Cerebral network disorders after stroke: evidence from imaging-based connectivity analyses of active and resting brain states in humans

Abstract

Stroke causes a sudden disruption of physiological brain function which leads to impairments of functional brain networks involved in voluntary movements. In some cases, the brain has the intrinsic capacity to reorganize itself, thereby compensating for the disruption of motor networks. In humans, such reorganization can be investigated in vivo using neuroimaging. Recent developments in connectivity analyses based on functional neuroimaging data have provided new insights into the network pathophysiology underlying neurological symptoms. Here we review recent neuroimaging studies using functional resting-state correlations, effective connectivity models or graph theoretical analyses to investigate changes in neural motor networks and recovery after stroke. The data demonstrate that network disturbances after stroke occur not only in the vicinity of the lesion but also between remote cortical areas in the affected and unaffected hemisphere. The reorganization of motor networks encompasses a restoration of interhemispheric functional coherence in the resting state, particularly between the primary motor cortices. Furthermore, reorganized neural networks feature strong excitatory interactions between fronto-parietal areas and primary motor cortex in the affected hemisphere, suggesting that greater top-down control over primary motor areas facilitates motor execution in the lesioned brain. In addition, there is evidence that motor recovery is accompanied by a more random network topology with reduced local information processing. In conclusion, Stroke induces changes in functional and effective connectivity within and across hemispheres which relate to motor impairments and recovery thereof. Connectivity analyses may hence provide new insights into the pathophysiology underlying neurological deficits and may be further used to develop novel, neurobiologically informed treatment strategies.

Anne K. Rehme studied psychology at the University of Trier/Germany. Her PhD thesis focused on the spontaneous reorganisation of the cortical motor network in acute stroke patients. She currently holds a post-doc position at the Max Planck Institute for Neurological Research in Cologne, Germany, where she continues to work on methodological aspects of fMRI-based connectivity in healthy and clinical populations. Christian Grefkes is a clinical neurologist with special interests in motor neuroscience and stroke recovery. He qualified as an MD after graduating from the University of Duesseldorf, Germany, in 2004 with clinical semesters at the University of Sydney (Australia) and UCL. His doctoral thesis on structural and functional properties of the human somatosensory system was supervised by Professor Karl Zilles. He worked as post-doctoral fellow with Professor Gereon Fink at the Institute of Neurosciences and Medicine at the Juelich Research Centre. He received his training in clinical neurology at the neurology departments of the universities in Aachen and Cologne. In 2007, Christian Grefkes was appointed by the Max Planck Society as head of the Neuromodulation & Neurorehabilitation research group at the Max Planck Institute for Neurological Research in Cologne.

Introduction

Voluntary movements depend on the well-tuned interplay of excitatory and inhibitory influences between neurons in primary motor cortex (M1), lateral premotor cortex (PMC) and supplementary motor area (SMA), as well as in subcortical areas such as basal ganglia, thalamus, cerebellum and brainstem nuclei (Middleton & Strick, 2000; Dum & Strick, 2002). In addition, fronto-parietal areas involved in somatosensory, visuospatial and executive processing are closely linked to this ‘primary motor network’ (Dum & Strick, 2005). Focal stroke lesions affecting either cortical and subcortical neurons or descending fibre tracts may critically disturb the neural processing in this network (Ward et al. 2006; Stinear et al. 2007). Stroke commonly results from ischaemia, that is, the lack of blood supply to cerebral tissue caused by (thrombo-)embolic occlusion of a cerebral artery. In the United States and Europe, brain lesions after stroke are a common cause for permanent disability (Taylor et al. 1996; Carandang et al. 2006; Heuschmann et al. 2011). However, particularly patients with initially mild to moderate motor impairments may show substantial motor recovery within the first 3 months of stroke onset (Duncan et al. 1994; Kwakkel et al. 2004, 2006). This early recovery of motor functions is mainly driven by neural plasticity and reorganization (Nudo, 2006; Cramer, 2008). Such processes may affect brain organization at the level of single neurons (microscale), neuronal populations (mesoscale) or at the level of interactions between different regions (macroscale) (Sporns et al. 2005).

Animal models of stroke provided evidence for a complex cascade of events enabling changes in structural connections and synaptic transmission. These changes occur not only in the vicinity of the lesion but also in remote brain regions (Carmichael, 2006). For example, tract-tracing studies in rats demonstrated that experimental lesions to primary sensory cortex (S1) are followed by axonal sprouting from cell bodies adjacent to the lesion into peri-infarct cortex (Carmichael et al. 2001). Likewise, new projections arise from S1 to ventral PMC after lesions to M1 (Dancause et al. 2005). In addition, sensorimotor cortex lesions are associated with a formation of new connections extending from the opposite, unaffected cortex to ipsilesional striatum and to peri-infarct cortex (Napieralski et al. 1996; Carmichael & Chesselet, 2002). Likewise, immunohistochemical and autoradiographic studies provided evidence for a downregulation of inhibitory γ-amino-butyric acid (GABA) receptors (Schiene et al. 1996; Redecker et al. 2002) and an increase in excitatory N-methyl-d-aspartate (NMDA) receptor binding sites in both hemispheres (Que et al. 1999). These processes, which are detectable within the first few hours after stroke and are most pronounced after 3–7 days, may facilitate the ‘rewiring’ of neurons and, hence, cortical reorganization (Nudo, 2006).

