вход по аккаунту



код для вставкиСкачать
[m5G;October 24, 2017;23:55]
Medical Engineering and Physics 0 0 0 (2017) 1–8
Contents lists available at ScienceDirect
Medical Engineering and Physics
journal homepage:
Prediction of ground reaction forces for Parkinson’s disease patients
using a kinect-driven musculoskeletal gait analysis model
Moataz Eltoukhy a, Christopher Kuenze b, Michael S. Andersen c, Jeonghoon Oh a,
Joseph Signorile a,d,∗
Department of Kinesiology and Sport Sciences, School of Education & Human Development, University of Miami, Coral Gables, FL 33143, USA
Department of Kinesiology, School of Education, Michigan State University, East Lansing, MI 48824, USA
Department of Materials and Production, Aalborg University, Fibigerstraede 16, 9220 Aalborg East, Denmark
Center on Aging, Miller School of Medicine, 1695 N.W. 9th Avenue, Suite 3204, Miami, FL, 33136, USA
a r t i c l e
i n f o
Article history:
Received 21 March 2017
Revised 28 August 2017
Accepted 4 October 2017
Available online xxx
Gait analysis
Parkinson’s disease
GRF prediction
a b s t r a c t
Kinetic gait abnormalities result in reduced mobility among individuals with Parkinson’s disease (PD).
Currently, the assessment of gait kinetics can only be achieved using costly force plates, which makes it
difficult to implement in most clinical settings. The Microsoft Kinect v2 has been shown to be a feasible
clinic-based alternative to more sophisticated three-dimensional motion analysis systems in producing
acceptable spatiotemporal and kinematic gait parameters. In this study, we aimed to validate a Kinectdriven musculoskeletal model using the AnyBody modeling system to predict three-dimensional ground
reaction forces (GRFs) during gait in patients with PD. Nine patients with PD performed over-ground
walking trials as their kinematics and ground reaction forces were measured using a Kinect v2 and force
plates, respectively. Kinect v2 model-based and force-plate measured peak vertical and horizontal ground
reaction forces and impulses produced during the braking and propulsive phases of the gait cycle were
compared. Additionally, comparison of ensemble curves and associated 90% confidence intervals (CI90) of
the three-dimensional GRFs were constructed to investigate if the Kinect sensor could provide consistent
and accurate GRF predictions throughout the gait cycle. Results showed that the Kinect v2 sensor has the
potential to be an effective clinical assessment tool for predicting GRFs produced during gait for patients
with PD. However, the observed findings should be replicated and model reliability established prior to
integration into the clinical setting.
© 2017 IPEM. Published by Elsevier Ltd. All rights reserved.
1. Introduction
Parkinson’s disease (PD) currently affects 10 million people
worldwide including one million Americans, and the number of
newly diagnosed cases is on the rise [1]. The particular reduction
in mobility experienced by patients with PD is commonly characterized by rigidity and gait impairments [2–4], which causes significant loss of independence and increased incidences of falls. These
gait impairments results from the progressive loss of dopamine
producing cells in the substantia nigra compacta of the basal ganglia [5,6].
Subjective gait assessment approaches, commonly used by clinicians to examine gait patterns of individuals with PD, have been
shown to be unreliable [7]. Therefore, achieving accurate gait anal∗
Corresponding author at: Department of Kinesiology and Sport Sciences, University of Miami, 1507 Levante Ave. Max Orovitz Building 114, Coral Gables, FL 33146
E-mail address: (J. Signorile).
yses for PD patients in clinical settings require combining both
subjective and objective assessments [8]. Furthermore, gait, which
is a particular semiautomatic motor task, is specifically sensitive to
the on and off Parkinsonian states. Thus, to precisely delineate the
temporal evolution of gait-related complications, their characteristics, and their severity, more objective instrumental methods are
needed. Currently, there is no marker-less and cost-effective gait
analysis tool that can be used on a regular basis to assess gait patterns of patients with PD during both on and off states in clinic
and home settings.
