We studied patients after a non-penetrating msTBI who were recruited from four different pediatric intensive care units in Los Angeles County. Inclusion criteria included a Glasgow Coma Scale (GCS) score between 3 and 12 (or a GCS above 12 with abnormalities on clinical imaging), aged between 8 and 18, normal or corrected-to-normal visual acuity, and proficient English competence to understand instructions for the neuropsychological inventories. Exclusion criteria included prior TBI, neurologic deficits that would preclude them from completing any of the planned testing, or prior neurologic, developmental, or serious psychiatric diagnoses. For comparison, 46 age- and sex-matched healthy controls (HC) were recruited through flyers, magazines, and school postings. Controls met the same inclusion and exclusion criteria as the TBI group other than those pertaining to the brain injury under investigation. All patients who underwent MRI were also required to be eligible for scanning (e.g., no metal implants or shrapnel). The institutional review boards of University of California, Los Angeles (UCLA) and the other sites of recruitment approved this study.

The study protocol has been previously described in detail (49). In the present study, we report the first analysis focused on rsfMRI data from this protocol. Of the 50 patients with TBI recruited, 19 had rsfMRI scans from the chronic time frame (TBI: 13–22 months). Clinical and demographic details are presented in Table 1.

All patients underwent MRI scans on a 3T Siemens Trio with a 12-channel Head Matrix Coil (Siemens AG). Patients were instructed to remain still throughout the scan and foam pads were placed around their heads to further minimize head motion. For rsfMRI, T2* weighted BOLD fMRI was acquired while Patients rested with their eyes open for 7 min and 55 s, utilizing an echo-planar imaging (EPI) pulse sequence (TR = 2400 ms, TE = 35 ms, FOV = 244 mm, matrix = 80 × 80, slice thickness = 3 mm, number of slices = 40, and flip angle = 90° giving an in-plane resolution of 3 × 3 mm). For anatomical reference, T1-weighted 3D MPRAGE images were acquired (TR = 1900 ms, TE = 3.26 ms, FOV = 250 mm, slice thickness = 1 mm, matrix = 256 × 256, and flip angle = 9°, giving an in-plane resolution of 1 × 1 mm). For microstructural metrics, diffusion weighted imaging (DWI) was acquired (GRAPPA mode; acceleration factor PE 2; TR 9500 ms; TE 87 ms; FOV 256 × 256 mm; isotropic voxel size 2 mm; 72 images per subject with 8 b0 and 64 diffusion-weighted, b 1,000 s/mm²).

Processing and analysis of the MRI data were performed using the CONN Toolbox version 19b (50) and Statistical Parametric Mapping version 12 (SPM12) in Matlab 2018b. Scans were preprocessed using the default pipeline, which includes realignment, slice-timing correction, outlier identification, co-registration, segmentation, normalization and smoothing (8 mm Gaussian kernel). Functional outlier detection was performed via Artifact Removal Tools (ART) in CONN using intermediate settings (97th percentile in normative sample). After scrubbing removed volumes with framewise displacement (FD) > 0.9 mm, we removed 2 subjects from the HC group from further analyses because their scrubbed time series had an overall mean FD > 0.9 mm. All rsfMRI and T1 images were visualized for lesions that would overlap with the networks of interest; no lesions or encephalomalacia were observed in close vicinity to the amygdala network masks used. Scrubbed time series were then passed into CONN's denoising steps, including band-pass filtering (0.008–0.09 Hz) and regression of noise components generated from individually segmented white matter and cerebrospinal fluid masks (5 components for each via CompCor) as well as motion parameters and their first-order derivatives (generated from ART). From the denoised times series for each subject, we computed rFC in the a priori defined amygdala networks (Figure 1) and performed second-level analyses on these networks, as described in detail below.

DTI data were processed as described previously (51–53). Briefly, we used a tract clustering and identification method developed in our laboratory, automated multi-atlas tract extraction (autoMATE). Raw DTI images were visually checked for artifacts. DTI volumes were corrected for eddy current-induced distortions using the FSL tool eddy_correct (http://fsl.fmrib.ox.ac.uk/fsl/) and skull stripped using BET. Fractional anisotropy (FA) and mean diffusivity (MD) maps were computed using dtifit. Whole-brain HARDI tractography was performed with Camino (http://cmic.cs.ucl.ac.uk/camino/). The maximum fiber turning angle was set to 35°/voxel to limit biologically implausible results, and tracing stopped when FA dropped <0.2, as is the standard in the field. The Eve atlas was registered, linearly and then non-linearly, to each patient's FA map using Advanced Normalization Tools (ANTs) and its ROIs were correspondingly warped to extract 19 tracts of interest, of which we focused on bilateral uncinate fasciculus given it connects the frontal lobe to medial temporal lobe. Five white matter tract atlases were created from healthy controls, and each patient's FA map was linearly and nonlinearly registered to each of these in order to refine fiber extractions, the results of these five tract identifications were fused for each patient for a final, cleaned fiber clustering. Registrations at each step were visually checked for quality. MD was averaged within ROIs to generate summary measures for the medial and ventrolateral amygdala subregions, using the cluster masks previously published (38).

