tone pitch (GS1) and the least similar tone (GS2). The all-or-nothing model predicts no change for GS1 and GS2 (weights: 2, −1, −1), whereas the generalization model predicts an increase for GS1 and a reduction for GS2 (weights: 1, 0.75, and − 1.75). The fits between the data and each of the two a priori models can be thought of as competing alternative hypotheses of all-or-nothing versus generalization patterns of reactivity.

Methods

The data presented in this report are part of a larger study design and set of hypotheses that were preregistered prior to data collection (https://osf.io/e26ad). The current study concerns findings regarding alpha-band power changes over an auditory stimulus generalization gradient across participants, focusing on within-subject effects.

Participants

Sixty-four participants were recruited using flyers, existing databases, and online advertisements. The participants received class credit or were paid $20 USD per hour for completion of the study. All procedures were approved by the University of Florida institutional review board, and participants gave informed consent prior to participating in accordance with the Declaration of Helsinki. All participants were 18 years of age or older, had normal or corrected-to-normal vision, and reported no personal or family history of seizures. Nine participants were excluded from this analysis due to withdrawal (n = 5), or having over 50% of EEG trial data contaminated by artifacts (n = 4). This resulted in a total of 55 remaining participants (36 female; M_age = 20.34, SE_age = 0.35) used for data analyses of ratings and alpha-band power.

For the secondary analyses of pupil dilation, 20 additional participants were removed due to unusable pupil recordings leaving 35 participants for the pupil dilation analyses (24 female; M_age = 20.21, SE_age = 0.49). Pupil data were considered unusable when over 50% of trials did not meet trial inclusion criteria described more thoroughly in the Pupillometry section. Trials were primarily lost because of excessive blinks or other oculomotor processes or when participants wore hard contact lenses. These artifacts are more readily corrected in EEG data compared to pupil data, where they result in signal loss.

Participants completed several other measures related to defensive engagement in aversive learning that are not reported here as the present focus is on overall reactivity not influenced by individual differences. These measures included questionnaires related to misophonia, fear, anxiety, and depression symptoms. Individuals reporting elevated misophonia symptoms (i.e. scores above 19 on the Misophonia Symptom Scale; Wu et al. 2014) were not recruited into this sample. Loudness discomfort levels were also acquired as an index of hyperacusis. No participants in the primary sample reported elevated loudness discomfort ratings (M = 23.90, SE = 2.22, on a scale from 5 to 50); the same was true for the sub-sample which was used for pupil analyses (M = 19.30, SE = 2.50).

Materials and procedures

Ratings of affect using the self-assessment manikin

Ratings for arousal and valence were collected using the Self-Assessment Manikin (SAM; Bradley and Lang 1994) during the 10th (early) and 90th (late) trials of the habituation and acquisition phases of the experiment. After each tone at this specific point of the experiment, participants self-reported their experienced emotional arousal and hedonic valence. Each dimension was presented with a set of five manikins ranging from completely calm/pleasant (i.e. low arousal or low valence scores) to aroused/unpleasant (i.e. high arousal or high valence scores). While viewing the manikins, participants reported how the previous tone made them feel by moving a monitor cursor with a mouse and clicking to make their rating. Responses were recorded as the pixel location along the horizontal axis of the SAM which ranged from 1 to 1,920 pixels.

Auditory stimuli

The auditory stimuli were three sinewave tones (sampled at 22,000 Hz) that were four seconds in duration. A cosine square window (20 points ramp-on and ramp-off) was applied at the beginning and end of the auditory presentation to minimize onset and offset sound spikes. Each sinewave tone had a specific frequency (i.e. 320, 541, or 914 Hz) based on an exponential function, creating three pitch conditions (i.e. CS+, GS1, and GS2). The number of tones and the pitch frequencies were chosen based on pilot studies, such that the majority of participants could discriminate the tones. Additionally, tones were multiplied by a 41.2 Hz cosine envelope to evoke auditory steady-state potentials, used in a separate set of hypotheses and analyses not reported here (see preregistration for additional information). Each tone was normalized based on their respective frequency’s amplitude. Finally, the loudness levels of each tone were reduced to 20% of their amplitude, resulting in loudness levels of 70 dBA for each tone stimulus, assessed with an audiometer.

