Functional connectivity changes in cerebral small vessel disease - a systematic review of the resting-state MRI literature

Inclusion criteria for articles considered in this review were as follows: (1) written in English, (2) analysing exclusively human study participants, (3) published after January 2010, (4) radiological evidence of sporadic cerebral small vessel disease with structural brain imaging showing manifestations of CSVD in the form of white matter hyperintensities in at least a subset of the study population, and (5) analysis of resting-state functional connectivity using functional MRI. We excluded review articles; descriptions of ongoing studies; functional imaging studies using only electroencephalography, magnetoencephalography, or positron emission tomography; and reports concentrating exclusively on patients with non-sporadic CSVD, e.g. of genetic origin, or non-vascular dementias, e.g. Alzheimer’s disease.

Following a prespecified search strategy, the PubMed online database was queried for studies published between January 2010 and November 2020 using the conjunction of keywords specific for pathology (‘small vessel disease’, ‘white matter lesion’, ‘leukoaraiosis’, ‘microangiopathy’), network science (‘connectivity’, ‘network’, ‘graph’, ‘module’), and imaging modality (‘MRI’, ‘BOLD’, ‘resting state’) as search criteria (see Additional file 1 for the exact search strategy). In addition, references of search results were screened for further eligible articles. Studies were discarded if the title or abstract indicated failure to meet all the specified inclusion or satisfaction of at least one exclusion criterion. The remaining articles were read in full and evaluated according to the stated criteria.

The risk of methodological bias in individual studies (PRISMA items 12 and 19) was assessed using the Appraisal tool for Cross-Sectional Studies (AXIS tool) [26], modified to not contain items related to presentation of the Results, the Discussion of findings, or the Funding of the study [27]. Detailed descriptions of individual items are presented in Additional file 1: Table S1. Based on the number of quality criteria satisfied, each study was assigned an integer score from 0 (no criteria satisfied) to 11 (all criteria satisfied). Trichotomising this ordinal scale, we classified the risk of bias as high (score 0–3), moderate (4–7), or low (8–11). We strived to cover the literature comprehensively, and even a high risk of basis was therefore not defined as an exclusion criterion for this review.

After screening, the following data were extracted from the articles: year of publication; sample size; average age and clinical characteristics of study populations including measures undertaken to minimise confounding by comorbidities; the employed operationalisation of cerebral small vessel disease and severity grading of WMH; details of the scanning parameters and pre-processing steps including the controversial topics of motion scrubbing and global signal regression; the analytical approach to functional connectivity; and key results regarding patterns of altered connectivity in patients with CSVD and, if reported, their relation to cognitive performance.

For ease of presentation, studies were classified according to clinical characteristics of the study population—manifest CSVD, healthy participants, or others, not primarily vascular clinical conditions—and their main approach to quantifying and analysing connectivity. These predefined analytical categories included short-range connectivity within a part of the brain, long-range connectivity between pairs of remote brain areas defined either a priori or using a data-driven approach, and global analyses of topological properties of the functional connectome. We also reviewed the cognitive tests applied in these studies and associations of cognitive ability with functional connectivity measurements.

The results of the search and selection process are summarised in Fig. 1. We identified a total of 493 potentially relevant papers, 471 of which were obtained by searching PubMed and 22 through personal communication or as references cited in other works. Four hundred seventeen papers were excluded based on their title or abstract. Of the remaining 76 studies, which were read in full by both MS and ES, 44 were included in this review. Details of individual studies are summarised in Table 1.

The number of subjects, including both patients with small vessel disease and controls depending on study design, varied between 11 and 1584 with a median sample size of 72.5 (interquartile range [IQR] 50.8–106.8). Mean age across studies ranged from 50.0 to 76.4 years, with a median of 66.0 years (IQR 62.4–69.8 years).

Regarding the underlying research questions, roughly half of the included studies (24/44) reported the investigation of altered functional connectivity patterns in the presence of cerebral small vessel disease and its relation to cognitive ability as their primary research objective. Of these, six reports focused on patients with CSVD exclusively, whereas the study designs of the remaining reports involved comparing groups of healthy controls, patients with non-vascular cognitive impairment, or both. Twelve studies reported measures of cerebral small vessel disease, often as part of a more comprehensive assessment of structural brain parameters, and functional connectivity in populations of healthy participants without clinically manifest vascular pathology or cognitive impairment. Eight articles addressed functional connectivity in the context of other clinical conditions not directly related to vascular pathology, such as tau pathology-associated cognitive impairment or depression, but included markers of small vessel disease as covariates.

