Group functional magnetic resonance imaging (fMRI) studies suggest that anxiety disorders are associated with anomalous brain activation and functional connectivity (FC). However, brain-based features sensitive enough to discriminate individual subjects with a specific anxiety disorder and that track symptom severity longitudinally, desirable qualities for putative disorder-specific biomarkers, remain to be identified. Blood oxygen level-dependent (BOLD) fMRI during emotional face perceptual tasks and a new, large-scale and condition-dependent FC and machine learning approach were used to identify features (pair-wise correlations) that discriminated patients with social anxiety disorder (SAD, N=16) from controls (N=19). We assessed whether these features discriminated SAD from panic disorder (PD, N=16), and SAD from controls in an independent replication sample that performed a similar task at baseline (N: SAD=15, controls=17) and following 8-weeks paroxetine treatment (N: SAD=12, untreated controls=7). High SAD vs HCs discrimination (area under the ROC curve, AUC, arithmetic mean of sensitivity and specificity) was achieved with two FC features during unattended neutral face perception (AUC=0.88, P<0.05 corrected). These features also discriminated SAD vs PD (AUC=0.82, P=0.0001) and SAD vs HCs in the independent replication sample (FC during unattended angry face perception, AUC=0.71, P=0.01). The most informative FC was left hippocampus-left temporal pole, which was reduced in both SAD samples (replication sample P=0.027), and this FC increased following the treatment (post>pre, t(11)=2.9, P=0.007). In conclusion, SAD is associated with reduced FC between left temporal pole and left hippocampus during face perception, and results suggest promise for emerging FC-based biomarkers for SAD diagnosis and treatment effects.
There is an increasingly recognized need for biomarkers in neuro-degenerative and psychiatric disorders for both early and differential diagnosis, personalized prediction of treatment response, and treatment and drug discovery (Gordon and Koslow, 2010). Biomarker research for anxiety disorders has received relatively little attention, despite the fact they are the most common psychiatric condition, with a lifetime prevalence of 29% (Kessler et al, 2005). Social anxiety disorder (SAD) is among the most common of these disorders (Jefferys, 1997), with a 7–13% lifetime prevalence (Kessler et al, 1994). SAD is characterized by heightened anxiety and avoidance during social interactions. It has an early onset (80% of cases occur before age 18 years) (Otto et al, 2001), and often precedes other anxiety, mood, and substance abuse/dependence disorders (Lampe et al, 2003; Randall et al, 2001). SAD is associated with a significant functional impairment and distress in work and social domains and usually persists unless treated (Schneier et al, 1994; Lochner et al, 2003).
Models of SAD (Clark and McManus, 2002; Rapee and Heimberg, 1997) have highlighted the role of sensitivity to perceived social threats, which is triggered by biased appraisals of social situations. These maladaptive appraisals transform innocuous social cues into interpersonal threats that induce a cascade of fears of negative evaluation, somatic concerns, inhibited behavior, and negative emotional reactivity. Behavioral studies have shown a negative interpretation bias in SAD, such that ambiguous facial expressions are more likely to be deemed as threatening (Veit et al, 2002; Yoon and Zinbarg, 2008a; Mohlman et al, 2007). Neuroimaging studies suggest exaggerated limbic responses to harsh or angry faces (Stein et al, 2002) as well as neutral faces (Gentili et al, 2008) in SAD, while a recent review of implicit face processing tasks in SAD suggests amplified responses to faces in general (Schulz et al, 2013) despite heterogeneity of findings for emotion specificity and activated brain regions. A meta-analysis of neuroimaging studies of anxiety (using tasks that mostly involved emotional face viewing) found overall greater activation of amygdala and insula, structures linked to negative emotional response, in SAD subjects relative to matched comparison subjects (Etkin and Wager, 2007).
