The effects of traumatic brain Injury, post-traumatic stress disorder on Amyloid-β associated network hyperconnectivity and progression of gray matter atrophy

Highlights

• •Amyloid-β associated hypersynchrony is an early sign of  preclinical AD.
• •AD risk factors TBI and PTSD increase the occurrence of  hypersynchrony.
• •AD pathology increases hypersynchrony dwell-time.
• •Progressing gray matter atrophy does not affect hypersynchrony dwell-time.

Abstract

Background

Amyloid-β associated network hypersynchrony is an early manifestation of pre-clinical Alzheimer’s Disease (AD). The overall goal was to investigate a. how TBI and PTSD influence hypersynchrony expression and b. how progressing gray matter atrophy affects hypersynchrony expression.

Methods

T1-weighted images, resting-state fMRIs and amyloid-β SUVRs were obtained from 234 DoD-ADNI subjects with or without TBI and/or PTSD. The denoised BOLD signal from 382 rois was extracted with CONN and dynamic resting state analysis was used to identify 8 states including one corresponding to a hypersynchrony state (HSS). SuStaIn with gray matter volumes and amyloid-β SUVR as inputs was used to identify 2 subtypes with progressive gray matter loss.

Results

HSS dwell-time correlated positively with amyloid-β (Kendall tau = 0.125,p = 0.047) and tau Braak stage 5&6 SUVR (Kendall tau = 0.200,p = 0.035). TBI increased the likelihood to observe the HSS (81 % with vs. 18 % wo TBI p < 0.001) as did a diagnosis of PTSD (67.4 % with vs. 32.6 % wo PTSD, p = 0.003). The SuStaIn subtypes differed mostly by the timing of the amyloid-β build-up but not by atrophy pattern. Subtype 2 had higher amyloid-β loads and longer HSS dwell-times than subtype 1 that had higher CAPS scores than subtype 2. Gray matter atrophy did not influence HSS dwell-time.

Conclusion

TBI and PTSD increased the likelihood to observe HSS. HSS dwell time was determined by AD pathology severity. The subtype characteristics indicate that PTSD drives gray matter loss in subtype 1 and AD pathology that in subtype 2. Severity of gray matter atrophy influenced neither HSS occurrence nor intensity.

1. Introduction

Network hypersynchrony is one of the earliest manifestations of incipient Alzheimer’s disease (AD) and is thought to play a critical role in the spread of AD pathology and the development of cognitive impairment (Kazim et al., 2021, Kamondi et al., 2024). The mechanisms underlying AD related hypersynchrony are complex and still not fully understood (Kamondi et al., 2024) but the currently available evidence suggests that a dysfunction of inhibitory interneurons probably plays a major role (Palop, xxxx, Verret et al., 2012). Network hypersynchrony can be demonstrated invasively in animal models of AD by single cell recordings, Ca2 + imaging etc., and non-invasively in patients with preclinical as well as clinically manifest AD by electroencephalography, magnetencephalography (MEG) or task-based and task-free functional fMRI (Kazim et al., 2021). These properties suggest that network hypersynchrony has the potential to become a biomarker for early AD or even a treatment target (Shah et al., 2016).

The overall goal of this study was to learn more about the mechanisms of network hypersynchrony and its impact on AD development and progression. This was done by addressing two aims. The first was to replicate the findings of a previous study in a large population with an increased risk for AD. This previous study had used dynamic resting-state analysis in a population of cognitively intact participants with and without increased amyloid-β load to demonstrate that network hypersynchrony, i.e., its graph analysis equivalent hyperconnectivity, was restricted to a single brain state whose dwell time correlated with amyloid-β load (Mueller, 2019). The participants of the new study were Vietnam veterans who had experienced at least one non-penetrating traumatic brain injury during their service (past TBI) and/or had been diagnosed with post-traumatic stress disorder (PTSD). Epidemiological studies have shown that past TBI and PTSD both increase the risk to develop AD in later life (Guo et al., 2000, Plassman et al., 2000, Meziab et al., 2014, Yaffe et al., 2010). Past TBI is of particular interest in the context of network hypersynchrony. TBI has not only been shown to impair inhibitory interneurons and thereby to induce short but also long-term cortical and subcortical hypersynchrony (Nolan et al., 2022, Carron et al., 2020, Pavlov et al., 2011, Hameed et al., 2023) but also to lead to an upregulation of amyloid precursor protein in the acute phase and to axonal amyloid-β accumulation and tau pathology in the long term (Iwata et al., 2002, Johnson et al., 2012). It was therefore assumed that a history of past TBI would not only increase the likelihood to observe the hyperconnectivity state in this population but also that the state would be more pronounced, i.e., last longer, than in veterans without TBI history.

The second aim was to explore the relationship between network hyperconnectivity and the development of gray matter atrophy and amyloid-β build-up in this at-risk population. Considering that hyperconnectivity is already observed in the early/preclinical AD stages, it was assumed that it would be most pronounced before the development of wide-spread gray matter atrophy in the early phase of the amyloid-β build-up and would subside once widespread atrophy had developed. The spatio-temporal pattern of gray matter atrophy in typical AD is quite well established thanks to the close relationship between gray matter atrophy with the progression of tau pathology that is captured by the Braak stages (Braak and Braak, 1996). However, TBI and PTSD cause gray matter loss on their own right (Wang et al., 2021, Belchev et al., 2022, Sandry and Dobryakova, 2021) which might alter the typical spatio-temporal progression pattern of AD related gray matter atrophy and introduce the possibility of more than one progression pattern. To account for this possibility, a recently developed unsupervised and data-driven machine learning algorithm called Subtype and Stage Inference or SuStaIn was used in this project (Young et al., 2018, Aksman et al., 2021). SuStaIn uses a spatiotemporal clustering approach to disentangle subtypes with distinct progression patterns from disease stages that capture the degree to which a particular subtype is expressed over time. To optimize statistical power and computational tractability as well as to emphasize the focus on the structure–function relationship in this study the staging and typing of gray matter volume loss with SuStaIn was done with four large regions of interest (rois) corresponding to four major functional networks that had been identified from the participant’s resting-state fMRI data (see Methods).

