Reduced Network Dynamics on Functional MRI Signals Cognitive Impairment in Multiple Sclerosis

Introduction

Multiple sclerosis (MS) is an inflammatory and neurodegenerative disease of the central nervous system, characterized by physical disability and the frequent presence of cognitive deficits (1), which is often not in line with the amount of structural damage on MRI scans. Functional MRI could help to further unravel the underlying mechanisms that lead to these cognitive deficits (2). Blood oxygen level–dependent signal coherence between brain regions (ie, functional connectivity) led to the discovery of spatially distributed but functionally connected brain regions or “resting-state networks” (3). These networks include the cognitively relevant frontoparietal, attention, and default-mode networks (3). The default-mode network is characterized by high activity at rest, while its activity is normally suppressed during externally oriented cognitive processing (4). Disturbed default-mode function has been observed in cognitively impaired patients with MS, including impaired activity suppression during cognitive task performance (5) and disturbed functional connectivity between and within default-mode network regions during rest (6,7). These findings are difficult to interpret, however, as both stronger and reduced connectivity relate to cognitive impairment (6,7).

To better understand the functioning of complex systems (eg, cognitive functions of the human brain), it is essential to not only study individual regions or connections in isolation but to consider how they work together as a network (8). This realization has led to the ascendance of network neuroscience (8). The few functional network studies that have been performed to investigate cognitive impairment in MS have shown a reduction in global efficiency (9,10), as well as a more modular architecture (11), both related to worse cognitive function (10,11). In a recent study (12), we investigated regional changes in functional network organization and observed a shift toward a more important position (ie, higher centrality) of regions that are part of the default-mode network in cognitively impaired patients with MS (12).

Such analyses on average network organization rely on the computation of average coupling strengths between brain regions during an entire functional scan. By only incorporating average coupling between brain regions, however, temporal fluctuations (ie, dynamics) in coupling strength are not taken into account. Emerging research indicates that these fluctuations could potentially be highly relevant for cognitive functioning (13,14). As such, we hypothesized that disturbances in functional network dynamics could be especially relevant for impaired cognitive functioning. In the current study, we therefore aimed to investigate potential disturbances in functional network dynamics in cognitively impaired participants with MS.

Materials and Methods

Study Participants

For secondary analysis of prospectively acquired data, approval was obtained from the institutional ethics review board of the Amsterdam UMC, Locatie VUmc, and study participants provided written informed consent. The MRI data from the consecutively recruited Amsterdam MS cohort (12,15) included in this study were acquired between 2008 and 2012 in the outpatient clinic of the Amsterdam UMC, Locatie VUmc. The data from this cohort were included in other recent studies (12,15), but network dynamics had not been studied before (see Appendix E1 [online] for overlap of participants with other studies). A total of 332 participants with MS and 96 healthy control subjects were included. Eligibility criteria for undergoing an MRI examination in the original cohort were a diagnosis of clinically definite MS according to the 2010 revised McDonald criteria (16) and a relapse-free period with no steroid treatment for at least 2 months prior to MRI. For this study, all participants and healthy control subjects from the original cohort were included with MRI and neuropsychological evaluation. Exclusion criteria for the cohort were the presence or history of psychiatric or neurologic disease (other than MS) and, for this study, any study participants with a framewise displacement of 0.5 mm or higher in more than 20% of functional MRI time points. Study participants who were classified as mildly cognitively impaired (see below) were also excluded from the analysis (17). The Expanded Disability Status Scale (EDSS) was used to assess overall disability.

Cognitive Assessment

Participants underwent extensive neuropsychological evaluation using an expanded Brief Repeatable Battery of Neuropsychological tests (18), with a total of seven cognitive domains assessed (see Appendix E1 [online]). Neuropsychological data were partially obtained by L.D. (a neuropsychologist with > 10 years of experience). We adjusted cognitive scores for normal effects of sex, age, and education based on healthy control subjects (19) and transformed them into Z-scores based on the healthy control scores. Next, we subdivided participants with MS into two cognitive groups (6): (a) participants who scored 2 standard deviations below healthy control subjects on at least two cognitive domains were categorized as cognitively impaired and (b) participants who scored 1.5 standard deviations below healthy control subjects on at least two domains but did not fulfill the cognitively impaired criterion as mildly cognitively impaired were excluded. The remaining participants were classified as cognitively preserved.

