Brain activation during suicide-specific cognition in trauma-exposed Veterans

Abstract

The suicide Implicit Association Test (S-IAT) captures the strength of the implicit identification between self and death and is one of the few suicide-specific behavioral tasks that uniquely predicts future suicide risk. Thus, identifying brain regions associated with the S-IAT provide insights into the neural mechanisms underlying suicidality. This study measured brain activation during the S-IAT with concurrent fMRI in a post-9/11 trauma-exposed veteran sample. In total, 37 post-9/11 veterans at low risk for suicide participated in this study as part of an ongoing longitudinal study. Behaviorally, participants were slower to categorize words during incongruent (death-me) contexts relative to congruent (life-me) contexts (p < 0.001). Whole-brain voxelwise fMRI contrasts revealed a brain network that was significantly more active during incongruent trials than congruent trials that included the bilateral occipital, posterior parietal, and cerebellum (corrected p < 0.05). This increased brain activation corresponded with task-performance, suggesting that more brain resources are needed to complete death-me identifications. These results suggest that death-me implicit identifications involve resolving conflict between self and death representations in the brain, and marks an important step towards characterizing neural mechanisms contributing to suicidality.

Introduction

Over the last 20 years, suicide rates have steadily increased¹ with veterans twice as likely to die by suicide compared to US civilians^1,2. This risk is even higher for veterans with traumatic brain injury or psychiatric diagnoses (e.g., posttraumatic stress disorder [PTSD] or depression). In the last 50 years, many studies have explored ways to predict future suicide risk with limited success³, which has subsequently impacted our ability to prevent and treat suicidal thoughts and behaviors⁴. One potential contributor to our limited progress is that prior research has largely been based on a restricted set of traditional predictors of suicide risk that rely on self-report (i.e., questionnaires or surveys, history of STBs, comorbid depression), which are often used in isolation³. Therefore, it is necessary to identify novel alternative predictors of suicide risk that can be used in conjunction with traditional measures (see^3,5,6 for more in-depth discussions on this topic). Specifically, neurobiological measures of suicide-specific cognition may offer potential novel alternative predictors of suicide risk.

Over the last 10 years, the number of studies investigating neurobiological predictors of suicide risk has grown exponentially⁷. Many of these studies have used functional magnetic resonance imaging (fMRI) to identify brain activation related to suicide, implicating brain regions involved in cognitive control^7–12 and in affect and rumination^{7–9,11–13}. Despite some convergence, functional brain markers of STBs remain inconsistent with limited predictive utility, which could be driven by the heterogeneity in the fMRI methodology and paradigms used across the literature (e.g., resting state fMRI, emotional faces tasks, continuous performance tasks). Further, these fMRI paradigms assessed less direct aspects of suicide-specific cognition, such as performance on a go/no-go task is related to but not necessarily targeting suicide-specific cognition⁶. Therefore, it is difficult to determine if these neurobiological predictors are specific to suicide risk or related to common comorbid psychiatric conditions like depression or PTSD. Assessing brain activation related to suicide-specific cognition would improve upon the previous literature by precisely investigating brain activation that is specifically related to STBs.

One promising measure of suicide-specific implicit cognition is the Suicide Implicit Association Test (S-IAT) developed by Nock and colleagues¹⁴. The S-IAT measures the implicit identification between oneself with death (death-me implicit identification) life (life-me implicit identification) by categorizing the reaction time of death, life, me, and other words. Stronger implicit identification of self words with death words has been shown to predict a 6-fold increase in suicide attempts 6 months after discharge from the psychiatric emergency department¹⁴. Similar patterns have been identified across multiple populations and age groups^15–17. Although implicit identifications are typically measured as a single score using all trials (i.e., d-score), there is also some evidence that specific word categories are particularly predictive of suicide risk¹⁸. Specifically, one study found that a weakened implicit identification between life-me was associated with having a history of a suicide attempt, more so than a stronger implicit identification between death-me. Therefore, the S-IAT may provide a promising means to predict suicide risk and augment suicide prediction in conjunction with traditional measures.

Despite the behavioral utility of the S-IAT in predicting suicide risk, at the time of this paper, only one study has investigated fMRI brain activation related to death-me implicit identifications. Ballard et al. (2019) identified a network of brain regions that was more active during death-me implicit identifications relative to life-me implicit identifications during the S-IAT. This network of brain regions included brain regions involved in salience (insula²⁰), attention (angular gyrus²¹), self-referential thought (middle temporal gyrus^22–26), and emotion regulation (ventrolateral prefrontal cortex^27,28). Together, these findings suggest that suicide-specific cognitions may involve a broad network of neural mechanisms distributed across the salience network, default mode network, and frontoparietal control network.

