Quiet standing postural control variables in subacute stroke: associations with gait and balance, falls prediction and responsiveness

Abstract Purpose To examine the construct validity, predictive validity and responsiveness of standing centre of pressure variables in subacute stroke. Materials and methods Seventy-nine ambulatory individuals were assessed before inpatient rehabilitation discharge and three months later. Measures were: gait speed (6-metre walk), dynamic balance (step test), and quiet standing (Wii Balance Board). Centre of pressure speed, amplitude, standard deviation, root mean square, wavelet decomposition, and detrended fluctuation analysis were examined. Falls data were collected over a 12-month period post-discharge. Results Moderate strength correlations (r = −0.505 to −0.548) with gait speed and step test scores were shown for 3/26 centre of pressure variables (mediolateral speed, low and moderate frequency wavelet). Twenty-two participants fell and the prediction was significant for gait speed and step test (IQR-odds ratio (OR) = 4.00 & 3.21) and 3/26 centre of pressure variables (mediolateral low-frequency wavelet: IQR-OR = 2.71; mediolateral detrended fluctuation analysis: IQR-OR = 3.06; anteroposterior detrended fluctuation analysis: IQR-OR = 2.71). Significant changes over time occurred for gait speed and step test scores and 20/26 centre of pressure variables. Conclusions Standing centre of pressure variables have limited validity to reflect dynamic balance and falls risk after stroke. Frequency and complexity measures warrant further exploration. Implications for rehabilitation Our findings indicate that quiet standing centre of pressure variables have limited validity to reflect dynamic balance tasks and predict falls after stroke. The mediolateral and higher frequency variables may be more strongly recommended than the commonly used total centre of pressure speed measure. Measures of signal frequency and complexity may provide insight into postural control mechanisms and how these change over time following stroke.


Introduction
Balance is frequently impaired in individuals with stroke and is associated with falls, injury and long-term disability [1-3]. A range of clinical and instrumented tools may be used to assess balance ability [3,4]. Centre of pressure (COP) variables, obtained using force platform technologies, have been used to evaluate treatment efficacy in research trials, and obtain information on balance mechanisms, recovery and risk of falls following stroke [5][6][7]. However, a large range of variables may be obtained from this type of assessment and there is currently no strong evidence to support the validity for falls prediction and responsiveness over time. The use of COP data has been largely limited to research settings as this method of assessment has practical limitations and there are gaps in our knowledge on the utility of the data [7].
While the selection of an assessment tool is dependent on the purpose and the individual's characteristics, measures of balance most strongly recommended for use after stroke include the Berg Balance Scale, Timed Up and Go (TUG), Fugl-Meyer Assessment and Postural Assessment Scale for Stroke [8][9][10]. Force platformbased assessments have been used to assess both quiet standing and dynamic balance activities. This may enable a more precise, quantitative assessment and provide insight into postural control mechanisms [7]. The construct validity of COP variables obtained from quiet standing balance testing post-stroke has demonstrated mixed results, with typically low-to-moderate correlations to other balance tests such as the Berg Balance Scale [11,12], step test [13], and TUG [12,13]. Research has shown significant differences between people with stroke who have or have not fallen (either once or more than once), for mediolateral velocity (i.e., speed) standard deviation (SD) and sway area [14,15]. However, other research found no significant difference between people with stroke who fell or not for COP speed measures [16]. The evidence for validating COP measures is further complicated by a lack of standardisation in testing protocols and differences in the types of variables obtained.
