Neural patterns differentiate traumatic from sad autobiographical memories in PTSD

Abstract

For people with post-traumatic stress disorder (PTSD), recalling traumatic memories often displays as intrusions that differ profoundly from processing of ‘regular’ negative memories. These mnemonic features fueled theories speculating a unique cognitive state linked with traumatic memories. Yet to date, little empirical evidence supports this view. Here, we examined neural activity of PTSD patients who were listening to narratives depicting their own memories. An intersubject representational similarity analysis of cross-subject semantic content and neural patterns revealed a differentiation in hippocampal representation by narrative type: Semantically similar sad autobiographical memories elicited similar neural representations across participants. By contrast, within the same individuals, semantically similar trauma memories were not represented similarly. Furthermore, we were able to decode memory type from hippocampal multivoxel patterns. Finally, individual symptom severity modulated semantic representation of the traumatic narratives in the posterior cingulate cortex. Taken together, these findings suggest that traumatic memories are an alternate cognitive entity that deviates from memory per se.

Introduction

Personal memory is at the core of post-traumatic stress disorder (PTSD)¹. Research on the neural mechanism of PTSD has largely focused on non-personal basic learning and memory paradigms². It is yet unclear how traumatic memories differ from negative non-traumatic autobiographical memories. Is a traumatic memory an exceptionally strong manifestation of autobiographical memory or a different neural representation altogether?

To examine this, we need to factor-in differences across individual traumatic narratives and the idiosyncratic experiences they evoke, and extract from them the common markers operating in trauma-driven state. With this in mind, we designed a study that examines PTSD patients’ neural responses to their own personal traumatic memory in the form of a structured, fully annotated, audio narrative. We compared traumatic memory, within each participant, to a negatively-valenced non-traumatic sad memory, and a calm positive memory.

Previous research has established the role of the hippocampus in the construction of relational cognitive maps, onto which events are bound across space and time to form episodic memories^3–5. It is through this tracking of sequences of events that the hippocampus is necessary for the construction of a narrative from discrete events^6,7. In turn, the hippocampus also governs the ensuing retrieval of such events^8,9. The hippocampus is in fact so central to maintenance of episodic memory that lesioning it results in grave deficits in mnemonic abilities, to the point of global anterograde amnesia in humans¹⁰.

Impairments to hippocampal processes are focal to PTSD pathophysiology¹¹. Evidence suggests that PTSD is associated with structural abnormalities (predominantly a reduction in volume), as well as reduced functional connectivity between the hippocampus and other regions of the default mode network (DMN) during rest^12,13. In the context of encoding of the traumatic memory itself, aberrations in hippocampal function are thought to contribute to the paradoxical mnemonic sequelae commonly observed in PTSD—difficulty in voluntary coherent recall alongside with detailed involuntary intrusions of the traumatic memory^14,15. All in all, our understanding of the impact of PTSD on the spectrum of hippocampus-mediated mnemonic processes, is still murky¹⁶.

Emotional memories—episodic memories that elicit emotions at retrieval—often engage the amygdala¹⁷. Functionally, the amygdala is considered one component in a broader neurocircuitry model implicated in PTSD, comprised of the hippocampus and medial prefrontal cortex¹⁸. Generally, the amygdala shows hyperresponsivity to both non-specific threat-related stimuli¹⁹ and personal trauma reminders in PTSD^20–22. Amygdala activation during memory encoding modulates the memory’s explicit subsequent strength²³, evaluated through its persistence, accuracy, and vividness²⁴. Whether the amygdala itself serves as an ‘auxiliary’ site to support the storage of emotional memory is still debated^25–27.

Here, we examined whether and how the brain differentiates between traumatic and sad autobiographical memories. We hypothesized that across PTSD patients, semantic similarity would correspond to neural similarity: if the personal memories of two participants are semantically close, their patterns of neural responses while listening to the audio recording of these memories should be similar as well. If traumatic and sad memories are just different cases of autobiographical memories, we should observe semantic-to-neural correspondence across pairs of traumatic memories and pairs of sad memories alike.

However, if traumatic autobiographical memories depart from—rather than being a version of—sad autobiographical memories, then we would observe the semantic-to-neural relationship only for sad, but not traumatic memories. Our hypothesis further suggests that a shared neural representation will allude to a shared underlying neural mechanism. If the effect does not extend to pairs of traumatic memories across participants, despite their semantic similarity, this may imply that traumatic memories deviate from the neurotypical mechanisms of other sad, non-traumatic, autobiographical memories.

If traumatic memories differ from ‘normal’ sad memories in their hippocampal representation, what might be their alternate representation? Based on qualitative depictions, traumatic memories are not experienced as memories per se, but more as a present experience. They are felt in the moment rather than as a remote event in the past^28–30. Traumatic memories are also more difficult to construe as a coherent narrative^31,32. Indeed, a great deal of psychotherapy is geared toward reconstructing the traumatic event as a narrative that is embedded within life-long memories, attempting to distance the past trauma from the current safe present¹⁵. Such felt internal reexperiencing, which is momentarily disconnected from the external world and is not strictly embedded in time and space, may be supported by the default mode network, particularly its central hub, the posterior cingulate cortex (PCC)^33–35. The PCC has been demonstrated to be heavily implicated in both narrative comprehension and autobiographical memory processing^33,36, and particularly in emotional memory imagery³⁷. Alterations in PCC function and connectivity are specifically focal in PTSD^38–40.

Considering this evidence, we hypothesized that if a mechanistic difference between traumatic and sad memories exists, it would be detected in hippocampal and PCC neural patterns. If traumatic memories diverge from normal memory representation, we expect semantic-to-neural mapping in the hippocampus to emerge only for sad but not traumatic memory. For traumatic memory, we set out to ask whether the relation between semantic and neural representation will emerge in the PCC given evidence supporting both its roles autobiographical memory processing and narrative comprehension, and conversely – evidence for dysfunction in PTSD. As a control comparison, we did not expect such differentiation, or any pattern representation related to the semantic content of memories in the amygdala, given the role of this region primarily in signifying emotional saliency rather than semantic content.

To test these hypotheses, we examined neural activity of PTSD patients who were listening to narratives depicting their own memories: traumatic, sad, and calm. We used an intersubject representational similarity analysis of semantic content and neural patterns across participants in order to reveal a differentiation in neural representation by narrative type. Such differentiation will indicate that traumatic event representation is a cognitive state that deviates from memory per se.

Results

Twenty-eight participants (age = 38.2 ± 10.4 years, 11 females) diagnosed with PTSD (CAPS score = 41.2 ± 8.3), underwent reactivation of autobiographical memory through script-driven imagery while undergoing functional magnetic resonance imaging (fMRI). First, in order to generate stimuli which are based on participants’ autobiographical memory, we used an imagery development procedure. Participants elaborated on three types of autobiographical memories: 1) the ‘PTSD’ condition: the traumatic event associated with their PTSD (DSM-5 criterion A; common examples were combat, sexual assault, domestic violence), 2) the ‘Sad’ condition: a sad meaningful, but non-traumatizing experience (common examples were death of family member or pet), and 3) the ‘Calm’ condition: a positive, calm event (common examples were memorable outdoor activities). These highly personal and variable depictions of autobiographical memory were then systematically arranged into an approximately 120-second audio clip (referred to henceforward as ‘script’ or ‘narrative’, interchangeably), narrated by a member of the research staff. All scripts were composed with ample attention to a common rigid structure, into which the individual autobiographical memory was incorporated. Notably, ‘PTSD’ and ‘Sad’ narratives were scripted to maximize their structural similarity to control for content and arousal (see Methods). Participants listened to this novel rendition of their autobiographical memory for the first time while undergoing functional magnetic resonance imaging (fMRI) (Figure 1a).

Figure 1: Experimental paradigm and semantic analysis framework.

(a) Experimental paradigm for script-based autobiographical memories reactivation. At some point prior to enrollment, an event was perceived, and an autobiographical memory was formed. This memory has since then undergone an unknown number of recalls and reconsolidation iterations. During the study, participants once again recalled this memory, this time while verbalizing it as part of the imagery development procedure. These recollections were incorporated into a novel narrative-form rendition of the autobiographical memory, which was played to the participants for the first time while undergoing functional imaging, in order to reactivate this autobiographical memory once again.

(b) Semantic similarity of autobiographical narratives using word embedding. Each indexable word in the script was assigned a 300-dimensional vector representation (e.g., ‘apartment’, ‘lay’ etc.). Sentence vectors were represented as the average of word vectors comprising them (e.g., tokenized e.g., ‘apartment lay couch watch television’) and scripts represented as the average of the sentences in them. Pairwise semantic similarity across participants was calculated using cosine similarity.

We provide word-cloud plots depicting word occurrence per each individual script (Figure S1) as well as bi-gram word-cloud plots clustering phrases common to each script type (‘Calm’, ‘Sad’, ‘PTSD’) (Figure S2).

