Download Supplementary Information (doc 1146K)

1
Pantazatos, et. al. 2013
Supplementary Material for
Reduced anterior temporal and hippocampal functional connectivity during face processing
discriminates individuals with social anxiety disorder from healthy controls and
panic disorder, and increases following treatment.
Authors: Spiro P. Pantazatos1,2,*,+, Ardesheer Talati3,7+, Franklin R. Schneier3,8, Joy Hirsch1,4,5,6,*
1
fMRI Research Center, Depts of 2Physiology and Cellular Biophysics, 3Psychiatry,
4
Neuroscience, 5Radiology, 6Psychology, Columbia University, New York, NY, USA; Divisions of
7
Epidemiology and 8Clinical Therapeutics, New York State Psychiatric Institute, New York, NY
* To whom correspondence should be addressed:
E-mail: [email protected], [email protected]
This file includes
Supplemental Methods
Supplemental Results
Supplemental Discussion
Tables S1, S2, S3, S4
Supplemental Figures S1, S2, S3, S4, S5, S6
Supplemental References
2
Pantazatos, et. al. 2013
Supplemental Methods
Subjects
Primary Sample
Potential subjects were first screened by a research assistant with the anxiety screening
modules of the Schedule for Affective Disorders and Schizophrenia-Lifetime Version, Modified
for Anxiety Disorders and Updated for DSM-IV (SADS-LA-IV (Mannuzza et al. 1986)); subjects
who screened positive for SAD then participated in a full SADS-LA-IV interview (see below).
Subjects with SAD were required to have a DSM-IV (Psychiatric Association 1994) diagnosis of
the generalized subtype of social anxiety disorder (GSAD, characterized by fear of most social
situations). Subjects in the comparison group were required to have a DSM-IV diagnosis of PD,
either with or without agoraphobia. Subjects in both groups were required to have first onset by
age 30, and have a first-degree relative with an anxiety disorder. HCs were required to have no
lifetime history of any psychiatric disorder, with exceptions for past minor depressive disorder,
adjustment disorders, or brief periods of substance abuse (not dependence) in adolescence or
college. HCs also could not have a history of an anxiety disorder in any first-degree relative.
Neither group could have a personal or family history of schizophrenia or bipolar disorder. All
subjects were free of psychotropic medications for 10 weeks preceding the scan.
Diagnostic assessments were administered by clinically trained mental health
professionals with the SADS-LA-IV(Mannuzza et al. 1986). Training and monitoring procedures
have been previously described (Talati et al. 2008). Family history was obtained with the Family
History Screen (Weissman et al. 2000). Final diagnoses were made by an experienced clinician
with the Best Estimate Procedure (Leckman et al. 1982).
Replication Sample
Exclusion criteria for GSAD participants included having a current Axis I disorder (other
than secondary diagnoses of generalized anxiety disorder, dysthymia, or specific phobia), major
3
Pantazatos, et. al. 2013
depressive episode in the past year, substance abuse in the past 6 months, and clinically
significant general medical conditions. HCs did not meet criteria for any lifetime Axis I disorder.
Health status was confirmed by a physical examination including drug toxicology screen. All
subjects were free of psychotropic medications for at least 4 weeks prior to study entry.
Data from four GSAD patients were excluded from analyses (one subsequently revealed
a recent history of major depression, one failed to follow imaging task instructions, and the
functional scans of the others suffered from technical issues), yielding 14 GSAD patients.
Secondary comorbid diagnoses in participants with GSAD consisted of current generalized
anxiety disorder (N=3), past major depression (N=6), and past alcohol abuse (N=1). Six GSAD
subjects had taken medication for anxiety or depression prior to the past 4 weeks. All subjects in
both samples provided written informed consent after discussion of study procedures.
Behavioral task
Primary Sample
Subjects performed a previously described task (Etkin et al. 2004; Pantazatos et al.
2012b) which consists of color identification of fearful, neutral, masked fearful and mask neutral
faces (F, N, MF and MN respectively) with in a blocked paradigm (four 20 second blocks for
each condition, 15 second baseline between each block). Stimuli: Black and white pictures of
male and female faces showing fearful and neutral facial expressions were chosen from a
standardized series developed by Ekman and Friesen (Pantazatos et al. 2012a). Faces were
cropped into an elliptical shape that eliminated background, hair, and jewelry cues and were
oriented to maximize inter-stimulus alignment of eyes and mouths. Faces were then artificially
colorized (red, yellow, or blue) and equalized for luminosity. For the training task, only neutral
expression faces were used from an unrelated set available in the lab. These faces were also
cropped and colorized as above.
4
Pantazatos, et. al. 2013
Each stimulus presentation involves a rapid (200 ms) fixation to cue subjects to fixate at
the center of the screen, followed by a 400 ms blank screen and 200 ms of face presentation.
Subjects have 1200 ms to respond with a key press indicating the color of the face. Behavioral
responses and reaction times were recorded. Unmasked stimuli consist of 200 ms of a fearful
or neutral expression face, while backwardly masked stimuli consist of 33 ms of a fearful or
neutral face, followed by 167 ms of a neutral face mask belonging to a different individual, but of
the same color and gender. Each epoch consists of ten trials of the same stimulus type, but
randomized with respect to gender and color. The functional run has 16 epochs (four for each
stimulus type) that are randomized for stimulus type. To avoid stimulus order effects, we used
two different counterbalanced run orders. Stimuli were presented using Presentation software
(Neurobehavioral Systems, http://nbs.neuro-bs.com), and were triggered by the first radio
frequency pulse for the functional run. The stimuli were displayed on VisuaStim XGA LCD
screen goggles (Resonance Technology, Northridge, CA). The screen resolution was 800X600,
with a refresh rate of 60 Hz. Prior to the functional run, subjects were trained in the color
identification task using unrelated neutral face stimuli that were cropped, colorized, and
presented in the same manner as the nonmasked neutral faces described above in order to
avoid any learning effects during the functional run. After the functional run, subjects were
shown all of the stimuli again, alerted to the presence of fearful faces, and asked to indicate
whether they had seen fearful faces on masked epochs.
Replication Sample
Stimuli consisted of faces of both genders expressing neutral, high valence angry or
happy expressions from the same standard series as above 26, during explicit and unattended
viewing conditions. During the explicit processing condition, subjects were asked to judge the
emotional facial expression (angry, neutral, happy) by using a keypad, and reaction times were
recorded. During the unattended processing condition, subjects were asked to identify gender of
5
Pantazatos, et. al. 2013
each face (male/female), responding via keypad. The stimuli were presented in a block design
consisting of two 6 min and 48 sec. runs (one run unattended, one run explicit) each containing
4 blocks of angry (A), neutral (N) and happy (H) faces. Each block lasted 20 seconds, followed
by 12-14 seconds of baseline (white crosshair against black backgroun d). Within each block,
10 stimuli (faces) were presented for 1 second, followed by 1 second crosshair between each
stimulus presentation. At the start of each run, an instruction screen was presented for 10
seconds, with instructions for using the keypad. Subjects had been trained prior to the scanning
session in the use of the keypad. Given that our primary sample performed an unattended face
processing task (i.e. identification of colors overlaid on emotional faces), we conducted the
replication analysis using the unattended condition from the replication sample. We note our
replication sample was not prospectively designed to be a replication of our experimental
paradigm, but rather it was an independent cohort of SAD and controls from a separate study
(PI: Schneier) which performed a similar, though not identical, implicit face processing task.
Due to a minor programming error, during the unattended runs, 11 baseline (pretreatment) subjects (6 controls, 5 cases) received a distribution (in no particular order) of 5/4/3
blocks of each condition, with 5 blocks tending to occur slightly more often for the A condition,
and 3 blocks slightly more often for N (over all subjects, mean #blocks per condition: A-4.22, H3.91, N-3.88). Five (1 control, 4 cases) post-treatment runs were similarly affected (over all
subjects, mean # blocks per condition: A-3.95, H-4.11, N-3.95).
Node definition and functional connectivity estimation:
Brain regions were parcellated according to bilateral versions of the Harvard-Oxford
Cortical and sub-cortical atlases and the AAL atlas (cerebellum) and were trimmed to ensure no
overlap with each other and to ensure inclusion of only voxels shared by all subjects
(Supplemental Figure 1A). For each subject, time-series across the whole run (283 TRs) were
extracted using Singular Value Decomposition (SVD) and custom modifications to the Volumes-
6
Pantazatos, et. al. 2013
of-Interest (VOI) code within SPM8 to retain the top 2 eigenvariates from each atlas-based
region. Briefly, the data matrix for each atlas-based region is defined as A, an n x p matrix, in
which the n rows represent the time points, and each p column represents a voxel within an
atlas-based region. The SVD theorem states:
Anxp= Unxn Snxp VTpxp,
where UTU = Inxn and VTV = Ipxp (i.e. U and V are orthogonal). The columns of U are the left
singular vectors (eigenvariates, or summary time courses of the region), S (the same
dimensions as A) has singular values, arranged in descending order, that are proportional to
total variance of data matrix explained by its corresponding eigenvariate, and is diagonal, and
VT has rows that are the right singular vectors (spatial eigenmaps, representing the loading of
each voxel onto its corresponding eigenvariate). Here we retain the top two eigenvariates
(nodes) from each region.