All of the 24 MRI studies reporting on resting-state functional connectivity in the context of clinically overt CSVD defined the presence of white matter hyperintensities on T2-weighted cerebral MR imaging as one of their inclusion criteria. In more than half of the studies (14/24), these were evaluated according to the ordinal Fazekas scale [55, 56]; in three studies, authors chose the Wahlund scale to assess age-related white matter changes [32]; white matter lesion load was also quantified volumetrically in eight studies; no precise definition of imaging criteria was reported in five articles.

When white matter disease was reported as a structural covariate in the investigation of functional connectivity, the extent of structural changes was quantified using either absolute or relative white matter hyperintensity volumes. Techniques for segmenting WMH on either T2 or FLAIR sequences included manual, semi-manual, and fully automated approaches; in one case, the algorithm was not described [57].

Beyond the presence of white matter lesions, evidence of lacunes or recent lacunar infarcts was considered in 13/24 studies; the distinction between the two entities was often imprecise, with only three articles referring to the STandards for ReportIng Vascular changes on nEuroimaging (STRIVE, [6]) consensus statement in this context [31, 50, 51]. Reflecting their conceptualisation as fluid-filled cavities, lacunes were defined as hypointense ovoid regions on T2- or FLAIR-weighted imaging with a diameter ranging from [2–3] to [15–20] mm. Three studies required patients with CSVD to have evidence of at least one lacuna or recent lacunar infarct [28, 42, 52], while one report excluded such patients [36]. Information on the number of lacunes contributed to the definition of a compositive CSVD score in one study [49]; in the remaining cases, it was either reported descriptively or used as a covariate in statistical analyses [50]. While most studies specified cortical or large subcortical infarcts as an exclusion criterion, one article included such patients specifically [43].

In addition to imaging findings, clinical characteristics were used to define patient cohorts. This was done to either separate patients and participants with and without cognitive impairment; to differentiate patients with CSVD from patients with non-vascular cognitive impairment, especially Alzheimer’s disease; or to grade the severity of vascular cognitive impairment, ranging from cognitively normal (CN) over mildly affected (variably called subcortical vascular mild cognitive impairment [svMCI], or vascular cognitive impairment no dementia [VICND]) to subcortical vascular dementia (SVaD). In addition to dedicated diagnostic criteria for different dementias [58, 59], cognitive assessment was based predominantly on scales such as the Mini Mental State Examination (MMSE), Montreal Cognitive Assessment battery (MoCA), or Clinical Dementia Rating scale (CDR). A minority of studies used the Petersen criteria and included functional activities and temporal evolution of cognitive abilities in their definition of mild cognitive impairment [60, 61]. Only three studies employed positron emission tomography (PET) to distinguish tau and/or amyloid pathology from purely vascular disease [49, 62, 63], and the risk of confounding by mixed disease seems therefore high in the majority of reported studies.

Magnetic resonance imaging was performed on scanners from a variety of vendors (Siemens, Philips, GE), usually at 3 Tesla. The use of specialised receiver head coils, multi-band, or multi-echo techniques was rarely reported. Repetition time (TR) and echo time (TE) were predominantly set at 2000 ms and 30 ms, respectively, with exceptional values ranging from 700 to 4500 ms and 13 to 84 ms, respectively. Reconstructed voxel sizes in the BOLD scans varied in the range [1.7–4] × [1.7–4] × [2–6] mm³, arranged in a three-dimensional matrix of dimensions varying in the range [64–128] × [64–128] × [20–64]. The number of acquired BOLD volumes varied between 100 and 700 (median 230, IQR 180–240). Participants were asked to keep their eyes open in 12 and closed in 20 of the reviewed studies; 12 articles provided no information. Description of functional MRI acquisition parameters was incomplete in 26 of the 44 analysed studies (61%); methods were judged as not-repeatable in these cases (AXIS item 11). Hemodynamic lags were not considered.

