Neural processing of emotional facial stimuli in specific phobia: An fMRI study

Patients with specific phobia (SP) show altered brain activation when confronted with phobia‐specific stimuli. It is unclear whether this pathogenic activation pattern generalizes to other emotional stimuli. This study addresses this question by employing a well‐powered sample while implementing an established paradigm using nonspecific aversive facial stimuli.


| INTRODUCTION
Specific phobia (SP) is among the most prevalent mental disorders with an estimated 12-month prevalence of 6.4% and 22.7 million affected people in Europe  and a high burden of disease (Kessler et al., 2012). The underlying neurofunctional alterations are not yet well understood (Stefanescu et al., 2018). A deeper understanding of these mechanisms could significantly broaden our knowledge of the development of the disorder and thus help in the improvement of treating it. The existing literature implies changed neurofunctional activity in several key brain areas in patients with SP during the confrontation with fear-inducing phobiaspecific stimuli (Del Casale et al., 2012;Etkin & Wager, 2007).
Most of these studies examine SP patients while exposed to phobiaspecific stimuli, yielding consistent results of heightened neurofunctional activity in the amygdala as well as the insula and the anterior cingulate cortex (ACC) compared to healthy controls (Dilger et al., 2003;Fredrikson & Furmark, 2006;Goossens et al., 2007;Peñate et al., 2017;Siegel et al., 2017;Straube, Glauer, et al., 2006;. For two current meta-analyses, see Gentili et al. (2019), who stress the importance of whole-brain analyses in addition to regions of interest (ROI) analyses, and Chavanne and Robinson (2021), who corroborate the importance of the insula and the cingulate gyrus.
It is not yet clear if heightened neurofunctional activity generalizes to nonspecific aversive stimuli. This could shed light on the specificity of these neurofunctional alterations (Wright et al., 2003) and have implications for clinical practice, as it would mean that exposure to stimuli other than the feared one is necessary to treat patients with SP. When comparing patients with SP (animal phobia) to HC, previous studies found heightened neurofunctional activity in the amygdala, insula, putamen, and middle temporal gyrus when exposed to fear-inducing images like shark or knife attacks (Schienle et al., 2005), and in the rostral ACC in an emotional counting Stroop paradigm, but lower amygdala and insula activation as well as heightened functional connectivity between the left amygdala and the rostral ACC (Britton et al., 2009). Finally, Schaefer et al. (2014) found a higher neurofunctional activity of the amygdala in patients with SP (snake phobia) compared to HC during the exposure to generally fear-inducing video clips, while Killgore et al. (2014) found a decreased activation of the ventromedial prefrontal cortex (vmPFC) when being confronted with masked emotional faces.
However, multiple other studies did not find activation differences between patients with SP (spider phobia) and HC in response to nonspecific aversive stimuli, specifically fear-inducing scenes (Schienle et al., 2007), mutilations, and animal attacks (Sabatinelli et al., 2005;Schweckendiek et al., 2011;Wendt et al., 2008): Furthermore, Wright et al. (2003) found no heightened neurological processing of emotionally valenced human faces in patients with small animal phobia compared to HC.
These conflicting findings might result from several methodological issues. First, all studies mentioned above examined comparably small groups of participants, typically less than 30 patients. Second, the utilized paradigms and stimuli were heterogeneous, further complicating a comparison between different studies. Third, inclusion criteria for phobic individuals were heterogeneous, ranging from subclinical selfreported phobic symptoms to a clinical diagnosis of SP.
Hence, we wanted to examine whether a well-powered sample of properly diagnosed SP patients (spider phobia) with minimal comorbidities shows altered neurofunctional activity in response to aversive face stimuli compared to HC. As stimuli, we used angry and fearful faces (Hariri et al., 2002), that have been shown to be aversive to different groups of participants, both with and without mental disorders Nikolova et al., 2014;Schneider et al., 2018). Furthermore, we investigated whether functional connectivity patterns of the amygdala differentiate between patients with SP and HC, as this region (and its connectivity to other brain areas) has been frequently shown to be of critical importance to the disorder (Lai, 2020;Stefanescu et al., 2018).
On the basis of these studies, we hypothesized (1) that aversive faces robustly activate the amygdala, the insula, and the ACC, compared to neutral stimuli (in our case, geometric shapes) across both study groups (main effect of task); (2) that this activation is higher in the SP group than in the HC group. Furthermore, we assumed (3) that we can also differentiate the two groups with a machine learning (ML) approach, which has not yet been done. Finally, based on the works of Britton and colleagues (2009) we hypothesized (4) that the functional connectivity of the amygdala and (because of results regarding hypothesis 2; see Section 3.3) the calcarine gyrus is altered in the SP group.