2. Materials and methods

2.1. Participants

Most of the data used in this project came from the DoD-ADNI data repository (n = 234). The DoD-ADNI project acquired MRI and PET to investigate if non-penetrating TBI and /or PTSD resulting from military service and other traumata increase the risk for later dementia in male Vietnam Veterans aged 60 to 80 years (Weiner et al., 2014). DoD-ADNI participants in this study were selected from the whole DoD-ADNI population based on the availability of a good quality T1 image at TP1, at least one good quality resting state fMRI (TP1 or later TP) and of a PET based amyloid-β load. In addition to imaging, all DoD-ADNI participants had also undergone medical and neurological examinations and extensive cognitive and mental health testing. From these, the Alzheimer’s Disease Assessment Scale – cognitive subscale (ADAScog) was selected as a measure of cognitive dysfunction and the Clinician-Administered-PTSD Scale for DSM 4 (CAPS) was used to assess current and lifetime PTSD. A CAPS current of ≥ 30 (CAPS < 30 used as threshold to define absence of PTSD in DoD-ADNI according to ClinicalTrial.gov) was used to identify 105 participants suffering from PTSD. TBI history was assessed with a structured interview that captured the number of non-penetrating TBIs (penetrating TBI was exclusion criterion) and TBI severity over lifetime based on presence/absence and duration of confusion/disorientation, memory gaps, loss of consciousness and hospitalization. Total number of non-penetrating TBI (npTBIcount) and a TBI severity score (npTBIseverity = no TBIs with loss of consciousness + no TBIs with hospitalization) were used to summarize TBI history. A npTBIcount > 0 was used to identify 117 participants with a history of TBI. Thirty-two of the 234 DoD-ADNI participants fulfilled the criteria of a typical aging control, i.e., had an amyloid-β SUVR of < 1.11, were cognitively intact (CDR-Total = 0) and did not meet the criteria PTSD and/or TBI outlined here.

In addition to the DoD-ADNI participants, 30 cognitively intact (CDR total = 0), amyloid-β negative (SUVR < 1.11) ADNI3 participants without a history of mood disorder or traumatic brain injury who had the same type of imaging data as the DoD-ADNI subjects were randomly selected from the ADNI3 data repository to be used as reference group to calculate z-scores for dynamic fMRI processing and subtype/stage identification. ADNI 3 is the continuation of the longitudinal ADNI project launched in 2004 that aims to identify biomarker(s) allowing for an early and accurate AD diagnosis and efficient monitoring of potential AD treatments (Mueller et al., 2005).

The project used completely de-identified data from the DoD-ADNI and ADNI data repository. Studies using completely de-identified data from data repositories where it is not possible to ascertain the subject’s identity are considered non-human subject studies and exempt from UCSF IRB review.

2.2. MRI acquisition

The DoD-ADNI and ADNI3 MR imaging data were acquired with the same harmonized protocol across all participating sites that had been optimized to provide comparable images from different platforms from Siemens, Philips and General Electric. The harmonization strategy and quality controls are described in detail in Mueller et al. (2005). The 3 T T1 weighted MPRAGE sequence with TR/TE/TI 2300/2.95/900  ms, sagittal, 1.1 × 1.1 × 1.2  mm resolution and the Task-free T2*weighted gradient echo EPI BOLD: TR 3000 ms, TE 30 ms, flip angle 92°, partial Fourier 1, voxel size 3.4 mm isotropic. 197 volumes were used for this project. 148 of the 234 DoD-ADNI with initial T1 weighted images and 30 of the 30 ADNI3 with initial T1 weighted images (TP1) had a follow-up exam approx. 12 months later (TP2). 119 of the 234 DoD-ADNI participants with a TP1 T1 weighted image also had a resting state fMRI acquisition at this TP that met the quality criteria for dynamic fMRI processing as did 122 of the 148 participants with TP2 T1 weighted image.

2.3. Image processing and analysis

2.3.1. PET

Amyloid-β and Braak SUVRs for DoD-ADNI participants were obtained from the DoD-ADNI data repository (UCBERKELEYAV45_20190808) and (UCBERKELEYAV1451_PVC_20201208). Amyloid-β SUVRs for the ADNI3 reference group were obtained from the ADNI data repository (UCBERKELEYAV45_8mm_02_17_23_15Jun2023).