Pre-processing steps common to most studies included slice-time correction; realignment to a reference volume to correct for head motion; normalisation to a template space (usually MNI EPI [64]) including resampling; temporal band-pass filtering (lower end 0.005–0.01 Hz; upper end 0.08–0.15 Hz); and smoothing with a Gaussian filter of full-width at half maximum (FWHM) between 4 and 8 mm.

Confound regression and motion scrubbing were performed and reported less uniformly, as detailed in Tables 1 and 2. Specifically, global signal regression (GSR), that is orthogonalisation of voxel-wise timeseries with respect to the average BOLD signals from the white and grey matter, or the whole brain, was undertaken in 27/44 studies. Twenty-five studies employed subject-wise censoring in which participants were excluded from further analysis if the maximum or average head translation or rotation during the scan exceeded a certain threshold ranging from 0.5 to 3 mm translation and 0.5 to 3° rotation. Ten studies performed volume censoring according to a framewise displacement (FD) or framewise translation/rotation cut-off, excluding participants with too few remaining uncontaminated volumes [31, 49]. Two studies used spike regression [39, 65].

The majority of studies investigated large-scale functional connectivity between remote brain areas, choosing full or partial temporal correlations between the BOLD time courses as a measure of connectivity. Regions of interest were defined a priori using external brain parcellations in 16 cases. Twenty-seven studies used a data-driven approach such as independent component analysis (ICA), seed-based connectivity analysis (SCA), or local BOLD activity (amplitudes of low-frequency fluctuations [ALFF]) to define regions of interest for further analysis. Many authors interpreted alterations in functional connectivity in the context of a small number of large-scale resting-state brain networks (RSNs), in particular the default mode (DMN) and frontoparietal control (FPCN) networks, but also the dorsal attention (DAN) and salience (SN) networks [22]. Eight reports used graph theoretical approaches, including global network parameters such as efficiency and clustering coefficient [31, 42, 43, 52, 53]; analysis of modularity structures [45, 50]; and self-referential quantification of region-specific centrality [35] to summarise the patterns of connectivity between multiple regions and to thus reflect global organisational principles of the brain networks.

Three studies investigated short-range connectivity [38, 66, 67], using regional homogeneity (ReHo) to quantify the similarity between BOLD signals as a marker of local connectivity [68].

The main findings of individual studies with respect to alterations in resting-state functional connectivity in the context of cerebral small vessel disease are summarised in Table 1. Patterns of altered connectivity were expressed either in comparison to healthy controls or along a gradient of increasing severity of CSVD imaging markers. For clinically or radiologically manifest CSVD, reduced functional connectivity dominated the findings on a global scale [43, 45, 53]. Within resting-state networks, lower functional coupling was repeatedly reported between components of the default mode network [28, 35, 44, 46, 47, 51], which is further supported by the co-occurrence of reduced DMN connectivity and increased WMH burden in patients with non-vascular cognitive impairment [62]. Within the FPCN, reduced connectivity was found in the left insula [46], whereas the right inferior parietal cortex appeared to be more strongly coupled to the rest of this network [51]. The average coupling between the DMN and FPCN was found to be reduced in patients with CSVD [51], even though a small number of inter-network edges showed increased connectivity [39]. The connectivity of the DAN was altered in relation to other networks with increased coupling to the FPCN and reduced coupling to the posterior DMN [39, 51]. The same pattern of altered inter-network connectivity was reported for the salience network [48]; intrinsic connectivity in the SN was increased in patients with CSVD and in association with the extent of white matter disease [83]. In healthy individuals or patients without symptomatic CSVD (Table 2), most studies did not report significant associations between FC and WMH burden [66, 79, 82, 86, 92]. Two studies found an association between higher FC, especially in occipital and frontal areas, and WMH burden [87, 93], whereas in patients with late-life depression the pattern was more similar to the one seen in patients with CSVD [69, 71].