Basic group inference in cognitive neuroscience, and neuroimaging in particular, relies on the detection of shared stimulus-induced signals of greater amplitude than the noise or idiosyncratic signals in these systems³⁶. Autobiographical memories are rarely generated in a controllable lab setting, and therefore differ in their content. However, at the same time, autobiographical memories recall may elicit common cognitive states (e.g., mental time travel) that are potentially subserved by common neural substrates across individuals. In this study, the sensory stimuli used for reactivation were based on idiosyncratic experiences in order to invoke a common cognitive state – the reactivation of traumatic autobiographical memories. During recruitment participants were not screened for a specific trauma type. This enabled us to span a wide range of themes, some of which were present in ‘PTSD’ and ‘Sad’ conditions. For example, a narrative describing the death of a loved one can meet PTSD criterion A classification for one participant and thus be associated with a traumatic autobiographical narrative yet be regarded as ‘Sad’ autobiographical memory (i.e., non-traumatizing) for another. Therefore, assigning a continuous, parametric relationship to similarity between autobiographical memories may allow comparison of neural representation of relatable autobiographical memories in light of their clinical outcomes., and critically – to compare the representation of relatable autobiographical memories in light of their clinical outcomes – ‘PTSD’ or non-traumatizing.

Semantic analysis of similarity in autobiographical memory

To quantify similarity between autobiographical memory-based narratives across individuals and conditions we applied a word embedding approach – a computational linguistic tool used to quantify distances between text-based semantics⁴¹. In brief, words are pre-sampled from gigantic text corpora and are then embedded in a high-dimensionality space according to local co-occurrences. The derived semantic space allows one to infer relational structure between concepts according to their distance. Such tools were previously used to uncover neural representations of semantic spaces⁴² both with functional imaging⁴³ and invasive recordings⁴⁴.

We used MATLAB’s word2vec with a pre-trained embedded space for one million words in the English language⁴⁵. Each word was assigned a 300-dimensional vector representation. In our analytical hierarchy, sentence vectors were represented as the average of word vectors comprising them. Similarly, scripts were represented as the average representation of their sentences (Figure 1b). The high dimensionality of the semantic dataset is difficult to interpret visually. We therefore applied t-distributed Stochastic Neighbor Embedding (t-SNE), a method for dimensionality reduction, to the data to cluster narratives based on the semantic similarity of their content and projected this dimensionality-reduced dataset onto a three-dimensional space.

We observed that both types of negatively-valenced narratives – ‘PTSD’ and ‘Sad’ – formed overlapping clusters in semantic space, whereas ‘Calm’ narratives were grouped in a separate part of the space (Figure 2a). Additional 2D projections of the semantic space are available in (Figure S3). This qualitative visualization affirmed that semantic content of ‘PTSD’ and ‘Sad’ autobiographical memories are comparable and thus ‘Sad’ scripts are poised to provide a valid control for the ‘PTSD’ scripts.

Figure 2: Semantic similarity across script types.

(a) Clustering of semantic similarity across script types using t-SNE. t-SNE embedding of scripts. Each dot represents a single script, projected onto a 3D space. Colored volumes are continuous areas in space occupied by each script type. Color denotes script type (‘PTSD’: red, ‘Sad’: blue, ‘Calm’: gray). Note overlap of ‘PTSD’ and ‘Sad’ semantic content. Text adjacent to data points are general titles of narrative content, generated by the researchers (abbreviations: Fam.: family, Sex. sexual. Mil. military). Some titles were omitted to prevent clutter.

(b) Pairwise similarity of semantic content of scripted narratives. Semantic pairwise cosine similarity presented for all scripted narratives. Within-category similarity is marked in colored triangles off the matrix diagonal (‘PTSD’: red, ‘Sad’: blue, ‘Calm’: gray). Across-category similarity is marked in colored circles in distinct square sectors of the matrix (‘PTSD:Sad’: yellow, ‘Calm:Sad’: magenta, ‘Calm:PTSD’: teal). Text labels next to matrix rows are general titles of narrative content, generated by the researchers.

(c) Semantic similarity across script types. Raincloud plots illustrating the distribution of pairwise semantic similarity between pairs of script types. Next to each raincloud is a box plot portraying the same data (N = 784). Median-centered boxes show 25^th and 75^th quantiles, whiskers extend to 2^nd and 98^th quantile (ANOVA: F(2,2351) = 1094.75, p < 1E-300. η_p² = 0.325, post-hoc two-sided two-sample t-tests (Tukey correction): ‘PTSD:Sad’: t = −7.43, p < 0.001; ‘Calm:Sad’: t = −43.15, p < 0.001; ‘Calm:PTSD’: t = −35.71, p < 0.001). *** p < 0.001.

We next measured semantic similarity using cosine similarity between the 300-dimension vectors representing the scripts. The resulting semantic similarity matrix was comprised of the three script types (‘PTSD’, ‘Sad’, ‘Calm’) of 28 participants yielding a 84 X 84 matrix in total. Script types were grouped to aid in visualization of ‘type’ clusters off the diagonal (Figure 2b).

We applied a second, complementary method of semantic contextual embedding to the scripts. This transformer model – BERT, implemented in MATLAB – was able to provide representations which were shaped by intra-sentence dynamics (Figure S4a,b). Both resulting semantic similarity matrices were highly correlated (Pearson’s correlation r = 0.88, p < 0.00001, N =3,486, Figure S4c), suggesting both models were able to capture similar semantic concepts.

In addition to dimensionality reduction, we calculated cross-category similarity and observed that the ‘PTSD’ and ‘Sad’ narratives showed a higher cross-category semantic similarity than other cross-category comparisons (Similarity (r): ‘PTSD’:’Sad’ = 0.059 ± 0.159, ‘Calm’:’Sad’ = −0.274 ± 0.147 ‘Calm’:’PTSD’ = −0.224 ± 0.150; ANOVA: F(2,2351) = 1094.75, p < 1E-300. η_p² = 0.325) (Figure 2c). This semantic resemblance can be attributed to the themes shared by both negatively valenced scripts, as well as the systematically controlled structure and the shared phrases that were used.

In a complementary analysis we quantified pairwise similarity within each of the three script categories. As expected, we observed that within-category scripts displayed high semantic similarity (Similarity (r): ‘PTSD’ = 0.164 ± 0.139, ‘Sad’ = 0.237 ± 0.174 ‘Calm’ = 0.413 ± 0.174; 1-sample t-test vs 0: all p < 1E-300). Further, an ANOVA conducted on script types revealed a main effect of ‘script’ (ANOVA: F(2,1133) = 231.80, p = 4.29E-85, η_p² = 0.225). Post-hoc t-tests further demonstrated that ‘Calm’ scripts had higher within-category similarity than ‘Sad’, and ‘Sad’ scripts had higher with-category similarity than ‘PTSD’ (both p < 0.0001, Bonferroni corrected).

These results suggest that sad and traumatic memories in the cohort overlapped in themes and semantic content. This analysis laid foundations for asking whether the neural patterns associated with these memories will differ by their clinical classification - traumatic or sad. Thus, any ensuing neural differences between ‘PTSD’ and ‘Sad’ reactivations may present themselves on top of a maximally identical pool of stimuli. We note that given their autobiographical nature and the use of a naturalistic paradigm, such stimuli may never be identical. However, establishing a ‘handle’ on the differences and commonalities of the narratives enables us to leverage those differences in a quantifiable manner to provide neural insight.

Validation of comparable psycholinguistics between scripts

We applied a series of validations to script content using external psycholinguistic rating databases (see Methods for details). We focused on concreteness, imageability, and valence, given that these factors may act as possible confounders in subsequent neural investigation that is based on imagery and memory reactivation (Figure S5). Repeated measures ANOVA indicated significant differences between the three conditions (’PTSD’, ‘Sad’, ‘Calm’) for the three variables (concreteness: rmANOVA F(2,54) = 14.02, p = 1.24E-05; valence: rmANOVA F(2,54) = 6.03, p = 0.004; imageability: rmANOVA F(2,54) = 12.37, p = 3.77E-05).

Post-hoc paired t-tests and Bayesian analyses confirmed no differences between ‘PTSD’ and ‘Sad’ in concreteness (t(27) = −1.69, p = 0.207 (Bonferroni corrected); BF₁₀ = 0.706) or valence (t(27) = 1.36, p = 0.707 (Bonferroni corrected); BF₁₀ = 0.461). We observe a difference in imageability ratings between ‘PTSD’ and ‘Sad’ scripts such that PTSD scripts had a higher imageability score (t(27) = 4.87, p = 3.8E-05 (Bonferroni corrected); BF₁₀ = 565). These results confirm that psycholinguistic features in general cannot explain the differences observed in the semantic-to-neural relationship between conditions. Specifically, lower imageability cannot explain the lack of hippocampal involvement in PTSD — despite higher level imageability scores, the PTSD scripts do not show increased neural similarity.

Intersubject Representational Similarity Analysis (IS-RSA)

All participants listened to each script type three times during the functional scan. We used general linear modeling (GLM) to measure brain activity during script reactivation compared to baseline (see Methods). Given our a-priori interest in the involvement of the hippocampus, amygdala and PCC in the processing of traumatic autobiographical memories in PTSD, we extracted signals from these structures using the Harvard-Oxford probabilistic atlas (hippocampus and amygdala) and a GLM-derived functional region-of-interest (ROI) in the PCC to conduct ROI-targeted analyses.