Neuroimaging – a non-invasive approach to brain organization

Obviously, invasive investigations of cortical reorganization on the microscale level are not feasible in humans. In contrast, neuroimaging techniques such as positron-emission tomography (PET) and functional magnetic resonance imaging (fMRI) represent non-invasive approaches to measure neural activity in humans in vivo. As these techniques work with spatial resolutions of a few millimetres, they enable the investigation of reorganization processes at the macroscale level (Sporns et al. 2005). PET measures changes in local perfusion and glucose metabolism by mapping the distribution of radioactive, positron-emitting tracers in brain tissue. Perfusion (i.e. regional cerebral blood flow, rCBF) is typically measured by injection of water or alcohols such as butanol labelled with the oxygen isotope ¹⁵O (Frackowiak & Friston, 1994). The metabolism of glucose – an important substrate for neuronal activity – can be mapped using ¹⁸F fluorodeoxyglucose (FDG) PET (Phelps et al. 1981). In contrast to PET, fMRI is free of ionizing radiation. This brain-mapping method is based on the blood-oxygenation-level-dependent (BOLD) signal, which arises from magnetic field inhomogeneities caused by the different magnetic properties of oxy- and deoxyhaemoglobin (Ogawa et al. 1990). Neuronal activity triggers several metabolic events leading to an increase in blood volume and a subsequent decrease of the deoxyhaemoglobin content of the activated tissue. These changes reduce local magnetic inhomogeneities followed by increases in the BOLD signal (Heeger & Ress, 2002). Neurophysiological studies showed that rCBF and BOLD signals are both related to local field potentials (Logothetis, 2002; Lauritzen & Gold, 2003).

Changes in brain activity after stroke

Functional neuroimaging studies with stroke patients frequently reported enhanced activity in a number of areas such as M1, PMC, SMA, parietal and prefrontal cortex, striatum, thalamus and cerebellum in both the affected and unaffected hemisphere during movements of the paretic limb (Chollet et al. 1991; Weiller et al. 1992; Ward et al. 2003; Gerloff et al. 2006). A recent meta-analysis revealed that particularly activation of premotor areas and contralesional M1 is a consistent finding across multiple experiments (Rehme et al. 2012) (Fig. 1). However, the functional meaning of this ‘over-activity’ for motor recovery is still a matter of debate. All these areas are crucially involved in motor planning and motor control of voluntary movements (Jenkins et al. 2000; Schubotz & von Cramon, 2003; Hoshi & Tanji, 2004, 2007). Reaching and grasping movements in particular depend on the activation of posterior parietal areas during sensorimotor integration and visuomotor control (Bodegard et al. 2003; Grefkes & Fink, 2005; Culham et al. 2006; Eickhoff et al. 2008). Dorsolateral prefrontal cortex has been reported to be engaged in action selection, motor learning and attentional control of behaviourally relevant stimuli (Ridderinkhof et al. 2004). Moreover, tract-tracing studies revealed dense intra- and interhemispheric cortico-cortical projections between PMC, SMA, posterior parietal areas and M1 that might facilitate the motor output to spinal cord neurons (Strick & Kim, 1978; Luppino et al. 1993; Tokuno & Tanji, 1993; Rouiller et al. 1994; Johnson et al. 1996).

Figure 1. Activation likelihood estimation (ALE) meta-analysis of motor-related neural activity after stroke.

The activation likelihood for affected upper limb movements as compared with rest is depicted in blue (based on 452 activation maxima from 54 experiments in a total sample of 472 stroke patients). Enhanced activation likelihood for movements of patients as compared with healthy subjects is depicted in yellow (based on 113 activation maxima from 20 experiments of 177 patients). Movements of the affected upper limb are associated with significant local convergence in primary sensorimotor cortices, lateral PMC, SMA, pre-SMA, parietal operculum and cerebellum of both hemispheres (P < 0.05, cluster-level familywise error (FWE) corrected for multiple comparisons). Compared with healthy controls, activation likelihood is specifically enhanced in contralesional sensorimotor cortex as well as in ventral PMC and SMA of both hemispheres. Hence, neural activity in these areas distinguishes patients most consistently from healthy subjects (adopted from Rehme et al. 2012, with permission).