In the majority of studies, cognitive testing on participants was performed and investigated in association with the extent of white matter disease and functional connectivity. In addition to scales covering the global level of cognitive functions and deficits (MMSE, MoCA, and CDR), impairments in specific cognitive domains were quantified by sub-scores of these global scales or specialised neuropsychological test batteries, operationalising, in particular, executive function, processing speed, and memory. Table 3 summarises key findings of individual studies in these different domains. Most studies were able to confirm known associations between CSVD and cognitive impairment on the one hand, and, albeit less robustly, between functional connectivity and cognitive impairment on the other hand. Only few articles, however, addressed the question of how structural white matter damage and functional connectivity interact to affect cognition. In one analysis of 127 clinically healthy participants of the Harvard Ageing Brain Study, it was shown that the extent of WMH-associated decoupling of structural and functional connectivity in the default mode network correlated with both executive function and memory [81]. Moreover, in a combined analysis of 140 healthy participants and 90 patients with both vascular and non-vascular cognitive impairment, the authors demonstrated that the association of higher WMH load with poorer executive function and memory scores was moderated by global functional connectivity in the FPCN and by local FC in the salience network [65].

Risk of bias was assessed using the AXIS tool for all 24 studies recruiting patients with clinical CSVD. We did not formally assess the risk of bias in studies reporting results on FC and WMH in the context of conditions different from vascular cognitive impairment or in longitudinal studies. According to the AXIS tool, all studies thus assessed had an at least moderate risk of bias (10/24 moderate, 14/24 high). The distribution of assessments of individual quality items of the tool is depicted in Fig. 2 a. The overall aim or objective of the study (Item 1) was deemed unclear in 14 studies, often because of a lack of distinctions between exploratory and confirmatory, and causal and correlational approaches. In 9 cases, where aims included the inference of causal effects or were too broad to be assessed, a cross-sectional design was judged as inappropriate (Item 2). The sample size was not satisfactorily justified in any study. The reference population (Item 4) was mostly adequately specified as patients with CSVD, qualified by lists of inclusion and exclusion criteria. In five studies, the definition of the target population was unclear or contradictory. All but one article reported results from single-centre studies that recruited a convenience sample from a clinical setting; in these cases, the sample frame (Item 5) was judged as inappropriate and the selection process (Item 6) as non-representative. The exceptions were an analysis of a formal clinical register [84, 99]. No article addressed non-responders. Risk factors and outcomes (Item 7) were mostly valid (see above); exceptions included one unvalidated method to quantify WMH load [87] and the use of global graph parameters such as efficiency and path length. Reliability of outcome measures (Item 9) was generally judged to be low given the poor reproducibility of FC estimates in the context of CSVD, except for studies who explicitly estimated reliability as part of the study design [31, 42]. There were two main problems with the statistical methods used: firstly, confusion of exploratory and confirmatory approaches (cf. Item 1) led to a lack of clearly specified hypotheses and thus to inappropriately controlled type-I error rates in the case of multiple testing; secondly, many papers employed multi-scale approaches, in which results from the first, often global, analyses informed hypotheses tested in later, often more local, analyses. It is known that this method can inflate the rate of false-positive findings if the entire analysis pipeline is not accounted for properly, for example in a bootstrap loop [100]. The quality of the description of methods varied considerably. No article provided links to the program code used in the analysis, but this was not required to satisfy Item 12. Specific shortcoming included incomplete reporting of MRI acquisition parameters, lack of description of structural image pre-processing, and lack of detail in the description of statistical methods, such as choice of covariates, method to determine of p-values, or correction for multiple testing. The distribution of aggregate AXIS scores is shown in Fig. 2b. Given that none of the included studies had a low risk of bias or was preregistered, the overall risk of bias in the reviewed literature seems high.

Cardiovascular risk factors such as age, hypertension, diabetes mellitus, and dyslipidaemia are known to be associated with imaging markers of CSVD [101]. They also affect cerebrovascular reactivity and the circulatory autoregulation in response to neuronal activity (neurovascular coupling) [29] and are thus potential confounders of the relation between WMH and BOLD-derived functional connectivity. Similarly, vasoactive medications, in particular antihypertensives, which are commonly prescribed to patients with CSVD as well as substances like nicotine or caffeine may alter neurovascular coupling [102]. Despite this, reporting of and adjustment for comorbidities and medication was poor in the reviewed studies. While information on the demographic variables age and sex was provided in all reviewed articles, only about half reported results of analyses adjusted for these factors. Nine articles gave details on cardiovascular risk factors, yet none attempted to control for their potential confounding effect. Effects of prescribed medication or caffeine intake were not considered.

