Reconfiguration of large‐scale functional connectivity in patients with disorders of consciousness

Abstract Introduction Functional connectivity alterations within individual resting state networks (RSNs) are linked to disorders of consciousness (DOC). If these alterations influence the interaction quality with other RNSs, then, brain alterations in patients with DOC would be characterized by connectivity changes in the large‐scale model composed of RSNs. How are functional interactions between RSNs influenced by internal alterations of individual RSNs? Do the functional alterations induced by DOC change some key properties of the large‐scale network, which have been suggested to be critical for the consciousness emergence? Here, we use network analysis to measure functional connectivity in patients with DOC and address these questions. We hypothesized that network properties provide descriptions of brain functional reconfiguration associated with consciousness alterations. Methods We apply nodal and global network measurements to study the reconfiguration linked with the disease severity. We study changes in integration, segregation, and centrality properties of the functional connectivity between the RSNs in subjects with different levels of consciousness. Results Our analysis indicates that nodal measurements are more sensitive to disease severity than global measurements, particularly, for functional connectivity of sensory and cognitively related RSNs. Conclusion The network property alterations of functional connectivity in different consciousness levels suggest a whole‐brain topological reorganization of the large‐scale functional connectivity in patients with DOC.


| INTRODUC TI ON
Disorders of consciousness (DOC) encompass a set of particular conditions occurring after coma (Bruno, Laureys, & Demertzi, 2013), including the minimally conscious state (MCS), in which patients exhibit signs of fluctuating, yet, reproducible remnants of nonreflex behavior, and the unresponsive wakefulness syndrome (UWS), related to patients who open their eyes, but remain unresponsive to external stimuli (Laureys & Schiff, 2012;Schnakers & Laureys, 2012). Due to the difficult communication imposed by these conditions, brain activity registered at rest (Biswal, 2012) is used to develop complementary diagnosis approaches. In particular, resting state functional magnetic resonance imaging (R-fMRI) protocols are used to understand brain activity while subjects are not exposed to stimuli Sporns, 2013), overcoming the need for their active participation. R-fMRI studies in healthy controls (HC) suggest that the brain is organized into large-scale resting state networks (RSNs) of sensory/cognitive relevance (Fox & Raichle, 2007;Rosazza & Minati, 2011). At least ten of these functional entities were identified in HC including auditory, cerebellum, default mode network (DMN), executive control left (ECN Left), executive control right (ECN Right), saliency, sensorimotor, visual lateral, visual media, and visual occipital (Damoiseaux et al., 2006). RSNs provide a suitable representation to study the preservation of sensorial and cognitive brain functions without any explicit stimulation (Rosazza & Minati, 2011) specifically for DOC studies.
First analyses of RSNs in patients with DOC focused on alterations of the functional connectivity inside the DMN. This is a functional structure that encompasses specific brain regions linked to the consciousness emergence phenomenon Demertzi, Soddu, & Laureys, 2013). Decreases in functional connectivity within this network are linked to modifications of the level of consciousness in these patients. Posterior studies showed that DOC conditions may affect functional connectivity within multiple RSNs Di Perri, Stender, Laureys, & Gosseries, 2014;Di Perri, Thibaut, et al., 2014;Guldenmund et al., 2013;Heine et al., 2012;Ribeiro de Paula et al., 2017). In particular, variations in intrinsic connectivity for specific RSNs were related to alterations in sensorial and awareness functions Demertzi et al., 2013;Di Perri, Stender, et al., 2014). Additional evidence indicates changes in the connectivity between RSNs, for instance, reductions of the connectivity strength between RSNs in patients with DOC compared to HC subjects (Rudas et al., 2014) and alterations in the level of anti-correlation between RSNs associated with the recovery of consciousness (Di Perri et al., 2016). In summary, these analyses focused on alterations within particular RSNs or between specific pairs of RSNs that may have functional relevance for consciousness emergence.
Nevertheless, these approaches may be limited because they do not consider a more general view of the brain, regarding, for instance, the existence of multiple functional units in the brain and the interactions among them (van den Heuvel & Hulshoff Pol, 2010). They are instead focused on specific consciousness-related circuits within the brain. A more general perspective is important because consciousness preservation in these patients would also require functional units related not only to consciousness processing but also to stimuli and response, and possibly systems to orchestrate them (Tononi & Koch, 2015). The understanding of interactions among these units may provide valuable information about these conditions (Tononi & Koch, 2015). Recently, a model of functional connectivity among RSNs has been proposed in the so-called functional network connectivity (FNC; Jafri, Pearlson, Stevens, & Calhoun, 2008), which considers the functional interaction between these large-scale units.
This model provides a network representation in which interactions between high-order functional systems can be characterized using network measurements (Bullmore & Sporns, 2009van den Heuvel & Hulshoff Pol, 2010). Lately, connectivity density decreases were associated with consciousness alterations in coma, providing a general description of FNC alterations (Malagurski et al., 2019).
However, the specific reconfiguration of FNC associated with consciousness states is not tackled.
In this study, we hypothesize that the FNC model may highlight reorganizations of connectivity related to the underlying pathology characterizing the DOC condition. These interaction patterns were studied by assessing modifications in integration, segregation, and centrality properties, which have been suggested to be highly relevant for consciousness emergence (Tononi & Koch, 2015). These properties were analyzed for three populations in different states of consciousness: healthy controls, subjects with MCS, and subjects with UWS. In contrast to previous studies that only focused on a limited set of RSNs, we considered the interactions among the whole set of functional units. To reach this objective, the FNC was computed for each subject obtaining a general brain functional representation with the interactions between RSNs. Next, we used a set of network measurements to assess the mentioned properties.
In particular, degree, strength, clustering coefficient, betweenness, and eigenvector centralities were used to understand key | 3 of 13 MARTÍNEZ ET Al.
This connectome corresponds to a large-scale network of functional relationships between functionally related brain regions.
Network-based measurements computed on this connectome provide a functional depiction of synchronized, spontaneous, and segregated activity (Rubinov & Bullmore, 2013;Sporns, 2013;van den Heuvel & Hulshoff Pol, 2010). Importantly, there is a methodological challenge in the characterization of the functional relationship between large-scale areas (RSNs) in patients with severe brain damage. Particularly, brain-injured patients may present functional and structural affectations that may change the connectome properties. Therefore, in this study, a particular processing pipeline that accounts for these alterations were considered, including, severe structural affectations, large head motions, and individual variability, among others. Figure 1 summarizes the process used to characterize functional connectivity alterations at the general brain level of interactions between RSNs.