Ballard and colleagues have laid important groundwork in uncovering brain activation related to suicide cognition, however, there were several limitations that are important to address. First, their sample consisted of healthy young adults carefully screened to rule out those with any diagnosed psychiatric disorder (i.e., non-clinical). Individuals at higher levels of suicide risk and trauma-exposed Veterans, often have one or more psychiatric disorder^29,30. Therefore, it is important to understand if the results identified by Ballard and Colleagues generalize to a clinical sample of Veterans with higher levels of psychopathology. Another limitation of Ballard et al.’s study design is that they did not use the standard S-IAT, instead they implemented a shorter modified S-IAT with an augmented word list. Therefore, the S-IAT and corresponding results from Ballard et al., are not comparable to the rest of the S-IAT literature.

To address these issues, the current study aims to replicate and extend Ballard’s findings by measuring brain activation during the standard S-IAT in a clinically heterogenous sample of trauma-exposed Veterans. Overall, we predict that we will find brain activation patterns similar to Ballard and colleagues. To extend these findings, we also explore if specific word categories (life vs. death) produced divergent results at the behavioral and neural level.

Methods

Participants

Participants were recruited as part of their participation in the longitudinal cohort study at the Translational Research Center for Traumatic Brain Injury and Stress Disorders (TRACTS) of the Veteran Affairs Boston Healthcare System. Inclusion criteria for enrollment were veterans ages 18–65, with service during OEF, OIF, OND, or scheduled deployment. Exclusion criteria included a history of neurobiological illness, history of seizures, or cognitive disorder not related to TBI, current diagnosis of schizophrenia spectrum, other psychotic disorders, current bipolar or related disorder not related to trauma or PTSD, or current active suicidal and/or homicidal ideation, intent, or plan, requiring crisis intervention. A detailed description of the recruitment, inclusion/exclusion criterion, and general characteristics of the TRACTS cohort are provided by McGlinchey and colleagues³¹. Broadly, participation in the longitudinal cohort study required a day of comprehensive medical, psychiatric and neuropsychological assessment, followed by an MRI session in the afternoon. For a subset of participants (n = 42; see the following text), this included completion of the S-IAT with concurrent fMRI. All procedures for this study were approved by VA Boston Healthcare System Institutional Review Board (No. 1645530).

Clinical Assessment:

Suicide risk was assessed using the Beck Scale for Suicide Ideation (BSS³²). Specifically, we reported the total score (sum of the first 19 items on the BSS) and question 20 which assessed the presence and frequency of past suicide attempts. The BSS total score was used to determine if a participant had STBs by reporting a total score on the BSS of three or more, or by endorsing question 20 was used to determine if a participant had a history of suicide attempt (SA). As part of their clinical assessments, participants completed the Structured Clinical Interview for the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) (SCID-5³³). To characterize our sample, we reported current mood disorder (current disorder from modules A or D), current anxiety disorder (current disorder from module F), and current substance use disorder (current disorder from module E). Lastly, since this is a trauma-exposed veteran sample, we also assessed current PTSD diagnosis with the Clinician-Administered PTSD Scale for DSM-5 (CAPS-5³⁴).

The Suicide Implicit Association Test (S-IAT)

The S-IAT was administered as close as possible to the standard administration to the S-IAT described by Nock and colleagues¹⁴, and used standard IAT practices for administration and scoring³⁵. Note, the previous fMRI S-IAT study (Ballard et al.) used eight additional words and ran two versions (counterbalanced for side of the screen the congruency effects were displayed and order of the congruent vs. incongruent blocks) of the S-IAT within participant, however we did not augment the word list and we ran the S-IAT once per participant.

Standard S-IAT;

Once the participants completed the practice IAT (description provided in the supplement), the participants began the MRI portion of the study (see Methods: MRI for details). The following section denotes any changes that were made in order to maximize the fMRI signal collected during the S-IAT. A full text description of the S-IAT is provided in the supplement and a visualization of the S-IAT is provided in Figure 1.

