The neural correlates of discrete gait characteristics in ageing: A structured review

Highlights • Gait impairments are linked to deterioration of the brain.• Evidence suggests that specific brain regions relate to different gait components.• Future gait velocity decline may be predicted from measurements of white matter.• Cognition can mediate associations between gait and neuroimaging parameters.• Functional neuroimaging will aid further interpretation of neural-gait correlates.


Introduction
Gait is considered to be an important indicator of overall health; poor gait performance in older adults is associated with greater morbidity, mortality, and fall risk (Hausdorff et al., 2001;Studenski et al., 2011;Verghese et al., 2011). The idea that a fully functioning motor system is solely responsible for effective gait has been replaced by a more encompassing sensory-cognitive-motor model which reflects a contemporary understanding of gait as a highly complex skill (Montero-Odasso et al., 2012). Safe and effective negotiation of complex environments encountered in real-world settings requires the integration of external sensory information with neural networks which involve cortical, subcortical, brainstem, and spinal cord structures (Bohnen and Jahn, 2013;Takakusaki, 2013). Additionally, the control of gait from higher-order regions responsible for cognition is becoming increasingly evident, with implications for ageing, mobility and cognitive decline (Morris et al., 2016).
Understanding the mechanisms of simple gait, which does not involve obstacles or slopes, is crucial to improve the health status of the older population, yet the precise nature of neural control of gait during typical healthy ageing is currently unclear. The brain has traditionally been assessed through the definition of specific localised regions of interest; however, it is now thought more appropriate to consider regions connected together, either structurally or functionally, as defined neural networks. This integrative brain network perspective should increase our understanding of the dynamic brain during gait, and may also help to address common and overlapping neural outputs such as gait and cognition. In spite of this, whilst topological associations with different motor and cognitive domains have been established, neither precise brain regions nor neural processes or networks linked to discrete characteristics of gait control have been well defined. This lack of clarity causes difficulty in discerning whether individual neural regions or networks should be targeted in a different manner when aiming to improve gait impairments. lack of subjectivity allows for more accurate comparisons to be made between and within subjects, as well as between studies. Gait speed is typically used as a global measure of gait, due to its ease of measurement and reliability (Wade, 1992). However, this provides a limited approach to gait assessment, as it is not reflective of the subtle and selective gait alterations that occur in response to ageing and disease, which are potentially controlled by different networks (Stolze et al., 2001;Verghese et al., 2007;Lord et al., 2014). Gait can be represented by multiple discrete gait characteristics, which are thought to represent different features of neural control, and are therefore assumed to respond in a selective manner to ageing and pathology. Several groups have devised models of gait which sort discrete gait characteristics into specific gait domains through data reduction techniques (Verghese et al., 2008;Hollman et al., 2011;Lord et al., 2013;Verlinden et al., 2013). Independent domains of gait, and the characteristics contained within them, can then be hypothesised to reflect independent neuroanatomical and functional substrates. Although these models have similarities, subtle differences differentiate them. Verghese et al. produced a model containing three gait domains; pace, rhythm and variability. Two additional domains were included within the model from Lord et al., asymmetry and postural control, through the inclusion of more gait characteristics. Other models such as that from Verlinden et al. contain domains relating to more complex gait tasks such as turning. Fig. 1 shows the model from Lord et al. in older adults (Lord et al., 2013), which has been adopted in this review as the most comprehensive model for simple gait, to provide a framework for rationalising and interpreting study findings. With this in mind, a more comprehensive approach to the characterisation of gait is needed to fully understand the neural correlates of gait control, beyond that of gait speed.

Neuroimaging techniques used to assess gait
A second major challenge in understanding the neural underpinnings of gait is that imaging gait in real-time is not easily achieved. Several approaches have been developed to address this limitation. Electroencephalography (EEG) can be used to record electrical activity within the brain, whereas functional near infra-red spectroscopy (fNIRS) measures brain activity through haemodynamic responses in relation to neuronal behaviour; both provide real time information relating to brain activity during walking, However, these techniques can only measure superficial cortical activity, are indirect measures, and lack spatial resolution, so cannot accurately measure responses at the neuronal level. An alternative approach to understand the neural substrates of gait is to adopt cross-sectional and longitudinal study designs to explore the relationships between discrete gait characteristics and brain structure and function. These are typically assessed independently, using a range of neuroimaging techniques including both structural and functional magnetic resonance imaging (MRI, fMRI), diffusion tensor imaging (DTI), positron emission tomography (PET) and magnetic resonance spectroscopy (MRS). These techniques can be used to analyse the brain on both a global and regional level, aiding our understanding of the general imaging parameters associated with gait as well as the more specific brain areas linked to different aspects of gait control. Structural MRI can also be used for the detection of white matter hyperintensities (WMH), cerebral infarcts and cerebral microbleeds; these lesions are manifestations of subclinical cerebrovascular disease, yet are common amongst typically ageing older adults (Choi et al., 2012). Given the dynamic nature of gait, it is important to consider a more global integrative model of the brain functions which may underpin gait and other overlapping behaviours, rather than simply assigning specific specialisms to discrete brain regions. One approach is through the use of techniques such as fMRI, which can be used to assess the brain as a series of networks involved in gait control through defining areas with commonalities in activity levels across time. However, as gait and imaging assessments are dissociated in time, correlative comparisons are relied upon during analyses, limiting these detailed approaches.

Review aims
Overall aims of this review are to assess the global and regional neural correlates of gait in ageing, taking a detailed and comprehensive approach. By forming a matrix of associations between individual imaging parameters and gait characteristics (grouped into their appropriate gait domain as outlined by Lord et al. (Lord et al., 2013)), we aim to understand the discrete nature of gait control, underpinned by a robust model of gait, to aid our interpretation of study findings. In addition to cross-sectional studies, longitudinal study types which focus on healthy older adults will be included, so that the effects of typical ageing on neural gait correlates may be identified. Additionally, the effects of cognitive test scores, which assess higher order brain functions such as attention and memory, on neural gait correlates identified in individual studies will be outlined. This will help us to determine whether any neural pathways or regions are shared between gait and cognition, as the aforementioned cognitive control of gait suggests a three-way interplay between gait, cognitive and neuroimaging parameters. This review leads on from two recently published articles which identify links between gait and brain imaging, but with a wider range of clearly defined gait characteristics which map to a validated gait model (Tian et al., 2017a;Wennberg et al., 2017b). Our specific aims are to: i) explore associations between discrete gait characteristics and brain structure and/or function in older adults, as identified through neuroimaging; ii) explore the longitudinal relationship between changes in gait and anatomical or functional imaging correlates, and; iii) identify recommendations for future areas of research. We hypothesise that independent gait characteristics will reflect discrete regional brain structure and functional brain activity in older adults. To the best of our knowledge, no other review has taken a structured approach to comprehensively map the neural correlates of gait to a full robust gait model. This review will therefore provide a clear representation of the current literature, provide a map of the neural correlates of gait control and highlight gaps for future research.  (Lord et al., 2013) for older adults. 16 gait characteristics map to 5 gait domains; Pace, Rhythm, Variability, Asymmetry and Postural Control.

