Brain-Based Biotypes of Psychiatric Vulnerability in the Acute Aftermath of Trauma

Participants were recruited from emergency departments as part of a multisite longitudinal study of adverse neuropsychiatric sequelae of trauma ( 7). Twenty-two emergency departments within the Northeast, Southern, Mid-Atlantic, and Midwest regions of the United States enrolled patients within 72 hours of trauma exposure. All participants were ages 18–75, able to speak and read English, oriented to time and place, and physically able to use a smartphone, and they had possessed a smartphone for more than 1 year. Potential participants were excluded if they had a solid organ injury greater than grade 1 or a significant hemorrhage, required a chest tube or general anesthesia, or were likely to be admitted for >72 hours. MRI scans and psychophysiology data were collected a mean of 18 days (SD=6) later at a laboratory visit at McLean Hospital (Belmont, Mass.), Emory University (Atlanta), Temple University (Philadelphia), or Wayne State University (Detroit), which were each located in proximity to multiple enrolling sites. Written informed consent was obtained as approved by each site’s institutional review board.

Data collection for the AURORA study is ongoing. The discovery cohort included an initial sample that was restricted to motor vehicle-related traumas for participants with at least 8 weeks of follow-up data by March 2019 (94 patients completed the MRI visit). Data for the replication cohort were separate from the discovery cohort because these data had not yet been released by the time of the initial analysis. A second freeze and release of the survey data was broadened to include all trauma types with at least 8 weeks of follow-up data by mid-October 2019. Unique participants in this second freeze made up the replication cohort (additional participants, N=108). After quality control, 69 participants in the discovery cohort and 77 in the replication cohort were retained for analyses. The study participants’ demographic characteristics are presented in Table 1.

Trauma severity was measured using an injury severity score, as well as participants’ subjective ratings of their chances of dying. Assessments of pretrauma risk factors included a general physical health status assessment, childhood maltreatment assessment, and demographic variables. Assessments of posttrauma outcomes, including PTSD symptoms, depression symptoms, dissociative symptoms, anxiety symptoms, and impulsivity, were assessed for the pretrauma period (queried in the emergency department) and at 2 weeks (days 7–21), 8 weeks (days 46–67), 3 months (days 77–104), and 6 months (days 168–195) posttrauma. Measures and scoring details are summarized in the online supplement.

Analyses were conducted in R, version 3.6.3, with RStudio, version 1.2.1335. All tests were two-tailed and used a significance threshold of 0.05, with family-wise error correction as noted in the Results section. Clustering was conducted on data from the regions of interest extracted from the three fMRI tasks, using hierarchical agglomerated clustering, with the cluster package, version 2.1.0, following Ward’s criterion (agnes function). This is a bottom-up method designed to preserve the existing structure of the data without imposing assumptions of linearity, appropriate for exploratory analysis. The optimal number of clusters was determined using silhouette ( 31) and distance ( 32) methods. Nonparametric bootstrapping using the fpc package, version 2.2.5, was applied to the cluster solutions, with 1,000 iterations. After the initial hierarchical clustering, the data were randomly resampled with replacement. In each bootstrap, clustering was performed on the resampled data, and the new cluster most similar to each original cluster was identified by saving the maximum Jaccard coefficient (indexing similarity) for each old-new comparison ( 33). This was repeated, and a mean permuted Jaccard coefficient was computed across all the bootstraps by cluster. Permuted Jaccard coefficient therefore represents the proportion of individuals from each original cluster solution that were again clustered together in the permuted data. A permuted Jaccard coefficient of 0.6–0.75 indicates stable clusters, >0.75 represents high stability, and <0.50 is thought to indicate cluster instability ( 33); clusters were considered reconstituted on any bootstrap with a permuted Jaccard coefficient >0.60.

The replication was assessed quantitatively using a train and test approach. We trained a simple k-nearest-neighbors (knn function) model ( 34) with the class, version 7.3, package using the discovery cohort data, labeled using the cluster labels from the hierarchical clustering solution. We applied this knn model (“test”) in the replication cohort to obtain a new set of labels. We then compared these new labels to the de novo hierarchical clustering of the replication cohort in caret, version 6.0 ( 35).

Because clustering produced some small cells within-cohort (Ns as low as 11), we combined the clusters that replicated between the discovery and replication cohorts for further characterization. Cluster assignments from the initial cluster solutions were retained, rather than reclustering in a combined data set.

Chi-square tests (categorical variables) or one-way analyses of variance (ANOVAs) (continuous variables) were used to assess whether demographic factors or trauma-related factors differed between the cluster groups.