Outcome measures in research and clinical practice are commonly employed to detect changes in performance over time or in response to interventions. Significant changes in COP variables have been demonstrated following balance training interventions in people with stroke [17,18]. Individuals undertaking post-stroke rehabilitation have also shown improvements in COP measures [e.g., 19,20], and this may be more pronounced in the mediolateral direction [6]. Specific information on balance deficits provided by COP variables may be useful to inform targeted training approaches and provide a sensitive measure to detect changes in balance. However, reliability studies of COP variables in people with stroke indicate that relatively large changes may be needed to be confident that true change has occurred, with minimal detectable change (MDC) in COP variable scores of between 38 and 52% [13,21,22]. Further, therapists have discussed the limitations in the feasibility of using force platform technologies when compared to tests requiring minimal time and equipment [23]. Alternative devices such as the Wii Balance Board (WBB), with greater portability and lower cost, may have increased clinical utility [24].
A comprehensive investigation of the measurement properties of COP variables will help to guide future research to better inform the appropriate variable selection, provide insights into the mechanisms of standing balance recovery, and facilitate more targeted and effective treatment approaches. Our study, therefore, aimed to determine: (1) construct validity, through the association of individual COP variables with gait speed (6-metre walk) and dynamic balance (step test); (2) the predictive validity for falls over 12 months; and (3) responsiveness, through the examination of change over three months post-discharge from inpatient rehabilitation. Secondary analysis included examining the strength of association of individual COP variables with each other.

Materials and methods
This study included data from 79 individuals consecutively recruited from inpatient rehabilitation facilities within Australia (n ¼ 24) and Singapore (n ¼ 55) between 2014 and 2016. This data was a subsequent analysis from research primarily investigating variables associated with falls and physical activity after stroke [16,25]. Eligibility criteria included: being in the early subacute phase (i.e., less than three months) post-stroke [26]; the ability to walk 10 m with no more than minimal assistance with or without gait aids; no other medical condition which could confound testing (e.g., progressive neurological disease); and no severe cognitive or language deficits. The study was approved by the institutions' ethics committees (Epworth HealthCare 643-14; Singapore General Hospital 2015/2010) and all participants provided written informed consent.
Participants undertook testing within one week prior to discharge and three months later. Assessment sessions were conducted by one of three trained physiotherapists or exercise physiologists. Where possible, the same assessor conducted both the first and second sessions for each individual. Baseline data also included the Functional Independence Measure [27] and Modified Rankin Scale [28], as assessed at inpatient rehabilitation discharge by the treating team and research team respectively. Prior falls were assessed by asking individuals whether they had fallen in the 12-month period prior to having their stroke.
Quiet standing balance data were collected using a WBB and customised software calibrated in accordance with Clark et al. [29]. Participants stood barefoot on a single WBB (heels 17 cm apart, toe-out 14 � [30], with eyes open and focused on a target at a distance of 2.2 m and height of 1 m. The standardisation of foot position was marked with lines indicating the heel width and toeout angles, and additional coloured images of feet adhered to the top of the WBB. Assistance was provided if needed to place the feet appropriately. Once in position, participants were asked to stand "as still as possible" for a duration of 30 s. The mean of two trials was used. This has been suggested to be sufficient for reliability in people with stroke [21] and reduced the testing burden for participants. Data were sampled from each load cell independently at their native frequency (around 40 Hz), lost packets (which occur in < 5% of trials and cause erroneous single point data spikes) were removed by replacement with the average of the samples immediately before and after the missing packet, lowpass filtered at 12.5 Hz and then converted to a COP coordinate. Data were then lowpass filtered again at 6.25 Hz using a Symlet-8 wavelet before being resampled to 100 Hz using spline interpolation [31] to enable a constant sampling rate and calculate cascading wavelet bands. WBB-derived variables have demonstrated excellent concurrent validity with other force platforms [24] and excellent testretest reliability in individuals with stroke (ICC ¼ 0.87-0.94) [13].