In addition to sample size, groups of studies also differed in their approaches to quantifying the severity of white matter disease. Clinically focused studies tended to rely on validated rating scales, such as the Fazekas or Wahlund scale, which assign an ordinal score based on the extent and distribution of white matter hyperintensities on T2-weighted MR imaging. A minority of studies considered the presence of lacunar infarcts as an additional marker for CSVD. The population-based studies of healthy participants, on the other hand, employed the cumulative volume of WMH as a continuous measure of disease burden. Numerical lesion load has the advantage of providing better resolution of inter-individual differences in groups of mildly affected participants; in addition, it can be determined reasonably reliably using automatic or semi-automatic image processing methods, although some degree of manual post-processing was usually done in the studies reviewed here [106]. Brain atrophy as a structural marker of both CSVD and neurodegenerative disease is known to be associated with changes in intrinsic brain connectivity [107]; it was included in many of the population-based studies using either the total intracranial volume to normalise observed WMH loads or region-specific grey matter volume, such as can be obtained from voxel-based morphometry (VBM) or cortical thickness measurements. Although methods have been developed to segment perivascular spaces (PVS) and cerebral microbleeds in an automated fashion [108‐114], none of the reviewed articles utilised enlarged PVS, and only one used microbleeds [49] as a marker of CSVD.

The variety of qualitative and quantitative analysis methods reflects the clinical heterogeneity of study populations comprising patients with CSVD at different stages of the disease. An attempt at standardising the assessment and reporting of imaging markers of CSVD was made in the STandards for ReportIng Vascular changes on nEuroimaging (STRIVE) position paper [6]. However, despite being published in 2013, the definitions and recommendations outlined in the STRIVE were referenced in only six of the 33 reviewed papers published after 2013 [31, 50‐52, 62, 108].

The analysis of recorded BOLD signals has not been standardised, with a broad variety of coupling measures and dimensionality reduction techniques being at the disposal of the researcher [20]. All reviewed studies used Pearson’s correlation coefficient to quantify the synchrony between BOLD time series in different parts of the brain. No clear distinction between full and partial correlations was often made, thus making the interpretation of direct or indirect connectivities difficult [115]. Similarly, the handling and interpretation of negative correlations was rarely reported or discussed [116‐118]. None of the included articles attempted to estimate directed [119‐121] or time-varying functional connectivities [122‐124], or to quantify patterns of synchronous activity involving more than two regions [125, 126]. Analytical approaches included whole-brain analyses (eigenvector centrality, connectivity density); the investigation of functional connectivities between region of interests, often components of well-defined intrinsic resting-state networks, which were either derived from the data themselves (independent component analysis) or specified a priori by an external brain parcellation; and combinations of the two (seed-based correlation analysis).

Brain parcellations for the region-of-interest-based analyses were mostly based on anatomically defined atlases, such as the automatic anatomical labelling (AAL) atlas [127], the Desikan–Killiany parcellation [128], or the H-1024 random parcellation [129], which do not take into account the functional architecture of the brain. Only three very recent articles [31, 49, 66] used the multimodal brain parcellations of Power [130] or Schaefer [131], or the Brainnetome atlas [132], which have been shown to better respect the functional organisation of the brain [22]. In addition to interpreting changes in functional connectivity directly, a few studies attempted to summarise patterns of FC by measures of global network organisation using graph theory. These approaches have been instrumental in the study of complex brain networks and include parameters to reflect the notions of integration, such as efficiency or characteristic path length; segregation, such as clustering coefficients; or community structure, quantified by modularity scores and participation coefficients [9, 133, 134]. With the exception of community detection, however, these network parameters have been defined only for structural brain networks and lack validation for networks derived from functional connectivity [135].