Search strategy
Three databases were used for the search: Medline, PsycInfo, and Scopus. Search terms relating to gait, neuroimaging, and older adults were included within each search; where possible, age limits and MESH headings were used. The search was limited to full journal articles only, written in the English language between 1990 and April 2018. Boolean operators were utilised in the search; "OR" was included between search terms within each section, whereas "AND" was included between the sections within each database. Table 1 includes the search terms used for each search.

Inclusion and exclusion criteria
Articles were included if they assessed gait in healthy older adults under single task conditions, and used at least one brain imaging technique. As we were interested in looking specifically at quantitative gait characteristics, complex paradigms which involved standing or turning, such as the Timed Up and Go task, were excluded, as were protocols which made use of imagined gait. Additionally, we did not consider walks performed under dual-task conditions (which involve a secondary task being completed whilst walking), as the methodologies of these dual-task paradigms greatly vary and may cause different impairments in gait (Beurskens et al., 2014;Doi et al., 2017), the details of which are not well explored and are beyond the scope of this review. Articles involving animal models, case studies, intervention studies, clinical trials or only nerve or brain stimulation were excluded, as were those that only assessed falls, freezing of gait or general physical activity.

Data extraction
Once duplicates were removed, one reviewer (J.W.) screened the titles from the initial search, and two reviewers (J.W. and L.A.) independently screened the abstracts to identify potential articles. The full-text of articles was retrieved if reviewers were unable to determine the eligibility of the study from the title and abstract alone. All fulllength articles were assessed by three reviewers (J.W., L.A., and R.M.A.). Data extraction forms were completed, which included information about population characteristics, whether the study was cross-sectional or longitudinal, the study inclusion and exclusion criteria, the gait analysis technique and variables measured, the imaging technique and variables measured, the statistical tools used, and the main study findings. A quality assessment was conducted separately by two reviewers (J.W and R.M.A) and overall quality scores were determined for each study (see Supplementary Table 1).

Search yield
The search yield is shown in Fig. 2. The search, completed on 04.04.2018, generated a total of 38,029 studies after search limits were applied. Once duplicates were removed, a total of 31,060 studies were yielded from the search. After the initial title screen, 502 studies were identified as being of interest; 105 studies were then eligible for data extraction after abstract screening. 59 studies were excluded during data extraction, as no single task gait was completed (n = 4); only the timed up and go task was completed (n = 2); the gait assessment involved a turn (n = 3); the gait measurement tool was not described (n = 1); derived gait measures were unsuitable (n = 21); access to the paper was unavailable (n = 2); the image analysis undertaken only involved brain or nerve stimulation (n = 1); the age range investigated was inappropriate (n = 3); the type of article was unsuitable (n = 2); only results relating to disease cohorts were presented (n = 11); only results representing group comparisons were presented (n = 1); the direction of association between the variables was not specified (n = 1), or there was no direct link between the two variables of interest (n = 7). Six additional studies have been identified outside of the search strategy since the search closed. Therefore, 52 studies are included in this review. Publication dates range from 1997 to 2018.

Table 1
Search terms used for the searches performed within each of the databases. All searches contained terms from the gait, neuroimaging and old age categories.