The COP variables were selected based on their potential to identify different mechanistic aspects of balance control. These COP variables were chosen as they are commonly reported and/ or represent measurement across different categories (i.e., speed, distance, area, frequency, complexity), and maybe relatively easily computed and interpreted. The variables were calculated in both the mediolateral (ML) and anteroposterior (AP) axes, which are one-dimensional, with two-dimensional total scores (i.e., movement on both axes combined calculated using trigonometry) reported for overall speed and variability (i.e., standard deviation). The variables were: 1. Path speed (cm/s). Total COP distance divided by trial duration. 2. Peak displacement amplitude (cm). Calculated as the distance between the highest and lowest values on each axis. Higher amplitudes indicate a greater sway range. 3. Standard deviation (SD) of the COP path (cm). Higher values indicate greater exploration of the mean COP position. 4. Root-mean-square (RMS) displacement (cm). Quantifies the mean magnitude of the absolute COP distance from the starting position. If the mean position of the COP trace is 0 it is analogous to SD, however the further the mean is from 0 the more it differs from SD. Therefore, similar to SD it quantifies greater exploration of the COP, but this effect is augmented by underlying trends in the movement. 5. Wavelet decomposition path length (cm). The trace is split into moderate (i.e., fast-moving; 3.13-6.25Hz), low (1.56-3.13Hz), very low (0.78-1.56Hz) and ultralow (<0.78Hz) frequency signals using a 9-level Symlet-8 wavelet [32], with the COP path length in each band recorded. 6. Detrended fluctuation analysis (DFA). This fractal measure examines the longer-term correlations in the trace to provide a measure of signal complexity [33]. DFA was calculated for small box sizes between 32 and 100 (DFA alpha 1) and larger box sizes above 100 (DFA alpha 2).
Gait speed was assessed using the 6-metre walk test (6mWT) [34]. Participants were asked to walk at a comfortable pace over a 10 m walkway, with the middle 6 m timed using a stopwatch and converted to m/s (mean of two trials). Shoes and gait aids could be used if needed for safety during this test. Dynamic balance was assessed using the step test [35]. Participants were asked to tap their foot on and off a 7.5 cm step without using support. Following one practice trial, one trial was performed. The number of taps completed over 15 s, with the more-affected leg in the supporting position (i.e., less-affected leg tapping), was recorded. Data on falls were collected over a 12-month period post-discharge via monthly calendars and follow-up phone calls [36]. A fall was defined as "an unexpected event in which the participant comes to rest on the ground, floor or lower level" [36]. A faller was defined as having one or more falls in the 12-month period. Further detail on falls data collection may be found in a prior paper [16].
Associations between COP variables and gait and dynamic balance tests were assessed using Spearman's rho (r). The strength of association was defined as excellent (r � 0.75), moderate (r ¼ 0.50-0.74), fair (r ¼ 0.25-0.49) or poor (r � 0.24) based on recommended thresholds [37]. Given a large number of proposed analyses, we determined the potential risk of erroneous findings. Performing one million iterations of correlations on random numbers with n ¼ 79 (our sample size) per array revealed that we could be > 95% confident that an absolute r value > 0.22 was not spurious, and that the odds of obtaining an absolute r value � 0.5 from these data were 0.0004%.
Differences between faller and non-faller groups for each variable were assessed using Mann-Whitney U tests. Each of the COP variables and clinical tests was included separately in an ordinal regression analysis, with falls as the dependent variable (i.e., predictive validity). The analysis was adjusted using the covariates of country, body mass and prior falls. These were selected to adjust for: (1) potential inter-country differences that may confound the results, (2) the influence of body mass on the centre of pressure data, and (3) pre-morbid function, with prior falls demonstrating a significant difference between faller and non-faller groups in this cohort [16].
Log transformation was used on variables found to have a significant positive skew (i.e., Shapiro-Wilk test < 0.05 and Skewness > 1). To enable a more clinically meaningful comparison between variables quantified on different scales, odds ratios (ORs) and 95% confidence intervals (CIs) were scaled to the interquartile range (IQR) of each predictor variable. Given the explorative nature of the study and the high inter-correlation between the COP measures, we did not adjust for multiplicity in our analyses.
