The effect of trait anxiety on attentional mechanisms in combined context and cue conditioning and extinction learning

Sensory processing and attention allocation are shaped by threat, but the role of trait-anxiety in sensory processing as a function of threat predictability remains incompletely understood. Therefore, we measured steady-state visual evoked potentials (ssVEPs) as an index of sensory processing of predictable and unpredictable threat cues in 29 low (LA) and 29 high (HA) trait-anxious participants during a modified NPU-paradigm followed by an extinction phase. Three different contextual cues indicated safety (N), predictable (P) or unpredictable threat (U), while foreground cues signalled shocks in the P-condition only. All participants allocated increased attentional resources to the central P-threat cue, replicating previous findings. Importantly, LA individuals exhibited larger ssVEP amplitudes to contextual threat (U and P) than to contextual safety cues, while HA individuals did not differentiate among contextual cues in general. Further, HA exhibited higher aversive ratings of all contexts compared to LA. These results suggest that high trait-anxious individuals might be worse at discriminating contextual threat stimuli and accordingly overestimate the probability and aversiveness of unpredictable threat. These findings support the notion of aberrant sensory processing of unpredictable threat in anxiety disorders, as this processing pattern is already evident in individuals at risk of these disorders.


Material and Methods
Participants. The sample included sixty subjects recruited via advertisements on the internet and on a local participant-platform. Participants were required to be between 18 and 35 years old, free of any family history of photic epilepsy, any psychiatric or neurological disorders, and have normal or corrected vision. All participants completed a pre-screening including four items of the STAI-Trait questionnaire, which showed the highest correlation with the STAI-Trait sum-score 47 in a large screening sample (n = 526), that was recruited within the context of the Collaborative Research Centre (SFB-TRR58) at the University of Würzburg (for further details see 48 ). On this basis, we used a regression approach to estimate the STAI-Trait score from four items of the STAI-Trait questionnaire that explained the most variance of the total scores to select 30 high and 30 low trait-anxious participants. Cut-offs were chosen based on the distribution of the local sample. The resulting cut-off scores were ≤33 (≤40%-quantile) for low trait-anxious and ≥41 (≥80%-quantile) for high trait-anxious participants. All participants completed the entire German version of the STAI-Trait questionnaire during the experimental session (see Table 1). The classification was mainly confirmed, only one participant of each group fulfilled the criterion of the respective opposite group and was consequently excluded, resulting in a final sample of 29 high trait-anxious (15 females) and 29 low trait-anxious (15 females) participants. To further characterize the final sample, participants were required to complete several questionnaires measuring symptoms of anxiety, including the Anxiety Sensitivity Index 3 49,50 , Intolerance of Uncertainty Scale 51,52 and the GAD-7 53 . All participants gave written informed consent and were paid 15 € or received an equivalent in course credits. The experiment was performed in accordance with relevant guidelines and regulations. All procedures were approved by the ethics committee of the University of Würzburg.
Materials and stimuli. The stimuli used were similar to those described in Wieser et al. 7 . Briefly, central cues were black-and-white Gabor patches, consisting of sinusoidal grating (Gaussian-windowed with maximal contrast at centre) with a spatial frequency of 1.4 cycles per degree. From a viewing distance of 100 cm, the Gabor patches spanned a visual angle of 5.6° horizontally and vertically. We used three different orientations (−45°, 0° and 45°) for the three contextual conditions. Visual stimuli were counter-balanced for conditions across participants. Contextual cues were an array of four triangles, squares or circles, presented peripherally ca. 4.7° of visual angle from the central grating, spanning visual angles of 2.4°. All visual stimuli were presented on a 19-inch monitor (resolution = 1024 × 768 pixels) with a vertical refresh rate of 60 Hz, located ca. 80 cm in front of the participant, using the Presentation software (Neurobehavioral Systems, Inc., Albany, CA, USA).
