The functional and structural neural correlates of dynamic balance impairment and recovery in persons with acquired brain injury

Dynamic balance control is associated with the function of multiple brain networks and is impaired following Acquired Brain Injury (ABI). This study aims to characterize the functional and structural correlates of ABI-induced dynamic balance impairments and recovery following a rehabilitation treatment. Thirty-one chronic participants with ABI participated in a novel rehabilitation treatment composed of 22 sessions of a perturbation-based rehabilitation training. Dynamic balance was assessed using the Community Balance and Mobility scale (CB&M) and the 10-Meter Walking Test (10MWT). Brain function was estimated using resting-state fMRI imaging that was analysed using independent component analysis (ICA), and regions-of-interest analyses. Brain morphology was also assessed using structural MRI. ICA revealed a reduction in component-related activation within the sensorimotor and cerebellar networks post-intervention. Improvement in CB&M scale was associated with a reduction in FC within the cerebellar network and with baseline FC within the cerebellar-putamen and cerebellar-thalamic networks. Improvement in 10MWT was associated with baseline FC within the cerebellar-putamen and cerebellar-cortical networks. Brain volume analysis did not reveal structural correlates of dynamic balance, but dynamic balance was correlated with time since injury. Our results show that dynamic balance recovery is associated with FC reduction within and between the cerebellar and sensorimotor networks. The lack of global structural correlates of dynamic balance may point to the involvement of specific networks in balance control.

MRI acquisition. Participants underwent two identical MRI sessions, pre-and post-intervention, each session including a 3D anatomical scan, Rs-fMRI scan, fMRI-Localizer scan, and DTI scan (not reported here).
High-resolution T1-weighted anatomical data were acquired with fast-spoiled gradient-echo (FSPGR) sequence, with a voxel size of 1 × 1 × 1 mm, (Repetition Time (TR) = 8165 ms, Echo Time (TE) = 3.74 ms, 256 × 256 acquisition matrix). The field of view (FOV = 192 mm) covered the entire cerebrum and the cerebellum. The duration of the scan was four minutes and 50 s. The fMRI data was acquired using a gradient echo EPI with voxel size of 3 × 3 × 3 mm (mm), TR = 2000 ms (ms), TE = 35 ms, flip angle = 77°, 35 slices, with a 0.6 mm gap, and lasted nine minutes and 50 s. The fMRI-Localizer scan was a block-design experiment containing six conditions (five movement-and one resting-condition); (1 + 2) left-and right-limb movements of dorsi/palmar flexion off palms, (3 + 4) left and right dorsi/plantar flexion for the ankles, and (5) bipedal ankle movements. The movement frequency of the limbs was equal to 1 Hz (Hz) and was demonstrated by the experimenter before the scan. Movement blocks of 12 s were separated by resting periods of 10 s, which were cued by a fixation cross ["+"] that was presented on a black screen. Visual cues for instructing hand and ankle movements were displayed on a screen during the experiment. Participants trained on the task before the scan using a dedicated apparatus located outside the scanner. The order of the blocks was random. In total, each localizer session included 25 movement blocks (five repetitions of each of the five movement conditions). This scan was conducted to define Regions of Interest (ROIs) based on functional activations.
The resting-state data acquisition parameters were similar to those in the fMRI-Localizer scan. During the resting-state session, a cross ["+"] was displayed in the middle of the screen, and participants were instructed to fixate on it during the scan. This scan lasted seven minutes and 26 s.
The magnetic resonance imaging and data acquisition were performed at the Imaging Center of Soroka Medical Center using a 3-Tesla Philips Ingenia whole-body MRI scanner (Philips Ingenia, Amsterdam, Holland). Imaging analysis. Functional, resting-state, and structural data were analysed by Brain Voyager 20.6 (Brain Innovation, Maastricht, The Netherlands). Brain segmentation was performed using FreeSurfer V 5.0 (developed by the Laboratory for Computational Neuroimaging at the Athinoula A. Martinos Center for Biomedical Imaging at Massachusetts General Hospital in Charlestown in Boston, MA).
Pre-processing of the localizer scan data. Pre-processing included removal of the first two functional images of each run series to allow stabilization of the BOLD signal; correction of the slice scan time acquisition (ascending-interleaved, using a cubic-spline interpolation algorithm); and head-motion correction (using a trilinear/sinc interpolation) and a temporal high-pass filtering using a cut-off frequency of 2 sine/cosine cycles. Functional images were aligned to the T1-weighted structural image and incorporated into the 3D datasets through trilinear interpolation. Data was not spatially smoothed to maintain maximal sensitivity of the selected voxels to their selection criteria. Head motions during the scan was inspected during the analysis to verify lack of excessive and abrupt head motions (> 1 mm).
Pre-processing of the resting-state scan data. The pre-processing included removal of the first two functional images of each run series to allow stabilization of the BOLD signal; correction for slice scan time acquisition (ascending-interleaved, using a cubic-spline interpolation algorithm); a trilinear interpolation approach in order to remove head motions; a high-pass (GLM-Fourier) frequency filter with a cut-off value of 2 sine/cosine cycles and a low-pass Gaussian-Full Width at Half Maximum (FWHM) of 1.9 data points 41 .
Further, in the ROI analysis we further used an 8th-order Butterworth filter with cut-off frequencies of 0.009 < f > 0.08 Hz 41 before projecting out averaged signals of the white matter, the cerebro-spinal fluid, and head motion parameters (containing 6 regressors: translations and rotations in the x, y, and z dimensions). This step was conducted by running a GLM regression analysis. The residuals of this analysis were free from the unwanted components and were used as inputs in the resting-state functional connectivity (FC) ROI analysis. Functional connectivity was computed using a correlation analysis (Pearson correlation coefficient) between the time courses of pre-defined ROIs (M1, cerebellum, thalamus, putamen, superior frontal, and superior parietal).
Both the fMRI-task and Rs-fMRI data sets of each participant were spatially aligned onto the corresponding anatomical scan (T1 weighted structural scan) using an automatic alignment procedure (implemented in Brain Voyager 20.6). The results of the automatic alignment were inspected during processing and manually adjusted when necessary. Subsequently, the co-aligned images were transformed into Talairach space 42 .
Definition of regions of interest. 12  www.nature.com/scientificreports/ the localizer scan (contralateral ankle movement vs baseline, p < 0.05, cluster size > 1000). The coordinates of each ROI were selected based on activation peaks of the above contrasts and were obtained from the localizer scan. The thalamus, putamen, superior frontal, and superior parietal were anatomically defined for each participant using FreeSurfer V 5.0. The size of each ROI was defined as the number of functional voxels (3 mm isovoxel). [Detailed characteristics (mean, SEM) of each ROI are presented in Table 1].