Associations between brain structure and function and gait characteristics
Associations between quantitative gait characteristics and structural and functional imaging parameters were explored. In order to examine the wide variety of gait characteristics assessed, all gait characteristics included within studies were mapped onto the five domains of gait outlined in Lord et al's gait model (Lord et al., 2013). Studies assessing either steps or strides were considered and grouped together for the purposes of this review; definitions of the gait terms used have been described in supplementary Table 2. Imaging parameters were also grouped by the imaging technique used for an effective interpretation of results. Imaging parameters derived from structural MRI were divided in to two groups, those assessing brain volumes and those investigating white matter changes that are common in ageing. Where possible, the specific brain regions in which associations were made have been reported, as well as any covariates included in statistical analyses. 47 studies described cross-sectional associations between gait and brain imaging parameters (Rosano et al., 2005a, b;Della Nave et al., 2007;Rosano et al., 2007a, b;Baezner et al., 2008;Rosano et al., 2008;Nadkarni et al., 2009;Novak et al., 2009;Soumaré et al., 2009;Zimmerman et al., 2009;Murray et al., 2010;de Laat et al., 2011a, b, c;Sorond et al., 2011;Choi et al., 2012;de Laat et al., 2012;Dumurgier et al., 2012;Manor et al., 2012;Moscufo et al., 2012;Shimada et al., 2013;Willey et al., 2013;Wolfson et al., 2013;Annweiler et al., 2014;Beauchet et al., 2014;Bolandzadeh et al., 2014;Bruijn et al., 2014;Callisaya et al., 2014;Nadkarni et al., 2014;Rosso et al., 2014;Sakurai et al., 2014;Beauchet et al., 2015;Ezzati et al., 2015;Holtzer et al., 2015;Yuan et al., 2015;del Campo et al. (2016);Fling et al., 2016;Rosario et al., 2016;Stijntjes et al., 2016;Verlinden et al., 2016;Beauchet et al., 2017;Nadkarni et al., 2017;Sakurai et al., 2017;Tian et al., 2017b;Wennberg et al., 2017a;Fling et al., 2018). Fig. 3 summarises the number of studies comparing each gait characteristic to each imaging parameter; detailed descriptions of each cross-sectional study are included in Table 2.
As anticipated, gait velocity was the gait characteristic most commonly assessed. Several characteristics of gait are yet to be associated with any brain imaging parameters, including gait speed variability, swing time asymmetry and stance time asymmetry. Only four gait characteristics, from three gait domains, have been assessed longitudinally; gait velocity, step length, cadence and step width. Gait velocity was the only characteristic considered longitudinally in all studies.
Step width, from the postural control domain, was assessed in only one study, and was not associated with any imaging parameter longitudinally (Callisaya et al., 2013).
Fifteen of these studies evaluating GM volume assessed characteristics within the pace domain of gait. Eight of eleven studies assessing gait velocity found that slower gait was associated with reduced GM volume (Rosano et al., 2007a;Novak et al., 2009;de Laat et al., 2012;Dumurgier et al., 2012;Callisaya et al., 2014;Nadkarni et al., 2014;Ezzati et al., 2015;Stijntjes et al., 2016). This association was made across many brain regions. Slower gait was associated with reduced global GM volume (Callisaya et al., 2014;Ezzati et al., 2015), reduced frontal GM (Rosano et al., 2007a;Novak et al., 2009;Callisaya et al., 2014), reduced GM in the occipital cortex (Callisaya et al., 2014), reduced hippocampal volume (Ezzati et al., 2015;Stijntjes et al., 2016), and, in the only study to assess cortical thickness, with cortical thinning in all regions except for the inferior temporal gyrus (de Laat et al., 2012). Subcortically, slower gait was associated with cerebellar atrophy (Rosano et al., 2007a;Callisaya et al., 2014;Nadkarni et al., 2014), and reduced basal ganglia volume (Dumurgier et al., 2012;Callisaya et al., 2014). Reduced step length was also associated with reduced grey matter in all four studies assessing both parameters (Rosano et al., 2008;Zimmerman et al., 2009;de Laat et al., 2012;Callisaya et al., 2014). The association was again made with many brain regions, including global GM volume (Callisaya et al., 2014), hippocampal volume (Zimmerman et al., 2009), and prefrontal, parietal, supplementary motor, sensorimotor, occipital and limbic regional volumes (Rosano et al., 2008;de Laat et al., 2012). The relationship between step time variability and GM volume is less clear. Three of the four studies assessing this relationship used similar datasets; whilst greater step time variability was associated with increased hippocampal volume (Beauchet et al., 2015) and reduced GM volume of the right parietal lobe (Beauchet et al., 2014) in two papers, no association was made between step time variability and hippocampal or somatosensory volume in the third (Beauchet et al., 2017). Similarly, no association was made between stride duration variability and global GM volume or cerebellar, dorsolateral prefrontal cortex or basal ganglia volumes in one study (Manor et al., 2012). Swing time variability was not associated with hippocampal volume in one study (Beauchet et al., 2015); as no other studies have investigated swing time variability with grey matter volume, no firm conclusions can be drawn.
Five studies have investigated associations between GM volume and characteristics from the rhythm domain of gait. Two studies assessed double support time and had conflicting results; one found that longer double support time associated with a reduction of GM volume in areas including dorsolateral prefrontal cortex, right parietal lobules, right motor cortex and sensorimotor cortex (Rosano et al., 2008), the other found no association between double support time and global,

Table 2
Descriptive information of all cross-sectional studies. Participant ages have been reported as mean ± standard deviation unless stated otherwise.
There was no association between step length and metabolic rates of glucose uptake when assessed at maximum gait speed, or with any gait characteristics assessed at comfortable pace.    (Manor et al., 2012). Similarly, the three studies assessing step time and cadence -the inverse of step time -had conflicting results. Whilst no association was made between step time and hippocampal volume (Beauchet et al., 2015), or between cadence and total GM volume (Callisaya et al., 2014), one study found that decreased cadence was associated with cortical thinning in left cingulate and visual regions, the left fusiform gyrus and the primary and premotor cortices (de Laat et al., 2012). Only one study assessed swing time; it was not associated with hippocampal volume (Beauchet et al., 2015). Within the variability domain, one study investigated the association between each of step length variability and step width variability; these were not associated with total grey matter atrophy (Rosso et al., 2014) or hippocampal volume (Beauchet et al., 2015). No gait characteristics within the asymmetry domain have yet been assessed with GM volume. Three studies assessed GM volume with step width, the only characteristic assessed from the postural control domain. Two of three studies found that wider steps were associated with reduced GM volume in the inferior parietal lobe (Rosano et al., 2008;de Laat et al., 2012). Several other brain regions were related to step width in only one of these two studies, including the orbitofrontal and ventrolateral prefrontal cortices, temporal gyrus, left fusiform gyrus and the dorsal anterior cingulate cortex (de Laat et al., 2012), as well as the pallidum and right dorsolateral prefrontal cortex (Rosano et al., 2008).

White matter volume.
Six studies assessed total, cortical and regional white matter (WM) volumes (Della Nave et al., 2007;Sorond et al., 2011;Beauchet et al., 2014;Callisaya et al., 2014;Ezzati et al., 2015;Stijntjes et al., 2016). All studies assessed characteristics from the pace domain of gait; one additionally assessed cadence from the rhythm domain. Four of five studies found no association between WM volume and gait velocity (Della Nave et al., 2007;Callisaya et al., 2014;Ezzati et al., 2015;Stijntjes et al., 2016). Similarly, no association was identified between WM volume and step length (Callisaya et al., 2014) or step time variability (Beauchet et al., 2014) from the pace domain, or with cadence from the rhythm domain (Callisaya et al., 2014). No other gait characteristics were assessed with WM volume.
3.3.1.1.3. Total brain volume. Total brain volume was assessed in three studies, either as an absolute value (Della Nave et al., 2007; de     Laat et al., 2011c) or through assessment of brain parenchymal fraction (Sorond et al., 2011). All assessed gait velocity; two of three studies found no association between gait velocity and total brain volume (Della Nave et al., 2007;Sorond et al., 2011). De Laat et al., however, found associations between total brain volume and gait characteristics from all domains except asymmetry (de Laat et al., 2011c) with increased total brain volume being associated with increased gait velocity, step length and cadence, and with reduced step time variability, double support time, step length variability and step width. 3.3.1.1.4. Ventricular volume. Four studies assessed ventricular volume (Rosano et al., 2005a, b;Annweiler et al., 2014;Ezzati et al., 2015). Again, all studies assessed gait characteristics within the pace domain of gait, and one additionally assessed double support time from the rhythm domain. Gait velocity was the only gait characteristic assessed in multiple studies; two of these three studies found no association between gait velocity and ventricular volume (Rosano et al., 2005a;Ezzati et al., 2015). For the other gait characteristics assessed, one study found that increased step time variability was associated with increased ventricular volume (Annweiler et al., 2014), whereas another study found no association between ventricular volume and either step length or double support time (Rosano et al., 2005a). No characteristics from the variability, asymmetry or postural control domains were assessed with ventricular volume.