The above step resulted in a total of 248 nodes with an associated time course (i.e.
eigenvariates) and spatial eigenmaps from the 124 initial atlas-based regions. Thus, each atlasbased region was comprised of two nodes. We note that this means it is possible that node 2 of
a particular region shows functional connectivity that differentiates SAD diagnosis and node 1 of
the same region has no differential connectivity. For clarity we therefore label each node using
its Harvard-Oxford atlas label appended by either “_PC1” for the first eigenvariate and “_PC2”
for the second. For display purposes, we calculated the MNI coordinates of the peak loading
weight (locations averaged across subjects) for each eigenvariate from its associated eigenmap
(Supplementary Figure 1B). Table S1 lists the average MNI coordinates for each node.
For each subject, functional connectivity matrices (i.e. where cell i,j contains the Pearson
correlation between region i and region j) were generated for non-masked (unattended) and
masked (subliminal) fearful (F, MF) and neutral (N, MN) conditions (primary sample), and for
unattended angry (A), happy (H) and neutral (N) conditions (replication sample), as well as over
7
Pantazatos, et. al. 2013
the full run (henceforth denoted as “Full”). The above time-series were segmented and
concatenated according to conditions of interest, incorporating a lag of 6 or 7 s from the start of
each block, before generating the correlation matrices. The lower diagonal of the above
preprocessed correlation matrices were then used as input features to predict subjects’
diagnoses.
SVM pattern analysis
Support vector machines (SVM) are pattern recognition methods that find functions of the
data that facilitate classification During the training phase, an SVM finds the hyperplane that
separates the examples in the input space according to a class label. The SVM classifier is
trained by providing examples of the form <x,c>, where x represents a spatial pattern and c is
the class label. In particular, x represents the fMRI data (pattern of correlation strengths) and c
is the condition or group label (i.e. c = 1 for SAD and c = −1 for control). Once the decision
function is determined from the training data, it can be used to predict the class label of new test
examples.
Our intent is not to estimate or maximize the true accuracy of prediction given a
completely new data set, but rather to test whether there exist information in the pattern of
functional connections relevant and specific to SAD, and to approximate the optimal number of
features that containing this information. We note that our approach in the primary sample
(plotting AUC vs. number of top N features) is not biased, since for each number of top N
features, and for each round of leave-one-out cross validation, the top N features were selected
from a training set that was completely independent from the testing set. If there is a true signal
present in the data, we expect, and in the current data observe, an initial rise in accuracy as
more informative features are added to the feature set, and a dip in accuracy as less informative
features (i.e. noise) are added to the feature set.
8
Pantazatos, et. al. 2013
For all binary classification tasks, a linear kernel SVM (Fu et al. 2008) with a filter feature
selection (t-test) and leave-one-out cross validation was used. During each iteration of leaveone-out cross validation (primary sample), one subject was withheld from the dataset 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 2 features identified in the analysis above, and this same model was used to predict SAD
vs. controls in the replication sample. Prior to 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 is reported with “area under the curve” (AUC) (i.e. area under
the receiver operator characteristic, or ROC curve (Hanley & McNeil 1982)). Since the predicted
labels are binary (not continuous scores), AUC is identical to the arithmetic mean of sensitivity
(true positive rate, or proportion of correctly classified cases) and specificity (true negative rate,
or proportion of correctly classified controls). AUC vs. number of features that have been ranked
by their t-score was plotted, and the performance significance was computed using nonparametric permutation tests (Golland & Fischl 2003) with 200 (1 to 40 features) or 10,000
(single reported peaks) iterations. The filter feature selection ranked features according to
absolute t-score, which identifies the strongest differing features but discards directional
information. Classification accuracy vs. every 5 features from the top 1 through 200 was first
examined (the maximum number was chosen heuristically based on (Dosenbach et al. 2010).
Other than a peak near 1 features, accuracies hovered near 50%. Therefore the range was
changed to every single feature from top 1 through 40, the same range which was recently used
in decoding supraliminal fear (Pantazatos et al. 2012b). Bonferroni correction was also applied
for the number of total Top N comparisons (in this case 40). Confidence intervals for AUC
estimate in the replication sample was also estimated using the 'bootstrap t' approach
9
Pantazatos, et. al. 2013
(Obuchowski & Lieber 1998) with 10,000 iterations using the Measures of Effect Size Toolbox
(http://www.mathworks.com/matlabcentral/fileexchange/32398-measures-of-effect-size-toolbox).
SVM learning and classification was done using the Spider v1.71 Matlab toolbox
(http://people.kyb.tuebingen.mpg.de/spider/) using all default parameters (i.e. linear kernel SVM,
regularization parameter C=1). Graphical neuro-anatomical connectivity maps of the top N
features were displayed using Caret v5.61 software
(http://brainvis.wustl.edu/wiki/index.php/Caret:About). We note that different features could have
been selected during the feature selection phase of each round of cross-validation. Therefore in
ranking the top N features, features were ranked by total number of times that feature was
included in each round of cross-validation, and then among these features, features were sorted
by absolute value of the average SVM weight.
PsychoPhysiological Interaction (PPI) analysis (see Supplemental Figure 4) was conducted
following a generalized PPI approach (McLaren et al. 2012). The PPI analysis measures the
extent to which regions are differentially correlated during a given task or between subjects. We
used ROI-based "seeds" for Left Hippocampus and Left Temporal Pole from the same atlas
used in the main text, and extracted the 1st eigenvariate from each region to be used as the
region's summary time course. The BOLD signal throughout the whole-brain was then
regressed on a voxel-wise basis against the product of this time course and the vectors of the
psychological variable of interest, (1*F + 1*N + 1*MF +1*MN), with the physiological and the
psychological variables serving as regressors of no interest (F, N, MF and MN pyschological
regressors were identical as in the main GLM analysis). Additional nuisance regressors included
9 motion parameters, and mean white and csf signal. Resulting beta maps for the N PPI
condition were subsequently passed to 2nd level random effects analysis (one way ANOVA with
3 levels: controls, SAD and PD, with age and gender as additional nuisance covariates).
Contrasts for control > SAD were then computed from this 2nd level model.
10
Pantazatos, et. al. 2013
Supplemental Results
Provisionary results included 7 HCs completing pre-post scans 8 weeks apart.
Additional analyses that incorporated 7 HCs confirmed that pre- to post-treatment FC
increases in the SAD group were significantly greater than FC changes over 8 weeks in the
control group (cases (post-pre) > controls (post-pre), t(17)=2.1, p=0.05, 2-sample t-test), and
that FC did not change significantly in the control group (t(6)=1.56, p=0.17, one sample t-test).
To further examine whether Left Hippocampus-Left Temporal Pole FC tracks social anxiety
symptom severity longitudinally, regardless of diagnostic or treatment status, the extent to which
changes in LSAS were associated with changes in this FC was tested. This analysis included
controls, because we were primarily interested in longitudinal symptom change that is not
necessarily specific to treatment, and in order to further test that small changes in LSAS in
controls are accompanied by small changes in Left Hippocampus-Left Temporal Pole FC. Given
that SAD subjects exhibited decreased FC relative to HCs at baseline, we hypothesized that,
across both HCs and SAD subjects, increases in Left Hippocampus-Left Temporal Pole FC
should be associated with decreases in symptom severity. This relationship was indeed
observed for Left Hippocampus-Left Temporal Pole FC, with significant correlations observed
for FC during angry and happy faces (pre-post ΔLSAS vs. post-pre ΔFC: angry R=0.55,
p=0.008, happy R=0.58, p=0.004, neutral R=0.33 p=0.08, and full run R=0.37 p=0.06)
(Supplementary Figure 6). These results held, particularly for FC during angry, happy and
neutral faces, after removal of the top 1 and top 2 outliers (indicated as boxed 1s and 2s in
Supplementary Figure 6) from each plot, (pre-post ΔLSAS vs. post-pre ΔFC top 1 removed:
angry R=0.59, p=0.01, happy R=0.63, p=0.005, neutral R=0.35 p=0.15, full run R=0.39 p=0.10;
top 2 removed: angry R=0.57, p=0.017, happy R=0.62, p=0.007, neutral R=0.51 p=0.018, and
full run R=0.36 p=0.16).
