UBC Faculty Research and Publications

Instrumenting gait assessment using the Kinect in people living with stroke: reliability and association… Clark, Ross A; Vernon, Stephanie; Mentiplay, Benjamin F; Miller, Kimberly J; McGinley, Jennifer L; Pua, Yong H; Paterson, Kade; Bower, Kelly J Feb 12, 2015

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
52383-12984_2015_Article_6.pdf [ 394.32kB ]
Metadata
JSON: 52383-1.0223530.json
JSON-LD: 52383-1.0223530-ld.json
RDF/XML (Pretty): 52383-1.0223530-rdf.xml
RDF/JSON: 52383-1.0223530-rdf.json
Turtle: 52383-1.0223530-turtle.txt
N-Triples: 52383-1.0223530-rdf-ntriples.txt
Original Record: 52383-1.0223530-source.json
Full Text
52383-1.0223530-fulltext.txt
Citation
52383-1.0223530.ris

Full Text

RESEARCH Open AccessInstrumenting gait assessment using the Kinect inpeople living with stroke: reliability andtestsbalance capacity of people living with stroke. This system provides a minimally intrusive method of examiningJ N E R JOURNAL OF NEUROENGINEERINGAND REHABILITATIONClark et al. Journal of NeuroEngineering and Rehabilitation  (2015) 12:15 DOI 10.1186/s12984-015-0006-8AustraliaFull list of author information is available at the end of the articlepotentially important gait characteristics in people living with stroke.Keywords: Rehabilitation, Measurement, Brain injury, Walking, 10 meter walk test* Correspondence: ross.clark@acu.edu.au1School of Exercise Science, Australian Catholic University, Melbourne,Conclusions: In conclusion, instrumenting gait using the KRoss A Clark1*, Stephanie Vernon1, Benjamin F Mentiplay1, Kimberly J Miller2, Jennifer L McGinley3, Yong Hao Pua4,Kade Paterson3 and Kelly J Bower1,5AbstractBackground: The Microsoft Kinect has been used previously to assess spatiotemporal aspects of gait; however thereliability of this system for the assessment of people following stroke has not been established. This study examinedthe reliability and additional information that the Kinect provides when instrumenting a gait assessment in peopleliving with stroke.Methods: The spatiotemporal variables of step length, step length asymmetry, foot swing velocity, foot swingvelocity asymmetry, peak and mean gait speed and the percentage difference between the peak and mean gaitspeed were assessed during gait trials in 30 outpatients more than three months post-stroke and able to standunsupported. Additional clinical assessments of functional reach (FR), step test (ST), 10 m walk test (10MWT) andthe timed up and go (TUG) were performed, along with force platform instrumented assessments of center ofpressure path length velocity during double-legged standing balance with eyes closed (DLEC), weight bearingasymmetry (WBA) and dynamic medial-lateral weight-shifting ability (MLWS). These tests were performed on twoseparate occasions, seven days apart for reliability assessment. Separate adjusted multiple regressions modelsfor predicting scores on the clinical and force platform assessments were created using 1) the easily assessedclinically-derived gait variables 10MWT time and total number of steps; and 2) the Kinect-derived variables which werefound to be reliable (ICC > 0.75) and not strongly correlated (Spearman’s ρ < 0.80) with each other (i.e. non-redundant).Results: Kinect-derived variables were found to be highly reliable (all ICCs > 0.80), but many were redundant. The finalregression model using Kinect-derived variables consisted of the asymmetry scores, mean gait velocity, affected limbfoot swing velocity and the difference between peak and mean gait velocity. In comparison with the clinically-derivedregression model, the Kinect-derived model accounted for >15% more variance on the MLWS, ST and FR tests andscored similarly on all other measures.inect is reliable and provides insight into the dynamicassociation with balance© 2015 Clark et al.; licensee BioMed Central. This is an Open Access article distributed under the terms of the CreativeCommons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, andreproduction in any medium, provided the original work is properly credited. The Creative Commons Public DomainDedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article,unless otherwise stated.Clark et al. Journal of NeuroEngineering and Rehabilitation  (2015) 12:15 Page 2 of 9IntroductionThe ability to ambulate independently in the communityis one of the most important rehabilitation outcomes forpeople following stroke [1]. Gait impairments are com-mon following stroke, resulting in restrictions in activ-ities of daily living and in the ability to integrate backinto the community after discharge from hospital [1,2].It is therefore important to assess gait function in thispatient cohort, as it is considered a proxy measure ofthe patient’s overall physical function status and capacityto return to their previous lifestyle.Clinical assessments of gait are often limited to measur-ing the total time taken to walk a set distance, for examplethe 10 m walk test (10MWT). From this information, gaitvelocity (or speed) can be calculated which allows forcomparison to normative standards and important thresh-olds such as the minimum velocity required to safely crossroads [3]. Additional information from these tests, such asthe number of steps performed, are also sometimes re-corded to gain greater insight into the person’s functionalability [4,5].Whilst these clinical assessments of gait function areinformative, they lack the precision and data richnessof instrumented methods that provide the kinematicand spatiotemporal aspects of the gait cycle [5]. Instru-mented systems include marker-based three dimen-sional (3D) motion analysis using multiple