Paired t-tests were used to detect significant changes over time. Responsiveness was calculated using: (1) standardised response mean (SRM), mean change score divided by its standard deviation; and (2) pooled effect size (ESp), SRM x ffi ffi ffi 2 p x ffi ffi ffi ffi ffi ffi ffi ffi ffi ffi 1 À r p where r is the correlation between the two time points [38].

Results
Participant characteristics are presented in Table 1. The mean age was 63 (SD 13) years and time post-stroke was 28 (SD 12) days. While all individuals were able to walk with no more than one person assisting, there was a wide variation in gait speed and step test scores. All 79 individuals had data collected at baseline and completed 12 months of falls follow-up. Seventy-three individuals had data from the 3-month post-discharge testing session. Data were missing due to non-attendance (n ¼ 4) or technical issues with the WBB data collection (n ¼ 2). There was a moderate strength association between the 6mWT and step test (r ¼ 0.729, p < 0.001). The strength of association between each COP variable and these tests is shown in Table 2. Moderate strength negative associations were demonstrated for ML speed (r ¼ À 0.548 and À 0.509 for 6mWT and step test respectively), ML moderate frequency wavelet (r ¼ À 0.512 and À 0.505) and ML low-frequency wavelet (À 0.517 for 6mWT). Detrended fluctuation analysis variables tended to have the poorest association with the gait speed and step tests. Ultralow wavelet variables also had no significant associations with these tests.
Over 12 months post-discharge, 22/79 (28%) of individuals fell at least once and 13/79 (16%) fell more than once. Further details on falls in this cohort can be found in a prior paper [16]. Of note, approximately half of the falls occurred inside the home, around one-third were related to stairs, and one fall resulted in hospital admission.
The step test, AP DFA alpha 2 and AP DFA speed alpha 2 were significantly different between the faller and non-faller groups (Table 3). After adjusting for the country, prior falls and body Presented as mean ± SD (range); a higher scores indicate better function; b lower scores indicate better function.  Table  4). Large responsiveness (> 0.80) [39] was seen for 3/26 (12%) COP variables calculated using the SRM. None of the COP variables demonstrated large responsiveness using the ESp. Greater responsiveness tended to be observed for COP speed and the higher frequency wavelet variables, with no clear trend for greater change in the AP, compared with the ML direction.
The associations between the COP variables ranged from poor to excellent (Supplementary Material 1). Of note, excellent strength associations were found between ML speed with ML amplitude, ML SD and the three highest frequency ML wavelet variables. Apart from excellent strength correlations with the two highest frequency AP wavelet variables, AP speed was less strongly associated with other AP variables (r ¼ 0.395-0.641). Compared to the higher frequency wavelet variables, the lowest frequency variables were less strongly associated with others, particularly those in the AP direction. DFA variables had mostly poor associations with other variables.

Discussion
This is the largest study to examine validity and responsiveness in a comprehensive range of quiet standing COP variables following stroke. Associations between these variables and tests of gait speed and dynamic balance ranged from non-significant to moderate in strength, with ML variables tending to show greater associations. Compared to gait speed and the step test, COP variables were generally less strongly predictive of falls, with only 3/26 variables examined demonstrating significance (ML low-frequency wavelet, ML DFA speed alpha 1 and AP DFA alpha 2). Responsiveness over three months post-discharge was more pronounced in some of the COP variables and these were larger than changes seen in gait speed and the step test. These findings highlight the limited validity of quiet standing COP variables to reflect dynamic tasks. While falls risk is multifactorial, the ability for these COP variables to contribute to falls risk prediction also appears to be more limited than tests of gait speed and dynamic balance. However, the variation seen when examining different aspects of COP suggests that it may be of value to extract data beyond typical measures of COP speed.