2.4. Task-free fMRI Preprocessing

Each subject’s T1 weighted image underwent tissue segmentation (gray matter, white matter, csf) with the new segment algorithm as implemented in SPM12 (Ashburner and Friston, 2005). The resulting gray matter maps were warped onto a symmetrical gray matter template using DARTEL as implemented in SPM12, and the resulting transformations applied to all three tissue maps. The first 6 timeframes of the resting state fMRI were discarded to allow the MRI signal to achieve T1 equilibrium. The remaining timeframes/subject underwent slice time correction, motion correction and realignment onto a mean EPI image in the T1 image subject space, spatial normalization using the DARTEL transformation matrices generated in the previous step with re-sampling to a 1.5 x 1.5 x 1.5 mm resolution. Conn 18.b (Whitfield-Gabrieli and Nieto-Castanon, 2012), a SPM based toolbox was used for further processing including detection of motion outliers (timeframes with motion > 0.9 mm) with its ART routine, linear detrending and band pass filtering (0.015–0.09 Hz) with simultaneous denoising. The latter includes the aCompCorr routine to reduce the effects of physiological noise (eroded white and csf maps, 5 components each) and motion regression (6 affine motion parameters and 6 first order temporal derivatives). Global signal removal was not performed since this is known to falsely increase anticorrelations (Murphy et al., 2009). 382 cortical and subcortical regions of interest from the AICHA atlas (Joliot et al., 2015) were used to extract the time series of denoised mean BOLD signals for each participant.

2.5. Stationary fMRI analysis

After scrubbing of the ART outliers, correlation matrices were calculated from the 382 de-noised BOLD time-series of each of the 30 ADNI reference participants and averaged across all subjects. The modularity_und algorithm (gamma = 1, 1000 iterations, community structure with highest Q) (Rubinov and Sporns, 2011) from the BCT toolbox was used identify 4 network modules. (Fig. 1.d).

Fig. 1.

Flow chart of dynamic fMRI processing.

2.6. Dynamic fMRI analysis

Each time series was divided into sections using a sliding windows approach (window size 20 timeframes/60 sec, advanced with 1 TR) and the 382x382 correlation matrix for each window calculated (Fig. 1 a). The window size was chosen based on observations that robust estimations of the functional connectivity without loss of potentially interesting fluctuations are possible with window sizes around 30–60 s. Graph analysis was used to describe the interactions between the different rois in each window (Fig. 1b) The positive (pos) strength outputs for each window were combined to obtain a map showing the fluctuations of positive (pos) strength over the acquisition time for each roi for each subject, concatenated across subjects to obtain population maps of pos strength for each roi (Fig. 1 c) and converted into z-scores using mean and standard deviation of the roi strength of the 30 ADNI reference participants as reference. As a data reduction step in preparation for the state identification by cluster analysis, the roi z-scores/window were averaged across the 4 network modules identified in the stationary analysis (Fig. 1d). Hierarchical cluster analysis (Ward’s minimum variance method with cubic clustering criterion to identify optimal cluster number) with each windows 4 mean z-scores as input identified 18 clusters or brain states (Fig. 1e). The ART outputs were used to identify windows with motion outliers and to calculate the % of motion outliers for each cluster. Ten clusters with more than 20 % motion outliers were identified as motion clusters and not further evaluated. The remaining 8 clusters or brain states were evaluated after excluding windows with timeframes identified as motion outliers (motion window). Residual motion was assessed by calculating mean framewise displacement (fwd)/window and mean fwd/cluster. Eliminating windows with excessive motion results in a more rigorous elimination of motion artifacts than just eliminating the motion affected timeframe alone because it also eliminates timeframes with subthreshold motion that usually accompany timeframes with suprathreshold motion. The duration of each brain state or dwell time in a participant was calculated as the number of non-motion affected windows assigned to this state by the cluster analysis. A “representative” window for each state observed in an individual was identified by calculating the Euclidian distance between each window’s correlation matrix to that of all other windows assigned to the same state in this individual and averaging these distances to obtain a representation index (RI) for each window. The window with the lowest RI was chosen as the best representation of this state in this subject and was used to characterize the networks of interest during this state (Fig. 1f). Brain state network connectivity was characterized by averaging all representative windows assigned to the cluster and calculating each node’s positive strength and modularity (modularity_und algorithm, gamma range: 1 – 2.0, 1000 iterations, community structure with highest Q). Cluster 3 or brain state 3 was the most common state. It also had the highest Q indicating a good separation, a stable 4 network module configuration across the whole gamma range and was found in all participants. Its module 1 encompassed the sensorimotor and salience network, module 2 encompassed the default mode network, module 3 encompassed the executive network and module 4 the visual network. These four modules were chosen as rois for the SuStaIn analysis.

2.7. Structural Preprocessing

The T1 weighted image underwent tissue segmentation with CAT12.7 (https://www.neuro.uni-jena.de/cat), a toolbox implemented into the SPM12 (https://www.fil.ion.ucl.ac.uk/spm/) software package. CAT12 performs a bias correction, spatial normalization to the MNI space using the shooting algorithm (Ashburner and Friston, 2011) and tissue segmentation into modulated gray and white matter maps 31. CAT12 segmentation accounts for partial volume effects (Rajapakse et al., 1997, Tohka et al., 2004) by applying adaptive maximum a posteriori estimations (Cuadra et al., 2005) and using a hidden Markov Random Field model (Chow et al., 2023). It also calculates intracranial volume and total white matter hyperintensity volumes. Image quality control was done using the CAT12 provided QC measures. For images with two timepoints the longitudinal processing routine was used.