Longitudinal associations from volumetric analysis
3.3.1.2.1. Grey matter volume. In contrast to the cross-sectional findings, only two studies assessed changes in GM volume with changes in gait (Callisaya et al., 2013;van der Holst et al., 2018), and their results conflicted. Van der Holst et al. found no association between GM volume change and a change in either gait velocity, step length or cadence (van der Holst et al., 2018), whereas Callisaya et al. identified that a decrease in total GM volume over time was associated with a decrease in gait velocity and a decrease with cadence over time, and hippocampal volume reduction was associated with slowing of gait velocity and shortening of step length over time (Callisaya et al., 2013).

White matter volume.
Five studies assessed WM volume longitudinally with gait (Wolfson et al., 2005;Frederiksen et al., 2011;Ryberg et al., 2011;Callisaya et al., 2013;van der Holst et al., 2018). Both studies that used baseline WM volumes to predict gait changes identified that smaller baseline WM volume were predictive of greater decline in gait velocity over time (Wolfson et al., 2005;Ryberg et al., 2011). However, of the three studies assessing a change in WM volume with a change in gait velocity (Frederiksen et al., 2011;Callisaya et al., 2013;van der Holst et al., 2018), only one found that greater reductions in WM volume over time related to greater decline in gait velocity (Callisaya et al., 2013). A reduction in step length over time was associated with a reduction in WM volume over time in both studies assessing it (Callisaya et al., 2013;van der Holst et al., 2018). Change in cadence, from the rhythm domain of gait, was assessed with WM volume change in two studies (Callisaya et al., 2013;van der Holst et al., 2018), yet results were conflicting. One study found that a reduction in total WM volume over time was associated with a reduction in cadence over time (Callisaya et al., 2013), whereas another study found no association between the parameters (van der Holst et al., 2018).

CSF and ventricular volume.
Two studies used baseline CSF volume (Wolfson et al., 2005) and ventricular volume (Rosano et al., 2005b) to predict gait changes; larger volumes of each of these at baseline were predictive of a greater decline in gait velocity over time.
Three studies have assessed WMH presence with characteristics from the rhythm domain of gait. Again, findings within the domain were inconsistent with each other. Whilst two studies found no association between WMH presence and cadence (Nadkarni et al., 2009;de Laat et al., 2011a), the other found that increased global WMH presence, and WMHs within the brainstem, were associated with increased double support time (Rosano et al., 2005a).
Two studies have assessed WMH presence with gait characteristics from the variability domain. One study found that increased WMHs were associated with an increase in step length variability, but not with step width variability (Rosano et al., 2007b). The other study found no association between WMH presence and step length variability (Rosso et al., 2014). No gait characteristics within the asymmetry domain have yet been assessed with WMH presence. Two studies assessed WMHs with step width, from the postural control domain of gait, and found conflicting results. Whilst one study found that increased WMH presence was associated with wider steps in several brain regions (de Laat et al., 2011a), the other found that increased WMH presence was associated with narrower steps within the basal ganglia (Nadkarni et al., 2009 (Rosano et al., 2005a, b;Rosano et al., 2007b;Choi et al., 2012;Stijntjes et al., 2016). Each of these studies assessed gait characteristics from the pace domain; all four studies assessing gait velocity found that slower gait related to a higher infarct presence (Rosano et al., 2005a, b;Choi et al., 2012;Stijntjes et al., 2016), one of which reported that infarcts in the basal ganglia were associated with gait velocity (Rosano et al., 2005a). Both studies assessing step length associated shorter steps with an increased infarct presence (Rosano et al., 2005a;Choi et al., 2012). One study assessed stance time variability, which was not associated with either the number of total or basal ganglia infarcts (Rosano et al., 2007b).
Associations between an infarct presence and characteristics from the other gait domains are sparse and inconclusive. Two studies assessed infarct presence with double support time, yet only one found that increased infarct presence associated with increased double support time (Choi et al., 2012). This study also found that a higher presence of infarcts was associated with increased step width, but not cadence. The only study to investigate gait characteristics from the variability domain found that an increased infarct presence, both globally and within the basal ganglia, was associated with step length variability, but not step width variability (Rosano et al., 2007b).

Microbleed presence.
Three studies assessed a presence of microbleeds with gait (de Laat et al., 2011b;Choi et al., 2012;Stijntjes et al., 2016). All three studies assessed gait velocity, yet only two found that an increased microbleed presence was associated with reduced gait velocity (de Laat et al., 2011b;Stijntjes et al., 2016). De Laat et al. specified that this association was made not only with a global presence of microbleeds, but with microbleeds within the temporal lobe (de Laat et al., 2011b). Two of the three studies assessed additional gait characteristics (de Laat et al., 2011b;Choi et al., 2012). Both found that an increased presence of microbleeds was associated with increased double support time, but their results conflicted whilst assessing associations with step length, cadence and step width.

Longitudinal associations from analysis of white matter change
3.3.2.2.1. WMH presence. WMH presence was the most investigated imaging parameter longitudinally; seven studies longitudinally assessed WMHs with gait (Rosano et al., 2005b;Soumaré et al., 2009;Moscufo et al., 2012;Callisaya et al., 2013;Willey et al., 2013;Wolfson et al., 2013;van der Holst et al., 2018). All seven studies assessed WMHs with gait velocity. In keeping with findings from cross-sectional evaluations, three of four studies that assessed baseline WMH presence with change in gait velocity found that increased WMH presence at baseline was associated with a greater decline in gait velocity over time (Rosano et al., 2005b;Soumaré et al., 2009;Willey et al., 2013). However, only one of three studies assessing whether a change in the number of WMHs over time was associated with a longitudinal change in gait velocity found that a higher accumulation of WMHs over time was associated with a greater decline in gait velocity (Wolfson et al., 2013). Two studies assessed a change in WMH presence with a change in step length; one found that increased WMH presence over time was associated with slowing of gait velocity over time (Callisaya et al., 2013), whereas the other found no association (van der Holst et al., 2018). Neither of these studies found an association between a change in the number of WMHs and a change in cadence.

