Disrupted Amygdalar Subregion Functional Connectivity and Evidence of a Compensatory Network in Generalized Anxiety Disorder

Context  Little is known about the neural abnormalities underlying generalized anxiety disorder (GAD). Studies in other anxiety disorders have implicated the amygdala, but work in GAD has yielded conflicting results. The amygdala is composed of distinct subregions that interact with dissociable brain networks, which have been studied only in experimental animals. A functional connectivity approach at the subregional level may therefore yield novel insights into GAD.

Objectives  To determine whether distinct connectivity patterns can be reliably identified for the basolateral (BLA) and centromedial (CMA) subregions of the human amygdala, and to examine subregional connectivity patterns and potential compensatory amygdalar connectivity in GAD.

Design  Cross-sectional study.

Setting  Academic medical center.

Participants  Two cohorts of healthy control subjects (consisting of 17 and 31 subjects) and 16 patients with GAD.

Main Outcome Measures  Functional connectivity with cytoarchitectonically determined BLA and CMA regions of interest, measured during functional magnetic resonance imaging performed while subjects were resting quietly in the scanner. Amygdalar gray matter volume was also investigated with voxel-based morphometry.

Results  Reproducible subregional differences in large-scale connectivity were identified in both cohorts of healthy controls. The BLA was differentially connected with primary and higher-order sensory and medial prefrontal cortices. The CMA was connected with the midbrain, thalamus, and cerebellum. In GAD patients, BLA and CMA connectivity patterns were significantly less distinct, and increased gray matter volume was noted primarily in the CMA. Across the subregions, GAD patients had increased connectivity with a previously characterized frontoparietal executive control network and decreased connectivity with an insula- and cingulate-based salience network.

Conclusions  Our findings provide new insights into the functional neuroanatomy of the human amygdala and converge with connectivity studies in experimental animals. In GAD, we find evidence of an intra-amygdalar abnormality and engagement of a compensatory frontoparietal executive control network, consistent with cognitive theories of GAD.

Generalized anxiety disorder (GAD) is a common anxiety disorder in adults, with estimated lifetime prevalence rates of around 5%.^1,2 Generalized anxiety disorder has a long average duration of symptoms^2,3 and is associated with significant quality-of-life impairment or disability.⁴ Despite its clinical importance, GAD has received considerably less study than other anxiety disorders. Predictions of which brain circuits are altered in GAD must therefore be extrapolated from findings in other anxiety disorders. A recent meta-analysis from our group⁵ showed that posttraumatic stress disorder, social anxiety disorder, and specific phobia—3 of the most frequently studied disorders—are all characterized by hyperactivity of the amygdala and insula of patients during the processing of negative emotion. Engagement of the amygdala and insula was also seen in this meta-analysis when healthy subjects experienced fear,⁵ suggesting that amygdalar and insular hyperactivation in patients reflects the neural correlates of a current fear or anxiety state or a trait vulnerability for excessive fear responses.

Two recent studies examined amygdala activity in adult patients with GAD in response to viewing fearful faces, a commonly used probe of amygdalar engagement.^6,7 One study found no difference in the amygdala between patients and control subjects,⁷ whereas the other study found less activity in the amygdala in patients.⁶ By contrast, similar approaches in adolescents with GAD have found hyperactivity in the amygdala.^8,9 Many factors may explain these differences, including differences in baseline amygdalar activity, the capacity for task-related activation, overall responses to emotional expressions, and the nature and relative levels of GAD psychopathology in each cohort. Nonetheless, these studies leave undefined the role the amygdala plays in GAD. Blair et al⁶ used their own data and those of related studies and suggested that an understanding of amygdalar dysfunction in GAD, along with discrepancies between individual studies, will require investigation of the brain network contexts in which the amygdala is involved.

Analyses of functional connectivity using functional magnetic resonance imaging (fMRI) data acquired while the subjects are at rest, unbiased by task demands, have been previously used to identify multiple, simultaneously operating, broadly connected networks of brain regions.¹⁰ Abnormalities in resting functional connectivity have also been identified in major depression.¹¹ Use of a resting-state approach may therefore provide a useful tool for examining dysfunctional amygdalar-based neural circuitry in GAD.