cameras,body-mounted inertial monitoring unit sensors, and in-strumented walkways such as the GAITRite [6]. Resultsderived from these systems, such as inter-limb gaitasymmetry [7-9], may provide important informationabout dynamic balance control and underlying impair-ments which are associated with long term outcomesbut cannot be derived from standard clinical assess-ments. Furthermore, these measures may better eluci-date the outcomes of gait training programs aimed atimproving symmetry and weight shifting ability [10]. Bymore precisely and accurately quantifying movement itmay be possible to observe subtle changes in physicalfunction that standard clinical assessments are not sen-sitive to. However, further population-level research isneeded to determine the additive benefits of instru-menting gait assessment in a clinical setting.Although these systems offer some benefits, real worldclinical feasibility is limited due to factors such as cost,portability, training and time requirements. Recent evi-dence indicates that the Microsoft Kinect can be used toassess some spatiotemporal aspects of gait [6,11], includ-ing validation against a camera-based 3D motion analysissystem for step length, foot swing velocity and inter-limbgait asymmetry in healthy young adults [6]. Although theKinect may not provide the precision of a multiple cameraor body-mounted sensor system, the low price (<US$100for the Kinect camera, plus the cost of a computer if bothdevices are needed), widespread availability, small size andmarker-less data collection and analysis capabilities offersunique potential for providing a more clinically feasiblemethod of instrumenting gait assessments.There is a paucity of research examining the reliabilityand potential usefulness of the Kinect for assessing gaitin clinical populations. One prior study used the Kinectfor instrumenting the timed up and go (TUG) test instroke survivors, reporting the data obtained were reli-able and provided additional and unique informationover and above the standard clinical outcome assess-ment information [12]. However, due to the 3 m restric-tion on walking inherent in the TUG test, there was onlya small focus on the gait aspects of the TUG. No exam-ination of inter-limb step length or velocity asymmetrieswas performed, which may provide important informa-tion related to a patient’s dynamic balance capabilities.Furthermore, the data reported are unlikely to reflecttypical walking patterns due to the acceleration and de-celeration occurring within this restricted range.The aim of the present study was to determine if instru-menting a gait assessment using the Kinect provides reli-able and potentially valuable information in comparisonwith the standard clinical 10MWT assessment. This wasexamined by determining 1) the inter-session reliability ofkey outcome variables, 2) the redundancy these variableshave with the clinical assessments and each other, and3) whether these variables provide unique and import-ant information about physical function to complementand extend the standard clinical assessment-derived out-come measures. We hypothesized that 1) the Kinect-de-rived measures would be highly reliable, 2) many variableswould be redundant, however a core group of Kinect-derived measures would give unique information re-garding physical function, and 3) when implementedinto a regression model to determine association withclinical and instrumented assessments of gait andbalance, the addition of specific combinations of non-redundant variables would outperform the clinically-derived results.MethodsParticipantsThirty participants living with stroke were consecutivelyrecruited from the Community Therapy Service at TheRoyal Melbourne Hospital. Participants must have beendiagnosed with a non-cerebellar ischemic or hemorrhagicstroke >3 months prior to recruitment, be attendingphysiotherapy for balance or mobility dysfunction, be ableto stand unsupported for >30 seconds, and have a MiniMental State Examination score ≥20. Exclusion criteriawere severe apraxia, severe dysphasia or any other medicalcondition that may impact their balance ability (e.g. severejoint pain, progressive neurological disorders or visualClark et al. Journal of NeuroEngineering and Rehabilitation  (2015) 12:15 Page 3 of 9impairment). This research was given ethical approvalfrom the Royal Melbourne Hospital and University ofMelbourne ethics committees, and all participants pro-vided written informed consent prior to testing. Partici-pants completed two testing sessions at the hospital oneweek apart, with each session consisting of 1) standardclinical assessments of gait and dynamic balance, 2) an in-strumented gait assessment, and 3) an instrumented as-sessment of static and dynamic balance. The participantsin this study are the same as those from a previously pub-lished article [12].Standard clinical assessments of gait and balanceIn order of completion, the clinical assessments were the10MWT, TUG, Step Test (ST) and Functional Reach(FR) test, and were undertaken in accordance with previ-ously published protocols [13-17]. These outcome mea-sures were chosen as they are commonly used withinAustralian rehabilitation settings and present a range ofdynamic balance activities. All assessments were per-formed twice on each of the two days, with the best trialused for analysis, with the exception of the ST whichwas performed once as per the standard testing protocol.The same standardized procedure was followed duringeach testing session. Participants wore shoes and coulduse gait aids and/or ankle-foot orthoses for the gait as-sessment and TUG only. Participants wore their usualclothing; however, if