2.8. Structural Analysis: Subtype and stage Inference (SuStaIn)

TP1 data was used as discovery cohort and −lacking an external data set- TP2 data was used as validation cohort in the SuStaIn analysis. Each cohort was processed separately. The mean cortical and subcortical gray matter volumes were extracted from the modulated gray matter maps using the four cluster 3 network modules (roi) and each of these volumes adjusted for intracranial volume and age effects using a linear regression analysis. The adjusted roi volumes as well as each DoD-ADNI participants amyloid-β SUVR were converted into z-scores (defined so that higher z-scores indicate a higher amyloid load and more pronounced atrophy) using the roi volumes and amyloid-β SUVR of the ADNI3 participants as reference. The piecewise linear z-score model of the SuStaIn algorithm (Young et al., 2018, Aksman et al., 2021) with 25 start points, 1,000,000 Markov Chain Monte Carlo iterations and evaluating up to 3 clusters or progression subtypes was ran to identify distinct progression subtypes and distinct stages within each of these progression subtypes. This was followed by 10-fold cross-validation to evaluate the optimal number of clusters/progression subtypes (defined as cluster number with lowest cross-validation information criterion (CVIC)) that best describe unseen data and to assess the stability of progression subtypes across folds. Each of the SuStaIn subtypes and each of the stages is expressed to some degree (indicated by a maximum likelihood value ranging from 0 to 1) in each participant. Using a maximum likelihood cut-off > 0.5 for subtype and identifying the stage with the highest maximum likelihood within this subtype, each participant’s gray matter roid and amyloid-β SUVR z-scores were assigned to a progression subtype and a stage. Participants without abnormalities in roi volumes and amyloid-β SUVR were assigned to a non-progression subtype 0 without staging indicating that they were not different from the reference population.

Finally, the quality of the SuStaIn model was characterized by comparing its ability a. to detect the same subtypes and number of stages in the validation cohort, and b. calculating the percentage of participants assigned to the same subtype in the validation and the discovery cohort. The model was considered to be successful if it detected the same subtypes and progression stages in both cohorts and assigned more than 80 % of the participants to the same progression subtype in both cohorts.

2.9. Statistics

Kolmogorov-Smirnov tests showed that many of the variables of interest were not normally distributed. One-sided Kendall tau tests were used to investigate the relationship between amyloid-β SUVR and hyperconnectivity cluster dwell time in participants displaying this brain state. Robust ANOVA tests were used to investigate differences of ADAScog, CAPS, TBI severity, hyperconnectivity cluster dwell time (log transformed), amyloid-β SUVR and tau SUVR between subtypes. Fisher’s exact test was used to investigate how diagnostic groups were distributed between SuStaIn subtypes.

Correction for multiple comparisons was not performed for two reasons: 1. The characteristics of the hyperconnectivity state were based on previous studies (Mueller, 2019) making it possible to formulate an a priori hypothesis regarding its features (highest global positive strength, lowest Q, dwell time positively correlated with amyloid-β load. 2. The second part (characterization of the relationship of the hyperconnectivity state, progression subtypes etc. with PTSD and TBI severity) was descriptive and not explorative. It is acknowledged that the majority of significant findings in this descriptive section would not have survived a stringent correction for multiple comparisons.

3. Results

3.1. Population

Table 1 characterizes the population. Participants with PTSD were younger than those in the other groups. As to be expected, participants with a history of PTSD had higher CAPS current and CAPS lifetime scores. Participants with PTSD and TBI&PTSD had higher ADAScog values than those with TBI only or without TBI or PTSD. Amyloid-β SUVRs were not different between groups. The validation cohort was not different from the discovery cohort.

Table 1.

Study Population Characteristics.

Discovery Cohort	No TBI, No PTSD	PTSD	TBI	TBI &PTSD
No of Subjects	47	36	59	92
Age	70.7 (6.03)*	67.1 (2.28)	71.2 (5.68)*	69.3 (3.47)*
ADAS cog	8.96 (4.31)$	10.1 (3.60)	8.73 (4.49)*$	11.3 (4.81)
CAPS current	1.83 (4.03)*$#	57.8 (13.9)	4.64 (6.44)*$	46.5 (19.5)*
CAPS lifetime	5.19 (8.36)*$#	78.9 (19.0)	9.53 (9.80)*$	63.3 (17.4)*
TBI severity	NA	NA	1.61 (1.41)	1.41 (1.27)
Amyloid-β SUVR	1.08 (0.15)	1.00 (0.05)	1.07 (0.18)	1.08 (0.15)
Braak 1 tau SUVRa	1.83 (0.35)	1.69 (0.26)	1.91 (0.42)	1.82 (0.34)
Braak 2 tau SUVRa	1.40 (0.17)	1.31 (0.17)	1.43 (0.22)	1.35 (0.18)
Braak 3/4 tau SUVRa	1.73 (0.16)	1.65 (0.12)	1.74 (0.12)	1.69 (0.16)
Braak 5/6 tau SUVRa	1.80 (0.14)	1.76 (0.16)	1.82 (0.15)	1.81 (0.16)
LN Dwell brain state 7b	1.90 (0.6)	1.31 (0.86)	1.82 (1.03)	1.61 (0.96)
Validation Cohort
No of Subjects	36	25	49	38
Age	71.3 (5.88)*	68.7 (2.61)	71.2 (5.44)*	70.2 (3.50)*
ADAS cog	8.33 (4.04)	9.48 (3.60)	8.08 (3.99)*$	9.68 (3.79)
CAPS current	1.86 (4.18)*$#	56.4 (13.8)	7.96 (8.59)*$	54.1 (13.1)
CAPS lifetime	5.47 (8.86)*$#	78.4 (20.3)	21.3 (21.8)*$	67.0 (20.1)
TBI severity	NA	NA	1.92 (1.40)	1.55 (1.64)
Amyloid-β SUVR	1.06 (0.12)	1.00 (0.05)	1.07 (0.16)	1.06 (0.10)
Braak 1 tau SUVRa	1.84 (0.36)	1.68 (0.26)	1.85 (0.44)	1.88 (0.40)
Braak 2 tau SUVRa	1.40 (0.18)	1.32 (0.19)	1.37 (0.16)	1.36 (0.16)
Braak 3/4 tau SUVRa	1.74 (0.16)	1.65 (0.13)	1.72 (0.14)	1.69 (0.17)
Braak 5/6 tau SUVRa	1.80 (0.14)	1.75 (0.17)	1.79 (0.16)	1.79 (0.16)
LN Dwell brain state 7b	1.74 (0.83)	1.88 (0.87)	2.02 (0.76)	1.58 (0.92)