During the acquisition phase, a 1-s 91 dBA white noise was played during CS+ trials, serving as the US, and accompanying the CS+ during the final second of the four second tone duration. The US duration and onset were selected based on previous research suggesting larger conditioning effects for durations over 500 ms (Sperl et al. 2016), and when a CS+ and US co-terminate following prolonged overlapping intervals (Kamin 1956). Levels were controlled with an audiometer between sessions. The white noise also included a short cosine-square ramp (5 points) at the beginning and end of this auditory presentation. The white noise was paired with either the 320 Hz or 914 Hz pitch, counterbalanced across participants. Thus, the paired pitch served as CS+ (100% reinforcement rate), with the other frequencies considered as GS1 or GS2, depending on their distance in frequency from the CS+ pitch. All auditory stimuli were presented through two Behringer Studio 50 speakers arranged symmetrically behind each participant at ear level, approximately 30 cm in distance from the participant.

Auditory aversive generalization task

The auditory aversive generalization task featured 240 total trials split into two 120 trial phases, habituation (Fig. 1A) and acquisition (Fig. 1B). A brief extinction phase followed at the end of the experiment, in which no EEG was recorded. This choice of experimental design reflected our primary interest in the acquisition of conditioned responses. Future work will focus on extinction-related processes, using targeted experimental designs toward that end. For each trial, one of the three tones (i.e. CS+, GS1, GS2) was played for 4 s while participants viewed a white fixation dot on a monitor occupying 0.5° of visual angle. The white noise US was only presented during the acquisition phase and was paired with the CS+ pitch (i.e. 320 or 914 Hz) at 100% reinforcement rate. The SAM ratings of arousal and valence for each pitch were acquired during trial 10 and 90 in the habituation and acquisition phases allowing for early and late assessments in each phase. During the acquisition phase, the first and third trials were CS+ trials to facilitate learning. All other trials were presented in a pseudorandomized order, such that no more than two CS+ trials would occur in sequence. All trials were separated by an inter-trial interval (ITI) randomly varying (uniform distribution) from 1.85 to 3.5 s in duration. All visual stimuli were presented using Psychtoolbox code (Brainard 1997) on a Cambridge Research Systems Display ++ monitor (1,920 × 1,080 pixels, 120 Hz refresh rate), positioned approximately 120 cm in front of the participant.

Following EEG data collection, a sequence of questions was used to assess how many participants differentiated the tones and if they learned the relationship between the CS+ and the US. Questions included how many distinct tones participants heard, an estimate of how many times they heard the US, and if there was a relationship between any of the sounds in the study as well as what that relationship was. If all questions were answered correctly, this was noted and the participant was classified as displaying evidence of contingency awareness.

Data acquisition and processing

E‌EG

EEG data were recorded at a 500 Hz sampling rate using an Electrical Geodesics (EGI) high input impedance system with a 128-channel (Ag-AgCl electrodes) HydroCel net, with impedances being kept below 60 kΩ. Online data were referenced to the vertex electrode (Cz), with a Butterworth filter applied during continuous recording. Following acquisition, EEG data were re-referenced to the grand average reference (i.e. averaged across all electrodes), and re-filtered with Butterworth low-pass (10th order, 3 dB point at 30 Hz) and high-pass (3rd order, 3 dB point at 1 Hz) filters. Next, continuous EEG data were separated into 3.6 s (1,801 sample points) trial epochs, spanning 600 ms (300 sample points) prior to tone onset and 3,000 ms (1,500 sample points) post-onset. The last 1,000 ms of each tone cue were not analyzed because the US played at that time for CS+ trials in the acquisition phase, rendering the EEG data during this period prone to motion and oculomotor artifact.