Neuroimaging studies have typically used emotional face paradigms to contrast average activation between patients and healthy control (HC) subjects, identifying significant differences in particular brain regions. However, clinical application of neuroimaging for the diagnosis and treatment of anxiety would require a quantitative measure of brain activity that can distinguish single patients with a specific disorder (eg, SAD) from healthy individuals as well as from individuals with a related disorder (eg, panic disorder, PD, which is characterized by panic attacks and anxiety symptoms that overlap those of SAD but are not exclusively related to social stimuli). To accomplish this, we shifted the focus of our data analysis from average differences (or similarities) in regional brain activity between groups to features of brain activity that maximize the probability of predicting the correct diagnosis within a single subject.
A novel approach based on multivariate machine learning-based pattern analysis of large-scale, condition-dependent functional connectivity (FC) recently demonstrated increased sensitivity of patterns of interactivity (ie, pair-wise FC from hundreds of nodes) relative to patterns of activity (ie, beta or contrast activation maps) in predicting subliminal and unattended viewing of fearful vs neutral faces in healthy subjects: functional connections that discriminated supraliminal fearful from neutral faces included amygdala, fusiform, thalamus, superior temporal sulcus, superior occipital cortex, hippocampus, angular gyrus, and cerebellum (Pantazatos et al, 2012b) whereas FC that discriminated subliminal fear from neutral faces included middle temporal gyrus, angular gyrus, cerebellum, superior frontal gyrus, and amygdala (Pantazatos et al, 2012a). On the basis of this and previous evidence that SAD subjects exhibit anomalies in the cognitive-emotional processing of emotional and ambiguous social stimuli, it was hypothesized that patterns of interactivity during fearful and neutral face processing would be sensitive in discriminating whether an individual subject has SAD. The current work examined (1) whether pattern classification of FC during unattended or subliminal processing of emotional or neutral faces can predict SAD diagnosis (control vs SAD, and SAD vs PD), (2) whether discriminating features from the above were also replicated in an independent sample, and (3) whether these same discriminating features tracked symptom severity longitudinally in both HCs and subjects that underwent 8 weeks of treatment with the selective serotonin reuptake inhibitor (SSRI) paroxetine. A treatment/longitudinal component of our study was included to test whether features that predict cross-sectional diagnosis are sensitive to changes in symptom severity, and hence may be potential neurobiological markers of disease state progression and/or response to treatment interventions. A secondary objective compared the classification performance achieved when using interactivity (pair-wise correlations) vs activity (ie, beta estimates from SPM maps). If FC-based features can be shown to reliably categorize subjects with a diagnosis of SAD, discriminate them from subjects with a closely related disorder such as PD, and demonstrate a tendency to normalize following treatment, then this would represent an important advance in the development of biomarkers for psychiatric diagnosis and treatment effects.
MATERIALS AND METHODS
All procedures and tasks were reviewed for ethical concerns and protection of human subjects by the Columbia University and New York State Psychiatric Institute Institutional Review Boards before subject recruitment and data collection.
Twenty HC subjects, eighteen subjects (ages 18–50) diagnosed with SAD, and sixteen patients diagnosed with PD were recruited through web advertisements (except for seven of the SAD subjects recruited from a genetic study of anxiety; Talati et al, 2008). Functional scans of two subjects (one control and one SAD) were unusable due to scanner technical issues, while a third subject was excluded because she was diagnosed with both SAD and PD. Recruitment and clinical procedures have been detailed elsewhere (Talati et al, 2013) and are described in more detail in Supplementary Methods.
Eighteen medication-free adults with a primary diagnosis of GSAD (age 20–52) and seventeen age, sex, and race-matched HCs were recruited through media notices and clinical referrals. Diagnoses were based on the psychiatric interview and confirmed by the Structured Clinical Interview for DSM-IV Axis I disorders. Data from four GSAD patients were excluded from analyses due to technical issues, yielding 14 GSAD patients and 17 HCs used for diagnostic classification analysis. Twelve GSAD patients and seven HCs completed a second scan following the 8-week paroxetine treatment (or non-treatment for HCs). See Schneier et al (2011) and Supplementary Methods for more details regarding subjects, recruitment, and treatment. Due to the small number of HCs that completed a second scan they were not included in longitudinal analyses (however, see Supplementary Results for provisionary analyses that include these subjects).