Importantly, longitudinal fMRI experiments showed that bilateral activity of these areas increases already within the first days and weeks after stroke concomitant with early motor recovery, particularly in patients with severe impairments (Rehme et al. 2011b). However, this over-activity typically returns to physiological levels after 6–12 months in well-recovered patients (Calautti et al. 2001; Loubinoux et al. 2003; Ward et al. 2003). Consequently, patients with persistent over-activity at chronic stages, particularly in contralesional motor areas, are often those with greater motor deficits (Ward et al. 2004; Loubinoux et al. 2007; Marshall et al. 2009), and with greater lesions of the corticospinal tract (Ward et al. 2007; Stinear et al. 2007; Schaechter et al. 2008). In addition, severely impaired patients with more bilateral activity patterns show a reduced integrity of transcallosal fibre tracts between motor areas that might result from a secondary degeneration of fibres connected to the lesion zone (Wang et al. 2012). Some authors suggested that enhanced contralesional activity in chronic stroke patients supports motor functions of the paretic hand (Lotze et al. 2006). This interpretation is, however, challenged by the finding that suppressing activity in the unaffected hemisphere by means of repetitive transcranial magnetic stimulation (rTMS) or transcranial direct current stimulation improves motor performance of the paretic hand (Fregni et al. 2005; Takeuchi et al. 2005; Nowak et al. 2008; Dafotakis et al. 2008). This evidence stimulated the debate that the unaffected hemisphere might be rather disturbing for motor performance after stroke.

In summary, neuroimaging data suggest that recovery of motor function after stroke leads to brain-wide changes in neural activity. However, the mere localization of activity does not explain how regional activations affect neural processing in the entire network. This question is important as even simple forms of behaviour rely on interactions of a number of brain regions forming a network (Friston, 2002). Furthermore, as outlined above, stroke lesions trigger several brain-wide processes to accommodate for tissue loss. Novel approaches in computational neuroscience enable the estimation of functional interactions between brain regions from functional neuroimaging data. Such connectivity models allow consideration of reorganization after stroke from a network perspective which seems to be closer to the neurobiology underlying recovery of function. In this review, we first provide an overview of different neuroimaging-based connectivity approaches. Then, we present recent data showing how stroke affects functional and effective connectivity in the motor system. We finally discuss these findings in a broader context and summarize how connectivity analyses may further our understanding of lesion-induced motor deficits and recovery thereof.

Brain connectivity

Connectivity models are based on the concept that the brain is organized by segregation of specialized and anatomically distinct brain regions that are functionally integrated in networks mediating cognitive, sensory or motor processing (Friston, 2002). Here, structural connectivity describes how spatially separated brain regions are physically linked, for example, as demonstrated by invasive tracing of single axons or by measuring diffusion along major fibre bundles non-invasively as in diffusion tensor imaging (DTI). In contrast, functional and effective connectivity describe how anatomically connected areas interact with each other. Both approaches, however, fundamentally differ in the way how these interactions are estimated. Functional connectivity is defined as temporal correlation between spatially remote neurophysiological events. In contrast to this non-directional, correlative nature of functional connectivity, effective connectivity refers to the causal influences that brain areas exert over another under the assumptions of a given mechanistic model (Friston, 1994; Stephan et al. 2007). Network configurations as defined by different connectivity measures can further be quantified by means of graph theory (Bullmore & Sporns, 2009).

Functional connectivity

Functional connectivity is usually computed from low-frequency (<0.1 Hz) resting-state fMRI data, that is, during wakeful rest, but in the absence of active task performance. Statistical analyses of resting-state fMRI either consist of seed-voxel correlations (Horwitz et al. 1998) or multivariate approaches including independent or principal component analyses (ICA or PCA) (Friston et al. 1993; Fox & Raichle, 2007). A seed-voxel analysis uses one or more pre-defined regions of interest (‘seeds’) to identify those voxels showing correlated fMRI signal time-courses with the respective seed region (Fig. 2). PCA and ICA are data-driven approaches to delineate spatially independent patterns of coherent signals. Both multivariate and seed-based approaches provided evidence for separable patterns of functional connectivity between areas with similar functional properties and known anatomical connections (Damoiseaux et al. 2006). For example, the resting-state BOLD time-series of a seed region in M1 shows correlated activity with M1 and premotor areas in both hemispheres despite any overt motor performance (Biswal et al. 1995). Resting-state functional connectivity has a modest to good test–retest reliability which is important for longitudinal study designs (Shehzad et al. 2009; Van Dijk et al. 2010).

Figure 2. Resting-state functional connectivity.

A, seed-voxel analysis in a group of 16 right-handed healthy volunteers (P < 0.05 false discovery rate (FDR) corrected). The colour scales indicate the correlation between a seed voxel located in left or right primary motor cortex (M1) and any other voxel time-course in the brain. B, Pearson correlation coefficients between BOLD signal time-courses from seed voxels of homologous motor areas in both hemispheres in 16 healthy subjects. Coordinates are reported in Montreal Neurological Institute (MNI) reference space. M1, primary motor cortex; PMC, premotor cortex; SMA, supplementary motor area.