Independent component analysis (ICA).
Resting-state data was analysed using ICA in Brain Voyager 20.6. This method allows the detection of a set of statistically independent spatial maps (networks) on a subjectby-subject basis during the resting-state scans and subsequently measures changes in the strength of these spatial maps after the intervention 43 . The analysis was composed of two stages. In the first, 30 ICA components were detected automatically in each scan based on the individual resting-state scans. In the second, consistent components within and across participants were automatically selected using the "self-organizing groups" feature in Brain Voyager that forms clusters (brain networks) across the brain according to similarity of their spatiotemporal structures 44 . The analysis was restricted to the sensorimotor network and the cerebellar networks that were shown to be sensitive to balance training 31 .
Processing of the anatomical data for the volume analysis. 3D anatomical scans were analysed using Free-Surfer software 5.0 45 .
Statistical analysis. for statistical calculations, we used SPSS statistics (SPSS for Windows, Version 16.0; SPSS Inc., Chicago). The significance level was set to p < 0.05 for all statistical tests. Normality assumption was tested by using the Kolmogorov-Smirnova test (p > 0.05). Paired t-test was used for within-subject analyses of the behavioral data (pre-intervention vs. post-intervention). The ES (Cohen's d) for the within-subject design was calculated by dividing the mean difference between pre-and post-intervention by the pooled SD. For the resting state functional connectivity analysis, we used z-Fisher transformation in order to normalize the distribution of the correlation coefficients. The associations between functional connectivity measures, brain volume parameters and functional behavioral measures were assessed using multivariate linear regression models. To be specific, we ran regression models between the behavioral parameters: baseline CB&M, ΔCB&M, baseline 10MWT or Δ10MWT, and baseline FC or ΔFC which were computed between the ROIs. Post-hoc analysis was conducted on the significant regression models. The post-hoc analyses were not corrected for multiple comparisons. Whole-brain ICA was corrected for multiple comparison using a cluster-size correction for family-wise error rate at p < 0.05. The association between brain volume and dynamic balance was assessed using Pearson's correlations.

Results
Data from five participants from the ABI-group were excluded from the analysis for the following reasons: 1) One participant refused to participate in the MRI scan. 2) Four participants refused to continue with the intervention study due to loss of interest and transportation difficulties. The total number of participants included in the imaging analysis was 31 (11 = TBI and 20 = stroke). Additionally, one participant out of the 31 did not undergo the behavioural assessment post-intervention due to loss of interest.
The average age of the participants who participated in the study was 60.16 ± 12.85 (SD) years, with an average time since injury of 92.74 ± 144.43 (SD) months. The time interval between pre-and post-assessments was 117 ± 14.14 (SD) days (see Table 2 for additional information).