The large-scale temporospatial organisation of neuronal activity in the brain is known to be supported and constrained by the anatomy of axonal projections that form structural connections between both adjacent and remote brain areas [136, 137]. This coupling between structure and function is particularly pronounced in the default mode network [138], possibly reflecting the long periods of time that the brain is engaged in inward-directed thought, memory formation and retrieval, or social estimation [139]. While the structural connectome thus contributes to maintaining stable neural activity patterns, it also means that normal functional connectivity is vulnerable to damage to white matter pathways as occurring in CSVD [140]. Most articles included in this review quantified the extent of white matter damage by using either neuroradiological rating scales or total lesion volume, as detailed above. Such global approaches are, however, not able to differentiate between lesions in functionally silent brain areas that can more easily be compensated by rerouting information through alternative redundant pathways, and lesions in functionally critical, strategic locations, where even spatially limited damage can be associated with substantial behavioural sequelae. In the context of cognition, damage to subcortical nuclei and tracts with a high density of neuromodulatory projections such as the dorsomedial and anterior thalamic nuclei or the anterior limb of the internal capsule appears to be particularly consequential [24, 52, 87, 141]. Advanced diffusion-weighted structural imaging modalities allow the spatial mapping of fibre tracts and the quantification of tract-specific white matter lesion loads [142, 143]. In combination with resting-state BOLD imaging, this approach has been used to show that leukoaraiosis disrupts functional connectivity in a spatio-topological non-uniform way that is shaped by the anatomy of the brain’s white matter scaffold [84]. The strongest association between tract-specific ischaemic damage and reduced FC was observed in the fronto-occipital fasciculus, which supports connectivity between the salience and frontoparietal control networks [144]. In addition to affecting functional connectivity directly, ischaemic white matter disease also seems to exert an indirect effect by modulating the coupling between structural and functional connectivity. Specifically, the association between mean diffusivity in the cingulum bundle and functional connectivity between the medial prefrontal and posterior cingulate cortices was significantly attenuated in patients with higher WMH burden, thus contributing to decoupling the anterior and posterior parts of the default mode network [81].

Both structural and functional connectomes share properties of complex networks, such as the presence of network communities, high-clustering with short path length (small-worldness), and hierarchical organisation [11, 145]. With cognition considered an emergent property of distributed neuronal activity in the brain [146, 147], understanding the behavioural sequelae of CSVD requires an understanding of how ischaemic lesions disturb not only specific fibre tracts and functional connections but also the global organisation of synchronous activity. Graph theoretical analyses have suggested that the global topology of functional brain networks in the presence of CSVD exhibits increased path length and modularity and reduced small-worldness that correlated with cognition [45, 53]. A similar effect was also observed in the structural networks of patients with CSVD and ischaemic stroke [13, 15, 148‐150].

An intriguing open question is the differentiation between altered functional connectivity as a direct consequence of damage to the supporting fibre tracts, and compensatory changes. The latter are thought to contribute to maintaining normal cognitive function in the early stages of the disease [83]. Indeed, increased coupling between brain areas has repeatedly been reported in cognitively normal individual with white matter hyperintensities [78, 87, 93].

The association of white matter hyperintensities of presumed vascular origin with cognition has been extensively described [151‐153], and indeed, cognitive impairment is one of the clinical hallmarks of manifest cerebral small vessel disease [154]. On the other hand, resting-state fMRI connectivity has been found useful in extracting neural correlates of cognitive function and mood disorders [155, 156]. Under normal physiological circumstances, patterns of coordinated activity within and between a small number of large-scale intrinsic brain networks have emerged as particularly relevant [146], including activation of the default mode network in brain states characterised by self-referential thought or rest that is anti-correlated with activation of the dorsal attention network; deactivation of the default mode network during focused attention on external stimuli [157]; and a modulating role of a frontoparietal control network with increased connectivity to the DMN as a correlate of working memory performance [158, 159]. Building upon these ‘cornerstones’ of functional connectivity under normal physiological circumstances, a disconnection hypothesis has been developed that postulates reduced DMN and FPCN connectivity, decoupling of neuronal activity along the anterior–posterior axis, and functional disconnection of the prefrontal cortex as neuronal correlates of cognitive impairment in CSVD [24]. This model is supported by several recent studies that reported decreased functional connectivity between the medial PFC and posterior components of the DMN in patients with CSVD [28, 44, 47], and observed an association with reaction times in the Stroop test [35]. A behaviorally relevant dissociation in functional resting-state fMRI activity and local connectivity was found between the anterior and posterior parts of the DMN with lower ReHo and ALFF values in the medial PFC and higher values in the precuneus and posterior cingulate cortex in patients with CSVD compared to healthy controls [38, 44]. Both increases and decreases of FC within the FPCN and DAN as well as their coupling with the DMN have been reported to be associated with CSVD [39, 46, 51], but the heterogeneity of these results and limited correlation with cognitive test scores makes it difficult to distinguish primary effects of disconnection from compensatory changes or sampling variability without physiological relevance.

