Abstract
Naturalistic observations show that animals pre-empt danger by moving to locations that increase their success in avoiding future threats. To test this in humans, we created a spatial margin of safety (MOS) decision task that quantifies pre-emptive avoidance by measuring the distance subjects place themselves to safety when facing different threats whose attack locations vary in predictability. Behavioral results show that human participants place themselves closer to safe locations when facing threats that attack in spatial locations with more outliers. Using both univariate and multivariate pattern analysis (MVPA) on fMRI data collected during a 2 h session on participants of both sexes, we demonstrate a dissociable role for the vmPFC in MOS-related decision-making. MVPA results revealed that the posterior vmPFC encoded for more unpredictable threats with univariate analyses showing a functional coupling with the amygdala and hippocampus. Conversely, the anterior vmPFC was more active for the more predictable attacks and showed coupling with the striatum. Our findings converge in showing that during pre-emptive danger, the anterior vmPFC may provide a safety signal, possibly via foreseeable outcomes, while the posterior vmPFC drives unpredictable danger signals.
Significance Statement
A common observation in nature is that under conditions of uncertain danger, animals will stay close to safety—a behavioral metric known as spatial margin of safety. We adapt this metric to examine risky and safety decisions to unpredictable attack distances. Using multivariate and univariate fMRI, we demonstrate a novel dissociation of vmPFC in decision-making: the posterior vmPFC encoded for the more unpredictable threat and showed functional coupling with the amygdala and hippocampus, while the anterior vmPFC was more active for more predictable attacks. Our findings suggest that when pre-empting danger, the anterior vmPFC may provide a safety signal associated with predictable outcomes, while the posterior vmPFC may drive uncertain danger signals.
Introduction
Staying close to safety is a key antipredator behavior as it increases the likelihood of future escape success (Lima, 1985; Mobbs et al., 2020). One metric used by behavioral ecologists to measure safety behavior is spatial margin of safety (MOS), where prey adopt locations that increase the likelihood of escape (Martindale, 1982; Lima, 1985; Wetterer, 1989). In turn, this provides the prey with a safety net, while also reducing stress, and energy consumption and promotes increased focus on other survival behaviors, such as feeding. Humans appear to use safety distance in similar ways. For example, when human subjects are placed close to a safety exit, measures of fear decrease, and when under threat, the sight of safety reduces fear and fear reinstatement (Christianson et al., 2008, 2011; Eisenberger et al., 2011). Here, we test the idea that when subjects are pre-empting threats of varying attack location probabilities, they will vary their spatial MOS decisions depending on predictability. We propose that MOS decisions reflect prospective spatial planning, which involves estimating safety by calculating the predator's attack locations (Cooper and Blumstein, 2015).
The prospective nature of spatial MOS decisions may elicit activity in a set of neural circuits involved in anxiety (Adhikari, 2014), which can be defined as a future-oriented emotional state and involves the behavioral avoidance of potential dangers. Two drivers in this spatial avoidance are the ventromedial prefrontal cortex (vmPFC) and the hippocampus (Adhikari, 2014; LeDoux and Pine, 2016; Mobbs, 2018; Mobbs et al., 2018; Mobbs and LeDoux, 2018; Qi et al., 2018). For example, the hippocampus is believed to play a key role in anxiety and guides decisions via memory and prospection (Benoit et al., 2014). Further, synchronization between the hippocampus and vmPFC is associated with anxiety-like behaviors (Adhikari et al., 2010; Padilla-Coreano et al., 2016; Fung et al., 2019), suggesting that the hippocampus is involved in signaling the threat significance of a stimulus. The vmPFC is a heterogeneous structure involved in information seeking, anticipation, and the organization of defensive and safety responses (Adhikari et al., 2010; Dixon et al., 2017; Iigaya et al., 2019).
Research has also shown that a safety stimulus during an aversive experience results in increased activity in the anterior vmPFC while decreasing threat also results in increased activity in the same region, suggesting that the anterior vmPFC may emit safety signals (Eisenberger et al., 2011; Åhs et al., 2015). Research also shows that attention set to safety signals, extinction, and downregulation of anxiety are associated with vmPFC activity, suggesting that it is a key node in what has been called the fear suppression circuit (Wilkinson et al., 1998; Xu et al., 2016; Sangha et al., 2020). Conversely, the posterior vmPFC, encompassing the subgenual and rostral anterior cingulate cortex (sgACC and rACC), receives dense projections from the amygdala (Amaral and Insausti, 1992) and is implicated in negative affective responses and behavioral expression of fear (Mobbs et al., 2007, 2010; Mobbs, 2018). How these and other brain regions are evoked during pre-emptive MOS decisions is yet to be tested.
To address these gaps in knowledge about human defensive circuits, we created a task to investigate spatial MOS decisions under uncertainty and elucidate: (1) How do changes in the threat's attack predictability, threat intensity, and reward value impact the subjects’ MOS decisions? And (2) do the hippocampus and vmPFC encode characteristics of threats that are central to MOS decisions? To create less predictable attack positions, we used leptokurtic distributions, which are evolutionarily novel and have been shown to increase the level of uncertainty and are more difficult to learn (d’Acremont and Bossaerts, 2016). Leptokurtic distributions are probability density curves that have higher peaks at the mean and are fatter-tailed where extreme outcomes (outliers) are expected more (Fig. 1C). We contrasted this with standard Gaussians (Fig. 1D,E), which are more computationally familiar to humans.
