Structure and Dynamics of Large‐Scale Cognitive Networks in Huntington's Disease

Huntington's disease is a neurodegenerative disorder characterized by clinical alterations in the motor, behavioral, and cognitive domains. However, the structure and disruptions to large‐scale brain cognitive networks have not yet been established.

behavioral features can be detected up to 15 years before the emergence of the first motor symptoms. 2 Subtle cognitive changes are present from the early stages, with the development of dementia as a likely outcome in the majority of the cases. 3 The progression of cognitive decline has been attributed to a pattern of whole-brain atrophy, with the basal ganglia and its fronto-subcortical connections as the most affected by disease progression. 4 However, recent works have shown that progressive atrophy of a wide range of cortical regions is linked to the clinical profile of the disease from the early stages. This cortical atrophy involves several regions that participate in cognitive functions such as the precuneus, 5 the anterior cingulate cortex, 6 and the supramarginal, fusiform, and lateral occipital gyrii. 7 The majority of these regions are integral to the functioning of large-scale cognitive networks, 8 such the salience network (SN) or default-mode network (DMN). Thus, cognitive manifestations of HD might be related both to structural brain damage and the disruptions in the architecture of large-scale brain cognitive networks.
Structural damage in the basal ganglia and motor and visual regions has been shown to correlate with alterations in both motor and visual functional networks, 9,10 even from the premanifest stage. 11,12 Studies focusing on the DMN using resting-state functional magnetic resonance imaging (fMRI) showed decreases in functional connectivity (FC) within the network in manifest HD but no changes in premanifest patients. 13 In this latter population, task-based fMRI found an increased connectivity within the DMN. 14 In the executive networks, a dissociated pattern was found between posterior cortical decreases in activity 10 and an increase in frontoparietal connectivity, 15 but again, studies in the premanifest population have shown no differences. 16 Finally, to our knowledge, no study has tackled the study of dynamic FC of cognitive networks in HD. Although static FC has proven to be demonstrable across task and resting-state conditions, 17 and even across species, 18 there is growing evidence that brain networks are not immutable across time, but inherently "multistable." 19 This includes the notion that cognitive hubs such as the posterior cingulate cortex (PCC) can dynamically engage with different networks. 20,21 Previous studies have established the relevance of the frequency of connectivity states in the cognition of neurodegenerative diseases such as Parkinson's disease, 22 showing that patients with an associated dementia have a higher frequency of a segregated connectivity state than healthy controls. On the other hand, high variability of these connectivity circuits is considered to be a feature of normal organization in the brain 23 that decreases with normal aging 24 ; it has also been found to be decreased within default-mode and frontoparietal networks in patients with schizophrenia. 25 It is conceivable that the subtler aspects of network disruption could become more apparent when considering the dynamic nature of brain circuits.
The main aim of the present study was to conduct a functional analysis of the different large-scale cognitive networks in premanifest HD (pre-HD) and symptomatic HD (sHD). Although cognitive manifestations may represent the earliest indicators of the disease, there is a lack of studies exploring the integrity of cognitive networks and its clinical impact on HD. Specifically, we performed classical averaged FC and graph theory analyses, and we also conducted a novel dynamic FC analysis through dynamic independent component analysis (dyn-ICA). We also performed a structural analysis of brain atrophy using cortical thickness while adopting both a standard whole-brain approach and a cognitive network-focused parcellation, similar to our previous work. 26

Participants and Clinical Assessment
A total of 65 gene mutation carriers (CAG ≥ 39) who attended the outpatient clinic at the Movement Disorders Unit and 23 healthy controls who were willing to participate were prospectively recruited. Patients with HD were classified as premanifest HD (pre-HD) and sHD. Individuals with a Unified Huntington's Disease Rating Scale (UHDRS) total motor score below 5 and a diagnostic confidence level (DCL) <3 were classified as pre-HD, whereas those with a DCL = 4 were classified as sHD. Cognitive measures were obtained using the UHDRS cognitive score (Cogscore). The disease burden score (DBS)-a measure of livelong exposure to mutant huntingtin-was calculated using the following formula based on age and CAG repeat length: age Â (CAG À 35.5). 27 However, healthy controls only had sociodemographic data available. Exclusion criteria based on image quality were (1) pathological magnetic resonance imaging (MRI) findings beyond mild white matter hyperintensities and (2) the presence of head motion or other MRI artifacts. After reviewing images for quality control, one patient was excluded because of excessive head motion.