Segmented trials were then evaluated for artifact contaminated trials using the Statistical Correction of Artifacts in Dense Array Studies (SCADS) procedure (Junghöfer et al. 2000). Eye artifacts were corrected with a regression-based electrooculography (EOG) correction method (Schlögl et al. 2007, 2009) using electrodes above and below the eyes as well as electrode located at the outer canthi. Participants with a substantial number of trials lost due to artifacts (i.e. >50% of all trials rejected) were excluded from further data analyses in both our primary (i.e. full behavioral and alpha-band power sample of 55) and secondary (i.e. reduced pupil sample of 35) data sets. The final sample had an average of 29.3 trials per pitch condition and phase (habituation, acquisition) retained. Importantly, a similar number of trials per condition within each phase were retained both our full behavioral and alpha-band power (habituation: CS+ = 31; GS1 = 32; GS2 = 32; acquisition: CS+ = 27; GS1 = 27; GS2 = 27).

A current source density (CSD) spatial transformation was applied to artifact-free trials, to increase spatial specificity of the signal. The method estimates the current flow at the cortical surface via the second spatial derivative (i.e. Laplacian) of the scalp topography. This was carried out with the equation proposed by Junghöfer et al. (1997), which is suitable for dense-array EEG montages. The recommend smoothing constant (λ) of 0.2 was applied. This CSD transformation heightens the interpretability of the topographical representation (Junghöfer et al. 1997; Kayser and Tenke 2015) and thus assists in addressing questions regarding the cortical origin of conditioning-related power modulations in the alpha-band. This was thought to be a necessary transformation of the data because it has been suggested that there are spatially distinct alpha-power changes associated with auditory processing (Weisz et al. 2011). In addition to the improved spatial resolution, a consequence of this transformation is that the CSD representation of the EEG signal is reference-free.

Following the CSD spatial transformation, single-trial data from each participant, condition, and sensor were transformed into the time–frequency domain through convolution with a family of complex Morlet wavelets. These wavelets had center frequencies (f) between 2.50 and 27.48 Hz, spaced in even intervals of 0.278 Hz. Although the focus of this study was on alpha-band oscillations, frequencies in the entire range between 2.5 and 27.48 Hz were examined as recommended by current guidelines (Keil et al. 2022), to establish specificity and robustness of differences observed. A Morlet constant (m) of 10 was selected based to optimize the trade-off between temporal (sigma_t) and frequency, i.e. sigma_f = 1/(2 × π × sigma_t) smearing in the alpha range, given the ratio m = f/sigma_f, for every frequency in the analysis, f. At the center frequency of interest (10.27 Hz, the center of the canonical alpha band in young adults), this resulted in temporal and frequency smoothing values of 155 ms and 1.03 Hz, respectively. The absolute value from the convolution between the EEG data and the complex wavelets served as an estimate of time-varying power (Tallon-Baudry and Bertrand 1999). Percent change in baseline alpha-band power for all EEG trials was calculated. The decision for this type of baseline adjustment was based on the presence of alpha-band activity in the baseline period, and because multiplicative changes in the alpha frequency range have been linked to meaningful differences in task performance (Foxe and Snyder 2011; Jensen and Mazaheri 2010). Specifically, the baseline period used for this correction was −422 to −202 ms prior to the onset of the stimulus, with this interval being selected to account for potential edge artifacts and the temporal smoothing value obtained from the wavelet transformation.

In addition to the trial-averaged time–frequency spectra, the trial-by-trial change in alpha power reduction was also examined in a post hoc analysis, to assess how changes in oscillatory activity evolve as learning progresses. To do this, the baseline-corrected single-trial power at 10.27 Hz was calculated per participant at the sensor and time period of maximum alpha reduction. This included six occipital–parietal sensors (including Pz and CPz) from 600 to 900 ms. This window was chosen based on visual inspection of the grand mean time course of alpha-band changes. To increase the signal to noise ratio, the resulting values (percent change in alpha power from baseline) were binned into 28 sequential groups of trials, 14 in habituation and 14 in acquisition, separately for each stimulus. Each participant could contribute a maximum of 8 trials per bin.