| Subjects and patients
Participants were healthy volunteers and patients with UWS or MCS following severe brain damage studied at least 5 days after acute brain insult. HC subjects were subjects free of psychiatric or neurological history. Clinical examination was performed using the French version of the Coma Recovery Scale-Revised (CRS-R; Giacino, Kalmar, & Whyte, 2004;Schnakers et al., 2008). The CRS-R is a standardized measure for characterizing the level of consciousness and monitoring recovery of neurobehavioral function (Giacino et al., 2004). It consists of 30 hierarchically arranged items that comprise six subscales addressing auditory (5 items), visual (6 items), motor (7 items), oromotor/verbal (4 items), communication (4 items), and arousal (4 items) processes. The scoring is based on the presence or absence of specific behavioral responses to sensory stimuli administered in a standardized manner, and the lowest item in each subscale represents reflexive activity while the highest item represents cognitively mediated behaviors (Giacino et al., 2004;Schnakers et al., 2008). Exclusion criteria were contraindication for MRI (e.g., presence of ferromagnetic aneurysm clips, pacemakers), MRI acquisition under sedation or anesthesia and large focal brain damage (>50% of total brain volume). Structural brain damage was assessed by visual inspection of two experts. Written informed consent to participate in the study was obtained from the healthy subjects and from the legal surrogates of the patients. The study was approved by the Ethics Committee of the Medical School of the University of Liège ).