Subjects performed a previously described task from our group (Etkin et al, 2004; Pantazatos et al, 2012b), which consists of color identification of Ekman fearful, neutral, masked fearful, and masked neutral faces (F, N, MF, and MN, respectively) within a blocked paradigm (four 20 s blocks for each condition, 15 s baseline between each block). See Supplementary Methods for further details regarding the task paradigm and stimuli.
Subjects performed gender identification of angry, happy, and neutral faces (A, H, and N, respectively) drawn from the same standard series as above (Ekman and Friesen, 1976), and within a blocked paradigm (four 20 s blocks for each condition, 12–14 s baseline between each block). See Supplementary Methods for further details
Image Acquisition and Analyses
Functional data were acquired on a 1.5 Tesla GE Signa MRI scanner in the functional MRI Research Center at Columbia University Medical Center, using a gradient-echo, T2*-weighted echoplanar imaging (EPI) with blood oxygen level-dependent (BOLD) contrast pulse sequence. Twenty-four contiguous axial slices were acquired along the AC-PC plane, with a 64 × 64 matrix and 20 cm field of view (voxel size 3.125 × 3.125 × 4 mm, TR=2000, TE=40, flip angle=60). Structural data were acquired using a 3D T1-weighted spoiled gradient recalled (SPGR) pulse sequence with isomorphic voxels (1 × 1 × mm) in a 24 cm field of view (256 × 256 matrix, ∼186 slices, TR=34 ms, TE=3 ms).
Functional data were preprocessed and processed in SPM8 (Wellcome Department of Imaging Neuroscience, London, UK). For preprocessing, the realigned T2*-weighted volumes were slice-time corrected, spatially transformed, and resampled to a standardized brain (Montreal Neurologic Institute, 2 × 2 × 2 mm3 cube resolution) and smoothed with a 8-mm full-width half-maximum Gaussian kernel. First-level regressors were created by convolving the onset of each condition (primary sample: MF, MN, F, and N, replication sample: A, H, and N) with the canonical HRF with duration of 20 s. Additional nuisance regressors included six motion parameters, white matter and csf signal, which were removed before time-series extraction. For the current work, the same GLM analysis served two main purposes: (1) facilitate removal of nuisance effects from time series before FC estimation using structurally (atlas-based) defined ROIs, and (2) produce beta estimates of each condition for case vs control classification analyses (primary sample) using spatial activity patterns.
Atlas-based parcellation (see Supplementary Figure 1) was applied and pair-wise correlations between 248 nodes (derived from 124 atlas-based brain regions) were computed using 40 total time points of fMRI data that were segmented and concatenated from four conditions; unattended and non-masked (ie, unattended) F and N faces, and subliminal, MF and MN faces (see Pantazatos et al, 2012b and Pantazatos et al, 2012a and Supplementary Methods for more details and analysis schematic). Correlations over the full run were also computed (Full). This resulted in 30 628 total functional connections (z-transformed Pearson’s correlations) for each condition of interest (F, N, MF, MN, and Full), which were used as features for diagnostic classification.
Pattern analysis of large-scale FC to predict SAD diagnosis
For all binary classification tasks, a linear kernel SVM (Vapnik, 1999) with a filter feature selection (t-test) and leave-one-out cross-validation was used. During each iteration of leave-one-out cross-validation (primary sample), one subject was withheld from the data set and (1) a 2-sample t-test was performed over the remaining training data, (2) the features were ranked by absolute t-score and the top N were selected, (3) these selected features were then used to predict the class of the withheld test examples during the classification stage. For classification in the replication sample, the SVM model was learned from the whole primary sample using the top two features identified in the analysis above, and this same model was used to predict SAD vs controls in the replication sample. Before learning, the effects of age and gender were regressed out from the features using a general linear model, and features were z-scored. Classification, performance assessment, and confidence interval estimation followed previously described procedures (Pantazatos et al, 2012a, 2012b), and are further described in Supplementary Methods.