Experimental structure. A, During the MOS decision task, participants were first presented with a series of information screens at the beginning of every 10 trials (one trial block), displaying the reward/shock level, color of the predator (leptokurtic, where a kurtosis is added to the normal distribution resulting in heavy tails, red; normmatch , where the variance is matched with the leptokurtic condition, green; normhalf , where the variance is half as compared to the leptokurtic condition, blue ), before being asked to rate their confidence in escaping. In the low shock/reward conditions, participants receive 1 shock and the base reward, respectively. In the high shock/reward conditions, participants received two shocks and twice the base reward. B, On every trial, for the first 4 s, participants were presented with a screen displaying the margin of safety runway and their initial location. They were told to make a choice of where to put themselves during this phase. However, they were not able to actually move in this phase. After a 4 s jitter, they were presented with the same screen again where they can move to the desired MOS with a varying bar of reward meter depending on the MOS location. In the next 2 s, the outcome of the chasing was revealed, including whether their escape was successful and how much reward was gained. C, Attack distributions for leptokurtic uncertainty; D, Gaussian distribution with matched variance, and E, half the variance Gaussian; F, the predator’s attack distances through all trials. Zero on the y-axis marks the mean of the distribution, while numbers represent how far away the drawn instance is away from the mean. G, Escape probability. The x-axis represents possible margin of safety choices, while the y-axis represents the corresponding probability of escape. H, Schematic representation of the experimental procedure. Participants undergo 4 × 30 min scans sessions over a 2 d period.
Materials and Methods
Experimental methods
We tested 24 subjects who were recruited according to the guidelines of the California Institute of Technology Institutional Review Board after providing informed consent. Data from two subjects were lost due to incomplete scanning sessions. Our final sample consisted of 22 subjects (10 females; age, 24.3 ± 8.1 years). Although an N = 22 might be low for traditional fMRI studies (e.g., experiments with short acquisition time), we decided to make sure we had more dense data with enough power for MVPA. For each individual participant, we have a total scan length of 2 h for each participant. Rather than increasing our sample size to maximize power, we instead minimize the noise of each subject's estimates, which has the same effect of increasing statistical power. This practice of long individual scans makes it different from traditional short length samples. Power analyses (Mumford and Nichols, 2008) on N = 24 resulted in above 80% power. The use of N = 22 is estimated to result in 80% power in the vmPFC (power, 80%; effect size, 0.31). All power calculations are based on a region of interest (ROI) using the regions implicated in previous studies (vmPFC) with an α = 0.001.
Attack distributions
An important property of frequency distributions is their kurtosis, which refers to their degree of peakedness. If a curve is more peaked than the normal distribution, it is said to be leptokurtic (Ibe, 2014). Compared with the normal distributions we used in the other conditions, a leptokurtic distribution is more peaked and has a flatter tail, making it more difficult to estimate the true mean. Leptokurtic distributions are also implicated in real-life scenarios of poor decision-making. For example, random changes in outcome variability in financial markets constantly produces frequent and salient outliers, leading investors to make wrong assumptions about the market and even causing financial crisis (d’Acremont and Bossaerts, 2016). The fat tails caused by the outliers in the leptokurtic distribution is a rare type of noise in humans’ natural environment (Bai et al., 2003), where most noise is either Gaussian or in its domain of attraction. Thus, in our paradigm, we define the leptokurtic condition to be more uncertain/less predictable than the normal distributions, when they have the same parameters in variance and mean.
To ensure that (1) the uncertainty of the attack distributions holds and (2) the outliers that help shape the distributions are what's driving differential behavioral and neural responses, the experiment is designed in a way where participants cannot view the movements of the predators and in a way where it would be unrealistic to “gauge” the probability of getting shocked by randomly placing themselves instead of trying to guess the distributions. Since the predator is invisible, and moves fast, such gauging will result in lots of shock trials and will then force the experiment to end (a fail-safe mechanism to both protect the participants and filter out participants who did not follow the structure of the task). More importantly, there were extensive instructions and training sessions with feedback to make sure that the participants were making MOS decisions based on the uncertainty distributions.
Participants were instructed and practiced beforehand that the attacks would follow certain distributions and were told to specifically form a hypothesis of these distributions to guide decisions. We collected feedback afterward and only included participants who followed the structure and design purpose of the task.
Stimuli, apparatus, and procedure
A complete pipeline of experimental procedures can be found in Figure 1. Participants completed a computer-based task while in an fMRI scanner. The task was set under the scenario where subjects place themselves at a desired location toward a safety exit while facing a potentially dangerous predator. The closer they place themselves to the safety exit, the more likely they will be able to escape from the predator after the trial starts, but the low the resulting reward will be. The goal of the task was to earn as much money as possible while avoiding being caught by the virtual predator. Prior to the beginning of the trial, the participants were presented with a 2 s cue indicating one of the three different predator types that would be presented in the upcoming trial. These predators differ in the location they speed up. These locations correspond to three distributions—a leptokurtic distribution, a normal distribution with matching variance, and a normal distribution with only half of the variance. (In the rest of the paper, we refer to the normal distribution with matching variance as normmatch and the normal distribution with half variance as normhalf.) The participants were then shown a two-dimensional runway (90 units distance, where a unit is the smallest increment on the runway), with a triangle icon representing the position of the participant toward the end of the runway (at 80 or 0 units distance, depending on which direction the trial goes. A random starting location is then assigned based on which direction they start), and a circle icon representing the position of a predator at the left side of the runway (at 1 unit distance). This predator had two distinct modes of movement.