MRI Acquisition and Preprocessing
All participants had available structural and restingstate fMRI. Details regarding image acquisition and preprocessing are available in Appendix S1 in the supplementary material. Briefly, T1-weighted scans and resting-state (blood-oxygen-level-dependent) BOLD images (12 minutes) were acquired in a 3 T Philips Achieva station (Philips Medical System, Best, the Netherlands). To assess cortical atrophy differences we applied a surface-based cortical thickness pipeline using FreeSurfer 6.0. (free software: Fisch et al, 10.1016/j. neuroimage.2012.01.021) Functional imaging processing was performed using CONN v19b software and its standard processing pipeline, 28 which includes functional preprocessing and brain parcellation and first-level and second-level analyses steps. For preprocessing, scans were functionally realigned and unwarped, slice-timing corrected, realigned, and spatially normalized to the Montreal Neurological Institute (MNI) space using coregistration with the associated anatomical data. For brain parcellation, we adopted both the Harvard-Oxford Atlas 29 provided within the CONN toolbox and the set of functional brain networks described in the Shirer atlas 30 to provide both whole-brain macroanatomical coverage and functionally oriented parcellations 31 ; time courses for all of these regions of interest (ROIs) were derived from MNI-normalized images. Thereafter, ROI time courses were submitted to CONN's standard denoising pipeline, which combines linear regression of potential confounding effects and temporal band-pass filtering. For the first-level signal processing, we adopted both time-averaged static FC measures and a dyn-ICA. Dyn-ICA matrices represent a measure of different modulatory circuits expression and rate of connectivity change between each pair of ROIs. For the second-level analysis, averaged metrics of FC and graph theoretical parameters were computed in ROI-to-ROI analyses as well as dyn-ICA parameters: frequency was defined as the recurrence of the modulatory circuits in each participant and variability as the standard deviation in bivariate, multivariate, or semipartial correlation or regression measures between pairs of ROIs. A detailed description of the preprocessing, denoising, and first-level and second-level analyses as well as an in-depth discussion of dyn-ICA parameters is provided in Appendix S1 in the supplementary material.

Statistical Analyses
Clinical and sociodemographic data were compared across groups using two-sample t test analyses for continuous variables and χ 2 for categorical variables. Differences were considered significant using a probability (P) value < 0.05. On the structural analyses, only surface clusters surviving P < 0.05 family-wise error corrected by permutation were considered.
On the structural analysis, we compared vertex-wise cortical thickness data between HD (pre-HD/sHD) and healthy controls using age, sex, and education as nuisance covariates. For this purpose, vertex-wise cortical thickness maps were normalized to a standard fsaverage space and smoothed using a Gaussian kernel of 10 mm full width at half maximum. Only surface clusters surviving P < 0.05 and family-wise error correction by permutation testing (10,000 permutations) were considered significant.
Functional and graph theoretical metrics were introduced into a generalized linear model within CONN to compare controls, pre-HD groups, and sHD groups using age, sex, and education as covariates of no interest. Furthermore, total gray matter volume was explored as a covariate of no interest to analyze the interaction between functional and structural changes. Results were considered significant using a threshold of P value < 0.05 corrected for multiple comparisons at the cluster level using the false discovery rate for all the functional and graph theoretical analyses. Temporal data regarding dyn-ICA analysis of frequency and variability of dynamic components were also introduced into a generalized linear model using age, sex, and education as covariates of no interest and considered significant using a P value < 0.05. Finally, regression analyses were performed using DBS, UHDRS, and Cogscore data in the relevant connections (which had previously been found as statistically significant between groups in a P < 0.05 corrected analysis) and dynamic components and considered significant using a P value of <0.05. Statistical analyses concerning clinical and sociodemographic comparisons as well as clinical imaging associations were conducted using the SPSS version 15 (IBM Corp. Armonk, NY) software package. All patients provided written informed consent according to the Declaration of Helsinki. The study was approved by the Ethics Comitee for Clinical Research at the Hospital de la Santa Creu i Sant Pau.  For full demographic data, see Table 1.