Recovery of dynamic balance and gait velocity following intervention. All 31 participants who
were included in the imaging analysis completed the 22 sessions of training using the mechatronic shoes 29 . Recovery was associated with a signal reduction in sensorimotor and the cerebellar networks. We first examined global changes in connectivity using an ICA approach. We focused our analysis on two resting-state spatial maps of interest that were shown to be involved in gait control 31 : the sensorimotor and the cerebellar spatial maps (Fig. 1A). For each network, we examined the strength of these networks at the voxel's level using contrasts. This analysis revealed a reduced component-related activation in both networks (p < 0.05, cluster-size correction) (Fig. 1B). Increases in component-related activation following training were not found in both networks (p > 0.05).
Recovery of dynamic balance was associated with reduced FC within the cerebellar network. Motivated by the ICA analyses, we aimed to better localize the networks that are associated with dynamic balance impairments and recovery. The first analysis focused on inter-hemispheric functional connectivity (IHFC) in the cerebellar and sensorimotor networks. Functional connectivity within these connections were not significantly modulated by the intervention. When searching for associations between connectivity and impairment and recovery, we found a significant association between ΔFC in the cerebellum and ΔCB&M [F (4, 25) = 3.41; β = − 0.57, p = 0.02 (uncorrected), R 2 = 0.35] (Table 3). www.nature.com/scientificreports/ Baseline resting-state Intra-hemispheric FC (IntraHFC) at the cerebellar-cortical and cerebellar-subcortical networks predicts dynamic balance recovery. Next, we examined the association between network connectivity within each hemisphere (IntraHFC) and dynamic balance and recovery. Here again, connectivity measures were not affected by the intervention. Regression models revealed the following: (1) Baseline IntraHFC in the cerebellar-cortical and cerebellar-subcortical networks was associated with ΔCB&M [F (10, 19) = 2.27; p = 0.02, R 2 = 0.58 (Table 4)]. Covariates contributing to the significant model were the left cerebellar-right thalamic network (β = − 0.51, p = 0.02, uncorrected) and left cerebellar-right putamen network (β = 0.62, p = 0.002, uncorrected).
Covariates contributing to the significant model were:   www.nature.com/scientificreports/ To study the association between brain volume and potential recovery, we ran the same models but this time with ∆CB&M as the dependent variable. This model was not significant as well [F (4,25) = 0.55, p = 0.69], suggesting that there is no linear dependency between the examined global structural brain variables and dynamic balance impairment and recovery. Lastly, in search of a possible degenerative mechanism for post-ABI dynamic balance impairment 2 , we examined the association between total gray matter ( Fig. 2A) and cortical white matter (left and right hemispheres) (Fig. 2B,C) and time since the brain injury. The results revealed a significant association beween time since injury and these global ROIs (r = − 0.4 p = 0.02, r = − 0.54 p = 0.001, r = − 0.46 p = 0.008, respectively).

Discussion
This study aimed to explore the neural substrates of dynamic balance by examining the association between dynamic balance impairments and recovery in chronic ABI participants and functional and structural brain measurements. The rationale for combining two subgroups of ABI (TBI and stroke) in the study was that both subgroups suffer from balance impairments due to a neuronal damage. Furthermore, both etiologies are characterized by initial neuronal loss, neurophysiological changes and disruptions of integration in structural and functional networks (diaschisis) 2 , resulting in sensorimotor impairments and loss of motor functions 2 . Lastly, in both groups, recovery is mediated by neural plasticity in intact cortical and sub-cortical regions 2 .
We report that dynamic balance recovery was associated with a reduction in connectivity in the sensorimotor and cerebellar networks. Furthermore, dynamic balance recovery was negatively associated with baseline connectivity within the cerebellar-thalamic network and with baseline connectivity in the cerebellar-M1 network. Dynamic balance recovery was also positively associated with baseline connectivity within the cerebellarputamen network and both the cerebellar-frontal and cerebellar-parietal networks. We also found that while morphological features were not correlated with dynamic balance impairment and recovery, they were associated with the time that passed since injury. www.nature.com/scientificreports/

