Dynamic functional connectivity patterns associated with dementia risk

We included baseline data from 127 cognitively unimpaired older adults from the Age-Well randomized controlled trial (RCT) of the Medit-Ageing European project, detailed in a previous publication [12] (Fig. 1). Participants were recruited from the general population, aged over 65 years old, native French speakers, retired for at least 1 year, had completed at least 7 years of education, and performed within the normal range on standardized cognitive tests. The main exclusion criteria were (i) any contraindication to MRI or PET scanning; (ii) evidence of a major neurological or psychiatric disorder, including alcohol or drug abuse; (iii) history of cerebrovascular disease; (iv) presence of a chronic disease or acute unstable illness; and (v) current or recent medication that may interfere with cognitive functioning.

The Preclinical Alzheimer’s Cognitive Composite (PACC5) [13] was computed for each participant. The PACC5 is a global cognitive composite score sensitive to detecting preclinical Alzheimer’s disease-related cognitive decline, detailed in the Supplementary Methods.

Cardiovascular, psycho-affective, lifestyle, and genetic risk and protective factors for dementia were selected based on existing evidence from the literature [1, 2] and are described in Table 1. Variables were treated as risk factors if higher values equate to higher risk for dementia, and protective factors if higher values equate to lower risk for dementia. It should be noted that (i) hypertension has been consistently shown to be associated with dementia risk at mid-life but have an unclear association at late-life, as both hypertension and declining blood pressure have been associated with increased dementia risk [18, 19], and (ii) high BMI in late-life has been found to be associated with decreased dementia risk (as opposed to high BMI in mid-life) [20, 21]. Among lifestyle factors, three questionnaires were used (available in the Supplementary data):

All participants were scanned at Cyceron Center (Caen, France) on the same MRI (Philips Achieva 3.0T) and PET (Discovery RX VCT 64 PET-CT, General Electric Healthcare) scanners. Neuroimaging acquisition has previously been published [22] and is detailed in the Supplementary Methods.

Resting-state data of all participants were analyzed using fully automated spatially constrained independent component analysis (ICA) [24] as implemented in the GIFT software [25] (http://​trendscenter.​org/​software/​gift). Spatially constrained ICA was performed using an ICA template from an independent cohort [26] to decompose the data into 30 components, containing 14 intrinsic connectivity networks. These networks are available to download at https://​findlab.​stanford.​edu/​functional_​ROIs.​html (note that they are referred as “90 fMRI ROI” as the 14 networks have also been further parceled to generate 90 functional ROI). The advantage of using a spatially constrained ICA approach is to enhance robustness to artifacts and noise compared to single-subject ICA denoising and regression-based back-reconstruction [27] and also facilitated automated component labeling and sorting. Images were masked with a gray matter mask (described in the Supplementary Methods). Among the 30 components, we selected 7 ICA components recapitulating the three intrinsic functional networks most involved in Alzheimer’s disease-related cognitive decline: the DMN, SN, and ECN [11], which are represented in Fig. 2. Following ICA, each time course was normalized using z-scaling. Time courses were detrended, despiked using 3Ddespike, and filtered by a fifth-order Butterworth low-pass filter with a high-frequency cut-off of 0.15 Hz [28].

DFNC was estimated using the sliding window approach implemented in the GIFT toolbox [28, 29], as described in previous publications [4, 5]. Resting-state data were divided into 182 windows of 18 repetition times (43 s) size, in steps of one repetition time (2.4 s). These time windows were convolved with a Gaussian of 7.2 s (r = 3 repetition times), given that a window length between 30 and 60 s is suitable to estimate DFNC [30]. Within each of these windows, we imposed a L1 norm of the precision matrix to promote sparsity. As dynamic functional connectivity analyses are sensitive to movement artifacts, we regressed out the covariates mean framewise translation and rotation. Finally, functional connectivity matrices were transformed into z-scores using Fisher’s Z-transformation to stabilize variance prior to further analyses.

A k-means clustering was applied [31] on window functional connectivity matrices to compute reoccurring functional connectivity patterns across time and subject space [4, 5]. Briefly, the optimal number of clusters (referred to as “states”) was determined as equal to four (k=4), using the elbow criterion. Each window of each subject was then categorized to one of these connectivity states based on the similarity with the cluster centroid.

The following temporal DFNC parameters were then extracted for each subject:

Statistical analyses were performed with R studio (version 4.0.3). Our goal was to assess the association between temporal DFNC parameters and dementia risk and protective factors as well as Azheimer’s disease cognitive and neuroimaging markers. To this aim, forward stepwise regression models were performed using the temporal DFNC parameters as the dependent variables:

For each model, the α for entry was set at 0.05 and statistical significance was defined as P less than 0.05. All analyses were performed on the entire sample of participants (n=127), considering the time spent in the state as zero if the participant did not visit the state at all. We replicated all analyses only including the participants who visited the state to ensure that the results remained consistent (Supplementary Results).