In addition to these established networks, connectivity patterns of the salience network (SN) have recently been investigated, with increased SN-FPCN and SN-DMN couplings associated with small vessel disease [51]; additionally, increased connectivity within the SN in patients with CSVD was associated with worse performance in the Stroop interference test [83]. In patients with mild cognitive impairment, the association between white matter disease and executive function was attenuated in the presence of increased local connectivity of the salience network [65]. The salience network includes the anterior insula, the dorsal anterior cingulate cortex, and subcortical components. Similar to the FPCN, it has a critical role in switching activity between different brain networks and has been implicated as a key component in network models of neuropsychiatric disorders [160‐162]. Specifically, increased connectivity within the SN and altered SN-DMN and SN-FPCN coupling have been described in patients with Alzheimer’s disease and mild cognitive impairment [163‐165].

Community-dwelling adults with early CSVD often perform normally on neuropsychological tests and only report mild subjective cognitive deficits [166]. This preclinical stage has been linked to compensatory mechanisms especially in patients who benefit from a larger cognitive reserve [167, 168]. Three recent studies provide further evidence for this hypothesis, linking increased functional connectivity to frontal and temporal areas to ischaemic white matter lesion load in cognitively normal subjects [78, 83, 93].

While it is possible to extract consistent themes from the reviewed articles that point toward physiologically relevant patterns of altered FC in the context of CSVD and cognitive impairment, the current literature is characterised by a high degree of heterogeneity of individual results. As discussed above, this may partly reflect variability in pre-processing and analytical approaches as well as heterogeneity in the clinical populations under investigation. However, given the absence of preregistered reports or high-quality multi-centre studies and the predominantly moderate-to-high risk of bias in individual studies, it must be assumed that selective reporting allowed the literature to be contaminated by a substantial number of false-positive findings, reflecting spurious associations and group differences. In addition, it is possible that reported results are confounded by the presence of other age-related pathology or neurodegenerative comorbidities, such as Alzheimer’s disease [169], which were considered specifically in only a small minority of studies.

Comparison and synthesis of individual study findings is further hampered by differences in data cleaning techniques, which are known to influence functional connectivity estimates [170]. Two important dimensions of BOLD pre-processing relate to removal of the global signal from the whole brain or tissue type compartments, and handling of subjects or frames with high motion. Global signal regression is known to be effective at mitigating the widespread inflation of connectivity estimates induced by subject motion, resulting in an elevated distance-dependence of residual motion artefacts [171]. Despite this theoretical prediction and the observation that GSR might improve associations between FC and behavioural measures [172], the use of GSR was not associated with specific patterns of altered connectivity or stronger relations with cognitive measures in the reviewed papers. Similarly, no clear effect of different motion scrubbing strategies, i.e. the censoring of subjects or individual volumes due to excessive average or framewise displacement, could be recognised. It seems likely that the myriad of unstandardised pre-processing choices is contributing to the heterogeneity of published results and that findings which have not been shown to be robust with respect to such choices should therefore be interpreted with great care.