Pupillometry

Eye-tracking data were recorded using an EyeLink 1000 Plus eye tracker system with a 16 mm lens placed in front of the monitor, but outside of visible range for the participants’ lower visual field. The data were recorded at a 500 Hz sampling rate. Pupil diameter was quantified by fitting an ellipse to each participant’s pupil mass threshold. The system’s infrared signal illumination level was initially set to 100%, and individual adjustments were made for each participant based on their respective pupil and corneal reflection. Eye-tracking data were calibrated and validated using a nine-point grid. During this process, a white circle (1° visual angle) was presented at each of these nine points against a black background. As the participants’ eyes fixated on each location presentation, the lens and pupil thresholds were adjusted if pupil diameter was lost.

Similar to EEG data, pupil diameter data were segmented into 3.6 s (1,801 sample points) epochs, with 600 ms (300 sample points) before the tone and 3,000 ms (1,500 sample points) after tone onset. Next, single trials underwent artifact correction and detection procedures. First, a Butterworth low-pass filter (sixth order, 3 dB point at 7.5 Hz) was applied to the data to attenuate large artifactual spikes. Then, two types of artifacts were addressed: (i) time segments with missing values (e.g. due to blinks) and (ii) rapid pupil diameter changes exceeding 2.5 units change in pupil diameter from a previous sample point. These time segments were marked as artifacts and subsequently corrected using piece-wise cubic interpolation. Participants with inconsistent pupil tracking (defined as more than 50% interpolated data) were excluded from this analysis. As with the EEG data, the percentage of trial data interpolated did not significantly differ between pitch conditions in the habituation (CS+ = 12.2%; GS1 = 14.1%; GS2 = 12.6%, F = 1.11, n.s.) or acquisition (CS+ = 17.1; GS1 = 15.9; GS2 = 15.9, F = 1.40, n.s.) phases in this reduced pupil data set. Finally, pupil data were baseline-corrected as the percentage change from −500 to 0 ms, relative to tone onset. Grand average pupil waveforms were computed across all pitch conditions in habituation and acquisition phases.

Statistical analyses

Overview

Affect ratings and physiological results were interpreted with the a priori models of predicted outcomes: all-or-nothing and generalization. Each model had different weights reflecting the expected pattern of effects for the CS+, GS1, and GS2 trials. The all-or-nothing model predicts a larger response only for the CS+ with no response for the GS1 or GS2 conditions; this pattern of effects is reflected in the weights of 2, −1, and −1, respectively. The generalization model predicts the GS1 to evoke a similarly large response as the CS+ stimuli with weights of 1, 0.75, and − 1.75. These weights were used to transform the valence and arousal ratings per person into a single value representing the model strength. For alpha-band power and pupil dilation, the model weights were used in the F-contrast permutation procedure. Both of these model fit analyses are described in the following sections.

Model fit analyses: affective ratings

The ratings from each phase of the study were transformed into model fit scores by taking the inner product of the results by each of the three models’ weights. Thus, each participant had a model fit score for the all-or-nothing and generalization models for both arousal and valence ratings per the early and late stages of the habituation and acquisition phases. All model fit scores were tested for statistical significance via single-sample t-tests against a null hypothesis of zero. The critical P-value was set at Bonferroni corrected threshold of 0.003 found by taking the typical threshold of 0.05 divided by the total number of the 16 significance tests performed. The standardized mean difference (Cohen’s D) measure of effect size was reported with each t-test. For this measure, a value of 0.01 is considered very small, 0.2 is small, 0.5 is medium, 0.8 is large, 1.2 is very large, and 2.0 is thought of as huge (Sawilowsky 2009; Cohen 2013).