Effective connectivity

Effective connectivity is based on mathematical models that are usually applied to task-based neuroimaging data, including psychophysiological interactions (Friston et al. 1997), structural equation modelling (SEM) (McIntosh & Gonzalez-Lima, 1994), Granger causality (Roebroeck et al. 2005) or dynamic causal modelling (DCM) (Friston et al. 2003). To date, most studies on effective connectivity of the motor system after stroke have used either SEM or DCM. Both are hypothesis-driven approaches that estimate interactions in a pre-defined network of brain regions based on anatomically motivated hypotheses about their connections (Penny et al. 2004). However, both methods differ in their underlying generative models and, therefore, in the estimation of connectivity parameters. SEM was first used in econometrics and social sciences and later implemented in the analysis of neuroimaging data (McIntosh & Gonzalez-Lima, 1991). SEM is based on the translation of a network model with pre-defined regions linked by a set of directional paths into a linear regression model. The path coefficients are subsequently estimated using an iterative maximum-likelihood algorithm to minimize the difference between observed and predicted covariance matrices (McIntosh & Gonzalez-Lima, 1994; Penny et al. 2004). While classical implementations of SEM assume that imaging data are stationary, recent developments such as extended unified SEM also enable the modelling of event-related inputs on regional activity (Gates et al. 2011).

In contrast, DCM – in its original form – relies on a deterministic model that treats the brain as an input–output system of hidden neural dynamics (Friston et al. 2003). This neural model describes changes in the system over time as a function of interactions between regional activity, known experimental inputs and neuronal parameters. DCM estimates three sets of neuronal parameters including (i) endo-genous coupling, (ii) condition-dependent changes, and (iii) direct experimental input (Fig. 3). DCM for fMRI uses an experimentally validated haemodynamic model (Buxton et al. 1998; Friston et al. 2000) to translate the modelled neural dynamics into a predicted BOLD response (Stephan et al. 2007). Both the neuronal and the haemodynamic parameters are estimated from the measured BOLD data using an iterative Bayesian algorithm to optimize an approximation of both the model evidence (i.e. likelihood of the model given the data) and the posterior density (i.e. likelihood of the data given the parameters for a particular model) (Friston et al. 2003). Statistical inference is drawn from the maximum a posteriori estimates and posterior covariances of the posterior density function. A recent study reported moderate to excellent test–retest reliability of DCM coupling parameters (Schuyler et al. 2010).

Figure 3. Effective connectivity as estimated using dynamic causal modelling (DCM).

A, top panel, example of a network model including three areas and a priori assumptions about their connectivity. Bottom panel, bilinear neuronal state equation which describes changes in the system over time based on: (1) endogenous coupling, which is computed independently from experimental conditions (A-matrix, grey); (2) context-dependent coupling (B-matrix, blue); and (3) direct experimental input to the system (C-matrix, red). B, context-independent (endogenous) and context-dependent connectivity between key areas of the cortical motor system for rhythmic fist closures of the left or right hand. Coupling parameters reflect the strength and direction of coupling between regions and are measured as rate of change (Hz). Positive coupling (green arrows) indicates promoting influences, whereas negative coupling (red arrows) indicates inhibitory influences on the activity in the target region (adopted from Grefkes et al. 2008a, with permission).

Graph theory

Graph theoretical analyses can be applied to structural, functional and effective connectivity matrices to quantify the topology of a network by means of different indices (Rubinov & Sporns, 2010). In graph theory, brain regions are referred to as nodes that are linked by edges (connections) (Sporns et al. 2005). All hitherto published studies on motor network topology after stroke focused on the clustering coefficient and the shortest path length to describe local and global network changes (Fig. 4). The clustering coefficient is based on the number of connections between the nearest neighbours of a given node relative to all possible connections and represents the efficiency of local information transfer. The shortest path length refers to the minimum number of connections that have to be traversed to get from one node to another. Thus, shorter paths within a given network imply a stronger potential for global information transfer (Rubinov & Sporns, 2010). In addition, graph theoretical analyses of brain networks after stroke used measures of centrality to describe the efficiency of information integration within a network. These measures encompass the node degree (i.e. the number of edges connected to a given node) and the betweenness centrality (i.e. the fraction of shortest paths that pass through a given node). Brain regions featuring a high node degree and high centrality are assumed to serve as ‘hubs’ mediating functional integration between regions or clusters of regions (Bullmore & Sporns, 2009). One key principle governing physiological brain function is economical information exchange which is achieved in networks with efficient parallel information transfer while maintaining relatively low wiring cost (Achard & Bullmore, 2007). This seems to be the case when networks display a ‘small-world’ topology characterized by a high local clustering and short paths that globally link all nodes of a network (Achard et al. 2006) (Fig. 4). In contrast, random networks are characterized by low local clustering and short paths, reflecting efficient information transfer, but limited local information processing (Sporns et al. 2007).