| METHOD
The SP group was part of a subproject of a Transregional Collaborative Research Center (CRC-TRR58 "Fear, Anxiety, Anxiety Disorders"). Complete information on the methods and goals of this BÖHNLEIN ET AL.
| 847 project can be found elsewhere (Schwarzmeier et al., 2019). The control group was acquired in parallel within the Marburg-Münster Affective Disorders Cohort Study (MACS; Kircher et al., 2019;Vogelbacher et al., 2018) using the identical magnetic resonance imaging (MRI) setup. All data is available upon request from the corresponding author. This analysis is a secondary analysis of a subsample of these two studies.
2.1 | Participants N = 270 participants responded to public notices and online advertisements for a spider phobia study and were invited following a short telephone interview. Inclusion criteria comprised a clinically significant SP, assessed with the structured clinical interview for DSM-IV (SCID; Wittchen et al., 1997; n = 123 exclusions), as well as adherence to other criteria like the ability to participate in the MRI procedure (n = 7 exclusions), minimal movement in the MRI (less than 3.3 mm of total movement or 3.3°of total rotation; n = 5 exclusions) and having no other comorbid mental diagnosis except for mild or moderate depressive episodes (also assessed with the SCID) and other SP from the animal subtype (n = 13 exclusions). N = 11 participants dropped out between measurement points. For full inclusion criteria, see Schwarzmeier et al. (2019). From the initial sample, n = 111 participants fulfilled these criteria. In this study, we only used the Münster Cohort of this bicentric study to ensure comparability with the HC group. For the HC group (n = 111), participants had to adhere to the same inclusion criteria as the SP group, except the diagnosis of SP. Out of the MACS sample, the HC group was chosen by matching the SP group for sex and age. Demographic and clinical characteristics can be found in Table 1. According to a power analysis using g*power 3.1.9.2 (Faul et al., 2007) our sample size resulted in sufficient power (1-beta=80%) to detect effect sizes as low as Cohen's d = 0.33 at α = .05 (family-wise error [FWE]-corrected for the search volume).

| Procedure
Participants were first invited for a diagnostic interview in which inclusion was secured and the SCID interview as well as the questionnaires were conducted. For the SP sample, one of these questionnaires was the Spider Phobia Questionnaire (SPQ; Hamm, 2017;Klorman et al., 1974) to measure the severity of spider phobia. It consists of 31 binary items. In our sample, Cronbach's α was .75. We included participants for the SP group with a sum score of 20 or higher. Additionally, we conducted the behavioral avoidance test (BAT) with a bird spider (Grammostola Rosea) in a plastic box in the SP sample. In a second appointment, the MRI examination was conducted. The study protocols were reviewed by the Ethics Committee of the Medical Faculty at Muenster University and are in accordance with the Declaration of Helsinki. Written informed consent was obtained.
We used a paradigm originally developed by Hariri et al. (2002) that has been used extensively in our research group and by others (Dannlowski et al., 2011(Dannlowski et al., , 2016Domschke et al., 2012;Nikolova et al., 2014;Redlich, Stacey, et al., 2015;Redlich et al., 2019;Schneider et al., 2018). The paradigm consists of four blocks of negative faces (either angry or fearful, each block lasting 47 s) from the Ekman and Friesen (1976) stimulus set and five blocks of a sensorimotor control task (circles and ellipses). Participants had to choose which of two pictures (either faces or shapes) displayed in the bottom row is identical to a picture displayed above. Each block consisted of six faces or shapes trios, shown for four seconds and with a variable interstimulus interval of 2-6 s. Each block lasted 35 s and the task lasted 363 s. All included participants had a success rate of 89% or higher, there was no difference between the two groups in terms of success, t(220)= −1.372; p = .171.