Model fit analyses: EEG alpha and pupil diameter

Both alpha-band power and pupil dilation were examined with general linear models using the model weights applied to the acquisition phase of the study. To understand at which points model fits were statistically meaningful, permutation-controlled mass univariate tests were performed at each temporal and spatial data point (Blair and Karniski 1993): F-contrasts for the two models were computed for each model fit producing a spatiotemporal map of each model’s F-values. This process was then repeated 5,000 times with each of the three conditions (i.e. CS+, GS1, and GS2) randomly permuted per participant. The maximum F-value for each electrode-by-time-by-frequency point permutation was used to create a nonparametric null distribution in which the 95th quantile served as the critical F-value (Blair and Karniski 1993; McTeague et al. 2015). For trial-averaged alpha-band power, this resulted in an F_crit of 11.18. For trial-by-trial binned alpha changes across the experimental session, the F_crit was 10.8. For pupil dilation, the maximum F-value per time point was used to form the null distribution with the same 95th quantile, resulting in a critical F-value (F_crit) of 9.3.

Results

Post-experiment, the majority of participants correctly reported the relationship between the CS+ and the US. Although only 15 out of 55 participants (27%) correctly reported that there were three distinct tones, 52 (94.5%) reported there was a relationship between the tones and the noise. Furthermore, 43 (78%) correctly stated that either the lowest (21 out of 29; 72.4%) or highest pitch (22 out of 26; 84.6%) was associated with the US. Although 12 participants did not correctly state that the lowest or highest pitch was associated with the US, post hoc analyses suggested they did not differ in their ratings or physiological reactivity compared to the rest of the sample (Supplemental Figs. 1 and 2). They consistently rated the CS+ as more arousing and showed similar physiological reactivity to the CS+ as well. Additional supplemental materials show the correlations between the CS+ versus the GS1 (Supplemental Fig. 3), and the time–frequency plots for each cue during the acquisition phase (Supplemental Fig. 4).

Arousal and valence ratings

The results for arousal and valence ratings can be seen in Fig. 2. On a scale of 1 being completely unarousing and 1,920 being completely arousing, the mean arousal rating in the early habituation phase for the CS+ stimulus was M = 735, SE = 30.2; GS1 was M = 672, SE = 26.3; and GS2 was M = 733, SE = 29.6. For the late habituation phase, CS+ was M = 687, SE = 34.3; GS1 was M = 704, SE = 29.5; and GS2 was M = 755, SE = 31.2. During the early acquisition phase, the CS+ was M = 1210, SE = 38.2; GS1 was M = 746, SE = 29.4; and GS2 was M = 750, SE = 31.5. In the late acquisition phase, the CS+ arousal rating was M = 1117, SE = 40.9; GS1 was M = 751, SE = 32.4; and GS2 was M = 737, SE = 31.5.

Valence ratings were on a similar scale as arousal self-reports, in which 1 was completely unpleasant and 1,920 is completely pleasant. In the early habituation phase, the CS+ stimuli were rated M = 913, SE = 23.5; GS1 was M = 923, SE = 25.1; and GS2 was M = 924, SE = 26.6. In the late habituation phase, CS+ was M = 919, SE = 36.1; GS1 was M = 975, SE = 32.5; and GS2 was M = 960, SE = 34.4. For the early acquisition phase, the CS+ was M = 1,302, SE = 29.9; GS1 was M = 938, SE = 28.1; and GS2 was M = 931, SE = 28.6. In the late acquisition phase, the CS+ arousal rating was M = 1,224, SE = 31.7; GS1 was M = 952, SE = 29.1; and GS2 was M = 922, SE = 34.6.

Arousal and valence model fit scores

As described in the Methods section, arousal and valence ratings were analyzed by means of general linear models with weights defined by the two a priori models for each phase of the study. The resulting fit scores were tested for significance via single-sample t-tests and a Bonferroni correct critical P-value of 0.003. The results of the t-tests can be seen in Table 1. The all-or-nothing fit scores for arousal ratings were not significant in the early habituation (M = 66.5, SE = 42.6) or late habituation phases (M = −83.6, SE = 57.0). The arousal model fit scores were also not statistically significant for the generalization model in both habituation phases (early ratings M = −42.6, SE = 39.7; late ratings M = −106, SE = 49.1). In the acquisition phases all model fit scores were statistically different with the largest effect sizes being for the all-or-nothing model. In the early and late acquisition phases respectively, the all-or-nothing fit scores were M = 923, SE = 79.9 and M = 746, SE = 68.8, and the generalization scores were M = 457, SE = 50.5 and M = 390, SE = 48.0.