The strength of association with clinical tests showed differences between the COP variables and some similarities with prior research. Bower et al. [13] found a similar strength association (r ¼ À 0.41) between WBB-derived total COP speed and the step test in 30 people with subacute stroke. While gait speed was not  [12,40], our study found typically stronger associations for the ML variables. This may be explained by the type of tests selected for our study, where ML control is needed for safe and effective weight-shift during stepping movements. Indeed, one study found associations were only significant between AP COP speed and the static components of the Berg Balance Scale [11]. Generally, associations with gait speed and the step test were low-to-moderate in strength, suggesting different aspects of postural control are being assessed. Some variables (e.g., DFA and AP ultralow wavelet) had no association with these tests, therefore their utility to reflect dynamic functional balance activities may be questioned. Mediolateral COP measures of speed or moderate frequency wavelet may best reflect functional balance performance. Of note, ML speed and moderate frequency wavelet were very highly correlated (r ¼ 0.926), therefore using only one of these measures would likely suffice, with the former computationally easier to derive. Very few of the COP variables obtained were found to be significantly predictive of falls, whereas the measures of gait speed and dynamic balance were significant. Of note, the difference between faller and non-faller groups for gait speed (0.17 m/s) was similar to the minimal clinically important difference values reported in people with stroke (0.16 m/s) [41] and the average step test score in the faller group (6.32 taps) was similar to the cut-off reported for falls risk following rehabilitation discharge after stroke (7 taps) [42].
The relatively higher frequency wavelet variable (1.56-3.13 Hz) in the ML direction and 2/8 DFA variables were found to have significance for falls prediction. This is somewhat consistent with prior literature, with studies demonstrating the ability to differentiate between faller and non-faller groups post-stroke using ML COP variables [14,15]. Prior research has shown DFA variables to differ between elderly and young control participants, with findings suggesting that older adults control their movement more tightly in the AP direction [43,44]. Furthermore, a study investigating balance training after stroke using WBB-derived DFA outcomes found a reduction in slow-scale ML dynamics, suggesting tighter ML control [45]. The DFA variables used in our study should be interpreted with caution due to the potential issues associated with this variable obtained using the WBB in the ML direction and for short-term scaling regions reported previously [46]. However, Meade et al. [46] sampled at a non-native, slower, fixed-rate (30 Hz versus around 40 Hz), which would have resulted in inconsistent timing errors between samples and may have impacted their findings. Regardless, DFA is known to be impacted by many factors such as down/up resampling, filtering and inconsistent sampling rate. Future research should examine the stability of this variable between multiple WBBs before it is used more widely.
Frequency analyses revealed a large reduction in the relatively higher-frequency postural sway in both the ML and AP directions when measured over a 3-month period from the early to late subacute phase post-stroke. This may indicate less reliance on proprioceptive input and subsequent muscular corrections. The large responsiveness of the relatively simple task of standing still with eyes open may reflect early recovery of motor control. Conversely, quiet standing balance ability may have improved through the employment of visual compensation strategies for somatosensory or vestibular deficits (i.e., increased visual dependency) [47]. While MDC or clinically-meaningful effect values were not available for most COP variables used in the current study, it should be noted that gait speed and step test scores just exceeded prior published MDC thresholds (i.e., 0.15 m/s [48] and 1.92 taps [13] respectively), whereas the MDC for COP speed measures were not met [13]. In contrast with the current study, Bernhardt et al. [49] demonstrated larger responsiveness of gait (SRM ¼ 0.80) and step test scores (SRM ¼ 0.92-0.95) compared with static balance (SRM ¼ 0.34-0.48) over four weeks of inpatient stroke rehabilitation. However, the static balance was assessed using the Clinical Test for Sensory Interaction and Balance, which is a simplified method involving timing different standing conditions and had substantial ceiling effects at eight weeks post-stroke [49]. This study supports the utility and further investigation of different COP variables. The COP is an indirect measure of postural sway and reflects the occurrence of small muscular contractions, or corrective movements, in quiet standing. Increases in sway, such as seen in various health conditions and with ageing, may indicate reduced neuromuscular control [50]. Nonetheless, it remains unclear as to precisely why postural sway occurs and what characteristics represent a balance deficit [50]. Indeed, smaller scores may reflect more postural stiffness and a reduced ability to respond dynamically to perturbations. Therefore, a range of COP measures reflecting the more complex underlying neuromuscular control dynamics of postural sway has been proposed and explored.