Univariate replication analyses
Features identified in the primary analysis were subjected to univariate statistical tests in the replication sample. SAD vs HC (pre-treatment) was assessed with the Mann–Whitney U-test, pre-post group changes in FC were assessed using the paired t-test, and longitudinal pre-post correlations (decrease ΔLSAS vs increase ΔFC) were assessed using Spearman’s Rho (rank correlation coefficient). Our hypotheses were based on the directions observed in the primary sample, so reported P-values are one-tailed.
The average response rates in the color discrimination task were: HCs 96.4% (σ=6.2%), SAD 99.9% (σ=0.15%), and PD 98.1% (σ=4.2%). Mean accuracies were HCs 96.3% (σ=4.4%), SAD 99.1% (σ=0.96%), and PD 97.0% (σ=4.3%). Mean reaction times were: HCs 0.66 s (σ=0.09 s), SAD 0.64 s (σ=0.16 s), and PD 0.70 s (σ=0.11 s), indicating that subjects performed the color discrimination task as instructed. The groups did not differ in RT (F50=1.75, P=0.18) or accuracy (F47=0.46, P=0.63, one-way ANOVA).
Discriminating Between SAD and HCs with Patterns of FC
For SAD (n=16) vs HC (n=19) classification in the primary sample (see Table 1 for demographics), a peak area under the ROC curve (AUC) of 0.88 (P<0.004 corrected, 0.81, 1.0 90% CI) was achieved when learning was based on the top two features in each training set, derived from the N condition (Figure 1a; see Supplementary Figure 2 for results from all conditions). The accuracy decreased to chance performance levels when more than three features were included in the feature set, indicating that additional features were irrelevant to SAD diagnosis. This is consistent with observed performance degradation of Support Vector Machine (SVM) classifiers when irrelevant features are included in the feature set (Judson et al, 2008). Anatomical display of these two features revealed functional connections between left hippocampus and left temporal pole, and between right anterior middle temporal gyrus and left orbitofrontal cortex (Figure 2b; Table 2A). When comparing classification performance of each feature alone, left hippocampus-left temporal pole was more discriminating than right anterior middle temporal gyrus-left orbitofrontal cortex (AUC=0.77 vs 0.68, data not shown).
Specificity of these features to the SAD diagnosis was tested by classifying SAD vs subjects with PD (n=16). Using only the above two features (no feature selection), an AUC of 0.81, P=0.0001 uncorrected was achieved in discriminating between SAD and PD patients (Table 2B), suggesting relative specificity of these features to SAD. These two features did not discriminate HCs from PD (AUC=0.47, data not shown).
We note that that this primary sample was not balanced for gender (HCs: 11 males, 8 females; SAD: 2 males, 14 females). We therefore tested whether the top two features identified above could predict gender among the combined group (13 males vs 22 females). Classification was not any greater than chance for this classification (AUC=0.50), verifying that classification performance was not an artifact of a sample mismatched for gender. In addition, there was only a slight decrease in performance after applying multiple regression to each feature and removing the effects of age and gender: original AUC/adjusted AUC=0.88/0.80. Importantly, replication of these findings (described more below) was tested in a sample that was matched for age and gender.
In addition to the exploratory, data-driven approach above, we examined FC previously identified to be anomalous in SAD, in particular reduced aINS-dACC (Klumpp et al, 2012) and amygdala-dACC and amygdala-dlPFC (Prater et al, 2012) in SAD during fear. Although FC differences were mostly consistent with these studies (particularly with the dACC, see Supplementary Results), including the above connections in addition to the two features identified above, individually or collectively, did not improve classification performance, whereas including only these connections resulted in poorer classification performance (AUC=0.53). In addition, SAD vs HC classification was also conducted using spatial patterns of activity (ie, beta and contrast estimates, which are considered summary measures of activation in response to each condition), and standard GLM/regression analysis and psychophysiological interaction (PPI) analyses seeding with left hippocampus and left temporal pole were conducted (see Supplementary Figures S3–S5, Supplementary Tables S3 and S4, Supplementary Results and Discussion for further details). In summary, during face processing, significantly greater activation of dorsolateral prefrontal cortex was observed in controls, while greater fusiform activity was observed in SAD. However, diagnostic classification performance using these activation patterns was inferior to classification performance using patterns of large-scale FC.