* p < 0.05 compared to PTSD.

# p < 0.05 compared to TBI.

a, 98 participants with tau PET in discovery, 72 with tau PET in validation cohort.

b, 86 participants with brain state 7 activity.

$ p < 0.05 compared to TBI&PTSD.

3.2. Dynamic resting state

The dynamic analysis identified 8 non-motion clusters or brain states (Fig. 2). Brain state 7 had the highest positive average module network z-score and the lowest Q indicating poor separation of its subnetworks, i.e., its features were consistent with a hyperconnectivity cluster (Table 2 and Fig. 2). 86 or 72 % of the TP1 DoD-ADNI participants with resting state fMRI and 85 or 70 % of the TP2 DoD-ADNI participants with resting state fMRI showed brain state 7 activity. The hyperconnectivity was not diffuse though but showed a pattern. Table 2 shows that hyperconnected regions were predominantly found within the sensorimotor/salience-like and visual-like networks. Regions with relative hypoconnectivity during brain state 7 were concentrated within the default mode-plus-like and executive-like networks. Brain state 7 dwell time was positively correlated with amyloid-β SUVR (Kendall tau B = 0.125, p = 0.047) and tau Braak stage 5 & 6 SUVR (Kendall tau = 0.200p = 0.035) but not with Braak stage 3 & 4 SUVR (Kendall tau = 0.156, p = 0.079), Braak stage 2 (Kendall tau = 0.093, p = 0.201) or Braak stage 1 SUVR (Kendall tau = 0.012, p = 0.458). Brain state 7 dwell time was not correlated with ADAScog (Kendall tau = 0.038, p = 0.62) but participants displaying brain state 7 activity had higher ADAScog scores than those who did not (10.79 (4.53) vs. 9.45 (4.33), p = 0.028).

Fig. 2.

fMRI brain states identified by cluster analysis. Each brain state is characterized by its connectivity pattern (whole brain positive strength maps in outer rows) and its community or module map (inner row). Brain state 7 had the highest overall whole brain connectivity strength, highest number of modules (Mod No) and highest Q value indicating low separation between modules which is consistent with a hyperconnectivity state.

Table 2.

Cluster/Brain State Characteristics.

Cluster/Brain State No	Counts	mean Z score spos Mod 1	mean Z score spos Mod 2	mean Z score spos Mod 3	mean Z score spos Mod 4	sum Z score spos Sum	Meanfwd	%afterScrub
1	6399	−1.20	0.17	−0.61	−0.55	−2.18	0.17	93.66
2	2016	−0.92	0.02	0.32	−1.97	−2.54	0.18	92.33
3	10,423	−1.77	−1.04	−0.94	−1.58	−5.33	0.16	98.86
4	2809	0.69	−1.21	−0.38	−0.64	−1.55	0.17	98.80
5	2054	1.64	−0.19	1.22	−0.23	2.44	0.19	90.56
6	2928	−0.63	−0.25	−0.20	0.72	−0.36	0.17	93.63
7	2520	1.93	−0.17	−0.25	2.09	3.60	0.17	89.99
8	1642	1.27	0.42	−0.07	−0.63	0.99	0.18	86.03

Counts, total no of windows in population after removing motion windows (DoD-ADNI,).

The community structure/module is identical with the network modules used for SuStaIn

Mod 4, visual network-like module.

81 % of the participants with a history of TBI (with or without history of PTSD) showed brain state 7 activity compared to only 18 % without a TBI history (p < 0.001) indicating that a diagnosis of TBI increased the likelihood to observe brain state 7 activity. The same was also true for a diagnosis of PTSD (with or without TBI) though (67.4 % of PTSD positive participants had brain state 7 activity vs. 32.6 % of PTSD negative participants, p = 0.003). This indicates that TBI and PTSD both increase the likelihood to develop brain state 7 activity. Indeed, 48 (expected: 33.2) or 57 % of the participants who had a history of TBI and PTSD but only 7 (expected: 17.1) or 8 % of the participants who never had experienced TBI nor PTSD showed brain state 7 activity (p < 0.001) while the frequency of participants with only a history of TBI or only a history of PTSD with brain state 7 activity was within the expected frequency. Brain state 7 dwell time was not correlated with TBI severity (Kendall tau B = −0.075, p = 0.813) or with CAPS current scores (Kendall tau B = −0.070, p = 0.824). Brain state 7 dwell time was also not different between the diagnostic groups (cf Table 1) nor was it influenced by ApoE4 carrier status (ApoE4 neg: 1.58 (0.91) vs. ApoE4 pos: 1.76 (1.08), p = 0.144).