11
Pantazatos, et. al. 2013
Changes in left hippocampus and left temporal pole activation following 8-weeks SSRI
treatment.
Pre-post activation differences in response to neutral and angry faces (vs. baseline)
were also assessed in the left hippocampus and left temporal pole. Using left hippocampus as
an ROI and thresholding at p<0.05 uncorrected, k=20, post > pre differences were observed at
MNI [-14 -10 -22], t=2.14, k=21, while pre > post differences were observed at MNI [-32, -22, 14], t=2.43, k=40 for neutral faces, and MN [-36 -32 -10, t=3.53, k=54 and [-20 -10 -22], t=2.83,
k=77 for angry faces. However no clusters survived cluster-extent correction at p<0.05 (k
threshold=97). Using left temporal pole as an ROI, pre > post differences were observed at
MNI=[-54, 6 -18],t=4.5, k=111 at p<0.05 uncorrected for neutral faces, which survived clusterextent correction at p<0.05 (k threshold=88), while post > pre cluster was observed for angry
faces at MNI=[-30 14 -28], t=2.88, k=23, which was not significant. Overall these results suggest
decreases in activation in left hippocampus and left temporal pole in response to angry and
neutral faces following 8-weeks SSRI treatment.
For angry and neutral faces, neither pre > post nor post > pre differences were observed
in right anterior middle temporal gyrus at p<0.05 uncorrected, k=20, while left OFC, post>pre
differences were observed for angry faces in [-30 12 -18], t=2.71, k=69 and pre>post differences
for neutral faces in [-24 8 -18], t=2.64, k=32, but neither cluster reached significance at p<0.05
after cluster-extent thresholding (k threshold =103). In addition, pre-post change in Left
Hippocampus and Left Temporal Pole grey matter volume were correlated with change in FC
between these regions. Positive, albeit non-significant, associations were observed with change
in Left Hippocampus volume with change in FC (r=0.39, p=0.11), change in Left Temporal Pole
volume with change in FC (r=0.24, p=0.34), and change in Left Hippocampus volume with
change in left Temporal pole volume (r=0.18, p=0.48).
12
Pantazatos, et. al. 2013
Examining previous SAD-related and emotion-related FC reported in the literature
In addition to the exploratory, data-driven approach above, we examined FC previously
identified to be anomalous in SAD, in particular reduced aINS-dACC (31) and amygdala-dACC
and amygdala-dlPFC (32) in SAD during fear. Using PPI analysis, a recent study observed less
aINS-dACC FC during fearful (> happy) in gSAD relative to controls (Klumpp et al. 2012). All FC
during both F and N conditions between bilateral Insula and Anterior Cingulate Gyrus was
queried at p < 0.05 uncorrected, and the following was observed: Control > SAD,
Left_Insular_Cortex_PC2-Left_Cingulate_Gyrus_anterior_division_PC1 t(33)=2.22/2.96 F/N,
and Right_Insular_Cortex_PC2-Left_Cingulate_Gyrus_anterior_division_PC1 t(33)=1.82/Not
significant F/N. The average peak location for Left Insula was anterior ([-36 16 2]), while peak
MNI location for the right was middle insula ([42 -4 6]). These results are consistent with the
aforementioned study.
A related study (Prater et al. 2012) used PPI and observed less connectivity between
amygdala-dACC and amygdala-dlPFC in SAD during fearful faces perception. As above, we
interrogated FC between these regions during F and N conditions at p<0.05 uncorrected,
Control > SAD and observed: Right_Amygdala_PC2-Right_Cingulate_Gyrus_anterior_division,
t(33)=2.53/Not significant F/N and Right_Ventral_Frontal_Pole_PC1 - Right_Amygdala_PC1 =
t(33) = 1.9208/Not significant F/N consistent with (Prater et al. 2012). However, many FC
differences between amygdala and dlPFC/precentral gyrus were in the opposite direction (i.e.
greater in SAD): Control > SAD, Right_Ventral_Frontal_Pole_PC1 - Right_Amygdala_PC2 = 1.94 Right_Ventral_Frontal_Pole_PC1 - Left_Amygdala_PC2 = -1.67, Left_Amygdala_PC1Left_Middle_Frontal_Gyrus_PC1, t(33)=-3.08, Left_Precentral_Gyrus_PC1Left_Amygdala_PC1, t=-2.63, Right, t=-1.72, Right_Ventral_Frontal_Pole_PC2 Left_Amygdala_PC1 , t=-2.5082.
Although FC differences were mostly consistent with these studies, including the above
connections (RAmygdala-RACC, Right_Ventral_Frontal_Pole_PC1-Right_Amygdala_PC1, Left
13
Pantazatos, et. al. 2013
Insula-dACC) with the top 2 connections identified in the main text did not improve classification
performance (data not shown), while including only these connections resulted in poorer
classification performance (AUC=0.53). It is important to note that FC was measure here using
Pearson correlation, while these previous studies applied "seed" based regression analyses,
which are different approaches for measure functional connectivity and their differences. See
(Kim & Horwitz 2008) for further discussion.
An additional analysis was conducted whereby the top 25 FC that discriminated
supraliminal fearful from neutral faces (Pantazatos et al. 2012b) and top 25 FC that
discriminated subliminal fearful from neutral faces (Pantazatos et al. 2012a) in a healthy sample
were used as initial input feature set for predicting SAD vs. controls. Classifications were
performed as in the main text from the top 1 to top 25 ranked features. For supraliminal
conditions, the following peak SAD vs. Control AUCs were achieved: F: 0.62, N: 0.68, F-N: 0.74.
For subliminal conditions, the following peak SAD vs. Control AUCs were achieved: MF: 0.69,
MN: 0.64, MF-MN: 0.82. The relatively high discrimination achieved when using the differences
between supraliminal and subliminal fearful and neutral faces (i.e. F-N and MF-MN) suggests
that SAD subjects do not process fearful vs. neutral faces in the same way as healthy subjects,
and instead use a different neural circuitry. However, it is important to note that this analysis is
biased due to the fact that the primary control sample constituted half of the healthy subjects
used in the above studies. Future studies using an additional independent control sample would
be necessary to acquire unbiased results.
Discriminating between SAD and healthy control subjects with patterns of spatial activity
To compare the information content of patterns of interactivity (i.e. functional
connections used above) vs. patterns of activity, SAD vs. Control classification was also
conducted using beta estimates, which are considered summary measures of activation in
response to each condition. This approach is conceptually similar to a recent study that used
14
Pantazatos, et. al. 2013
pattern classification of whole-brain activity (BOLD averaged over several TR’s of an event
minus baseline activity immediately preceding the event) during sad face viewing to predict
diagnosis (normal vs. clinically depressed)(Fu et al. 2008). In order to make featureselection/leave-one-out cross validation and SVM learning more computationally tractable,
preprocessed functional data were resized from 2x2x2 mm voxel resolution to 4x4x4 mm
resolution, and subject-specific GLM models were re-estimated, resulting in a reduction of total
feature space per example from ~189,500 betas to ~23,500. Feature selection, leave-one-out
cross validation and SVM learning proceeded exactly as above for FC data. When using the
contrast F-N, we observed a peak AUC of 0.88 (p<0.0001 uncorrected) with 8 voxels (within
cerebellum and middle occipital gyrus), and when using the F beta weights, a peak AUC=0.83,
p=0.0008 uncorrected was observed with ~170 voxels (Supplementary Figure S3,
Supplementary Table S2). However the AUC using F betas dropped to 0.49, and AUC using FN contrast dropped to 0.58 after regressing out the effects of age and sex prior to classification.
Classification of SAD vs. PD using the same features as above was then attempted, and
a decrease in classification performance was observed; when using F>N contrasts, AUC=0.59,
p=0.03 uncorrected, and when using F beta weights, AUC=0.56 p=0.26 uncorrected, data not
shown). Classification of SAD vs. PD using top 10:10:500 F-N contrast estimates over the
whole-brain only achieved a peak AUC of 0.66 (data not shown). Thus, although peak
classification performance for SAD vs. Controls using contrast estimates as features matched
that of using pair-wise functional connectivity, under the current analysis these activation
differences appear to be less specific to SAD.