Infarct and microbleed presence.
Two studies assessed the longitudinal associations between gait and infarct and/or microbleed presence. One study found that a higher presence of infarcts at baseline was associated with a greater decline in gait velocity (Rosano et al., 2005b), whereas the other found no association between a change in either infarct or microbleed presence and a change in gait velocity, step length or cadence (van der Holst et al., 2018).
3.3.3.1.1. Fractional anisotropy. All studies utilizing DTI determined fractional anisotropy (FA); high values of FA indicated a good integrity of the tracts assessed. Five investigated gait characteristics from the pace domain; three of these five found that slower gait velocity was associated with reduced FA (de Laat et al., 2011a, c;Verlinden et al., 2016) in most white matter tracts, except for the brainstem, in one study (Verlinden et al., 2016) and expect for periventricular regions in another (de Laat et al., 2011c). Both studies assessing step length found that shorter steps were associated with FA (de Laat et al., 2011a, c), and the one study that assessed step time variability found that increased variability was associated with reduced FA (de Laat et al., 2011c). Two studies assessed gait characteristics from the rhythm domain (de Laat et al., 2011a, c); both found that less cadence was associated with reduced FA. Only one of these studies assessed double support time, and found that it was not associated with FA (de Laat et al., 2011c). Two studies assessed characteristics from the variability domain. Whilst one found that increased step length variability was associated with reduced FA within grey matter (Rosso et al., 2014), the other found no association between FA in white matter and either step length variability or step width variability (de Laat et al., 2011c). The recent study from Fling et al. was the only one to assess step time asymmetry, from the asymmetry domain, or step length asymmetry from the postural control domain (Fling et al., 2018). Neither characteristic was associated with FA. Four studies assessed step width with FA, and their results conflicted. Two studies found that wider steps were associated with reduced FA (de Laat et al., 2011a, c), one found that narrower steps associated with reduced FA (Bruijn et al., 2014) and the other study found no association between the parameters (Fling et al., 2016).

Mean, radial and axial diffusivity.
Four studies additionally used measures of mean diffusivity (MD) within white matter tract, where high MD is a sign of poor white matter integrity (Della Nave et al., 2007;de Laat et al., 2011a, c;Verlinden et al., 2016). Three studies found that increased MD was associated with both slower gait velocity and shorter step length from the pace domain, and with increased step length variability from the variability domain (de Laat et al., 2011a, c;Verlinden et al., 2016). Associations with MD were made in more regions than those with FA assessment; associations were additionally made in subcortical frontal and parietal regions in one study (de Laat et al., 2011c), and within the brainstem in another study (Verlinden et al., 2016). Two studies assessed additional gait characteristics (de Laat et al., 2011a, c); both found that increased MD was associated with reduced cadence from the rhythm domain and increased step width from the postural control domain. Additionally, one of these studies identified that increased MD was associated with increased step time variability (pace) and increased double support time (rhythm), but was not associated with step width variability (variability) (de Laat et al., 2011c). Another study found that increased MD was associated with short stance time (rhythm) (Verlinden et al., 2016). Two of the studies assessing MD additionally reported associations relating to radial diffusivity (RD) and axial diffusivity (AD) (de Laat et al., 2011a;Verlinden et al., 2016), which related to gait in the same way as MD in both studies.

Longitudinal associations with DTI parameters.
One study utilised DTI (van der Holst et al., 2018) to determine longitudinal neural imaging correlates of gait. Van der Holst et al. found that changes in DTI parameters over time were not associated with decline in gait velocity or cadence, but that a decrease in FA over time was associated with a reduction in step length over time (van der Holst et al., 2018). Similarly, reduced step length over time was associated with increased MD and RD over time, but not increased AD.

PET 3.3.4.1. Cross-sectional associations with PET imaging.
Four studies assessed amyloid beta, Aβ, burden through PET imaging (del Campo et al. (2016);Nadkarni et al., 2017;Tian et al., 2017b;Wennberg et al., 2017a). Three of these utilised [ 11 C] Pittsburgh Compound B (PiB) PET scans (Nadkarni et al., 2017;Tian et al., 2017b;Wennberg et al., 2017a), whereas del Campo et al. used [18 F] Florbetapir PET scans (del Campo et al., 2016. Three of the studies found that an increased burden of Aβ was associated with reduced gait velocity (del Campo et al., 2016;Nadkarni et al., 2017;Wennberg et al., 2017a). Specifically, Aβ burden within basal ganglia regions, the precuneus, the temporal cortex and the anterior cingulate was negatively associated with gait velocity in more than one of these studies. No association was identified between gait velocity and Aβ burden within the hippocampus, pons, thalamus or parietal lobe. Additionally, Wennberg et al. (Wennberg et al., 2017a) found that increased Aβ burden was associated with increased stance time variability, reduced cadence and increased double support time, in prefrontal, temporal and cingulate regions; no associations were made between step length and Aβ burden.
Three studies used FDG PET imaging to determine cerebral glucose uptake (Shimada et al., 2013;Sakurai et al., 2014Sakurai et al., , 2017. All three of these studies recruited only female participants from similar databases of elderly volunteers, limiting an overall interpretation of the associations between cerebral glucose uptake and gait in the general population. Both of the studies authored by Sakurai et al. (Sakurai et al., 2014;Sakurai et al., 2017) found that lower uptake of glucose was associated with reduced gait speed and reduced cadence, but not step length. Associations were made in the posterior cingulate and parietal cortex in both studies; other regions were investigated, but either no association was made or the associations was not consistent across both studies. These associations were made with gait characteristics assessed during fast walking only; no associations were made between glucose uptake and gait characteristics measured at a comfortable walking pace. In the study from Shimada et al. (Shimada et al., 2013), participants were split in to two groups, those with low step length variability (LSV) and those with high step length variability (HSV). The LSV group had relatively increased glucose uptake in the primary sensorimotor area in comparison to the HSV group, and the HSV group had comparatively decreased uptake in the middle and superior temporal gyrus and hippocampus in relation to the LSV group; note that the direction of association is consistent in both sets of analyses. In this study, step length variability was assessed during walking at a fixed pace of 2 km/h, a relatively slow walking speed (Bohannon and Williams Andrews, 2011).

Longitudinal associations with PET imaging.
One study utilised PET imaging to determine longitudinal neural gait correlates. Gait velocity change was the only gait characteristic assessed alongside baseline PET imaging; Tian et al. found that higher baseline Aβ burden was associated with a greater slowing of gait velocity over time (Tian et al., 2017b).

fNIRS, MRS and fMRI
Only one study using each of fNIRS , fMRI (Yuan et al., 2015) and MRS (Zimmerman et al., 2009) imaging techniques in relation to single-task gait characteristics were identified in this review. All studies were performed in cross-section, and assessed gait characteristics from the pace domain. These studies identified that gait velocity was associated with functional connectivity in sensorimotor, visual, vestibular, and left fronto-parietal cortical areas (Yuan et al., 2015), but not with activation strength of the prefrontal cortex , and stride length was not associated with either prefrontal cortex activation or N-acetylaspartate:creatine ratio within the hippocampus (Zimmerman et al., 2009). However, the sparsity of studies focussed on these approaches and gait means that no firm conclusions can be drawn about associations between gait and parameters derived from these functional imaging techniques.