MRI Examination

Participants were imaged with a 3-T General Electric MRI system (Signa-HDxt; Milwaukee, Wis) using an eight-channel phased-array head coil (partially collected by M.M.S., neuroscientist with > 10 years of experience). The protocol included a resting-state functional MRI sequence covering the entire brain (echo planar imaging, 202 volumes; repetition time msec/echo time msec, 2200/35; 80° flip angle; 3-mm contiguous axial sections; in-plane resolution, 3.3 × 3.3 mm), a three-dimensional (3D) T1-weighted fast spoiled gradient-echo sequence (repetition time msec/echo time msec/inversion time msec, 7.8/3 /450; 12° flip angle; 1.0-mm sagittal sections; in-plane resolution, 0.9 × 0.9 mm²) for volumetric measurements, and a 3D fluid-attenuated inversion-recovery sequence (8000/125/2350; 1.2-mm sagittal sections; in-plane resolution, 0.98 × 0.98 mm²) for lesion detection.

Image Processing

Images from resting-state functional MRI were preprocessed by A.J.C.E. and K.A.M. (medical doctor and neuroscientist, respectively, both with 5 years of experience) and M.M.S., using the automatic MELODIC pipeline (part of FSL 5, FMRIB 2012; Oxford, United Kingdom) (see Appendix E1 [online]). We used an independent component analysis–based strategy for Automatic Removal of Motion Artifacts (ICA-AROMA, v0.4-beta 2017, Nijmegen, the Netherlands) (20) to further reduce effects of motion. After preprocessing, we automatically registered functional MR images to lesion-filled (21), 3D T1-weighted images using boundary-based registration and subsequently nonlinearly registered them to Montreal Neurologic Institute standard space.

Centrality Dynamics

We (A.J.C.E., A.M.W. [computational scientist with > 15 years of experience], L.D., M.M.S.) assessed functional network dynamics by using eigenvector centrality, a measure of regional functional “importance” based on the strength of a region’s connections with other regions and importance of the connected brain regions themselves (22). We computed centrality for each gray matter functional MRI voxel as described previously (12) by using fastECM (created by AMW, version 3 2007, Amsterdam, the Netherlands) (23) (see Appendix E1 [online]). To assess centrality dynamics, individual functional MR scans were split into sliding windows of 20 time points (44 seconds) (14,24) with a shift length of one volume, resulting in 181 time windows per functional MRI acquisition. Next, eigenvector centrality maps were computed for each window and concatenated to obtain a four-dimensional “dynamic centrality map.” To quantify changes in a region’s functional importance over time, the variability (ie, standard deviation) of a voxel’s centrality across all time windows was computed, based on dynamic functional connectivity literature (14).

Hierarchical Organization of Centrality Dynamics

To assess whether fluctuations in regional importance occur in spatial patterns, we performed a standard group probabilistic independent component analysis (25) on the concatenated centrality maps (see Appendix E1 [online]). To further assess how these spatial patterns are temporally and hierarchically related, the Pearson correlation coefficient was computed between component centrality time series followed by a hierarchical cluster analysis.

Statistical Analysis

We performed statistical analyses of the demographic, clinical, and global MRI variables by using SPSS (version 22, 2013; IBM, Chicago, Ill). We checked all demographic, clinical, and global MRI variables for normality by using the Kolmogorov-Smirnov test and histogram inspection. We used nonparametric testing to assess group differences in demographic variables and EDSS, and we performed multivariable general linear model analyses to assess group differences in MRI and cognitive variables, with sex, age, and education entered as covariates. To test for regional differences in centrality dynamics, we compared the voxel-wise centrality dynamics (ie, variability) maps between cognitively impaired and cognitively preserved participant groups by using a clusterwise permutation-based method (26) at P less than .05, with sex, age, and education as covariates. We assessed differences in the coherence of centrality fluctuations between spatial components among cognitive groups by using general linear model analysis, with sex, age, and education entered as covariates. We applied the Bonferroni correction according to the number of tests performed. Corrected P values less than .05 were considered indicative of statistical significance.