Standard GLM Activation analysis
A standard GLM analysis was also run to identify SAD vs. Control differences in overall
activation and differential activation to F, N, MF or MN conditions in the primary sample. For a 2way ANOVA 2nd level analysis was conducted (1st independent factor diagnosis: 2 levels, 2nd
15
Pantazatos, et. al. 2013
dependent factor face conditions: 4 levels-F, N, MF, MN). Omnibus F-tests were used to identify
main effects of diagnosis and diagnosis X face condition interactions. Whole-brain results were
thresholded at p<0.05 corrected using cluster-extent threshold using the 3dClusterSim program
in AFNI (v.2011). Briefly, 1000 Monte Carlo simulations of whole-brain fMRI data were
generated using the applied smoothing (8 mm fwhm), voxel size (2x2x2 mm) and whole-brain
mask to determine the cluster size at which the false positive probability was below a desired
alpha level of p < 0.05 corrected. For an uncorrected p-value threshold of 0.005, this yielded a
cluster size required of 148.
Greater activation to faces in the SAD group were observed in the fusiform, consistent
with higher salience of faces in SAD, while greater activation to faces in the control group were
observed in dorsolateral prefrontal cortex, consistent with reduced cognitive control during face
processing in SAD (see Supplementary Table S4 and Supplemental Figure S5). These clusters
were used to define ROIs to test for main effects of diagnosis in the replication group (in this
case overall activation to A, H and N faces). Within these ROIs, no voxels survived a lenient
threshold of p<0.05 uncorrected.
In addition, the above ROIs were used as a mask to preselect voxels to be used in an
additional, biased SVM analysis that used averages of these beta estimates across each face
condition as well as beta estimates from each face condition as features. Performance was
assessed across the range of top 10 through 650 (every 10) selected voxels. Despite biased
feature selection, peak AUCs were still lower than those achieved using large-scale functional
connectivity as features and unbiased feature selection (beta estimates averaged across all
face conditions: 0.71, F: 0.67, N: 0.74, MF: 0.71, MN: 0.71, data not shown).
Grey matter volume and pre-post activation analyses
Voxel-based morphometry analyses (Ashburner & Friston 2000) were used to correlate
pre-post changes in FC with pre-post changes in local grey matter (GM) volume. Intra-subject
16
Pantazatos, et. al. 2013
realignment, bias correction, segmentation, and DARTEL normalization, and modulation of
anatomical data were conducted using batch processing for longitudinal data within VBM8
toolbox (no smoothing was applied). Processed post-images were subtracted from processed
pre-images to create a pre-post subtraction image for each subject. Briefly, the coordinate of the
peak loading factor from the first PC spatial eigenmap within Left Hippocampus and Left
Temporal Pole was identified for each subject, and the estimates of local GM volume pre-post
value was extracted from a 6 mm radius sphere about the coordinate. These values were then
correlated with pre-post changes in Left-Hippocampus-Left Temporal Pole FC. These analyses
included the same subjects as in the analysis of pre-post changes in FC except for one case
subject whose anatomical data was discarded due to technical issues with the scan (total 18
subjects: 11 cases, 7 controls).
See methods, main text for description of GLM analysis. For pre-post activation
analyses, cluster-extent correction was applied using the 3dClusterSim program in AFNI
(v.2011). Briefly, 1000 Monte Carlo simulations of whole-brain fMRI data were generated using
the applied smoothing (8 mm fwhm), voxel size (2x2x2 mm) and ROI masks (Left Hippocampus
and Left Temporal Pole) to determine the cluster size at which the false positive probability was
below a desired alpha level of p < 0.05 (i.e. an effective threshold of p < 0.05 corrected for
multiple comparisons).
Supplemental Discussion
A study using effective connectivity (Granger causality analysis) of a few a priori
selected regions (amygdala, visual and association cortex) suggested increased connectivity
between amygdala and visual cortex, and decreased connectivity with OFC during resting-state
in SAD (Liao et al. 2010), while atlas-based functional connectivity analysis of resting-state data
suggested decreased functional connectivity between frontal and occipital lobes in SAD (Ding et
al. 2011). More recent work has focused on condition-dependent functional connectivity (i.e.
17
Pantazatos, et. al. 2013
while viewing fearful faces or during perception of scrutiny) and/or during rest of amygdala,
anterior cingulate, frontal cortex and other regions using PPI analysis, which assess FC
between an a priori selected seed region and the rest of the brain (Prater et al. 2012; Klumpp et
al. 2012; Giménez et al. 2012; Pannekoek et al. 2012).
When we focused on specific FC differences of the amygdala and insula, we observed
some consistency with the above studies (Prater et al. 2012; Klumpp et al. 2012) . However,
these were not the strongest observed FC differences observed in the current data. One
possible explanation for this speaks to an advantage of the current approach in that it assesses
the multivariate pattern of the strongest FC differences selected from among many thousands of
estimated pair-wise FCs, and these previous approaches may have missed this FC because
they did not employ hippocampus nor temporal pole as seed regions. However, it is also
important to point out that significance FC differences estimates using correlation varies much
more for different scanning and hemodynamic parameters relative to FC differences estimated
using regression approaches (i.e. PPI) (Kim & Horwitz 2008), and SAD vs. Control comparison
using PPI analysis seeded with either Left Hippocampus or Left Temporal Pole in the current
data produced only moderately significant results (current approach, peak T-value = 5.39; PPI
peak T-value = 3.43, see Supplementary Table S3, Supplementary Figure S4). Finally, we note
that our findings that activity (fearful vs. neutral faces contrast) in middle occipital gyrus
distinguished SAD vs. controls Supplemental Figure 3, is consistent with (Doehrmann et al.
2012), in which greater response to CBT treatment correlated significantly with greater
pretreatment activation (angry vs. neutral faces contrast) in middle occipital gyrus.
In addition, positive, albeit non-significant, correlations between increases in Left
Hippocampus-Left Temporal Pole FC with increases in GM volume in each of these structures
were observed, suggesting increases in grey matter volume may be associated with increased
functional connectivity between these regions. Future studies with larger sample sizes are
needed to confirm whether this is indeed a true association.
18
Pantazatos, et. al. 2013
It may seem counterintuitive that the most predictive FC was during neutral faces in
primary sample. However this is consistent with evidence suggesting that SAD is characterized
by negative interpretation bias, particularly when presented with ambiguous social cues (i.e.
neutral faces) (Winton et al. 1995; Yoon & Zinbarg 2007). Other studies demonstrate abnormal
reactivity to emotional, and in particular harsh (i.e. angry, disgust), faces (Klumpp et al. 2010). In
the current study, case vs. control and pre-post treatment differences in Left Hippocampus-Left
Temporal Pole FC during neutral faces in the replication sample was observed on a trend level
(Table 3, Neutral column: SAD>Control, p=0.09, SAD pre>post p=0.10). However the strongest
effects in this sample were observed for this FC during angry faces (Table 3, Angry column:
SAD > Control, p=0.027, pre>post p=0.007). One possible interpretation is that angry faces
(relative to neutral) are more salient in SAD, and larger differences in Left Hippocampus-Left
Temporal Pole FC might have observed in the primary sample if angry faces had been used.
Alternatively, neutral (relative to angry) faces could be a more salient in SAD, but the signal was
not apparent in the replication sample due to a minor technical issue that caused slightly fewer
blocks of neutral face conditions relative to angry (see Methods: Replication Sample, last
paragraph). Future studies using a balanced block design with both angry and neutral faces can
facilitate a direct comparison that should help resolve this ambiguity.
Interestingly, we observed increased FC between Left Hippocampus-Left Temporal Pole
concomitant with symptom improvement following 8-weeks SSRI treatment, yet there was a
trend-level decrease in activity in each of these structures in response to angry and neutral
faces following treatment (see Supplementary Results). Previous PET and SPECT studies have
also shown reduced perfusion and cerebral blood flow (rCBF) in these regions following 8weeks SSRI treatment. PET imaging during a public speaking paradigm in SAD subjects
demonstrated that regardless of treatment approach (SSRI citalopram or behavioral therapy),
improvement was accompanied by a decreased rCBF-response to public speaking bilaterally in
the amygdala, hippocampus, and the periamygdaloid, rhinal, and parahippocampal cortices
19
Pantazatos, et. al. 2013
(Furmark et al. 2002), while a related SPECT study demonstrated reduced cerebral perfusion in
left hippocampus following 8 or 12 weeks of citalopram in a combined group of SAD, obsessive
compulsive disorder and post-traumatic stress disorder patients (Carey et al. 2004). A related
SPECT study observed reduced perfusion in anterior and lateral temporal cortex in SAD
subjects following 8-weeks citalopram treatment (Van der Linden et al. 2000), while in a recent
fMRI BOLD study, temporal pole activity during successful understanding of others' mental
states correlated with neuroticism (Jimura et al. 2010). Taken together, these results suggest
that while increased activation of hippocampus and temporal pole may be associated with
increased social anxiety symptom severity, increased functional connectivity between these two
structures is associated with decreased symptom severity.