| Data description
Acquisitions from 75 subjects were used for this study: 27 HC subjects (14 women, mean age 47 ± 16 years), 24 patients with MCS (eight women, mean age 47 ± 16 years; nine of nontraumatic etiology: two anoxic, three with cerebrovascular accident, three with hemorrhage, one with seizure; 14 of traumatic, and one of mixed etiology), and 24 with UWS (12 women, mean age 50 ± 18 years; 18 of nontraumatic etiology: nine anoxic, six with cerebrovascular accident, two with hemorrhage, one metabolic; five of traumatic, and one of mixed etiology). Thirty-one patients with UWS and MCS were assessed in the chronic setting, that is, ≥50 days postinsult. Further details about the patients' demography can be found in Table S1.
The three initial volumes were discarded to avoid T1 saturation effects. In addition, for anatomical reference, a structural T1-weighted image was acquired. Patients were scanned in sedation-free condition, and healthy volunteers were instructed to close their eyes, relax without falling asleep and refrain from any structured thinking (e.g., counting and singing), as commonly performed in resting state paradigms (Beckmann, Luca, & Devlin, 2005;Guldenmund et al., 2013).

| Data preprocessing
Data preprocessing was performed using the Statistical Parametric Mapping (SPM8) (Friston, 2007) toolbox for Matlab (The Mathworks, Inc.). SPM preprocessing stages included realignment and adjustment for movement-related effects, coregistration of functional onto structural data, segmentation of structural data, normalization into standard stereotactic MNI space, and spatial smoothing with a Gaussian kernel of 8 mm. To evaluate the data acquisition quality, the frame-wise displacement (Power, Barnes, Snyder, Schlaggar, & Petersen, 2012) was assessed on each population, further details Supplementary Material Section 2. Motion correction (e.g., small, large and rapid motions, noise spikes and spontaneous deep breaths) was applied by using ArtRepair toolbox for SPM Mazaika, Hoeft, Glover, & Reiss, 2009).

| Resting state networks identification
For each subject, the resting state networks (RSNs) were selected as follows: First, the rs-fMRI signal was decomposed into maximally independent spatial maps using spatial ICA (McKeown et al., 1998).
ICA decomposition was performed with 30 components (Jafri et al., 2008) and the infomax algorithm as implemented in GroupICA toolbox (Calhoun et al., 2001). Each spatial map (source fMRI signal) has an associated time-course, which corresponds to the common dynamic exhibit by the component. Second, RSNs were identified at individual level  by using a two-fold process: template matching and neuronal/artifactual classification (see Supporting Information Section 4). Template Matching is an approach that aims to identify each RSN directly from the single subject sICA decomposition . It is a matching problem with two constraints: (a) a template had to be assigned to one of the 30 ICs and (b) an IC could be labeled as an RSN or not. These two conditions ensure that all the templates (one for each RSN) have to be assigned and a unique identification of each IC, which deal with the potential concurrent component assignation. The pair between the template and the IC with the highest goodness-of-fit score was selected . Later, a neuronal/artifactual classification of independent components (ICs) was performed by using a machine learning-based labeling method . It consists of a binary classification approach by means of support vector machine (SVM) classifier trained on 19 independently assessed healthy subjects. This SVM uses the fingerprints obtained from ICA decomposition (n = 30 components) as the feature vector containing both spatial (i.e., degree of clustering, skewness, kurtosis, spatial entropy) and temporal information (i.e.,

| Functional network connectivity estimation
For each subject, a FNC matrix was computed by using a measure of dependency between pairs of representative time-courses, resulting in a matrix with strengths of the interactions between the identified RSNs. The strength for edges pointing to RSNs marked as F I G U R E 1 Illustration of the methodological procedure defined as the sequence of the following processes: data acquisition consisting of 300 volumes of functional magnetic resonance imaging (MRI) at rest and a structural MRI for each subject; data preprocessing including brain extraction, alignment, registration, Gaussian smoothing, motion correction, and normalization; extraction of the resting state networks (RSNs) using spatial independent component analysis and a template matching strategy under a data-driven approach; computation of functional network connectivity (FNC) between RSNs by the lagged distance correlation method; and finally, computation of integration, segregation and centrality measurements to characterize the populations in different states of consciousness no-neuronal was set as zero, indicating no interaction. The measures of dependency level were computed using the distance correlation (DC) between time-courses (Székely, Rizzo, & Bakirov, 2007). DC aims to measure nonlinear dependencies between two random variables X and Y with finite moments in arbitrary dimension. In order to account for time delays, a circular shifted lagged version of the DC was used (Jafri et al., 2008;Rudas et al., 2014). Once the FNC was computed, it induces a functional connectivity matrix, which was used to characterize alterations of functional connectivity. In particular, a 10 × 10 weighted matrix was computed to model interactions between different RSNs. Each one of them models a brain region associated with specific arousal and awareness regions related to consciousness emergence. An entry c ij in this matrix corresponds to the interaction between the RSN i and RSN j assessed by using the lagged DC. Further details of this approach can be found in Supporting Information Section 5.