In “approach” mode, the predator would proceed rightward along the runway at 4 units per second. At a randomly chosen distance (i.e., the attack distance) according to the leptokurtic, normmatch and normmatch distribution, the predator would switch to “chase” mode, at which point it would advance at 10 units per second. The position where it switches to the “chase” mode is drawn either from the leptokurtic, normmatch or normhalf distributions depending on the actual attacking condition. Before the abovementioned chasing sequence starts, the participants were told to make a decision of where they want to start by pressing left or right arrows, to move from their randomly assigned initial location to a location they desire. The direction of the chase was counterbalanced by adjusting the relative location of the predator, participant, and the safety zone so that half of the chase was from the left to the right, and the other half were opposite. After participants responded with their preferred margin of safety choice (MOS choice), they skip the actual animation of the chase (which was shown in full during the practice session) and were shown the final result of the trial: whether they got caught or not and how much reward they earned.
The experiment starts with the subjects being shown that if captured, they will receive one or two shocks and high or low reward if they escape (Fig. 1B). They will then be presented with one of three different colored spheres, each representing different attack distributions of the virtual predators. They will then be asked to rate how confident they are of escape. Next, the subject will be asked to make safety decisions by either staying or switching to a riskier position that is further away from the safety exit or stay or move closer to the safety exit. To motivate risky decisions, the subject will acquire more money if they are riskier (i.e., move further from safety), which follows a simple linear relationship as a function of MOS choice (10 cents minimum, 20 cents maximum). They will then be asked to move the cursor to the decided safety position. After a jittered ITI, the subject will observe the outcome. If caught, they will receive a shock(s) and lose their money on this trial. This will repeat for another nine trials, before the subject is introduced to a new set of reward and shock contingences as well as a new virtual predator. The virtual predator attack distribution is either (1) normal distribution with half variance, (2) leptokurtic (positive kurtosis with fatter tails), or (3) normal distribution with matched variance with the leptokurtic distribution. Leptokurtic distributions are rare in the natural environment, where distributions are often normally distributed and easier to learn.
A total number of 460 trials (400 experimental trials and 60 control trials) were administrated throughout four sessions (two sessions per day with 2 d). The computer task was programmed in Pygames with Python.
fMRI data acquisition
We collected the fMRI data using a 3 T Prisma scanner in the Caltech Brain Imaging Center with a 32-channel head receive array. BOLD contrast images will be acquired using a single-shot, multiband T2*-weighted echoplanar imaging sequence with the following parameters: TR/TE, 1,000/30 ms; flip angle, 60°; 72 slices; slice angulation, 20° to transverse; multiband acceleration, 6; no in-plane acceleration; 3/4 partial Fourier acquisition; slice thickness/gap, 2.0/0.0 mm; FOV, 192 mm × 192 mm; matrix, 96 × 96. Anatomical reference imaging will employ 0.9 mm isotropic resolution 3D T1w MEMP-RAGE (TR/TI/TE, 2,550/1,150/1.3, 3.1, 4.0, 6.9 ms; FOV, 230 mm × 230 mm) and 3D T2w SPACE sequences (TR/TE, 3,200/564 ms; FOV, 230 mm × 230 mm). Participants viewed the screen via a mirror mounted on the head coil, and a pillow and foam cushions were placed inside the coil to minimize head movement. Electric stimulation was delivered using a BIOPAC STM100C.
Data analysis
All statistical analyses for the behavioral data were carried out in R, using the packages “ezANOVA,” “coxme,” and “lme4”. Where appropriate, the Greenhouse–Geisser corrections were performed to account for violations of sphericity, and the correction factor values and original degrees of freedom are reported. Partial eta-squared effect sizes are reported only for significant analyses. Where appropriate, we corrected for multiple comparisons using the Holm–Bonferroni method.
Analysis of fMRI data was carried out using scripted batches in SPM8 software (Wellcome Trust Centre for Neuroimaging) implemented in Matlab 7 (The MathWorks). Structural images were subjected to the unified segmentation algorithm implemented in SPM8, yielding discrete cosine transform spatial warping coefficients used to normalize each individual's data into MNI space. Functional data were first corrected for slice timing difference and subsequently realigned to account for head movements. Normalized data were finally smoothed with a 6 mm FWHM Gaussian kernel.
Multivariate pattern analysis were performed using PyMVPA (Hanke et al., 2009), where we looked at both the representational similarities (Popal et al., 2019) and searchlight decoding (Etzel et al., 2013).
We extracted the beta values associated with experimental conditions of all the voxels in each ROI, removing the mean intensity for each multivoxel activity pattern. For each participant, the brain response pattern analyses of classification training and testing with linear support vector machines (SVMs) were conducted using a leave-one-run-out cross-validation procedure. Furthermore, to evaluate whether predator type contrast reliably modulates brain response patterns across sessions, with data collected in distinct runs, we performed cross-validations that use run-1 data for training and run-2 data for testing and vice versa. ANOVAs were then conducted to compare classification accuracies. The construction of the decoding ROIs and the RSAs are described below.
To localize the functional ROIs, a whole-brain searchlight was first performed to identify brain regions representing the MOS decision information, where a classifier predicting each trial's association with one of the six MOS decision category was constructed. For each voxel in native space, we built a spherical ROI (radius, 6 mm) centering on the voxel, extracted t values in this ROI to each of the 50 MOS decisions, and calculated one minus Spearman rank correlations of all decision pairs within this ROI to construct a neural RDM. The relationship between the neural RDM and the theoretical RDM was then assessed using partial Spearman correlation, which produced a correlation coefficient for this voxel. Moving the searchlight center throughout the cortex, we obtained a whole-brain r-map in the native space. Note that the searchlight analysis was restricted to the voxels with a probability higher than 1/3 in the native gray matter image generated from the segmentation step. For a group-level random-effects analysis, the r maps in the native space were Fisher z-transformed, normalized to the MNI space using the forward deformation field, and spatially smoothed using a 6 mm full-width at half maximum Gaussian kernel. Clusters surviving the cluster-level FWE correction at p < 0.05 were reported. For each subject, we then identified the voxels whose neural RDMs showed a significantly positive correlation with the RDM in the abovementioned searchlight analysis. These voxels together with their adjacent voxels within a 6-mm-radius sphere were considered as individual subjects’ functional ROI (Fig. 2B–E).