In order to enhance signal to noise for the detection of the common pattern elicited by the repeated stimuli across different presentations, the time courses were averaged across the three repeats of the script per ROI within the course of a single scan. The voxel time course of this average run was collapsed across time to generate a spatial pattern associated with the reactivation of each autobiographical memory. Finally, we generated a neural similarity matrix by calculating the pairwise Pearson correlation coefficient between the spatial patterns of all scripts. The dimensions of this matrix were identical to the one storing the narratives’ semantic similarity (28 participants X 3 script types) (Figure 3a).

Figure 3: Semantic-to-neural similarity analysis of hippocampal patterns.

(a) Extraction of spatial patterns. Each script narrative (‘PTSD’, ‘Sad’, ‘Calm’) was played three times in the scanner. spatial patterns were extracted from regions of interest and averaged across repeated presentations to generate an average pattern associated with each script reactivation.

(b) Intersubject representational similarity analysis (IS-RSA). We carried out three independent IS-RSA in which intersubject variability during script reactivation was captured using two subject by subject matrices: one depicting neural pattern pairwise similarity, and the other depicting semantic pairwise similarity. Spearman rank correlation was calculated for each pair of vectorized similarity matrices to provide a correlation coefficient tying semantic and neural representation of either ‘PTSD’, ‘Sad’ or ‘Calm’ narratives.

(c) Hippocampus – neural similarity matrix. Script-by-script neural cosine similarity matrix for spatial patterns extracted from the hippocampus during script reactivation. Within-category similarity is marked in colored triangles off the matrix diagonal (‘PTSD’: red, ‘Sad’: blue, ‘Calm’: gray).

(d) Semantic similarity matrix Script-by-script semantic similarity of scripted narratives. Within-category similarity is marked in colored triangles off the matrix diagonal (‘PTSD’: red, ‘Sad’: blue, ‘Calm’: gray).

(e) Hippocampus – semantic-to-neural IS-RSA. Intersubject representational similarity analysis conducted on pairwise similarity of semantic content and neural patterns in the hippocampus. Each datapoint is one pairwise comparison (N = 378 per condition). Analysis was iterated per script type (‘PTSD’: red, ‘Sad’: blue, ‘Calm’: gray). Histograms along axes depict similarity distribution, thick trace depict estimated density, colors correspond to main legend. Regression lines are approximate visualization of Spearman correlation ρ coefficients for IS-RSA in ‘PTSD’ (red, ρ = −0.117, p_{(FDR corrected)} = 0.104 ) and ‘Sad’ scripts (blue, ρ = 0.177, p_{(FDR corrected)} = 0.005 ). Two-sided tests, error bands indicate 95% confidence bounds. * p < 0.05; *** p < 0.001.

To account for the idiosyncratic nature of autobiographical memories, we used intersubject representational similarity analysis (IS-RSA) to relate the personalized semantic content of the scripts with neural representations acquired during script reactivation. Generally speaking, IS-RSA assesses how intersubject variability in neural patterns relates to individual differences in behavioral measures or traits, and in this case – autobiographical memories. The semantic-neural similarity coefficient indicates that the more similar participants’ semantic content is, the more similar their neural patterns, while listening to their own autobiographical memories.

In IS-RSA, the neural and semantic similarity matrices were vectorized and correlated (Spearman’s correlation) within each category of script type, yielding three correlation coefficients (‘rho’, ρ) per ROI, one for each script type (Figure 3b). The first IS-RSA we conducted related between-participant neural similarity of hippocampal patterns (Figure 3c) with the between-participant semantic similarity matrix computed before (Figure 3d).

In the hippocampus, the IS-RSA uncovered a differentiation in semantic representation: semantic similarity scaled positively with neural similarity for ‘Sad’ narratives (‘Sad’, ρ = 0.177, p_{(FDR corrected)} = 0.005) but not for ‘PTSD’ narratives (‘PTSD’, ρ = −0.117, p_{(FDR corrected)} = 0.104) (Figure 3e). (We further verified the strength of link between semantic and neural matrices in the ‘Sad’ condition using Mantel’s test for matrices correlation^46,47 which suggested a strong effect, p = 0.001). A detailed breakdown of the three overlaid conditions in the hippocampal IS-RSA figure is available in Figure S6.

We then tested whether the two correlation coefficients of IS-RSA conducted on ‘PTSD’ and ‘Sad’ scripts, and associated hippocampal patterns differed significantly. The correlation coefficients underwent a z-score transformation and the absolute difference between them was assigned with a p value (see Methods). We observed that indeed hippocampal IS-RSA representations of semantic content significantly differed as a function of script type (Coefficient comparison (two-tailed) PTSD’ vs. ‘Sad’, hippocampus: p_{(FDR corrected)} = 0.00045.

No semantic-to-neural representation in amygdala and PCC

To examine whether the lack of semantic representation of traumatic autobiographical memories was specific to the hippocampus, we repeated the same IS-RSA with neural patterns extracted from the amygdala and PCC (Figure S7). Signals from the amygdala did not demonstrate a significant link between semantic content and neural patterns for either ‘Sad’ (ρ = 0.066, p_{(FDR corrected)} = 0.404) or ‘PTSD’ narratives (amygdala: ρ = −0.057, p_{(FDR corrected)} = 0.269). In addition, univariate analyses of GLM-derived parameter estimates suggested comparable levels of activation across script types in the amygdala (rmANOVA F(2,54) = 0.96, p = 0.39, Bayesian rmANOVA: Null model: BF_M=3.75; ‘script’ model: BF_M/BF₁₀ = 0.27).

To verify that this null effect is not a result of averaging left and right amygdala signals, we conducted IS-RSA and univariate analysis left and right amygdala separately, and observed no differences between ‘PTSD’ and ‘Sad’ conditions (IS-RSA: Left amygdala: ‘PTSD’, ρ = −0.03, p = 0.56, ‘Sad’, ρ = 0.024, p = 0.643; Coefficient comparison (two-tailed) ‘PTSD’ vs. ‘Sad’ p = 0.46); Right amygdala: ‘PTSD’, ρ = −0.064, p = 0.21, ‘Sad’, ρ = 0.078, p = 0.129; Coefficient comparison (two-tailed) ‘PTSD’ vs. ‘Sad’ p = 0.051); parameter estimates: Left amygdala (F(2,54) = 1.19, p = 0.31; Bayesian rmANOVA: Null model: BF_M=3.29; ‘script’ model: BF_M/BF₁₀ = 0.304; Right amygdala (F(2,54) = 0.541, p = 0.585; Bayesian rmANOVA: Null model: BF_M=5.25; ‘script’ model: BF_M/BF₁₀ = 0.19).

To define the PCC functionally, we generated a contrast comparing neural activity during narrative playback (all three script types were included) to inter-trial interval baseline. This contrast therefore factored activations for both positively and negatively valenced scripts. Signals from the amygdala and PCC did not demonstrate a significant link between semantic content and neural patterns for either ‘Sad’ (ρ = 0.049, p_{(FDR corrected)} = 0.495) or ‘PTSD’ narratives (ρ = 0.079, p_{(FDR corrected)} = 0.374). The univariate approach uncovered a main effect of script (rmANOVA F(2,54) = 6.24, p = 0.003, Bayesian rmANOVA: Null model: BFM=0.08; ‘script’ model: BF_M/BF₁₀ = 12.2) but no difference between ‘PTSD’ and ‘Sad’ was observed in a post-hoc planned comparison (p = 0.558, BF₁₀ = 0.236). Finally, additional exploratory analysis on two other DMN regions that were functionally active during the task (medial prefrontal cortex, mPFC) and left angular gyrus, (both defined functionally by contrast: all conditions > baseline) did not reveal any effects (Figure S8)

Control Analyses

We verified that the difference in semantic-to-neural mapping is not due to differences in mean amplitude between the conditions. To this end we conducted general linear modeling (GLM) on the functional imaging data in which regressors related to all event types were created. The GLM included three separate regressors denoting the full duration of the three types of scripts – ‘PTSD’, ‘Sad’ and ‘Calm’ (see Figure S9 for activation maps). We extracted univariate parameter coefficients from the hippocampus and compared them using a repeated-measures ANOVA (rmANOVA) which uncovered no differences between script types (rmANOVA F(2,27) = 0.67, p = 0.528, Bayesian rmANOVA: Null model: BFM=5.7; ‘script’ model: BF_M/BF₁₀ = 0.177). Additionally, a whole-brain GLM contrasting positive- and negative-valence scripts ([‘PTSD’+’Sad’] > ‘Calm’) uncovered no significant differences in hippocampal signals (see Table S1 for clusters of activation). No clusters were observed also for the opposite contrast ([‘PTSD’+’Sad’] < ‘Calm’).

To control for possible habituation over the course of script playback, we repeated this approach while modeling the early and late parts of the script separately (roughly one minute each). Focusing on the early part, we again observed comparable parameter estimates across script types (rmANOVA F(2,27) = 1.59, p = 0.212, Bayesian rmANOVA: Null model: BF_M=2.8; ‘script’ model: BF_M/BF₁₀ = 0.36). Taken together, these results mitigate concerns regarding confounding effects of different arousal levels across conditions.