The importance of kinetics, and therefore vertical ground reaction forces (GRFv) to assessment of gait is evidenced in the literature. Morris et al. [6] indicated that analyzing changes in kinetics is necessary to provide a full understanding of the primary determinants of PD gait dysfunction. They noted that although kinematic and spatiotemporal gait parameters can be provided using
visual cues, the kinetics of gait abnormalities is a necessary component in determining the reason for unexplained falls, as noted
by researchers who have assessed fall probability in the elderly [9].
1350-4533/© 2017 IPEM. Published by Elsevier Ltd. All rights reserved.
Please cite this article as: M. Eltoukhy et al., Prediction of ground reaction forces for Parkinson’s disease patients using a kinect-driven
musculoskeletal gait analysis model, Medical Engineering and Physics (2017),
[m5G;October 24, 2017;23:55]
M. Eltoukhy et al. / Medical Engineering and Physics 000 (2017) 1–8
Fig. 1. The experimental setup with the two floor-impeded force plates and the Microsoft Kinect v2 sensor placed in front of the subject’s direction of progression and at a
height of 0.75 m from the ground.
Kinetics measures, such as the pattern of the vertical ground reaction forces (GRFV ) during gait in PD patients, can be used as an indicator of the stage the disease [10]. For instance, both GRFV peak
values are reduced in the early stages of the disease, while in more
advanced stages the GRFV pattern is characterized by only a single
narrow peak [11].
Additionally, patients with PD have difficulty with both rapid
force production and declination, since rapid gait termination requires concurrent increases in braking impulse and decreases in
propulsion impulse. Bishop et al. [12] investigated the net braking impulse using horizontal ground reaction forces (GRFAP ) when
braking impulse rapidly increases and propulsion impulse decreases during unplanned gait termination in patients with PD.
They showed that patients with PD are less capable of generating sufficient net braking impulse in time-critical situations than
age-matched healthy subjects.
Furthermore, a number of researchers used ground reaction
forces to develop automatic classifiers of gait patterns in patients with PD. Manap et al. [13] used GRFV and other kinematic parameters to classify PD gait; and in a follow-up study
[14], they utilized only the GRFV peak values during initial contact, mid-stance, and toe off to detect gait irregularities in PD patients. Additionally, Dubey et al. [15] used artificial neural networks to distinguish between normal and PD gait patterns using GRFV ; while Zhang et al. [16] developed a sparse representation method utilizing the patient’s GRFV to detect Parkinsonian
gait. The main drawback of these studies, designed to analyze
gait patterns and develop auto-classifiers of Parkinsonian gait using GRF, was that they required the use of floor-embedded force
platforms, which makes them an impractical solution for clinical
Despite the alarming increase in the number of PD patients in
the US [1]; clinicians still rely on subjective tools such as the Unified Parkinson’s Disease Rating Scale [17] to determine the presence and severity of the disease, and these methods often return inconsistent diagnoses [18]. Alternatively, gait analysis using
laboratory-based motion capture systems can quantitatively assess
the severity and progress of the disease [19]; yet, due to the financial burden, technical challenges, lack of portability, need for
relatively large space, and comprehensive setup requirements associated with its use, this approach remains difficult to adopt in
clinical settings [20,21]. On the other hand, wearable sensors used
to measure joints kinematics such as skin-attached optical tracking devices [22] and Inertial Measurement Units [23] can interfere
with the movement performed, especially for those already suffering from gait impairment, and IMUs are known to suffer from technical issues such as interfering with surrounding magnetic fields.
Also, both systems suffer from motion artifacts, where a translational displacement between bony landmark and the attached sensor is likely to occur, causing errors in estimating position during movement, which in turn leads to inaccuracies in determining
joint angles. Although the IMUs measure orientation rather than
position, errors related to movement can however still occur because of rotational displacement of the sensor relative to the segment [24]. Additionally, assessment of the kinetic measures produced during gait using floor-impeded force plates is affected by
on the patient’s foot placement over the force plate, while also the
use of pressure mats limits the patient to walk within a limited
width of the mat [25].