The effect of cognition on the relationship between gait and imaging correlates
The impact of cognition on the relationship between gait and imaging correlates was considered, through assessment of studies which assessed cognition and included cognitive test scores as covariates. Twelve studies within this review included a measure of cognition as a potential confounding factor (Rosano et al., 2008;Dumurgier et al., 2012;Annweiler et al., 2014;Bolandzadeh et al., 2014;Nadkarni et al., 2014;Rosso et al., 2014;Sakurai et al., 2014;Ezzati et al., 2015;Stijntjes et al., 2016;Beauchet et al., 2017;Nadkarni et al., 2017;Tian et al., 2017b). These cognitive measures included global and primarily executive outcomes: the mini mental state exam (MMSE), the modified mini mental state exam (3MS), trail making tasks A and B (TMT-A, TMT-B), the digit symbol substitution test (DSST), the frontal assessment battery (FAB) and the abbreviated Stroop test, although one study specifically included a free recall score task (Stijntjes et al., 2016).
Gait velocity was frequently assessed as a measure in these studies. Adjusting for cognition caused gait velocity to no longer be associated with frontal and parietal lobe GM volume (Dumurgier et al., 2012), hippocampal volume (Ezzati et al., 2015), cerebellar volume (Nadkarni et al., 2014), amyloid beta burden (Nadkarni et al., 2017), WMHs in the anterior thalamic radiation and anterior corpus callosum (Bolandzadeh et al., 2014) and almost all regional associations with baseline amyloid beta burden (Tian et al., 2017b). In contrast, gait velocity remained associated with some imaging parameters after the inclusion of cognition as a covariate, in particular, with basal ganglia and caudate volumes (Dumurgier et al., 2012) as well as global microbleed and infarct presence (Stijntjes et al., 2016).
The effect of cognition on the neural correlates of other gait characteristics were also considered. Associations between regional grey matter volume and step length, stance time and step width (Rosano et al., 2008) were largely unaffected by the inclusion of cognition in statistical models, as were associations made between mean diffusivity in grey matter and step length variability (Rosso et al., 2014) and the association between step time variability and ventricular volume (Annweiler et al., 2014).

Discussion
To our knowledge, this is the first structured review to comprehensively map many discrete gait characteristics, defined by a validated objective gait model, to their structural and functional imaging neural correlates. We identified many associations between gait characteristics and the brain, however, our findings demonstrate a limited understanding of the overall neural control of gait. This is as a result of studies mostly reporting associations with gait velocity, and the majority of studies focussing on structural, rather than functional, neuroimaging parameters. Additionally, network-based approaches were scarcely utilised within the neuroimaging methodologies, despite the importance of networks becoming increasingly evident (O'Sullivan et al., 2001). There is emerging evidence linking other gait characteristics to a wider array of neuroimaging parameters, and of the effect of the ageing process on associations through longitudinal observations, although this remains limited.

Global neuroimaging correlates of discrete gait characteristics
The overarching findings from the studies included in this review is that a 'healthy brain' is associated with better gait. Larger volumes of healthy grey matter and WM integrity, determined through WMH presence in addition to FA and MD, were consistently associated with quality of gait performance, as demonstrated by gait characteristics related to pace, rhythm, variability and postural control. Increased step width, from the postural control domain, was associated with both greater and less white matter integrity in different studies, consistent with the idea that optimal step width is neither too wide nor too narrow. Although wide steps provide greater stability (Gabell and Nayak, 1984), they could indicate compensation for poor balance control. Non-linear optimisation of step width may also begin to explain why step width variability has not yet been associated with any imaging parameter.
The global presence of common white matter changes in ageing was associated with worse gait. Several of the studies in this review found that associations between gait and either brain volumes or DTI parameters were no longer significant once the presence of white matter changes such as WMH, infarct or microbleed presence were included as covariates, either in isolation or with a combination of other confounding factors. Some studies specified the presence of small vessel disease (through WMH or infarct presence) within their inclusion criteria (de Laat et al., 2011a, b, c;de Laat et al., 2012;van der Holst et al., 2018), and none of the studies included assessments related to ageing factors such as frailty or muscle mass. Overall, this causes great difficulty in discerning whether global neural gait correlates are due only to the presence of white matter artefacts, or if in fact these artefacts are an unrelated sign of ageing, and we are simply identifying associations between ageing and gait. Clearer assessment of the direct effects of agerelated white matter changes on associations between gait and the brain should be completed in future to resolve this issue. The global imaging correlates of gait presented here suggest that gait may not controlled by discrete regions, instead many different regions may be working harmoniously for effective gait.