Our understanding of the organization of the amygdala is derived in large part from investigations of the functions of subnuclei within the amygdala of animals and the distinct brain networks in which they participate. This extensive body of research has established a model wherein sensory information across multiple modalities enters the amygdala through the nuclei of the basolateral complex (BLA) (consisting of lateral, basal, and accessory basal nuclei).^12-16 Neurons in the BLA encode fear memories related to these sensory stimuli, signal the threat value of a stimulus, and can modulate memory encoding and sensory processing in other brain regions.^13,14 The BLA in turn activates the central nucleus, which is essential for the basic species-specific defensive responses associated with fear.^12-16 The central nucleus achieves these functions through projections to brainstem, hypothalamic, and basal forebrain targets.^12-16

Central to this outlined circuit are anatomical connectivity findings in rodents and nonhuman primates, which differentiate the largely cortical connectivity pattern of the BLA from the largely subcortical connectivity pattern of the central nucleus.^12,15,16 The anatomical connectivity of human amygdalar nuclei, however, is currently unknown. In this study, we therefore examined the differential connectivity patterns of these amygdalar subregions in healthy control subjects and GAD patients to investigate the functional brain networks in which the amygdala is involved and that underlie the distinct functions of these amygdalar subregions.

We first sought to establish whether the functional connectivity of amygdalar subregions can be distinguished in healthy subjects by using resting-state fMRI. Cytoarchitectural, myeloarchitectural, and chemoarchitectural studies of the human amygdala support a differentiation between the BLA and a centromedial subregion (CMA) composed of the central and medial nuclei.¹⁷ We therefore derived regions of interest (ROIs) from cytoarchitectonically determined probabilistic maps of the human BLA and CMA.^18,19 To demonstrate the reliability of this approach, we derived functional connectivity maps from 2 independent control cohorts. We then compared these functional connectivity patterns with those in a matched GAD patient cohort. Finally, we complemented our functional connectivity analyses by performing voxel-based morphometry (VBM), an independent and commonly used structural imaging approach, focusing on gray matter volumes of the BLA and CMA.

A total of 64 subjects participated in this study (see Table 1 for demographics) after giving their informed consent according to institutional guidelines for the protection of human subjects at Stanford University. The first control cohort (GAD controls) and the GAD patients were recruited through local online advertisements. Psychiatric diagnoses based on the DSM-IV²⁰ were determined through an informal clinical interview with a psychiatrist and the structured diagnostic Mini International Neuropsychiatric Interview.^21,22 Exclusion criteria were bipolar, psychotic, or posttraumatic stress disorders or substance abuse diagnoses. Generalized anxiety disorder was the primary diagnosis for all patients. Patients with comorbid major depression were included if the onset of GAD was clearly primary to that of depression. Other exclusion criteria included a history of a neurological disorder, head trauma or loss of consciousness, claustrophobia, and regular use of benzodiazepines, opiates, or thyroid medications. Of the 16 patients, 4 were taking regular antidepressants at a stable dose. No subject used an as-needed dose of a benzodiazepine within 48 hours of the scan. Comorbid Axis I conditions included major depression (n = 4), social anxiety disorder (n = 5), panic disorder (n = 2), dysthymia (n = 2), and obsessive-compulsive disorder (n = 1). Of the patients, 6 had no comorbidities, 6 had 1 comorbidity (depression in 3, dysthymia in 2, and social anxiety in 1), and 4 had 2 comorbidities (social anxiety and panic disorder in 2, social anxiety and obsessive-compulsive disorder in 1, and depression and social anxiety in 1). None had more than 2 comorbidities. All GAD controls were free of any current or past Axis I condition and psychiatric medication. No GAD patient had undergone structured psychotherapy. All GAD controls and GAD patients completed the Spielberger State-Trait Anxiety Inventory,²³ the Penn State Worry Questionnaire,²⁴ the Beck Anxiety Inventory,²⁵ the Beck Depression Inventory II,²⁶ and the Mood and Anxiety Symptoms Questionnaire.^27,28 The 31 healthy subjects in the second control cohort were recruited as part of an unrelated study and denied a history of psychiatric or neurological disorders but did not undergo formal evaluation with a structured interview. Data from 22 of these subjects were used previously in an unrelated study.²⁹

All participants underwent an 8-minute resting-state fMRI scan in which they were told to keep their eyes closed, hold still, try not to fall asleep, and allow their minds to wander. Images were acquired on a 3-T scanner (GE Signa scanner; GE Healthcare, Milwaukee, Wisconsin) using a custom-built head coil. A total of 29 axial slices (4.0-mm thickness) covering the whole brain were acquired using a T2-weighted gradient-echo spiral-pulse sequence (repetition time, 2000 milliseconds; echo time, 30 milliseconds; flip angle, 80°; and 1 interleave).³⁰ The field of view was 22 cm for the GAD controls and GAD patients and 20 cm for the second control cohort, and the matrix size was 64 × 64. To reduce blurring and signal loss arising from field inhomogeneities, an automated high-order shimming method based on spiral acquisitions was used before acquiring fMRI scans.³¹ A high-resolution T1-weighted spoiled grass gradient-recalled inverted recovery 3-dimensional MRI sequence was used with the following parameters: time after inversion pulse, 300 milliseconds; repetition time, 8 milliseconds; echo time, 3.6 milliseconds; flip angle, 15°; field of view, 22 cm; 124 slices in the coronal plane; 256 × 192 matrix; number of excitations, 2; and acquired resolution, 1.5 × 0.9 × 1.1 mm. The images were reconstructed as a 124 × 256 × 256 matrix with a 1.5 × 0.9 × 0.9-mm spatial resolution. Structural and functional images were acquired in the same scan session.