Figure 4. Examples of different graph network topologies based on a model with 16 nodes and 32 edges.

The average shortest path length L refers to the minimum number of edges that have to be traversed to get from one node to another. Hence, shorter paths (i.e. a few number of edges indicated by a low shortest path length L) represent a greater ability for global information exchange. The clustering coefficient C represents an index for the number of edges between nearest neighbours of a given node relative to the maximum number of possible connections (Rubinov & Sporns, 2010). A, regular network, which is highly clustered or ‘cliquish’ and in which high number of edges has to be traversed to get from one node to another. Thus, regular networks feature a high efficiency of local information processing at each node, but less global information transfer. B, random network with low local clustering, but short paths between nodes. This topology provides a high global efficiency, but reduced local information processing. C, small-world network representing an intermediate network configuration with some highly clustered nodes and a medium number of paths to enable both local and global information processing.

Comparison of connectivity models

Functional and effective connectivity models for neuroimaging data differ in several ways. Resting-state functional connectivity has been assumed to measure intrinsic brain connectivity which is independent from task-specific activations. Thus, resting-state fMRI can be used in an exploratory fashion and enables simultaneous testing of several functional systems (‘resting-state networks’). Resting-state analyses are not biased regarding task performance which is particularly useful for the investigation of patients with different degrees of neurological impairment. In contrast to the inherently non-directional nature of resting-state functional connectivity, analyses of effective connectivity provide a description of directional influences between areas. However, effective connectivity approaches like SEM and DCM require a priori defined network structures to estimate connectivity from neuroimaging data. Hence, these approaches are useful to test a particular hypothesis or neuropsychological model.

Findings from recent studies suggest that disturbed haemodynamic signals after stroke may lead to an underestimation of brain activity (Bonakdarpour et al. 2007; Mazzetto-Betti et al. 2010). However, as DCM estimates individually fitted haemodynamic parameters separately for each area, this approach appears to be especially suited for stroke patients with abnormal neurovascular coupling. In addition to neuroimaging-based connectivity analyses, methods like transcranial magnetic stimulation (TMS) can be used to analyse how excitation of one brain region interferes with the excitability of another region, thereby enabling the investigation of causality from a different methodological perspective (for a review see Shafi et al. 2012). Alternatively, functional and effective connectivity can be computed from electroencephalography (EEG) or magnetencephalography (MEG) data (Dubovik et al. 2012; Westlake et al. 2012), which account for the fact that neural connectivity may occur at faster time scales than measured by neuroimaging techniques. However, EEG signal coherence within and between hemispheres shows a good correspondence with functional connectivity of resting-state fMRI data in a rat stroke model (van Meer et al. 2012), implying that both methods assess a similar neuronal correlate driving these interactions. In addition, the high spatial resolution of neuroimaging data is an important advantage for the localization of interactions between well-defined anatomical regions.

Resting-state functional connectivity after stroke

Findings from resting-state functional connectivity approaches can be summarized by two major patterns of changes after stroke: (1) Reduced interhemispheric functional connectivity between cortical motor areas, which correlates with the severity of motor deficits; and (2) Reduced global network efficiency even in patients with good clinical recovery (Fig. 5A).

Figure 5. Synopsis of changes in motor networks after stroke.

The figure summarizes those areas which were included in network models of functional and effective connectivity studies: cerebellum (Cereb), primary motor cortex (M1), prefrontal cortex (PFC), lateral premotor cortex (PMC), supplementary motor area (SMA), superior parietal cortex (SPC) and thalamus (Thal). The numbers displayed on connections refer to the respective publication which reported connectivity disturbances. Correlations with motor impairment are marked in green. A, resting-state functional connectivity after stroke. The figure shows connections which either positively correlate with motor impairment or were related to disrupted anatomical connectivity. The most consistent finding pertains to reduced interhemispheric interactions, particularly between primary motor cortices, which correlate with more severe impairments. For reasons of clarity, performance-related disturbances in interhemispheric connections between cerebellum and SMA, as well as between ipsilesional thalamus and contralesional premotor areas reported in the study of Wang et al. (2010) are not depicted here. B, effective connectivity after stroke. The figure depicts changes in excitatory (left panel) and inhibitory (right panel) interactions relative to healthy subjects as revealed by DCM and SEM analyses of fMRI motor tasks and one resting-state fMRI study (Inman et al. 2012). For intervention studies (numbers 4 and 7), only findings from baseline assessments are presented. Strongest convergence across studies was observed for reduced positive influences between premotor areas and ipsilesional M1. Likewise, inhibitory interhemispheric influences between the primary motor cortices are often attenuated, which suggests that motor deficits after stroke are maintained by a disinhibition of contralesional M1 activity. At the subacute stage, there is evidence that contralesional M1 exerts a positive (i.e., supportive) influence onto ipsilesional M1 activity in patients with severe impairments. Together, there seems to be a time-dependent role of contralesional M1 in motor recovery after stroke.