Limitations
The estimate of AUC=0.89 in the primary group is not an estimate of the true accuracy
given new data of our machine learning approach, since we did not know a priori how many top
N features to select. Although significantly above chance, classification performance in the
replication sample for the features selected in the primary sample was lower (primary
sample/replication sample AUC=0.89/0.71). There are several possible reasons for this. Firstly
the task in the replication sample was slightly different (identify gender vs. color), and eleven
subjects received only 3 blocks of neutral face blocks (as opposed to 4) due to a minor
programming error (see methods). Secondly, the peak MNI location for Left Orbitofrontal Cortex
was very inferior, bordering the dorsal tip of the left temporal pole. Since parcellation was done
in 3D, and smoothing was applied, it is possible this signal originated from the left dorsal
temporal pole. Future studies should derive functional connectivity from 2D cortical surface
maps that are then registered across subjects, in order to ensure that regions are more
precisely labeled.
Previous simulations have raised concerns regarding the use of atlas-based approaches
20
Pantazatos, et. al. 2013
for parcellating the brain (Smith et al. 2011). Because the spatial ROIs used to extract average
time-series for a brain region do not likely match well the actual functional boundaries, BOLD
time-series from neighboring nodes are likely mixed with each other. While this hampers the
ability to detect true functional connections between neighboring regions, it has minimal effect
on estimating functional connectivity between distant regions. This perhaps explains why in this
study most of the functional connections that discriminated between SAD and controls were
long-distance. Future experiments using non-atlas based approaches would likely lead to better
estimates of shorter-range functional connections. We also note that the current atlas-based
approach may have under-sampled the prefrontal cortex, and that possible future improvements
could break up the prefrontal regions into smaller pieces in order to sample more nodes from
this area. Finally, we note that using Pearson correlation, it is possible that any association
between two brain regions is the result of a spurious association with a third brain region.
Conclusion
Here we applied a whole-brain, data-driven approach that combines pattern analysis
with atlas-based condition-dependent FC to identify a subset of FC features that can
discriminate individual SAD subjects from both HCs and PD subjects. These features
discriminated SAD subjects in an independent replication sample, which performed a similar
emotional face viewing task, with significantly higher than chance performance. Finally, the most
discriminative feature at baseline normalized following effective SSRI treatment and also
correlated with change in symptom severity. We propose a valuable, exploratory approach to
identify FC-based emerging biomarkers for psychiatric diagnosis and treatment effects.
21
Pantazatos, et. al. 2013
Supplemental Table S1
ROI
X
Y
Z
Name
1
-32
-56
-26
Cerebelum_34567b_L_PC1
2
-14
-38
-22
Cerebelum_34567b_L_PC2
3
32
-56
-24
Cerebelum_34567b_R_PC1
4
12
-42
-18
Cerebelum_34567b_R_PC2
5
-4
-48
-34
Cerebelum_8910_L_PC1
6
-6
-66
-32
Cerebelum_8910_L_PC2
7
6
-48
-36
Cerebelum_8910_R_PC1
8
10
-64
-34
Cerebelum_8910_R_PC2
9
-40
-64
-28
Cerebelum_Crus1_L_PC1
10
-18
-74
-30
Cerebelum_Crus1_L_PC2
11
40
-62
-28
Cerebelum_Crus1_R_PC1
12
28
-74
-34
Cerebelum_Crus1_R_PC2
13
-8
-80
-30
Cerebelum_Crus2_L_PC1
14
-6
-72
-32
Cerebelum_Crus2_L_PC2
15
8
-78
-32
Cerebelum_Crus2_R_PC1
16
6
-70
-32
Cerebelum_Crus2_R_PC2
17
2
0
-12
Hypothalamus_PC1
18
-6
-4
-8
Hypothalamus_PC2
19
-8
12
-6
Left_Accumbens_PC1
20
-10
12
-4
Left_Accumbens_PC2
21
-20
0
-20
Left_Amygdala_PC1
22
-24
-8
-14
Left_Amygdala_PC2
23
-52
-60
26
Left_Angular_Gyrus_PC1
24
-48
-56
44
Left_Angular_Gyrus_PC2
25
-8
10
8
Left_Caudate_PC1
26
-12
4
18
Left_Caudate_PC2
27
-46
-14
10
Left_Central_Opercular_Cortex_PC1
28
-56
-20
16
Left_Central_Opercular_Cortex_PC2
22
Pantazatos, et. al. 2013
29
0
40
6
Left_Cingulate_Gyrus_anterior_division_PC1
30
0
10
34
Left_Cingulate_Gyrus_anterior_division_PC2
31
0
-52
24
Left_Cingulate_Gyrus_posterior_division_PC1
32
0
-28
42
Left_Cingulate_Gyrus_posterior_division_PC2
33
-2
-80
28
Left_Cuneal_Cortex_PC1
34
0
-84
24
Left_Cuneal_Cortex_PC2
35
-4
62
24
Left_Dorsal_Frontal_Pole_PC1
36
-38
46
18
Left_Dorsal_Frontal_Pole_PC2
37
-28
-72
52
Left_Dorsal_Lateral_Occipital_Cortex_superior_division_PC1
38
-34
-72
52
Left_Dorsal_Lateral_Occipital_Cortex_superior_division_PC2
39
-2
50
-10
Left_Frontal_Medial_Cortex_PC1
40
-4
42
-12
Left_Frontal_Medial_Cortex_PC2
41
-42
18
2
Left_Frontal_Operculum_Cortex_PC1
42
-36
18
8
Left_Frontal_Operculum_Cortex_PC2
43
-36
16
-22
Left_Frontal_Orbital_Cortex_PC1
44
-32
16
-16
Left_Frontal_Orbital_Cortex_PC2
45
-46
-16
6
Left_Heschls_Gyrus_H1_and_H2_PC1
46
-44
-22
10
Left_Heschls_Gyrus_H1_and_H2_PC2
47
-16
-16
-22
Left_Hippocampus_PC1
48
-22
-26
-12
Left_Hippocampus_PC2
49
-54
16
14
Left_Inferior_Frontal_Gyrus_pars_opercularis_PC1
50
-54
14
20
Left_Inferior_Frontal_Gyrus_pars_opercularis_PC2
51
-52
30
8
Left_Inferior_Frontal_Gyrus_pars_triangularis_PC1
52
-50
28
18
Left_Inferior_Frontal_Gyrus_pars_triangularis_PC2
53
-50
-40
-18
Left_Inferior_Temporal_Gyrus_posterior_division_PC1
54
-42
-14
-24
Left_Inferior_Temporal_Gyrus_posterior_division_PC2
55
-50
-58
-16
Left_Inferior_Temporal_Gyrus_temporooccipital_part_PC1
56
-50
-50
-18
Left_Inferior_Temporal_Gyrus_temporooccipital_part_PC2
57
-44
0
-6
Left_Insular_Cortex_PC1
58
-36
16
2
Left_Insular_Cortex_PC2
23
Pantazatos, et. al. 2013
59
-2
-72
8
Left_Intracalcarine_Cortex_PC1
60
-6
-82
6
Left_Intracalcarine_Cortex_PC2
61
0
0
54
Left_Juxtapositional_Lobule_Cortex_Supp_Motor_cortex_PC1
62
-4
-6
48
Left_Juxtapositional_Lobule_Cortex_Supp_Motor_cortex_PC2
63
-38
-78
-18
Left_Lateral_Occipital_Cortex_inferior_division_PC1
64
-50
-70
0