Clinical and Sociodemographic Data
Patients with pre-HD were comparable in age and sex to controls but had a slightly higher education level (14 vs. 12 years for controls). Patients with sHD were older than controls (49 vs. 39 years in controls). All of these differences were accounted for in imaging analyses by using these variables as covariates of no interest.

Structural Analysis: Cortical Thickness
On the structural analysis, patients with HD showed reduced widespread cortical thinning in frontal, parietal, and temporo-occipital regions. These changes were more pronounced when comparing patients with sHD with healthy controls, whereas patients with pre-HD showed no differences in cortical thickness relative to controls. Figure 1 and Appendix S1 in the supplementary material display the cortical thickness differences between these groups.

Whole-Brain FC
On a whole-brain approach, patients with HD showed decreased FC on a wide range of cortical regions, chiefly the motor cortex (Fig. 2, left). Precentral, postcentral, and supplementary motor areas showed decreased FC with visual associative cortices (P = 0.002 corrected) and fronto-opercular regions, including the bilateral insular and anterior cingulate cortices (P = 0.01 corrected). Importantly, these decreases were clinically associated with both the UHDRS scores and with the DBS. The regression analysis showed that reduced FC between the postcentral gyrus and the occipital fusiform gyrus was associated with higher DBS (t = À2.57, P = 0.006) and UHDRS (t = À5.27, P < 0.001), and similarly, decreased FC between the right insular and supplementary motor cortex was correlated with higher DBS (t = À2.83, P = 0.003) and UHDRS (t = À4.73, P < 0.001). The basal ganglia also showed reduced connections with the superior and mid frontal gyrii (P = 0.01 corrected). Specifically, reduced caudate-mid frontal gyrus connections were associated with higher DBS (t = À2.98, P = 0.002) and UHDRS (t = À2.60, P = 0.004) for the caudate-mid frontal gyrus connections. The main increase in FC among patients with HD compared with controls was found between visual regions-fusiform cortex, occipital pole-and the PCC and precuneus, both cortical hubs of the DMN (P = 0.003 corrected).
The regression analysis showed that higher FC between the precuneus and the right fusiform cortex was associated with higher DBS (t = 1.99, P = 0.02).

FC in Cognitive Networks
Considering cognitive networks, patients with HD showed decreased connectivity within posterior hubs of the DMN such as the precuneus and PCC, and of these hubs with anterior regions of the same network such as the medial prefrontal cortex (P < 0.001 corrected; Fig. 2, right). A key aspect is that the posterior cingulate cortex (PCC)-Med frontal connection showed a positive correlation with cognitive scores as measured by the Cogscore (t = 2.25, P = 0.01) and an inverse correlation with DBS (t = À2.42, P = 0.009). The salience network of patients with HD showed decreased FC between insular, supramarginal, and anterior cingulate cortices (P = 0.009), and the connection between the right insular and supramarginal regions also showed a positive correlation with the Cogscore (t = 2.11, P = 0.02) and an inverse correlation to DBS (t = À2.35, P = 0.01).

FC in Pre-HD
Patients with pre-HD showed similar results both on a whole-brain and cognitive networks approach, albeit in a reduced fashion. Differences with healthy controls were centered on motor and visual associative regions on a whole-brain approach (P = 0.02 corrected), and connectivity increases between the PCC/precuneus and visual associative regions were also observed (P = 0.04 corrected). Meanwhile, in cognitive networks, changes were mainly observed as decreases of connectivity between anterior and posterior nodes of the DMN (P = 0.02). Importantly, after controlling for gray matter volume (GMV), all results remained significant with the following exceptions: in patients with pre-HD, increased precuneus/PCC connectivity with visual associative regions in a whole-brain approach (not present after correcting for GMV) and decreases in connectivity between anterior and posterior nodes of the DMN (P = 0.06 after correcting for GMV).