The use of the Kinect v2 in performing three-dimensional spatiotemporal and kinematics analysis was previously validated in
various applications such as treadmill gait [26] and over-ground
gait [27] analyses. Therefore, this paper will validate a full-body
musculoskeletal model, driven by the Microsoft Kinect v2, to derive the ground reaction forces produced during over-ground gait
for patients with PD. This will be achieved by comparing a number of calculated measures of the ground reaction forces to the values measured using floor-imbedded force plates. These measures
will include; the peak braking and propulsive vertical and horizontal forces commonly used as automatic Parkinsonian gait classifiers
[14−16], patterns of the three ground reaction forces using ensemble curves, and a single step GRF symmetry index [28] computed
Please cite this article as: M. Eltoukhy et al., Prediction of ground reaction forces for Parkinson’s disease patients using a kinect-driven
musculoskeletal gait analysis model, Medical Engineering and Physics (2017),
[m5G;October 24, 2017;23:55]
M. Eltoukhy et al. / Medical Engineering and Physics 000 (2017) 1–8
Fig. 2. (A) The Kinect-driven musculoskeletal model (KinectAB) including the AnyBody mannequin and the Kinect skeleton as well as the aligned global coordinate systems
of AnyBody (yellow) and Kinect (purple). The locations of the Kinect’s body joint centers (grey dots), the AnyBody and Kinect anatomical markers, green and red dots,
respectively, (B) The 25 contact points created under each foot, and (C) The KinectAB model during one gait trial with the calculated GRF (red arrows). (For interpretation of
the references to color in this figure legend, the reader is referred to the web version of this article.)
as the ratio between the braking and propulsive impulse of the
horizontal forces.
To the best of our knowledge, this paper will be the first to
utilize a commercially available cost-effective videogame accessory,
Kinect v2, to drive a full-body musculoskeletal model to obtain kinetic measures produced during pathological gait of patients with
PD, thus exploring the potential of its possible adoption in clinical
2. Methods
2.1. Subjects
Parkinson’s disease (age = 71.0 ± 5.6 yrs, height = 165.4 ± 11.3 cm,
weight = 70.8 ± 16.7 kg), were recruited for this study. Subjects
were mildly to moderately impaired (H&Y stages I-III) [29] and
had a score of 24 or above on the Folstein Mini-Mental State
Examination [30]. All patients were capable of ambulation for
at least 50 feet without an assistive device. Additionally, subjects with greater than stage III symptoms on the H&Y Scale,
spinal fusion or other orthopedic surgery in the past six months,
severe visual deficits, or dementia were excluded from the
2.2. Experimental setup
Kinematic data were collected using a single Kinect sensor v2
(Microsoft Corp. Redmond, WA). The Kinect v2 used in this study
is an upgraded version of its predecessor, Kinect v1, which utilizes the time of flight technology for the depth data extraction
[31]. This improved technology enables accurate distance estimation of the tracked objects using the light pulse’s time of flight
from the sensor to the tracked object [32]. Additionally, the Kinect
v2 has higher resolution of 512 × 424 pixels as compared to the
Kinect v1 resolution (320 × 240 pixels) resulting in higher tracking accuracy including smaller objects [33]. The Kinect was placed
2.5 m from the subject, at a height of 0.75 m from the ground. This
positioning was based on previous published work involving measurement of gait parameters using the Kinect sensor [26]. Kinetic
data were collected using two floor-imbedded force plates (Type
9286AA; Kistler Instrument AG, Winterhur, Switzerland), as shown
in Fig. 1.