Figure 1:

A visualization of the Suicide Implicit Association Test (S-IAT). A.) Word List from the S-IAT. This provides a list of all the words and categories used to measure implicit identification between self and death. B.) An illustration of the S-IAT. Words are presented one by one in the center of the screen for 2 seconds. Participants must categorize the word into either the left or right category at the top of the screen. During the ITI, a cross hair is presented in the center of the screen. C.) A visualization of the standard S-IAT seven-block design. S-IAT = Suicide Implicit Association Test; ITI = Inter Trial Interval;

Modifications for fMRI:

The S-IAT was modified from its original form to maximize the signal we could measure from the concurrent fMRI. The original S-IAT does not have a standard trial length, instead the next trial is initiated upon response to the previous trial. Event-based designs in fMRI benefit from set trial lengths and jittered ITIs, which standardizes the amount of data collected for each subject, optimizes the deconvolution of overlapping events, and minimizes time-on-task differences between subjects that could impact the quality of the BOLD signal. Changing the task so that each participant experiences each stimulus for the same amount of time (two-second trial length) with jittered ITI will optimize this version of the S-IAT to measure BOLD activation during fMRI.

Calculating the d-score:

The d-score is the primary outcome from the S-IAT. A positive score indicates that a participant is faster during death-me implicit identification (higher suicide risk) while a negative score indicates that a participant is faster during life-me implicit identifications (lower suicide risk). The d-score was used to explore the relationship between brain activation during death-me implicit identification and performance on the S-IAT. A full description of how the d-score is calculated is provided in the supplement.

Congruency effects by word category:

We also explored potential differences in the congruency effect (incongruent-congruent conditions) by word category (Death, Life, Me, and Not Me). For each word category, a difference score was calculated between the mRT during that word’s incongruent trials and the mRT during the same word’s congruent trials. For these difference scores, a larger value indicates slower reaction times during the incongruent trials compared to the congruent trials. These difference scores were used to explore potential relationships between brain activation and behavior related to differences in implicit identification for specific word categories.

MRI acquisition and preprocessing

Neuroimaging data was collected during participation in TRACTS at the Neuroimaging Research for Veterans (NeRVE) center on a 3T Siemens MAGNETOM Trio System (n = 9) until the scanner was upgraded to a 3T Siemens MAGNETOM PrismaFit (n = 33). Structural (MPRAGE) scans were obtained using a 12-channel head coil and the following parameters: repetition time (TR) = 2350 ms, echo time (TE) =3.32 ms, flip angle =7◦, acquisition matrix =256 ×256 ×176, voxel size =1 mm3, resolution =3.0 ×3.0, and slice thickness =3.75 mm. The task fMRI scan with concurrent S-IAT was obtained using a 32-channel head coil and the following parameters: multiband acceleration factor = 8m, 2.0mm voxels, TR = 0.750s, TE = 34ms, 64 oblique axial slices. The study stimuli for the S-IAT were presented using a mirror in the scanner bore via a back projection screen. We used fMRIprep⁶⁶, which combines best-practice preprocessing techniques, to preprocess our structural and functional MRI data. The specific details of our preprocessing pipeline are provided in the supplement.

MRI Data Analytic Approach

First Level Models:

After preprocessing, first level general linear models were calculated using AFNI 3dDeconvolve^36,37. These first level GLMs included nuisance regressors, block regressors, and event regressors. The nuisance regressor includes six rigid body motion regressors, mean white matter and CSF timeseries, and five polynomial trends to account for linear and non-linear scanner drift. We included one type of block regressor which indexed six, 10 second instruction blocks. The event regressors included incorrect trials, and 12 additional trial regressors which correspond to each word category (death words, life words, me words, not me words) during practice trials, congruent trials, and incongruent trials.

Second Level Models:

We conducted a linear mixed-effects (3dLME³⁸ in AFNI) to investigate group-level brain activation related to death-me implicit identifications across the whole brain. This model included two within-subject factors: congruency (congruent implicit identification [i.e., Me-Life or Not Me-Death pairing] vs incongruent implicit identification [i.e., Me-Death or Not Me-Life pairing]) and word (Death, Life, Me, and Not Me). This allowed us to observe a main effect of congruency, and an interaction between congruency and word to determine if certain word categories show unique congruency effects in the brain. A random effect was included to account for subject-level deviations from the group average. To correct for multiple comparisons, we used a voxel-wise and cluster threshold approach. We obtained our cluster size threshold at α = 0.05 using the 3dClustSim function in AFNI^39,40 with a nominal threshold of p = 0.001 and observed smoothness of estimated residuals using 3dFWHMx (ACF parameters = 0.83 3.93 13.50). After completing 10,000 Monte Carlo simulations, the 3dClustSim function identified a 38-voxel cluster level threshold corrected for multiple comparisons (FWE p < 0.05).