| Second-level analysis
The following calculations were performed within SPM12, first whole brain (with a minimum cluster size of k = 10) and then for our three ROIs: the bilateral amygdala (k = 462), insula (k = 3627), and ACC (k = 2712), three regions having shown to be central in the neurofunctionality of anxiety (Wager, 2012). These three ROIs were obtained using the Wake Forest PickAtlas (Maldjian et al., 2003) which follows the AAL atlas (Rolls et al., 2020;Tzourio-Mazoyer et al., 2002) definitions. We corrected for multiple testing with the FWE-method and a significance voxel threshold of p FWE < .05 and repeated these analyses with a more lenient threshold of p unc < .05.
As proof of general activation as expected, we analyzed activation for the contrast image faces>shapes, regardless of group (hypothesis 1). Subsequently, we conducted one ANOVA with group as a between-subjects factor to test for group differences (hypothesis 2). If this group contrast yielded significant results, we conducted post-hoc directed t tests (SP > HC and HC > SP).

| Machine learning approach
In addition to standard univariate analyses, we employed a multivariate machine learning (ML) approach to differentiate the two groups using identical data (faces > shapes contrast, once whole brain and once with the amygdala, insula and ACC ROI; hypothesis 3). We used the PHOTON toolbox developed in our workgroup (Leenings et al., 2020; see https://www.photon-ai.com).
To get a non-biased performance estimation, we applied a nested cross-validation scheme. This scheme consisted of 10 inner validation loops to optimize hyperparameters of the ML model and ten outer validation loops to estimate model performance, as proposed by Poldrack et al. (2020).
For the whole-brain analysis, all voxels within a whole-brain mask with a resolution of 4x4x4 mm were used. For the ROI analysis, we used all voxels within the three ROIs as input to the machine learning analysis.
For detailed information, see Supporting Information A.

| Group comparison: SP>HC
When contrasting SP > HC, the whole-brain analysis of the faces>shapes contrast yielded no significant above-threshold results. Furthermore, none of the ROIs showed significant activation differences, both when applying a significance threshold of p FWE < .05 and even when applying an uncorrected threshold of p unc < .05.

| Group comparison: HC > SP
When comparing HC > SP, the whole-brain analysis showed significant activation differences in the occipital lobe, specifically in the right calcarine gyrus (k = 56; peak voxel: x = 8, y = −92, z = −4; t(220) We are confident that the lack of univariate difference is not a result of type II errors due to insufficient statistical power. Our paradigm elicited strong reactions in brain regions thought to play key roles in the processing of fear and anxiety, namely the amygdala, the insula, and the ACC. Furthermore, even lowering the statistical threshold for the amygdala ROI to an overly lenient value of 0.05, uncorrected (resulting in a corrected α of .37, according to a Monte-Carlo simulation by means of the REST toolbox) still yielded no significant results in the selected contrast (SP > HC).

| Machine learning analysis approach
The ML analyses were able to classify our participants to their respective groups with above-chance balanced accuracies of 73% for the whole-brain analysis and 59% for the ROI analysis (p < .001 and p = .015, respectively). For more detailed results, see Tables S1 and S2.

