Clinical interviews

Clinician-Administered PTSD Scale for DSM-5 (CAPS-5) – Doctoral-level clinicians administered the CAPS-5 [15], the gold-standard interview to determine PTSD diagnostic status and symptom severity. We derived a CAPS-5 composite score representing the severity of all non-intrusion symptoms by summing the scores for symptom clusters C (avoidance), D (negative thoughts and feelings), and E (arousal/reactivity).

Mini International Neuropsychiatric Interview (M.I.N.I.) – Doctoral-level clinicians administered the M.I.N.I. [67] version 7.0 in order to assess DSM-5 disorders including those relevant to the exclusion criteria.

Ecological momentary assessments (EMA)

EMA was conducted using the MetricWire smartphone application (MetricWire Inc., Kitchener, Ontario, Canada). At the conclusion of the first study visit, participants received an orientation to the app and were assisted in downloading it onto their phone. Participants then were asked to complete five daily surveys over the following fourteen consecutive days, for a maximum of seventy surveys. The first daily survey was sent to participants at a random time within a three-hour window, commencing one hour before their usual wake time. The survey queried the occurrence of trauma-related nightmares the night prior. The three subsequent surveys were also distributed randomly within each consecutive three-hour block throughout the day and queried the occurrence of TR-IMs (“Since the last time you completed a survey, did you have any unwanted memories of your trauma? (If “Yes”: How many?”). Finally, the fifth daily survey evaluated the severity of PTSD symptoms experienced over the past twenty-four hours using an adapted version of the PCL-5 [66]. Each survey was available for completion for 1.5h. Participants’ compliance was monitored daily by a research assistant logging into the MetricWire website. In instances where a participant had not completed any surveys for two consecutive days, the research assistant contacted them via phone, aiming to ensure there was no technical issue preventing survey completion and/or to provide reminders for continued engagement. For this study, only the data from the three surveys assessing TR-IMs were analyzed, for a maximum of 42 TR-IM surveys over the two-week EMA period. We calculated TR-IM frequency for each participant as the total number of TR-IMs reported across all surveys during the two-week EMA period.

Imaging acquisition

Participants were scanned at the McLean Hospital Imaging Center on a 3T Siemens MAGNETOM Prisma scanner (Siemens Healthineers, Erlangen, Germany) with a 64-channel head coil. We used the Human Connectome Project (https://www.humanconnectome.org/) imaging protocol, including (a) a high-resolution MPRAGE T1-weighted sequence and (b) an ultra-high-resolution T2-weighted turbo-spin-echo sequence with slices oriented perpendicular to the long axis of the HPC (from the anterior margin of the amygdala to the most posterior part of the hippocampal tail). The high in-plane resolution of the T2-weighted sequence allows the identification of hippocampal subfields and amygdalar nuclei [68, 69]. More details on image acquisition and preprocessing can be found in the Supplementary Materials.

Study-specific probabilistic templates

Because most atlases are generated from data acquired in healthy volunteers and HPC volumes are altered in PTSD compared to healthy controls [30–33], we created study-specific aHPC and pHPC templates using the following steps: (a) preprocessing and segmentation of T1-weighted images using Freesurfer software version 7.2 (https://surfer.nmr.mgh.harvard.edu); (b) segmentation of hippocampal subfields with the hippocampal subfield segmentation module [70] in Freesurfer using both T1- and T2-weighted sequences; (c) quality check of the segmentations; (d) aggregation of the hippocampal subfields to create an anterior subregion (head) and a posterior subregion (body + tail) according to previously published aggregation scheme [71–73]; (e) registration of the T1-weighted image of each participant to the Montreal Neurological Institute 152 (MNI152) 1mm template; (f) application of the registration parameters to the aHPC and pHPC masks; (g) merging of aHPC and pHPC masks across participants to create probabilistic aHPC and pHPC templates. More details can be found in the Supplementary Materials.