Analysis Plan

First, we described performance on the S-IAT, as well as determined if congruency effect varied by word type with a repeated measures ANOVA (congruency by word). Then we examined a main effect of congruency in the brain to identify brain activation related to incongruent trials (i.e., death-me implicit identification) relative to congruent trials (i.e., life-me implicit identification). Next, we investigated if there was an interaction between the congruency and word category. Lastly, we determined if the brain activation or behavior during the S-IAT was related to STBs using Welch’s two sample t-tests comparing brain and behavior between those with and without STBs.

Controlling for confounds and exploring clinical covariates: We explored the potential impact of potential confounds and clinical covariates on behavior and brain activation. First, we investigated if average movement in the scanner (average framewise displacement) and change in scanner predicted the main effect of congruency in the brain (averaged whole brain map) using linear regression. We also simultaneously explored within the same linear model the role of clinical comorbidities (i.e., PTSD, mood disorder, anxiety disorder, or substance use disorder) in predicting the main effect of congruency in the brain. We conducted a parallel analysis to determine if average movement in the scanner, change in scanner, or comorbid diagnosis predicted performance on the S-IAT (d-score).

Results

Demographics and Clinical Characteristics

This study included 42 veterans that completed the S-IAT with concurrent fMRI. After removing participants who had excess motion (n = 2) and removing those who had not completed the clinical assessment portion of this study (n = 3), the following results included 37 veterans (mean age = 43.00, SD = 9.11). We defined STBs in this study as either reporting a history of a suicide attempt on the BSS and/or reporting a score of three or more the BSS total score, consistent with prior work^41,42. Of these 37 participants, five reported STBs (BSS total score: mean = 4.4, median = 3, range: 0–12). A full breakdown of the demographics and clinical characteristics of this sample is provided in Table 1.

Table 1:

Demographics and Clinical Characteristics of Study Sample. Education: This scale ranges from 1 (first grade) to 20 (Doctorate), with a score of 12 indicating a high school education.

Demographics and Clinical Characteristics
	All Participants (n = 37)	No STBs (n = 32)	With STBs (n = 5)
Age: mean (SD)	43.00 (9.11)	43.72 (9.48)	38.00 (5.16)
Gender (% Male)	95%	93.75%	100%
Education (SD)	15.76 (2.22)	15.88 (2.28)	15.00 (1.73)
Race/Ethnicity
American Indian	0.0%	0.0%	0.0%
Asian	3.0%	0.0%	25.0%
Black	11.0%	12.5%	0.0%
Pacific Islander	3.0%	0.0%	25.0%
Other	3.0%	3.0%	0.0%
White	89.0%	90.6%	75.0%
Hispanic	8.0%	6.3%	25.0%
d score (SD)	−0.80 (0.44)	−0.81,(0.44	−0.75 (0.29)
SA History	11.0%	0.0%	80.0%
Endorsed STBs	14.0%	0.0%	60.0%
PTSD Diagnosis	50.0%	48.4%	60.0%
Mood Disorder	14.0%	9.4%	40.0%
Anxiety Disorder	11.0%	6.3%	40.0%
Substance Use Disorder	24.0%	25.0%	20.0%

SD = standard deviation, STB = Suicidal Thoughts and Behaviors, SA = Suicide Attempt, STB = suicidal thoughts and behaviors.

S-IAT Behavior

Overall, most participants (n = 36) exhibited a greater implicit identification between self and life words (M = −0.80, SD = 0.42, Range = −1.51 – 0.37). Only one participant exhibited a greater implicit identification between self and death (i.e., a positive d-score; d-score = 0.37). As reported in Table 1, those with STBs had more positive d-scores than those without STBs, numerically speaking. The percentage of non-high-risk samples that displayed implicit death-me identifications is around 7%^19,43. In our sample, 3% displayed implicit death-me identifications and 14% reported STBs (i.e., previous suicide attempts or current suicidal ideation). Therefore, our sample had relatively low implicit biases towards self and death, which was similar to previous studies that investigated the S-IAT in non-clinical samples.