Discriminating SAD vs HCs in an Independent Replication Sample
To see how well classification using the top two FC features identified above generalized to new data, these features (left hippocampus-left temporal pole and right anterior middle temporal gyrus-orbitofrontal cortex) were tested in a second, independent sample of SAD (n=14) vs HCs (n=17). This sample also performed unattended perception of emotional faces, but instead of color identification of fearful and neutral faces (primary sample), they identified the gender of angry, happy, and neutral faces. For this, an SVM model (line) was learned using the full primary data set for the above two features (Figure 2a). To increase generalization ability of the model, features were corrected for the effects of age and gender and z-scored before SVM learning (under these preprocessing steps classification performance was only slightly reduced to AUC=0.86). This learned model was applied to the independent replication sample, and the highest AUC was achieved when using FC during angry faces (angry: sensitivity=0.71, specificity=0.71, AUC=0.71, P=0.01 (0.59, 0.88) 90% CI; Figure 2b; Table 3D; happy: AUC=0.47, P=0.77; neutral: AUC=0.54, P=0.43; full: AUC=0.57, P=0.31, data not shown).
Univariate group comparisons over each feature revealed that left hippocampus-left temporal pole FC was significantly reduced in SAD vs HCs, particularly during the angry condition (mean difference=−0.25, P=0.027) with a trend-significant reduction in the neutral (mean difference=−0.21, P=0.092) condition (Table 3, first row). Among the cases, this FC did not correlate with symptom severity (as assessed through the Liebowitz Social Anxiety Scale, or LSAS, all P>0.3, data not shown). In addition, there were no significant differences observed between SAD and HCs in the right anterior middle temporal gyrus–orbitofrontal cortex FC (all P>0.2, data not shown).
Changes in Left Hippocampus-Left Temporal Pole FC Following the Treatment
We examined whether left hippocampus-left temporal pole FC could be considered as a possible biomarker for treatment effects. A subset of the replication sample completed an additional scan following 8 weeks of paroxetine treatment (SAD n=12). At the group level, social anxiety symptom severity (LSAS) was significantly reduced following the 8-week SSRI (paroxetine) treatment (cases (pre-post)>controls (pre-post) t(17)=3.4, P=0.003, Table 1B). Pre minus post comparisons within the SAD subject group (n=12) revealed the left hippocampus-left temporal pole FC, particularly during angry faces, increased following the treatment (angry pre>post, mean change in R=−0.41, t(11)=−2.9, P=0.007 paired t-test, Table 3, second row). Thus, a paired t-test within only the SAD group reveals that this FC increased significantly following the 8-week treatment.
When pre-post ΔLSAS vs post-pre ΔFC (left hippocampus-left temporal pole) was correlated among only SAD cases (n=12), positive, yet non-significant, correlations were observed (angry: R=0.31, P=0.16, happy: R=0.38, P=0.11, neutral: R=0.16, P=0.31, full: R=0.11, P=0.36, data not shown). Provisionary analyses of pre-post ΔLSAS vs post-pre ΔFC correlations that combined both cases and seven HCs (ie, correlating longitudinal symptom change with change in FC regardless of treatment status) are presented in Supplementary Results. In addition, left hippocampus-left temporal pole FC at baseline was not associated with ΔLSAS symptom improvement (all P>0.2, data not shown). Change in right anterior middle temporal gyrus-left OFC FC did not correlate with change in symptom severity (angry: R=−0.10, P=0.34, happy: R=−0.20, P=0.2, neutral: R=0.04, P=0.44). Additional analyses suggest decreases in activation in left hippocampus and left temporal pole in response to angry and neutral faces following the treatment, although effects were subthreshold (see Supplementary Results).