Regional neural correlates of gait
In addition to global associations between gait and the brain, we aimed to identify regional neural structures and/or functions which related to specific gait characteristics. Studies which have attempted regional analyses were too disparate in approach to allow any firm conclusions to be drawn. A number of studies have focussed only on one region and most studies have each investigated different regions. Very few studies within this review assessed regions connected together through a network-based approach, either through the assessment of functional networks (Yuan et al., 2015) or by assessing the specific structural tracts which connect different brain regions (Verlinden et al., 2016). We have identified several trends of gait domains relating to specific brain regions; it is important to highlight, however, that greater efforts should now be made to assess these regions together as one or several networks which may be responsible for gait control. Within the pace domain, the most consistent trends were towards an association of gait velocity with GM volume in frontal, basal ganglia, hippocampal and cerebellar regions, and with WMHs within frontal regions, the basal ganglia and the corpus callosum, although most brain regions have been associated with gait velocity in at least one study.
Step length was consistently associated with several GM volumes, however, not all the regions associated with gait velocity also related to step length. This suggests specificity in findings; gait velocity, as the global measure of gait, may be associated with global brain features, whereas step length may be related more so to cortical, rather than subcortical brain regions. For gait characteristics from the other domains, few studies specified the regions in which associations were made, and most regions were identified in only one study. Fig. 5 shows maps of the regional associations between GM volume and all gait characteristics in which GM regions were specified; gait velocity, step length, step time variability, step width, cadence and double support time. The entirety of the gait velocity maps are in colour, to indicate that most brain regions have been associated with gait velocity in at least one study; regions in which several associations have been made have additionally been highlighted.
It is clear from Fig. 5 that some brain regional volumes have been associated with several gait characteristics. This may be due to the volumes assessed being too crude, which would not allow for the specificity of associations between smaller, better defined, regions and gait to be apparent. Alternatively, it may be that brain regions are responsible for several of the gait characteristics, which could be grouped in to their own region-dependent domains in future. Before this can be clarified, however, further regional analyses assessing many of the gait characteristics through the same methodologies are required.
In studies assessing infarct and microbleed presence, the topographic location of these brain insults were rarely reported, although from the few studies that did detail regional associations it appears that insults within the basal ganglia have a negative impact on gait. Similarly, amyloid beta burden within the basal ganglia was associated with gait velocity. It is not yet clear whether these disturbances to the basal ganglia impact on all aspects of gait; it may be hypothesised that the basal ganglia is responsible for less complex gait characteristics, such as step length, as it is a lower level structure, although there is currently not enough literature to determine this with any certainty.
Two other reviews have recently been published which sought to determine neural gait correlates. Tian et al. focussed on associations made with variability characteristics of gait, and identified a particular association with temporal variability measures and the right hemisphere (Tian et al., 2017a). Here, we identified that the right hemisphere may also have a greater responsibility for double support time (Rosano et al., 2008), whereas cadence may be more dependent on the left hemisphere (de Laat et al., 2012); this asymmetry gives further evidence for specificity in the brain regions responsible for gait, and should be considered further. Wennberg et al. found that frontal and parietal regions of grey matter were most commonly associated with gait, as measured through more complex tests such as the timed up-andgo in addition to some of the gait characteristics identified here (Wennberg et al., 2017b)., Here we additionally found that the involvement of subcortical areas such as the basal ganglia and limbic system, as well as the hippocampus, were related to several gait characteristics. Several of the gait characteristics appeared to rely on involvement from both types of brain area, strengthening the argument that brain areas associated with both motor tasks and cognition are heavily involved in coordinating gait.
This review has demonstrated that, although many brain regions have been assessed with gait, there is a lack of cohesion between studies about the brain areas of most importance in gait. The non-specificity of findings between gait and structural volumes suggests that the impact of gait impairment on neural networks, connected either structurally or functionally, should be considered in future.

Longitudinal neural correlates of gait
A reduction in gait characteristics from the pace domain over time was not only associated with changes in WM volume over time, but also with baseline measures of WM volume. However, WM volume was not associated with gait during cross-sectional analyses. This highlights the importance of longitudinal study types when assessing associations Fig. 5. Map of the regional associations between GM volume and gait characteristics; gait velocity (A), step length (B), step time variability (C), step width (D), cadence (E) and double support time (F). Areas which are darker in colour indicate regions that were associated with the characteristic in multiple studies. Panel A shows the entire brain in an orange colour, to indicate that the volume of most brain regions have been associated with gait velocity. between gait and the ageing brain, yet relatively few studies assessed the longitudinal neural correlates of gait. In general, we identified more associations between gait changes over time and imaging parameters when imaging was used as a predictive measure of gait decline, rather than where changes in imaging parameters were related to gait changes. This may indicate that predictive models utilising only one set of brain imaging data should be considered more regularly in future, particularly given that gait assessment is relatively inexpensive and is more accessible to older adults than imaging. Clinically, improved understanding of the brain changes associated with a poorer gait performance would allow us to utilise gait assessment more readily as a predictive measure for future neurodegeneration.

Methodological critiques
Most studies included in this review utilised structural MRI and/or DTI protocols, perhaps due to the major issue of cost surrounding functional imaging techniques such as PET and fNIRS and the intensity of task-based fMRI studies. Most studies that have utilised fMRI to assess associations between the brain and gait have done so through finding the neural correlates of imagined walking and leg movements whilst in an fMRI scanner. These were not considered in this review, as these protocols do not directly explore discrete gait characteristics. The use of virtual reality and foot pedals within functional protocols may provide further insight in to the functional networks utilised during walking, particularly as recent work has started to produce surrogate measures of gait characteristics such as step time variability (Gilat et al., 2017). The recent development of new radiotracer elements and enhanced PET scanning techniques may contribute to our understanding of neural activity associated with gait. For example, in healthy adults aged 21-85 years, gait velocity, cadence, and stance time have been slowed in those with lower striatal dopamine activity (Cham et al., 2008), and in PD, acetylcholinesterase ([11C]PMP) PET has been used to show that cholinergic deficit is associated with slow gait (Müller et al., 2015). Additionally, nerve stimulation can be used to assess cholinergic activity through assessment of short-latency afferent inhibition (Rochester et al., 2012); consideration of nerve or brain stimulation was beyond the scope of this review. To the best of our knowledge, no studies have utilised functional imaging of neurotransmitter activity in a group of healthy older adults to assess neural correlates of gait. Four of the seven studies which used PET imaging in this review assessed amyloid burden which, in effect, is a marker of structural rather than functional pathology and does not enhance our understanding of the functional neurochemical correlates underlying gait impairments. Although the assessment of CSF biomarkers through lumbar puncture was not covered here, studies of this nature can provide us with additional information about the effects of an accumulation of pathologies such as amyloid and tau on motor performance .
Only one study within this review completed brain imaging in real time during the assessment of discrete gait characteristics. This is due to most fNIRS and EEG studies comparing differences in brain states between tasks, such as between single and dual task gait, rather than assessing single task gait characteristics (Hamacher et al., 2015;Vitorio et al., 2017). Electrophysiological responses, as measured through EEG, are consistent with haemodynamic responses assessed through both fNIRS and fMRI (Anwar et al., 2016), demonstrating their reliability as promising technologies for future research in to the functional neural correlates of gait. Additionally, functional imaging both during realtime gait assessment and during the resting state can be used to assess neural networks; it is becoming increasingly evident that network analyses, particularly those of a dynamic rather than static nature, will be key to furthering our understanding of the system-wide neural substrates which underpin dynamic gait control. There is evidence of an interaction between gait velocity and inter-network connectivity between the default mode network and supplementary motor as assessed through fMRI in older adults with mild cognitive impairment (Crockett et al., 2017), yet no study has assessed this in a healthy ageing population; these areas could provide a starting point for the neural networks to assess in association with gait. Decoupling between imaging and gait measurements introduces the potential for increased variance and noise, undermining potential correlation. Although structural imaging approaches are likely to be robust to this, functional imaging may be more sensitive due to its reliance on brain state and performance. To avoid this issue, further efforts should be made to develop strong protocols assessing real-time brain function with single task gait characteristics.
Several of the gait protocols within this review included use of a stopwatch during corridor walking. Although measures of gait velocity can be calculated through this technique, and its components in some instances, more subtle gait characteristics cannot be assessed. This limits our understanding of the neural control of variability and asymmetry; measures within the variability domain particularly have a high prevalence within gait research as they can be used as markers of fall risk and cognitive decline (Montero-Odasso et al., 2012;Tian et al., 2017a), therefore it is crucial that we develop a greater understanding of their underlying mechanisms. Additionally, non-linear approaches to analysis, such as the fractal analysis of stride-to-stride fluctuations in walking, are of increasing interest and prevalence within human gait research as they take in to consideration the structure and complexity of these large data sets Li et al., 2018). Studies within this review varied not only by the gait measurement tools, but by the number, speed and type (continuous or intermittent) of walks performed, limiting our interpretation of findings. If a standard robust single-task gait protocol were developed and used, findings would be more comparable between studies, and standardised reporting of characteristics other than gait velocity would be carried out more frequently.