Left_Lateral_Occipital_Cortex_inferior_division_PC2
65
0
-76
0
Left_Lingual_Gyrus_PC1
66
-2
-88
-22
Left_Lingual_Gyrus_PC2
67
-48
16
40
Left_Middle_Frontal_Gyrus_PC1
68
-40
34
38
Left_Middle_Frontal_Gyrus_PC2
69
-52
-8
-16
Left_Middle_Temporal_Gyrus_anterior_division_PC1
70
-56
-8
-14
Left_Middle_Temporal_Gyrus_anterior_division_PC2
71
-58
-36
-6
Left_Middle_Temporal_Gyrus_posterior_division_PC1
72
-58
-24
-10
Left_Middle_Temporal_Gyrus_posterior_division_PC2
73
-56
-54
6
Left_Middle_Temporal_Gyrus_temporooccipital_part_PC1
74
-56
-56
-4
Left_Middle_Temporal_Gyrus_temporooccipital_part_PC2
75
-30
-78
-20
Left_Occipital_Fusiform_Gyrus_PC1
76
-24
-74
-14
Left_Occipital_Fusiform_Gyrus_PC2
77
2
-98
4
Left_Occipital_Pole_PC1
78
-28
-98
-2
Left_Occipital_Pole_PC2
79
-20
-6
-2
Left_Pallidum_PC1
80
-24
-12
0
Left_Pallidum_PC2
81
-2
52
2
Left_Paracingulate_Gyrus_PC1
82
0
28
36
Left_Paracingulate_Gyrus_PC2
83
-16
-18
-26
Left_Parahippocampal_Gyrus_anterior_division_PC1
84
-20
-8
-30
Left_Parahippocampal_Gyrus_anterior_division_PC2
85
-16
-30
-22
Left_Parahippocampal_Gyrus_posterior_division_PC1
86
-12
-36
-14
Left_Parahippocampal_Gyrus_posterior_division_PC2
87
-50
-32
16
Left_Parietal_Operculum_Cortex_PC1
88
-50
-34
24
Left_Parietal_Operculum_Cortex_PC2
24
Pantazatos, et. al. 2013
89
-46
-2
-14
Left_Planum_Polare_PC1
90
-48
-4
-2
Left_Planum_Polare_PC2
91
-58
-26
10
Left_Planum_Temporale_PC1
92
-62
-32
16
Left_Planum_Temporale_PC2
93
-40
-26
58
Left_Postcentral_Gyrus_PC1
94
-58
-12
30
Left_Postcentral_Gyrus_PC2
95
-52
4
34
Left_Precentral_Gyrus_PC1
96
0
-24
48
Left_Precentral_Gyrus_PC2
97
0
-58
22
Left_Precuneous_Cortex_PC1
98
0
-76
56
Left_Precuneous_Cortex_PC2
99
-24
6
-2
Left_Putamen_PC1
100
-28
-8
4
Left_Putamen_PC2
101
-2
8
-14
Left_Subcallosal_Cortex_PC1
102
-2
10
-6
Left_Subcallosal_Cortex_PC2
103
-2
54
36
Left_Superior_Frontal_Gyrus_PC1
104
-2
18
54
Left_Superior_Frontal_Gyrus_PC2
105
-42
-44
60
Left_Superior_Parietal_Lobule_PC1
106
-28
-46
62
Left_Superior_Parietal_Lobule_PC2
107
-48
-2
-16
Left_Superior_Temporal_Gyrus_anterior_division_PC1
108
-54
-2
-10
Left_Superior_Temporal_Gyrus_anterior_division_PC2
109
-60
-34
6
Left_Superior_Temporal_Gyrus_posterior_division_PC1
110
-56
-18
-6
Left_Superior_Temporal_Gyrus_posterior_division_PC2
111
0
-78
12
Left_Supracalcarine_Cortex_PC1
112
-12
-66
16
Left_Supracalcarine_Cortex_PC2
113
-58
-32
40
Left_Supramarginal_Gyrus_anterior_division_PC1
114
-60
-32
32
Left_Supramarginal_Gyrus_anterior_division_PC2
115
-62
-44
20
Left_Supramarginal_Gyrus_posterior_division_PC1
116
-56
-46
38
Left_Supramarginal_Gyrus_posterior_division_PC2
117
-32
-6
-36
Left_Temporal_Fusiform_Cortex_anterior_division_PC1
118
-40
-8
-30
Left_Temporal_Fusiform_Cortex_anterior_division_PC2
25
Pantazatos, et. al. 2013
119
-28
-40
-22
Left_Temporal_Fusiform_Cortex_posterior_division_PC1
120
-36
-24
-22
Left_Temporal_Fusiform_Cortex_posterior_division_PC2
121
-36
-60
-22
Left_Temporal_Occipital_Fusiform_Cortex_PC1
122
-24
-52
-14
Left_Temporal_Occipital_Fusiform_Cortex_PC2
123
-40
10
-26
Left_Temporal_Pole_PC1
124
-44
6
-18
Left_Temporal_Pole_PC2
125
0
-4
2
Left_Thalamus_PC1
126
0
-18
12
Left_Thalamus_PC2
127
-44
48
2
Left_Ventral_Frontal_Pole_PC1
128
-2
58
-4
Left_Ventral_Frontal_Pole_PC2
129
-50
-68
30
Left_Ventral_Lateral_Occipital_Cortex_superior_division_PC1
130
-32
-86
22
Left_Ventral_Lateral_Occipital_Cortex_superior_division_PC2
131
0
-24
-8
Midbrain_PC1
132
2
-36
-10
Midbrain_PC2
133
-14
-32
-26
Pons_PC1
134
-16
-32
-28
Pons_PC2
135
8
12
-6
Right_Accumbens_PC1
136
12
16
-8
Right_Accumbens_PC2
137
20
0
-20
Right_Amygdala_PC1
138
26
-4
-18
Right_Amygdala_PC2
139
58
-54
28
Right_Angular_Gyrus_PC1
140
54
-50
44
Right_Angular_Gyrus_PC2
141
10
12
8
Right_Caudate_PC1
142
12
6
18
Right_Caudate_PC2
143
50
-8
6
Right_Central_Opercular_Cortex_PC1
144
44
-12
16
Right_Central_Opercular_Cortex_PC2
145
2
38
2
Right_Cingulate_Gyrus_anterior_division_PC1
146
2
36
16
Right_Cingulate_Gyrus_anterior_division_PC2
147
2
-50
20
Right_Cingulate_Gyrus_posterior_division_PC1
148
2
-22
36
Right_Cingulate_Gyrus_posterior_division_PC2
26
Pantazatos, et. al. 2013
149
4
-80
32
Right_Cuneal_Cortex_PC1
150
2
-72
22
Right_Cuneal_Cortex_PC2
151
34
54
20
Right_Dorsal_Frontal_Pole_PC1
152
0
62
24
Right_Dorsal_Frontal_Pole_PC2
153
34
-70
52
Right_Dorsal_Lateral_Occipital_Cortex_superior_division_PC1
154
14
-82
46
Right_Dorsal_Lateral_Occipital_Cortex_superior_division_PC2
155
0
50
-10
Right_Frontal_Medial_Cortex_PC1
156
10
48
-10
Right_Frontal_Medial_Cortex_PC2
157
44
20
0
Right_Frontal_Operculum_Cortex_PC1
158
38
18
8
Right_Frontal_Operculum_Cortex_PC2
159
42
22
-12
Right_Frontal_Orbital_Cortex_PC1
160
32
18
-24
Right_Frontal_Orbital_Cortex_PC2
161
46
-14
4
Right_Heschls_Gyrus_H1_and_H2_PC1
162
44
-22
14
Right_Heschls_Gyrus_H1_and_H2_PC2
163
18
-18
-20
Right_Hippocampus_PC1
164
26
-22
-14
Right_Hippocampus_PC2
165
56
16
8
Right_Inferior_Frontal_Gyrus_pars_opercularis_PC1
166
54
16
24
Right_Inferior_Frontal_Gyrus_pars_opercularis_PC2
167
52
26
2
Right_Inferior_Frontal_Gyrus_pars_triangularis_PC1
168
54
26
18
Right_Inferior_Frontal_Gyrus_pars_triangularis_PC2
169
46
-36
-20
Right_Inferior_Temporal_Gyrus_posterior_division_PC1
170
44
-26
-20
Right_Inferior_Temporal_Gyrus_posterior_division_PC2
171
54
-54
-20
Right_Inferior_Temporal_Gyrus_temporooccipital_part_PC1
172
56
-52
-10
Right_Inferior_Temporal_Gyrus_temporooccipital_part_PC2
173
46
-4
-2
Right_Insular_Cortex_PC1
174
42
-4
6
Right_Insular_Cortex_PC2
175
2
-74
8
Right_Intracalcarine_Cortex_PC1
176
10
-80
6
Right_Intracalcarine_Cortex_PC2
177
2
0
54
Right_Juxtapositional_Lobule_Cortex_Supp_Motor_cortex_PC1
178
6
-10
50
Right_Juxtapositional_Lobule_Cortex_Supp_Motor_cortex_PC2
27
Pantazatos, et. al. 2013
179
42
-72
-22
Right_Lateral_Occipital_Cortex_inferior_division_PC1
180
48
-70
8
Right_Lateral_Occipital_Cortex_inferior_division_PC2
181
2
-76
0
Right_Lingual_Gyrus_PC1
182
8
-92
-14
Right_Lingual_Gyrus_PC2
183
48
18
40
Right_Middle_Frontal_Gyrus_PC1
184
34
28
48
Right_Middle_Frontal_Gyrus_PC2
185
48
2
-24