needed, tape was used to makepants more closely fitted at the knees and ankles inorder to increase the accuracy of joint center detectionusing the Kinect.Instrumented assessment of gaitA Microsoft Xbox360 Kinect camera was used duringthe additional gait assessment, which was performed afterthe 10MWT, to obtain spatiotemporal and kinematic in-formation from the participant. The Kinect integrates in-formation from video and depth-sensing cameras tocreate a 3D representation of the field of view [18]. Anartificial intelligence algorithm provided freely by Micro-soft is then used to automatically locate and track the jointcenters and major anatomical landmarks of the body [19].This enables the camera to provide information on the 3Dmovement of the participant in close to real-time.The Kinect gait assessment could not be performedconcurrently with the 10MWT, as the end location ofthat test was in a position which did not allow for place-ment of the Kinect camera or access to an externalpower supply without creating a trip hazard. Conse-quently the Kinect gait assessment walkway was setup inan alternative space in the rehabilitation center, with awalkway length of 6 m. The Kinect camera placementand field of view was setup at the end of this walkway,and calibrated using a custom written software programprior to each testing session using a protocol describedpreviously [6,20].The participant was instructed to start at the begin-ning of the walkway and walk towards the Kinect cam-era, stopping 0.5 m in front of it. For this study theacceptable field of view was restricted to a range from1.5 to 3.5 m from the Kinect. This distance allowed for aminimum of one full gait cycle (i.e. a complete stride)per limb to be recorded per walking trial. The 3D skel-eton position data for each ankle and shoulder center(ie. the position in the middle of the sternum) were re-corded throughout the two trials, and expressed relativeto the Kinect camera. These data were acquired at theirnative sampling frequency, which is irregular and fluctu-ates around 30 Hz. To overcome the sampling irregular-ity issues of the Kinect, spline interpolation was used toresample the Kinect data to 100 Hz. Data was loadedinto a custom program and filtered using a Daubechies4 undecimated wavelet 3.125 Hz lowpass filter. The gaitevent time points for toe-off and ground contact wereidentified using a supervised automated analysis algo-rithm described previously [6], which is based on thevelocity of the movement of the ankle joint center.The outcome measures analysed and reported in thisstudy are based on those previously observed to be reli-able in healthy young people [6]. Specifically, these vari-ables included step length and foot swing velocity forthe affected (deemed the limb contralateral to the sideof the stroke) and unaffected limb. Measures of meanand peak gait velocity during the trial were derived fromthe anterior displacement of the shoulder center through-out the field of view of the Kinect. The difference betweenthe peak and mean velocity was also calculated as a meas-ure of the forward progression variability, and expressedas a percentage score. Additionally, given the level ofinter-limb asymmetry often present during gait afterstroke, asymmetry ratio scores were calculated for eachmeasure which compared the affected and unaffected limb[9]. These asymmetry scores were converted to an abso-lute value and expressed as a percentage. All Kinect-derived outcome measures were averaged across the twotrials.Instrumented assessment of static and dynamic balanceForce platform-derived measures of static (quiet stance)and dynamic (movement over a fixed base of support)balance were performed after the standard clinical tests.Static balance tests involved standing double-legged bal-ance with eyes closed (DLEC), standing weight bearingasymmetry (WBA) and dynamic balance involved amedial-lateral weight shifting (MLWS) test. Each test wasperformed on one (DLEC) or two (WBA and MLWS)Nintendo Wii Balance Boards (NWBB) with custom soft-ware used to record data, and in the case of the MLWSClark et al. Journal of NeuroEngineering and Rehabilitation  (2015) 12:15 Page 4 of 9test, provide visual feedback. These tests have demon-strated high test-retest reliability (ICC = 0.82 to 0.98) inpeople with stroke [21]. The DLEC test was performed onthe NWBB with feet a fixed distance apart (17 cm betweenheels and toe-out angle of 14°) [22] and eyes focused on atarget on the wall in front. Data was collected for three tri-als of 30 seconds duration and the median score was usedfor analysis. The outcome measure for this test was totalcenter of pressure path velocity in cm/s, and this was de-rived using a technique described previously [23]. WBAwas assessed by having the participant stand still for30 seconds with a foot on each NWBB with heels 17 cmapart and toe-out angle of 14°. The difference in force dis-tributed between the lower limbs expressed as a percent-age of body mass was deemed the WBA. This datacollection and analysis technique is similar to previouslypublished articles examining asymmetry with multipleNWBBs [24,25]. The MLWS test was designed to measurethe ability to repeatedly shift body weight distribution be-tween the lower limbs to follow a visual feedback targetfor 30 seconds. Participants were required to shift theirbody weight alternatively between the left and right sidesto reach and hold (for one second) in a target area (equat-ing to 60-80% body weight) displayed as columns on atelevision screen. This test was based on a previous designshown to be responsive to change post-stroke [26] andhas been used as an outcome measure