Given the multiple overlapping adverse mental health outcomes of trauma, we used multivariate analysis of variance (MANOVA) to test whether the pattern of subsequent mental health outcomes varied across the cluster groups. The outcome was a vector of standardized scores for PTSD symptoms, depression symptoms, dissociation, anxiety, and impulsivity. Predictors included cluster, assessment time point (time-invariant term for the 2-week, 8-week, 3-month, and 6-month posttrauma assessments; linear and quadratic terms), cluster-by-time point interaction, cohort, and a random effect for participant. Post hoc tests separating each outcome type were conducted using linear mixed models in the lme4 package, with the same set of predictors used in the MANOVA. To test whether the fMRI-based clustering provided incremental information above and beyond pretrauma symptom levels, we conducted secondary models including pretrauma symptoms that participants reported in the emergency department. Initial AURORA study findings indicated that among sociodemographic risk factors, pretrauma symptom levels were the strongest predictor of later PTSD and depression symptom severity ( 36, 37).

Finally, we tested whether a dimensional model of the 2-week fMRI data outperformed cluster assignment in predicting posttrauma outcomes using the first three principal components from the principal component analysis of the nine regions of interest in the combined discovery and replication data set as dimensional predictors of later outcomes. These models used the same structure as the cluster-based analyses. Model fit for dimensional and cluster-based models were directly compared.

Fear-potentiated startle during fear conditioning and extinction was also collected on the same day as the MRI scan (acquisition details are presented in the online supplement). ANOVAs tested whether the cluster groups varied in fear-potentiated startle responses during either acquisition or extinction as a function of cluster, cohort, block, and conditioned stimulus (CS) type and interactions between cluster, block, and CS type.

To assess for feature redundancy, we examined the covariance structure between the tasks and regions of interest. Regions of interest showed positive within-task covariance but not between tasks ( Figure 2A,B). The small to moderate correlations suggested that each task and region would contribute unique variance to a clustering analysis. Interestingly, similar regions were uncorrelated from one task to another; for example, participants’ amygdala reactivity to threat was not correlated with amygdala reactivity to reward (r=0.00, p=0.97). This suggests that the “crud” factor (everything correlates with everything) ( 38) was very low across this set of tasks.

Hierarchical clustering was first applied in the discovery sample, with 69 survivors of motor vehicle accidents. A four-cluster solution was identified ( Figure 2C; see also Figure S1A,B in the online supplement). Silhouette results suggested an optimal clustering with two groups (k=2) but with only a small decrement in width for k=4, whereas Hartigan’s distance metric showed an optimal gain in cluster cohesiveness at k=4. Examination of fMRI activation patterns ( Figure 2E,G) indicated that individuals in cluster 1 (permuted Jaccard coefficient=0.52) showed high reactivity to both threat and reward, with little engagement of regulatory regions in threat or inhibition. Thus, we classified the cluster 1 group as “reactive/disinhibited.” Individuals in cluster 2 (permuted Jaccard coefficient=0.54) showed threat responsivity predominated by the sgACC but low reward reactivity, and we classified this group as “low reward/high threat.” Individuals in cluster 3 (permuted Jaccard coefficient=0.52) showed no reactivity to threat, nor engagement of the vmPFC or hippocampus during inhibition, but very high responsivity to reward, and we classified this group as “high reward.” Finally, individuals in cluster 4 (permuted Jaccard coefficient=0.56) showed marked deactivation to threat in the amygdala, dACC, and insula, some activation of the hippocampus in the inhibition task, and little reactivity to reward, and we classified this group as “inhibited.”

The replication cohort included 77 participants with a variety of different trauma types, including interpersonal traumas. Here, the most favorable clustering solution included three groups, with agreement between the silhouette and distance metrics at k=3 ( Figure 2D; see also Figure S1C and D in the online supplement). The groups appeared to be consistent with cluster 1 (reactive/disinhibited; permuted Jaccard coefficient=0.53), cluster 2 (low reward/high threat; permuted Jaccard coefficient=0.73), and cluster 4 (inhibited; permuted Jaccard coefficient=0.55) from the discovery sample ( Figure 2F,H). There was a striking absence of a high reward-like phenotype; individuals who showed high reward reactivity also showed high threat reactivity.

The inclusion of higher-impact traumas may have pushed reward-responsive individuals toward higher threat reactivity. To test this, we combined both cohorts and examined effects of either injury severity or interpersonal violence on threat reactivity in the amygdala. Injury severity positively predicted amygdala reactivity (F=4.58, df=1, 144, p=0.03; see also Figure S2 in the online supplement), whereas interpersonal compared with noninterpersonal trauma did not (p=0.61). High reward was therefore likely subsumed under the reactive/disinhibited phenotype, related to higher-acuity traumas.