In the present study, we used templates with a probability of 0.75 (i.e., only voxels present at least in 75% of our sample were retained) to avoid including inconsistent voxels (Fig. ​(Fig.1a).1a). Because the middle portion of the HPC has been associated with a mix of aHPC and pHPC functions, it was excluded from Freesurfer’s template, which includes it in the definition of pHPC. Instead, the pHPC Freesurfer template was cut according to a previously used criterion (y=−32) [45, 74]. The Freesurfer aHPC template did not need to be modified, as it already respected the previously used criterion (y=−21) [45, 74] (Fig. ​(Fig.1b1b).

Open in a separate window

Fig. 1

Study-specific templates of the anterior hippocampus (aHPC) and posterior hippocampus (pHPC).

a Illustration shows left hemisphere aHPC (red) and pHPC (blue) templates in MNI space. b pHPC template was cut at y=−32 to remove the middle portion of the hippocampus. MNI Montreal Neurological Institute.

Volumetric and structural covariance network (SCN) analyses

T1-weighted images were preprocessed using the Computational Anatomy Toolbox version 12.8.2 (CAT12; https://neuro-jena.github.io/cat/). Subsequently, an average gray matter (GM) template was created, and the probabilistic templates for aHPC and pHPC were registered to this average GM template. For more details, see Supplementary Materials.

For each participant, the volumes of aHPC and pHPC were determined by multiplying the average GM density within the subregion by the number of voxels composing that subregion and then summing left and right volumes.

Regarding the SCN analysis, a seed-based partial least squares (PLS) analysis was performed using PLSgui version 6.15 (https://www.rotman-baycrest.on.ca/index.php?section=84) in Matlab version R2022b (MathWorks Inc., Natick, MA, USA) to construct the SCN [75]. Seed-based PLS is a data-driven multivariate approach that identifies voxels across the whole brain whose GM density correlates with the GM density of a seed region across participants. This approach is suitable for analyzing large-scale SCNs and has been used previously in multiple studies to examine structural network integrity [76, 77], including studies using aHPC and pHPC as seed regions [78, 79]. One advantage of seed-based PLS is that it considers all voxels simultaneously, thereby avoiding issues associated with multiple statistical comparisons.

To create the SCN, the average GM density was computed for each participant within the bilateral aHPC and pHPC using the individual normalized, modulated but unsmoothed GM images and the respective seed templates. A between-subject matrix representing the covariance between the average GM density of the seeds and all the other brain voxels was then computed. The number of voxels included in this matrix was constrained by a binary mask derived from the average GM template generated using the individual GM images.

This matrix was then decomposed into a latent variable (LV) that identified a pattern of structural covariance. Importantly, because the focus was on identifying a pattern that distinguished aHPC and pHPC, we performed a non-rotated PLS analysis. This approach allowed the specification of an a priori contrast to identify voxels that showed significant differences in salience (i.e., weight) between aHPC and pHPC. The significance of the LV was determined through non-parametric permutation tests (n=1000) using resampling without replacement. The reliability of each voxel’s contribution to the LV was established by 1000 bootstraps with replacement, estimating the standard error of each voxel’s weight on the LV. A threshold of p < 0.05 was considered significant for the LV. A bootstrap ratio (BSR) (i.e., the ratio of each weight to its standard error) greater than 3.3 or less than −3.3 (corresponding to a p-value of 0.001) was considered reliable for each voxel. The automated anatomical labeling atlas version 3 was used to identify the brain regions included in the structural pattern [80].

Finally, a composite structural covariance score, referred to as the “brain score”, was calculated for each participant. This score represented the strength with which each participant expressed the pattern identified by the LV and was calculated as the dot product of the GM voxel density in each participant’s normalized, modulated and smoothed GM image with the corresponding voxel weight in the LV pattern.