in a recent strokerehabilitation trial [27]. Data (number of shifts completedin 30 seconds) were collected for five trials and the me-dian of the last three trials was used [17].Statistical analysisIntraclass correlation coefficients (ICC2,k) were used asindices of relative reliability of the Kinect-derived vari-ables, and these coefficients were calculated in a 2-wayanalysis of variance (ANOVA) based on absolute agree-ment. Absolute reliability was represented by the stand-ard error of measurement (SEM), which was derivedfrom the square root of the mean square error termfrom the respective repeated ANOVA, and the mini-mum detectable change (MDC95) score.To compare the abilities of the standard clinical as-sessment of gait versus a Kinect instrumented gait as-sessment for predicting static and dynamic balancescores, we created two independent regression modelsusing data from the first testing session. The first modelonly included the clinically-derived gait assessment out-come measures, and included two independent variables:1) the time taken to complete the 10MWT, and 2) thenumber of steps taken during the test. The second modelconsisted of the Kinect-derived outcome variables. Tominimize model overfitting and to avoid multicollinear-ity from correlated variables, we a priori specified thatKinect-derived variables which were strongly related(Spearman ρ’s ≥ 0.80) to either of the clinical gait as-sessment outcome variables would be deemed redun-dant and excluded from the model. The exception wasmean gait velocity recorded during the trial, which wasretained as the proxy Kinect-derived measure for 10MWTtime. We compared the two (non-nested) models usingthe model unadjusted and adjusted R2 values [28], andadditionally the Akaike Information Criterion (AIC) [29]which is a penalized measure of model fit.ResultsAll 30 participants attended both testing sessions. Table 1provides the participant demographics. Ten of the thirtyparticipants required the use of gait aids during the10MWT. Of these ten, two could not perform the testwithout using a 4-wheeled walking frame. The data forthese two subjects was retained for assessing the reliabil-ity of gait velocity derived from the Kinect, but were ex-cluded from the regression and redundancy analysisbecause accurate information regarding all Kinect-derivedvariables except those calculated from the shoulder centercould not be obtained. The reliability results for theKinect-derived gait measures are provided in Table 2. Allvariables were found to have high reliability (all ICCscores ≥ 0.80).Redundancy analysis of the Kinect-derived gait vari-ables with the clinical measures of 10MWT time andnumber of steps revealed that four of the nine variableswere highly correlated (i.e. redundant), and thereforethese measures were excluded from any further analysis(Table 3). Redundancy analysis of the remaining Kinect-derived outcome measures with each other is providedin Table 4. This revealed that foot swing velocity for theaffected and unaffected limbs was highly correlated. Adecision was made to discard the unaffected limb footswing velocity from any further analysis, as the affectedlimb score may be a more likely target for interventionssuch as botulinum toxin injection [30]. Consequently,only four variables remained to be inputted into the re-gression model, namely 1) step length asymmetry, 2)foot swing velocity asymmetry, 3) the difference betweenpeak and mean gait velocity, and 4) affected limb footswing velocity. A fifth input, mean gait velocity, was in-cluded into the model to represent the standard clinicalmeasure of gait speed.Table 5 describes the results for the regression models.Both the clinically-derived and Kinect-derived modelsproduced similar adjusted R2 values for the TUG, DLECand WBA tests (R2 value differences of 0.06, 0.01 and0.06 respectively). In contrast, the Kinect-derived modeloutperformed the clinically-derived model for predictingscores on the MLWS, ST and FR by R2 value differencesof 0.25, 0.21 and 0.19 respectively. Furthermore, the AICvalues were lower in four of the six Kinect-derivedhypothesis 2, many of the outcome measures were re-dundant within device or when compared to the 10MWTtime and the total number of steps performed. Interest-ingly, when a combination of non-redundant Kinect-derived variables were inputted into a multiple regressionmodel to predict dynamic balance assessments, this modelhad higher adjusted R2 values than one built using the10MWT time taken to complete and the total number ofsteps performed. Specifically, the Kinect-derived modelexplained >15% more variance for the FR (Kinect-derivedadjusted R2 = 0.58, clinically-derived adjusted R2 = 0.39;19% more variance explained), ST (Kinect-derived ad-justed R2 = 0.79, clinically-derived adjusted R2 = 0.58;Table 1 Participants’ characteristics recorded during thefirst assessment sessionDemographic Variable ResultAge (yrs) 68 ± 15Height (cm) 166.7 ± 9.4Body Mass (kg) 72.5 ± 11.9Gender (n =male) 21Mini-mental State exam score (/30) 27.1 ± 2.7Lesion type (% Infarct) 73Lesion side (% Right) 63Time since stroke (months) 21 ± 19Hypertension (%) 70Clark et al. Journal of NeuroEngineering and Rehabilitation  (2015) 12:15 Page 5 of 9Dyslipidaemia (%) 47Type 2 Diabetes mellitus (%) 2710 m Walk test speed (m/s) 0.97 ± 0.3510 m Walk test steps (n) 20 ± 9Timed up and go (s) 17.7 ± 10.5Functional reach (cm) 28.5 ± 7.4Step test (n) 9.3 ± 4.9Static double leg eyes closed balance – Path velocity 2.06 ± 1.05models, indicating that the Kinect-derived models hadbetter overall model fit.DiscussionThe present study examined the instrumentation of agait assessment in a clinical