We used an automated segmentation and probabilistic region-of-interest (ROI) labeling technique (FreeSurfer, http://surfer.nmr.mgh.harvard.edu) to perform a quantitative morphometric analysis of the T1 MRI scans (54). Forty different brain regions are segmented and assigned neuroanatomic ROI labels using probabilistic estimations based on a manually labeled atlas dataset of 40 individuals. All ROIs were divided by the total intracranial volume to control for differing head size, as performed previously (55).

The Behavior Rating Inventory of Executive Function (BRIEF) (56) is a parent rating of “real world” behavioral manifestations of executive functioning problems. The Child Behavior Checklist (CBCL) is a 112-item parent rating scale targeting various domains of behavioral functioning (57). The current study used standardized subscales of the BRIEF and CBCL in which higher T scores indicate greater problems. On both inventories, parents rated their child's current functioning. At the first assessment, parents were also asked to retrospectively rate their child's pre-injury behaviors in the 6 months predating the TBI.

We were most interested in investigating associations between resting-state connectivity and ratings of mood and behavior regulation. As such, our primary behavioral outcome measures were Emotional Control and Externalizing subscales of the BRIEF and CBCL, respectively. We also assessed the Working Memory subscale of the BRIEF to test whether amygdala network connectivity was specific to mood and behavior regulation rather than more general to frontal regulation phenomena.

All variables of interest were assessed for normality and outliers, removing values greater than 3 standard deviations from the mean. We first compared TBI and HC groups on potential confounding variables using independent samples t-tests using SPSS, accepting alpha of 5% as significant.

For whole network analysis, we performed a series of independent first order Pearson correlational analyses in the TBI group to discern whether connectivity for the medial and ventrolateral amygdala networks predicted externalizing and internalizing problems, respectively and specifically, as hypothesized. For the externalizing behavior domain, we used two distinct measures to test for convergent validity, the Emotional Control subscale on the BRIEF and the Externalizing subscale on the CBCL. For the internalizing behavior domain, we did not have a similar measure to test for convergent validity. For both externalizing and internalizing domains, we also examined the discriminant validity of the brain-behavior findings by testing whether amygdala network connectivity correlated with a less affective domain, the Working Memory subscale of the BRIEF. We also tested the specificity of the amygdala network-behavior associations by assessing whether macro- and microstructural metrics within limbic regions and circuits, including uncinate fasciculus FA, medial amygdala MD, and amygdala volume, accounted for any of the predicted variance in the behavior domain of interest. Finally, we tested whether potential confounds accounted for any of the predicted variance, including age, sex, handedness, time since injury, and severity of injury (GCS on arrival), presence of gross injuries on hospital CT (as described in Table 1). We performed each of these tests in SPSS and considered them as independent, accepting alpha of 5% as significant.

Subsequently, we performed voxelwise analyses within the network of interest to determine voxels of greatest correlation with the behavioral measure. We performed these analyses in the CONN Toolbox (50), using the following parameters for multiple comparison correction, p-uncorrected ≤ 0.01 with cluster-size p-FDR corrected ≤ 0.05. Given that the voxelwise map of the association between medial amygdala connectivity and Emotional Control problem ratings included anterior and posterior cingulate regions that are considered part of the default mode network, we also tested whether connectivity between the anterior and posterior cingulate ROIs from the Brainnetome Atlas (58) (A32sg and A23d, respectively) correlated with Emotional Control problem ratings using the CONN Toolbox ROI to ROI regression analysis, accepting alpha of 5% as significant. Finally, we extracted Fisher r-to-z values for peak voxels from the voxelwise analyses to compare TBI and HC groups using independent samples t-tests in SPSS, accepting alpha of 5% as significant.

Patients in the TBI groups incurred injuries through a variety of mechanisms, as detailed in Table 1. From the computed tomography (CT) scan that patients received at the hospital prior to study enrollment, the types of neuropathology were detailed for each patient. The TBI group showed no significant differences from the HC group on potential confounding variables (Table 1).