Right_Middle_Temporal_Gyrus_anterior_division_PC1
186
58
-6
-22
Right_Middle_Temporal_Gyrus_anterior_division_PC2
187
62
-18
-12
Right_Middle_Temporal_Gyrus_posterior_division_PC1
188
62
-30
-6
Right_Middle_Temporal_Gyrus_posterior_division_PC2
189
60
-48
6
Right_Middle_Temporal_Gyrus_temporooccipital_part_PC1
190
62
-44
-4
Right_Middle_Temporal_Gyrus_temporooccipital_part_PC2
191
30
-74
-20
Right_Occipital_Fusiform_Gyrus_PC1
192
24
-70
-12
Right_Occipital_Fusiform_Gyrus_PC2
193
6
-100
6
Right_Occipital_Pole_PC1
194
16
-92
32
Right_Occipital_Pole_PC2
195
20
-2
-2
Right_Pallidum_PC1
196
24
-10
-2
Right_Pallidum_PC2
197
0
50
-2
Right_Paracingulate_Gyrus_PC1
198
4
26
38
Right_Paracingulate_Gyrus_PC2
199
18
-20
-26
Right_Parahippocampal_Gyrus_anterior_division_PC1
200
6
-6
-24
Right_Parahippocampal_Gyrus_anterior_division_PC2
201
18
-28
-22
Right_Parahippocampal_Gyrus_posterior_division_PC1
202
16
-34
-16
Right_Parahippocampal_Gyrus_posterior_division_PC2
203
54
-28
18
Right_Parietal_Operculum_Cortex_PC1
204
42
-32
20
Right_Parietal_Operculum_Cortex_PC2
205
46
0
-12
Right_Planum_Polare_PC1
206
48
-2
-2
Right_Planum_Polare_PC2
207
60
-24
12
Right_Planum_Temporale_PC1
208
52
-28
14
Right_Planum_Temporale_PC2
28
Pantazatos, et. al. 2013
209
58
-14
40
Right_Postcentral_Gyrus_PC1
210
54
-14
42
Right_Postcentral_Gyrus_PC2
211
54
8
34
Right_Precentral_Gyrus_PC1
212
48
-8
54
Right_Precentral_Gyrus_PC2
213
2
-56
24
Right_Precuneous_Cortex_PC1
214
4
-70
62
Right_Precuneous_Cortex_PC2
215
26
6
0
Right_Putamen_PC1
216
28
-4
6
Right_Putamen_PC2
217
0
26
-4
Right_Subcallosal_Cortex_PC1
218
2
16
-8
Right_Subcallosal_Cortex_PC2
219
2
54
36
Right_Superior_Frontal_Gyrus_PC1
220
4
20
54
Right_Superior_Frontal_Gyrus_PC2
221
38
-48
58
Right_Superior_Parietal_Lobule_PC1
222
30
-46
62
Right_Superior_Parietal_Lobule_PC2
223
50
-2
-16
Right_Superior_Temporal_Gyrus_anterior_division_PC1
224
60
-2
-6
Right_Superior_Temporal_Gyrus_anterior_division_PC2
225
50
-12
-8
Right_Superior_Temporal_Gyrus_posterior_division_PC1
226
62
-12
-4
Right_Superior_Temporal_Gyrus_posterior_division_PC2
227
2
-78
12
Right_Supracalcarine_Cortex_PC1
228
18
-64
16
Right_Supracalcarine_Cortex_PC2
229
64
-28
32
Right_Supramarginal_Gyrus_anterior_division_PC1
230
56
-28
48
Right_Supramarginal_Gyrus_anterior_division_PC2
231
62
-40
34
Right_Supramarginal_Gyrus_posterior_division_PC1
232
64
-44
16
Right_Supramarginal_Gyrus_posterior_division_PC2
233
34
-6
-36
Right_Temporal_Fusiform_Cortex_anterior_division_PC1
234
38
-2
-40
Right_Temporal_Fusiform_Cortex_anterior_division_PC2
235
30
-34
-24
Right_Temporal_Fusiform_Cortex_posterior_division_PC1
236
38
-28
-22
Right_Temporal_Fusiform_Cortex_posterior_division_PC2
237
34
-56
-22
Right_Temporal_Occipital_Fusiform_Cortex_PC1
238
36
-52
-26
Right_Temporal_Occipital_Fusiform_Cortex_PC2
29
Pantazatos, et. al. 2013
239
40
12
-26
Right_Temporal_Pole_PC1
240
48
8
-10
Right_Temporal_Pole_PC2
241
2
-4
2
Right_Thalamus_PC1
242
10
-22
6
Right_Thalamus_PC2
243
38
54
0
Right_Ventral_Frontal_Pole_PC1
244
0
58
0
Right_Ventral_Frontal_Pole_PC2
245
50
-66
28
Right_Ventral_Lateral_Occipital_Cortex_superior_division_PC1
246
34
-84
26
Right_Ventral_Lateral_Occipital_Cortex_superior_division_PC2
247
2
-46
-18
Vermis_PC1
248
0
-54
-34
Vermis_PC2
30
Pantazatos, et. al. 2013
Supplemental Table S2
SAD vs. Controls
F betas, top 170
features
Region
x
y
z
Fastigium
Cerebelum_4_5_R
Cerebelum_Crus1_L
Sub-Gyral
Frontal_Sup_Orb_R
Midbrain
Temporal_Mid_R
Extra-Nuclear
Lingual_R
Frontal_Mid_Orb_L
Temporal_Mid_L
Frontal_Inf_Tri_L
Occipital_Sup_L
SupraMarginal_L
Lateral Ventricle
Insula_L
Cuneus
Frontal_Inf_Tri_L
Posterior Cingulate
Sub-Gyral
Cuneus_L
Cingulate Gyrus
Extra-Nuclear
Frontal_Inf_Tri_L
Cingulum_Mid_L
Precentral_L
SupraMarginal_R
Precentral_R
Postcentral_R
Frontal_Sup_Medial_L
Paracentral_Lobule_L
F>N contrast values,
top 8 features
Region
10
14
-26
42
18
10
66
6
22
-38
-58
-50
-22
-54
-6
-22
-18
-42
-2
30
-10
-10
-10
-34
-6
-42
62
38
30
-2
-6
-60
-52
-88
-36
32
-28
-20
0
-52
56
-48
40
-64
-24
16
24
-80
24
-32
16
-88
-16
0
32
-32
8
-24
-20
-28
28
-24
-30
-22
-22
-14
-14
-10
-10
-10
2
-2
10
10
26
14
14
14
14
22
22
22
26
30
26
26
34
42
42
58
46
54
58
x
y
z
Cerebelum_6_R
Middle Occipital Gyrus_R
Extra-Nuclear
10
26
34
-76
-64
-8
-22
14
26
Cluster
size
2
12
2
1
3
1
3
2
8
7
17
2
13
14
2
3
1
19
1
1
1
6
3
1
1
11
1
12
6
13
1
SVM weight
Cluster
size
1
6
1
T-value
0.11766
0.14319
0.093878
0.16173
-0.1472
0.13876
0.079332
0.14802
0.0921
-0.08209
-0.12757
-0.11533
0.14933
0.19237
-0.12109
-0.12363
0.1218
-0.23092
-0.069371
-0.064564
0.03237
-0.14666
-0.17125
-0.071431
-0.098303
0.076975
-0.24465
-0.15092
-0.1556
-0.11001
0.1246
-1.5561
1.2832
1.5478
31
Pantazatos, et. al. 2013
Supplemental Table S3
Whole brain results corresponding to Supplemental Figure 4: PPI analysis, Left Temporal
Pole (top) and Left Hippocampus (bottom) as seed, N condition, Control > SAD
LTP, P-value = 0.001, Cluster threshold = 4
Region
x y z Cluster size T-value
Pons
0 -32 -30
11 -3.6385
Cerebelum_6_L
-10 -60 -20
9 -3.4956
Fusiform_L
-34 -58 -20
15 -3.9095
ParaHippocampal_L
-20 -18 -20
4 3.4287
Midbrain
4 -24 -10
13 3.7302
Temporal_Mid_L
-58 -56 -4
30 3.8853
Temporal_Mid_R
58 -46 2
21 3.856
Hippocampus_L
-36 -32 2
4 3.3426
Sub-Gyral
32 -42 8
7 3.6794
Cingulum_Mid_L
-8 -22 44
13 -3.7999
Precentral_L
-50 0 40
6 -3.5892
LHipp, P-value = 0.001, Cluster threshold = 4
Region
x y z Cluster size T-value
Thalamus_R
8 -16 8
13 -4.0177
Sub-Gyral
28 22 16
24 4.2935
Frontal_Inf_Tri_L
-52 26 14
6 3.5373
Precuneus_L
-8 -48 14
8 -3.8489
Temporal_Sup_R
40 -34 18
6 3.6271
Sub-Gyral
-32 22 18
4 3.5652
Cingulum_Post_R
12 -48 22
4 3.5628
Sub-Gyral
32 -4 30
4 3.887
32
Pantazatos, et. al. 2013
Supplemental Table S4
SAD vs. control activation differences at p<0.05 corrected (p=0.005, cluster threshold >
148), based on F-test from 2-way ANOVA (1st factor diagnosis: 2 levels, 2nd factor face
conditions: 4 levels-F, N, MF, MN). For directionality of results see supplemental figure S5. No
significant diagnosis X face condition interactions were observed at this threshold.