Results

Demographics

Among all 332 participants, 87 were classified as cognitively impaired, 65 as mildly cognitively impaired, and 180 as cognitively preserved. After excluding all patients with mild cognitive impairments, a total of 267 participants with MS (83 men, 184 women; mean age, 47.9 years ± 10.8; mean symptom duration, 14.7 years ± 8.5) and 96 healthy control subjects (40 men, 56 women; mean age, 45.9 years ± 10.4) were included, all of whom underwent complete functional MRI examinations and complete cognitive evaluations; none were excluded based on excessive motion (see Table). Compared with the cognitively preserved participants, the cognitively impaired participants included a higher proportion of men (48 of 180 [27%] vs 35 of 87 [40%], respectively; P = .02), and had an older mean age (46.3 years vs 51.1 years, respectively; P < .01) and a lower median level of education (level 6 vs level 4, respectively; P < .01).

Clinical Function

Among the 267 participants with MS, 197 were diagnosed with relapsing-remitting MS, 47 with secondary progressive MS, and 23 with primary progressive MS. Disease-modifying treatments included β-interferons (n = 56), glatiramer acetate (n = 10), natalizumab (n = 22), or other immunosuppressive therapy (n = 5). Participants with MS were mildly disabled with a median EDSS score of 3.0 and an average cognitive function Z-score of −0.80.

Reduced Network Dynamics in Cognitively Impaired Participants with MS

As this analysis included motion regression with ICA-AROMA (20), we first investigated whether this affected a previous report on stationary centrality (12) and found highly comparable results (Fig 1, A and B). Next, we compared centrality dynamics in cognitively impaired versus cognitively preserved participants, and we observed reduced dynamics in cognitively impaired participants with MS in regions that are part of the default-mode network (ie, posterior cingulate cortex and right angular gyrus), frontoparietal network (ie, middle frontal gyrus and superior parietal gyrus), visual network, and thalamus (Fig 1,C and D). These regions partially overlapped with abnormal stationary centrality regions, with both the more central default-mode and frontoparietal networks as well as the less central visual network all showing reduced dynamics. These analyses were repeated with window lengths of 10, 30, and 40, showing similar results (see Fig E1 [online]).

Figure 1:

Download as PowerPoint

Centrality Dynamics Follow Resting-State Network Patterns

Above-mentioned bottom-up analyses indicated a loss of dynamics within regions belonging to specific resting-state networks. These analyses could not, however, indicate whether this loss of dynamics of specific networks also impacted their normally alternating behavior. To specifically assess this question, we first evaluated whether fluctuations in the functional importance of brain regions over time indeed occurred in such functional network patterns. The independent component analysis on the concatenated dynamic centrality maps resulted in the extraction of 13 spatial components closely resembling classic resting-state networks, including the default-mode, frontoparietal, sensorimotor, and visual networks (Fig E2 [online]), all of which overlapped with clusters found in Figure 1 (as shown in Table E1 [online]). The fluctuations in functional importance of brain regions have been visualized in a representative healthy control subject in Movie E1 (online) and for a single time window in Figure 2 with a strikingly low centrality of the default-mode network at this particular time window.

Figure 2:

Download as PowerPoint

Network Dynamics Are Hierarchically Organized into Opposing Functional Systems

As some networks, like the default-mode network, normally show strong negatively correlated relation with other networks over time, we next investigated whether the aforementioned reduced network dynamics could have affected this aspect of normal network dynamics. The normal coherence of centrality dynamics between individual resting-state networks was first investigated in healthy control subjects, indicating two main clusters as shown in the heat map in Figure 3, A. Networks within the same cluster mostly showed positively correlated fluctuations over time, whereas networks in opposing clusters mostly demonstrated negatively correlated behavior. The first cluster included the default-mode and frontoparietal networks, while the second included the sensorimotor and visual networks.