Aversive unconditioned stimuli (US) were 20 ms electric pulse trains, composed of 5 pulses of 2 ms and intervals of 2 ms between the pulses, which were delivered to the left calf through surface bar electrodes consisting of two gold-plated stainless-steel disks of 9 mm diameter and 30 mm spacing. The electric stimuli were generated by a constant current stimulator (Digitimer DS7A, Digitmer Ltd., Welwyn Garden City, UK). Prior to the actual experiment, the US were adjusted to the individual pain-threshold. Thus, participants received two series of increasing and decreasing intensities until they reached a level they described as "just noticeable pain" -corresponding to 4 on a scale from 0 (no pain at all) to 10 (unbearable pain). The individual US intensity was determined by calculating the mean of the four series' final intensities and then adding 30% to avoid habituation. The resulting intensities and subjective pain ratings were 2.2 ± 1.4 mA (mean ± SD) for LA individuals and 1.9 ± 1.2 mA for HA individuals, t(56) = 0.91, p = 0.366, respectively 5.4 ± 0.8 for LA individuals and 5.7 ± 1.5 for HA individuals, t(56) = 0.86, p = 0.394.
Procedure. Participants completed the questionnaires before being seated in a sound-attenuated, dimly lit testing room, where the EEG-net was applied. Participants then were instructed to fixate a centrally presented fixation cross and to reduce any movements including eye movements and blinks during the experimental session. We used a modified version of the original NPU-threat task 17 , similarly as reported previously 7 . In brief, three different conditions were indicated by the presentation of the contextual cues for 32 s (one block), flickering at a frequency of 15 Hz (Fig. 1). During each block, two or three Gabor patches were presented centrally for 3 s, flickering at 20 Hz. Each Gabor patch was randomly presented between 6 s and 27 s after context onset, with an inter-stimulus-interval of at least 3 s. In the predictable condition (P), Gabor patch offset was paired with an US presentation (100% reinforcement rate). In the unpredictable condition (U), two or three US were presented www.nature.com/scientificreports www.nature.com/scientificreports/ unpredictably between 5 s and 30 s after context onset, with a temporal lag of at least 3 s, but never during the presentation of a Gabor patch. In the neutral condition (N), no US were presented. Participants were fully instructed about the US-contingency. To avoid that context onset and central cue presentation during the U-condition acquired safety-signal properties, we implemented a practice phase of 6 blocks (each condition twice) prior to the actual acquisition phase, in which one US in the U-condition was always presented during the first 5 s and another during the central cue presentation. After the practice phase, participants were also familiarized with the rating scales. The acquisition phase consisted of 36 blocks overall (12 per N, P, or U-condition) with half of each block containing two and three central cue presentations respectively, resulting in 30 central cue presentations per condition. In sum, 30 US were also presented during the U-condition to obtain the same amount of US as during the P-condition. Blocks were separated by inter-trial-intervals (ITIs) of random durations between 2,500 ms and 3,500 ms. To investigate extinction learning, we implemented an additional phase, which equalled the acquisition phase in number of blocks and stimulus timing, except no US were delivered. Participants were not instructed about the absence of the US during extinction. After every 12 th block, participants were asked to rate the contextual cues and the combination of context and central cue separately regarding perceived threat ("How threatening do you perceive this picture?", NRS, 1 = not threatening at all -9 = very threatening) and US-contingency ("To what extent do you expect an electrical stimulus meanwhile this picture is present?", VAS, ranging from 0 to 100%). EEG recording and data analysis. EEG data analysis was conducted and is reported according to published guidelines 54 . Electrocortical brain activity was recorded using a 129 electrodes Electrical Geodesics System (EGI, Eugene, OR) referenced to the vertex electrode (Cz), with a sampling rate of 250 Hz and an on-line band-pass filter of 0.1-100 Hz. Electrode impedances were kept below 50 kΩ. Subsequent data processing occurred off-line using the EMEGS software for Matlab 55 . In a first step, all data were filtered using a 40-Hz low-pass filter (cut-off at 3 dB point; 45 dB/octave, 19 th order Butterworth), before extracting epochs from 600 ms pre-to 2,900 ms post-onset for central cue and 600 ms pre-to 4,900 ms post-onset for contextual cue responses. Following the guidelines for the statistical correction of artefacts in dense array studies procedure (SCADS 56 ), we first detected individual channel artefacts based on the original recording reference (Cz), before data were re-recorded to the average reference to identify global artefacts. Bad sensors within individual trials were identified based on rejection criteria for the distributions of the maximum absolute amplitude, standard deviation and gradient. Contaminated trials were removed, if they included more than 20 bad sensors. After rejection, contaminated sensors of the remaining epochs were interpolated using weighted spherical splines fit to all remaining sensors. The mean rejection rate for contextual cue and central cue responses were 28% ± 7 (mean ± SD) and 23% ± 10 (mean ± SD), respectively. Also, there were no group differences for the total number of usable trials www.nature.com/scientificreports www.nature.com/scientificreports/ associated with the central and contextual cues 57 . The time-varying amplitude of the ssVEP signal was then extracted as the modulus of the filtered empirical signal and the Hilbert-transformed analytic signal.