We also sought to rule out a possible confounding factor in narrative similarity that may have been driven by similarity of low-level acoustic properties of the auditory stimuli. To this end we generated an acoustic similarity matrix (see Methods) and conducted an IS-RSA using the acoustic and neural similarity matrices. We did not observe any significant representation of acoustic features in the hippocampal patterns, both in ‘PTSD’ and ‘Sad’ conditions (Figure S10).

Finally, we applied a battery of control analyses to test whether the low correlation coefficient values in the semantic-to-neural IS-RSA conducted on the ‘Calm’ scripts was due to insufficient variance in the semantic content of the full set of this type of autobiographical memories.

To do so, we comprehensively mapped the link between semantic and neural similarity across script types. We iteratively generated 5-script subsets of the original cohort and conducted IS-RSA on all possible combinations separately per script type (N = 98,280 X 3). Next, we sorted the subsets by their semantic cosine similarity and directly compared the values of IS-RSA correlation coefficients of the hundred highest- and lowest- ranking subsets. We observed that in both ‘Calm’ and ‘Sad’ scripts, higher semantic similarity was significantly associated with higher hippocampal IS-RSA values (‘Calm’: mean ‘Lowest’ = −0.142, ‘Highest’ = 0.026, 2-sample t-test t(198) = −3.21, p = 0.0015; ‘Sad: mean ‘Lowest’ = 0.067, ‘Highest’ = 0.219, 2-sample t-test t(198) = −3.22, p = 0.0014, Figure S11a,b). Conversely, ‘PTSD’ scripts exhibited an opposite effect where higher semantic similarity was significantly associated with lower hippocampal IS-RSA (‘PTSD: mean ‘Lowest’ = 0.069, ‘Highest’ = −0.171, 2-sample t-test t(198) = 5.09, p = 6.14E-07, Figure S11c).

The same analysis conducted on PCC patterns showed a consistent and robust link between semantic similarity and neural similarity, regardless of script type (‘Calm’: mean ‘Lowest’ = −0.306, ‘Highest’ = 0.052, 2-sample t-test t(198) = −8.97, p = 2.31E-16; ‘Sad: mean ‘Lowest’ = −0.044, ‘Highest’ = 0.513, 2-sample t-test t(198) = −13.08, p = 1.41E-28; ‘PTSD: mean ‘Lowest’ = −0.051, ‘Highest’ = 0.382, 2-sample t-test t(198) = −7.49, p = 2.24E-12, Figure S11d–f). See Table S3 for a detailed summary of central tendency measurements for this analysis.

Involvement of hippocampal subregions in representation

Evidence suggests hippocampal subregions along its longitudinal axis are recruited differently during tasks involving recall autobiographical memory and scene construction^48,49 with particular implications in PTSD^50–52. With this in mind we conducted an exploratory analysis to further delineate the differences in patterns of hippocampal representation of traumatic autobiographical memory narratives. We conducted IS-RSA on neural patterns extracted separately from the anterior and posterior extremities of the hippocampus (split into three segments: anterior, middle, and posterior, see Methods).

We observed the semantic representations tended to be more pronounced in the posterior part of the hippocampus (Posterior hippocampus: ‘PTSD’, ρ = −0.0506, p = 0.327; ‘Sad’, ρ = 0.169, p_{(FDR-corrected)} = 0.0004; Coefficient comparison (two-tailed) ‘PTSD’ vs. ‘Sad’ p_{(FDR-corrected)} = 0.0005; anterior hippocampus: ‘PTSD’, ρ = −0.0503, p = 0.329; ‘Sad’, ρ = 0.065, p = 0.204; Coefficient comparison (two-tailed) ‘PTSD’ vs. ‘Sad’ p = 0.112) (Figure S12). When tested directly, the difference between the correlation coefficients was not significant (p = 0.141).

Memory type can be decoded from hippocampal patterns

To further support our finding that hippocampal neural patterns were linked to the semantic content of the autobiographical memories in ‘Sad’, but not in ‘PTSD’ script reactivation, we sought a complimentary approach by which these two conditions may be teased apart. With this in mind, we trained a regularized li7near discriminant analysis (rLDA) model to decode narrative conditions (‘PTSD’ or ‘Sad’) from the multivoxel spatial patterns extracted from the hippocampus during script reactivation. Using 25-fold cross-validation we were able to decode at 66.2 % accuracy whether hippocampal spatial patterns belonged to a ‘PTSD’ or ‘Sad’ narrative. To assess the power of this predictive ability, we iterated the procedure with shuffled labels (N = 2,500) to obtain a surrogate distribution of decoding accuracy (mean ± SD = 52.7 % ± 4.9). Nonparametric testing confirmed that decoding accuracy for the empirical data was well above chance (p = 0.0028) (Figure 4a). To further validate our finding, we then attempted to decode ‘PTSD’ from ‘Calm’ scripts, expecting those to be teased apart more easily given their different mental state. Indeed, we were able to decode these conditions at 80.9 % accuracy, which was well above chance (mean ± SD = 54.2 % ± 5.0, p < 0.0001) (Figure 4b). For comparison, spatial patterns derived from the amygdala could not be used as robustly to distinguish between ‘PTSD’ and ‘Sad’ scripts (decoding accuracy for empirical data: 58.2 %, surrogate: mean ± SD = 52.2 % ± 4.7, p = 0.097) (Figure 4c–d). These findings provide additional support to the idea that hippocampal activity during traumatic autobiographical memory recall represents elements of the narrative, and these patterns are observably different from activity during non-traumatizing autobiographical memory recall.

Figure 4: Memory type can be decoded from hippocampal patterns.

(a) Decoding accuracy of scripted narrative type from neural patterns. Vertical black line denotes decoding accuracy of ground truth script type (‘PTSD’ vs. ‘Sad’) from hippocampal spatial patterns in the empirical condition. Colored histogram is a surrogate distribution comprised of decoding accuracy for the same neural data with shuffled labels. p value is derived non-parametrically through a two-sided permutation test (N = 2,500, p = 0.0028).

(b) Same as (a) but for decoding accuracy of ‘PTSD’ vs. ‘Calm’ from hippocampal spatial patterns (N = 2,500, p =0.0004).

(c) Same as (a) but for decoding accuracy of ‘PTSD’ vs. ‘Sad’ from amygdala spatial patterns (N = 2,500, p = 0.097).

(d) Same as (a) but for decoding accuracy of ‘PTSD’ vs. ‘Calm’ from amygdala spatial patterns (N = 2,500, p = 0.002).

Symptom level modulates PCC representation of ‘PTSD’ memory

Having observed that similar non-traumatic semantic content is associated with similar neural patterns in the hippocampus, but not in the PCC, we asked whether individual differences in PTSD symptoms severity may have attenuated the group effects observed through IS-RSA. We therefore carried out an exploratory analysis in which we asked whether PTSD severity, operationalized as CAPS-5 score evaluated during screening, will explain the between-subject variability in the extent by which narratives will be represented in the PCC. To this end, we split the cohort into two sub-groups (both n = 14) labeled ‘High’ and ‘Low’ according to the median of their CAPS-5 score. The total scores in the study ranged between 26 – 60 (mean ± SD = 41.2 ± 8.3). Setting a cutoff CAPS-5 score of 38 (non-inclusive), the two resulting sub-groups significantly differed in their CAPS-5 score (‘High’ = 47.8 ± 3.4, ‘Low’ = 34.7 ± 6.5 t(26) = 6.68, p < 0.0001).

First, we verified that univariate BOLD responses in these ROIs were comparable. We used a whole-brain two-sample t-test design to compare the ‘High’ and ‘Low’ symptoms groups’ response across script types. In both ‘PTSD’ > ‘Calm’ and ‘PTSD’ > ‘Sad’ contrasts we were not able to detect differences between the groups using cluster size threshold corrected with p_FWE < 0.05. Next, we verified that semantic similarity within these sub-groups was comparable, that is, that script similarity within each script category did not differ between the groups.

The rmANOVA uncovered a main effect of script type (F(2,180) = 108.56, p < 0.00001), a main effect of ‘symptoms severity’ (F(1,90) = 11.02, p = 0.00131) but no interaction between the two (F(2,180) = 2.15 p = 0.119). Post-hoc comparisons (Bonferroni corrected) revealed that in ‘PTSD’ scripts and ‘Sad’, the average semantic similarity did not differ between the ‘High’ and ‘Low’ symptoms groups (two-sample t-test, t(180) = −0.73, p = 1; BF₁₀ = 0.207); ‘Sad’: t(180) = −2.03, p = 0.22, BF₁₀ = 1.09. In the ‘Calm’ condition, scripts similarity was significantly higher in the ‘Low’ group than the ‘High’ one (‘Calm’: t(180) = −3.18, p = 0.0043, BF₁₀ = 16.23). We therefore applied ensuing IS-RSA only to the ‘PTSD’ and ‘Sad’ conditions.