stroke rehabilitation settingusing the Microsoft Kinect. All outcome measures de-rived from the Kinect were found to be highly reliable(ICC ≥ 0.80) in accordance with hypothesis 1. Supporting(cm/s)Static weight-bearing asymmetry (% Body mass) 46.3 ± 8.5Medial-lateral weight shifting ability (no. of shifts in30 seconds)10.0 ± 3.7Values are reported as mean ± standard deviation unless otherwise indicated.Table 2 Test-retest reliability measures for the Kinect-derivedVariables Test ReteMean ± SD MeaAffected step length (mm) 507 ± 147 513Unaffected step length (mm) 520 ± 166 520Step length asymmetry (%) 15.0 ± 15.6 17.4Affected foot swing velocity (m/s) 3.18 ± 0.91 3.16Unaffected foot swing velocity (m/s) 3.08 ± 1.3 3.09Foot swing velocity asymmetry (%) 12.6 ± 13.7 14.0Mean velocity (m/s) 0.89 ± 0.33 0.90Peak velocity (m/s) 1.22 ± 0.39 1.22Peak – Mean velocity difference (%) 41.8 ± 18.4 39.8Abbreviations: SD = standard deviation; ICC = intraclass correlation coefficient; CI = cdetectable change.21% more variance explained) and MLWS assessments(Kinect-derived adjusted R2 = 0.70, clinically-derivedadjusted R2 = 0.44; 26% more variance explained). Con-versely, the model results were similar (≤6% more vari-ance explained by either model) for the static balancetests DLEC (Kinect-derived adjusted R2 = 0.36, clinically-derived adjusted R2 = 0.37; 1% less variance explained),WBA (Kinect-derived adjusted R2 = 0.38, clinically-derivedadjusted R2 = 0.44; 6% less variance explained), and theoverall measure of physical function test the TUG(Kinect-derived adjusted R2 = 0.94, clinically-derived ad-justed R2 = 0.88; 6% more variance explained). This pro-vides partial support for hypothesis 3, but only for thedynamic balance assessments.In regard to reliability, all outcome measures recordedacceptable ICC scores, and seven of the nine variablespossessed ICC values > 0.90. This was expected, as thevariables were chosen based on those reported previ-ously to be reliable using the Kinect to assess gait, aswell as other systems assessing spatiotemporal gait pa-rameters in people living with stroke [6,8,31]. However,in the present study the participants were not requiredto perform the assessments barefoot and while wearingshorts. Consequently, this is the first study to show that theKinect can provide reliable spatiotemporal gait informationgait variablesst ICC(2,k) SEM MDC95n ± SD (95% CI)± 169 0.97 (0.95 to 0.99) 29 80± 144 0.98 (0.96 to 0.99) 21 58± 19.4 0.89 (0.76 to 0.95) 5.2 14.3± 0.90 0.97 (0.94 to 0.99) 0.16 0.44± 1.29 0.99 (0.97 to 0.99) 0.13 0.36± 13.0 0.80 (0.58 to 0.90) 6.1 17.0± 0.32 0.98 (0.96 to 0.99) 0.05 0.13± 0.39 0.98 (0.95 to 0.99) 0.06 0.15± 18.1 0.96 (0.92 to 0.98) 3.7 10.2onfidence interval; SEM = standard error of measurement; MDC =minimumcorrelations between these two aspects of physical func-tion in people with Parkinson’s disease or the community-dwelling elderly [33,34]. Instrumenting a gait assessmentusing the Kinect may allow for large scale screening whichprovides insight into balance whilst also obtaining infor-mation on gait from a single test. However, the utility ofthis assessment would be highly context dependent andTable 3 Redundancy analysis of the Kinect-derived gaitvariables compared to the clinical measures of manuallyassessed time (MT) and manually assessed number ofsteps (MS) using Spearman’s correlationOutcome Measure MT MSAffected step length (mm) −0.86* −0.93*Unaffected step length (mm) −0.93* −0.96*Clark et al. Journal of NeuroEngineering and Rehabilitation  (2015) 12:15 Page 6 of 9in a stroke population assessed in a clinical setting withminimal interference to the patient. This is important, be-cause instrumenting gait assessments using the Kinect al-lows for examination of inter-limb step length and footswing velocity asymmetries – measures which cannot cur-rently be readily assessed in a clinical setting. Additionally,while pressure-mat systems such as the GAITRite can pro-vide some spatiotemporal measures of gait such as stridelength, unlike the Kinect they cannot be used to examineother potentially important factors such as foot swing vel-ocity or variability in trunk motion.Instrumenting a gait assessment using the Kinectallowed for improved prediction of scores on the FR, STand MLWS tests compared to the non-instrumented gaitassessment. When using the often reported R valuethresholds of poor (R < 0.40), modest (R = 0.40 – 0.74) orStep length asymmetry (%) 0.45* 0.39Affected foot swing velocity (mm/s) −0.71* −0.76*Unaffected foot swing velocity (mm/s) −0.78* −0.80*Foot swing velocity asymmetry (%) 0.25 0.20Mean velocity (mm/s) −0.92* −0.92*Peak velocity (mm/s) −0.92* −0.93*Peak – Mean velocity difference (%) 0.66* 0.65**significant at P < 0.05.The correlation between MT and MS is 0.93*.excellent (R ≥ 0.75) [32], the clinical assessment modelwas rated as fair for the FR and MLWS tests and narrowlyexceeded the threshold for excellent for the ST. In con-trast, the Kinect-derived model was rated as excellent forall three of these tests of dynamic balance capability dur-ing controlled movements. The large difference in predict-ive strength between the two regression models indicatesthat the standard clinical assessment of gait is not suitableas a strong proxy measure of dynamic balance. This sup-ports prior research showing only moderate to lowTable 4 Redundancy analysis of the Kinect-derived gait variaVariables Step length AffectedAsymmetry Swing vePeak – Mean velocity difference 0.53* −0.59Foot swing velocity asymmetry 0.02 0.06Unaffected foot swing velocity −0.26 0.91Affected foot swing velocity −0.31*significant at P < 0.05.not suitable as a replacement for a