Relationship of TR-IM frequency with aHPC and pHPC volumes

In the first Poisson regression model (Nagelkerke R²=0.092), aHPC volume was not a significant predictor of TR-IM frequency (incidence rate ratio (IRR)=0.98; p=0.492; 95% confidence interval (CI)=[0.94; 1.03]), nor were the effects of age (IRR=1.01; p=0.610; 95% CI = [0.97; 1.06]) and sex (IRR=0.94; p=0.342; 95% CI = [0.84; 0.94]). However, the total number of TR-IM surveys completed was a significant covariate in the model (IRR=0.94; p=0.008; 95% CI = [0.90; 0.98]).

The second Poisson regression model (Nagelkerke R²=0.111) revealed that pHPC volume was not a significant predictor of TR-IM frequency (IRR=1.04; p=0.123; 95% CI = [0.99; 1.09]), and that neither age (IRR=1.02; p=0.417; 95% CI = [0.97; 1.06]) nor sex (IRR=0.92; p=0.159; 95% CI = [0.82; 1.03]) were significant covariates. In contrast, the total number of TR-IM surveys completed was a significant covariate (IRR=0.94; p=0.008; 95% CI = [0.90; 0.98]).

Characterization of the pattern of structural covariance captured by the LV

PLS analysis identified a significant LV corresponding to the specified contrast (i.e., aHPC vs. pHPC volume) (p=0.013). Pearson correlations showed that LV brain scores were significantly associated with aHPC volume (r=0.60; p<0.001) and not significantly associated with pHPC volume (r=−0.11; p=0.288). Therefore, the LV brain scores represented a pattern of covariance with aHPC volume only. Brain regions showing positive covariance with aHPC included the left inferior and superior temporal gyri, bilateral superior frontal gyri, left middle frontal gyrus, bilateral fusiform gyri, bilateral precuneus, and bilateral cerebellar regions (Fig. ​(Fig.22 and Table ​Table2).2). Fewer brain regions demonstrated negative covariance with aHPC volume, with the pHPC and right middle occipital gyrus being the most prominent regions (Fig. ​(Fig.22 and Table ​Table22).

Open in a separate window

Fig. 2

Structural covariance pattern identified by the seed-based partial least squares analysis.

Color scales represent bootstrap ratios (proportional to z-scores). Brain regions that covaried positively with aHPC volume have positive bootstrap ratios (warm colors), while brain regions that covaried negatively with aHPC volume have negative bootstrap ratios (cold colors). aHPC anterior hippocampus.

Table 2

Brain regions encompassed in the structural covariance pattern identified by the seed-based partial least squares analysis.

Label	Peak MNI coordinates			Number of voxels	Bootstrap ratio
	X	Y	Z
Positive saliences
Left parahippocampal gyrus	−28	−2	−30	7740	9.40
Right hippocampus (anterior)	31	−11	−24	6236	7.79
Left inferior temporal gyrus	−51	−15	−42	1141	5.32
Left inferior temporal gyrus	−43	8	−43	775	5.31
Left insula	−25	16	−17	230	4.73
Left superior temporal gyrus	−44	−10	−7	881	4.56
Left precuneus	−11	−59	52	178	4.46
Right precuneus	6	−40	52	525	4.44
Left superior frontal gyrus, dorsolateral	−13	67	9	245	4.43
Left fusiform gyrus	−22	−49	−12	197	4.34
Right superior frontal gyrus, dorsolateral	28	63	20	343	4.33
Left middle frontal gyrus	−31	28	40	552	4.28
Left fusiform gyrus	−37	−14	−34	258	4.26
Left postcentral gyrus	−53	−9	40	182	4.09
Right fusiform gyrus	38	−16	−40	762	4.04
Left lobule IV, V of cerebellar hemisphere	−6	−60	−16	416	4.00
Right lobule IV, V of cerebellar hemisphere	20	−50	−16	359	3.92
Right lobule VI of cerebellar hemisphere	17	−70	−17	518	3.88
Left Crus I of cerebellar hemisphere	−17	−90	−23	102	3.77
Right inferior frontal gyrus, opercular part	42	8	23	102	3.70
Left Crus I of cerebellar hemisphere	−17	−71	−35	260	3.65
Negative saliences
Left hippocampus (posterior)	−26	−37	−3	1886	−8.46
Right hippocampus (posterior)	25	−37	−1	1753	−7.65
Right middle occipital gyrus	27	−90	6	328	−4.76