Network reduction as a biomarker of impairment post-ABI and of intervention-induced plasticity.
Using the ICA approach, we found a reduction in component-related activation in the sensorimotor and cerebellar networks post-training. Furthermore, the results of the regression analysis showed that reduction in FC within the cerebellar network post-training and low baseline FC within the cerebellar-cortical and cerebellar-putamen were associated with better dynamic balance recovery. Therefore, we suggest that FC within the cerebellar-cortical and cerebellar-subcortical networks can provide insight into the neural substrates and mechanisms that support the recovery of dynamic balance. Here we propose that the reduced component-related activation and FC in sub-networks can be considered as a manifestation of increased network modularity (where the modules of the networks are the ROIs) 46 , which was described as the neural basis of complex behaviours in health 32,46 , and in disease 47,48 and was shown to be associated with more flexible and adaptable behaviour, which is needed for benefit from training 32 . Future experiments should test this conjecture directly.
The role of the cerebellar-cortical and cerebellar-subcortical in dynamic balance control. The resting state FC analysis revealed a correlation between dynamic balance recovery and the cerebellum, putamen, and thalamus. These inferred connections are consistent with the increasing evidence of the existence of subcortical loops that reciprocally connect the cerebellum with the putamen through the thalamus 49,50 , and functional interconnection between the cerebellum and the putamen 31 , the role of the putamen in gait and gait kinematics was demonstrated in studies in healthy 31,51 , and stroke participants 21,52 . Lastly, the resting-state FC analysis revealed a correlation between cerebro-cortical areas and dynamic balance recovery which is also consistent with previous studies highlighting the involvement of the cerebellum and the cerebral cortices in gait and balance control in healthy subjects 31,53 , and in those with post-brain injury 54,55 . Our results highlight the potential contribution of these cortical and subcortical brain areas to dynamic balance and gait-related recovery following brain injury.
Chronicity of ABI and brain volume. Our findings exhibit a negative correlation between global brain volume reduction and time since injury, indicating a diffused atrophy that progresses with time. These findings are consistent with previous studies which showed that the spatial pattern of TBI-related atrophy affects multiple grey and white matter areas 35,56 and also are consistent with a previous longitudinal study demonstrating that enhanced reductions in total brain volume with time reduced white matter integrity post-ABI 56 . Importantly, we did not find a direct association between brain atrophy and dynamic balance. The lack of association can be explained by the involvement of multiple brain areas in dynamic balance control and the variability among subjects in terms of damage and/or recovery mechanisms. Thus, there might be an indirect association between brain atrophy and dynamic balance which is manifested by the alteration in functional networks, as has been www.nature.com/scientificreports/ reported previously in cognitive impairments post-stroke 57 . Furthermore, since atrophy is a process associated with time, conducting longitudinal assessments over time will provide a better understanding of the relationships between atrophy and dynamic balance.
The integration of structural and functional brain measures (the multimodal approach). Brain functional connectivity and structural morphology are clearly not independent. Two emerging perspectives in the neuroimaging literature attempt to explain the association between structural damage and network function. Firstly, studies indicate a negative association between the extent of the brain damage (post-stroke/TBI/ Tumours) and FC depicting a depression of neural activity in brain regions remote from the initial site of brain damage due to reduced FC (diaschisis) 3,58 . Secondly, studies report an increase in FC within brain regions that had reduced structural connectivity, which may represent the reorganization of the system following the brain damage 57,59 . Future studies should adapt a multi-modal approach of brain network assessments post brain damage, as this approach can open new perspectives into the consequences of brain damage and its impact on impairment and recovery.
Several limitations of this study should be acknowledged: (1) Resting-state data was collected only for post-ABI with no control reference. This limitation confines the discussion about abnormal patterns of connectivity but does not affect the interpretation of the longitudinal data. (2) The absence of follow-up assessments limits the estimation of the efficacy of the treatment and its longterm neural outcomes. (3) There is an increase in the risk of type-1 bias due to multiple comparisons, especially in the post-hoc analysis of the regression models that were not corrected for multiple comparisons. We suggest taking these results as preliminary exploratory evidence and call for their replication. (4) While all subjects suffered from dynamic balance impairment due to ABI, the different etiologies (TBI and stroke) increased the inter-subject variability with respect to the location of damage. (5) We could not reproduce the ICA results with the ROI analysis. This might be due to the different methodologies and the procedure that were adopted in order to define the ROIs.
In conclusion, functional and structural mapping of brain networks reveals widespread alterations of networks following brain injury and rehabilitation. We suggest that these alterations were more likely to have resulted from the training than from the chronic time post-injury, since spontaneous changes are less prominent at the chronic phase post brain injury. Our study demonstrates functional connectivity changes in the cerebellar and sensorimotor networks following a rehabilitation treatment for post-ABI participants and suggests neural markers for the treatment's gains. The lack of inter-subject correlations with structural atrophy suggests that dynamic balance is an emergent feature of functional and composite networks. Our results contribute to the understanding of the neural correlates of dynamic balance and depict several markers of recovery that should be further investigated.