We conducted IS-RSA separately for each subgroup, using its corresponding semantic and neural similarity matrices. Matrix dimensions were 14 X 14, which when vectorized, resulted in 91 values. We compared the resulting Spearman coefficients that were derived for each of the two symptom groups using a nonparametric test. We split the full cohort into two random sub-groups iteratively (N = 25,000) and computed a surrogate distribution to which the statistics of the true, CAPS-5-based split, were compared.

The PCC displayed a strong discriminative utility where higher symptom severity was associated with stronger semantic representation of ‘PTSD’ scripts (Figure 5a). Representation of the ‘Sad’ scripts also showed stronger representation in the ‘High’ symptoms group, however to a much lesser extent (Figure 5b). (PCC ‘PTSD’: ‘High’: ρ = 0.266; ‘Low: ρ = 0.0687; ‘Sad’: ‘High’: ρ = 0.0756; ‘Low: ρ = −0.0537). In tandem with the functional ROI, the PCC was also defined anatomically in two supporting ROI analyses to account for ROI definition method (see Methods and Figure S13)

Figure 5: Symptoms severity modulate PCC representation of PTSD memory.

(a) Intersubject representational similarity analysis in the PCC differed by symptom severity IS-RSA conducted on pairwise similarity of semantic content of ‘PTSD’ narratives and neural patterns in the PCC on subgroups differing in symptom severity (‘Low’ and ‘High’). Each datapoint is derived from a pairwise comparison. Analysis was iterated per subgroup. Regression lines are approximate visualization of Spearman correlation ρ coefficients for IS-RSA in ‘Low’ and ‘High’ symptoms (light and dark red, respectively). Error bands indicate 95% confidence bounds.

(b) Same as (a) but conducted on semantic content of ‘Sad’ narratives and neural patterns in the PCC on subgroups differing in symptom severity (‘Low’ and ‘High’ are light and dark blue, respectively). Error bands indicate 95% confidence bounds.

(c) Permutation test for differences in IS-RSA by symptom severity. Vertical black line denotes z transformed difference (High-Low) in correlation coefficients in semantic-to-neural IS-RSA Colored histogram is a surrogate distribution comprised of randomly generated with shuffled severity labels. p value is derived non-parametrically via permutation test, uncorrected (N = 25,000). From left to right: Hippocampal patterns during ‘PTSD’ narratives (z = 0.661, p = 0.074). PCC patterns during ‘PTSD’ narratives (z = 1.359, p = 0.00016). Hippocampal patterns during ‘Sad’ narratives (z = 0.05, p = 0.885). PCC patterns during ‘Sad’ narratives (z = 0.881, p = 0.0036).

By contrast, in the hippocampus, the extent of the link between neural patterns and semantic representations of both ‘PTSD’ and ‘Sad’ scripts did not differ by symptom severity (hippocampus ‘PTSD’: ‘High’: ρ = −0.254; ‘Low: ρ = −0.163; ‘Sad’: ‘High’: ρ = 0.199; ‘Low: ρ = 0.206). Nonparametric permutation tests asserted the Spearman coefficients significantly differed by symptom severity in the IS-RSA based on PCC, but not the hippocampal neural patterns (PCC ‘PTSD’: coefficient difference_{(High - Low)} = 0.197, nonparametric permutations p_{(FDR corrected)} = 0.001; ‘Sad’: coefficient difference_{(High - Low)} = 0.129, nonparametric permutations p_{(FDR corrected)} = 0.011; hippocampus ‘PTSD’: coefficient difference_{(High - Low)} = −0.074, nonparametric permutations p_{(FDR corrected)} = 0.875; ‘Sad’: coefficient difference_{(High - Low)} = −0.007, nonparametric permutations p_{(FDR corrected)} = 0.885) (Figure 5c).

We further observed that symptom severity modulation of the link between neural patterns and semantic representations of ‘PTSD’ but not ‘Sad’ scripts was specific to the PCC and did not extend into the amygdala (‘PTSD’: ‘High’: ρ = 0.037; ‘Low: ρ = −0.082; ‘Sad’: ‘High’: ρ = 0.054; ‘Low: ρ = 0.120; PTSD: coefficient difference_(High-Low) = 0.1184, nonparametric permutations p_{(FDR corrected)}= 0.02, note that this effect is driven by the small negative correlation in the ‘Low’ condition; ‘Sad’: coefficient difference_(High-Low) = −0.0657, nonparametric permutations p_{(FDR corrected)}= 0.286) (Figure S14).

This result suggests that the severity of PTSD symptoms is linked to the semantic representation of the traumatic narrative in the PCC, and to a lesser extent, in the amygdala, whereas the differentiation in representation observed in the hippocampus persisted regardless of symptom severity.

Discussion

Despite continuous effort, the nature of intrusive traumatic autobiographical memories and the mechanisms underlying their unique perceptual attributes in PTSD remain largely unknown⁵³. Here we used individualized traumatic autobiographical memory narratives in a script reactivation paradigm, in which PTSD patients listened to a novel rendition of their traumatic memory. We set out to ask whether, and how the hippocampus, amygdala, and PCC differentiate traumatic autobiographical memories from sad ones. Given the duration and richness of the stimuli, our paradigm was at the intersection of autobiographical memory reactivation and naturalistic narrative comprehension tasks.

We leveraged the variance between idiosyncratic memories by quantifying their semantic similarity to ask whether their neural representations are altered during the processing of personal trauma narratives, compared to negative non-traumatic narratives of the same individuals (‘Sad’). Using intersubject representational similarity analysis, we found that hippocampal patterns showed a differentiation in semantic representation by narrative type; Sad scripts which were semantically similar (e.g., death of a loved one) across participants, elicited similar neural representations. Conversely, thematically-similar traumatic autobiographical memories (of DSM-5 Criterion A event) did not elicit similar representations. Unlike the hippocampus, the amygdala did not represent semantic information in a significant manner, suggesting a poorer representational space for semantic content.

Finally, in an exploratory analysis, we focused on the PCC to ask whether unlike the hippocampus, this region—recently conceptualized as a cognitive bridge between the world events and representation of the self^33,54—will demonstrate a positive relation between semantic content and neural patterns of the traumatic narratives. We indeed observed such a relation in the PCC, with individual symptom severity mediating the extent of semantic-to-neural representation. Conversely, this differentiation by PTSD severity was not evident in the hippocampus. An exploratory analysis did not reveal any additional regions of the DMN evincing the same function.

Semantic-to-neural mapping has been demonstrated comprehensively, both in the hippocampus^55–57 and cortex^58,59. Therefore, we expected to observe a link between narratives’ semantic similarity and elicited neural similarity. That said, here we make two advances: First – we extend this understanding into an underexplored domain: real-life traumatic autobiographical memories in PTSD. Second – we observe that within the same brains, hippocampal representations differed considerably between two types of autobiographical memory of comparable content and valence.

In a complementary approach, we decoded condition identity from the hippocampal patterns. The fact that we were able to tease the two negative conditions apart suggests that these signals hold some shared high-dimensional pattern implying a common cognitive state shared across participants.

Our key findings therefore are twofold: first – that the emotional content of autobiographical memories is represented differently in the two major systems subserving autobiographical memory - the hippocampus and the PCC. Second – that traumatic autobiographical memories undergo a parallel, or a dissociable mode of representation suggesting they profoundly differ from neurotypical autobiographical memories of comparable content and valence.

Why would traumatic autobiographical memories be represented differently than non-traumatic ones? We discuss several explanations: PTSD patients may develop a highly detailed and very personal memory of their traumatic event, and thus their semantic representations become highly idiosyncratic (i.e., unique to the individual). This interpretation is in line with a recent study⁶⁰ reporting that the more concepts were perceived as self-relevant, the more person-specific became their neural representation of valence. Thus, it is possible that PTSD phenomenology generates an overly personal autobiographical memory, which despite being semantically similar to other memories, is linked to highly idiosyncratic representations. Another possibility is that traumatic memory reactivation is not experienced as memory per se, but is rather disconnected from time and space and from current surroundings, and thus experienced as an intrinsic mental event, akin to the internal processing that typically engages the DMN.

Lastly, an intriguing possibility is that patients attempted to block or suppress the reactivation of the traumatic content, and by doing so, exhibited brain activity that was incongruent with the semantic content presented to them^61,62. Re-examining the nature of these memories after successful trauma-focused psychotherapy may shed further light on the observed results.

In clinical settings, the evaluation of traumatic memory organization is often reliant mostly on meta-memory: the patient’s self-report about memory coherence and meaning of the traumatic experience ⁶³. Semantic representation of idiosyncratic autobiographical memories using IS-RSA may allow a more objective neural marker for PTSD. For example, we may observe the emergence of semantic-to-neural mapping of the traumatic memory within the hippocampus in the course of treatment. Trauma-focused psychotherapy could help restore a narrative and also align the narrative with a more normative and less idiosyncratic meaning.

Finally, we would like to acknowledge several limitations of our study: the sample size of 28 patients may have limited the generalizability of the results. This was particularly evident in the ad-hoc analysis involving symptom severity, where the sample was split in two halves. Further, despite the careful construction and thorough analysis of the memory scripts, differences not accounted for may have remained, particularly given their naturalistic nature. For example, we found that the variance of the ‘Calm’ scripts dataset had low variance, likely obscuring the semantic-to-neural relationship. Lastly, our within-subject design, where the traumatic narratives were compared to sad ones within the same individuals, controlled for potential between-cohort effects such as the global alterations in brain structure and connectivity in PTSD. Still, it would be interesting to further delineate the factors that influence semantic-to-neural mapping of autobiographical memories within the neurotypical range, such as parametrically modulating relevancy to self and narrative coherence.

