Alteration of spatial patterns at the network-level in facial synkinesis: an independent component and connectome analysis.

BACKGROUND: The treatment of post-facial palsy synkinesis (PFPS) remains inadequate. Previous studies have confirmed that brain plasticity is involved in the process of functional restoration. Isolated activation has been well studied, however, the brain works as an integrity of several isolated regions. This study aimed to assess the alteration of the brain network topology with overall and local characteristics of information dissemination. Understanding the neural mechanisms of PFPS could help to improve therapy options and prognosis.
METHODS: Patients with facial synkinesis and healthy controls (HCs) were estimated using functional magnetic resonance imaging (fMRI) of resting-state. Subsequently, an independent component analysis (ICA) was used to extract four subnets from the whole brain. Then we used the measurements of graph theory and calculated in the whole-brain network and each sub-network.
RESULTS: We found no significant difference between the patient group and the HCs on the whole-brain scale. Then we identified four subnetworks from the resting-state data. In the sub-network property analysis, patients' locally distributed properties in the sensorimotor network (SMN) and ventral default mode network (vDMN) were reduced. It revealed that γ (10,000 permutations, P=0.048) and S (10,000 permutations, P=0.022) within the SMN progressively decreased in patients with PFPS. For the analysis of vDMN, significant differences were found in γ (10,000 permutations, P=0.019), Elocal (10,000 permutations, P=0.008), and β (10,000 permutations, P=0.011) between the groups.
CONCLUSIONS: Our results demonstrated a reduction in local network processing efficiency in patients with PFPS. Therefore, we speculate that decreased characteristics in the intra-vDMN and intra-SMN, rather than the whole-brain network, may serve distinct symptoms such as facial nerve damage or more synkinetic movements. This finding of the alteration of network properties is a small step forward to help uncover the underlying mechanism. 2021 Annals of Translational Medicine. All rights reserved.

Introduction

Post-facial palsy synkinesis (PFPS) is a sequela of facial nerve injury. It manifests as involuntary abnormal facial movements accompanied by volitional movement in another area of the face, such as eye closure, triggered by the volitional movement of the mouth (1,2). However, the common therapeutic modalities for synkinesis, including facial neuromuscular retraining with physical therapy, targeted chemodenervation, and surgical procedures, are not sufficiently efficacious. Therefore, synkinesis remains a significant clinical challenge owing to its unclear causative mechanisms (3-5). In particular, it is unclear how abnormal information exchanges occur within the architecture of the sub-networks. This hinders our understanding of the full spectrum of pathophysiologic interactions in this debilitating movement disorder.

Previous neuroimaging studies on movement disorders have demonstrated brain regions responsible for sensorimotor control, aberrant excitation, inhibition of the basal ganglia circuitry, and abnormal integration of the basal ganglia, thalamus, insula, and cerebellum, with the primary sensorimotor engaged in disorder-specific alterations (6-9). Also, many studies of task-dependent on facial synkinesis lead us to conclude that there is one particular module that focuses on the projection area in the brain map, which specifically corresponds to ocular-oral or oral-ocular synkinesis (10). This isolated activation has been well studied. However, the brain works as an integrity of several regions but not several isolated modules (7,11). The brain network recruits neurons to complete various functional activities, such as sensorimotor, speech, and cognition, in an integrated circuit (12). Instead of exploring large-scale and distributed properties for information processing, research on facial synkinesis research has started to focus on regional alterations patterns or deficits (13).

To examine the brain plasticity changes associated with facial paralysis, we used an independent component analysis (ICA) to compare patients with PFPS and healthy controls (HCs) in resting-state functional magnetic resonance imaging (fMRI). We built a graph theory analysis based on the results of the ICA that identified the topological changes in patients with PFPS compared to HCs on the whole-brain network scale and functional sub-network range, to expand our understanding of global and focal alterations. PFPS is a sequela of facial palsy, which generally does not cause a strong reaction in the whole brain. However, most researchers still prefer to target an area (14). We hypothesize that functional sub-networks, rather than the whole-brain network, can better explain the abnormal movement, collectively inducing neural plasticity of pathologic involuntary activity. This study used a high-accuracy network analysis that was able to detect mild changes to a limited extent from the perspective of the connectome; this is useful for increasing our understanding of PFPS.