In the current work, a novel and exploratory approach based on multivariate pattern analysis of large-scale, condition-dependent FC (Pantazatos et al, 2012b) was used to identify FC that discriminated individual subjects with SAD. FC features that discriminated SAD from HCs in the primary sample also discriminated SAD from HCs and from subjects with the closely related diagnosis of PD with significant sensitivity and specificity. Additionally, following 8 weeks of treatment of SAD with paroxetine, the most discriminative FC feature was increased. Results suggest promise for FC-based biomarkers for psychiatric diagnosis and treatment effects.
Related diagnostic classification studies have applied pattern analysis to condition-dependent activation to a particular probe or stimulus relevant to the disorder (ie, responses to a sad face to predict depression; Fu et al, 2008). While pattern analysis of activation maps takes into account multivariate interactions among regions, activation maps are usually beta maps (summary statistics of activation) or signal that has been averaged across multiple successive scans. Thus, interactions are at a grosser temporal scale. In contrast, the current approach explicitly takes into account scan-to-scan covariation between regions. Here, pattern analysis of FC was more sensitive and specific in discriminating SAD than was multivariate pattern analysis of activation (when using the canonical HRF to model activity), likely due to the fact that it captures information inherent in the interactions among brain regions. Previous large-scale FC approaches capture this information, but only during resting state (ie, resting-state fMRI BOLD to predict schizophrenia (Yu et al, 2013) and age (Dosenbach et al, 2010)). The current approach measures condition-dependent FC (ie, large-scale FC during emotional face viewing), combining the sensitivity of multivariate machine-learning analysis with the advantages of both task-based and resting-state FC approaches.
A meta-analysis of both PET and fMRI studies of SAD before 2007 showed increased activation in amygdala and insula during negative emotional processing (Etkin and Wager, 2007), while many recent fMRI studies of SAD have applied activation analyses focused on the amygdala, and insula (Yoon et al, 2007; Klumpp et al, 2010; Schmidt et al, 2010) to show increased activation in these areas to intense vs low or negative vs neutral emotional stimuli. In addition, univariate analyses showed decreased activation of the amgydala and increased activation in the vmPFC in response to social threat stimuli following 12 weeks of SSRI treatment (Phan et al, 2013). However, although these univariate approaches can identify areas that respond more or less to a particular stimulus, they ignore interactions, or FC, between regions, which are thought to measure information transfer underlying complex cognitive-emotional processing such as during threat or facial effect perception and appraisal (Friston, 2002).
Although previous studies have demonstrated differences in activation and FC (Ding et al, 2011; Liao et al, 2010; Prater et al, 2012; Hahn et al, 2011; Klumpp et al, 2012) between SAD and HCs, the current work is the first to use FC (and activation) to discriminate SAD vs HC diagnostic membership. The current approach of combining machine learning with large-scale, condition-dependent FC is more exploratory and data driven in identifying FC differences than previously used techniques such as PPI Analysis, which only assess FC with a single, a priori specified, ‘seed’ region at a time. The additional discrimination of SAD from PD is particularly notable given that these disorders have a significant overlap of both symptoms and neurobiology, such as amygdala hyperactivation and decreased frontal regulation (Damsa et al, 2009; Rauch et al, 2003).
The most discriminative feature was significantly reduced left hippocampus-left temporal pole FC in SAD, which increased concomitant with improvement following 8 weeks of SSRI treatment. Our finding of left hippocampus-left temporal pole FC during face perception in healthy subjects is consistent with the observation of intrinsic FC between anterior hippocampus and anterior temporal pole in humans and non-human primates (Kahn et al, 2008), and with increased FC between hippocampus and left temporal pole during successful retrieval of memory for face-name associations (Tsukiura et al, 2010). Interestingly, we observed that this FC is reduced in subjects with SAD, particularly during neutral (primary sample) and angry face (replication sample) processing (see Supplementary information for further discussion regarding differences between these samples).