thorough assessmentof balance where indicated. Of particular interest from thepresent study is the disparity in the strength of the associ-ation that the two different models have with MLWS per-formance. Previous studies have highlighted the potentialfor this assessment to provide insight into dynamic bal-ance [26,27]; however, it requires two force platforms anda computer monitor to provide visual feedback, which re-duces its clinical feasibility. Having a single Kinect setupto acquire data during a gait trial without the need to pro-vide visual feedback is much simpler to implement. Whilstthe Kinect-derived model was capable of predictingperformance on the dynamic balance assessments, nei-ther regression model was able to accurately predictperformance on the static balance assessments of WBAand standing still with eyes closed. This provides fur-ther evidence that static balance capability has only alimited association with gait function, and supports thework by Lewek et al. [35] which found that spatiotemporalgait parameters are more closely related to dynamic ratherthan static balance in people with chronic stroke.This study builds on our previous work which instru-mented the TUG using the Kinect in a stroke population[12]. Similar to this present study, the prior work exam-ined individual components of the TUG and observedthat this provided additional information over-and-abovethat which can be obtained from just recording the totaltime taken to complete the test. The present study dif-fers in that it focuses on the spatiotemporal componentsof steady-state gait, which cannot be derived from theshort walking distance of the TUG. Given our smallsample size, we were unable to statistically contrast thepredictive validity of the Kinect-TUG and Kinect-Gaitvariables. However, we noted that both sets of Kinect vari-ables (i) were non-redundant with their correspondingclinical measures, (ii) provided incremental predictivevalues over their corresponding clinical measures, and (iii)bles using Spearman’s correlation matrixfoot Unaffected foot swingvelocityFoot swing velocityasymmetrylocity* −0.60* 0.21−0.16*mR2****f sttryareClark et al. Journal of NeuroEngineering and Rehabilitation  (2015) 12:15 Page 7 of 9produced similar model R2 values (ST R2: Kinect-TUG=0.73, Kinect-gait = 0.79; FR R2: Kinect-TUG= 0.53, Kinect-gait = 0.58). The major advantages of instrumenting a gaitassessment in comparison with the TUG are that it can bedone with no additional equipment (i.e. the standardisedheight chair of the TUG), increasing its potential utility,and by any patient who can walk a relatively short dis-tance, making it potentially more generalizable to an inpatient severe stroke population. The present study alsobuilds on the prior study by examining the associationof the Kinect-derived variables with both instrumentedand clinical assessments of static and dynamic balance,with the finding that spatiotemporal aspects of gait aremore strongly associated with dynamic rather thanstatic balance.This study had limitations. The study sample consistedTable 5 Results for the clinically-derived and Kinect-derivedFunctional assessment ClinicalR2 AdjOverall physical functionTimed up and go 0.89 0.88*Static balanceStatic double leg eyes closed balance 0.41 0.37*Static weight-bearing asymmetry 0.49 0.44*Dynamic balanceFunctional reach 0.43 0.39*Step test 0.61 0.58*Medial-lateral weight shifting ability 0.48 0.44*Clinically-derived model included 2 input variables: 10MWT time; total number oswing velocity asymmetry; Peak – Mean velocity difference; Step length asymme*p-values <0.01; **p-values <0.001.R2, Ezekiel equation adjusted R2 and the Akaike Information Criterion (AIC) scoreof a relatively small and heterogenous group of strokesurvivors recruited from a single outpatient facility. Theallowance of gait aid use and inclusion of participantswith minimal gait deficits may have impacted on thestrength of associations found. Outcome measures werelimited to spatiotemporal variables which have been pre-viously identified as being reliable. Potentially importantoutcome measures such as medial-lateral center of masssway [36] were not examined, and if reliable their inclu-sion may have strengthened the regression models. Thesevariables were not analyzed for a number of reasons re-lated to potential data error. For example, medial-lateralcenter of mass sway is highly reliant on accurate identifi-cation of the hip center by the skeleton tracking algo-rithm, however in this patient population this landmarkwas often occluded due to reasons such as elbow flexionof the stroke-affected arm or handles on gait assistive de-vices. Furthermore, additional Kinect-based measureswere not analyzed because of the already high number ofvariables derived from the Kinect for this form ofregression research. Another limitation was the accuracyand precision of the Kinect, which is unlikely to be as highas body-mounted sensors or marker-based 3D motionanalysis camera systems. However, given the potentialclinical feasibility and mass physical function screeningadvantages that this system offers, we believe that its pos-sible benefits warrant further investigation. The use of themanually assessed time and manually assessed number ofsteps together in the clinically-derived regression model isalso a limitation, as correlation analysis revealed these twovariables to be somewhat redundant (Spearman’s ρ =0.93). This may have unfairly penalised the results of theclinically-derived model given the two somewhat similarinput variables, biasing the study towards a positive out-come for the Kinect assessment. However, performing theregression analysis