We present the following article in accordance with the MDAR reporting checklist (available at http://dx.doi.org/10.21037/atm-20-4865).

Methods

Participants

Thirty-one patients with PFPS (20 females; age range: 12–49 years; mean age: 33.61 years) and 19 HCs (17 females; age range: 23–42 years; mean age: 33.21 years) participated in this cross-sectional study (). Patients with Bell’s palsy were included, who were suffered from unilateral facial synkinesis, as well as participants that were recruited via two experienced plastic surgeons from the Department of Plastic Surgery. Inclusion criteria were medical history of 9 months or over and signed informed consent for the study. Participants with one or more of the following characteristics were excluded: Hunt syndrome or other causes (e.g., trauma, iatrogenic, or parotid tumors); central nervous system pathology; recurrent BP or concurrent peripheral neuropathy; a history of nerve transposition, and contraindications to an investigation by MRI. All fMRI measurements and clinical assessments were performed on the same day. The whole procedure lasts about 75 minutes. Besides, a neurological examination (normal vision and the capability to follow instructions) and an interview (the Sunny-Brook facial grading system for patients) were conducted. The HCs had to be free of facial palsy or other neurological impairments. All participants signed informed consent forms before proceeding to the trial according to the Helsinki declaration (as revised in 2013). The study was approved by the local ethics committee (approval number 2017-365-T267), and the study protocol was registered in the Chinese Clinical Trial Registry (ChiCTR1800014630).

Table 1

Demographic material of facial synkinesis patients and healthy controls

Items	PPFS (n=31)	HC (n=19)	P value
Age	33.61±8.55	33.21±6.75	0.863a
Sex (male/female)	11/20	2/17	0.105b
Duration	30.82±9.79 m	–	–
Affected side	13 (right)/18 (left)	–	–
Sunny-brook scores	41.45±14.07	–	–

a, two independent sample t-test; b, Chi-square test. PFPS, post-facial palsy synkinesis; HC, healthy controls.

Data acquisition and preprocessing

All subjects went through a whole-brain scan of resting-state fMRI (rs-fMRI). fMRI data were taken on a 3 T MRI scanner (GE Signa VH/I) using a 32-channel phased-array head coil. The rs-fMRI in whole-brain used a gradient echo-planar imaging sequence (interleaved scanning order, slice number =43, TR =2,000 ms, matrix size =64×64, FOV =220×220 mm, voxel size 3.4×3.4×3.2 mm3, number of acquisitions =240). The preprocessing and analysis of fMRI data was performed using SPM12 (http://www.fil.ion.ucl.ac.uk/spm) and the graph-theoretical network analysis was developed on the GRaph thEoreTical Network Analysis (GRETNA) (http://www.nitrc.org/projects/gretna/) toolbox (15). Briefly, after discarding the first 10 volumes of each fMRI run, slice timing was performed to correct the inconsistency of temporal collection. Then the data were motion-corrected, normalized to stereotactic Montreal Neurological Institute (MNI) space via a standard EPI template, and spatially smoothed (6×6×6 mm3 Gaussian kernel). Point-to-point head motion and mean head motion was estimated for subjects to control the motion-induced artifacts. Therefore, one patient with excessive head motion (cumulative translation or rotation >3 mm or 3°) was excluded (16).

Node and edge definition

Topological analyses in whole-brain network and functional sub-network were conducted in this study. We used an AAL90 structure-labeled atlas for parcellation in the whole-brain network analysis. Thus, nodes were defined as the parcellated 90 brain regions. Subsequently, Pearson’s correlations of the meantime series between all pairs of the 90 nodes were characterized as the edges.