In the quantitative assessment of replication for clusters 1, 2, and 4, we assessed the extent to which a model trained on the clustering solution from the discovery cohort could predict the clustering solution within the replication cohort. The model trained on the discovery cohort data had 65.0% (95% CI=53.2, 75.5) accuracy in predicting the original hierarchical clustering-based labels in the replication cohort, compared with a 45.4% no-information rate (p=0.0005, kappa=0.45). This indicated that the clustering solution in the replication cohort could be recapitulated above and beyond chance levels using only the features of the discovery cohort solution.

The clusters were unrelated to demographic, health-related, trauma-related, or site-specific factors in follow-up testing (see the online supplement) and thus appeared to reflect covert neurocognitive features.

We next assessed trauma-related outcomes across the clusters in a combined sample of 125 individuals from the clusters that replicated across both cohorts (see Figure S3 in the online supplement). The clusters showed different multivariate symptom profiles posttrauma (F=2.25, df=2, 948, p=0.013) ( Figure 3A). Although assessment time point was included as a factor in the model, there was no interaction of cluster by time point on the symptom profile (p=0.82). Follow-up tests were then performed for each symptom type separately. First, there was an effect of cluster on the longitudinal model of PTSD symptoms (Wald χ ²=6.47, p=0.039), with the highest symptoms in the reactive/disinhibited cluster ( Figure 3B). The effect of cluster was reduced when pretrauma PTSD symptoms were added as a predictor in the model (cluster effect: χ ²=5.10, p=0.078; pretrauma PTSD severity effect: χ ²=20.19, p<0.001).

Second, there was an effect of cluster on the longitudinal model of anxiety symptoms (Wald χ ²=6.23, p=0.044) that was higher in the reactive/disinhibited cluster ( Figure 3E). This effect held after including pretrauma anxiety symptoms as a predictor in the model (cluster effect: χ ²=6.07, p=0.048; pretrauma anxiety severity effect: χ ²=72.37, p<0.001), suggesting that cluster information provided unique predictive value above baseline symptoms. Finally, cluster did not predict depressive (p=0.19) or dissociative symptoms (p=0.86) or impulsivity (p=0.96) ( Figure 3C,D,F).

We tested the utility of the discrete clusters against a dimensional model of the fMRI data for predicting longitudinal trajectories of stress-related symptoms. Dimensional fMRI predictors were continuous covariates reflecting threat reactivity, reward reactivity, and inhibition. Models with these covariates as predictors of later posttrauma symptoms showed negligible improvement in the model fit over the cluster-based models (see Table S3 in the online supplement). The individual fMRI dimensions were not linearly associated with any posttrauma outcome, with the exception of a negative association between the inhibition-related fMRI dimension and later dissociative symptoms (p=0.044). In models that included both clusters and dimensions competing for the variance in posttrauma outcomes, cluster assignment still predicted subsequent PTSD symptoms (p=0.012), whereas the three fMRI dimensions did not (all p values >0.05). For anxiety, neither cluster (p=0.080) nor dimensions (all p values >0.05) were significant in the head-to-head model. In summary, the dimensional model did not provide better predictive value than the cluster-based models.

On the day of the MRI scan, participants also completed a fear-potentiated startle paradigm that included fear conditioning and extinction. During fear conditioning, effects of CS (F=12.73, df=1, 468, p=0.0004) and the CS-by-block interaction (F=6.06, df=1, 468, p=0.003) suggested that fear conditioning occurred and that discrimination between the CS+ and CS− developed across acquisition. There was a significant cluster-by-block interaction (F=4.13, df=4, 468, p=0.003), such that the low reward/high threat cluster showed the highest fear-potentiated startle responses to both the CS+ and CS− at the beginning of fear conditioning, but this cluster was comparable to the other cluster groups by the end of the task ( Figure 3G). There was no cluster-by-CS interaction (p=0.82). During fear extinction, startle responses to the CS+ and CS− showed the expected decline over time (block effect: F=28.79, df=3, 623, p=0.02 × 10 ⁻¹⁴), indicating the presence of extinction learning, but there was no interaction of CS by block (p=0.64), indicating no difference in the extinction pattern for CS+ compared with CS−. This was consistent with findings from previous studies of chronic PTSD using this startle paradigm ( 39). There was again an interaction of cluster by block, indicating different rates of extinction in the different cluster groups (F=2.35, df=6, 623, p=0.03). The low reward/high threat cluster showed the highest fear-potentiated startle responses to both the CS+ and CS− at the beginning of extinction, decreasing to become comparable to the other cluster groups by the end of the task ( Figure 3H).