| Network characterization
Functional network connectivity matrix contains a measure of dependency between pairs of RSNs time-courses. To assess functional connectivity alterations, three network properties were computed for each FNC matrix, namely, integration, segregation and centrality of the functional connectivity between RSNs. FNC integration was assessed by degree and strength (Bullmore & Sporns, 2009Rubinov & Sporns, 2010). FNC segregation was characterized by clustering coefficient (Bullmore & Sporns, 2009Rubinov & Sporns, 2010), and FNC centrality was estimated by betweenness centrality and eigenvector centrality (Lohmann et al., 2010; see a brief description of the network measurements in Table S3). These computations were performed using the brain connectivity toolbox (Rubinov & Sporns, 2010

| Statistical analysis
To assess the discrimination power of the network properties, an unpaired-sample t test (Welch, 1947

| RE SULTS
In this study, a set of network measurements were used to assess the integration, segregation, and centrality of the FNC between RSNs to characterize connectivity variations in different states of consciousness. were also found when comparing HC versus subjects with UWS and when comparing HC and subjects with DOC in auditory network, DMN, ECN Left, visual medial network, and ECN Right. Also, degree values for subjects with MCS were greater than the UWS in all RSN but no significant differences were observed. Table S4 reports statistical details of these assessments.

| Loss of functional network connectivity integration in DOC
As observed in Figure 2b

| Alterations of functional network connectivity centrality in DOC
Betweenness centrality and eigenvector centrality values are reported in Figure 5. Betweenness centrality values were higher for HC in contrast to subjects with altered states of consciousness for DMN, ECN Left, ECN Right, salience network and cerebellum network, as observed in Figure 5a. These centrality values of subjects with UWS were higher when comparing to subjects with MCS and when comparing to HC subjects, for auditory, sensorimotor, visual lateral, visual medial and visual occipital networks. Also, ECN Right and salience network has values of zero of betweenness centrality for UWS patients, indicating that these nodes were not part of any shortest path in the network. Table S7 reports statistical details about these comparisons.
As observed in Figure 5b,

| D ISCUSS I ON
In this paper, we studied whole-brain functional connectivity changes in different states of consciousness: HC, subjects with MCS and subjects with UWS. Unlike previous approaches, which mainly focus on functional units that seem to be associated to consciousness, that is, DMN, the proposed model considers a more general set of functional units that represents connectivity between RSNs, including sensory and cognitive-related ones. So, this model provides a general perspective of the cognitive and sensory RSNs connectivity variations in altered states of consciousness. We assessed the interactions between these RSNs, resulting in a FNC model that allows describing whole-brain system-level interactions. We used a model of functional connectivity among RSNs, which corresponds to large spatial scale segregated functional units. In this model, the interaction between pairs of representative time-courses of each RSN is computed (Biswal, 2012), resulting in a FNC model that allows describing whole-brain interactions (Jafri et al., 2008). The proposed model provides a general perspective which overcomes the specific or single consciousness-related region studies Di Perri, Thibaut, et al., 2014;Guldenmund et al., 2013).
It allows describing the entire brain functional reconfiguration in a broad scale of regions related to sensory and cognitive processes.
This reconfiguration description goes further than previous general density studies in coma (Malagurski et al., 2019)   to other RSNs than in subjects with DOC ( Figure 2). This reduction of the degree values in altered states of consciousness could be understood as a reduction of relationships between RSN time-courses, that is, representative time-courses are less or not correlated, suggesting an alteration of the functional connectivity structure in this patients. This result corroborates the disruption of external and internal awareness networks  and the decrease in anti-correlated connectivity previously observed in subjects with DOC (Di Perri, Thibaut, et al., 2014). Also, functional connectivity of salience network was reported as diminished in altered states of consciousness (Guldenmund et al., 2013). This network is usually associated with the orchestration between internal attention and task-related-states, and its alterations were linked to consciousness disorders (Heine et al., 2012). In our experiment degree values for salience network support this observation. A more detailed analysis of the integration phenomena can be obtained by studying strength values alterations (Figure 3). These values were also reduced in altered states of consciousness indicating that the amount of information that different time-courses share is lower for subjects with DOC.