The first 8 volumes were not analyzed to allow for signal equilibration effects. A linear shim correction was applied separately for each slice during reconstruction using a magnetic field map acquired automatically by the pulse sequence at the beginning of the scan.³⁰ The fMRI data were then preprocessed using SPM5 software (available at: http://www.fil.ion.ucl.ac.uk/spm) implemented in a MATLAB suite (Mathworks, Inc, Natick, Massachusetts). Images were realigned to correct for motion, corrected for errors in slice timing, spatially transformed to standard stereotaxic space (based on the Montreal Neurologic Institute coordinate system),³² resampled every 2 mm, and smoothed with a 6-mm full-width half-maximum gaussian kernel. There were no participants with movement greater than 3 mm of translation or 3° of rotation. There were also no significant differences between the total range of movement across any axis of translation or rotation between groups. Data were then bandpass filtered using FMRIB Software Library tools (available at: http://www.fmrib.ox.ac.uk/fsl/index.html) to retain frequencies from 0.008 to 0.1 Hz, to remove high- and low-frequency noise sources, and to allow for better discrimination of low-frequency resting-state networks. Most of the power in resting-state networks is found in this low-frequency band.³³

Seeds for the connectivity analysis corresponded to the BLA and CMA subregions³⁴ and were constructed from maximum probability maps of these 2 subregions, defined through the Anatomy Toolbox in SPM5.¹⁹ Maximum probability maps allow the definition of nonoverlapping regions from underlying probability maps that are inherently overlapping.^18,19 Voxels were included in the maximum probability maps if the probability of their being assigned to the BLA or CMA was higher than for other nearby structures, such as other amygdalar subregions or medial temporal lobe structures, with no less than 40% likelihood.¹⁹ The BLA and CMA maximum probability maps were then converted to ROIs using a toolbox for SPM (MarsBaR [MARSeille Boîte À Région d'Intérêt]; available at http://marsbar.sourceforge.net/), and the individual time series within each ROI was obtained from bandpass-filtered images. Each time series was then put into a first-level fixed-effects general linear model in SPM5,³⁵ and 4 separate connectivity analyses were then performed for each subject (2 ROIs per side). A global signal regressor and the 6 motion parameters for each subject were included as covariates of no interest in each model. For direct correlations of ROI time courses, variance associated with the global mean or motion parameters was removed before performing the correlations. A similar approach was taken in a control connectivity analysis in which the primary auditory cortex was the seed region, using a previously defined ROI.³⁶

We performed group-level analyses using the contrast images from the individual functional connectivity analyses in random-effects t tests or analyses of variance (ANOVAs). The ANOVAs were performed with a flexible factorial model in which a subject factor was additionally used to model the sphericity of the data appropriately. All group-level analyses were masked with a map from 1-sample t tests of each relevant group, thresholded leniently at P = .05, uncorrected, and then combined across side and region to restrict connectivity analyses to positively correlated voxels only. This procedure avoided potential interpretational confounds related to apparently negative connectivity resulting from correction for global signal changes.³⁷ Statistical thresholds were set at q < 0.05, whole-brain corrected for the false discovery rate. Connectivity foci were labeled by comparison with neuroanatomical atlases.³⁸ Reported voxel coordinates correspond to standardized Montreal Neurologic Institute space. Because the BLA cortical targets formed 1 very large cluster at q < 0.05, whole-brain corrected, we report representative differential connectivity peaks separately for each lobe in Table 2 and limit the number of peaks to no more than 3 per region for clarity of presentation. For the displayed sections, the right side of the image corresponds to the right side of the brain.

To determine the connectivity specificity of amygdalar subregions with their respective targets, average connectivity (β weights) in volumes of interest were extracted from CMA- or BLA-seeded connectivity analyses (separated by side) using masks generated from the ANOVA effect of region for the GAD control cohort (separately for BLA or CMA targets). The volume of interest data then underwent ANOVA using commercially available statistical software (SPSS; SPSS Inc, Chicago, Illinois). Identical masks were used for both control cohorts and the GAD patient group. All error bars or standard error ranges reported refer to standard errors of the mean.