We used a repeated measures two-way (congruency by word category) ANOVA to determine if the behavioral congruency effect varied by word categories. Both main effects, congruency (F(1,36) = 91.46, p < 0.001) and word category (F(3,108) = 43.42, p < 0.001), were significant, as well as the interaction (F(3,108) = 18.26, p < 0.001). Post-hoc analysis used paired samples t-tests to determine which word categories showed larger differences in the reaction time between incongruent and congruent contexts. This analysis determined that the difference between Incongruent and Congruent trials was significantly larger for life words, compared to Death (t(36)=4.83, p < 0.001), Me(t(36)=5.88, p < 0.001), and Not Me words (t(36)=6.75, p < 0.001; Figure 2.A). However, it is important to note that all word categories exhibited a significant difference between incongruent and congruent contexts (ps < 0.001), which suggests a quantitative, rather than qualitative, difference in Congruency effects across word types.

Figure 2:

A.) Differences in Reaction Time by Word Category. B.) Visualization of the brain regions that showed increased activation during incongruent trials relative to congruent trials. C.) Significant interaction in brain activation between congruency and word. The right cuneus, had a significant congruency by word interaction, showing a larger congruency effect for death and me words compared to life and not-me words. RT = Reaction Time.

S-IAT Brain Activation

Our primary aim was to examine differences in brain activation between incongruent trials (e.g., death-me and life-not me) and congruent trials (e.g., life-me and death-not me). Greater activation was seen across a wide number of brain regions (22 clusters). To aid in the interpretation of results and visualize discrete brain clusters, we used a highly conservative nominal threshold (p<1×10⁻⁷) while keeping the same cluster size threshold (p < 0.05, cluster size >38 voxels). Significant clusters were distributed across the whole brain, including clusters in the occipital lobe, parietal lobe, temporal lobe, frontal lobe, and cerebellum (Figure 2.B and Table 2). Next, we identified only a single cluster in the brain that had a significant congruency by word interaction in the right cuneus (nominal threshold (p<0.001, cluster threshold p < 0.05, cluster size >38 voxels; Figure 2.C). It showed a greater congruency for death and me words, and smaller congruency effects for life and not me words. These results suggested robust brain activation related to the context of congruency, and little evidence that this effect varied by word type.

Table 2:

Significant clusters in the brain that show increased activation during incongruent trials compared to congruent trials.

Index	Hemisphere	Brain Region	Cluster Size	MNI Coordinates for center of mass			Correlation with d-score
				x	y	z
1	Left/bilateral	IPS and precuenus	2018	−35	−65	45	−0.30
2	right	IPS	1563	27	−73	45	−0.35*
3	bilateral	cerebellar crus II	811	7	−79	−35	−0.46**
4	right	cerebellum (VIII)	425	27	−71	−53	−0.37*
5	left	precentral (premotor eye fields)	355	−43	−1	39	−0.27
6	left	fusiform	174	−37	−63	−17	−0.36*
7	right	anterior insula	169	41	23	−5	−0.27
8	right	middle temporal gyrus	158	69	−33	−11	−0.08
9	right	cerebellum VI	122	27	−61	−31	−0.35*
10	left	DLPFC	121	−35	−1	69	−0.09
11	right	DLPFC	118	41	3	53	−0.16
12	right	inferior parietal lobule	115	49	−47	51	−0.10
13	left	DMPFC	80	−5	13	53	−0.34*
14	right	lingual gyrus	77	11	−81	1	−0.47**
15	left	DLPFC	69	−51	27	31	−0.27
16	left	cerebellum (VIII)	67	−31	−65	−53	−0.21
17	right	inferior parietal lobule	59	33	−45	39	−0.15
18	right	lingual gyrus	58	9	−73	−7	−0.37*
19	left	middle temporal gyrus	51	−55	−55	17	−0.04
20	right	inferior occipital gyrus	46	37	−83	−15	−0.40*
21	left	DLPFC	43	−49	29	23	−0.34*
22	right	middle frontal gyrus	38	55	25	35	0.08

^{* =}p < 0.05,

^{** =}p < 0.01,

Interparietal Sulcus = IPS, Dorsal Lateral Prefrontal Cortex = DLPFC, Dorsal Medial Prefrontal Cortex (DMPFC)

S-IAT Brain and Behavior

We explored if the brain activation that differed between incongruent and congruent contexts was also related to performance on the S-IAT. First, we examined the correlation between activation computed for participants across the congruency brain map with S-IAT d-scores. This correlation was significant (r = −0.37, p = 0.023), indicating that faster responding to life-me implicit identifications (negative d-score) was associated with increased brain activation during death-me implicit identifications. Ten of the 22 clusters within the congruency map were significantly negatively correlated with the d-score (see Table 2), indicating that increased activation in these clusters corresponded with more negative d-scores. This suggests that there were significant general relationships between brain activation across the congruency map and performance on the S-IAT, with some specific neural regions exhibiting stronger correlations with S-IAT d-scores than others.