For the sub-network analysis, preprocessed rs-fMRI data were decomposed into independent components (ICs) using GIFT software (Group ICA of fMRI Toolbox, version 4.0, http://icatb.sourceforge.net) to construct the sub-networks. Generally, the group ICA for this multi-subject analysis utilized a concatenation approach plus back-reconstruction (17). The images were first dimensionally reduced using principal component analysis (PCA), then temporally concatenated and reduced to extract 40 spatial components using the expectation-maximization algorithm at a group level. In addition, we performed the infomax ICA algorithm in ICASSO for 100 repetitions to validate its robustness. After the aggregated spatial maps were estimated, subject-specific spatial patterns and time courses were extracted through the back-reconstruction approach. A time course for all subjects as well as spatial weight maps were produced, which revealed the possibility of one voxel belonging to one particular component. Then we thresholded these maps at group level after a Z-transformation of the spatial weight map.

We selected meaningful ICs through spatial sorting and visual inspection. Four ICs-of-interest were identified in this study: the sensorimotor network (SMN) (), frontoparietal network (FPN) (), dorsal default mode network (dDMN) (), and ventral default mode network (vDMN) (). For these sub-networks, we specified each voxel as the node and functional connectivity of the voxel–voxel as the edge. Besides, we resampled the data and adjusted the voxel size to 6×6×6 mm3 to decrease the large scale of the connectivity matrix, which reduced the computational work significantly.

Figure 1

The spatial distribution pattern of four potential sub-networks extracted from resting-data by ICA. Four components (A-D) resemble the RSNs described in a previous study (18) and consist of regions known to be involved in the sensorimotor network (A), FPN (B), dDMN (C), and vDMN (D). Images (coronal, sagittal, and axial views) are t-statistics overlaid on the average high-resolution scan transformed into MNI152. Red to yellow represents t values, ranging from 6.0 to 12.0. The left hemisphere of the brain corresponds to the right side of the image. ICA, independent component analysis; RSNs, resting-state networks; MNI, Montreal neurological institute.

After the functional connectivity matrices were obtained, we calculated the network metrics of the whole-brain and the sub-networks separately using a pre-selected sparsity value (the ratio of the number of actual edges divided by the maximum possible number of edges in a network) to ensure the relative network organization. Specifically, the topological organization of whole-brain networks was analyzed over a wide range of network sparsity (0.05–0.45) (19), where the small-world metrics were analyzed (20). Network density was applied to each adjacent matrix with an increment of 0.01 for the whole-brain network and 0.02 for the sub-network metrics to reduce the computational dimension.

Network analysis

Global properties of the whole-brain network and sub-networks

Graph theory analysis is a suitable method to characterize the brain network topology properties (12,21). The commonly used network metrics, including the small-world property (clustering coefficient C, characteristic path length L, normalized clustering coefficient γ, normalized characteristic path length λ and small-worldness σ), network efficiency (local efficiency E and global efficiency E), assortativity r, hierarchy β and synchronization S), were calculated in this study. All the calculations of network metrics were completed with GRETNA. gives detailed descriptions of the above metrics (15,22).

Table 2

Descriptions of global network metrics examined in the present study

Parameter	Character	Descriptions
Clustering coefficient	Cp	The extent of local clustering of a network
Characteristic path length	Lp	The extent of overall routing efficiency of a network
Gamma	γ	The deviation of Cp of a network from those of surrogate random networks
Lambda	λ	The deviation of Lp of a network from those of surrogate random networks
Sigma	σ	The small-worldness indicating the extent of a network between randomness and order
Local efficiency	Elocal	How efficient of information propagation over a node’s direct neighbors
Global efficiency	Eglobal	How efficient of information propagation through the whole network
Assortativity	r	The tendency of nodes to link those nodes with similar number of edges
Hierarchy	β	How likely it is that all nodes oscillate with the same wave pattern
Synchronization	S	How likely that all nodes fluctuate in the same wave pattern