Even ignoring potential biases inherent in study design and publication practice, the study of FC in the context of CSVD may be limited by more fundamental obstructions. One concern is that the reliability of estimating functional connectivity may be negatively affected by the presence of white matter lesions itself. Two of the reviewed studies reported results from repeated measurements on participants in longitudinal designs [31, 42]. Worryingly, in both cases, resting-state fMRI measures were found to be poorly reproducible, indicating a further need to evaluate their robustness as an imaging biomarker. In one case, this might have been a consequence, in part, of using a brain parcellation that does not respect the functional boundaries between brain areas, which is known to be damaging to network estimation [119]. However, the persistence of low reliability measures for a range of network characteristics across network densities and atlas resolutions, as well as the particularly poor reproducibility of functional network measures in patients with CSVD compared to controls, suggests more fundamental problems beyond the choice of parcellation. The finding of poor reproducibility of RSNs and graph metrics in CSVD contrasts with high reproducibility reported in healthy participants [173‐176] and patients with stable multiple sclerosis [177‐179]. It has been suggested that age and confounding age-related pathologies could be responsible for reduced reliability of functional connectivity estimates [180, 181]; however, specific methodological challenges arise in patients with cerebral small vessel disease as a consequence of microvascular pathology, that are absent in other conditions.

As a measure of synchronous brain activity, the interpretation of BOLD-derived functional connectivity is contingent upon an understanding of the relation between neuronal activity and local blood flow. This neurovascular coupling, however, is known to be altered in normal ageing as well as the presence of ischaemic disease [102, 182, 183], and attributing differences in BOLD-derived measures of connectivity to either vascular or neuronal factors is therefore challenging [184]. More specifically, white matter lesions of presumed vascular origin are known to be associated with subcortical hypoperfusion [185], possibly reflecting observed rarefaction of the microcirculation in a mouse genetic model of CSVD [186]. The later stages of neurovascular coupling involve dynamic upregulation of regional blood flow mediated by increased CO₂ concentration in areas of increased neuronal activity [102]. This mechanism appears to be affected in the presence of CSVD as demonstrated by a diminished cerebrovascular response to hypercapnia in an early study involving 24 patients with leukoaraiosis [187], and an association between WMH load and sonographically assessed measures of pulsatility and dynamic autoregulation in a cohort of elderly patient with cardiovascular risk factors [188]. These findings are further complicated by differences in age-related changes in cerebrovascular reactivity between grey and white matter [189]. BOLD-derived functional connectivity is a function of BOLD activity in remote brain areas, and spatial variations in age- or disease-related changes in neurovascular coupling might therefore affect FC estimates in unpredictable ways [190]. A small study of 25 subjects with WMH found that while cardiovascular risk factors are associated with cerebrovascular reactivity, no such association was observed for resting-state functional connectivity in the default mode network [29]. One potentially testable hypothesis about the effects of impaired neurovascular coupling on functional connectivity estimates derives from the observation that BOLD-derived measures of synchronous brain activity are a composite of true coincident neuronal activation (‘signal’) and shared noise, where the latter tends to be more dominant for short-range connections [171]. Reduced ‘signal’ strength as a consequence of a lower vascular response would therefore be expected to result in weaker and less precise FC estimates, especially in long-range connections.

While being comprehensive in our inclusion of primary research articles from electronic databases and other sources, we cannot exclude the possibility that additional findings from the grey literature, such as blogs or unpublished conference abstracts, have not been covered by this review. In order to keep the scope of the work focused, we have not included reports of task-based connectivity or resting-state connectivity derived from electrophysiological recordings. Findings obtained using these alternative paradigms and modalities might lend further support to the themes of disturbed connectivity patterns outlined above. This review attempted a qualitative synthesis of the recent literature; the heterogeneity of study designs and populations did not permit the extraction and quantitative analysis of numerical effect estimates beyond sample size and age of participants. From a meta-analytical perspective, it can be noted, however, that all studies of patients with clinically manifest CSVD report significant FC alterations, while that is the case for only 30% of the population-based studies despite larger sample sizes. This discrepancy could reflect larger effect sizes in clinically preselected patients or indicate selective reporting in the sense of publication bias [191].

For conciseness, we have concentrated our attention on cognitive impairment as one of the main clinical sequelae of CSVD. Associations of altered patterns of functional connectivity with depressive symptoms, apathy, or gait imbalance were rarely reported and have not systemically been explored here. As an entry point to the recent literature, we note that abnormal functional coupling has been observed as a correlate of late-life depression in the context of the vascular depression hypothesis [192‐195]; while apathy has been investigated using resting-state fMRI in various clinical contexts [196‐198], results on gait disorders are scarce [34, 199]. Functional connectivity does not seem to interact with race or socio-economic status as possible contributing factors to neurodegeneration [200].