Functional connectivity – regional changes

The most consistent finding from animal and human resting-state studies is that stroke leads to changes in interhemispheric functional connectivity particularly between homotopic areas in both hemispheres (Carter et al. 2010, 2012; Wang et al. 2010; Park et al. 2011). Van Meer et al. (2010) analysed seed-voxel correlations of resting-state fMRI data to investigate the functional connectivity of the primary sensorimotor cortex in rats recovering from experimental stroke. In these animals, interhemispheric resting-state connectivity between ipsilesional primary sensorimotor cortex and its contralesional homologue is significantly diminished in the first few days, and subsequently increases in the ensuing 70 days while sensorimotor functions recover. However, interhemispheric connectivity remains reduced compared with assessments obtained prior to stroke (van Meer et al. 2010). In line with these findings, resting-state studies in human stroke patients provided evidence for disturbed interhemispheric connectivity. Carter and colleagues (2010) computed seed-based correlations between resting-state fMRI time-series in a sensorimotor network consisting of M1, SMA, secondary somatosensory cortex, cerebellum, putamen and thalamus in both hemispheres. The authors reported that particularly interhemispheric M1 connectivity positively correlates with motor performance at the subacute stage after stroke. Thus, patients with relatively poor hand motor function feature reduced interhemispheric M1–M1 connectivity. Disturbed interhemispheric connectivity of ipsilesional M1 was also reported for connections to contralesional PMC and posterior parietal cortex (Wang et al. 2010). In addition, stronger interhemispheric connectivity of ipsilesional M1 and contralesional areas such as thalamus, SMA and middle prefrontal cortex within the first few days predicts better motor recovery in the next 6 months post-stroke (Park et al. 2011). Such phenomena were not exclusively reported for the motor system. Also patients with attention deficits (He et al. 2007; Carter et al. 2010) and aphasia (Warren et al. 2009) feature reduced interhemispheric connectivity between attention-related areas in parietal cortex or language areas in inferior frontal cortex. Importantly, there seems to be a tight structure–function relationship as, for example, motor impairments are not related to interhemispheric connectivity among attention-related areas (Carter et al. 2010). Rather, the integrity of corticospinal fibres correlates with disturbed interhemispheric M1 resting-state connectivity after stroke in both humans and rats (Carter et al. 2012; van Meer et al. 2012). Reduced interhemispheric resting-state connectivity also coincides with secondary myelin degeneration of transcallosal fibres between primary motor cortices of rats (van Meer et al. 2012). A similar structure–function relationship was found for the cortico-ponto-cerebellar tract by Lu et al. (2011) who reported reduced functional connectivity between M1 and contralateral cerebellum in patients with pontine lesions.

Altogether these findings imply that disturbances in resting-state connectivity result from both damage to cortical areas and remote disruptions of their fibre pathways. These disturbances, however, remain within the functional network affected by the lesion, and do not spread to other functional domains.

Functional connectivity – global changes

Stroke-induced changes in network efficiency can be quantified by means of graph theoretical analyses. For example, Honey & Sporns (2008) modelled oscillatory cortical interactions between areas based on anatomical connectivity data as established in macaques. The authors found that focal lesions of hub regions with long-range connections like parietal and frontal areas lead to greater and more widespread disturbances of cortico-cortical interactions than lesions of regions linking neighbouring areas. Similarly, Alstott et al. (2009) simulated the effect of cortical lesions on resting-state connectivity based on human structural connectivity data. In this model, the removal of parietal, frontal and midline hub regions with high regional centrality results in reduced global network efficiency and greater changes of resting-state connectivity patterns. These effects of simulated lesions are paralleled by graph theoretical analyses of resting-state fMRI data in stroke patients. For example, Wang et al. (2010) used graph theory to characterize the efficiency of information exchange in the motor network of severely impaired subcortical stroke patients. Here, the data showed that the regional centrality particularly of ipsilesional M1 and contralesional cerebellum is reduced within the first few days and increases during motor recovery across 1 year. However, these changes in the relative importance of a few central nodes are associated with decreased local clustering, that is, a less efficient information exchange between neighbouring regions. These findings implicate that the motor network gradually shifts towards a random configuration in patients with better recovery. Interestingly, the lower clustering in this reorganized network suggests that neural reorganization after stroke evokes a non-optimal network structure in terms of information exchange. Wang et al. (2010) hypothesized that axonal sprouting in intact brain regions leads to a number of new, but unspecific and, hence, random connections. From the perspective of graph theory, random networks are also characterized by shorter paths enabling rapid information transfer between distant nodes (Fig. 4). Here, a recent MEG study found that stroke patients with such a greater capacity for global information integration achieved better performance in a sensorimotor skill training (Buch et al. 2012). Thus, one hypothesis is that faster global information exchange may facilitate new functional network configurations (McIntosh et al. 2008; Deco et al. 2011; Beharelle et al. 2012).