For the valence ratings, model fit scores showed a similar pattern as the arousal ratings (Table 1). In the early and late habituation phases, none of the model fit scores were statistically meaningful. In the acquisition phases, all model fit scores were statistically significant with the all-or-nothing transformed scores again featuring the largest effect sizes. In the early and late habituation phases respectively, the model fit scores for the all-or-nothing model were M = −20.6, SE = 42.8 and M = −97.1, SE = 47.4; for generalization scores M = −11.70, SE = 39.1 and M = −30.3, SE = 46.7. In the early and late acquisition phases, the all-or-nothing valence scores were M = 735, SE = 59.0 and M = 575, SE = 60.6; for generalization scores M = 376, SE = 37.9 and M = 325, SE = 46.4.

Alpha-band power

For the trial-averaged results, alpha-band power was found to be reduced during the presentation of each tone cue in an overall quantification of alpha from all sensors, cues, and phases of the study (Fig. 3). It was primarily found over occipital–parietal areas even when baseline-corrected (Fig. 4). The F-contrast permutation procedure only found a significant alpha-band power reduction for the all-or-nothing model during the acquisition phase. The F-values corresponding to that model exceeded the critical threshold at parieto-occipital electrode sites encompassing CPz and Pz from 670 to 820 ms (Fig. 5). This selective reduction in response to the conditioned threat cue was present in 40 out of the 55 participants (Fig. 6). The generalization model never crossed the significance threshold. To ensure that the failure to observe evidence in favor of this model was not due to the specific weights, we repeated the analyses using different coefficients predicting less generalization for GS1, i.e. 1.5, 0.5, −2 and 4, 1, and −5. None of these weights resulted in support for the generalization model.

When examining the trial-by-trial evolution of this selective alpha reduction across the experimental session (Fig. 7), a significant F-contrast for the all-or-nothing model was observed in trial bins 21 to 27, corresponding to the time between the 30th and 96th trial of the acquisition phase (out of 120 total trials). Thus, selective alpha reduction in response to the conditioned threat cue appeared around mid-acquisition and lasted into the late phase of acquisition. Near the end of acquisition, the 27th and 28th trial bin no longer supported the all-or-nothing model and instead, the data showed a generalization pattern for the 27th bin.

Pupil dilation

The average pupil waveforms by cue type are shown in Fig. 8 for the habituation and acquisition phases. Consistent with the pattern visible in the grand mean waveforms, the permutation procedure showed that the only fit with the all-or-nothing model surpassed the critical threshold at approximately 2,500 ms (Fig. 9). Paralleling alpha band power, the generalization model did not surpass the critical threshold.

Discussion

In the present study, a differential aversive conditioning task was used to examine the hypothesis that selective alpha power reduction represents an index of ongoing cortical engagement with a conditioned auditory threat cue. Additionally, the study examined the extent to which alpha power reduction in response to the CS+ generalizes to the most perceptually similar cue that was never paired to an aversive outcome (the GS1). The present study found strong support for the hypothesis that the selective reduction of parietal alpha oscillations reflects the engagement with conditioned auditory cues. The timing and topography of the alpha-band power reduction in the present study strongly resembled the occipital–parietal power changes observed with visual cues (Yin et al. 2020; Friedl and Keil 2021; Bacigalupo and Luck 2022), suggesting that suppression of alpha-band oscillations in parietal cortices represents a supra-modal index of aversive conditioning. Unlike previous studies in the visual domain (Friedl and Keil 2021), the present study did not show strong evidence of generalization effects. Among all variables, only trial-wise alpha power changes late in acquisition evinced generalization. This effect was driven by stronger alpha power enhancement for the GS2 stimulus, compared to GS1, consistent with the overall trend of increased baseline alpha power during the experimental session, shown in Fig. 7, left panel. This trend was accompanied by steadily increasing alpha power reduction in response to the CS+, throughout the acquisition session.