Voxel based morphometry was performed using the VBM5 toolbox in SPM5 (Wellcome Department of Cognitive Neurology, London, England; available at: http://dbm.neuro.uni-jena.de/vbm/vbm5-for-spm5/). The individual T1-weighted images from the GAD controls and the matched GAD patient group were segmented using a unified segmentation routine through the VBM5 toolbox.³⁹ The segmented, modulated gray matter images were then smoothed at 6 mm full width half maximum.⁴⁰ The volumes of interest corresponding to gray matter volume (modulated images) were extracted for the BLA and CMA on both sides for statistical analyses in SPSS.

We initially examined a cohort of 17 well-characterized healthy subjects (the GAD controls) (Table 1) by conducting resting-state fMRI connectivity analyses separately for each of the 2 amygdalar subregion ROIs across both sides of the brain while controlling for variations in global signal (described in the “Functional Connectivity Analyses” subsection of the “Methods” section). In this cohort, the BLA and CMA ROIs explained 21% of the variance in the other ROI during our resting-state fMRI scans, suggesting that these ROIs have enough nonshared variance to allow for identification of differential connectivity patterns.

Next, we performed a set of connectivity analyses, using the BLA and CMA ROIs as seeds, which we entered into a voxelwise ANOVA. Directly contrasting BLA and CMA connectivity (collapsing across sides) yielded robust, spatially coherent patterns of differential connectivity (Figure 1A). The connectivity patterns were bilateral and similar for ROIs from the right and left sides (data not shown), and there were no significant connectivity clusters in the region × side ANOVA interaction analysis (q < 0.05; data not shown).

The BLA was associated with exclusively cortical differential connectivity, encompassing bilaterally the entire occipital lobe, large extents of the ventral temporal lobe, superior temporal gyrus and sulcus, precentral and postcentral gyri, intraparietal sulcus, lateral posterior orbitofrontal cortex, dorsal and ventral medial prefrontal cortex (PFC), parahippocampal gyri, and a small region in the inferior lateral PFC (Figure 1A). This broad neural network includes primary and higher-order association cortices for the visual, somatosensory, and auditory systems, as well as motor regions, areas important in memory formation, and higher-order areas in the PFC, most notably along its medial and inferior aspects. The CMA was associated almost exclusively with subcortical differential connectivity, including the thalamus (in particular along the midline), midbrain (in the region of the substantia nigra, ventral tegmental area, and periaqueductal gray), medulla, cerebellum (in particular in the vermis), and caudate (Figure 1A).

To determine the robustness and replicability of these connectivity patterns, we examined a second, unrelated group of 31 healthy adults (second control cohort) and replicated the differential connectivity patterns for the BLA and CMA that had been identified in the GAD controls (Figure 1B). The only difference in connectivity results between these cohorts consisted of insular, dorsal anterior cingulate, and midcingulate CMA connectivity in the second control cohort. To directly and quantitatively identify those regions with the most robust and reproducible pattern of differential BLA and CMA connectivity, we conducted a formal conjunction analysis across the 2 independent control cohorts (Figure 1C and Table 2).⁴¹ This analysis confirmed the exclusively cortical pattern of differential BLA connectivity and the subcortical pattern of differential CMA connectivity, as detailed in the preceding paragraphs.

To further quantify the differential connectivity patterns, we constructed a separate mask for BLA and CMA targets derived only from the GAD controls, removing amygdalar voxels to avoid the potentially confounding effects of correlations of ROIs with themselves. We determined the average connectivity of the BLA or CMA target networks with the BLA and CMA source ROIs from both sides (Figure 2A). As expected, in the GAD control cohort, we found a strong reciprocal difference in BLA and CMA connectivity with their respective targets (eg, stronger connectivity of BLA ROIs with BLA targets than CMA targets; ANOVA region × target interaction, P < .001), which was consistent across both sides (ANOVA side × region × target interaction, P = .38). Strikingly, we found highly similar results for the second control cohort, using the target network defined only from the GAD controls (ANOVA region × target interaction, P < .001; region × side × target interaction, P = .59). The 2 control cohorts, moreover, did not differ in their pattern of differential BLA-CMA connectivity (ANOVA group × region × target interaction, P = .63). This dissociable pattern of amygdalar subregional connectivity in the second control cohort was captured by target networks independently defined from the ANOVA on just the GAD controls. Performing the analysis in this way would, if anything, be expected to reduce our sensitivity for finding subregional connectivity differences in the second control cohort and thus speaks to the robustness of these findings.