Cognition as a covariate
Cognition had a mediating effect on some of the associations made between gait velocity and volumes of the frontal lobe and hippocampus. Executive function (the function of which is subsumed by frontal regions) and memory (controlled by the hippocampus as well as portions of the prefrontal cortex) have both been associated with gait velocity (Watson et al., 2010;Morris et al., 2016), therefore suggesting a broad three-way interplay between cognitive function, gait velocity and grey matter volume. It is currently unclear whether gait velocity is directly impacted by cognition, or whether impairments in both gait and cognition are as a result of changes within the brain. Our understanding of the interaction between gait, cognition and the brain, and whether it applies to gait characteristics other than velocity, is limited due to the relative scarcity of studies assessing cognition in addition to gait and neuroimaging parameters.

Current limitations and recommendations for future work
The literature is currently dominated by assessments of gait velocity, most likely due to its ease of measurement. Also, it is frequently assessed alongside other motor parameters in studies assessing "mobility" as opposed to pure gait. Gait velocity is a global measure of gait, which has been assessed through a wide range of techniques; its lack of specificity cannot reflect subtle gait changes that occur during ageing and disease, and its use increases the likelihood of chance findings and different clinical interpretations of results (Graham et al., 2008). There is a relative scarcity of studies assessing gait characteristics other than gait velocity, therefore it cannot be concluded with any certainty whether characteristics within the same gait domain have similar neural underpinnings. Similarities in the grey matter regions associated with gait velocity and step length from the pace domain indicate that these may involve similar mechanisms, although it should be emphasised that this may be due to the high correlation between these characteristics. Until such a time that the neural correlates of several characteristics from each of the gait domains are well understood, to confirm whether characteristics within each domain are controlled in a similar manner, we should not attempt to reduce the number of gait characteristics to be associated with imaging parameters moving forward.
The literature mostly consists of assessments of neural structure rather than function, and the specific regions assessed differed across most studies, suggesting a high degree of heterogeneity. Associations between neuroimaging parameters and gait characteristics were typically conducted through correlational approaches. Correlations were not necessarily reflective of causation and can be affected more strongly by sample size; therefore any identified correlation should be interpreted with some degree of caution. It is justifiable to perform correlational analyses in the first instance, particularly given our limited understanding of the neural control of gait. However, correlations alone do not allow for a strong understanding of the neural control of gait; moving forward it is crucial to develop methodologies which allow for the concurrent measurement of discrete gait characteristics and brain functionality with a good degree of spatial resolution to further our knowledge of gait control.
The mini-mental state examination (MMSE) was most commonly used as an assessment of cognition (where cognitive tests were completed); by using an assessment of global cognition, it is difficult to discern whether the neural substrates associated with different cognitive domains, such as frontal regions with executive function, match those neural areas associated with gait. Confounding factors included in analyses generally differed between studies. Additionally, some papers highlighted results from unadjusted models whereas others only reported results from models including a full set of confounding factors. This may have caused some discrepancy when comparing studies. Gait assessed under dual-task conditions has not been considered, which may have moderated our findings. We feel that a more robust understanding of the influence of cognition on neural gait correlates is required before findings from dual-task protocols can be sufficiently interpreted. A limitation of this review is that we have focussed on the neural correlates of gait in healthy older adults. We could have found a greater number of associations if young or middle-aged adults were included, which may not have been confounded by age. We were confident in restricting our search criteria to older adults, so that we could clearly present the current understanding of gait mechanisms during typical ageing, which should better relate to the mechanisms underpinning gait dysfunction in disease. Nonetheless, it is important to highlight that the mechanisms of gait during earlier life stages may not be fully encompassed by the neural gait correlates presented here.
Overall, most of the studies considered within this review were only of average quality (see supplementary Table 1). We have therefore summarised some key recommendations for future studies within this field, see Fig. 6.

Conclusion
In summary, our structured review has demonstrated that global imaging markers of a 'deteriorating brain', namely grey matter atrophy, high volume of white matter hyperintensity lesions, and worsening of diffusion tensor imaging measures of white matter integrity, are correlated with poor gait performance. Additionally, we found that gait velocity decline over time can be predicted by an initial global assessment of imaging markers of white matter. Regionally, we identified a predilection for both the volume of and white matter hyperintensity presence within frontal and basal ganglia regions to influence gait velocity, although many brain regions have been related to gait velocity in some capacity. Beyond this our conclusions are more limited, largely due to the small range of discrete gait characteristics included within each study design, as well as a lack of consistency in the brain regions investigated between studies. We hypothesised that discrete gait characteristics may have more discrete neural correlates; although global associations have been more concretely assessed, there is an emerging specificity of associations between gait and the brain, evidenced by differences in the neural regions that have been associated with different gait characteristics thus far. This review has also demonstrated a relative scarcity of functional imaging correlates with formal gait measures; only one study utilised resting state functional networks in analyses, and only one functional near infra-red spectroscopy study has identified neural gait correlates through single task walking. This is somewhat surprising given the dynamic nature of gait, and given that the association between gait decline and a loss of generalised white matter integrity highlighted in this review points strongly towards an association between gait decline and reduced functional connectivity. Further work should take greater consideration of the covariates included within analyses, and particularly assess the influence of cognition on any associations found, so that a strong model of the three-way interplay between gait, cognition and the brain can be developed.

Conflicts of interest/disclosure statement
The authors have no conflict of interest to report.