Previous findings indicate that the temporal pole has a role in both social and emotional processes including face recognition and theory of mind (Wong and Gallate, 2012), and memory for face-name pairs (Damasio et al, 1996), and it has been proposed that the temporal pole binds complex, highly processed perceptual inputs to visceral emotional responses (Olson et al, 2007). It is also thought to be involved in access to social knowledge during mentalizing, the implicit attribution of intentions and other mental states (Frith and Frith, 2003). The left hippocampus is a key region for memory (ie, autobiographical memory retrieval) (Spreng and Mar, 2010), and FC between this region and the temporal pole may reflect an integration of stored memory with social knowledge during face perception and mentalizing in healthy subjects. This neural process (left hippocampus-left temporal pole FC) is presumably disrupted or under-utilized in SAD, consistent with observations of deficits in memory and recognition of facial identity in social anxiety (Davis et al, 2011; Pérez-López and Woody, 2001), and excessive self-focused attention and fears of negative evaluation in interpersonal situations (Clark and Wells, 1995). The results suggest that this neural disruption is prominent when processing faces that carry signals of negative evaluation, such as angry faces (Stein et al, 2002; Staugaard, 2010), or neutral faces, which persons with SAD are more likely to respond to as threatening (Yoon and Zinbarg, 2008b). The strongest effects for left hippocampus-left temporal pole FC in the replication sample (case vs control differences and longitudinal changes) were observed during the angry condition, with a trend-significant effect for the neutral face condition. For the primary sample there was no angry face condition, so the FC to angry faces is unknown, but it is possible that stronger effects (relative to neutral face condition) might have been observed in the primary sample as well, had the task included an angry face condition. The apparent discrepancy between samples in the significance level of findings for neutral faces may be due to the limitation of small samples yielding marginal power for these analyses, or could be due to the difference in tasks that required attention focused on gender (replication sample task) or on color (primary sample task). Future studies that include both emotional expressions and types of tasks would be required to clarify this issue.
In the absence of a placebo or comparison therapy group, we cannot infer to what extent changes in left hippocampus-left temporal pole FC were specific to SSRI treatment or to clinical improvement. In a recent PET study in which SAD subjects responded to either placebo or SSRI treatment, reduction in (amygdala) brain activity was similar in both groups (Faria et al, 2012). Hence, we cannot rule out that left hippocampus-left temporal pole FC would increase in response to any treatment. From the current longitudinal analysis based on only two time points, we also cannot infer whether the changes in left hippocampus-left temporal pole FC preceded, or instead followed, changes in symptom severity. A future study could be designed to employ mediation analysis to more explicitly test whether changes in left hippocampus-left temporal pole FC mediate changes in symptom severity, or vice versa.
Although baseline (pre-treatment), left hippocampus-left temporal pole FC did not predict outcome in response to 8 weeks of SSRI treatment, power to detect predictors of outcome was limited by limited heterogeneity in outcome, as most of the SAD patients improved during treatment. FC (as well as brain activity features identified in related studies, Ding et al, 2011; Doehrmann et al, 2012), may be a useful pre-biomarker to refine the diagnostic classification of psychiatric disorders and advance the development of personalized treatment approaches. For example, it is possible that left hippocampus-left temporal pole FC could be modulated by particular conditions (eg, social threat stimuli under various cognitive reappraisal strategies) within a single scan session, and the extent of this modulation may then be predictive of treatment outcome. Alternatively FC features might be used to identify targets for direct modification by techniques such as transcranial magnetic stimulation. If the FCs identified in this study are further validated by independent replications, then future studies could examine the clinical features of individuals, regardless of diagnosis, that exhibit these features, who might thus also benefit from treatments designed to modulate this circuitry.