Figure 3:

Download as PowerPoint

Altered Network Coherence in Cognitively Impaired Participants with MS

The four networks that showed reduced dynamics, that is, the default-mode, frontoparietal, sensorimotor, and visual networks, were subsequently analyzed. The negative correlation between the default-mode and visual networks (r = −0.23 in healthy individuals) was reduced in cognitively impaired participants (r = −0.04) compared with cognitively preserved participants (r = −0.14; corrected P = .02 [Fig 3, B]). We further assessed the cognitive relevance of this negative correlation between default-mode and visual networks as a posthoc analysis by relating it to average cognitive function in the healthy control subjects. This analysis showed a negative correlation (Pearson r = −0.21; P = .04), indicating that a stronger negative correlation between these two networks is related to better cognitive performance in healthy individuals. In cognitively impaired participants, the loss of this negative correlation was related to lesion volumes (r = 0.23; P = .03), but not whole-brain volumes (r = −0.18; P = .10). The key differences between cognitively impaired and cognitively preserved participants are illustrated dynamically in Movies E2 and E3 (online), and for a single time window in Figures 4 and 5 using two representative participants with MS. Compared with the cognitively preserved participant (Fig 4), the cognitively impaired participant (Fig 5) demonstrated a higher average centrality of the default-mode and a lower average centrality of the visual network, lower centrality dynamics in both networks, and a reduced negative correlation between these networks.

Figure 4:

Download as PowerPoint

Figure 5:

Download as PowerPoint

Discussion

Previous studies in network neuroscience have shown that the brain is a complex functional system showing a dynamic interplay between functional brain networks (8). The brain network becomes less efficient in multiple sclerosis (MS) (9,10), with drastic implications for cognitive function. To further study how the altered behavior of such networks could underlie cognitive impairment in MS, we investigated disturbances in functional network dynamics by using dynamic eigenvector centrality, a measure of how the functional importance of specific brain regions fluctuates over time, in a large cohort of MS participants with and without cognitive impairment and in healthy control subjects. We showed that fluctuations in the functional importance of brain regions follow classic functional network patterns. We observed reduced dynamics in the default-mode, frontoparietal, and visual networks in cognitively impaired participants with MS, with an almost complete loss of the negative correlation between the dynamically opposed default-mode and visual systems.

Combined with the previously observed higher overall importance of the default-mode and frontoparietal networks and reduced importance of the sensory networks in cognitively impaired patients with MS (12), our finding of reduced dynamics in these same networks in cognitively impaired participants with MS may indicate that networks become less dynamic or “stuck” in their high- and low-centrality positions, respectively. Previous studies have indicated the relevance of high default-mode network dynamics during cognitive task performance (27) and impaired working memory function associated with reduced default-mode network suppression in cognitively impaired patients with MS (5). A more central and less dynamic (ie, stuck) default-mode network could therefore potentially prevent dynamic suppression of default-mode network activity during (externally oriented) cognitive processing (4). It has been hypothesized that attenuation of the default-mode network is necessary for adequate goal-directed action and impaired default-mode attenuation may result in lapses of attention through intrusion of introspective thought, reducing cognitive task performance (28).

The observation that centrality patterns fluctuate along classic resting-state network patterns was previously suggested in a smaller healthy control study that used k-means clustering on dynamic centrality maps to show “representative dominant patterns” resembling resting-state networks (29). The finding that resting-state networks are hierarchically organized into multinetwork clusters demonstrating positively and negatively correlated behavior agrees with findings of previous studies that showed negatively correlated resting-state network activity fluctuations (30). To our knowledge, however, this has not yet been shown for a higher-order network measure of regional functional importance. Although the exact functional underpinnings of these negative correlations over time remain unknown, a common view is that these negatively correlated networks process opposing functions and compete for processing resources (31,32).