The small-world parameters hold on both modularized/specialized and distributed/integrated information processing and greatly amplify the efficiency of information transfer at a low-cost. Therefore, small-world networks have a shorter L than regular networks (high C and long L) but greater E than random networks (low C and short L). To characterize the small-world properties of the target network, we calculated its C and L with the corresponding value, which derive from the average of 100 random networks with the equal scale of nodes and degree distribution (23). Compared with the random networks, small-world networks have a relatively high normalized clustering coefficient γ (C/Crand) >1 and relatively low normalized characteristic path length λ (L/Lrand) ≈1 (24).

Global efficiency measures the efficiency of information transfer at a network level, which is the inverse of the harmonic mean of the minimum path length (25). While the local efficiency of the network measures the fault tolerance in a network, showing how the efficient communication exchanges when the first neighbors of a given node is phased out (25).

Additionally, we evaluated the hierarchical nature of networks using the β parameter (26), which defines the size of the power–law relationship between the clustering coefficient (C) and degree (k): Cp ≈ k (27). In a network with hierarchy organizations, some high-degree related nodes form a densely connected cluster. These generated clusters act as elements in the next level of the network and merge into a larger-scale interconnected cluster (28). We computed the parameter β for the network through the plot of log(C) versus log(k) applied with regression line fitting.

In addition, assortativity reflects the tendency of nodes to link those nodes with a similar number of edges, which measures the correlation between the node degree and the mean degree of its adjacent neighbor (29). Positive assortativity indicates that the densely connected nodes are more likely to associate with other nodes with the same degree. Synchronization, which is described as the ratio of the second smallest eigenvalue to the largest eigenvalue of the coupling matrix of the network, measures the possibility of the fluctuation of all nodes in the same wave pattern.

Regional properties of the whole-brain network

For nodal metrics, we also calculated the regional properties, including the nodal degree, nodal efficiency, and betweenness centrality of the whole-brain network. The nodal characteristics of the brain networks may demonstrate the importance of specific areas in brain network by measuring the extent to which a given node connects to all other nodes of a network (25,30).

Data analysis

The area under the curve (AUC) over the sparsity range was used to conduct group comparisons of the metric. Statistical tests of topological measures between the groups were performed using permutation testing to assess alterations of the topological parameters (10,000 permutations) for each topological parameter over a wide range of connection densities. For each permutation, participants were randomly assigned to patients and HCs with the same size as the original two groups. Next, the randomized groups difference of the metric was measured to obtain a null distribution. Depended on the generated null distribution, a P value was charged to evaluate the extent of alteration between groups. A P value of <0.05 was considered statistically significant (31). In the nodal properties analysis, a correction of false discovery rate (FDR) for multiple comparisons was performed at a q value of 0.05 (32).

Results

Demographic and clinical characteristics

The patients with PFPS (females; age: 33.61±8.55 years) and 19 HC (females; age: 33.21±6.75 years) were included in the statistical analysis. We observed non-significant differences between PFPS and HC concerning age (P=0.86) and gender (P=0.11). Demographic and clinical characteristics of these subjects is provided in . One patient was excluded because his head-motion beyond 3 mm or 3°.

Alterations in whole-brain organization of the functional connectome

In the initial test of global topological parameters, we found λ ≈1, γ >1, and σ = γ /λ >1, which indicates the networks are of small-worldness property. As displayed by , the topological metrics in whole-brain propagation between groups were non-significant. In the nodal analysis, none of the nodes in the regional properties’ analysis (including nodal degree, nodal efficiency, and betweenness centrality) were significant after FDR correction. That means, there was no significant difference between the patient group and the HCs on the whole-brain scale.