COP speed, amplitude and RMS are relatively easily interpreted, with decreases typically indicating better postural control [50]. Amplitude measures may be less useful as the reliance on individual data points can cause substantial inter-trial variability [51]. The wavelet frequency ranges have been purported to represent different sensory systems [51] or balance control mechanisms [52] and maybe helpful in identifying underlying causes of balance deficits. Indeed, wavelet analysis has been used to identify shifts in neuromuscular strategies with dual-tasking, vision occlusion and neck muscle fatigue [32]. Further, Shinkel-Ivy et al. [52] found individuals with greater RMS wavelet frequencies of between 0.4 and 3.20 Hz (i.e., higher frequencies) had worse responses to external postural perturbation testing. This may indicate impaired reactive balance control reflected by higher frequency movements in static standing. Interestingly, our study found the relatively higher frequency wavelet variables were more closely related to the gait and dynamic balance tests, had a stronger prediction for falls and had a greater magnitude of change over time. Nonetheless, the attribution of different frequency ranges to underlying systems should be interpreted cautiously as this area remains relatively unexplored. Changes in DFA variables, where smaller scores indicate sway is occurring within a more compact boundary [33], are more difficult to conceptualise in a way that allows clinicians to readily interpret and guide decision-making. While our study found that some DFA variables were predictive of falls and changed over time, they had a poor correlation with gait speed and the step test.
This study had several limitations. Our study only examined a small select range of potential COP variables. While research has validated the use of the WBB for obtaining simple COP measures, the concurrent validity of more complex variables has not been comprehensively explored. More challenging stance positions and the use of dual force platforms may also have resulted in different findings. Participants in our study were recruited at different poststroke time points, but all were within the subacute window of recovery. The findings may be limited to a select group of individuals who are being discharged from rehabilitation and able to walk independently or with no more than one person assisting.
There were also limitations related to our selected methods of analysis and interpretation. The strength of association between tests was defined using previously recommended thresholds [37]. Nonetheless, we recognise that there are no widely accepted criteria for defining a strong versus moderate or poor association and the strength of association will be affected by sample size and measurement error [37]. The regression analyses and calculation of odds ratios used for falls prediction were limited in the ability to accurately quantify the prediction of a fall event. However, our sample size did not allow for a sensitive calculation of the Area Under the Curve (AUC) values and these may be more difficult to interpret clinically [53]. The selection of different covariates in the regression analyses may also have altered the findings. Initial stroke severity, for example, would have been considered as a covariate in this data were available to us. Changes in balance were not anchored to other outcomes, such as gait speed, and therefore the clinical relevance is less clear. Furthermore, we did not control for the levels of physical activity or therapy provided post-discharge.
While recommendations for the assessment of quiet standing COP variables in clinical practice remain limited, our findings may be used to guide future research. Future studies could involve larger cohorts, subgroup analyses and assessment at set time points after stroke. We recommend further exploration of measures of signal frequency and complexity in different planes as these demonstrate a potentially stronger relationship with balance and falls. Further validation of lower-cost devices such as the WBB may assist the ease of research and clinical application. Research should consider exploring the clinical feasibility, where barriers may include resource limitations, technical expertise and workload prioritisation [7,54].
In conclusion, our findings highlight the limited validity of quiet standing COP variables, measured in the early subacute phase post-stroke, to reflect dynamic tasks and predict falls. The ML speed and higher frequency variables could be recommended, rather than just a broad measure of COP speed. However, our findings do not support their use for falls prediction in clinical practice over other easier to conduct mobility or balance tests. Nonetheless, COP measures can provide insight into postural control mechanisms and how these change over time. In particular, the measures of signal frequency and complexity warrant further exploration.