| Integration alterations in DOC
This observation confirms the functional disruption associated with the severity of the pathological condition, as was reported for highly detailed networks in distinct consciousness states (Ribeiro de Paula et al., 2017). Further, this reduction could result from a deterioration process of the connectivity between RSNs, which can be an effect of the connectivity drops in small regions (Hannawi et al., 2015). Also, averaged integration measurements, both degree and strength, suggest that preserved levels of consciousness seem to be related to narrow distributions for integration values. In particular, patients with UWS seem to exhibit a larger variety of connectivity values including hyperconnectivity (increment of connectivity) and disconnections, when compared to healthy subjects. Subjects with altered states of consciousness not only reduce the number of connections between RSNs, but also degrade the ones that remain, suggesting a reduction of the synchronization level associated with the communication between networks. Importantly, the measures herein proposed were computed in large spatial regions that contain previously studied areas, such as the thalamo-cortical circuit . Therefore, breakdowns in integration seem to appear not only for small brain areas, as reported for strength reductions in the connectome computed from EEG (Cacciola et al., 2019)

| Segregation alterations in DOC
High values in clustering coefficient of the DMN seem to be related to the level of synchronization of this network with other RSNs. This result was previously reported in specific awareness circuits involving the DMN . Segregation measurement assessed by clustering coefficient confirms that consciousness could be a phenomenon involving segregated functional units that work in an integrated manner (Tononi, Sporns, & Edelman, 1994). RSNs could be understood as segregated regions that execute specific tasks (Biswal, 2012) but share information in consciousness phenomena (Heine et al., 2012). Sensory and cognitive-related networks appear to be more clustered for HC. In contrast, the segregation increases for sensorimotor in DOC with no significant differences between MCS and UWS. A similar finding was reported in an experiment with altered states of consciousness and anesthesia  where an increment of functional connectivity between thalamus and sensorimotor network was found in altered states of con-

| Centrality alterations in DOC
Centrality measurements indicate how central a node is in the network. High centrality scores in auditory and sensorimotor networks suggest that these functional units play a central role in patients with altered states of consciousness, revealing a behavior alteration phenomenon even if these regions exhibit a functional connectivity reduction in patients with altered states of consciousness, as was previously reported Kirsch et al., 2017).
This observation could be further explored to understand the kind of variation induced by DOC that reveals a centrality increment.
Similarly, higher scores in sensorimotor network suggest that this network also change its nature, becoming more important in subjects with altered states of consciousness. Interestingly, even if the sensorimotor input-output loops were reported as not required for consciousness (Tononi & Koch, 2008), the circuits involving these RSN were altered by the pathology (Di Perri, Stender, et al., 2014).
A surprising finding is the increment of centrality values for this external awareness network in subjects with DOC, which is not expected due to its associated behavior to sensory stimuli and motor reaction. A similar finding was reported by Cacciola et al. (2019), where an increment of the betweenness centrality in posterior cingulate and visual areas were stated for patients with UWS. However, this finding at RSN level can be a result of a brain reconfiguration in response to not-conscious stimuli response (Tononi & Koch, 2015).
Increases of centrality values of functional connectivity between RSN in altered states of consciousness suggest a modification of their time-courses nature, becoming more relevant in subjects with DOC. Nevertheless, this new central role of some RSNs would be not suitable for consciousness phenomena, where a sort of equilibrium between segregation and integration is required (Cacciola et al., 2019;Tononi & Koch, 2008Tononi et al., 1994). Summarizing, centrality alterations describe a reconfiguration of the relevant functional units in altered consciousness states that seem to be not suitable for the emergence of consciousness.