For statistical analyses, the ssVEP amplitudes were averaged across time points between 100 ms to 2,900 ms for central cue onset and between 100 ms to 4,900 ms for contextual cue onset. Only the onsets of the contextual cue presentations were analysed, since later time intervals were confounded with electrical stimulations that lead to strong artefacts in the EEG signal. We used the same cluster of electrodes as in our previous study 7  was selected for contextual cue analysis, with the latter covering a larger area, as the peripherally presented contextual cues elicited a more wide-spread electrocortical activation (see Fig. 2). Statistical analysis. The mean ssVEP amplitude reactions to central and contextual cues were analysed separately for acquisition and extinction with mixed-measure analysis of variances (ANOVA) with the within-subject factor condition (N vs U vs P) and the between-subjects factor group (LA vs HA). Averaged threat and US-contingency ratings were examined using the same procedure. Significant effects were followed up using t-tests. A significance level of 0.05 was used for all analyses and Greenhouse-Geisser correction was applied where appropriate 58 . Throughout this manuscript, the uncorrected degrees of freedom, the corrected p values and the partial η² (η p 2 ) or Cohen's d (d) are reported 59 . Ratings of central and contextual cue combinations. The combination of context and central cues were rated differentially with regard to perceived threat. During acquisition, the ANOVA yielded a significant main effect of condition, F(2,112) = 154.71, p < 0.001, η p 2 = 0.734. The groups did not differ on any level, ps > 0.208. As expected, post-hoc t-tests revealed that during acquisition cues in the predictable, t(57) = 16.64, p < 0.001, d = 2.184, and in the unpredictable condition, t(57) = 11.58, p < 0.001, d = 1.521, were rated as more threatening than the cues in the neutral condition. Moreover, cues in the predictable condition were rated as more threatening compared to cues in the unpredictable condition, t (57)  As for threat ratings, the ANOVA for US-contingency during acquisition showed a significant main effect of condition, F(2,112) = 283.22, p < 0.001, η p 2 = 0.835, without any effects of group, ps > 0.463 (see Fig. 5). Again, post-hoc t-tests showed that cues in the predictable, t(57) = 30.56, p < 0.001, d = 4.012 and in the unpredictable condition, t(57) = 11.93, p < 0.001, d = 1.567, were associated with higher shock expectancy than in the neutral condition. While the cues in the predictable condition had also higher likelihoods than the cues in the unpredictable condition, t(57) = 9.75, p < 0.001, d = 1.281. During extinction, the differences between conditions stayed  Fig. 6). The high anxious group generally rated contextual cues as more threatening compared to the low anxious group. To follow up on the main effect of condition, several t-tests were calculated. The unpredictable context was rated with higher threat ratings compared to the neutral and predictable context, t(57) = 18.14, p < 0.001, d = 2.381 and t(57) = 7.34, p < 0.001, d = 0.964. The predictable context was perceived as more threatening than the neutral context as well, t(57) = 8.41, p < 0.001, d = 1.104. During extinction, these differences remained significant,

Discussion
The purpose of the present study was to compare electrocortical correlates of attention allocation during predictable and unpredictable threat acquisition and extinction learning in high vs. low anxious individuals. For this goal, we investigated steady-state visual evoked potentials along evaluative ratings in a combined cue and context conditioning paradigm.
Ratings of threat and US-contingency demonstrated that predictable and unpredictable threat was acquired as expected. Both groups showed elevated evaluative responses to the central cue in the predictable threat condition and to contextual cues in the unpredictable threat condition, replicating our earlier results 7 . Moreover, verbal responses to predictable and unpredictable threat cues decreased during extinction.
