Anxiety and the Neurobiology of Temporally Uncertain Threat Anticipation

Abstract

When extreme, anxiety-a state of distress and arousal prototypically evoked by uncertain danger-can be debilitating. Uncertain anticipation is a shared feature of situations that elicit signs and symptoms of anxiety across psychiatric disorders, species, and assays. Despite the profound significance of anxiety for human health and wellbeing, the neurobiology of uncertain-threat anticipation remains unsettled. Leveraging a paradigm adapted from animal research and optimized for fMRI signal decomposition, we examined the neural circuits engaged during the anticipation of temporally uncertain and certain threat in 99 men and women. Results revealed that the neural systems recruited by uncertain and certain threat anticipation are anatomically colocalized in frontocortical regions, extended amygdala, and periaqueductal gray. Comparison of the threat conditions demonstrated that this circuitry can be fractionated, with frontocortical regions showing relatively stronger engagement during the anticipation of uncertain threat, and the extended amygdala showing the reverse pattern. Although there is widespread agreement that the bed nucleus of the stria terminalis and dorsal amygdala-the two major subdivisions of the extended amygdala-play a critical role in orchestrating adaptive responses to potential danger, their precise contributions to human anxiety have remained contentious. Follow-up analyses demonstrated that these regions show statistically indistinguishable responses to temporally uncertain and certain threat anticipation. These observations provide a framework for conceptualizing anxiety and fear, for understanding the functional neuroanatomy of threat anticipation in humans, and for accelerating the development of more effective intervention strategies for pathological anxiety.SIGNIFICANCE STATEMENT Anxiety-an emotion prototypically associated with the anticipation of uncertain harm-has profound significance for public health, yet the underlying neurobiology remains unclear. Leveraging a novel neuroimaging paradigm in a relatively large sample, we identify a core circuit responsive to both uncertain and certain threat anticipation, and show that this circuitry can be fractionated into subdivisions with a bias for one kind of threat or the other. The extended amygdala occupies center stage in neuropsychiatric models of anxiety, but its functional architecture has remained contentious. Here we demonstrate that its major subdivisions show statistically indistinguishable responses to temporally uncertain and certain threat. Collectively, these observations indicate the need to revise how we think about the neurobiology of anxiety and fear.

Keywords: Research Domain Criteria (RDoC); affective neuroscience; anxiety and fear; bed nucleus of the stria terminalis; emotion; extended amygdala.

Figure 1.

MTC paradigm. As shown schematically in a, the MTC paradigm takes the form of a 2 (Valence: Threat/Safety) × 2 (Temporal Certainty: Uncertain/Certain) repeated-measures design (for details, see Materials and Methods). Subjects provided ratings of anticipatory fear/anxiety for each trial type during each scan. Skin conductance was continuously acquired during scanning. Simulations were used to optimize the detection and deconvolution of task-related hemodynamic signals (variance inflation factors <1.54). Middle panels, Structure of each trial type. Trial valence was continuously signaled during the anticipatory epoch by the background color of the display. Safety trials were similar but terminated with the delivery of benign stimuli (e.g., just-perceptible electrical stimulation). Trial certainty was signaled by the nature of the integer stream. Certain trials always began with the presentation of 30. On Uncertain trials, integers were randomly drawn from a uniform distribution ranging from 1 to 45 to reinforce the belief that uncertain trials could be much longer than certain ones. To mitigate potential confusion and eliminate mnemonic demands, a lowercase 'c' or 'u' was presented at the lower edge of the display throughout the anticipatory epoch (not depicted). As shown in b and c, threat anticipation robustly increased subjective symptoms (in-scanner ratings) and objective signs (skin conductance) of anxiety, and this was particularly evident when the timing of aversive stimulation was uncertain (Valence × Certainty, p values < 0.001; Uncertain Threat > Certain Threat, p values < 0.001). b, c, Data (black points; individual participants), density distribution (bean plots), Bayesian 95% highest density interval (colored bands), and mean (black bars) for each condition. Highest density intervals permit population-generalizable visual inferences about mean differences and were estimated using 1000 samples from a posterior Gaussian distribution. TR, the time required to collect a single volume of fMRI data.

Figure 2.

Interpretive ambiguities of canonical HRF modeling. The canonical approach to fMRI analysis models the amplitude of anticipatory activity (solid black line) under the assumption that it approximates a boxcar-like square-wave shape (dotted line; convolution of a canonical HRF with task duration). In some cases, such as the top left panel, the hemodynamic signal and the model will match. But in others, it will not. Importantly, a variety of physiologically plausible hemodynamic responses can produce similarly strong and statistically significant results (T = 52.556 in this example), highlighting the importance of modeling the BOLD signal on a finer temporal scale.

Figure 3.

Amygdala and BST ROIs. BST: The probabilistic BST ROI (green) was described by Theiss et al. (2017) and was thresholded at 0%. The seed mostly encompasses the supra-commissural BST, given the difficulty of reliably discriminating the borders of regions below the anterior commissure using T1-weighted images (Kruger et al., 2015). Amygdala: The Harvard-Oxford probabilistic amygdala (cyan) was described by Frazier et al. (2005) and Desikan et al. (2006) and conservatively thresholded at 50%. Analyses used ROIs decimated to the 2 mm resolution of the EPI data. For illustrative purposes, 1 mm ROIs are shown. Single-subject data were visually inspected to ensure that the ROIs were correctly aligned to the spatially normalized T1-weighted images.