Finally, we examined the relationship between target region connectivity and cytoarchitectonic probability of individual amygdalar voxels by performing seeded connectivity analyses for both control cohorts using the entire BLA or the CMA target network, defined from just the GAD controls, as source ROIs. We then extracted the connectivity strength and cytoarchitectonic probability for each voxel in the amygdala. The probability of belonging to the BLA or CMA was found to significantly and positively predict that voxel's connectivity with the BLA or CMA target network, respectively, on both the right and left sides for the GAD controls (P < .001 for all comparisons). This relationship also held true for the second control cohort, despite the fact that the target networks were derived from the ANOVA data of the GAD controls (P < .001 for all comparisons).

We next wanted to determine whether an amygdalar subregional connectivity analysis could provide insight into GAD. In 16 GAD patients who were age-, sex-, and education-matched to the GAD controls (see Table 1), we examined the connectivity strength of the BLA and CMA ROIs with the target networks identified from the GAD controls (Figure 2A). Whereas the 2 independent control cohorts showed similarly robust differential subregional connectivity, the GAD patients showed less distinct subregion-target connectivity patterns (ANOVA group × region × target interaction, P = .003; partial η² = 0.25), although these patients still showed some degree of dissociation between BLA and CMA target connectivity (ANOVA region × target interaction for GAD patients only, P = .001). This group difference in specificity of subregional connectivity was driven by decreased connectivity in patients between the BLA or CMA and their respective targets and by increased connectivity with the other subregion's targets (Figure 2A). This effect can also be clearly seen in Figure 3A and B, in which the ANOVA effects of group are shown separately for each subregion. In Figure 3A, decreased connectivity of the GAD patients' BLA with BLA targets is shown in blue (eg, occipitotemporal cortex, superior temporal gyrus, somatomotor cortices, and medial PFC). Similarly, increased connectivity of GAD patients' CMA with the same BLA targets is shown in red in Figure 3B. This voxelwise analysis also revealed greater BLA connectivity in GAD patients with normal CMA targets (the thalamus, brainstem, and cerebellum) and decreased CMA connectivity with its normal targets (data not shown). Thus, the target network approach in Figure 2A, in which amygdalar connectivity with all significant voxels was averaged into a single summary value, yielded the same results as the voxelwise analysis and thereby provided a simple graphic representation of the findings. Moreover, the group × region × target interaction finding remained significant after controlling for the regular use of psychiatric medication (P = .01, partial η² = 0.19) or the presence of comorbid major depressive disorder (P = .008, partial η² = 0.21), a psychiatric condition closely related to GAD. Furthermore, comparisons of subgroups based on the use of psychiatric medication or the presence of comorbid depression did not yield significant differences.

We next examined whether the group differences in amygdalar connectivity reflected a global alteration of resting-state connectivity in patients, by examining connectivity within a control network previously reported by our group.³⁶ We found that the primary auditory cortex was strongly connected to its contralateral homologue in both control cohorts (left seed target peak voxel coordinates, 54, −18, 8 [z = 6.0]; right seed, −46, −25, 4 [z = 4.87]) and patients (left seed, 46, −16, 8 [z = 6.03]; right seed, −52, −16, 6 [z = 5.51]), with no difference between groups at q < 0.05. The amygdalar connectivity findings are therefore not due to global changes in resting-state connectivity.

Differences in subregional connectivity in patients using the whole-network masks may be due to an intra-amygdalar alteration (eg, in subregional size, shape, or cellular composition) or a change in the connectivity of the amygdala with specific target brain regions. To distinguish between these 2 possibilities, we performed a voxelwise ANOVA group × region analysis and displayed it at a lenient threshold to avoid false-negative results. Significance in this analysis would mean that there is a group difference in the dissociability of BLA-CMA connectivity, because of decreased connectivity with the correct targets or increased connectivity with the incorrect targets in 1 group. We found similar ANOVA significance diffusely across all BLA or CMA target regions (Figure 2B). These data argue, therefore, that there is an intra-amygdalar perturbation in subregional organization in GAD rather than an alteration in the connectivity of the amygdala selectively with some targets but not others. In additional support of this idea, we found that time courses from the BLA and CMA ROIs were more highly correlated in the GAD patients than in the control cohorts (P = .002, Cohen's d = 1.2), accounting for nearly twice as much of each others' variance as in the controls (37% vs 21%), thus consistent with GAD patients having less distinct subregional connectivity patterns.

To further test the idea that amygdalar structure is altered at the subregional level in GAD, we sought convergent evidence from a complementary neuroimaging modality sensitive to changes in brain structure. To do so, we conducted a VBM analysis of gray matter volume within the amygdala, using the same maximum probability maps as for the functional connectivity analysis because this was the greatest level of neuroanatomical detail for which we sought to make conclusions. We found a significant increase in gray matter volume in GAD patients (modulated images; ANOVA effect of group, P = .01, partial η² = 0.55), which was most notable within the right CMA (Figure 3C; 2-sample t test, P = .03; Cohen's d = 0.78). These data therefore provide convergent evidence for an intra-amygdalar perturbation at the level of individual subregions in GAD patients, using an independent measure of brain structure from anatomical (nonfunctional) brain scans.