To assess whether DFNC/dementia risk factor associations were stronger than with static functional connectivity, we replicated models 1 and 2 with the mean static functional connectivity within the 3 networks (DMN, SN, and ECN) as the dependent variable.

To check the reproducibility of the current findings, we replicated the analyses using another network template, the Neuromark atlas [32].

Participants had a mean age of 68.9 ± 3.82 years old and were composed of 63.8% of women. Participant characteristics are summarized in Table 2.

Four states were identified from the DFNC k-mean clustering analysis. The connectivity matrix of each state and its overall frequency (i.e., the total proportion of this state across subjects and acquisition time) are represented in Fig. 3. State 1 (“weakly connected state”), had a frequency of 51% and was characterized by low (positive or negative) or neutral (zero) connectivity between and within the three networks (DMN, SN, and ECN). This state was visited by 126/127 participants. State 2 (“SN-negatively connected” state) had a frequency of 18% and was characterized by high negative connectivity between the SN and the other networks (DMN and ECN) and a strong positive connectivity between the DMN and ECN and within each network. This state was visited by 105/127 participants. State 3 (“strongly connected state”) had a frequency of 16% and was characterized by high positive connectivity between and within the DMN, the SN, and the ECN. This state was visited by 94/127 participants. Finally, state 4 (“DMN-negatively connected” state) had a frequency of 15% and was characterized by high negative connectivity between the DMN and the other networks (SN and ECN) together with a strong positive connectivity between and within the SN and ECN. This state was visited by 78/127 participants.

For each state, we found a strong correlation between the mean dwell time and the total time (P < 10⁻¹⁶). Moreover, the mean and total times spent in states 3 and 4 were positively correlated (all P < 0.003). The mean and total times spent in state 1 were negatively correlated with the mean and total times spent in each other state (all P < 0.001). The average number of transitions between states across subjects was 9 ± 4.

Results from the stepwise regression showing the dementia risk and protective factors associated with the mean and total times spent in each state are provided in Table 3, model 1. A scatterplot of the main results of model 1 is represented in Fig. 4.

Longer times spent in both states 1 and 2 were associated with increased dementia risk (lower early-life and mid-life CAQ and a higher level of LDL cholesterol for state 1; higher systolic blood pressure, higher GDS score, and lower BMI for state 2). In contrast, longer times spent in states 3 and 4 were associated with reduced dementia risk (higher early-life CAQ and higher BMI for state 3; higher mid-life LEQ and lower LDL cholesterol for state 4). The number of transitions was not significantly associated with any dementia risk or protective factor.

Results of the stepwise regression showing the Alzheimer’s disease cognitive and neuroimaging markers associated with the mean and total times spent in each state are provided in Table 3, model 2. Longer time spent in state 3 was associated with higher Preclinical Alzheimer’s Cognitive Composite score-5. Times spent in states 1, 2, and 4 were not associated with any Alzheimer’s disease cognitive or neuroimaging markers. Finally, the lower number of transitions between states was associated with lower brain perfusion (β = 25.4 [CI 95% 12.7−38.1], P = 0.0001). A scatterplot of the main results of model 2 is represented in Fig. 5.

We found no significant association between the mean functional connectivity within the three networks and dementia risk factors or AD cognitive and neuroimaging markers.

The results of the replication analyses with the Neuromark atlas are in line with these main results, showing links with risk factors for states similar to states 1 and 2 and with protective factors for states 3 and 4, with yet subtle differences (see Supplementary data).

In line with previous studies [4], our analyses demonstrated a dominant state (state 1), characterized by weak intra- and inter-network connections. Similar weakly connected states have been found to be increased in Alzheimer’s disease [32] and in other neurodegenerative diseases [33]. The time spent in this weakly connected state was associated with a lower score in mid- and early-life cognitive activities, which is consistent with studies reporting an association between cognitive reserve and global network efficiency [34, 35]. This suggests that longer time spent in a weakly connected state is associated with poor cognitive reserve, possibly reflecting a lack of network efficiency. Alternatively, subjects spending more time in state 1 could have a lower amplitude global signal resulting in weaker global functional connectivity. The lower amplitude of the global signal could be related to degeneration of the basal forebrain which has been shown to be involved in global resting-state fMRI fluctuations [36] and is altered in Alzheimer’s disease [37].