Against expectations, none of the other variables examined here—self-reported aversiveness and arousal, as well as pupil dilation—supported generalization. Accordingly, we found consistent support only for the all-or-nothing model, when applied to ratings, pupil, and averaged parietal alpha power data. Future work will use smaller differences in pitch between CS+ and the GS, to examine the extent to which this absence of generalization reflects that the tones were too different. Another limitation to the present analyses were the fixed weights that were used to model generalization. While post hoc tests with different coefficients also did not lead to an overall generalization effect, future efforts will try to implement a more flexible means for quantifying the shape of generalization gradients.

At parietal sensor locations, 40 out of the 55 participants displayed a selectively larger reduction in alpha power for the CS+, compared to the GS cues, showing that alpha reduction is a robust correlate of auditory conditioning. This effect for the CS+ emerged around mid-acquisition and lasted until the final trials of the experiment. This finding has several conceptual implications: Participants reported heightened displeasure and US expectancy for the CS+ after just the 3rd CS+ trial out of the 10th total trial in the acquisition phase, well before the emergence of a statistically significant selective alpha power reduction, which occurred after the 15th CS+ trial. Thus, the present finding supports the notion that alpha reduction in response to conditioned threat is related to the formation and/or active maintenance of aversive memory associations between the CS+ and the US (Yin et al. 2020). This is at odds with interpretations of selective alpha power reduction as reflective of attention to the CS+. Processes related to focusing attention on conditioned cues are expected to be largest during the emergence of contingency awareness early on in the experimental session (Bacigalupo and Luck 2022), and decrease in magnitude during subsequent trials, perhaps reflecting processes such as adaptation and repetition suppression (Yin et al. 2020). Post hoc analysis of the GS stimuli in the present study shows that the difference between the safest (GS2) and the threat cue (CS+) remain robust until the end of the experimental session. In contrast, the more similar GS1 increasingly prompted alpha reduction as learning progress, consistent with ongoing generalization learning. Recent models of aversive conditioning, rooted in animal and human work, emphasize the steadily changing neural dynamics as associative learning progresses (Li and Keil 2023).

Consistent with findings using fMRI (Yin et al. 2020) and event-related potentials (Thigpen 2017), the present trial-by-trial analyses suggest that initial discrimination between threat and safety cues involves higher-order cortices involved in parietal alpha power changes. The observation that the discrimination between threat and safety cues is attenuated rather than increased later in the acquisition phase supports the notion that sustained aversive conditioning prompts discrimination at increasingly lower levels of the cortical hierarchy, eventually leading to threat representations in sensory areas (Li and Wilson 2023).

Compared to previous research, the present results suggest that specific experimental conditions may facilitate auditory versus parietal alpha-band power changes. Hartmann et al. (2012) found an auditory cortex alpha reduction when participants thought they heard a “high”-pitched tone that they were told predicted a noxious noise punishment; however, there was actually only one tone pitch presented to participants with pseudorandom feedback. In the present study, because the tones were mostly differentiable, perhaps the CS+ is quickly categorized and resources are diverted to processes related to the anticipation of the aversive US. In a preliminary report that included some of the current participants, primarily parietal alpha power changes were also found (Ward et al. 2022). However, that study also included people with misophonia, a disorder characterized by a heightened sensitivity to some noises.