Given the function of regions in the medial amygdala network (Figure 2A) – regulating affective responses to meet the goals and demands of a situation – and prior work in adolescents implicating regions within this circuit in externalizing behaviors (42, 43), we tested whether connectivity in this network correlated with the Emotional Control subscale of the BRIEF. Patients with the greatest medial amygdala connectivity were rated by their parents to have the most problems with emotional control (Figure 2B). For convergent validity, we examined an independent scale tapping a similar domain, Externalizing Behavior from the CBCL, and again found that patients in the TBI group with the highest connectivity had the highest ratings of externalizing problems (Figure 2C). None of these findings could be explained by potential confounds, including age, sex, handedness, time since injury, severity of injury (GCS on arrival), presence of gross injuries on hospital CT (as described in Table 1), uncinate fasciculus FA, medial amygdala MD, or amygdala volume, in that these variables did not correlate with both amygdala connectivity and the behavioral variables. Further, supporting the specificity of these brain-behavior relationships to the affective domain, we found no significant correlation between medial amygdala connectivity and ratings on the Working Memory subscale of the BRIEF (Figure 2D).

In a voxelwise analysis of the medial amygdala network, patients in the TBI group with the highest emotional control problem ratings had the greatest connectivity between the medial amygdala and voxels in the rostral anterior and posterior cingulate cortices (rACC and PCC, Figure 3).

Given that the resultant clusters from the voxelwise analysis are part of the canonical default mode network, we investigated whether their connectivity to one another – a good proxy for default mode connectivity – correlated with these behavioral ratings. Again, supporting the specificity of the main finding, we found no significant correlation between rACC to PCC connectivity and the Emotional Control or Externalizing Behavior scales (p = 0.33 and 0.59, respectively).

Based on prior work showing that amygdala rsfMRI connectivity to the ventromedial prefrontal cortex contributes to internalizing problems (45–47), we tested whether rFC in the ventrolateral amygdala network (Figure 4A) predicted internalizing ratings in the TBI group (N = 16). Indeed, patients with the greatest ventrolateral amygdala connectivity were rated by their parents to have the most internalizing problems (Figure 4B). This finding could not be explained by potential confounds, including age, sex, handedness, time since injury, severity of injury (GCS on arrival), or presence of gross injuries on hospital CT (as described in Table 1), uncinate fasciculus FA, ventrolateral amygdala MD, or amygdala volume, in that these variables did not correlate with both amygdala connectivity and the behavioral variables. Further, supporting the specificity of this brain-behavior relationship to the affective domain, we found no significant correlation between ventrolateral amygdala connectivity and ratings on the Working Memory subscale of the BRIEF (Figure 4C).

In a voxelwise analysis of the ventrolateral amygdala network, patients in the TBI group with the highest internalizing problem ratings had the greatest connectivity between the ventrolateral amygdala and voxels in the ventromedial prefrontal cortex (vmPFC), explaining 57% of the variance in internalizing problems (Figure 5).

Supporting the clinical significance of these brain-behavior correlations, all behavioral ratings were higher (less well-controlled) in the TBI group as compared to the HC group (Emotional Control: T = 2.26, p = 0.028; Externalizing: T = 2.34, p = 0.02), although not Internalizing (T = 0.36, p = 0.71). The reported brain-behavior associations were specific to the TBI population in that healthy controls did not show a correlation in amygdala connectivity with emotional control, externalizing or internalizing problems (Pearson correlation p-values ranged from 0.78–0.93). In addition, medial and ventrolateral amygdala connectivity with the peak voxels from the voxelwise analyses did not differ across groups (Independent t-test p-values ranged from 0.11–0.35).

This study is unique in that it combined functional connectivity from rsfMRI with measures of macro- and microstructure to investigate the mechanisms of mood and behavioral dysregulation after msTBI in adolescents. Findings from this study suggest that specific functional circuits within the amygdala's broad connectome, particularly its frontoamygdala circuitry, explain significant variance in affective and behavioral problems that are hallmarks of chronic TBI. We showed that the functional connectivity predictors of externalizing and internalizing problems were differentiable at the amygdala subregional level and could not be better explained by injury severity, gross lesion burden, time since injury, patient demographics, or structural measures of amygdala circuit integrity, diffusivity, and volume.