Longer time spent in the SN-negatively connected state (state 2) was related to a high subclinical depression score, itself known to be associated with increased dementia risk [2]. This link is in line with previous studies showing the key role of SN regions in depressive disorders, including subclinical depressive symptoms [38, 39]. The time spent in state 2 was also associated with lower late-life BMI, which has also been associated with increased dementia risk, although it is still unclear whether it is a prodromal symptom for dementia or a risk factor per se [20, 21].

In addition, states 1 and 2 were both associated with higher cardiovascular risk factors (either higher levels of LDL cholesterol or systolic blood pressure). Those factors increase dementia risk [2] and they have both been shown to disrupt cognitive brain networks in healthy elderly subjects [40‐42]. Our results highlight for the first time that cardiovascular risk factors are associated with the modification of dynamic connectivity states. To ensure that the BOLD signal amplitude was not related to the blood pressure, we performed a complementary analysis which showed no association between these two measures (Supplementary Fig. 1).

Longer time spent in the strongly connected state (state 3) was associated with a higher PACC5 score, a cognitive composite score sensitive to cognitive decline, and particularly relevant in preclinical Alzheimer’s disease [13]. More particularly, the strongly connected state was characterized by a positive connectivity between the DMN, SN, and ECN, which all support cognitive functions [10]. Alteration of the static functional connectivity of those three networks has been linked with longitudinal PACC5 decline in cognitively unimpaired older adults [11]. Here, we expend this association to dynamic connectivity of the DMN, SN, and ECN. The time spent in the strongly connected state was also associated with early-life cognitive activities. This is in line with the previous findings and with studies showing that greater engagement in cognitive activities in early life is associated with better cognitive functioning in late-life and reduced the risk of dementia [43‐45]. Additionally, a higher prevalence of the strongly connected state was associated with higher BMI. Contrary to mid-life BMI, higher late-life BMI has been found to be associated with a lower risk of dementia, described as the “obesity paradox” [20, 21]. Thus, this result does not appear to be contradictory with the previous findings. However, it is important to keep in mind that, due to the possible reverse causation effect [46, 47], this association does not necessarily mean that a high BMI has a protective effect and remains thus more difficult to interpret in regard to dynamic functional connectivity change.

Longer time spent in the DMN-negatively connected state (state 4) was associated with higher early- and mid-life LEQ, a proxy of cognitive reserve including educational, occupational, leisure, social, and physical activities [15]. Similar DMN-negatively connected patterns have been linked with sustained attention tasks in both static and dynamic functional connectivity studies [48, 49]. This suggests that increased time in DMN-negatively connected is one of the mechanisms underlying attention-related cognitive reserve processes [50, 51]. Longer time spent in the DMN-negatively connected state was also associated with lower late-life cardiovascular risk factors including lower LDL cholesterol and systolic blood pressure. This last result is more difficult to interpret as it is unclear whether lower blood pressure is protective against dementia when assessed at late-life, since opposite results have also been reported [18, 19].

Finally, the times spent in states 3 and 4 were both negatively correlated with the times spent in states 1 and 2, strengthening the hypothesis that they have opposite relationships with dementia risk factors.

Interestingly, the times spent in the different states were not associated with Alzheimer’s disease neuroimaging biomarkers. This might indicate that, while being associated with dementia risk/protective factors, they are not directly associated with Alzheimer’s disease-specific pathological processes per se. More precisely, modification of DFNC could be a mechanism underlying cognitive reserve (i.e., modulating the effect of Alzheimer’s disease pathology on cognitive functions), but not a mechanism underlying brain reserve (i.e., preventing Alzheimer’s disease lesions to occur) [3, 52, 53].

The lower number of transitions between states (i.e., less state changes) was strongly associated with lower brain perfusion in Alzheimer’s disease-sensitive regions. This suggests that our ability to transition from one stage to another is reduced as a consequence of, or resulting in, decreased perfusion in the posterior cingulate and temporo-parietal regions known to characterize early Alzheimer’s disease changes. Interestingly, the lower number of transitions was associated with poorer cognitive performance in cognitively unimpaired older adults in a previous study [54]. Thus, the lower number of transitions could be a preclinical biomarker of hypoperfusion in brain regions typically impaired in Alzheimer’s disease.

There are limitations to this study. First, its cross-sectional and observational design does not allow for the assessment of the causal associations between risk factors and DFNC changes. Second, although we found some consistency across our results (i.e., risk and protective factors were associated with different states), caution in interpreting the results is warranted due to the number of models conducted, which could lead to spurious findings. Third, our analyses focused on cognitive networks, while other networks could also be altered in the dementia preclinical stage, which could be investigated in future studies. Fourth, the findings of this study would need to be replicated in an external fMRI dataset to test for its reproducibility. Finally, other dementia risk factors, such as air pollution and hearing impairment, were not available in this study but could also be of interest for further studies.