3.3. SuStaIn

The model with two progression subtypes with 15 stages each and a non-progression subtype had with 4638.7 the lowest CVIC (one progression subtype CVIC = 4789.9, three progression subtype CVIC = 4668.0) in the discovery data set and in the validation data set (two progression subtypes = 2986.0, one progression subtype = 3011.1, three progression subtypes = 3000.2). Fig. 3a shows the original and cross-validated positional variance diagrams of the two progression subtypes for each cohort. 125 or 84.5 % of the 148 participants with two TP were assigned to the same subtype. Seventeen or 11.5 % of the participants assigned to the non-progression subtype in the discovery cohort were re-assigned to a progression subtype in the validation cohort. Six or 4.1 % of the participants assigned to a progression subtype in the discovery cohort were re-assigned to the non-progression subtype in the validation cohort.

Fig. 3.

Results of SuStaIn analysis. Upper panel (A) shows positional variance diagrams of the two progression subtypes for the discovery cohort (timepoint 1 images of 234 DoD-ADNI and 30 ADNI3 participants) and the validation cohort (timepoint 2 images of 148 DoD-ADNI and 30 ADNI3) on the left side and the positional variance diagrams from the cross-validation on the right side. The insert in the upper left corner shows the functional rois used to extract age and ICV corrected gray matter volumes. The lower panel (B) shows the progression of the amyloid-β SUVR increase (upper row, SUVR gray < 1.11, red > 1.2, plum > 1.3, blue > 1.4), gray matter atrophy (middle row, same color scheme as upper panel) and brain state 7 dwell time (stages for which data is available, intensity indicates mean dwell time for stage ranging from 3.14 windows to 15.9 windows) over the 15 stages for each subtype as well as the average amyloid (all DoD participants) and tau maps (98 DoD-ADNI participants who have undergone tau PET) for each of the subtypes. White squares indicate stages that were inferred by the SuStaIn algorithm, all other stages are supported by imaging data.

In the discovery cohort 83 participants were assigned to the non-progression subtype, 122 to progression subtype 1 and 28 to progression subtype 2. (Fig. 3b). The most prominent difference between the two progression subtypes was the stage when the amyloid-β SUVRs started to raise above that of the reference group. Whereas this was a late event (stage 13–15) in progression subtype 1 that only occurred after widespread severe atrophy had developed, it was the earliest event in progression subtype 2 (stage 1–3) and preceded the development of atrophy. In the case of the progression subtype 1 atrophy started in module 1 (combined sensorimotor/salience network-like module) at stage 1, followed by module 3 (executive network-like module) at stage 3 and then expanded into module 2 (default mode network-like module) at stage 4 and finally into module 4 (visual network-like module) at stage 6. Modules 1 and 3 were the first modules to develop atrophy in progression subtype 2 at stages 4 and 5 followed by the development of atrophy in module 2 at stage 7–8 and finally in module 4 at stage 11.

85 participants or approx. 30 % of the participants assigned to each of the three subtypes (non-progression (n = 34), progression type 1 (n = 45) and type 2 (n = 6)) had brain state 7 activity at TP 1. The mean dwell time in the non-progression subtype was 1.70 (1.06). In progression subtype 1, brain state 6 activity was observed in stages 1–7, 9 and 11. Its dwell time (mean no of windows) ranged from 1.03 (0.53) at stage 4 to 1.93 (0.94) at stage 1. In progression subtype 2, brain state 6 activity was observed in stages 1–3 and 5. Its mean dwell time/stage was highest in stage 1 with 2.76 (0.19) followed by 2.16 in stage 2, 1.93 (0.91) in stage 3 and 2.08 in stage 5.

Table 3 summarizes the clinical characteristics of each subtype. Participants assigned to progression subtype 1 had higher CAPS current and CAPS lifetime scores and accordingly also a higher percentage of participants with a diagnosis of PTSD 65.6 % (PTSD&TBI: 46.7 %, PTSD: 18.9 %) vs 34.5 % (TBI: 18.9 %, noTBI/noPTSD: 15.6 %), (Fisher’s exact test p = 0.01) than those assigned to the other two subtypes (subtype 0: 44.6 % (PTSD&TBI: 28.9 %, PTSD: 15.7 %) vs. 55.4 % (TBI: 31.3 %, noTBI/noPTSD: 24.1 %); subtype 2: 35.7 % (PTSD&TBI: 37.7 %, PTSD: 0 %) vs. 64.3 % (TBI: 35.7 %,noTBI/noPTSD: 28.6 %). Participants assigned to progression subtype 2 had higher ADAScog scores, higher tau and amyloid-β SUVRs and longer brain state 7 dwell times than those assigned to the other subtypes. Subtype 2 was observed in all four diagnostic groups.

Table 3.

Clinical Characteristics of Subtypes.

	subtype 0	subtype 1	subtype 2
No of Subjects	83	122	28
ApoE4 pos	22	28	12
Age	69.7 (5.28)	69.1 (3.88)	72.5 (5.60)
ADAS cog	9.11 (4.48)*	10.1 (4.18)*	12.4 (5.69)
CAPS current	25.5 (28.6)#	33.0 (26.1)	20.2 (25.0)#
CAPS lifetime	36.8 (37.0)#	45.6 (30.9)	29.2 (32.2)#
TBI severity	0.952 (1.21)	0.893 (1.18)	1.26 (1.79)
Amyloid-β SUVR	1.01 (0.06)*	1.02 (0.07)*	1.4 (0.14)
Braak 1 tau SUVRa	1.77 (0.36)*	1.75 (0.25)*	2.31 (0.58)
Braak 2 tau SUVRa	1.37 (0.22)*	1.34 (0.17)*	1.51 (0.13)
Braak 3/4 tau SUVRa	1.68 (0.12)*	1.69 (0.14)*	1.90 (0.30)
Braak 5/6 tau SUVRa	1.78 (0.12)	1.79 (0.16)	1.94 (0.33)
LN Dwell brain state 7b	1.70 (1.06)	1.54 (0.87)*	2.27 (0.57)

* p < 0.05 compared to subtype 2.