Limitations of this study include that the primary data set was not gender matched, and there were differences in ascertainment, diagnostic assessments, and paradigm (ie, color vs gender identification) with the replication sample. However, the fact that replication was significant despite these differences suggest robustness of our approach, and encourage further refinement of the approach and replication in larger samples. The left hippocampus-left temporal pole FC identified here appears to be symptom severity dependent, and as such may not useful as a marker for early detection or a test of risk in those who are pre-symptomatic. However, the current results suggest that future SAD prodrome longitudinal studies should investigate the utility of pattern analysis of large-scale FC during face processing as an approach to identify putative risk biomarkers. Future studies should also include a larger, longitudinal control group to more accurately assess the reliability of condition-dependent FC measures across multiple time points.
Our primary and replication samples contained similar distributions of race/ethnicity in SAD and HC groups: (African American/White/Hispanic/Asian/Unspecified: primary HCs 3/9/1/5/1, primary SAD 2/10/3/0/1, replication HCs 0/14/1/2, replication SAD 0/10/3/1/0), while the viewed stimuli were all Caucasian (Ekman) faces. Previous studies have demonstrated differences in amygdala responses to Black vs White faces during indirect race evaluation (Phelps et al, 2000), while priming race in biracial observers affects visual search for Black and White faces (Chiao et al, 2006). Unless future studies with larger, multiracial samples confirm that these findings are robust across gender and race/ethnicity, race/ethnicity of the facial stimuli could pose an obstacle in eventually translating the current findings and methodology toward clinical use. In this case, use of schematic face stimuli could be explored to preclude any confounds due to racial bias (Evans et al, 2008).
Further improvements in the spatial and temporal resolution of fMRI data acquisition should improve the sensitivity/specificity of whole-brain functional-connectivity measures. The use of higher field strength (3T) combined with instrumental choices and technical improvements has recently allowed for whole-brain coverage with 2 mm isotropic resolution with TR=0.7 s for fMRI (Ugurbil et al, 2013). Furthermore, since structural parcellation introduces arbitrary boundaries between regions (Smith et al, 2011), future methodological improvements could include functional parcellation so that the signal from each node region is adequately captured (Ugurbil et al, 2013). Future studies might also utilize a variety of imaging paradigms that activate different regions, such as speech anticipation, eye gaze (Schneier et al, 2009), and other structural imaging modalities (Liao et al, 2011; Talati et al, 2013), as the best discrimination may ultimately result from combining several paradigms and imaging modalities that tap into various neural facets of the disorder.
FUNDING AND DISCLOSURE
In the past 2 years, FRS has received an honorarium from GlaxoSmithKline for speaking at two conferences. Neither present a conflict of interest with the present study.
SPP had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. We wish to thank Stephen Dashnaw and Andrew Kogan for technical assistance with image acquisition; Aviva Olsavsky and Melissa Sy for assistance with subject recruitment and project management; Xian Zhang, Paul Pavlidis, and Tor Wager for helpful discussion and guidance. Nico Dosenbach provided scripts which aided in the 3D network visualizations using the Caret software. This work was supported by an NIMH predoctoral fellowship (NRSA) F31MH088104-02 (SPP/mentor JH), a K01 DA029598 (National Institute of Drug Abuse) and NARSAD Young Investigator Award (AT), a US Army TARDEC W56HZV-04-P-L (JH) and the Sycamore Fund (FRS). Subject scanning and recruitment was also funded by the following grants: PO1MH60970 (PI: MM Weissman) and R21 MH077976 (PI: FRS).
About this article
Conceived, designed and conducted the experiments: SPP AT. Analyzed the data: SPP. Contributed reagents/materials/analysis tools: JH FRS. Drafted the paper: SPP. Supervised the study: JH FRS. All authors read and approved the final paper. Research was conducted in partial fulfillment of the requirements for a PhD degree in Physiology and Cellular Biophysics at Columbia University (SPP/JH mentor).
Supplementary Information accompanies the paper on the Neuropsychopharmacology website (http://www.nature.com/npp)