In examining the group differences in BLA or CMA connectivity (Figure 3A and B), we also found that several regions consistently showed increased or decreased connectivity in patients using seeds in either of the 2 subregions—an effect orthogonal with the decreased subregional specificity of connectivity that we have already noted in patients. To explore these findings, we identified those voxels in which GAD patients had increased (red) or decreased (blue) connectivity with both amygdalar subregions (Figure 3C). We were particularly intrigued by the finding of increased dorsolateral PFC (DLPFC)–amygdalar connectivity in the GAD patients because we did not find amygdalar connectivity with this region or the posterior parietal cortex in our control cohorts. Indeed, extraction of data for the DLPFC revealed no connectivity with either subregion in the controls but consistently increased connectivity with both subregions in the GAD patients (Figure 3D). Moreover, average DLPFC-amygdalar connectivity in the GAD patients was negatively correlated with all anxiety measures, most notably with the Beck Anxiety Inventory (Figure 3E). These data therefore suggest that a lateral prefrontal executive control network is abnormally coupled with the amygdala in GAD patients. That this connectivity is strongest in the least anxious patients strongly suggests a compensatory neural adaptation.

To study amygdalar subregional connectivity in GAD, we first investigated whether the functional connectivity patterns of the BLA and CMA can be robustly discriminated using resting-state fMRI in healthy controls and whether these abide by predictions of anatomical connectivity derived from experimental animals. As with these anatomical studies, our study focused on differential connectivity, thus being sensitive to subregional differences in connectivity even if there is some anatomical connectivity of the BLA and CMA with the same targets. Anatomical tract tracing studies in nonhuman primates have determined that a wide range of cortical regions, primarily with sensory functions or in the medial PFC, provide input preferentially into the BLA and frequently receive reciprocal projections from the BLA.^12,16 Studies in rodents have generally supported this view,¹⁵ although some discrepancies exist across species. We found robust and highly reproducible differential connectivity between the BLA and the entire occipital lobe and large parts of the temporal lobe. These targets represent primary and higher-order association cortices for the visual and auditory systems, with likely equal sensitivity to afferent and efferent connections. We also found strong differential connectivity of the BLA to the ventromedial PFC, posterior orbitofrontal cortex, and parahippocampal gyrus. In nonhuman primates, the medial and orbital PFC and the entire temporal lobe are primarily connected with the BLA, although portions of these regions also have light connectivity with the central or medial nuclei.^42,43 Our data are therefore consistent with predictions from connectivity studies in experimental animals.

By contrast with animal connectivity studies, we found robust and highly reproducible differential BLA connectivity bilaterally with the primary motor and primary somatosensory cortices throughout their mediolateral course. In nonhuman primates, no evidence supports amygdalar connectivity with either region,^44,45 although in rodents limited evidence supports amygdalar connectivity with both regions.⁴⁶ Additional work will be necessary to determine whether these findings reflect species-specific differences in amygdalar connectivity, greater sensitivity of our functional connectivity method compared with anatomical connectivity methods, or connectivity of the amygdala and sensory/motor cortex via a third region.

Finally, anatomical studies in nonhuman primates have generally reported absent or very sparse connectivity of the amygdala with lateral PFC, with connectivity most notable within the posterior inferior frontal gyrus.^47,48 Within the lateral PFC, we found differential BLA connectivity most robustly also in the posterior inferior frontal gyrus. The connectivity of the BLA with this region of the lateral PFC, as well as with the ventromedial PFC, thus provides evidence of a tight relationship among these 3 regions—a relationship that is thought to be important for emotion regulation through cognitive control.^49-52

The CMA is composed of the central and medial nuclei, which are thought to constitute a closely related functional subsystem within the amygdala.¹⁷ In humans, the central nucleus represents most of the volume of the CMA,³⁸ and thus we focus primarily on its connectivity and functions, in particular in light of a role of the central nucleus in the expression of fear, regulation of arousal, and patterning of behavior^13,14 and the overall similar connectivity of these 2 nuclei.^12,15,53 Input into the central nucleus comes from a range of subcortical regions, including midline thalamic and anterior gustatory nuclei, hypothalamus, midbrain and pontine reticular formations, and viscerosensory and neuromodulatory brainstem nuclei.^12,15,53 The central nucleus projects to many of the regions from which it receives input, as well as to the pulvinar,^12,15,16,53 the mediodorsal thalamic nucleus,⁵⁴ and the effector nuclei in the brainstem.^12,15,16