Similar behavioral, pupil, and alpha-band power changes between visual and auditory conditioning paradigms are consistent with theoretical notions of how physiological changes are related to cognitive and behavioral processes. Physiological measures that index salience, attention, or emotion may occur as part of an adaptive coordinated response to motivationally relevant situations (Lang et al. 1997; Bradley 2000). A threatening situation typically triggers concurrent physiological responses preparing the body to fight or flee, whether or not that action is carried out. So, if different paradigms evoke a similar action disposition, then similar physiological changes could be expected. For example, if it is presumed that pupil dilation can occur as a part of an adaptive general response whenever threat is anticipated, then it is not surprising that it is modulated by emotional scenes (Bradley et al. 2017), sounds (Partala and Surakka 2003), emotional imagery (Henderson et al. 2018), and the anticipation of electrical shocks (Bitsios et al. 1996, 2004).

While generalization appears to emerge in the later trials of acquisition, it is unclear why the overall alpha power reduction showed very little generalization to the GS1 when compared to visual-based studies (Friedl and Keil 2021). Further experiments with different pitch gradients are likely necessary to examine the role of sensory distinctness in auditory generalization learning. There is considerable research on tone discrimination in humans (see, e.g. Moore 2012). However, this work has not systematically been applied to aversive conditioning. Tone differentiation tasks typically involve presenting tones in quick succession and can differ widely based on the individual differences and features of the tone besides its frequency (Sek and Moore 1995; Smith et al. 2017). These past studies suggest that participants should be able to differentiate tones much closer in pitch than the present research. However, during piloting for the present study, participants failed to correctly report CS-US contingencies when pitches were more proximal than in the present study. Interpreting pitch discrimination here is further complicated by the fact that the tones also contained a separate frequency aimed at evoking an auditory steady state response (not analyzed here) possibly making recognition more difficult. To clarify if generalization occurs, future research could either use tones more similar in pitch or add additional tones to see if this elicits a generalized response. Future research is also necessary to understand how the effects of auditory conditioning extend into the extinction phase which was not carried out during EEG data collection in the present study.

Better understanding alpha power changes in differential aversive conditioning could have clinical significance in a variety of ways. Many psychiatric disorders are thought to feature overgeneralization of fear responses, and this appears to be measurable for some conditions such as generalized anxiety disorders in differential conditioning paradigms (Lissek et al. 2010, 2014). However, as has been noted by others, auditory conditioning may sometimes be necessary as visual stimuli are not applicable for all participants such as children or those with disorders of consciousness (Kotchoubey 2014; Rossetti and Laureys 2015). Auditory conditioning paradigms could also provide unique insights into auditory specific symptoms. Noise sensitivity is associated with many psychiatric disorders (Stansfeld 1992), particularly for the disorder of misophonia characterized by sensitive to loud and bodily noises that was recently defined in Diagnostic and Statistical Manual of Mental Disorders (DSM). More generally, noise pollution itself appears to contribute to the prevalence of some mental health conditions (Hardoy et al. 2005; Zaman et al. 2022). Thus, future research to refine auditory conditioning paradigms and comparing the results against paired visual cues may lead to important new clinical findings.

In conclusion, baseline alpha-band power changes were studied in a differential aversive paradigm with auditory cues. Occipital–parietal alpha power changes were found in most participants similar to visual cue studies, showing clear discrimination between conditioned threat and safety cues. There was very little evidence of generalization to the two other unpaired tones. Future work will carefully manipulate the psychophysical and aversive properties of tone cues to examine the relationship between sensory discriminability and aversive generalization. Overall, the present results are consistent with conceptual and mathematical models of aversive learning (Miskovic and Keil 2012; Anderson 2019), emphasizing that sustained aversive conditioning heightens sensory and attention changes, following a predictable time course (Li and Keil 2023). Theoretical models and extant data suggest that aversive learning prompts initial increases in attentive processing of the CS+, followed by the emergence of sparse sensory representations of threat and safety features and a decrease in higher-order cortical involvement in threat processing. Future studies using mathematical modeling and multimodal imaging studies may refine and expand those models, ultimately informing a framework of aversive learning that includes specific cortical dynamics as a key mechanism.

Funding

This research was supported by a grant from the Misophonia Research Fund and by a grant from the Office of Naval Research N00014-21-2324.

Conflict of interest statement: None declared.

References