Whole-Brain Graph Metrics
On a whole-brain analysis, patients with HD showed decreased node degrees compared with controls in bilateral postcentral and precentral gyrii (P = 0.007 corrected) and the supplementary motor area (P = 0.04 corrected) and decreases in other nonmotor regions such as the bilateral supramarginal gyrus (P = 0.04 corrected) and the PCC (P = 0.006 corrected; Fig. 3, left). All of these regions also showed a reduction in global efficiency, and in the left postcentral gyrus, decreases in node degree and global efficiency were correlated with higher UHDRS scores in patients with manifest HD (t = À2.31, P = 0.01), whereas a lower node degree was correlated with higher DBS in pre-HD (t = À1.98, P = 0.03).

Graph Metrics in Cognitive Networks
Using a cognitive networks approach, decreases in node degree were found on the bilateral precuneus (P = 0.02 corrected), bilateral angular (P = 0.02 corrected), occipital (P = 0.03 corrected), and parahipoccampal gyrii (P = 0.03 corrected)-all DMN nodes-and the right anterior insula from the SN (P = 0.04 corrected; Fig. 3, right). Clustering coefficient was also reduced in the PCC (P = 0.001 corrected) and left anterior insula (P = 0.02 corrected). Finally, global efficiency was decreased in the right precuneus (P = 0.03 corrected). This decrease correlated directly with the Cogscore (t = 2.23, P = 0.01).

Graph Metrics in Pre-HD
On a whole-brain approach, the pre-HD group showed similar decreases in node degree compared with controls, centered around bilateral precentral and postcentral gyrii and the PCC (all P = 0.03 corrected). We also found that some of these nodes-the bilateral precentral and postcentral gyri-also showed decreased global efficiency compared to controls. Furthermore, path length was increased in motor and nonmotor regions of the cortex, including the bilateral insular cortex, bilateral fusiform gyri, and the bilateral amygdalae (all P = 0.04 corrected). Considering cognitive networks, pre-HD already showed a loss of clustering coefficient in the PCC (P = 0.01 corrected) and the left anterior insula (P = 0.03 corrected).
After controlling for GMV, all results remained significant with the following exceptions: node degree decreases in the supplementary motor area (P = 0.06 corrected after controlling for GMV); clustering coefficient decreases in the left anterior insula (P = 0.09 corrected after controlling for GMV); global efficiency decreases in the left precuneus (P = 0.06 corrected after controlling for GMV); and in pre-HD, path length increases in bilateral insular cortex and bilateral amygdalae (not present after controlling for GMV) and clustering coefficient decreases in the PCC and left anterior insula (not present after correcting for GMV).

Dynamic Independent Component Analysis
A total of 12 dynamic factors corresponding to withinnetwork and between-network connectivity patterns were identified in the whole sample: 4 components captured SN-DMN interactions, 4 identified CEN-DMN interactions, 1 corresponded to CEN-SN interactions, and 3 corresponded to within network connections. The outlook of the 12 dynamic factors is displayed in Appendix S1 in the supplementary material and Figure 2.
Comparing the HD group with healthy controls, patients had higher frequencies of DMN-CEN coupling, with the strongest connections between parahippocampal and precuneus DMN and orbitofrontal Central executive network (CEN) (factor 7, P = 0.04), whereas the controls showed an increased frequency of dynamic SN-DMN connections, highlighted by the links between the anterior cingulate and frontal cortex with the bilateral precuneus of the DMN (factor 2, P = 0.03; Fig. 4, bottom left). On the variability analysis, all of the dynamic factors showed increases in controls, related to DMN-CEN couplings, in which the right caudate of the CEN shows increased dynamic connectivity with DMN structures such as the bilateral precuneus, bilateral angular, and bilateral occipital cortices (factor 7, P = 0.04; factor 3, P = 0.03; factor 8, P = 0.002; Fig. 4, right). After performing a post hoc correction for the number of dynamic factors, only the variability of factor 8 remained statistically significant (P < 0.05 corrected). Importantly, the variability of factor 8, which connects frontal nodes of the CEN and posterior hubs of the DMN such as the PCC and precuneus, was directly correlated to the Cogscore (t = 1.98, P = 0.02) and inversely with the DBS (t = À2.32, P = 0.02).
Comparing pre-HD with controls, the controls preserve the higher frequency in the previously mentioned SN-DMN component (factor 2, P = 0.04), and variability remained higher for the DMN-CEN coupling in the control group, with higher connectivity between the frontal CEN hubs and the bilateral precuneus (factor 8, P = 0.02). The variability of factor 8 also correlated inversely with DBS in patients with pre-HD (t = À2.60, P = 0.02). Comparing pre-HD with sHD, the latter group presented a higher frequency of within-SN connections, highlighted by insular-precuneus connections (factor 4, P = 0.04), whereas patients with pre-HD showed an increased variability in SN connectivity, mainly insular cortex, with frontal regions of both the DMN and CEN (factor 9, P = 0.04).