# p < 0.05 compared to subtype 1.

a, 98 participants with tau PET.

b, 86 participants with brain state 7 activity at TP1.

4. Discussion

The study had two major findings. 1. Dynamic resting state fMRI analysis identified a hyperconnectivity state (brain state 7) in a population of veterans with an increased risk for AD due to of a history of non-penetrating TBI and/or PTSD. Its features showed a good correspondence with the core characteristics of a hyperconnectivity state described in a population of cognitively intact older participants with and without increased amyloid-β load in a previous study (Mueller, 2019). As hypothesized, a history of past TBI facilitated the occurrence of this hyperconnectivity state but contrary to our expectations TBI severity did not influence its dwell time or intensity. An unexpected finding was that PTSD also facilitated the occurrence of this state. 2. SuStaIn identified two progression subtypes in the TP1 data or discovery data set that were confirmed in the TP2 or validation data set. The hyperconnectivity state could be observed in both subtypes. The main difference between the two progression subtypes was the timing of the amyloid pathology in relation to the development of gray matter atrophy. In subtype 1 the amyloid-β build up only occurred after the development of widespread gray matter loss. This and the higher CAPS scores and higher percentage of participants with PTSD suggest that PTSD and not AD pathology is the driving factor for the development of atrophy in subtype 1. In subtype 2 the amyloid-β build up preceded the development of gray matter atrophy. The higher amyloid-β and tau SUVRs as well as the longer hyperconnectivity dwell times in progression subtype 2 compared to subtype 1 indicate that AD pathology is likely to cause the gray matter loss in this subtype.

In conclusion, the findings suggests that past TBI and PTSD both increase the likelihood to observe the hyperconnectivity state but that the severity of the accompanying AD pathology determines its intensity. The development of gray matter atrophy influences neither its occurrence nor its intensity. The following paragraphs will discuss these findings in more detail.

Brain state 7 had the highest overall positive strength and the lowest Q value consistent with a poorly segregated hyperconnectivity state. The hyperconnectivity was most pronounced in regions belonging to the sensorimotor, salience and visual networks, i.e. regions free of AD pathology (Fig. 3b) whereas regions with beginning AD pathology, e.g., the default mode network, were characterized by a relative hypoconnectivity. This observation and the fact that its dwell time became longer with increasing amyloid load and increasing tau load in Braak 5 and 6 regions suggest that AD pathology plays a role shaping the expression of the hyperconnectivity state. Salience/sensorimotor network hyperconnectivity and default mode network hypoconnectivity in amyloid positive participants have been described in other studies that used a stationary resting state fMRI analysis approach (Chow et al., 2023, Schultz et al., 2017, Sepulcre et al., 2017). If it is assumed that brain state 7 represents the dynamic equivalent of the phenomenon observed in these previous studies, the new insight gained from the dynamic analysis is that the hyperconnectivity in brain regions spared from AD pathology is not always present but restricted to a single state, i.e., it has a paroxysmal character.

A history of TBI, particularly when combined with a history of PTSD, increased the likelihood to observe the hyperconnectivity state. As mentioned in the introduction, this association between TBI and hyperconnectivity state occurrence was expected. Finding the same association between hyperconnectivity state occurrence and PTSD was not expected though. However, PTSD associated hypersynchrony/hyperconnectivity has been described previously in MEG studies. Mišić et al (Mišić et al., 2016) found an interregional hypersynchrony at high frequencies in veterans with PTSD at rest and Dunkley et al (Dunkley et al., 2018) a maladaptive subcortical-cortical hyperconnectivity when veterans with PTSD were presented with an angry face condition during a face matching task. However, the participants in these two studies were in their twenties – forties, i.e., considerably younger than the DoD-ADNI participants. More research will be needed to understand if or how these findings of these two studies are related to the hyperconnectivity state described in this study. In contrast to our expectations though, TBI severity and also PTSD severity had no effect on the hyperconnectivity state’s dwell time. There are two possible explanations for this observation:1. TBI and/or PTSD have no direct influence on the occurrence of the hyperconnectivity state by themselves but instead are able to promote a condition or environment that facilitates its occurrence, e.g., amyloid-β build-up. 2. TBI and PTSD have a direct influence on hyperconnectivity state occurrence but the specific mechanisms that facilitate its appearance are not sufficiently captured by the CAPS and TBI severity scores used here or influenced by additional, not accounted for parameters, e.g., genetic disposition or TBI mechanism, for this relationship to become clear.