A negative correlation between the default-mode and the (external visuospatial attention–related) dorsal attention network has been shown previously (33). Our finding of a reduced negative correlation between the default-mode and visual networks could represent a disruption in the intrinsic wiring between these internally and externally focused systems. It is currently unclear how accumulating pathology in MS could drive such disruption, but reduced cofluctuations have frequently been reported in stroke (34) and traumatic brain injury (35), commonly interpreted as resulting from damage to connecting tracts. Our finding that the loss of negative correlation between default-mode and visual networks in cognitively impaired patients with MS was related to lesion volume could suggest that damage to the underlying tracts due to MS pathology disrupts the negatively correlated dynamics between opposing functional systems, thereby possibly hampering normal cognitive functioning.

Our study has some limitations. First, the cognitive groups differed on demographic characteristics including sex, age, and level of education, which are known to relate to cognitive impairment in MS, and we therefore entered these as covariates in our analyses. Second, the experimental design was cross-sectional and retrospective, which hampers conclusions on longitudinal evolution. Finally, we used a sliding-window approach and the choice of window length is a matter of extensive debate (13). The use of time-frequency analyses allows the adaptation of the observation window to the frequency content of the original time courses, thereby circumventing the need to use fixed window lengths (14).

To conclude, by using a large cohort of participants with multiple sclerosis (MS) and healthy individuals, we have further elucidated cognitive disturbances in MS by studying abnormalities in functional network dynamics. Combined with the previously observed shift toward more central default-mode and frontoparietal and less central visual and sensorimotor networks, the reduced network dynamics and loss of normal interplay of these networks could reflect a “stuckness” in the high- and low-centrality positions of these networks, respectively. This defective network condition may represent a biologic substrate of impaired cognitive functioning, as stuck networks hamper adequate transitioning between internally and externally oriented states.

Disclosures of Conflicts of Interest: A.J.C.E. Activities related to the present article: disclosed receipt of institutional grant from Stichting MS Research (grant numbers 08-650, 13-820, and 14-358e). Activities not related to the present article: disclosed no relevant relationships. Other relationships: disclosed no relevant relationships. A.M.W. Activities related to the present article: disclosed institutional grants from EuroPOND (ERC Horizon grant 666992), EPAD (EU IMI grant 115736), and AMYPAD (EU IMI grant 115952). Activities not related to the present article: disclosed disclosed institutional grants from EuroPOND (ERC Horizon grant 666992), EPAD (EU IMI grant 115736), and AMYPAD (EU IMI grant 115952). Other relationships: disclosed no relevant relationships. K.A.M. Activities related to the present article: disclosed no relevant relationships. Activities not related to the present article: disclosed institutional grant from Biogen. Other relationships: disclosed no relevant relationships. L.D. disclosed no relevant relationships. J.J.G.G. Activities related to the present article: disclosed institutional grant from Dutch MS Research Foundation. Activities not related to the present article: disclosed employment with Amsterdam University Medical Center and Netherlands Organisation for Scientific Research; receipt of payment from Biogen Idec, Novartis Pharma, and Sonofi Genzyme to institution for investigator-initiated trial or sponsor-initiated trial. Other relationships: disclosed work as an editor for the Multiple Sclerosis Journal. M.M.S. Activities related to the present article: disclosed institutional grant from Dutch MS Research Foundation (grant numbers 08-650, 13-820, and 14-358e). Activities not related to the present article: disclosed payment received by institution from Biogen and Sanofi Genzyme for consultancy; institutional research grant from Biogen; and payment to institution from ExceMed for lectures, including service on speakers’ bureaus. Other relationships: disclosed no relevant relationships.

Author Contributions

Author contributions: Guarantors of integrity of entire study, A.J.C.E., M.M.S.; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; approval of final version of submitted manuscript, all authors; agrees to ensure any questions related to the work are appropriately resolved, all authors; literature research, A.J.C.E., K.A.M., L.D., M.M.S.; clinical studies, A.J.C.E., L.D., M.M.S.; experimental studies, J.J.G.G., M.M.S.; statistical analysis, A.J.C.E., A.M.W., K.A.M., J.J.G.G., M.M.S.; and manuscript editing, all authors