| Psychophysiological interaction
Regarding the psychophysiological connectivity of the left amygdala across both groups, several areas (an occipital cluster including the fusiform gyrus, hippocampus, inferior frontal gyrus, insula, inferior parietal lobus, and calcarine lobus) showed significant results, see Table 3. When contrasting SP > HC, no significant connectivity differences could be found. Contrasting HC > SP, a few regions showed significant connectivity differences: the inferior parietal cortex, the fusiform gyrus, the middle cingulate cortex, the postcentral cortex, and the insula (see Table 3, Figure 1). HC showed a stronger increase of amygdala connectivity while processing faces versus shapes with these regions, hence a tighter coupling, compared to patients with SP.
The results of the PPI for the right amygdala and the calcarine gyrus across both groups were similar to the ones of the amygdala (see Tables 4 and 5). There were no significant connectivity differences in the group contrasts.

| DISCUSSION
The present study addressed the question of whether patients with specific phobia show altered activation when confronted with nonspecific aversive stimuli, in our case emotional faces. Our results are threefold: First, we could not find any activation differences between a large group of patients with SP and an HC group except for a heightened activity of the calcarine gyrus in the HC group. Second, a multivariate machine learning approach could discriminate the two groups based on our contrast images significantly above chance level.
Third, several areas of the brain like the middle cingulate cortex and the inferior parietal cortex displayed significant connectivity with the left amygdala in response to the faces > shapes contrast across both groups and the HC group showed significantly greater connectivity with the inferior parietal cortex, the insula, and the fusiform cortex.

| Activity and connectivity of the amygdala and the calcarine gyrus
The central role of the amygdala in processing fear is well documented and widely accepted (Adolphs, 2013;Davis, 1992;Panksepp et al., 2011;Tovote et al., 2015). Patients with SP show heightened activation when confronted with phobia-specific stimuli (Dilger et al., 2003). Our results show a strong reaction of the amygdala to our paradigm across both groups but no differences between the groups. This could be due to a ceiling effect in our data or indicate that the amygdala of patients with SP does not show heightened responsiveness to nonspecific emotional facial stimuli. Regarding the PPI results across both groups, several areas showed a significant psychophysiological interaction with the amygdala. These results emphasize the importance of the amygdala in the processing of aversive stimuli and particularly emotional faces.
All three utilized seed regions showed very similar psychophysiological interactions, potentially because of their high interconnectivity.
Additionally, while we did not find areas of significantly higher connectivity with the left amygdala in SP compared to HC, we did find significant differences between the HC > SP contrast in clusters known to be active in the processing of emotional faces (Kropf et al., 2019;Radua et al., 2010): Both posterior parietal areas and the insula have been proposed to be part of a more complex, conscious processing of compared to a more direct, less elaborated route (Dbiec & LeDoux, 2009;LeDoux, 1996). In our study, HC participants did not show higher activation of these areas, but a stronger connection with the amygdala.

| Activity of the insula, ACC, and other brain areas
Similarly, our results regarding the insula and the ACC fit into the broader context of current research: Although both regions have been found to play a central role in the processing of anxiety and fear (Etkin & Wager, 2007;Goossens et al., 2007;Straube, Glauer, et al., 2006), our results and those of other studies (Sabatinelli et al., 2005;Schienle et al., 2007;Schienle et al., 2013;Schweckendiek et al., 2011;Wendt et al., 2008;Wright et al., 2003) seem to indicate that these differences in activation of the insula and the ACC in patients with SP cannot be seen in the processing of nonspecific aversive stimuli-except for the already discussed altered connectivity with the amygdala.
In addition, we found a hypoactivation of the calcarine gyrus in the SP group compared to HC in the faces>shapes contrast.
Contradicting this, other researchers found a hyperactivation of the occipital lobe, and specifically the calcarine gyrus, in the same comparison (Van Strien & Van der Peijl, 2018;Wiemer & Pauli, 2016) which was suggested to be a transdiagnostic trait of heightened processing of threatening stimuli (Feldker et al., 2017).
These divergent results could be explained by the nature of our stimuli: maybe patients with SP only show this hyperactivation when confronted with more fear-inducing or phobia-specific stimuli.
F I G U R E 1 Activity difference and PPI results of the HC versus SP contrast. (a) Results of the faces>shapes contrast across both groups (x = −32, y = 13, z = −5). b) Results of the faces>shapes contrast in the HC > SP contrast (x = 9, y = −89, z = −5). c) Results of the PPI analysis (faces > shapes, left amygdala as seed region) across both groups (x = 28, y = −28, z = 24). d) Results of the PPI analysis (faces > shapes, left amygdala as seed region) in the HC > SP contrast (x = −40, y = −12, z = 8). In the coronal and the sagittal view, the significant parietal inferior cortex is shown. Additionally, the sagittal view shows the insula. In the axial view, the fusiform gyrus is shown.