Low and high anxious individuals did not differ regarding ratings related to predictable threat, but during acquisition, high anxious individuals demonstrated higher threat and contingency ratings for contextual cues than low anxious participants irrespectively from the condition. These results suggest elevated sensitivity to contextual threat in high anxious individuals and are in line with a growing body of literature suggesting that heightened sensitivity to unpredictable threat may be a hallmark for anxiety disorders [21][22][23] and extend these effects to a subclinical sample. Post-hoc analyses during the extinction phase revealed higher threat ratings in  www.nature.com/scientificreports www.nature.com/scientificreports/ the unpredictable threat condition for high anxious compared to low anxious individuals, which is in line with a recent meta-analysis, indicating reduced or slowed extinction learning in patients with anxiety disorders 46,60 . Even though the meta-analytical results are mainly based on predictable threat conditioning paradigms, in our experiment high trait-anxious individuals demonstrated reduced extinction learning for unpredictable threat only. This lends further support for the notion that high anxious individuals are especially sensitive to uncertain threat.
During predictable threat conditioning, sensory processing was facilitated for the central cue in the P-condition compared to the U-condition, supporting the idea of selective attention to specific threat stimuli 8,9,31 . However, in contrast to our earlier study, we only found a descriptive, non-significant difference between the Pand the N-condition, which may be due to methodological variations in the present design.
Regarding contextual cue onset analysis, different patterns were observed for high and low anxious individuals. Low anxious individuals showed increased activity for the unpredictable and predictable threat contexts compared to the neutral context. This confirms results of Kastner-Dorn et al. 8 . Thus, contextual cues, associated with an aversive event -independent of threat predictability -draw more attentional resources than the safety context. This result supports the notion that threat relevant stimuli gain enhanced visual processing. After all, the onset of the predictable and unpredictable contextual cue both signal aversive events during the next 32 seconds, www.nature.com/scientificreports www.nature.com/scientificreports/ while only the neutral context reliably signals no upcoming aversive events. During the predictable context, however, aversive events are further signalled by a discrete cue, but not in the unpredictable context. Paradoxically, the onset of the discrete cue during the predictable condition is unpredictable, which in turn serves as a source of uncertainty about when the aversive event appears. Note that Wieser et al. 7 found enhanced visual processing of the unpredictable contextual cues only. However, in contrast to this previous study, we this time analysed not only the first second but an interval of five seconds after contextual cue onset. In the same way, other studies report that during the first second after stimulus onset the ssVEPs does not discriminate as strongly between CS+ and CS− as in a later interval 35,61,62 , which could explain the found discrepancies.
High anxious individuals did not differentiate in ssVEP activity to the contextual cues among conditions. A very similar pattern was recently observed in a study by Ahrens et al. 43 who associated different neutral faces with neutral, positive or negative verbal comments. In contrast to low socially anxious individuals, high socially anxious individuals did not differentiate between the three conditions in the ssVEP amplitudes, indicating a potentially impaired ability to discriminate between relevant and irrelevant social stimuli. The authors interpreted this finding as a possible hyperactivation of the amygdala and a consequent overgeneralization of conditioned social threat in social anxious individuals. In the current study, high anxious individuals might similarly be characterised by a discrimination deficit for contextual cues, explaining the diminished visual discrimination of these cues. In addition, high anxious individuals also showed attenuated attentional processing of the threatening contextual cues compared to low anxious individuals, which might suggest a potential indicator of perceptual avoidance in early stages of visuocortical processing. Even though there is a vast body of literature describing attentional hypervigilance-avoidance biases in anxiety 63 , empirical evidence for perceptual avoidance in ssVEP measures is sparse 44 . Notably, a recent study observed reduced ssVEP amplitudes to aversive relative to neutral facial expressions for severely impaired social anxiety patients 64 . In this study, visuocortical responding to threatening stimuli in anxiety patients depended on the severity of the disorder: patients diagnosed with circumscribed social anxiety showed increased vigilance to aversive facial expressions, while attentional processing of the most impaired patients was actually decreased, reflecting a possible indicator of perceptual avoidance 64 . Although, we investigated solely high anxious, subclinical individuals, accumulating evidence indicates that initial hypervigilance and consequent perceptual avoidance of aversive stimuli might be a potential risk factor characterizing highly anxious individuals 65 . However, it should be noted that due to early non-stationary components of ssVEPs 35,37 , it might be difficult to capture the initial short-lasting hypervigilance, which is already evident in early electrocortical components (about 50-100 ms) of visual event-related potentials 66,67 . For that reason, it is important to mention, that we cannot definitely distinguish between impaired discrimination abilities or perceptual avoidance in high anxious individuals. Both mechanisms seem plausible. Also, they are not necessarily mutually exclusive and could occur together. Future research may address this issue by using a combination of eye tracking and ssVEPs to record measures of initial hypervigilance and sustained visuocortical responding within the same paradigm. Further, these sensory processing biases could also explain the elevated threat and contingency ratings for the contextual cues independent of the conditions. Since high anxious individuals might be worse at discriminating or perceptually avoid contextual threat stimuli, they accordingly overestimate the probability and aversiveness of unpredictable threat.