Discussion
In the present study, we analyzed the landscape of FC changes associated with pre-HD and manifest HD and how they compare to healthy controls. Adopting both a whole-brain and a cognitive networks approach, we found commonalities with previous descriptions of brain networks in HD, but also a number of salient points that are derived from newer approaches such as dynamic FC analysis. We also examined the brain atrophy of the patients with HD compared with healthy controls to frame this functional analysis.
In patients with HD, widespread cortical thinning was observed compared with healthy controls. This pattern of extensive atrophy has been previously reported by our group 7,32 and others 2,4 and shows that the loss of cortical thickness occurs in frontal, temporo-parietal, and occipital cortices in patients with HD. Patients with pre-HD showed no differences compared with healthy controls using stringent statistical thresholds, a fact that has been shown in previous studies in patients with early HD, 33 although others have shown cortical thinning in the early stages. 5 However, this could be explained by the long time to onset in the patients with pre-HD in our sample (14.05 AE 6 years; see Table 1) because previous studies have shown almost no cortical thinning in patients with an estimated onset to disease higher than 5 years. 2 Furthermore, we observed that the structural changes did not have a pronounced effect in the functional analyses when gray matter volume was used as a covariate of no interest.
Using a whole-brain parcellation scheme, the changes in FC observed in our sample were mainly related to visual associative and motor regions. These changes involved the fusiform gyrus, occipital pole, and precentral/postcentral cortices and were also found when comparing patients with pre-HD with healthy controls. In previous studies, reduced FC in the somatomotor cortex has been related to the severity of motor symptoms. 9,10 Similarly, the visual networks have shown reduced FC in HD, 12 and this decrease was related to impairment in visual scanning and motor speed. Therefore, the reduced FC between motor and visual associative regions in our sample could underlie these clinical changes that are present, albeit in a reduced fashion, even in the premotor stage. 34 Importantly, these changes in our sample were linked to increasing disease burdens and worse motor scores. Graph theoretical findings also support these results, showing decreased degree and efficiency in precentral and postcentral hubs even in the premotor stage, and a widespread increase in path lengths across the network, signaling a decrease in the efficient transfer on information across its nodes. Of note, some cognitive regions were highlighted in this global analysis, mainly via an increase of FC between the PCC and the visual associative cortex, which correlates with higher DBSs. However, these changes are much less prominent compared with the visuomotor impairment.
The cognitive networks approach showed that patients with HD lose connectivity in the major largescale networks in the resting state. These decreases were particularly consistent regarding the DMN, which had already been reported in sHD. 10,13,35 In our sample, decreases of FC affected key nodes of the DMN, and this reduced connectivity translated into lower Cogscores as well as higher DBSs, especially for connections between the anterior and posterior components of the DMN. Furthermore, these decreases were accompanied by lower node degree, clustering coefficients, and global efficiency properties in these nodes, the latter showing a correlation with cognitive scores. Changes outside of FC had previously been documented only in relation with depression 36 or using regional metrics such as amplitude of low-frequency fluctuations. 37 More significant for us were the decreases found within the SN, which had not been found in previous studies of cognitive networks in HD. 38 We found a loss of FC within SN hubs such as the insula, the supramarginal gyrus, and the anterior cingulate and a further decrease in node degree and efficiency in the anterior insula. These changes were again correlated with clinical outcomes both in cognitive and DBSs. These findings are important given that the SN plays a critical role in the attribution and signaling of relevant inner and external information and thus in the engagement of goaldirected behavior and emotional processing. 39 Therefore, our findings may provide new insights on the functional mechanisms leading to the reduction on goal-directed behavior and emotional processing observed in HD. 34 We also found disruptions in these networks in the premanifest group, with lower FC within the DMN and decreased clustering coefficients both in the SN and DMN.