Open in a separate window

Labels, coordinates, and bootstrap ratios are reported for the peak voxel of each cluster. Positive saliences (i.e., weights) represent voxels covarying positively with aHPC, and negative saliences voxels covarying negatively with aHPC. Only clusters with at least 100 voxels are reported.

aHPC anterior hippocampus, MNI Montreal Neurological Institute, pHPC posterior hippocampus.

PLS analysis

The PLS analysis yielded a significant structural covariance pattern uniquely associated with the volume of the aHPC. This finding aligns with a previous study that used the same statistical approach and identified a structural covariance pattern distinguishing aHPC and pHPC in healthy volunteers [82]. These results suggest that the aHPC displays greater unique whole-brain covariance than the pHPC.

The observed structural covariance pattern encompassed multiple temporal and frontal lobe gyri, which was expected given the known interaction between the aHPC and these regions during the initial construction phase of autobiographical memory retrieval [48]. Furthermore, a significant cluster was found in the right dorsolateral prefrontal cortex, displaying significant covariance with the aHPC. This prefrontal region has been implicated in top-down inhibition of aHPC activity during attempts to suppress intrusive memories in healthy volunteers [53]. Hence, the structural pattern identified in the present study closely mirrored the functional networks associated with autobiographical memory, aligning with previous research showing that inter-regional structural covariance partly contributes to functional connectivity [57].

The SCN identified in our study also involved other brain regions, particularly multiple cerebellar regions. Interestingly, a previous study of healthy individuals examining the pattern of structural covariance with aHPC volume did not identify any significant cluster in the cerebellum [82]. Nevertheless, the presence of significant structural covariance between the HPC and the cerebellum is not surprising. A growing body of evidence highlights the cerebellum’s role in emotional and cognitive processing [83, 84], including in retrieving autobiographical memories [85]. In addition, preclinical and clinical studies have demonstrated structural and functional connectivity between the cerebellum and HPC [86, 87]. Our findings extend these observations by revealing significant positive covariance between multiple cerebellar regions and the aHPC in a PTSD sample. Further investigations in healthy volunteers will be necessary to determine whether this relationship is specific to PTSD or trauma-exposed groups, or applies more broadly.

In line with our hypothesis, the main finding of the present study revealed a significant negative association between TR-IM frequency and LV brain scores. This negative association remained significant after controlling for the severity of PTSD symptoms other than intrusion symptoms, namely avoidance, negative alterations of cognitions and mood, and arousal/reactivity. This indicates that the observed relationship is specific to trauma-reexperiencing. This finding suggests that individuals who exhibit lower expression of the identified structural covariance pattern have more frequent TR-IMs, indicating that lower structural synchronization between brain regions involved in autobiographical memory is associated with a higher occurrence of intrusive memories related to a traumatic event. Interestingly, a previous study in healthy volunteers reported a significant positive correlation between the expression of aHPC SCN and associative memory performance [82]. Notably, associative learning and autobiographical memory involve overlapping brain regions, including the HPC, which is crucial in encoding contextual information associated with the traumatic experience [88, 89]. Collectively, these findings suggest that decreased structural synchronization between brain regions involved in associative learning and memory may be linked to dysfunction of these cognitive processes. However, it is important to note that our study did not include associative memory tasks, warranting further investigations to explore the relationship between reduced aHPC structural covariance and memory performance. Moreover, our cross-sectional study design cannot determine the directionality of the relationship between structural brain covariance and TR-IMs; it is possible that lower covariance contributes to more frequent TR-IMs and/or that higher TR-IM frequency has an impact on brain covariance patterns. Further studies are needed to elucidate potential causal relationships between TR-IM frequency and decreased structural covariance of the aHPC with other brain regions involved in autobiographical memory.