Figure 2

Global topological properties of functional connectome between the groups. The statistical comparison of area under the curve (AUC) showed no significant difference in C, L, γ, λ, E, E, assortativity, hierarchy, or synchronization between groups, and the bin-width of sparsity was 0.01. PFPS group: red symbols and lines; HC group: gray symbols and lines.

Alterations in sub-network organization of the functional connectome

We identified four subnetworks from the resting-state data after the ICA procedure (): SMN (component A), FPN (component B), dDMN (component C), and vDMN (component D). A comparison of component A between groups revealed that γ (10,000 permutations, P=0.048) and synchronization (10,000 permutations, P=0.022) within the SMN progressively decreased in patients with PFPS (). A comparison of components B and C found a non-significant difference between the groups. The FPN and dDMN show analogous patterns in PFPS and HC (). For component D, significant differences were found in γ (10,000 permutations, P=0.019), E (10,000 permutations, P=0.008), and hierarchy (10,000 permutations, P=0.011) between the groups. The PFPS group showed lower γ, E, and hierarchy of the vDMN than did the HCs ().

Figure 3

Functional brain network properties of the SMN with a bin-width sparsity of 0.02. The area under the curve (AUC) displayed no significant differences in C, L, γ, λ, E, E, assortativity, hierarchy, or synchronization between the groups. While the index γ was significantly lower in the PFPS group (P=0.048) and the index of synchronization was significantly decreased in patients (P=0.022). PFPS group: yellow symbols and lines; HC group: gray symbols and lines. *, indicates a significant difference between the PFPS group and the HC group.

Figure 4

Functional brain network properties of the FPN with a bin-width sparsity of 0.02. The area under the curve (AUC) displayed no significant differences in C, L, γ, λ, E, E, assortativity, hierarchy, or synchronization between the groups. PFPS group: blue symbols and lines; HC group: gray symbols and lines.

Figure 5

Functional brain network properties of the dDMNs with a bin-width sparsity of 0.02. The area under the curve (AUC) displayed no significant differences in C, L, γ, λ, E, E, assortativity, hierarchy, or synchronization between the groups. PFPS group: blue symbols and lines; HC group: gray symbols and lines.

Figure 6

Functional brain network properties of the vDMN with a bin-width sparsity of 0.02. The area under the curve (AUC) displayed no significant differences in C, L, γ, λ, E, E, assortativity, hierarchy, or synchronization between the groups. The index γ, E, and hierarchy were significantly lower in the PFPS group (P values were 0.019, 0.008, and 0.011, respectively). PFPS group: orange symbols and lines; HC group: gray symbols and lines. *, indicates a significant difference between the PFPS group and the HC group.

Discussion

In recent years, cortical plasticity in PFPS has been extensively discussed (33,34). To date, most work has attempted to identify neuro markers that reveal mechanisms and neurophysiological processes that are significant in PFPS (9,35). Considering that the experience of synkinesis has diverse influences, from a deficit in control to social context, neuro marker candidates will be found to reflect all aspects of involuntary movements in all contexts and help to afford clinical prognostic indicators (36-38). We utilized a graph theory analysis to examine the resting-state fMRI data to characterize the global and local features after facial synkinesis, instead of localization information. Beyond regional activity, PFPS in our study was found to influence intrinsic brain networks (also known as resting-state networks), mainly the somatosensory networks, and default mode networks, which subserve homeostatic, attentional, cognitive, executive, and sensory functions together and alter their functional connectivity. Therefore, facial synkinesis is a major complication of facial paralysis while it does not induce extensive plasticity of the whole brain. By contrast, it only leads to limited topological changes in some functional sub-networks.