Ending with our initial question about the very nature of PTSD phenomenology: is traumatic memory an extreme case of ‘standard’ negative emotional processing or a different cognitive entity altogether? Our main finding, that hippocampal patterns of PTSD patients showed a differentiation in semantic representation by narrative type during memory reactivation, supports the idea of a profoundly separate cognitive experience in the reactivation of traumatic memories. This is consistent with the notion that traumatic memories are not experienced as memories per se. Rather, these are fragments of prior events, subjugating the present moment to evade the comfort of belonging to the past.

Methods

Experimental paradigm

The study was registered at clinicaltrials.gov (NCT02727998T) and approved by the Yale University Institutional Review Board (IRB). Twenty-eight participants (mean age = 38.1 ± 10.5 years, range = 24 – 63; females (n = 11), males (n = 17) took part in this study after providing a statement of informed consent. All participants had chronic PTSD (see Table S2 for cohort demographics). PTSD diagnosis was established using the Clinician-Administered PTSD Scale (CAPS-5)¹. The cohort in this study is part of a larger, longitudinal study focused on effects of ketamine-aided extinction in PTSD, for which the participants were compensated for a total sum of up to $700 for their participation in the original clinical trial. Our data are based on the baseline session assessment only, with no drugs being administered. Therefore, blinding was not relevant and data collection and analysis were not performed blind to the conditions of the experiments. No statistical methods were used to pre-determine sample sizes but our sample sizes are similar to those reported in the publication of the aforementioned longitudinal study⁶⁴. Exclusion criteria included a diagnostic history of bipolar disorder, borderline personality disorder, obsessive-compulsive disorder, schizophrenia or schizoaffective disorder, dementia, current psychotic features, or suicide risk, moderate or higher severity of substance use disorder and history of traumatic brain injury. Participants who were currently engaged in trauma focus therapy were also ineligible to participate in the study. Lastly, patients were excluded for acute medical illness. Psychotic features were determined by the Structured Clinical Interview for DSM-IV (SCID)⁶⁵. All participants who were prescribed psychiatric medication had to maintain a stable dose for four weeks prior to the assessment of PTSD and other study inclusion/exclusion criteria. Therefore, if they met PTSD diagnosis, this was established even in the presence of medication. Randomization was handled by the Yale Investigational Drug Services and participants were randomized in counterbalanced blocks of 10-subjects each stratified by gender.

Of the 28-patient cohort mentioned, no data exclusions were made. All script-reactivation blocks were included in the IS-RSA. Exclusion of specific words from script content in detailed in the Semantic analysis section.

We used the Sinha method – a previously established script-driven imagery development method of autobiographical events, which has been developed as a way to personalize stimuli that induce traumatic imagery, and has been widely used as a method of memory reactivation in PTSD research^66–70. The method aims to produce ecologically valid audio narratives by activating comparable physiological, subjective, and behavioral responses across patients. This methodology allows comparison across scripts and across subjects while avoiding confounds such as length, vocabulary, sentence structure, and experiential features. The narrative is built based on several hours of clinical sessions, in which patients were asked to describe the events in as much detail as possible. It employs a structured form that is similar across scripts but at the same time highly detailed with specifics (e.g., military jargon, nicknames, personal slang) to provoke the individualized experience.

Specifically, patients completed an imagery-development procedure in which they were asked to describe the traumatic event associated with their PTSD, as well as a significant sad, but not traumatizing event (‘Sad’) and a positive, low-arousal, event in which they felt relaxed (referred to as ‘Calm’). The imagery scripting procedure followed a procedure presented by Sinha et al.⁷¹ Participants were asked to describe the events in as much detail as possible. They were then asked to select at least three physiological responses corresponding to each specific event to be later embedded in the narrative. Using this information, we developed audio scripts, approximately 120-second long for each event (Duration (sec): ‘PTSD’ = 120.19 ± 1.20; ‘Sad’ = 119.64 ± 1.48; ‘Calm’ = 118.69 ± 2.72), narrated by a male member of the research staff. The narratives were all in second-form pronouns (‘you’/’your’), mostly in the present tense.

The scripts were comprised of three main elements – the episodic unfolding of events (time and place, scenery description, actions, dialogues), description of mental state (e.g., ‘you feel helpless’, ‘a feeling of peace comes over you’), and a vocabulary of sensory and somatic phrases used to promote reexperiencing (e.g., ‘your heart beats faster’). The somatic vocabulary consisted of references to heartbeat, respiration, muscle tone, perspiration, tearing etc. Both negatively-valenced script types – ‘PTSD’ and ‘Sad’ – were conveyed using the same vocabulary, to control and maximize the similarity of reactivation between these conditions. The positive scripts (‘Calm’) often mentioned similar autonomic functions, but with opposite reactions (e.g., muscle relaxation, slow breathing).

Semantic analysis

For the purpose of semantic analysis some names, for example places with greater geographical resolution than a US state, commercial brands or firms and specialized military jargon or acronyms were removed. Sentences were defined according to full stops, question, and exclamation marks as they appear in the text read by the narrator. As part of the reactivation procedure aimed to heighten autonomic arousal, the pace (words per script) of both negatively valenced scripts was intentionally higher and their duration slightly longer than the ‘Calm’ ones.

During preprocessing of the semantic input, punctuation marks were erased, texts were transformed to lower-case and tokenized. Stop words, defined by MATLAB R2020a’s default NLP vocabulary, were removed. Words underwent lemmatization and were then assigned a 300-dimension vector representation. The semantic space we used was a pre-trained embedding for 1M English words (16B tokens) available through MATLAB’s word2vec NLP tools. Vector representations were calculated for each single word. Words in the scripts which were not indexed in the pre-trained space were not included in the analysis. The semantic representation of the next level in text hierarchy – the sentence – were calculated as the average representation of words in it. Similarly, the semantic representation of each entire script was calculated by averaging the representation of its sentences.

Additionally, we carried out the same analytical pipeline using a pre-trained contextual embedding transformer model, BERT, implemented in the MATLAB R2021a NLP toolbox. Briefly, contextual embedding departs from static semantic representations by being able to account for intra-sentence dynamics. As such, here the first level of data input in the hierarchy was sentences, not individual words. Again, representation of scripts was calculated by averaging the representation of its sentences. For both methods, before computing similarity between scripts, we scaled the vector representations by subtracting the average norm of all words embedding in the entire vocabulary that was used in the study (based on all script types), following a normalization procedure described in⁷². The semantic similarity was derived from a transformed distance matrix of pairwise cosine distances of the vectorial representations. We used cosine distance as it is considered a better fit for semantic analysis than Euclidean distance because it does not consider vector magnitude, which is often biased in datasets involving text corpora due to differences in word occurrence^73,74.

Linguistic evaluation of script content

We applied a series of validations to script content using external psycholinguistic rating databases. We focused on factors which may act as possible confounders in subsequent neural investigation, with specific attention to comparisons between the two negative script types – ‘PTSD’ and ‘Sad’. The following measures were tested: 1) Concreteness ratings from the database by Brysbaert et al.⁷⁵, which was defined as the degree to which a concept refers to a perceptible entity 2) Imageability ratings from the Glasgow Norms database by Scott et al.⁷⁶, defined as the degree of effort involved in generating a mental image. 3) Valence ratings derived from NRC-VAD, a database by Mohammad et al which was defined as position of concept on dimensions of positiveness-negativeness / pleasure-displeasure⁷⁷.

Dimensionality reduction

Dimensionality reduction of the semantic information was done using t-distributed stochastic neighbor embedding (t-SNE)⁷⁸. Similarly to principal component analysis (PCA), t-SNE also reduces dimensionality of the input data set. It is superior however to PCA, in its ability to preserve local structure. Projection of t-SNE clusters to 3-dimensional space was carried out in MATLAB, with an intermediate step of a PCA into 100 components. Learning rate was set to 1000 and perplexity was set to 30.

Experimental Procedure

All participants listened to each script type three times in the scanner. Participants were naïve to this new scripted rendition of their autobiographical memories and not familiar with the voice of the narrator. The order of scripts was fixed and identical across all participants but one. The order was: T-C-S-C-T-S-C-T-S in all scans but one, whose order was S-C-T-C-S-T-C-S-T (T = Trauma, S = Sad, C = calm). The fixed order reflects constraints on the methodology of script reactivation: the need to avoid random back-to-back presentation of two identical scripts to prevent habituation, as well as the need to never end with the trauma script for the sake of the patient’s wellbeing following the scan. The pre-defined sequence factored-in all these considerations. Scripts were presented in the scanner using E-prime 2. Each script began with a slide instructing the participant to press a button to initiate the next playback and its type. No visual information was displayed during script playback.