| Machine learning results
Our ML results show a discrimination success of over chance, which is similar or slightly less successful than comparable approaches that tried to discriminate between SP and HC on the basis of structural MRI data  or to predict therapy outcome in patients with several different mental diseases (Ball et al., 2014;Frick et al., 2020;Hahn et al., 2015;Hilbert et al., 2020;Leehr et al., 2021;Månsson et al., 2015). We find that our ML paradigm utilizing activity in the whole brain is more successful than one limiting itself to our ROIs, once again corroborating our theory that differences in neurofunctional activity of patients with SP are not limited to these areas. Although these results are promising, their clinical relevance still must be verified. This might be due to the poor discriminative power of the underlying data in our study. However, another potential explanation could be that earlier studies had comparably small sample sizes, which might have led to overfitting of the data (Flint et al., 2019) and thus replication with larger samples is needed (Rashid & Calhoun, 2020).

| Limitations
Some limitations of the presented study should be mentioned. First, emotional faces may not be best suited to examine the reaction of patients with SP to nonspecific aversive stimuli. Almost all studies that are in line with our results used paradigms utilizing emotionally valenced but not graphic stimuli, commonly pictures of faces (Wright et al., 2003), while studies that found differences used more graphic stimuli (Sabatinelli et al., 2005;Schienle et al., 2005;Schweckendiek et al., 2011;Wendt et al., 2008). Additionally, the Hariri paradigm has not been used in a study with SP before. Thus, we are not able to offer a fully satisfying solution to the heterogeneity problem in our field of research. Another potential critique on the Hariri paradigm could be that it is in fact too successful, resulting in a ceiling effect that prevents the SP group from showing even higher activation than the HC group.
Furthermore, the difference in activation in the calcarine gyrus could also be explained by a hypoactivation to shapes in the HC group. As our paradigm did not include a baseline condition, we cannot rule this out.
Finally, we cannot generalize our data to all subtypes of SP, as our sample only consisted of patients with spider phobia. Studies have shown that there are neurofunctional differences especially between patients with animal phobia and dental phobia (Lueken et al., 2011(Lueken et al., , 2014Stefanescu et al., 2018). Second, our sample size consisted primarily of female participants (87%, see Table 1), which is not unusual (Britton et al., 2009;Schienle et al., 2005;Wendt et al., 2008), but limits the generalizability to male patients suffering from SP.

| Conclusion
Although we could find strong neurofunctional differences in our paradigm across both groups, SP patients' activity did only differ from HC's in the calcarine gyrus in a univariate approach. However, we did find altered task-related connectivity between the amygdala and regions typically associated with more elaborated processing of emotional stimuli. Together with the ability to differentiate the two groups above chance in a multivariate ML approach based on our paradigm, these results might point to a small difference in processing nonspecific aversive stimuli that do not show themselves when processing emotional faces but might potentially be found in other paradigms using phobia-specific or different aversive nonspecific stimuli. Some methodological questions need further research, mainly whether the nature of the stimuli plays a role in differentiating the neurofunctional reactions to nonspecific aversive stimuli of patients with SP from those without.

ACKNOWLEDGMENT
This study was funded by the Deutsche Forschungsgemeinschaft (DFG)-Projektnummer 44541416-TRR 58 (CRC-TRR58, Projects C09 and Z02 to UD and UL). Open Access funding enabled and organized by Projekt DEAL.