Behavioral results. Choice frequencies for (A) leptokurtic, (B) matched variance, and (C) half variance attacking threats. The MOS decision phase and the outcome. D, Confidence ratings for leptokurtic distribution, matched variance normal distribution, and normal distribution with half variance. Post hoc analysis revealed that participants were less confident in the leptokurtic condition compared with the other two conditions (p < 0.001). Leptokurtic attack locations are in red; normal distribution with matching variance is in green; and normal distribution with half variance is in blue.
Classification accuracy
To explore the regions involved in the decision-making process under threat within the MOS framework, we examined MVPA classification accuracies using both whole-brain searchlight analysis and ROI analysis. We extracted voxel-wise fMRI responses to MOS trial (decision phase) as classification samples. For each participant and each run, we designed a general linear model (GLM). The GLM contained three regressors indicating the decision phases (duration, 4 s) of the three distribution types, as well as four regressors indicating the indication phase (duration, reaction time), motor phase (duration, 4 s), and feedback phase (duration, 3 s). All the regressors were convolved with a canonical hemodynamic response function. In addition, the motion-correction parameters and their first temporal derivatives were included as regressors of no interest to account for motion-related artifacts. To account for excessive motions, we calculated the mean framewise displacement (FD) for each participant and excluded those with an FD >0.5 mm from further analysis. For each voxel, the parameter estimates of the three regressors corresponded to the fMRI responses to each of the three distributions in each run. The fMRI responses to each distribution item were then entered into the classification analysis as classification samples.
Naturally, there are two main questions we prioritized. First, what brain regions are involved in determining which distribution type the participant is facing and second, what brain regions are involved in determining the MOS decision the participant is making. Thus, we used two sets of classification labels corresponding to the two questions: (1) normmatch, normhalf, and leptokurtic distribution (2) the 50 possible discrete MOS choice options.
We employed a linear SVM with a cost parameter C = 1 as a classifier. Classification accuracy was estimated using a leave-one-run-out cross-validation: for each of the four runs, a classifier was trained on the other three runs and tested on the remaining focal run; and the procedure was repeated for the four runs (accuracy scores were averaged).
To validate whether the classification performance was significantly above chance, we further conducted Monte Carlo permutation-based statistical tests. This method entailed running a classification analysis 1,000 times with randomly permuted experimental condition labels, allowing us to construct null distributions that were used to examine whether a classification accuracy was significantly above chance at an α of p < 0.05.
Univariate analyses
We also ran a univariate analysis pipeline to decompose the neural circuits employed when facing uncertain and stable threats. Preprocessed imaged were subjected to a two-level GLM using SPM8. The first level contained the following regressors of interest, each convolved with the canonical two gamma hemodynamic response function: a 2 s box-car function for the onset of the trial (where the color of the incoming attack is shown); a 4 s (duration jittered) box-car function for the decision period; and a 4 s box car (function for the time window where participants actually select their starting positions). In addition, regressors of no interest consisted of motion parameters determined during preprocessing, their first temporal derivative, and discrete cosine transform-based temporal low frequency drift regressors with a cutoff of 192 s. To account for excessive motions, we calculated the mean FD for each participant and excluded those with an FD >0.5 mm from further analysis.
Beta maps were used to create linear contrast maps, which were then subjected to second-level, random-effects one-sample t tests. In addition, a flexible factorial model was used to examine the main effects of attack type, reward level, and shock level. Interaction effects between attack type, reward level, and shock level were also examined using the factorial model. The resulting statistical maps were thresholded at p < 0.05 corrected for multiple comparisons [false discovery rate (FDR) corrected].
Connectivity analysis
Based on the key regions obtained during MVPA searchlight analysis, we further performed connectivity analysis using gPPI (gPPI; http://www.nitrc.org/projects/gppi), which is configured to automatically accommodate more than two task conditions in the same PPI model by spanning the entire experimental space, compared with the standard implementation in SPM8.
Power analysis
The fMRI-based power analysis employed novel methods developed Mumford and Nichols (2008) that are implemented in the fMRI power software package (fmripower.org). This method estimates power for detecting significant activation within specific regions of interest, with the assumption that the planned studies will have the same number of runs per subject, runs of the same length, similar scanner noise characteristics, and data analysis with a comparable model. The effect sizes have been expressed in standard deviation (SD) units, which is analogous to Cohen’s D measure. All power calculations are based on a ROI using the regions implicated in previous studies (vmPFC, hippocampus, MCC) with a p value threshold of 0.001 for a one-sided hypothesis test.
With 22 subjects, we will have at least 80% power to detect an effect size of 0.31 (SD units) in the vmPFC.
With 23 subjects, we will have at least 80% power to detect an effect size of 0.45 (SD units) in the hippocampus.
With 20 subjects, we will have at least 80% power to detect an effect size of 0.30 (SD units) in the MCC.
Results
MOS choices are less risky in the less predictable threat environment
MOS choice in the task represents the position participants selected relative to the safety refuge. In order to investigate how uncertainty of predator attack modulates MOS choices, we first examined how MOS decisions vary across distributions types, with a repeated-measures, one-way ANOVA. The result showed a main effect of distribution type (F(2,44) = 61.33; p < 0.001). A Tukey post hoc test revealed that participants’ MOS choice was significantly closer to the safety zone in the leptokurtic distribution condition (0.74 ± 0.06) than that in the normmatch condition (0.68 ± 0.03) and normhalf condition (0.67 ± 0.01). This indicates that participants perceived leptokurtic distributions as riskier, resulting in overall safer choices. Interestingly, there was no significant difference in mean MOS choices between the two normal distributions. This suggests that only a fundamental change in the statistical structure of the target distribution can impact participants’ decision under threat, rather than a change in the variance of the distribution (Fig. 3A–D).