S-IAT Brain and Behavior relationship with Suicidal Thoughts and Behaviors (STBs)

STBs, in this study, were defined as either a history of a suicide attempt on the BSS and/or reporting a BSS total score of three or more. Based this definition, five participants reported STBs. Welch’s two-sample t-tests determined that there were no differences between those with and without STBs in their brain activation related to congruency (with STB mean = 0.12, without STB mean = 0.11, t = 0.19, p = 0.85) or their d-score (with STB mean = −0.75, without STB mean = −0.81, t = 0.35, p = 0.73). These results indicate that STBs do not predict brain or behavior during the SIAT, however our sample was not powered to detect differences, and our sample included veterans who were relatively at low risk for STBs (see exclusion criteria). These results should be interpreted with caution.

Controlling for Confounds and Exploring Clinical Covariates

Lastly, we examined if other characteristics of our sample were related to behavior or brain activation during the S-IAT. Using a linear regression, we found that none of the potential confounds (framewise displacement, change in scanner) or clinical comorbidities (PTSD, mood disorder, anxiety disorder, or substance abuse disorder) predicted the main effect of congruency in the brain (ps>0.47; β range: 0.06–0.43) nor predicted the behavior during the SIAT (d-score; ps>.0.16; β range: 0.15–1.08). These results suggest that our results were not influenced by participant motion or the switching between two different MRI scanners. Further, these regressions did not provide evidence that behavior and brain activation during the S-IAT was related to these clinical diagnoses, although our sample was not necessarily powered to detect such effects.

Summary of Results

This study showed strong incongruency effects, indicative of a “healthy” (i.e., non-suicidal) response pattern, as well as robust brain activation to incongruency relative to congruency during the S-IAT. Word-specific effects were observed, to lesser degree. Life words showed the largest difference in mean RT between incongruent and congruent trials; however, all four word categories displayed significant slowing in reaction time during incongruent contexts. When comparing the BOLD signal between incongruent (e.g., death-me) and congruent trails (e.g., life-me), there was a robust difference in brain activation across a distributed set of brain regions. Two of the brain regions we identified had previously been implicated in death-me implicit identifications: the right insula and right MTG. In addition, we observed a significant relationship between congruency-related brain and behavior across individuals that was not previously been identified in the literature. The results of this study demonstrate that there are consistent patterns of behavior and brain activation related to death-me implicit identifications in clinical sample with low-risk for suicide.

Discussion

This study is one of the only two studies to investigate brain activation involved in implicit death-me identification and the first to investigate this brain activation in trauma-exposed Veterans. The aims of this study were to replicate and extend the previous study that investigated brain activation during the S-IAT¹⁹, alongside methodological and analytic extensions. We partially replicate the previous study in that the right anterior insula and right middle temporal gyrus (MTG) are active during death-me implicit identifications (incongruent vs congruent contexts). In addition to this overlap with the previous study, we also identify brain activation across a widely distributed network of brain regions that exhibits greater activation during death-me contexts. Unlike Ballard and colleagues, however, we did observe a relationship between performance on the S-IAT (as measured by the d-score) and brain activation. Specifically, stronger life-me implicit identification is correlated with greater relative brain activation during incongruent trials (i.e., death-me trials). This suggests that stronger neural activation during death-me trials may be adaptive, given its relationship with stronger life-me implicit identification. Together, the results of this study provide important information regarding brain activation related to implicit death-me identifications in a clinically heterogeneous veteran sample.

Behavior during the S-IAT

In this clinically heterogeneous veteran sample, 3% (n = 1) had implicit death-me identifications (a positive d-score). Other studies that investigate death-me implicit identifications in non-clinical samples report around 7–8% of their sample showing an implicit death-me identification^19,43. In a sample of veterans at high risk for suicide (i.e., post psychiatric hospitalization^15,44), the average d-score is negative (mean ≈ −0.46, SD ≈ 0.44). Therefore, a positive d-score is relatively rare, even in high-risk populations. In addition to measuring the d-score, we also investigated if there was any word category specificity in the congruency effect. We found that the difference between incongruent and congruent trials of all word categories significantly differed, and this difference score was significantly larger for Life words. Despite the many studies that used the d-score, only one has investigated differences in congruency by word category in the S-IAT. O’Shea and colleagues (2020) found that weakened life-me implicit identification was predictive of a previous suicide attempt, more so than stronger death-me implicit identifications, while stronger death-me implicit identifications were predictive of the recency of a previous suicide attempt¹⁸. Both the current study and the work by O’Shea suggest that certain word categories may contribute differentially to the omnibus d-score and could differentially predict suicide risk¹⁸. Thus, it is critical to examine if the brain differentially activates by the congruency of specific word types.