Conclusion – stroke and resting-state connectivity

The studies reviewed above consistently showed that stroke-induced lesions cause an imbalance in synchronized spontaneous signals between sensorimotor areas of both hemispheres. Although resting-state connectivity was shown to be related to anatomical connectivity (Honey et al. 2007), significant resting-state connectivity has also been detected in the absence of direct anatomical connections (Damoiseaux & Greicius, 2009; Honey et al. 2009). However, stroke lesions do not globally reduce connectivity in all functional systems of the brain but specifically alter connectivity of areas connected to that lesion. As implied by graph analyses, such alterations typically affect the communication efficiency in a given functional network which is closely related to behavioural deficits after stroke. Accordingly, the damage of hub regions has the strongest impact on local and global information transfer. In addition, more random network architectures with less local but high global efficiency seem to promote the re-learning of sensorimotor skills but may also explain why performance is often less stable, even in well-recovered patients.

Effective connectivity after stroke

In contrast to functional connectivity, models of effective connectivity estimate the causal influences that one area exerts over another (Friston, 1994; Friston et al. 2003). Basically all published imaging studies on effective connectivity after stroke used either SEM or DCM. Two key findings may be summarized from effective connectivity studies: (1) Reduced promoting influences from fronto-parietal areas, particularly from premotor cortex on M1 activity in the affected hemisphere; and (2) Disturbed interhemispheric connectivity between primary motor cortices that depends on time since stroke and motor impairment (Fig. 5B).

Effective connectivity in the affected hemisphere

While functional resting-state correlations revealed a reduction of interhemispheric connections after stroke, the most consistent finding from effective connectivity analyses pertains to reduced intrahemispheric interactions in the ipsilesional hemisphere. Grefkes et al. (2008b) applied DCM to fMRI data which were recorded during rhythmic fist closures with the left or right hand to estimate effective connectivity between cortical motor areas. In healthy subjects, fist closures with the left or right hand are associated with positive couplings lateralized towards M1 in the hemisphere contralateral to the moving hand with strong influences from SMA and ventral PMC. At the same time, areas in the ipsilateral hemisphere, particularly ipsilateral M1, are inhibited. Stroke patients were demonstrated to show a number of pathological changes in the coupling of motor areas in both hemispheres. In the first days after stroke, patients suffering from motor impairments feature a strong reduction of excitatory influences from premotor regions onto ipsilesional M1 although the lesions do not directly affect these regions but rather corticospinal fibres in deep white matter (Rehme et al. 2011a). In the following 3–6 months, motor recovery in these patients was associated with an increase in coupling between premotor areas and ipsilesional M1. Such changes in inter-regional coupling might result from growth-related neurobiological processes enabling the formation of new synapses (Carmichael, 2006; Cramer, 2008).

In addition to disturbed interactions between ipsilesional premotor areas and M1, Sharma et al. (2009) reported reduced effective connectivity between ipsilesional SMA and dorsal PMC in a sample of well-recovered patients. Notably, there was no difference between patients and controls at the level of motor-related BOLD activity while effective connectivity among these areas as estimated with SEM was significantly altered. Hence, connectivity analyses seem to be more sensitive in detecting pathological changes than conventional analyses of regional activity. Sharma and colleagues also explicitly modelled influences from prefrontal cortex on SMA and dorsal PMC. The authors found significantly enhanced positive influences from prefrontal cortex to both premotor areas in patients relative to healthy subjects that positively correlate with motor performance. Another study (Inman et al. 2012) included inferior frontal and superior parietal areas in addition to M1, SMA and dorsal PMC into their network model to investigate effective connectivity of resting-state fMRI data. An SEM analysis showed that compared with healthy controls, path weights from superior parietal cortex to both M1 and SMA in the affected hemisphere were significantly reduced. However, this ‘hypo-connectivity’ was not directly related to poor motor performance, probably because effective connectivity was estimated from resting-state data whereas the other studies computed effective connectivity from task-based fMRI. In contrast, the crucial role of premotor–M1 connectivity for motor impairments and recovery thereof is also supported by intervention studies. For example, interference with contralesional M1 activity by means of rTMS results in increased ipsilesional SMA–M1 couplings and significant motor improvements (Grefkes et al. 2010). Similar effects on ipsilesional SMA–M1 couplings were observed when enhancing motor performance by means of noradrenergic stimulation in chronic stroke patients (Wang et al. 2011). Together, these data suggest that motor deficits after stroke are not only caused by direct disruption of descending motor pathways, but may also depend on a less effective communication between premotor areas and M1 in the lesioned hemisphere.