All of these concepts, however, are based on the premise that FC is "static" across resting-state acquisition, a concept that seems at odds with the dynamic nature of brain functioning. Reconciling classical resting-state network knowledge with a dynamic approach is challenging, but canonical large-scale networks could be considered as approximate static representations of the underlying dynamic brain. 40 Using dyn-ICA, we tried to retain the spatial identification of classical cognitive networks while adding temporal data such as frequency and variability, and we obtained two main conclusions. First, our results show that the dynamic connection between SN and DMN is less frequent in HD. These two networks were the most affected by the loss of intrinsic connectivity and topographical features in time-averaged analyses, and our results show that their interplay is also disrupted in a dynamic framework. An altered relationship between the SN and DMN can be found in previous static analyses of neurodegenerative diseases as a loss of anticorrelations, 26,41,42 but so far, to our knowledge, not in a dynamic framework. Our results lead us to believe that a time-dependent connection between these supposedly anticorrelated networks might be a trait of healthy participants and become lost in HD. This effect might probably be related to the SN's role as a "network fulcrum"-pivoting between executive and default-mode predominant settings-a concept proposed since the independent description of this network 43 that has been further confirmed using dynamic causal modeling. 44 Second-and perhaps more important given its clinical repercussion-although patients with HD show a higher frequency of DMN-CEN connections, their variability seems to be decreased across different dynamic factors. Above all, this loss in the dynamic interplay of the CEN and the DMN correlates with worse cognitive scores and higher DBS and can be found even in the premanifest group. Variability in brain networks is related to cognitive performance 45 and decreases with age, 24 and it could be conceived of as the repertoire of possible configurations of a brain network. 23 Thus, patients with HD seem to have a decreased range of network configurations that link DMN hubs, such as the precuneus, with CEN hubs such as the caudate nucleus. This altered dynamic resulted in worse clinical outcomes and is related to a more severe disease in our sample. We speculate that this lack of variability might be compensated by a higher frequency of this state, suggesting that CEN-DMN connections may assume maladaptive configurations that trade reduced efficiency for higher coupling.
We acknowledge some limitations to this study. First, cognition was explored mainly via the Cogscore, a widely employed and validated tool for cognitive assessment in HD. However, detailed neuropsychological evaluation might reveal more subtle disruptions of cognitive function, especially in patients with pre-HD, in which we found fewer cognitive correlations with networks disruption. Further studies will need to evaluate the link between more precise clinical evaluation and network dysfunction. Second, our resting-state acquisition time might not sufficiently represent all the dynamic variability of the HD brain; however, given the intrinsic difficulties of imaging acquisition in patients (especially given the disease-specific symptoms such as chorea) we believe this is a realistic rendering of the dynamic brain states in these patients. Third, healthy controls only had sociodemographic data available in this study. Finally, the choice of cortical parcellation could influence the results, and it is still up for debate as recent publications show. 46 However, we think this work strikes a balance between whole-brain, microanatomical, and functionally oriented parcellations as has been suggested in recent reviews. 31

Conclusions
To sum up, our results show that cognitive networks in HD are particularly affected, even in the premotor stages of the disease. These disruptions follow wellknown patterns in neurodegenerative disease, with decreases in the intrinsic connectivity of the networks and the loss of graph-theoretical properties. Furthermore, a dynamic approach uncovers network interactions unapparent in static analyses of FC that extend into the premanifest stage. Further studies will be needed to replicate these findings and link them to the diverse phenotypical and neuropsychological aspects of HD.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available from the corresponding author upon reasonable request