Brain Activation during the S-IAT

Right Insula and Right Middle Temporal Gyrus (MTG):

This study showed a widely distributed and robust brain activation related to implicit death-me identifications, as reflected by a contrast of incongruent vs. congruent trials. Two of these brain regions overlapped with the study by Ballard et al. (2019): the right anterior insula and right middle temporal gyrus (MTG). The insula is a part of the salience network (SN; Menon & Uddin, 2010). The theorized role of the insula, and in particular the anterior insula, in the SN is to act as a “gatekeeper of executive control⁴⁶.” According to this theory, the anterior insula triages and integrates multisensory stimuli so it can initiate a switch from “resting-state” brain systems to higher-order cognitive brain networks to meet external demands. In addition to its role in executive control, the insula has also been implicated in social-emotional processing⁴⁷, pain perception^48,49, interoception^47,50,51 and death-me implicit identifications¹⁹, which are all processes that have been linked to suicide risk^8,51,52. The insula has also been implicated in suicide and suicide risk⁷, suggesting that not only is the insula multifaceted, but it may play a role in suicide specific cognitions. The MTG is involved in language memory processing^53,54, as well as visual⁵⁵ and auditory processing⁵⁶. The MTG is also a part of the default mode network^57–59, which is a network of brain regions that are active when engaging in a wide variety of internal mental processes⁶⁰. As such, the MTG has also been implicated in tasks that involve future thinking and remembering the past⁶¹, and self-referential processing⁶². Resting state connectivity studies of MTG and DMN have been associated with history and future suicide attempt^11,12. This suggests that MTG is involved in self-referential processing, which may play a role in suicide-specific cognitions. Along with the current study, the literature indicates that the MTG and insula play a role in suicide risk.

Whole Brain Activation during the S-IAT:

In addition to regions identified in previous research, we found an expansive network of brain regions that underlie incongruent implicit identification of self and death including the interparietal sulcus (IPS), precuneus, dorsal lateral prefrontal cortex (DLPFC), dorsal medial prefrontal cortex (DMPFC), cerebellum, cuneus, lingual, and the fusiform gyrus. These visual and cognitive control brain regions exhibited greater activation when participants responded more slowly during incongruent trials, suggesting additional processing demands are required for incongruent trials. Additionally, some of these brain regions have previously been implicated suicide risk. For instance, increased DLFPC activity has been found in adults with STBs while completing “cold” cognitive tasks, like inhibitory control^9,63–65. While the role of the cerebellum and occipital gyri in suicide risk are not well understood, their structure and function have previously been linked to STBs (cerebellum^7,66, occipital gyrus⁶⁷). Future studies in higher risk sample, using a range of suicide-specific and non-specific tasks, will help determine the degree to which these activations are associated with suicide-related cognitive processes or more general information processing. In an additional exploratory analysis, the right cuneus had a significant congruency by word interaction, displaying more activation during incongruent categorization of death and me words (Figure 2.C.). While this right cuneus cluster was not apart of the main effect congruency brain map, this result suggests there is some unique brain activation sensitive to congruency by word category interactions. Future work will need to investigate these interactions in larger samples with suicide risk to determine the potential sensitivity of the cuneus to implicit death-me identifications.

Brain and Behavior during the S-IAT

Our study is the first to observe that brain activation during death-me implicit identifications corresponds to behavior reflecting the strength (or weakness) of implicit death-me identification (i.e., the d-score). Increased brain activation during implicit self-me identifications corresponded to slower reaction times. This supports the above interpretation that stronger implicit life-me identifications (negative d-scores) require more brain resources to complete death-me implicit identifications, leading to increased RT alongside increased brain activation during death-me implicit identifications, especially in brain regions involved in attention and visual processing.