Disturbed effective connectivity between hemispheres

Similar to resting-state findings, disturbances in effective connectivity were also found between hemispheres, particularly for transcallosal M1–M1 interactions. Grefkes et al. (2008b) showed that compared with healthy subjects, stroke patients with relatively poor motor performance exhibit an enhanced inhibitory influence from contralesional to ipsilesional M1 during movements of the paretic hand. These findings from DCM analyses are supported by paired-pulse TMS studies which revealed abnormal levels of interhemispheric inhibition from contralesional on ipsilesional M1 during movement preparation (Murase et al. 2004; Duque et al. 2005). The hypothesis that this inhibition might contribute to the motor deficit of the patients is further substantiated by findings from intervention studies which demonstrated that reducing inhibitory influences from contralesional M1 via rTMS induces significant improvements in hand motor performance (Grefkes et al. 2010). However, rTMS effects may considerably vary across studies with some patients showing no response to contralesional M1 inhibition (Werhahn et al. 2003; Talelli et al. 2012) or even deteriorated performance (Lotze et al. 2006; Bradnam et al. 2011). One factor that seems to determine the functional role of contralesional M1 for motor performance of the stroke-affected hand is the time that has elapsed since stroke onset. A longitudinal DCM study with recovering stroke patients showed that interhemispheric inhibitory influences from ipsilesional motor areas to contralesional M1 are significantly diminished in the first few days (Rehme et al. 2011a). After 2 weeks, this apparent disinhibition of contralesional M1 is accompanied by a promoting influence from contralesional to ipsilesional M1, particularly in patients with severe motor deficits. Hence, in the subacute phase, contralesional M1 seems to support activity of motor areas in the lesioned hemisphere. However, after 3–6 months, this supportive influence may turn into inhibition in those patients with incomplete motor recovery, confirming the results of Grefkes et al. (2008b).

Summary – stroke and effective connectivity

The restoration of effective communication in the lesioned hemisphere seems to be an important factor driving motor recovery after stroke. Here, particularly interactions between premotor areas and M1 are directly related to recovery and final outcome. Also, interactions between fronto-parietal areas and M1 are altered after stroke, although the relationship with motor impairment is not always clear. The positive correlation of premotor–prefrontal cortex interactions and motor outcome suggests that enhanced effective connectivity supports motor function after stroke. The meaning of contralesional M1 activity for motor recovery after stroke remains a matter of debate. Here, connectivity analyses provided data that the functional role of contralesional M1 depends on time after stroke and the neurological deficit. There seems to be a supportive role in the first few weeks which may turn into inhibition in patients with persistent motor deficits. Interestingly, microstructural changes in neuronal wiring after experimental stroke seem to follow a very similar timeline. For example, studies in rats demonstrated that axonal sprouting from contralesional M1 into perilesional cortex is initiated in the first 3 days by a trigger phase of synchronized neuronal discharges which is followed by the initiation and maintenance of sprouting 7–14 days before newly formed anatomical connections can be detected at about 28 days post-stroke (Carmichael, 2003). However, it remains to be elucidated whether similar neurobiological mechanisms hold true for changes observed in fMRI connectivity.

Conclusion – network effects after stroke

In summary, functional resting-state and effective connectivity studies provide converging evidence that stroke patients with motor impairments feature changes in interhemispheric interactions and reduced top-down control of motor execution in the affected hemisphere (Fig. 5). Importantly, changes in cortico-cortical interactions were observed in samples of patients with predominantly subcortical lesions, which rules out that cortico-cortical dys-connectivity simply results from the damage of cortico-cortical fibres. Such changes in inter-regional interactions which also affect brain areas that are remote albeit connected to the primary site of damage have also been referred to as ‘diaschisis’ (von Monakow, 1914; Andrews, 1991). Neuroimaging and models of connectivity have extended this concept by showing complex changes in network topology and information transfer depending on which area was affected by a lesion (Honey & Sporns, 2008; Alstott et al. 2009; van Meer et al. 2012; Carter et al. 2012).

The finding that the normalization of disturbed intra- and interhemispheric interactions is accompanied by motor improvements implies that connectivity is an important pathophysiological factor contributing to neurological deficits after stroke. Hence, the network perspective of connectivity analyses provides insights into the pathophysiology underlying neurological deficits (James et al. 2009; Grefkes et al. 2010; Wang et al. 2011). Taken together, evidence from animal models and neuroimaging studies in humans suggests that motor disability after focal stroke lesions results from changes in the entire motor network rather than from the removal of a single component (Lemon, 2008). The application of connectivity analyses in the therapeutic realm is still in the very beginning. Therefore, intervention studies are now needed to provide further insights into causal relationships between behaviour and connectivity (Grefkes & Fink, 2011).

Glossary

BOLD

blood-oxygenation-level-dependent

DCM

dynamic causal modelling

DTI

diffusion tensor imaging

EEG

electroencephalography

ICA

independent component analyses

MEG

magnetencephalography

PCA

principal component analyses

PMC

premotor cortex

rCBF

regional cerebral blood flow

SEM

structural equation modelling

SMA

supplementary motor area

TMS

transcranial magnetic stimulation