Figure 4.

The anticipation of temporally Uncertain and Certain Threat recruits broadly similar neural systems. Key regions (cyan arrowheads) showing significantly elevated activity during the anticipation of Uncertain Threat (left column) and Certain Threat (middle column) compared with their respective control conditions. Right column, Voxels showing significantly increased activity in both contrasts. BST and dorsal amygdala images are masked to highlight significant voxels in extended amygdala (green). Coronal insets, The thresholded statistical parametric maps without the additional mask. Together, these observations indicate that these regions are sensitive to both temporally certain and uncertain threat. For additional details, see Extended Data Figures 4-1, 4-2, 4-3, 4-4, 4-5. Ant., Anterior; FDR, false discovery rate; WB, whole-brain-corrected.

Figure 5.

Regions sensitive to temporally Uncertain Threat show sustained hemodynamic activity. Mean responses to the anticipatory epoch were estimated on a TR-by-TR (1.25 s) basis for Uncertain Threat (red) and Uncertain Safety (blue) trials, using data from the local maxima of key clusters (black-and-white asterisks in the left panels) identified using a canonical analytic approach. Given the temporal resolution and autocorrelation of the hemodynamic signal, data were averaged for 4 windows (TR-1 to TR-5, TR-6 to TR-10, TR-11 to TR-15, and TR-16 to TR-24), spanning a total of 24 measurements (30 s). Broken vertical lines indicate windows. Shaded envelopes represent SEM. Ant., Anterior; TR, the time needed to acquire a single volume of fMRI data.

Figure 6.

Regions sensitive to temporally Certain Threat show sustained hemodynamic activity. Mean responses to the anticipatory epoch were estimated on a TR-by-TR (1.25 s) basis for Certain Threat (red) and Certain Safety (blue) trials, using data from the local maxima of key clusters (black-and-white asterisks in the left panels) identified using a canonical HRF GLM approach. Given the temporal resolution and autocorrelation of the hemodynamic signal, data were averaged for 3 windows (TR-1 to TR-5, TR-6 to TR-10, and TR-11 to TR-15), spanning a total of 15 measurements (18.75 s). Broken vertical lines indicate windows. Shaded envelopes represent SEM. Ant., Anterior.

Figure 7.

The threat anticipation network can be fractionated into subdivisions. The MCC, AI/FrO, and dlPFC showed greater activity during the anticipation of Uncertain Threat (left column), whereas the BST and dorsal amygdala showed greater activity during the anticipation of Certain Threat (middle column). Thresholds and other conventions are identical to Figure 4. For additional details, see Extended Data Figures 7-1 and 7-2. Right column, TR-by-TR (1.25 s) hemodynamic responses during the anticipation of Uncertain Threat (solid red line) and Certain Threat (broken red line). Data were extracted from the local maxima of key clusters (black-and-white asterisks in the left and center columns) identified using a canonical HRF GLM approach. Given the temporal resolution and autocorrelation of the hemodynamic signal, data were averaged for 3 windows (TR-1 to TR-5, TR-6 to TR-10, and TR-11 to TR-15), spanning a total of 15 measurements (18.75 s). Broken vertical lines indicate windows. Shaded envelopes represent SEM. Ant., Anterior.

Figure 8.

The BST and dorsal amygdala regions recruited by the MTC paradigm show statistically indistinguishable responses during threat anticipation. While it is impossible to demonstrate that the true difference in regional hemodynamic activity is zero, the TOST procedure provides a well-established and widely used statistical framework for testing whether mean differences in regional activity are small enough to be considered equivalent (Lakens, 2017; Lakens et al., 2018). Using the subset of voxels that were most sensitive to each threat contrast (see Materials and Methods), results revealed significant equivalence for all contrasts (Extended Data Fig. 8-1). Figure represents the data (black points; individual subjects), density distribution (bean plots), Bayesian 95% highest density interval (colored bands), and mean (black bars) for each condition. Highest density intervals permit population-generalizable visual inferences about mean differences and were estimated using 1000 samples from a posterior Gaussian distribution. Inset ring plots, Percentage of subjects showing greater activity in the BST compared with the dorsal amygdala for each contrast.

Figure 9.

Certain and uncertain threat anticipation elicit broadly similar patterns of neural activity. Figure summarizes the results of two coordinate-based meta-analyses (CBMA) of functional neuroimaging studies. Top left inset, Results for 27 “fear conditioning” studies (N = 677), highlighting regions showing consistently greater activity during the anticipation of certain threat (CS⁺ > CS^–; https://neurovault.org/collections/2472) (Fullana et al., 2016). Bottom right inset, Results for 18 “threat-of-shock” studies (N = 693), highlighting regions showing consistently greater activity during the anticipation of uncertain threat (Threat > Safe; https://neurovault.org/collections/6012) (Chavanne and Robinson, 2020). Visual inspection of the results (red clusters) suggests that the anticipation of certain and uncertain threat elicits qualitatively similar patterns, including heightened activity in the region of the BST. This impression is reinforced by the substantial correlation between the two whole-brain patterns (r = 0.69). Consistent amygdala activity was not detected in either meta-analysis. The pattern correlation was estimated in Neurovault using a brain-masked, 4 mm transformation of the publicly available, vectorized meta-analytic maps (Gorgolewski et al., 2015). For illustrative purposes, every 10th voxel is depicted in the scatter plot.