The second major finding was the identification of two progression subtypes each with 15 stages by SuStaIn. When interpreting the SuStaIn staging, it is important to keep in mind that it does not allow to infer the timing of the stage progression, i.e., neither is the time to progress from one stage to the next fixed within a subtype nor are the stages synchronized between subtypes, e.g., stage 1 might occur at an earlier age in one subtype than in the other. The clinical characteristics associated with the two progression subtypes suggest a link between subtype 1 with PTSD and a link between subtype 2 with AD pathology. Interestingly atrophy started in both subtypes in module 1, nearly simultaneously with (subtype 2) or followed by module 3 (subtype 1). This order is conceivable for a PTSD driven process that is known to be associated with parietal, frontal and insular gray matter loss (Wang et al., 2021, Siehl et al., 2023) but unexpected for atrophy associated AD pathology that typically starts in mesial-temporal structures. There are two possible explanations for this discrepant subtype 2 atrophy pattern. The first is that the four large rois used to extract the gray matter volumes were based on functional data that assigned the mesial temporal region to the default mode network-like module which could have obscured the atrophy in the mesial-temporal region. To investigate this possibility the SuStaIn analysis was repeated using large rois based on an anatomical parcellation (AAL3 (Rolls et al., 2020) that included a parietal, frontal, insular and mesial-temporal roi with the latter being modified to include an isolated hippocampus or an isolated parahippocampal roi. This had no effect on the order though, atrophy in the parietal roi always preceded that of the hippocampal or hippocampal-mesio-temporal roi (data not shown). The second explanation is that subtype 2 does not include the mesio-temporal stage of the AD pathology because it was already atrophied in the ADNI 3 reference population. This explanation is supported by two observations. The first is the stage 1 mean 1.27 (0.02) of amyloid-β SUVR in subtype 2 that was already above the threshold for amyloid-β positivity and rose further with each stage. The second were the amyloid-β SUVR and tau map of participants assigned to the non-progression subtype (Fig. 3b) that corresponds to the ADNI 3 reference population. Even though the mean amyloid-β SUVR in that group was with 1.01(0.06) well below the threshold for amyloid positivity, the tau map showed increased uptake of AV1451 in the mesio-temporal region. This indicates that tau related gray matter atrophy in the mesio-temporal region was already developing in participants assigned to the non-progression subtype and thus very likely also in the ADNI 3 reference population.

Hyperconnectivity episodes were not restricted to progression subtype 2 where their occurrence was expected considering the high amyloid-β loads characterizing that subtype but also occurred in progression subtype 1 and even the non-progression subtype although they were shorter than in subtype 2. There are two ways to interpret this observation. 1. The hyperconnectivity state represents a non-physiological state in elderly populations. It is not only an inherent manifestation of the AD pathology but also facilitates amyloid-β build up and distribution across the brain and by extension the development of AD pathology and ultimately gray matter loss and cognitive impairment (Bero et al., 2011, Cirrito et al., 2005). In this scenario, participants assigned to the non-progression subtype and progression subtype 1 in whom the hyperconnectivity state is observed would be likely to have a higher risk to develop amyloid-β positivity in the near future than those in whom this is not the case. In this case its early detection and suppression could have a major impact on AD development and progression. 2. Short hyperconnectivity state episodes are normal in elderly populations. However, increasing amyloid-β deposition due to incipient AD prolongs their dwell time which eventually could interfere with normal cognitive function, i.e., the prolongation of hyperconnectivity state dwell time represents an epiphenomenon of the AD disease process (Busche et al., 2012, Kellner et al., 2014). Investigating how the time course of amyloid-β build-up in an individual influences the individual’s hyperconnectivity state dwell time could help to decide which one of the two scenarios is more likely. Unfortunately, this is not possible in the DoD-ADNI population because amyloid-β PET scans were acquired only once. This question has therefore to be addressed in future, longitudinal studies that ideally focus on just one risk factor, e.g., ApoE4 or TBI, and also include middle-aged subjects.

In contrast to the a priori hypothesis outlined in the introduction, there was no evidence that the gray matter atrophy severity had a significant influence on hyperconnectivity state dwell time in the two progression subtypes. In subtype 1, hyperconnectivity state dwell time was similar to that in non-progression subtype 0 and persisted at that level during and after the development of widespread gray matter atrophy. In subtype 2, hyperconnectivity state dwell time was longer than in non-progression subtype 0 before the development of widespread atrophy (stages 1–3) and decreased once atrophy started to develop but this difference was not significant. It should be kept in mind though that the number of participants assigned to subtype 2 was small and therefore that the analysis in this subgroup was likely under-powered.

The study has several limitations. 1. The task-free fMRI protocol used by the DoD-ADNI project was optimized for the standard stationary fMRI analysis. fMRI protocols designed for dynamic analyses typically acquire the fMRI data over a longer time which increases the chances to observe the brain state of interest. It is therefore possible that longer acquisition times would have allowed to detect significant associations between hyperconnectivity state dwell time and clinical variables such as TBI severity etc.. 2. Although large for an imaging study, the number of participants with increased amyloid-β SUVRs who were assigned to subtype 2 by SuStaIn was relatively small in the DoD-ADNI population which limited the possibilities to investigate more complex associations, e.g., interactions between dwell time and clinical variables. 3. DoD-ADNI enrolled Vietnam veterans which means that the population in this study was male. This did not allow to investigate sex differences. 4. To the best of the authors knowledge, there exists no other publicly shared data repository that recruited the same or a similar population as DoD-ADNI. It was therefore not possible to verify the SuStaIn subtypes and staging in an independent data set.

CRediT authorship contribution statement

Susanne G. Mueller: Writing – review & editing, Writing – original draft, Visualization, Methodology, Funding acquisition, Formal analysis, Conceptualization.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

data came from DoD-ADNI data repository