Neural representation of pre-emptive MOS decisions. Avoidance decisions decoded in the vmPFC and hippocampus. A, Whole-brain searchlight map displaying regions representing information from the MOS choice classifier, trained to predict subjects’ MOS choices (FDR corrected; p < 0.05), indicating regions that encoded the upcoming MOS choice. B, Classification accuracy of the MOS choice classifier across subjects. Each dot represents data from a single subject. Average accuracy was significantly higher than the simulated chance level (p < 0.001). Box and whisker plots display accuracies from the same classifier, at various regions of interest [the hippocampus, vmPFC (posterior), vmPFC (anterior), and MCC]. C, In the hippocampus, classification accuracy from all three attack conditions was significantly higher than their corresponding chance levels. D, Classification accuracy was only significantly higher than chance in the leptokurtic distribution in vmPFCpost. E, Classification accuracy was only significantly higher than chance in the normmatch condition in vmPFCant. F, Behavioral similarity structure among MOS choices, as used in the representational similarity analysis. The behavioral similarity structure represents how similar MOS choices are on the behavior level. For example, MOS choices 1 and 2 are closer in distance compared with choices 1 and 6, thus more similar in the structure. Naturally, choices are more similar when in close spatial distance and more dissimilar when in sparse spatial distance. G, Actual pattern similarity within the regions of interest. The neural RDM in the hippocampus was significantly correlated with the theoretical model (leptokurtic: t = 0.593, p < 0.001; normmatch : t = 0.403, p < 0.001; normhalf : t = 0.377, p = 0.012). Similar correlation effects were also found in (H) vmPFCpost and (I) vmPFCant (vmPFCpost, leptokurtic condition, t = 0.754, p < 0.001; vmPFCant , normmatch condition, t = 0.482, p < 0.001). The RDMs were obtained in ROIs identified through the MVPA searchlight.
MOS choices are less risky in threat environment with higher punishment
To further disentangle how shock and reward levels could interact with predator attack type as additional external incentives, we examined participants’ MOS choices within high/low shock conditions and high/low reward conditions. In the low shock/reward conditions, participants receive one shock and the base reward, respectively, whereas in the high shock/reward conditions, participants received two shocks and twice the base reward. While there was no significant difference in their MOS decisions when facing different levels of rewards (t(21) = 1.378; p = 0.182), their MOS choices were significantly more conservative in the high shock condition (0.75 ± 0.07), compared with the low shock condition (0.69 ± 0.05): t(21) = 21.21; p < 0.001. This suggests that participants were sensitive to the level of danger and adjusted their MOS decisions accordingly (Fig. 4).
Psychophysiological interactions seeding from regions of interest and meta-analytical decoding. A, Example of Brodmann areas (BA) that distinguish the posterior–anterior axis. For example, the posterior vmPFC reflects BA 25, 24, 32(ACC), 10m, and 14, while the anterior encompasses BA 10p, 10 r 11, and 32 (non-ACC). This is made clearer by the dotted line. Connectivity analyses were first performed on the anterior and posterior vmPFC seeds, which are 6 mm spheres centered on the peak voxel of the corresponding clusters in the MVPA searchlight. B, For the posterior vmPFC seed, in all three attacking conditions, the connectivity maps showed significant activation in the hippocampus (leptokurtic: p < 0.001, T = 4.06; normmatch: p < 0.001, T = 3.62; normhalf: p = 0.011, T = 3.18). Interestingly, only in the leptokurtic attacking condition, the amygdala was found significant on the connectivity map (p < 0.001, T = 4.60). C, On the other hand, with the anterior vmPFC seed, all three attacking conditions showed significant connectivity toward the caudate (leptokurtic: p < 0.001, T = 3.87; normmatch: p < 0.001, T = 4.23; normhalf: p < 0.001, T = 4.59). We constructed two parametric modulators indicating whether the participants’ final MOS choices is a (D) safety choice or a (E) risky choice (compared with their randomly assigned initial location). The parametric modulation of univariate data thus reveals what regions were associated with risky/safety choices under different levels of predictability. On the resulting statistical maps, using SVCs from the previously constructed vmPFCpost and vmPFCant ROIs, we found that the “move to danger” and “move to safety” modulations were significant in the vmPFCpost and vmPFCant ROIs, respectively (move to danger: p < 0.001, T = 6.44; move to safety: p < 0.001, T = .39). F, Meta-analytical decoding with Neurosynth. Red and green radar bars represent correlation strength between key words and the anterior (x = 0; y = 26; z = −12) and posterior (x = −2; y = 46; z = −10) vmPFC ROIs.
More confident participants made riskier MOS decisions
Having shown that uncertainty in the attack distribution influences observable behavior, we asked whether it also affects subjective confidence in escape success. We collected participants’ confidence ratings before every unique trial block (shown in Fig. 1A,B; every 10 trials consist a unique trial block). An ANOVA on the confidence ratings also revealed that participants were generally more confident on trials in the two normal distributions compared with trials in the leptokurtic distribution. A main effect of distribution type was found (F(2,44) = 27.32; p < 0.001), and a Tukey post hoc test showed that confidence rating in the leptokurtic condition (1.42 ± 0.42) was significantly lower than those in the normmatch condition (2.43 ± 0.68) and the normhalf variance (2.65 ± 0.62; p < 0.001; Fig. 3E). We also examined the relationship between participants’ MOS choices and confidence ratings. Interestingly, a significant correlation was only observed in the leptokurtic condition, where individuals who were more confident made riskier MOS choices (r = −0.54; p = 0.04). This effect was not observed for either the normmatch condition (r = 0.25; p = 0.37) or the normhalf condition (r = −0.31; p = 0.27).