During extinction learning, only the context which indicated safety in the acquisition phase gained enhanced attentional resources. A reason for that could be that participants realized the sudden absence of aversive events and consequently grew suspicious of the other conditions, since the neutral condition may have become the best candidate for a change in the US-contingencies. Accordingly, the increased visuocortical processing of the former neutral stimuli may reflect the electrocortical correlate of an attentional reorientation process. A very similar pattern was found in McTeague et al. 42 , where visuocortical responses to former safety stimuli increased and responses to former threat stimuli decreased during extinction. The reason why responses to the former threatening stimuli did not decrease during extinction seems in contrast to the results of McTeague et al. 42 . However, it should be considered that we instructed participants only in regard to the acquisition phase, but not about the absence of the US during extinction. Also, another explanation for the increased ssVEP amplitudes to the neutral contextual cues during extinction learning seems plausible: As the neutral contextual cues predict -and have always predicted -absence of shocks, they stay the last reliable predictor during extinction learning. Stimuli with a strong predictive value usually are also motivationally significant, which in turn is associated with stronger visuocortical processing 31 . To elucidate potential underlying mechanisms, a recent steady-state visual evoked field (ssVEF) study used a predictive modelling approach to explain increased ssVEF amplitudes to the conditioned safety stimulus during extinction learning 68 . The authors showed that changes in visuocortical processing reflect the associative strength of the CS 69 and can be predicted by estimates of the Rescorla-Wagner model 70 .
One limitation of our current study is that we could only analyse the onset of the contextual cue presentation, since later time intervals were confounded with electrical stimulations that lead to strong artefacts in the EEG signal. Consequently, we can make no statements about visuocortical processing of the contexts during the whole trial. Even so attentional mechanisms seem to be stable over time 8,71 , differences between low and high anxious individuals as well as extinction processes should be considered throughout the whole context presentation intervals in future studies. Finally, since the central cue during the unpredictable condition gained salience, though it was never directly associated with an US, the question arises, whether threat cues are differently processed in contexts that were associated with aversive events compared to contexts that are perceived as safe. In the NPU-threat task, predictable and unpredictable threat are fully separated by conditions, making it a perfect tool to simultaneously investigate fear-and anxiety-like processes in the laboratory. However, it might be important to consider an interplay between cue and context conditioning and hence a possible interaction between fear and anxiety 72 . Future studies may also test if contextual discrimination deficits contribute to a poorer treatment response in anxiety disorders 73 and whether they can be targeted with attentional discrimination training 74 to potentially improve the treatment of anxiety disorders. (2019) 9:8855 | https://doi.org/10.1038/s41598-019-45239-3 www.nature.com/scientificreports www.nature.com/scientificreports/ Altogether, our results support the idea that individuals at risk for anxiety disorders are characterized by aberrant processing of unpredictable threat cues 21,23 . In line with recent findings, attention is selectively focused onto the specific threat stimulus during predictable threat, while during unpredictable threat attentional vigilance for the context is initially increased for healthy low anxious individuals 7,8 . On the contrary and strikingly, high anxious individuals are less able to discriminate between aversive contexts and might even perceptually avoid threatening relative to safe contexts on an early visuocortical level. Increased visual processing of former safety stimuli during extinction learning may reflect electrocortical correlates of an active attention reorientation process after the change of contingencies 42 .

Data Availability
The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.