Brain Activation during the IAT

Brain activation during other non-suicide-related IATs showed a consistent pattern of brain activation. Namely, these studies reported brain activation in the dorsal lateral prefrontal cortex (DLPFC), anterior cingulate cortex (ACC), and the ventral medial prefrontal cortex (VMPFC) during incongruent trials relative to congruent trials^68–70. The previous literature is partially consistent with current results, as we observed brain activation in a similar DLPFC region during incongruent trials relative to congruent trials. This suggests that prefrontal brain regions may be recruited during these more difficult (slower) categorizations. One IAT study observed increased insula activation, like that observed in S-IAT, when viewing pain inflicted on the racial in-group relative to the racial out-group⁷¹. Thus, it is possible that the insula activation we observed during self-death identifications may reflect reactivity to the association between oneself and physical harm. Other brain regions activate during the S-IAT may be more specific to self-death implicit identifications. For instance, MTG (which was not identified in the general IAT literature) may process information that is unique to both suicide risk^11,12, and self-death associations¹⁹, and could be an important target for future studies investigating brain markers of suicide risk.

Strengths, limitations, and future directions

Overall, we observed a brain-behavior relationship that previously had not been identified. This may be partially due to this study’s larger sample size, which improved our power to detect individual differences. Further, this study used the standard S-IAT, which more congruent and incongruent trials to average BOLD signal for statistical analysis compared to the modified S-IAT used by Ballard and Colleagues (2019). In addition, our study also included a clinically heterogeneous sample that may have had more distributed behavior during the S-IAT, so there were individual differences to detect. Lastly, our study modeled each trial in the S-IAT separately, which may have provided important trial-level information that was critical for identifying brain-behavior relationships. Regardless of the differences between study designs, more research is necessary to replicate and validate the results identified in this study, and future work should investigate how brain-behavior relationships may change as a function of suicide risk.

While this study replicated and extended the literature related to brain activation during the S-IAT, there are important limitations to this study. While we investigated brain activation during death-me implicit identifications in a clinical trauma-exposed Veteran sample, the sample size in our study was relatively small, especially for exploring individual differences related to clinical characteristics like depression or PTSD. Another limitation is that this study did not include any participants at high risk for suicide, and we had few participants in our sample with STBs, therefore, we could not make any inferences in the brain activation related to suicide risk. This highlights an important future direction, as exploring this brain activation in a high-risk sample will be a clear next step in this line of research. Another important future direction is to investigate how other symptoms and diagnoses (like PTSD and depression) play a role in the brain activation and behavior of the S-IAT. While, we did not detect any relationships related to other diagnostics, our study was not powered to do so. Relatedly, our study’s results are limited in its racial, ethnic, and gender diversity, and our results are primarily representative of white male veterans who are all at least high school educated. This highlights an additional future direction that adjusts recruitment to better target diverse populations, determining the generalizability of our results.

This study improves our understanding of brain mechanisms underlying self-death implicit identification, however, we have yet to determine how this brain network and activation pattern may change in people who are at risk for suicide. If this brain activation is critical for resolving the incongruency of self-death implicit identification, then those with high suicide risk could exhibit reduced activation (i.e., hypoactivation) in this task, indicating reduced conflict in implicit self and death identification. If these brain regions are critical for resolving global incongruency of implicit self identifications, this network will show hyperactivation corresponding to the level of incongruence of self-identification, irrespective of the context of life or death. Further, we don’t know if this network supports global conflict resolution –indicating it could be hypoactive in those with suicide risk across more domain-general control tasks –or if it supports specific self-referential processing. Investigating this network across a variety of cognitive tasks in those at risk for suicide will aid in determining which of these hypotheses, if any, are supported.

Conclusions

The purpose of this study was to replicate and expand upon the initial work by Ballard and colleagues in a clinically heterogeneous veteran sample. Similarly to Ballard, we found brain activation in the right anterior insula and right middle temporal gyrus during death-me implicit identifications. In addition, we identified brain activation during death-me implicit identifications in a widely distributed set of visual and cognitive control regions. This brain activation was strongest in those that had the weakest implicit identification between self and death, indicating an increased neural effort to associate self and death in people with strong implicit identifications between self and life. These results further demonstrate that brain activation related to suicide-specific cognition can be identified and evaluated. In future work that investigates death-me implicit identifications in those with a history of suicide attempt, current ideation, or recently hospitalized populations will improve our understanding of the neural mechanism that support suicide-specific cognitions and lead to opportunities to develop novel treatment and interventions for those with STBs.