MOS decisions are represented within the prefrontal and subcortical regions
Building on our behavioral results, we next sought to identify neural systems underlying MOS decisions in response to varying levels of attack uncertainty. Due to the design feature of the behavioral experiment, the decision phase consists of both a cognitive (perception of the threat) and decision component, making the univariate analysis insufficient to capture the underlying parametric of the neural process (Norman et al., 2006; Davis et al., 2014). The MVPA analysis here thus serves two main purposes: to identify the key regions involved in decision-making under the current threat and to distinguish the underlying neural mechanism among threats with different levels of uncertainty. Results of this analysis can then be used to inform ROIs for subsequent connectivity and parametric modulation analysis. To accomplish this, we used a searchlight cross-decoding approach using linear support vector regression (SVR) and leave-one-run-out cross-validation (Extended Data Fig. 1).
Extended Figures
Download Extended Figures, DOCX file.
Two separate whole-brain searchlight analyses were performed to answer the following questions, respectively: which regions are critically involved in processing (1) different attacking distributions and (2) MOS choices. The first classifier predicted which distribution type a given trial belonged to (see Extended Data Fig. 2 for analyses pipeline). This showed that regions including the right insula and the mid-cingulate cortex (MCC) encoded the distribution type, with a decoding accuracy significantly higher than the Monte-Carlo simulated chance level accuracy (overall accuracy: t(21) = 2.82; p = 0.010). The whole-brain decoding map was thresholded at p < 0.05 (FWE; Fig. 2A).
Next, for the analysis of MOS decision types, each trial was labeled according to the MOS decision the subject made, and a classifier was trained to predict which trials fall into which decision categories. The categories were created by grouping MOS choices that are close in spatial distance together (i.e., the entire MOS choice runway is divided to six segments from left to right). Each choice category thus represents a level of how close participants place themselves to the safety. Decoding of choices was found in regions including the right hippocampus, vmPFCpost, and vmPFCant with a decoding accuracy significantly higher than chance level (t(21) = 2.47; p = 0.022). These results suggested that both the distribution type and MOS decision-making process are robustly represented in the abovementioned regions (Fig. 2A,B).
vmPFC subregions differentially encode MOS decisions according to uncertainty
The regions implicated in the whole-brain searchlight overlap with ROIs in previous literature shown to be critically involved in the process of decision-making under threat. We thus performed MVPA analysis within each ROI, namely, the hippocampus, vmPFCpost, and vmPFCant to investigate how they uniquely contributed to the MOS decision process. Within each specified ROI, we investigated classification accuracy for the MOS decisions labels, separately for each distribution conditions. Thus, by comparing how well the process is decoded within each ROI, we can examine how the involved regions drive behavioral change depending on levels of uncertainty in different predator conditions.
Within the vmPFCpos, only choice decoding for the leptokurtic condition was significantly above the Monte-Carlo simulated chance level (Monte-Carlo simulated baselines: leptokurtic, 36.7%; normmatch, 34.8%; normhalf, 33.7%; leptokurtic distribution, p < 0.001; normmatch, p = 0.410; normhalf, p = 0.868). Within the vmPFCant, only classification for the normmatch condition was significantly above chance level (leptokurtic distribution, p = 0.341; normmatch, p = 0.004; normhalf, p = 0.156). Within the hippocampus, classification for all three distribution types was significantly above chance level (leptokurtic distribution, p < 0.001; normmatch, p = 0.011; normhalf, p = 0.038). A follow-up ANOVA did not reveal a significant difference among the decoding accuracies (Fig. 2B–E).
Univariate overlap with vmPFC regions involved in “fear” and “extinction”
To compare the activated regions with past studies, we constructed ROIs from neurosynth using the key words “fear” (for comparison with posterior vmPFC/sgACC) and “extinction” (for comparison with vmPFCant, ROIs were constructed using 6 mm spheres from the peak coordinate). Small volume corrections were performed. Extinction maps were used as we hypothesized that the extinction and reduced threat would overlap. We then performed SVC with the “fear” ROI on vmPFCpos with the leptokurtic contrast [p < 0.001; T = 5.07; cluster size = 31, (0, 26, −12)] and SVC with the “extinction” ROI on vmPFCant [p = 0.010; T = 4.35; cluster size = 11, (−2, 46, −10)]. The SVC was performed within the entire neurosynth map, which included regions beyond the overlap between the vmPFC defined by the MVPA analyses and the neurosynth map. For a full list of activated regions, please refer to Table 1. These coordinates overlap with the corresponding ROIs taken from the searchlight analysis, indicating that information processing and learning through both fear and safety are potentially presented in MOS decision-making through vmPFCpost and vmPFCant, respectively.
Activation table for SVC corrections
vmPFC activity encodes MOS decisions
Having demonstrated that vmPFC activity patterns encode MOS decisions, the next step was to ask whether overall BOLD activity levels in the vmPFC also covaried with MOS decision (Extended Data Fig. 2). To test this, we constructed two univariate parametric modulators indicating whether the participants’ final MOS choices is a safety choice or a risky choice (compared with their randomly assigned initial location). The parametric modulation of univariate data thus reveals what regions showed activity associated with risky/safety choices under different levels of predictability. Inspection of the resulting statistical maps, using SVCs from the previously constructed vmPFCpost and vmPFCant ROIs, showed that the “move to danger” and “move to safety” modulations were significant in the vmPFCpost and vmPFCant ROIs, respectively (move to danger: p < 0.001, T = 6.44; move to safety: p < 0.001, T = 4.39; Table 2).