MRI Scan

MRI data were collected with a Siemens 3T Prisma scanner, using a 32-channel receiver array head coil. High-resolution structural images were acquired by Magnetization-Prepared Rapid Gradient-Echo (MPRAGE) imaging (TΡ = 1 s, TE = 2.77 ms, TI = 900 ms, flip angle = 9°, 176 sagittal slices, voxel size = 1 ×1 × 1 mm, 256 × 256 matrix in a 256 mm FOV). Functional MRI scans were acquired while the participants were listening to the narrated scripts, using a multi-band Echo-Planar Imaging (EPI) sequence (multi-band factor =4, TR= 1000 ms, TE= 30 ms, flip angle = 60°, voxel size = 2 × 2× 2 mm, 60 2 mm-thick slices, in-plane resolution = 2 × 2 mm, FOV = 220 mm).

MRI preprocessing

Data were preprocessed with fMRIPrep, version 20.2.0⁷⁹. For a complete preprocessing procedure please refer to the relevant Methods section. Functional images were motion- and slice-time corrected, aligned to T1 anatomical images, and then warped to MNI space. Analysis of the functional data included the following regressors: 6 movement variables (translation and rotation), framewise displacement, the first 6 anatomical components-based noise correction (CompCor) and the 6 first discrete cosine regressors. Subsequent preprocessing and statistical contrasts were done using standard statistical parametric mapping (SPM12, Wellcome Department of Imaging Neuroscience) algorithms (fil.ion.ucl.ac.uk/spm) and custom MATLAB R2018b/R2020a/R2021a code.

ROI analysis

Given our a-priori interest in the function of amygdala and hippocampus in PTSD, we defined masks for ROI analysis of these structures, bilaterally using the probabilistic Harvard Oxford atlas⁸⁰ thresholded at 25%⁸¹. In an exploratory analysis, the posterior cingulate cortex ROI was defined functionally through a contrast comparing neural activity during narrative playback (all three script types were included) to inter-trial interval baseline. Following cluster size thresholding with a family-wise error rate of p_(FWE) < 0.01 an ROI consisting of 361 voxels was defined. Prior to ROI analysis, functional images underwent spatial smoothing using a Gaussian kernel of 1 mm full-width at half maximum (FWHM) to enhance signal-to-noise ratio and classification accuracy⁸². Time courses were extracted from the entire session and were applied with a discrete cosine transform high-pass filter (cutoff of 128 sec)⁸³. ROI data were then normalized using z-score. Spikes in the data, exceeding 4 times the voxel’s standard deviation were applied de-spiking and were interpolated using the mean of one TRs engulfing each side of the outlier data point. In line with previous studies, functional data were shifted 5 seconds (5 TRs) to account for the delay in the hemodynamic response compared with the audio stimuli^72,84,85. The segments containing the script narratives were extracted with time indices rounded to include the nearest TR interval to prevent omission of functional data.

Each script was played three times during the scan. To enhance signal to noise in identifying the recurring pattern elicited by the script, the time course for each ROI was averaged across the three repeats of each script. The three repetitions were always considered a unit and we did not treat them as individual trials. The voxel time course of this average run was collapsed across time to generate a spatial pattern associated with the reactivation of the specific autobiographical memory. This approach also aided in circumventing the slight mismatches in script durations across participants that would have been detrimental for similarity based on temporal fluctuations.

Intersubject representational similarity analysis (IS-RSA):

To relate the semantic content with neural representation and to determine whether this representation differs in PTSD-related narratives, we conducted an intersubject representational similarity analysis (IS-RSA)^73,86–88. The neural similarity matrix was derived from each ROI separately. The semantic similarity matrix was fixed in all analyses (for some analyses it was broken down to smaller matrices, for example, sub-cohorts differing in symptom severity).

A neural similarity matrix was then generated by calculating pairwise Pearson correlation for each pair of the 84 scripts. Metrics that are represented as distance by default (rather than similarity) were transformed to similarity using

$$Similarity=\frac{1}{1+Distance}$$

Since this study comprised of 28 participants, who each listened to 3 different scripts, the narrative-based similarity matrices (e.g., semantic, acoustic) consisted of the lower triangle of a 28 X 28 matrix, extracted separately per narrative type (‘PTSD’, ‘Sad’, ‘Calm’). This corresponded to 378 unique combinations when vectorized ((28 X (28 – 1)) / 2 = 378). Similarly, trait-based similarity matrices (e.g., CAPS) consisted of the lower triangle of a 28 X 28 matrix, corresponding to 378 unique combinations. Of note, in our design both the neural responses and the naturalistic stimuli varied across individuals.

IS-RSA uses non-parametric Spearman’s correlation since it does not assume normal distribution of similarity coefficients of the semantic or neural signals. For other analyses data distribution was assumed to be normal but this was not formally tested.

We tested the significance of IS-RSA by subjecting p-values of Spearman’s correlation coefficients amassed across ROIs and conditions (e.g., 3 ROIs X 3 script types) by using a false-discovery rate (FDR) at q = 0.05 implemented in MATLAB R2018a as function ‘fdr_bh’.

To compare correlation coefficients between narrative types (e.g., ‘PTSD’ vs. ‘Sad’) we used two methods depending on groups’ dependency: In cases where one variable is shared (e.g., correlation between CAPS and neural similarity in ‘PTSD’ narratives compared with correlation between the same CAPS data and neural similarity in ‘Sad’ narratives) we used Steiger test⁸⁹ as implemented in ‘r_test_paired.m’ in MATLAB. Briefly, each correlation coefficient is converted into a z-score using Fisher's r-to-z transformation. Next the asymptotic covariance of the estimates is computed and are then used in an asymptotic z-test. We reported p-values from a two-tailed probability distribution. In contrast, in cases where the two comparisons had no shared components (e.g., correlation between semantic and neural similarity in ‘PTSD’ narratives and correlation between semantic and neural similarity in ‘Sad’ narratives), we used the ‘corr_rtest’ function in MATLAB to convert both correlation coefficient into z-score using Fisher's r-to-z transformation and calculate their absolute difference. This value was assigned a p value from a cumulative normal distribution function (‘normcdf’ in MATLAB). We report p-values from a two-tailed probability distribution.

We included Bayesian paired tests where appropriate, to provide additional information regarding the evidence in favor of the null hypothesis⁹⁰. A Cauchy prior of 0.707 was used as set in software defaults⁹¹. The output Bayesian statistic BF₁₀, was interpreted according to standard recommendations by which BF₁₀ ranges of 1 – 3, 3 – 10 and 10 –30 imply anecdotal, substantial, or strong evidence, respectively. All Bayesian statistical analyses were conducted in JASP (2019) version 0.11.1 and 0.16.4. In Bayesian repeated-measures ANOVA designs we compare two models: the null model, which suggests no difference between conditions, and the alternative model, ‘script’, which suggests scripts do differ from one another. The two models had equal priors (P(M) = 0.5). We report BF_M per model, which quantifies the change from prior odds to posterior odds following the introduction of the evidence (i.e., the data), as well as BF₁₀ which was fixed to always compare the ‘script’ model to the null model.

IS-RSA visualization

IS-RSA is usually calculated using nonparametric tests (Spearman, Kendall). However, the limitation of Spearman, being a ranked test, is that it does not yield a regression in the same manner linear correlations do. Due to this technical limitation in visualizing the relationship between data matrices in the IS-RSA, we plotted slopes that are computed based on Pearson correlation. Note that despite this difference in visualization, throughout this study, the correlation coefficient reported and discussed are Spearman’s ρ values.

Longitudinal parcellation of the hippocampus

To date, several definitions for segmentation of the long axis of the human hippocampus exist. We followed a percentile-based segmentation of the hippocampus into three regions along its long axis, described by Poppenk et al.⁴⁹. To avoid discrepancies between boundaries defined by the various segmentation methods, we omitted the medial part from our analyses and focused only on the two extremities – the most anterior and posterior thirds.

Neural general linear modeling (GLM) and contrasts

We conducted general linear modeling (GLM) of the functional scans of each participant by modeling the observed BOLD signals and regressors to identify the relationship between the task events and the hemodynamic response. First, functional data underwent spatial smoothing using a 6 mm FWHM kernel (note that a different kernel was used in IS-RSA). Next, regressors related to all events were created by convolving a train of delta functions representing the sequence of individual events with the default basis function in SPM12, which consists of a synthetic hemodynamic response function composed of two gamma functions. The GLM included three separate regressors for the onset of the three types of scripts – ‘PTSD’, ‘Sad’ and ‘Calm’. We carried out linear contrasts of parameter estimates to identify effects in each participant. Statistical maps from all participants were then entered into a second-level group analysis to implement a random-effects statistical model. Statistical maps were rendered using MRIcroGL v1.12.20.

Classification analysis

We decoded narrative conditions (‘PTSD’ or ‘Sad’ script) from multivoxel spatial patterns data using a rLDA (regularized linear discriminant analysis classifier, ‘fitdiscr’ function in MATLAB) which shows superior performance over LDA in high-dimensional imaging data which may present issues of multicollinearity and overfitting. Data from all script repeats of the two negatively-valenced conditions, along with corresponding condition labels, were used to train the rLDA. Input data consisted of 28 participants X 2 conditions X 3 repeats yielding a 168-samples in total. To test the model’s performance on a testing data set we iteratively repeated this process with permuted data partitions (N = 2,500) per ROI and condition and then applied cross-validation.