2-way ANOVA
Main effect of diagnosis
Region
L/R x y z Cluster size F-value(1,130) Z score
DLPFC
L
-50 6 60
152
28.28
4.91
Fusiform
L
-32 -76 -14
503
23.04
4.45
33
Supplemental Figure S1
Pantazatos, et. al. 2013
34
Supplemental Figure S2
Pantazatos, et. al. 2013
35
Supplemental Figure S3
Pantazatos, et. al. 2013
36
Supplemental Figure S4
Pantazatos, et. al. 2013
37
Supplemental Figure S5
Pantazatos, et. al. 2013
38
Supplemental Figure S6
Pantazatos, et. al. 2013
39
Pantazatos, et. al. 2013
References
Ashburner J & Friston KJ. 2000. Voxel-based morphometry--the methods. Neuroimage, 11(6 Pt
1), 805-21.
Carey PD, Warwick J, Niehaus DJ, van der Linden G, van Heerden BB et al. 2004. Single
photon emission computed tomography (SPECT) of anxiety disorders before and after
treatment with citalopram. BMC Psychiatry, 4, 30.
Ding J, Chen H, Qiu C, Liao W, Warwick JM et al. 2011. Disrupted functional connectivity in
social anxiety disorder: a resting-state fMRI study. Magn Reson Imaging, 29(5), 701-11.
Doehrmann O, Ghosh SS, Polli FE, Reynolds GO, Horn F et al. 2012. Predicting Treatment
Response in Social Anxiety Disorder From Functional Magnetic Resonance Imaging.
Arch. Gen. Psychiatry, 1-11.
Dosenbach NU, Nardos B, Cohen AL, Fair DA, Power JD et al. 2010. Prediction of individual
brain maturity using fMRI. Science (New York, N.Y.), 329(5997), 1358-1361.
Etkin A, Klemenhagen KC, Dudman JT, Rogan MT, Hen R et al. 2004. Individual differences in
trait anxiety predict the response of the basolateral amygdala to unconsciously processed
fearful faces. Neuron, 44(6), 1043-1055.
Fu CH, Mourao-Miranda J, Costafreda SG, Khanna A, Marquand AF et al. 2008. Pattern
classification of sad facial processing: toward the development of neurobiological
markers in depression. Biological psychiatry, 63(7), 656-662.
Furmark T, Tillfors M, Marteinsdottir I, Fischer H, Pissiota A et al. 2002. Common changes in
cerebral blood flow in patients with social phobia treated with citalopram or cognitivebehavioral therapy. Arch. Gen. Psychiatry, 59(5), 425-33.
Giménez M, Pujol J, Ortiz H, Soriano-Mas C, López-Solà M et al. 2012. Altered brain functional
connectivity in relation to perception of scrutiny in social anxiety disorder. Psychiatry
research.
Golland P & Fischl B. 2003. Permutation tests for classification: towards statistical significance
in image-based studies. Inf Process Med Imaging, 18, 330-41.
Hanley JA & McNeil BJ. 1982. The meaning and use of the area under a receiver operating
characteristic (ROC) curve. Radiology, 143(1), 29-36.
Jimura K, Konishi S, Asari T & Miyashita Y. 2010. Temporal pole activity during understanding
other persons' mental states correlates with neuroticism trait. Brain Res, 1328, 104-12.
Kim J & Horwitz B. 2008. Investigating the neural basis for fMRI-based functional connectivity
in a blocked design: application to interregional correlations and psycho-physiological
interactions. Magnetic resonance imaging, 26(5), 583-593.
40
Pantazatos, et. al. 2013
Klumpp H, Angstadt M, Nathan PJ & Phan KL. 2010. Amygdala reactivity to faces at varying
intensities of threat in generalized social phobia: an event-related functional MRI study.
Psychiatry Res, 183(2), 167-9.
Klumpp H, Angstadt M & Phan KL. 2012. Insula reactivity and connectivity to anterior
cingulate cortex when processing threat in generalized social anxiety disorder. Biol
Psychol, 89(1), 273-6.
Leckman JF, Sholomskas D, Thompson WD, Belanger A & Weissman MM. 1982. Best estimate
of lifetime psychiatric diagnosis: a methodological study. Arch. Gen. Psychiatry, 39(8),
879-83.
Liao W, Qiu C, Gentili C, Walter M, Pan Z et al. 2010. Altered effective connectivity network of
the amygdala in social anxiety disorder: a resting-state FMRI study. PLoS ONE, 5(12),
e15238.
Mannuzza S, Fyer AJ, Klein DF & Endicott J. 1986. Schedule for Affective Disorders and
Schizophrenia—Lifetime Version modified for the study of anxiety disorders (SADSLA): rationale and conceptual development. Journal of psychiatric research.
McLaren DG, Ries ML, Xu G & Johnson SC. 2012. A generalized form of context-dependent
psychophysiological interactions (gPPI): a comparison to standard approaches.
Neuroimage, 61(4), 1277-86.
Obuchowski NA & Lieber ML. 1998. Confidence intervals for the receiver operating
characteristic area in studies with small samples. Acad Radiol, 5(8), 561-71.
Pannekoek JN, Veer I, van Tol MJ, van der Werff S, Demenescu LR et al. 2012. Resting-state
functional connectivity abnormalities in limbic and salience networks in social anxiety
disorder without comorbidity. European neuropsychopharmacology : the journal of the
European College of Neuropsychopharmacology.
Pantazatos SP, Talati A, Pavlidis P & Hirsch J. 2012a. Cortical functional connectivity decodes
subconscious, task-irrelevant threat-related emotion processing. NeuroImage, 61(4),
1355-1363.
Pantazatos SP, Talati A, Pavlidis P & Hirsch J. 2012b. Decoding unattended fearful faces with
whole-brain correlations: an approach to identify condition-dependent large-scale
functional connectivity. PLoS Comput. Biol, 8(3), e1002441.
Prater KE, Hosanagar A, Klumpp H, Angstadt M & Phan KL. 2012. ABERRANT
AMYGDALA-FRONTAL CORTEX CONNECTIVITY DURING PERCEPTION OF
FEARFUL FACES AND AT REST IN GENERALIZED SOCIAL ANXIETY
DISORDER. Depress Anxiety.
41
Pantazatos, et. al. 2013
Psychiatric Association American. 1994. Diagnostic and statistical manual of psychiatric
disorders. Washington, DC.
Smith SM, Miller KL, Salimi-Khorshidi G, Webster M, Beckmann CF et al. 2011. Network
modelling methods for FMRI. NeuroImage, 54(2), 875-891.
Talati A, Ponniah K, Strug LJ, Hodge SE, Fyer AJ & Weissman MM. 2008. Panic disorder,
social anxiety disorder, and a possible medical syndrome previously linked to
chromosome 13. Biol. Psychiatry, 63(6), 594-601.
Van der Linden G, van Heerden B, Warwick J, Wessels C, van Kradenburg J et al. 2000.
Functional brain imaging and pharmacotherapy in social phobia: single photon emission
computed tomography before and after treatment with the selective serotonin reuptake
inhibitor citalopram. Prog. Neuropsychopharmacol. Biol. Psychiatry, 24(3), 419-38.
Weissman MM, Wickramaratne P, Adams P, Wolk S, Verdeli H & Olfson M. 2000. Brief
screening for family psychiatric history: the family history screen. Arch. Gen. Psychiatry,
57(7), 675-82.
Winton EC, Clark DM & Edelmann RJ. 1995. Social anxiety, fear of negative evaluation and the
detection of negative emotion in others. Behav Res Ther, 33(2), 193-6.
Yoon KL & Zinbarg RE. 2007. Threat is in the eye of the beholder: social anxiety and the
interpretation of ambiguous facial expressions. Behaviour research and therapy, 45(4),
839-847.