using just manually assessed time toultiple regression modelsKinectAIC R2 Adj R2 AIC81 0.95 0.94** 63−8 0.47 0.36* −5102 0.50 0.38* 108109 0.65 0.58** 10070 0.83 0.79** 5260 0.75 0.69** 46eps. Kinect-derived model included 5 input variables: Mean velocity; Foot; Affected foot swing velocity.provided for comparison of model strengths.complete the test as the sole input variable did not createa meaningful change in the adjusted R2 values (R2 valuechange range = −0.02 to 0.02), and therefore we areconfident in our observations. The restricted field of viewof the Kinect does not allow for multiple gait cycles to beexamined from a single test. This is a limitation when itcomes to assessing asymmetry, or measures of gait vari-ability. Also potentially related to this restricted field ofview, we observed a difference in the gait speed derivedfrom the Kinect compared to the stopwatch, with theKinect-derived value being slower (median [inter-quartilerange] difference between the 10MWT and Kinect gaitspeeds for each individual = 7.8 [1.5 to 17.5] %). Visualexamination of our data revealed that for some partici-pants deceleration occurred as they approached the cam-era, but in the majority the mean gait speed was relativelystable. Consequently, in the former case this disparity mayhave been a result of patients beginning gait decelerationwhile they were still in the field of view. In the latter ex-ample however, the participant may have simply beenClark et al. Journal of NeuroEngineering and Rehabilitation  (2015) 12:15 Page 8 of 9walking slower with a camera located at the end of the testin contrast to the 10MWT which finished multiple metersfrom any obstruction. Both of these strategies may impactgait velocity and step length data, and therefore should beconsidered when comparing results obtained from theKinect with those of other methods. Finally, our patientcohort was reasonably high functioning, with an averagegait speed well above that considered critical for commu-nity ambulation. Whether the Kinect provides useful in-formation in people with slower walking speeds (i.e.poorer function) cannot be generalized from our find-ings. This is an important question as these people maybe at most risk of adverse outcomes related to physicalfunction.In conclusion, these findings provide support for thepotential usefulness of implementing a Kinect instru-mented gait assessment in a clinical setting. This single,easy to implement and reliable assessment may allow forgreater insights into a person's gait and dynamic balanceability. Variables derived from this assessment may allowfor better monitoring of change in gait performance overtime in both clinical and research settings, as well asproviding information to inform targeted treatment ap-proaches. It is important that future research examinesways to optimize the feasibility and utility of this systemin a clinical setting to ensure translation into standardclinical practice.Competing interestsThe authors declare that they have no competing interest. Author RC mayrelease the software used in this study in the future.Authors’ contributionsRC was involved in the study design, software design, data collection,coordination, manuscript drafting and overall supervision of the study. KBwas involved in the study design, data collection, coordination andmanuscript drafting. SV was involved in the study design, data collection,coordination and manuscript drafting. BM was involved in data analysis andmanuscript drafting. KM, JM and KP were involved in study design andmanuscript drafting. PYH was involved in data and statistical analysis andmanuscript drafting. All authors read and approved the final manuscript.AcknowledgementsThe authors would like to thank the Royal Melbourne Hospital for theirassistance with the study. This project was partially supported by a NationalStroke Foundation Research Honours Grant.Funding sourcesNo external funding was received for this project.Author details1School of Exercise Science, Australian Catholic University, Melbourne,Australia. 2Department of Physical Therapy, University of British Columbia,Vancouver, Canada. 3School of Physiotherapy, University of Melbourne,Melbourne, Australia. 4Department of Physiotherapy, Singapore GeneralHospital, Singapore, Singapore. 5Department of Physiotherapy, RoyalMelbourne Hospital, Melbourne, Australia.Received: 8 October 2014 Accepted: 28 January 2015References1. van de Port IG, Kwakkel G, Lindeman E. Community ambulation in patientswith chronic stroke: how is it related to gait speed? J Rehabil Med.2008;40:23–7.2. Salbach NM, Mayo NE, Higgins J, Ahmed S, Finch LE, Richards CL.Responsiveness and predictability of gait speed and other disabilitymeasures in acute stroke. Arch Phys Med Rehabil. 2001;82:1204–12.3. Asher L, Aresu M, Falaschetti E, Mindell J. Most older pedestrians are unable tocross the road in time: a cross-sectional study. Age Ageing. 2012;41:690–4.4. Sturt R, Holland A, New P. Walking ability at discharge from inpatientrehabilitation in a cohort of non-traumatic spinal cord injury patients. SpinalCord. 2009;47:763–8.5. Coutts F. Gait analysis in the therapeutic environment. Man Therap.1999;4:2–10.6. Clark RA, Bower KJ, Mentiplay BF, Paterson K, Pua YH. Concurrent validity ofthe Microsoft Kinect for assessment of spatiotemporal gait variables.J Biomech. 2013;46:2722–5.7. Hendrickson J, Patterson KK, Inness EL, McIlroy WE, Mansfield A. Relationshipbetween asymmetry of quiet standing balance control and walking post-stroke. Gait Posture. 