Our study shows that the activity induced by synkinesis is related to many brain regions that belong to different functional brain networks, rather than the dedicated areas within the brain. These components of networks belong to the default mode, and somatosensory networks that control the cortical and subcortical representations of synkinesis movements, which have been clarified in large-scale resting-state studies, extensively overlap with the regions related to PFPS (39). The functional network of the brain is complicated and includes areas associated with coding information regarding muscular control and reactions to muscular control, the recognition and modulation of spatial information, and somatosensory information and attention (40). This might be the reason for the deactivated SMN, which plays an important role in motor control and consists of somatosensory (post-central gyrus) and motor (pre-central gyrus) regions, and extends to the supplementary motor areas. The vDMN, which is spontaneously active during passive performance and associated with effective processing, memory, and self-projective thinking (41,42), consists of the posterior cingulate cortex, medial prefrontal cortex, angular gyrus, hippocampus, parahippocampus, retrosplenial cortex, and posterior inferior parietal lobe. Consequently, these intrinsic networks and dedicated connectomes may be engaged during facial synkinesis. The primary aim of this study was to define the scale affected by facial synkinesis.

Although regular networks and random graphs are both useful idealizations, many real networks lie somewhere between order and randomness. Small-world networks have been extensively researched since the fundamental work by Watts and Strogatz, who indicated that random mutation of a few connections in a regular or lattice network extensively lessened the mean minimum path length between any pair of nodes while holding strong local connectivity or “cliquishness” (43). The coexistence of integration and separation further reflects the small worldness characteristics of the networks.

The local efficiency of the network evaluates how efficient communication is among the first neighbors of a given node when it is eliminated. A lower E in SMN and vDMN means that the possibility of “off track” increases, suggesting a trend of “separation” of these functional sub-networks. The decrease in E also indicates the randomization of the sub-networks, reflecting the decrease in overall efficiency and change of the integration-separation balance in different neural functional sub-networks. Hierarchy is the tendency of hub regions to connect with nodes outside the hubs (44). In this study, it was found that the subjects’ functional complex sub-networks had small-world attributes, and the efficiency of regional information transmitting in SMN and vDMN decreased. It is suggested that patients with facial synkinesis can cause changes in the mode of brain information processing.

The declined propagation characteristics in the sub-network represent less active neural transmitting activity in facial paralysis. Therefore, the connection of the subnetworks is diminished. The results imply a disruption of the efficiency of local information in the sub-networks, which is a manifestation of maladaptive alteration in facial synkinesis. According to previous studies, facial synkinesis seems to be a facial mismatch on neural activity rather than deactivation (39,45). We found that facial synkinesis tends to reduce the cluster efficiency to lower than normal, which may illustrate the effects on cortical plasticity from the perspective of neural mechanisms. Furthermore, we discovered many measurements, including network efficiency and synchronization, that showed neither significant nor trending differences in any sub-network. Therefore, we believe that they may not be involved in the neuroplasticity of facial synkinesis. Such interpretations are not difficult as post-facial paralysis synkinesis only causes changes in topological properties to a limited extent. This would not be significant if the contrast were executed on a global scale. These findings should be confirmed with a larger sample size in the future.

The present study investigated the topological structure of functional brain networks between HCs and patients with PFPS. The rs-fMRI images were examined using a graph theory analysis to study the brain functional connectomes. No significant differences in the whole-brain scale were observed between patients with PFPS and HCs. However, patients with PFPS had decreased small-worldness and synchronization in the SMN and decreased small-worldness, local efficiency, and hierarchy structures in the vDMN. Moreover, patients with PFPS had altered network metrics in the SMN and vDMN brain regions, mainly involved in information transport, attention, memory, and emotion.

Conclusions

The current results provide the first evidence that reduced intrinsic connectivity represents regionally restricted information transmission efficiency in the SMNs and vDMNs of patients with PFPS. PFPS produces changes in neuroplasticity in sub-networks rather than the whole brain network. These findings indicate that functional disruption in patients with PFPS may be associated with instabilities in the local topological organization of the brain functional connectome. In addition, this study provides new insights into the pathophysiological alterations to further reaching the mechanism of PFPS in-depth.

The article’s supplementary files as