| Limitations and future directions
The analysis developed in this experiment presents some methodological limitations. A potential confounding factor is related to head motion. In order to study this potential bias source, frame-wise displacements (FWD; Power et al., 2012) were computed for each group to evaluate data quality acquisition. According to FWD, images with major displacements have to be removed for the func- Therefore, it is reasonable to assume a small influence of large head motions in the results herein reported. Another potential confounding result is related to the brain gray matter reduction in patients with DOC. Gray matter volume can influence the functional connectome measurements by reducing the amount of voxels which are considered to be in a region, that is, in a RSN, (Table S2). In addition, the proposed analysis was made for functional connectivity between RSNs, that is, for a 10 × 10 matrix representing a broad picture of the brain functionality. This feature limits more specialized analyses as those made to study the topology for larger networks, that is, hubs (Bullmore & Sporns, 2012), and small-world (Rubinov & Bullmore, 2013;Sporns, 2003 represented by an averaged time-course which explains the neuronal activity of the entire region. This representative time-course was built from a data-driven approach, a combination of spatial ICA of neuronal nature (Jafri et al., 2008). Functional connectivity between the representative time-courses of each RSN was computed by using the lagged distance correlation (Rudas et al., 2014 Other graph measurements, like small-world, rich club, efficiency and shortest paths, have been successfully applied to analyze functional connectivity alterations associated to different pathologies (Rubinov & Sporns, 2010). These measurements provide a better understanding of network topological alterations; however, they require large networks, that is, networks with a large number of nodes. Statistical analysis was based on the family-wise approach that does not require interpretation of any property. In this work Network-Based Statistic (Zalesky, Fornito, & Bullmore, 2010;NBS) was not used to identify differences between networks due to the following considerations: (a) the network size; each FNC has 10 nodes that means a maximum of 45 comparisons per network, (b) the power of the contrast ratio suggested in NBS was not suitable. It looks for preserved connections between nodes at different threshold computations which impose a minimum strength in the relationships. These conditions were also indicated by the NBS author, who highlights that the connections comprised the contrast of interest might form components, that is, regions with high power. If they do not form components, or if the extent of the components formed are too small, the NBS is ineffective. This is the case for the 10 × 10 FNC. Finally, in order to get a clear idea about integration, segregation, centrality and topological alterations related to specific areas or circuits in the brain, functional connectivity networks in different scales for the entire brain can be explored. This permits further understanding of time-courses alterations in sensorimotor and auditory networks to capture the essence of modifications induced by the altered state of consciousness.

| CON CLUS IONS
We use a general model of brain functionality to study its modifications in different states of consciousness. We found that this general model built from large-scale areas exhibits connectivity alterations induced by the pathology. In particular, we use network measurements to observe modifications of the FNC linked to consciousness level. Our results suggest that the FNC is better integrated and segregated for healthy subjects than for patients with DOC except by the sensorimotor network. Besides, FNC centrality indicates that there exists a role alteration in sensorimotor and auditory networks for patients with DOC where these RSNs become more important.

ACK N OWLED G M ENTS
We would like to thank Johann Martínez for his useful suggestions in network analysis, Camilo Salomón for his support in pipeline preprocess, and Alejandra Parra for her writing assistance and language editing. J.R. is a doctoral student at Universidad Nacional de Colombia supported by funds from the agreement between the Gobernación del Magdalena and COLCIENCIAS (Instituto Colombiano para el Desarrollo de la Ciencia y la Tecnología) for the education of high quality human capital in the department of Magdalena, Colombia.

CO N FLI C T O F I NTE R E S T
The authors report no competing interests.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data and source code that support the findings of this study are available from the corresponding author upon reasonable request.