2014;39:177–81.8. Lewek MD, Randall EP. Reliability of spatiotemporal asymmetry duringoverground walking for individuals following chronic stroke. J NeurologicPhysical Ther. 2011;35:116–21.9. Patterson KK, Gage WH, Brooks D, Black SE, McIlroy WE. Evaluation of gaitsymmetry after stroke: A comparison of current methods andrecommendations for standardization. Gait Posture. 2010;31:241–6.10. Patterson KK, Mansfield A, Biasin L, Brunton K, Inness EL, McIlroy WE.Longitudinal Changes in Poststroke Spatiotemporal Gait Asymmetry OverInpatient Rehabilitation. Neurorehab Neural Repair. 2014;29:153–62.11. Stone E, Skubic M. Evaluation of an inexpensive depth camera for in-homegait assessment. J Ambient Intell Smart Environ. 2011;3:349–61.12. Vernon S, Paterson K, Bower K, McGinley J, Miller K, Pua Y, et al. QuantifyingIndividual Components of the timed up and go using the Kinect in peopleliving with stroke. Neurorehab Neural Repair. 2014;29:48–53.13. Bernhardt J, Ellis P, Denisenko S, Hill K. Changes in balance and locomotionmeasures during rehabilitation following stroke. Physiother Res Int.1998;3:109–22.14. Thrane G, Joakimsen RM, Thornquist E. The association between timed upand go test and history of falls: the Tromso study. BMC Geriatr. 2007;7:1.15. Mercer VS, Freburger JK, Chang SH, Purser JL. Step test scores are related tomeasures of activity and participation in the first 6 months after stroke. PhysTher. 2009;89:1061–71.16. Duncan P, Weiner D, Chandler J, Studenski S. Functional reach: a newclinical measure of balance. J Gerontol. 1990;45:192–7.17. Bower K, Clark RA, Martin C, Miller K. Clinical feasibility of the Nintendo Wiifor balance training post-stroke: A phase II randomised controlled trial in aninpatient setting. Clin Rehabil. 2014;28:912–923.18. Menna F, Remondino F, Battisti R, Nocerino E. Geometric Investigation of aGaming Active Device. Bellingham, WA, ETATS-UNIS: Society of Photo-Optical Instrumentation Engineers; 2011.19. Shotton J, Fitzgibbon A, Cook M, Sharp T, Finocchio M, Moore R, et al. Real-Time Human Pose Recognition in Parts from Single Depth Images. In:Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference;20–25 June 2011. 2011. p. 1297–304.20. Clark RA, Pua Y-H, Fortin K, Ritchie C, Webster KE, Denehy L, et al. Validity ofthe Microsoft Kinect for assessment of postural control. Gait Posture.2012;36:372–7.21. Bower K, McGinely J, Miller K, Clark RA. Instrumented static and dynamicbalance assessment after stroke using Wii Balance Boards: Reliability andassociation with clinical tests. PLOS One. 2014;9:e115282.22. McIlroy WE, Maki BE. Preferred placement of the feet during quiet stance:Development of a standardized foot placement for balance testing. ClinBiomech. 1997;12:66–70.23. Clark RA, Bryant AL, Pua Y, McCrory P, Bennell K, Hunt M. Validity andreliability of the Nintendo Wii Balance Board for assessment of standingbalance. Gait Posture. 2010;31:307–10.24. Clark RA, McGough R, Paterson K. Reliability of an inexpensive and portabledynamic weight bearing asymmetry assessment system incorporating dualNintendo Wii Balance Boards. Gait Posture. 2011;34:288–91.25. Foo J, Paterson K, Williams G, Clark R. Low-cost evaluation and real-timefeedback of static and dynamic weight bearing asymmetry in patientsundergoing in-patient physiotherapy rehabilitation for neurologicalconditions. J NeuroEng Rehab. 2013;10(1):74.26. de Haart M, Geurts AC, Dault MC, Nienhuis B, Duysens J. Restoration ofweight-shifting capacity in patients with postacute stroke: a rehabilitationcohort study. Arch Phys Med Rehabil. 2005;86:755–62.27. Bower K, Clark RA, McGinley J, Martin C, Miller K. Feasibility and efficacy ofthe Nintendo Wii™ gaming system to improve balance performance post-stroke: Protocol of a phase II randomised controlled trial in an inpatientrehabilitation setting. Games Health. 2013;2:103–8.28. Ezekiel M. Methods of Correlation Analysis. New York: Wiley; London,Chapman. 1930.29. Akaike H. A new look at the statistical model identification. AutomaticControl, IEEE Transactions on. 1974;19:716–23.30. Gracies JM, Elovic E, McGuire J, Simpson D. Traditional pharmacologicaltreatments for spasticity part I: Local treatments. Muscle Nerve. 1997;20:S61–91.31. Høyer E, Opheim A, Strand LI, Moe-Nilssen R. Temporal and spatial gaitparameters in patients dependent on walking assistance after stroke:Reliability and agreement between simple and advanced methods ofassessment. Gait Posture. 2014;40:101–6.32. Fleiss J. The Design and Analysis of Clinical Experiments. New York, U.S.A.:John Wiley & Sons; 1986.33. Desai A, Goodman V, Kapadia N, Shay BL, Szturm T. Relationship betweendynamic balance measures and functional performance in community-dwelling elderly people. Phys Ther. 2010;90:748–60.34. Yang YR, Lee YY, Cheng SJ, Lin PY, Wang RY. Relationships between gait anddynamic balance in early Parkinson's disease. Gait Posture. 2008;27:611–5.35. Lewek MD, Bradley CE, Wutzke CJ, Zinder SM. The relationship betweenspatiotemporal gait asymmetry and balance in individuals with chronicstroke. J Appl Biomech. 2014;30:31–6.36. Hodt-Billington C, Helbostad JL, Moe-Nilssen R. Should trunk movement orfootfall parameters quantify gait asymmetry in chronic stroke patients? GaitPosture. 2008;27:552–8.Submit your next manuscript to BioMed Centraland take full advantage of: • Convenient online submission• Thorough peer review• No space constraints or color figure charges• Immediate publication on acceptance• Inclusion in PubMed, CAS, Scopus and Google Scholar• Research which is freely available for redistributionClark et al. Journal of NeuroEngineering and Rehabilitation  (2015) 12:15 Page 9 of 9Submit your manuscript at www.biomedcentral.com/submit

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.52383.1-0223530/manifest

Comment

Related Items