Finally, to identify brain regions outside the primary regions of interest included in the clustering, we conducted whole-brain analysis in the combined sample of 125 participants ( Figure 4; see also Table S4 in the online supplement), comparing the three replicated cluster groups within the threat, reward, and inhibition fMRI tasks. The reactive/disinhibited cluster showed greater activation than the other two cluster groups in a mesopontine cluster overlapping with the median raphe nucleus and ventral tegmental area, as well as the hypothalamus, dACC, and insula in response to threat cues. In contrast, the low reward/high threat cluster showed greater activation in the left and right amygdala, hippocampus, and insula in response to threat cues. The reactive/disinhibited cluster also showed greater reactivity than the other two cluster groups in the amygdala, hippocampus, and rostral anterior cingulate cortex in response to reward. The inhibited cluster showed no region of greater activation compared with the other two cluster groups.

The reactive/disinhibited cluster was the most interesting candidate as a risk-related biotype. Individuals in this cluster showed threat hyperreactivity, particularly in the insula and dACC, accompanied by high NAcc reward reactivity, as well as higher subsequent PTSD symptoms. This was partly contrary to previous findings showing that lower NAcc reward response predicted subsequent PTSD ( 16). Threat and reward reactivity have rarely been assessed concurrently in previous studies of trauma and related outcomes. However, preclinical findings indicate that interacting pathways regulate both threat and reward reactivity. For example, stress-related hyperactivity of the basolateral amygdala can directly influence NAcc function via direct efferent projections ( 42) and change reward-seeking behaviors ( 43). Participants in the reactive/disinhibited cluster also showed greater activation in the reticular nuclei (median raphe, ventral tegmental area) during the threat task. The reticular formation stimulates wakefulness and arousal ( 44, 45). The role of these nuclei in dopamine and serotonin synthesis may point toward tailored intervention opportunities for the future; early studies are exploring dopaminergic modulation to treat PTSD ( 46).

The fMRI features observed in the low reward/high threat cluster included moderate responsivity to threat dominated by sgACC activation, along with markedly low reactivity to reward. Reduced reward responsivity in the NAcc, amygdala, and OFC is characteristic of major depressive disorder ( 47), as is sgACC hyperreactivity to sadness-inducing stimuli. The heightened fear-potentiated startle shown by this group during early fear acquisition and extinction is consistent with patterns previously observed in comorbid PTSD and major depressive disorder ( 27), which is more common posttrauma than each disorder alone ( 48). This cluster group also showed greater threat reactivity in the amygdala, insula, and hippocampus in whole-brain analyses compared with the other cluster groups. Together, the findings suggest the possibility that PTSD-related symptom groups may be divided into a low reward/high threat group driven more by cortical function, and a reactive/disinhibited group, driven more by brainstem nuclei.

Participants in the inhibited cluster appeared to be most consistent with active coping, with low threat reactivity accompanied by relatively high vmPFC and hippocampus engagement during inhibition. Individual features of this pattern have previously been associated with resilience. For example, lower amygdala threat reactivity predicts lower future PTSD symptoms ( 12, 13). Similarly, greater vmPFC and hippocampal activation during inhibition has been associated with resilience ( 22, 26). Our findings indicate that some individuals show the combined profile in the acute posttrauma period. However, this was the smallest cluster, and it is possible that inhibition as an adaptation to the stress of the index trauma had not yet fully emerged by 2 weeks posttrauma. As data collection continues, the AURORA study will include additional neuroimaging 6 months posttrauma, allowing a window into the further development of these profiles.

Participants in the high reward cluster showed high reward reactivity, low threat reactivity, and low inhibition, a pattern suggesting preserved positive affect in the context of low top-down regulation. However, this pattern may only be observed when the emotional impact of trauma is relatively low, and this cluster was not observed in the replication cohort.

To reduce participant burden, symptom assessments were abbreviated and based on self-report. This limited our ability to directly compare the outcomes to gold-standard assessments of psychiatric disorders. However, future work could apply these biotypes to archival data with interview-based assessments to extend our findings. Additionally, with sample sizes of 69 participants in the discovery cohort and 77 in the replication cohort, the study may be underpowered. However, we are encouraged that three clusters were replicated, indicating generalizability even across different types of trauma.