Consistent with this pattern of anatomical connectivity in rodents and nonhuman primates, we found prominent differential CMA connectivity across most of the thalamus (most notably along the midline) and in the pulvinar. Activation of the midline thalamic nuclei is seen during endogenous and learned fear in rodents^55,56 and is associated with viscerosensory functions.⁵⁷ Finally, we found differential CMA connectivity with a midbrain region containing the ventral tegmental area and substantia nigra and extending posteriorly to the periaqueductal gray. Our findings of functional connectivity of the CMA with these regions are, therefore, consistent with findings of anatomical studies in animals.^12,15,16

Strong differential connectivity of the CMA was also seen with the cerebellum, most notably with the vermis—connectivity that is not generally discussed in animal tract-tracing studies. The vermis has been implicated in endogenous and learned fear,^58-60 and vermis stimulation leads to amygdalar synaptic potentials with latencies consistent with monosynaptic connections.^61,62 Our results, therefore, strengthen evidence of a role of the cerebellum—and in particular the vermis—in emotional processes⁶³ through connectivity with the CMA.

Generalized anxiety disorder has received less intense neurobiological study than other anxiety disorders, in which amygdalar hyperactivity in patients during the processing of negative emotion has been a consistent finding.⁵ The issue of whether amygdalar abnormalities exist in adults with GAD has been unresolved, in part because the amygdala's context within distributed brain networks appears to be important⁶ but has not previously been a focus of investigation. In this study, we found evidence of an intra-amygdalar perturbation in GAD and abnormal connectivity of the amygdala with at least 2 distinct brain networks.

By first focusing on differential amygdalar subregional connectivity patterns, we found that, although GAD patients still showed evidence of differential BLA-CMA connectivity, this difference was considerably less robust than in controls. Further investigation revealed that this finding was due to disorganization at an intra-amygdalar subregional level in the GAD patients. This conclusion is supported by (1) a higher correlation between BLA and CMA time series in the GAD patients than in the controls, (2) the lower connectivity of the BLA or the CMA of GAD patients to all of their normal targets and increased connectivity to all of the other subregions' targets, and (3) our finding of increased amygdalar gray matter volume in GAD patients (in particular within the right CMA) using a VBM approach with structural scans acquired from the same subjects. These findings therefore provide evidence that GAD, like other anxiety disorders, involves important alterations in the amygdala, even if these have not been consistently detected during emotional activation paradigms. Moreover, our data raise new questions about how to interpret hypoactivation or the absence of differences in amygdalar activation in GAD in light of subregional disorganization within the amygdala.

Another striking finding in GAD patients was the presence of significantly increased or decreased connectivity of several regions with both amygdalar subregions. We found significant decreased amygdalar connectivity bilaterally with the insula, dorsal/midcingulate, supplementary motor area, thalamus, caudate, putamen, superior temporal gyrus, and ventrolateral PFC. Extraction of data from the insula and cingulate revealed that, in controls, these regions were connected with both amygdalar subregions (data not shown). This connectivity pattern closely resembled a previously identified resting-state network, implicated in salience processing, which consists of limbic and paralimbic regions and of which the amygdala is normally a part.⁶⁴ Activation in these regions, most notably in the frontoinsular and dorsal anterior cingulate cortices, tends to track with sympathetic nervous system arousal.⁶⁵ Aberrant autonomic nervous system activity has been noted in GAD, and in particular a decrease in autonomic flexibility,^66,67 suggesting that decreased coupling of the amygdala to the salience network may be related to abnormalities in modulation of the autonomic nervous system. Further work, however, will be needed to explore this possibility, which at present remains speculative.

Significantly increased amygdalar connectivity was seen bilaterally with the dorsolateral, ventrolateral, dorsomedial, and ventromedial PFCs; posterior parietal cortices; and inferior occipitotemporal cortices. This pattern closely resembles the canonical frontoparietal executive control network identified in many studies of cognitive control over emotional⁴⁹ and nonemotional^68,69 material. This coordinated network has also been observed using resting-state fMRI and has been reliably dissociated from the salience network.⁶⁴

At rest, the executive control network in healthy controls does not normally include the amygdala,⁶⁴ and thus coupling of the amygdala with this network in patients likely reflects a network-level neural adaptation in GAD. Extraction of data from the DLPFC and posterior parietal cortices bilaterally confirmed that none of these regions was connected with either amygdalar subregion in controls (Figure 3; other data not shown). Amygdalofrontoparietal coupling in GAD patients may thus reflect the habitual engagement of a cognitive control system to regulate excessive anxiety. Indeed, cognitive theories of GAD suggest that the use of compensatory cognitive strategies, such as worry, reflects attempts at diminishing the impact of emotions, with which GAD patients are otherwise less able to deal.^70,71 The negative correlation between amygdala-DLPFC coupling and anxiety furthermore suggests a relationship between this adaptation and successful control of anxiety.