For adolescents in the chronic phase after msTBI, medial amygdala network rFC predicted problems in emotional control and externalizing behavior that were not accounted for by measures of structural aberrancy or injury severity. Medial amygdala connectivity to voxels in the rACC and PCC drove this relationship, predicting upwards of 48% of the variance in externalizing problems (Figure 3). This link was specific to medial amygdala connectivity to these structures rather than their connectivity with each other. The brain-behavior association was also specific to affective rather than cognitive problems.

The regions in the medial amygdala network play a critical role in integrating memories and goals with salient information from the environment, particularly the salient actions of others, to regulate affective, approach, and affiliative behaviors (39). In prior work, we found that atrophy specific to the medial amygdala network, over and above atrophy to other amygdala networks, predicts decreased empathy and warmth as well as increased disregard and sometimes cold or cruel behavior towards other people in a sample of patients with frontotemporal dementia (41).

Consistent with this role and the findings from the current study, prior work in typically developing adolescents found that elevated rFC between the amygdala and rACC predicted greater externalizing behaviors (42) Connectivity between these structures also increases through development and correlates with increases in externalizing behaviors, a finding that was unique to the amygdala's connectivity to the rostral portion of the ACC (42) The positive direction of this brain-behavior association extends to a sample of adolescents with conduct disorder and callous-unemotional traits who had increased rFC between the amygdala and a cluster extending from the rACC to the PCC as compared to typically developing controls, and increased connectivity between the amygdala and rACC correlated with higher callous-unemotional traits in the adolescents with conduct disorder (43). Additional work in adolescents with conduct disorder also links heightened amygdala rFC with rACC and PCC to greater degrees of aggression (44). These prior findings suggest heightened rFC between the amygdala and portions of the cingulate cortex may result in weaker regulatory control over externalizing behaviors, such as aggressiveness, disinhibition, or a disregard for other people.

For adolescents in the chronic phase after msTBI, ventrolateral amygdala network rFC (yellow network in Figure 1), particularly connectivity to voxels in the ventromedial prefrontal cortex (vmPFC), explained 57% of the variance in internalizing problems (Figure 5). This is consistent with a meta-analysis of studies using rsfMRI of the amygdala to predict internalizing problems, which showed that there were two distinct anterior cingulate clusters related to mood dysregulation, the more ventral cluster specific to studies of at-risk youth who had been exposed to aversive childhood events (48). Considering this prior work alongside our findings suggests that heightened rFC in frontoamygdala pathways may translate to mood dysregulation or a decreased ability to downregulate negative moods, such as anxiety, in adolescents after msTBI.

Pre-clinical work demonstrates parallel and more causative insights into the mechanisms linking connectional differences after injury with internalizing problems. Specifically, prior studies have shown cellular structural differences with increased dendritic arborization in the basolateral amygdala (59) as well as increased excitatory proteins after TBI that was associated with increased fear learning (60). Another study found increased activation in the auditory thalamo-amygala pathway during noxious auditory stimuli that was associated with heightened sensitivity and defensive behavior to the stimuli in rodents after moderate TBI but not in uninjured sham animals (61). This suggests a hyperconnected and overactive arousal circuit due to TBI and fits well with the functional and structural anatomy predicting internalizing symptoms in our human cohort.

Our study had many limitations in common with other studies in TBI but also unique to this specific cohort. In general, studies of patients with TBI suffer from a heterogeneity in injury characteristics, including different severities, mechanisms, intracranial lesions, neuropsychological history, and more. We attempted to control for many of these and other obvious demographic confounds as well as measures of macro- and microstructure that could account for the observed findings in rsfMRI. None explained our main findings. While the sample size is relatively comparable to other similar studies of this population in the literature, it is quite small, and does not allow subgroup analyses by sex, age, severity, mechanism, or other potentially confounding variables. In future work, larger samples will enable investigation of more subtle effects of potential confounds, particularly sex given its role in amygdala development and function. These findings will clearly need replication as recent work demonstrates poor replicability in brain-wide association studies involving less than thousands of subjects (62). Further discussion of this replicability problem suggests that a practical alternative to such large samples is to perform a focused study in which the phenotype of interest is deeply characterized (63). We believe one of the strengths in the present cohort is the detailed, and hypothesis driven, behavioral and imaging data that we used to define the behavioral and neural phenotype of interest while ruling out other, closely related potential confounds. Altogether, at this point, rsfMRI does not seem to be a usable measure on the individual subject level, but it does continue to explain variance in behavior between people and cohorts at the group level and may serve to guide brain-specific treatments, such as forms of neuromodulation.