Activation Table for Parametric modulation with escape choices
Representational similarity analysis of the vmPFCpost, vmPFCant, and hippocampus
The MVPA searchlight analysis offers insights into what key regions are involved in coding MOS decision process. However, it is left unclear how different MOS choices were actually neurally represented in the ROIs mentioned above. Thus, a representational similarity analysis was conducted to investigate the underlying geometry of the neural encoding of the MOS decision variables in the abovementioned ROIs.
A behavioral RDM (representational dissimilarity matrices), together with RDMs from the neural data within the hippocampus, vmPFCpost, and vmPFCant were constructed to investigate the potential MOS decision information and perceived distribution information embedded in the activity patterns of these ROIs.
A high level of similarity between the theoretical structure and the actual brain activity in a certain ROI will indicate that task-relevant information is encoded in a way that is consistent with the behavioral structure of the during the MOS decision process. Figure 2 illustrates the theoretical/behavioral RDMs constructed by the pairwise relations of the 50 MOS decisions. Distinctive clustering in the RDM structure also helps further validate the original behavioral paradigm, showing how sensitive participants were to all the possible MOS choices. Spearman correlation coefficients were used to calculate the distance between the model and neural data matrices. The neural RDM in the hippocampus was significantly correlated with the theoretical model (r = 0.593; p < 0.001) across all conditions. Similar correlation effects were also found in vmPFCpost and vmPFCant (r = 0.754, p < 0.001; r = 0.482, p < 0.001), but these were specific to the leptokurtic and normmatch conditions, respectively (Fig. 2F–I).
Converging evidence from the previously mentioned searchlight, univariate parametric modulation, and RSA analysis has shown that the vmPFC subregions (vmPFCpost and vmPFCant) play a role in the encoding of MOS decisions under environments with different levels of predictability. Next, we further investigate the connectivity structure seeding from these regions.
Differences in vmPFCpost and vmPFCant connectivity
With vmPFCpost and vmPFCant identified as key regions associated with risky and dangerous choices, we were interested in how these regions regulate MOS decisions in concert with subcortical structures. To test this, we performed connectivity analysis using gPPI (see Materials and Methods), to reveal regions that showed covarying activity with our vmPFC seed regions. From the MVPA analysis, we took the vmPFCpost and vmPFCant as seed regions for the leptokurtic distribution contrast and normal distribution contrasts, since they were identified as regions representing the process where participants make risk decisions under the corresponding predator conditions. Importantly, as these seed regions were defined based on their multivariate representations (i.e., relative activity levels across voxels) rather than overall activity levels, this analysis represents an entirely orthogonal test. PPI analyses were first performed on the moving to safety/danger contrast, respectively, on the vmPFCpost and vmPFCant ROIs (Extended Data Fig. 2B,C). For the vmPFCpost seed, in all three attacking conditions, the connectivity maps showed significant activation in the hippocampus (leptokurtic: p < 0.001, T = 4.06; normmatch: p < 0.001, T = 3.62; normhalf: p = 0.011, T = 3.18). Interestingly, only in the leptokurtic attacking condition did the amygdala show significant coupling with the vmPFCpost (p < 0.001; T = 4.60). On the other hand, with the anterior vmPFC seed, all three attacking conditions showed significant connectivity toward the caudate (leptokurtic: p < 0.001, T = 3.87; normmatch p < 0.001, T = 4.23; normhalf p < 0.001, T = 4.59). For a full list of activated regions, please refer to Table 3.
Activation table for PPI analysis
Discussion
We found evidence in support of the hypothesis that less predictable predatory attack locations with outliers result in participants adjusting their distance to be closer to safety (Mobbs et al., 2015). We also show that when encountering more uncertain threats, participants decreased confidence in escape success. MVPA analysis shows that the vmPFCPost is associated with the avoidance of more uncertain threats and consequently the decision to stay closer to safety. The vmPFCPost also showed increased functional coupling with the hippocampus and amygdala, supporting the known connectivity with this region as well as its role in the control of fear (Nili et al., 2010; Mobbs and Kim, 2015). On the other hand, the vmPFCAnt was associated with more certain attack locations and thereby executing safer decisions. These results are congruent with the idea that vmPFC subregions play distinct roles in both danger and safety signals that reflect action plans in the context of positive or negative outcomes with a threat.
Our results also suggest that when the attack location is relatively predictable (i.e., normmatch and normhalf Gaussian distributions), participants make riskier MOS choices. That is, participants choose to place themselves further away from the safety exit to earn more reward. On the other hand, when the attack location is more unpredictable (i.e., leptokurtic distribution), participants tended to place themselves closer to safety and thus displayed more protective actions. Critically, despite significant differences in variance, there were no differences in MOS decisions between the two Gaussian distributions. This suggests that participants’ decision patterns facing uncertain threats were not swayed by a simple change in distribution variance but by a total structural change in the predictability of the distribution. While this may appear intuitive, this is not a trivial distinction; subjective first-order outcome uncertainty is typically assumed to be equivalent to the variance around an estimate (Bach and Dolan, 2012) and has been a focus of investigation in studies of psychopathology (Wise and Dolan, 2020); however, our results indicate that subjective uncertainty, and resulting avoidant behavior, is not dependent upon variance but on the structure of the outcome distribution. This was echoed in participants’ subjective rating of their confidence, a reflection of how likely they felt they were to escape (Fig. 3E).