Aberrant coupling of the BLA and CMA with the DLPFC is also consistent with several other recent findings. Greater right DLPFC activation has been found in GAD patients during emotional processing.⁶ A magnetic resonance spectroscopy study focusing on the DLPFC found a higher N-acetylaspartate to creatinine ratio, a marker of neuronal viability, in GAD patients.⁷² Moreover, the right DLPFC has recently been successfully targeted in a small open-label treatment study of GAD using transcranial magnetic stimulation.⁷³ In sum, our findings demonstrate abnormal engagement of the amygdala in at least 2 distinct brain networks and argue that one of these networks reflects a compensatory neural adaptation, consistent with cognitive theories of GAD.

Several limitations are also important to mention. We used conventional fMRI methods in this study, and, as such, one concern may be that the limited spatial resolution of conventional fMRI may be insufficient for discriminating between amygdalar subregions. We and others^74-76 have used activation task-based approaches to demonstrate that amygdalar subregions may be functionally differentiated during emotional processing. In the present study, despite the use of conventional fMRI acquisition parameters, the BLA and CMA were associated with dramatically different connectivity patterns, which were highly replicable and consistent with predictions from anatomical studies in animals, thus strongly suggesting that fMRI is sufficiently sensitive to discriminate between amygdalar subregions.

Second, although the use of resting-state fMRI data precludes conclusions about specific ongoing cognitive processes in subjects at rest, our results provide proof that this approach has significant utility in understanding network-level abnormalities in psychiatric disorders. Although the degree to which resting-state connectivity reflects real-time, ongoing state properties vs longer-standing trait properties remains uncertain, we and others¹⁰ favor the trait interpretation for several very compelling reasons. A number of studies have examined the topography of resting-state networks during different human behavioral states, including performance of tasks,^77-81 sleep,^82,83 and anesthesia,⁸⁴ and in other species,⁸⁴ and have found that resting-state networks are largely unaltered despite the highly divergent neural and behavioral contexts being assessed or manipulated. In addition, a large number of simultaneously operating resting-state networks have been identified, and it is unlikely that ongoing mental activity could account for moment to moment fluctuations across so many different brain networks. As a result, it is likely that resting-state networks reflect intrinsic activity possibly related to synaptic maintenance in critical networks or priming of these networks to allow for rapid activation rather than transient task-related variations in activity.¹⁰ Therefore, group differences in resting-state connectivity most likely reflect group differences in brain structure and connectivity that importantly relate to symptoms, present whether or not the individual is at rest. Moreover, the groups did not differ in connectivity within a control primary auditory network, demonstrating that the group difference in amygdalar connectivity is not due to a global alteration of resting-state connectivity in patients.

Finally, given the correlational nature of fMRI, 1 possible explanation for the connectivity data is that they reflect the confounding effects of physiological variation not related to neural connectivity, arising from cardiac, respiratory, or vascular sources. There are several reasons, however, why these are unlikely to account for our findings. First, we directly accounted for the effects of global signal variation in our regression models and eliminated high-frequency sources of noise by restricting the resting fluctuation frequency band analyzed. In addition, the differentiation between BLA and CMA connectivity was heavily motivated by a large body of anatomical connectivity studies in experimental animals, the results of which were consistent with our functional connectivity findings that we replicated in 2 cohorts of healthy subjects. Moreover, we note that the highly divergent patterns of differential BLA-CMA connectivity arose from source ROIs that abut each other within the small structure of the amygdala. Ours was a study of differential connectivity, comparing the BLA with the CMA, and thus any common effects should be subtracted out in this comparison.

Correspondence: Amit Etkin, MD, PhD, Department of Psychiatry and Behavioral Sciences, Stanford University School of Medicine, 401 Quarry Rd, Stanford, CA 94305 (amitetkin@stanford.edu).

Submitted for Publication: December 5, 2008; final revision received February 11, 2009; accepted March 17, 2009.

Author Contributions: Dr Etkin 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.

Financial Disclosure: None reported.

Funding/Support: This study was supported by funds for a residency-research program of the Veterans Affairs–Palo Alto Health Care System (Dr Etkin) and grants NS048302 (Dr Greicius), HD047520 (Dr Menon), and NS058899 (Dr Menon) from the National Institutes of Health.

Additional Contributions: Eric Kandel, MD, provided helpful comments on the manuscript, and Jennifer Keller, PhD, assisted with planning of the study.