Volumetric analysis

Contrary to our hypothesis, we did not uncover a significant association between TR-IM frequency and volumes of the aHPC or pHPC. Although meta-analyses have established that smaller HPC volume is a robust finding in PTSD [31–33], the relationship between HPC volume and intrusion symptoms remains unclear. Some studies have reported a significant negative correlation between HPC volume and intrusion symptom severity in PTSD adults [34–38], while others have not [39–41]. These inconsistencies may be partly explained by methodological differences across studies, including with the assessment of intrusion symptom severity and the neuroimaging techniques to segment the HPC.

Indeed, previous studies used different tools to estimate intrusion symptom severity. These have included clinician-administered interviews such as the CAPS-5 [15] and self-report questionnaires such as the PCL-5 [66]. Importantly, most of these previous studies estimated the overall severity of intrusion symptoms rather than specifically examining TR-IM frequency [34–38]. To our knowledge, only one study has directly investigated the relationship between TR-IM frequency and HPC volume [90]. Using a neurocomputational model, this study reported that the perceived danger of a traumatic event was positively associated with the likelihood of experiencing intrusive memories and that higher TR-IM frequency was linked to smaller HPC volume [90]. However, it is important to note that this study relied on simulated data rather than clinical data and did not control for potentially confounding demographic and clinical variables. Furthermore, most of the questionnaires used in previous studies include other intrusive phenomena, such as nightmares, within the reexperiencing/intrusion symptom category. These questionnaires typically asked participants to rate the level of distress or bother caused by intrusion symptoms using summary Likert scales, without directly estimating TR-IM frequency. Consequently, the specific relationship of HPC volume with TR-IM frequency, distinct from the overall severity of intrusion symptoms, has not yet been fully characterized in people with PTSD.

Regarding the segmentation techniques used for HPC volume analysis, it is worth noting that previous studies reporting no significant associations between intrusion symptom severity and HPC volume primarily employed the automatic segmentation provided by Freesurfer software [39–41]. Conversely, studies that observed significant associations mainly used manual segmentations [35–38]. In this study, we used the Freesurfer module for HPC subfields and amygdalar nuclei segmentation [70], which has been shown to provide reliable anatomical segmentations overall, except for the smallest sub-structures like the hippocampal fissure, which was not included in our aggregation scheme [91]. In addition, we used a multispectral approach known to optimize HPC segmentation, by combining a high-resolution T1-weighted sequence with an ultra-high-resolution T2-weighted sequence [92]. Therefore, the absence of significant associations between TR-IM frequency and HPC volumes in the present study aligns with previous investigations employing a similar segmentation technique [39–41].

Finally, it is important to highlight that prior studies of HPC volume in PTSD have rarely distinguished between aHPC and pHPC, leading to inconsistent findings [54–56, 93]. To our knowledge, only one previous study has investigated the relationship of intrusion symptom severity with aHPC and pHPC volumes [54]. This study reported a significant negative association with aHPC volume in a pediatric PTSD sample, but it considered an overall score of intrusion symptom severity and did not directly estimate TR-IM frequency [54]. Thus, our study contributes novel findings by demonstrating that TR-IM frequency is not significantly associated with aHPC and pHPC volumes in a sample of adults with PTSD. Together, the findings of the present study suggest that a higher TR-IM frequency is not related to the morphology of a singular brain structure but rather to the structural synchronization within a widespread brain network centered on the anterior hippocampus. In other words, a higher TR-IM frequency was not significantly associated with smaller or larger hippocampal volumes but with a lower synchronization of volumes between the aHPC and other brain regions involved in autobiographical memory: the volume of these brain regions covaried significantly less in participants experiencing more TR-IMs.

This work was supported by NIMH award R01-MH120400 (IMR). Additionally, IMR was partially supported by NIMH award P50-MH115874 (Project 4 PIs: IMR, Scott L. Rauch; PDs: William A. Carlezon Jr., Kerry J. Ressler).