When dissecting the defensive circuitry, it is critical to understand which brain regions are involved in the avoidance of forthcoming danger. Our MVPA searchlight identified three key regions, namely, the hippocampus, the vmPFCPost, and the vmPFCAnt. Interestingly, when looking at the classification accuracies, we found that within the vmPFCAnt, classification accuracy was above the chance level only for the normhalf, in line with our prediction that this region would be involved in the most predictable attack locations. On the other hand, within the vmPFCPost, the classification was more accurate than the chance level only for the more unpredictable, leptokurtic distribution condition. This suggests a separation of vmPFC subregions in terms of functional roles. While the vmPFCAnt is correlated with more predictable decision environments, the vmPFCPost seems to be associated with more unpredictable counterparts. Interestingly, the hippocampus classification accuracies revealed no differences between attack location distributions, suggesting a more general role in avoidance decisions.
The vmPFCPost may function as a hub when the environment is more uncertain and where more information gathering is needed. Further evidence for this comes from our parametric modulation analysis using relative MOS from the starting position, which showed that more dangerous choices are associated with activation in the vmPFCPost. This suggests a tentative role for the vmPFCPost to be responsible for computations concerning a more unpredictable environment or a riskier choice. In our connectivity analysis seeding from the vmPFCPost, we observed activations in the amygdala and hippocampus only in the uncertain attacking locations. Previous research has shown a role for the amygdala-mPFC as a pathway for modulating threat avoidance behavior and the hippocampus as a center for representing predictive relationships between environmental states (Lisman and Redish, 2009; Stachenfeld et al., 2017). This is in line with the idea that for decision-making under threat with less predictability, more predictive computations are required. Notably, while we did not directly observe a differential encoding within the amygdala, the posterior vmPFC's significant coupling with the amygdala during uncertain attacking suggests a more nuanced bridging role of the amygdala in optimizing behaviors under the MOS.
The vmPFCAnt modulates behavior when the environment is relatively easy to predict during the spatial MOS decisions. Interestingly, using relative MOS from the starting position as a modulator in the parametric modulation analysis, the vmPFCAnt was also activated when the choice is categorized as “safe.” In previous studies, this region has been implicated in both safety learning through extinction and safety learning through active avoidance (Eisenberger et al., 2011; Harrison et al., 2017). For example, studies using the lever press avoidance task in rodents have shown activation of the prelimbic regions of MPFC (the rodent homolog of human anterior vmPFC) during the expression of active avoidance (Bravo-Rivera et al., 2015; Diehl et al., 2018). These regions partially overlap with the identified clusters of vmPFCAnt in our task. Further, when looking at functional connectivity seeding from the vmPFCAnt, the caudate was significant only in the two more predictable predator conditions, although there may be other explanations (action selection; Lau and Glimcher, 2007). This resonates with previous studies where vmPFC not only functions as a center for signaling safety but also in reward-related processes, because safety processing may be “intrinsically rewarding or reinforcing” (Eisenberger et al., 2011). This is also supported by a parametric modulation analysis showing that shifts toward safety activate the vmPFCAnt. Also involved in this process is the striatum, which has been shown to be responsible for fear memory extinction (Maren and Quirk, 2004; Alexander et al., 2019). For example, previous research on rodents has shown that in rats, the dopamine level in the striatum was unchanged after exposure to novel environmental stimulus but follows more closely to the expression of conditioned responses (Wilkinson et al., 1998). Importantly, this orchestrates with our finding where the striatum is only responsive to the high predictability threats together with the vmPFCAnt. Finally, it is important to note that the vmPFC is also involved in reward processes including reward value (Chib et al., 2009) and the rewards of avoiding an aversive outcome (Kim et al., 2006). While the current study might lack the mechanics to do so, hence effect from the reward levels is absent, future research should aim to dissociate these reward processes from safety.
The hippocampus also emerged as a central region involved in MOS decisions. First, decoding of choice was higher than chance level in the hippocampus, regardless of how uncertain the attacking locations were. However, the hippocampus only showed functional connectivity with the vmPFCPost in the uncertain, leptokurtic attacking condition. The first finding resonates with the idea that the hippocampus has long been thought of as a predictive map and center for planning when considering future actions based on immediate feedback from the environment (Lisman and Redish, 2009; Bach et al., 2014; Stachenfeld et al., 2017). It was thus universally involved regardless of the uncertainty level of the attacking environment. However, our results indicate that activity in the hippocampus becomes more coordinated with the vmPFCPost in situations that require more intensive planning, as evidenced by the distinct functional connectivity to the hippocampus when the subjects are encountering a more unpredictable, leptokurtic, attacking threat. Indeed, our finding corresponds to previous studies using rodents where the hippocampus has been shown to specifically contribute to model-based planning that may include also memory-based decision-making (Miller et al., 2017).
In sum, the current study offers the first insight into how spatial MOS decisions are determined in threat environments with different levels of predictability. It also establishes the posterior and anterior vmPFC subregions as centers modulating the push and pulls between risky and safe choices, where the hippocampus is involved in both processes in a more universal manner. More work is needed to further validate the functional separation of vmPFC subregions in terms of their roles during decision-making under threat. These new insights, however, suggest a dissociable role of the vmPFC in anxiety, where the vmPFCPost is involved in heightened threat signals, while the vmPFCAnt may be involved in downregulation of threat via safety signals.
Footnotes
The authors declare no competing financial interests. D.M. is supported by an award NIMH R01MH133730.
- Correspondence should be addressed to Dean Mobbs at dmobbs{at}caltech.edu or Song Qi at sqi{at}caltech.edu.