Whole-brain investigation

The main contrast used in the whole-brain investigation was “All Scripts > Baseline” where all script event types (i.e., ‘PTSD’, ‘Sad’ and ‘Calm’) were contrasted with the baseline interval between scripts. Statistical inference was made based on whole-brain statistical maps corrected for multiple comparisons using cluster size threshold family-wise error rate of p_(FWE) < 0.01 for the identification and extraction of regions of interest (ROIs).

Acoustic similarity analysis

Acoustic similarity of the scripts was computed based on acoustic landmarks envelope. We used a custom MATLAB script, adapted from Oganian et al.⁹² to extract the analytic envelope of the speech signal filtered within critical bands based on the Bark scale which is a psychoacoustical measure of loudness.

Subsets analysis

We mapped the link between semantic and neural similarity across subsets by iteratively generating 5-script subsets of the original cohort and conducted IS-RSA on all possible combinations separately, implemented in ‘nchoosek(N,K)’). This procedure was repeated separately per script type. Next, we sorted the subsets by their within-subject semantic similarity and directly compared the values of IS-RSA correlation coefficients of the hundred highest- and lowest- ranking subsets, using a two-sample t-test. Values underwent Fisher Z-transformation before they were compared.

Symptom severity analysis

CAPS - PTSD diagnosis was established using the Clinician-Administered PTSD Scale (CAPS-5)¹. CAPS was administered within one month of the imaging session. The questionnaire data of one participant were missing and were interpolated using the group’s mean per questionnaire item.

Additional definitions posterior cingulate cortex region of interest (Corresponding to Figure 5)

Willard atlas – functional delineation: We used an atlas-derived ROI to generate the PCC mask – the PCC from the Willard functional atlas⁹³, which is extensively used for a-priori demarcation of the PCC. Notably, this was also the ROI in the study by Chen et al⁹⁴. investigating shared memory representations across individuals. Compared to the functionally-derived ROI, the Willard atlas ROI is considerably bigger. The Willard-PCC displayed a strong discriminative utility where higher symptom severity was associated with stronger semantic representation of ‘PTSD’ scripts. Representation of the ‘Sad’ scripts also showed stronger representation in the ‘High’ symptoms group, however to a lesser extent. (Willard-PCC ‘PTSD’: ‘High’: ρ = 0.249; ‘Low: ρ = 0.031; ‘Sad’: ‘High’: ρ = 0.182; ‘Low: ρ = 0.109, Figure S13a and b). Nonparametric permutation tests asserted the Spearman coefficients significantly differed by symptom severity in the IS-RSA (PTSD: coefficient difference (High - Low) = 0.218, nonparametric permutations p = 0.00044; ‘Sad’: coefficient difference (High - Low) = 0.073, nonparametric permutations p = 0.2476)

4mm spherical ROI – anatomical center: To further corroborate the contribution of PTSD symptom severity to autobiographical memory representations in the PCC, we sampled a considerably smaller area of the PCC as well. Here we used a 4mm sphere, sampled from the anatomical center of the PCC. This ROI was obtained from a study by Simony et al. who investigated the role of the DMN in representation of narrative across individuals³⁶. The spherical ROI PCC too, displayed a strong discriminative utility where higher symptom severity was associated with stronger semantic representation of ‘PTSD’ scripts. Representation of the ‘Sad’ scripts also showed stronger representation in the ‘High’ symptoms group, however to a lesser extent (spherical ROI PCC: ‘PTSD’: ‘High’: ρ = 0.293; ‘Low: ρ = 0.086; ‘Sad’: ‘High’: ρ = 0.167; ‘Low: ρ = 0.066, Figure S13c and d). Nonparametric permutation tests asserted the Spearman coefficients significantly differed by symptom severity in the IS-RSA (PTSD: coefficient difference (High - Low) = 0.207, nonparametric permutations p < 0.0001; ‘Sad’: coefficient difference (High - Low) = 0.101, nonparametric permutations p = 0.1085).

fMRIprep Preprocessing

Results included in this manuscript come from preprocessing performed using fMRIPrep 20.2.0⁹⁵, which is based on Nipype 1.5.1⁹⁶. Anatomical data preprocessing: A total of 1 T1-weighted (T1w) images were found within the input BIDS dataset. The T1-weighted (T1w) image was corrected for intensity non-uniformity (INU) with N4BiasFieldCorrection⁹⁷, distributed with ANTs 2.3.3⁹⁸ and used as T1w-reference throughout the workflow. The T1w-reference was then skull-stripped with a Nipype implementation of the antsBrainExtraction.sh workflow (from ANTs), using OASIS30ANTs as target template. Brain tissue segmentation of cerebrospinal fluid (CSF), white-matter (WM) and gray-matter (GM) was performed on the brain-extracted T1w using fast (FSL 5.0.9⁹⁹. Volume-based spatial normalization to one standard space (MNI152NLin2009cAsym) was performed through nonlinear registration with antsRegistration (ANTs 2.3.3), using brain-extracted versions of both T1w reference and the T1w template. The following template was selected for spatial normalization: ICBM 152 Nonlinear Asymmetrical template version 2009c¹⁰⁰.

Functional data preprocessing: For each of the BOLD runs found per subject (across all tasks and sessions), the following preprocessing was performed. First, a reference volume and its skull-stripped version were generated using a custom methodology of fMRIPrep. Susceptibility distortion correction (SDC) was omitted. The BOLD reference was then co-registered to the T1w reference using flirt (FSL 5.0.9)¹⁰¹ with the boundary-based registration cost-function¹⁰². Co-registration was configured with nine degrees of freedom to account for distortions remaining in the BOLD reference. Head-motion parameters with respect to the BOLD reference (transformation matrices, and six corresponding rotation and translation parameters) are estimated before any spatiotemporal filtering using mcflirt (FSL 5.0.9)¹⁰¹. BOLD runs were slice-time corrected using 3dTshift from AFNI 20160207¹⁰³. The BOLD time-series (including slice-timing correction when applied) were resampled onto their original, native space by applying the transforms to correct for head-motion. These resampled BOLD time-series will be referred to as preprocessed BOLD in original space, or just preprocessed BOLD. The BOLD time-series were resampled into standard space, generating a preprocessed BOLD run in MNI152NLin2009cAsym space. First, a reference volume and its skull-stripped version were generated using a custom methodology of fMRIPrep. Several confounding time-series were calculated based on the preprocessed BOLD: framewise displacement (FD), DVARS and three region-wise global signals. FD was computed using two formulations following Power (absolute sum of relative motions), and Jenkinson (relative root mean square displacement between affines^104,105. FD and DVARS are calculated for each functional run, both using their implementations in Nipype. The three global signals are extracted within the CSF, the WM, and the whole-brain masks. Additionally, a set of physiological regressors were extracted to allow for component-based noise correction (CompCor)¹⁰⁶. Principal components are estimated after high-pass filtering the preprocessed BOLD time-series (using a discrete cosine filter with 128s cut-off) for the two CompCor variants: temporal (tCompCor) and anatomical (aCompCor). tCompCor components are then calculated from the top 2% variable voxels within the brain mask. For aCompCor, three probabilistic masks (CSF, WM and combined CSF+WM) are generated in anatomical space. The implementation differs from that of Behzadi et al. in that instead of eroding the masks by 2 pixels on BOLD space, the aCompCor masks are subtracted a mask of pixels that likely contain a volume fraction of GM. This mask is obtained by thresholding the corresponding partial volume map at 0.05, and it ensures components are not extracted from voxels containing a minimal fraction of GM. Finally, these masks are resampled into BOLD space and binarized by thresholding at 0.99 (as in the original implementation).

Components are also calculated separately within the WM and CSF masks. For each CompCor decomposition, the k components with the largest singular values are retained, such that the retained components’ time series are sufficient to explain 50 percent of variance across the nuisance mask (CSF, WM, combined, or temporal). The remaining components are dropped from consideration. The head-motion estimates calculated in the correction step were also placed within the corresponding confounds file. The confound time series derived from head motion estimates and global signals were expanded with the inclusion of temporal derivatives and quadratic terms for each¹⁰⁷. Frames that exceeded a threshold of 0.5 mm FD or 1.5 standardized DVARS were annotated as motion outliers. All resamplings can be performed with a single interpolation step by composing all the pertinent transformations (i.e., head-motion transform matrices, susceptibility distortion correction when available, and co-registrations to anatomical and output spaces).

Data availability

Data supporting the findings of this study are deposited in https://osf.io/dc7jb/.

The Harvard-Oxford atlas is available at https://neurovault.org/collections/262.

The Willard atlas is available at: https://pyhrf.github.io/manual/parcellation_mask.html

The set of analyses described in this study was not preregistered

Code availability

The scripts used for data analysis are available in https://osf.io/dc7jb/

fMRI preprocessing was done in fMRIPrep, analyses were conducted primarily in MATLAB R2018b, R2020a and R2021a (MathWorks, Natick, MA)