Please cite this article as: M. Eltoukhy et al., Prediction of ground reaction forces for Parkinson’s disease patients using a kinect-driven
musculoskeletal gait analysis model, Medical Engineering and Physics (2017),
[m5G;October 24, 2017;23:55]
M. Eltoukhy et al. / Medical Engineering and Physics 000 (2017) 1–8
2.3. Experimental procedures
After providing informed written consent, subjects were familiarized with the experimental setup, and their heights and body
mass were recorded. Subjects then performed a number of familiarization gait trials to determine the starting point for each subject. This was done to ensure that the subjects would consistently
contact the force plate with no interference to their walking patterns. Following familiarization, subjects performed three walking
trials at their normal walking speeds, and all subjects performed
the walking trials barefoot, without wearing any type of footwear.
A trial was considered successful if the limb of interest made contact with the force plate during the complete stance phase of a gait
cycle, the entire foot contact took place on the force plate, and the
subject’s entire body was recorded within the field of view of the
Kinect sensor.
During each gait trial, data were collected simultaneously from
the Kinect and the two force plates. Force plates data were collected at a sampling rate of 10 0 0 Hz, while the data from the
Kinect sensor were sampled at 30 Hz. The Kinect’s depth data
stream was analyzed by isolating the background depth information and tracking the subject’s movement using anthropometric
models to extract 26 joint trajectories using a customized MATLAB
code (MathWorks, Massachusetts, USA) as described by Skals et al.
2.4. Kinect-driven full-body musculoskeletal model
The musculoskeletal GaitFullBody model (AnyBody Technology
A/S, Aalborg, Denmark), which is based on three separate studies
[35–37], was used in this research. Subject-specific scaling of the
model was based on the method by Rasmussen et al. (2005) [38].
As described in [34], the stick figure fitted in each frame of the
Kinect’s depth data was used to automatically scale the segment
lengths in the GaitFullBody model according to the joint-to-joint
distances of the Kinect’s stick figure, e.g. hip-to-hip distance. In
the case of segments that are defined differently between both systems, such as the trunk, hands, and feet, the segment lengths were
not obtained automatically. Instead, these segments were scaled
by first, adding additional nodes to the musculoskeletal model at
locations corresponding to points identifiable on the stick figure.
Then, the un-scaled model was placed in a neutral position and
the distance between the added nodes were computed and saved
together with the un-scaled segment lengths. Subsequently, the
ratio between the un-scaled segment lengths and the nodes was
multiplied onto the segment length measurements on the stick
figure before being used to scale the respective segments in the
musculoskeletal model. Additionally, as shown in Fig. 2A, virtual
markers (green dots: AnyBody Model and red dots: Kinect) around
the corresponding virtual joint centers (grey dots) were used to
drive the musculoskeletal model using the three-dimensional position data obtained by the Kinect. This was achieved by solving a
nonlinear least-square optimization problem to minimize the leastsquare difference between the virtual markers on the stick figure
and those on the musculoskeletal model according to the method
described by Andersen et al. [35−37]. This method ensures that the
musculoskeletal model is optimally driven by the Kinect’s stick figure despite the discrepancies in the segment and joint definitions
between the two models. Furthermore, the total body mass was
distributed to the individual segments according to the method
of Winter et al. [39]. The Kinect-driven AnyBody musculoskeletal
model used in this paper will be referred to as “KinectAB” and is
available at
Fig. 3. (A) The braking and propulsive measured vertical GRF points, GRFV1 and
GRFV2 , respectively, (B) The braking and propulsive measured horizontal GRF points,
GRFAP1 and GRFAP2 , respectively, and the corresponding impulses of force (IB and IP ).
2.5. Ground reaction forces prediction and impulse of force
The ground reaction forces generated during gait was calculated using a modified version of the method proposed by Fluit
et al. [40] and Skals et al. [34]. This is achieved by introducing artificial muscle-like actuators at 25 contact points under each foot
(pink dots shown in Fig. 2B). Contact (including Coulomb friction)
between each node and the ground was established when the
node was sufficiently close to the ground (located within 50 mm of
the ground plane) and relatively stationary, as defined by a minimal relative velocity (Vrel ) across the ground with a maximum
Vrel of 1.1 m/s. At each contact point, five unidirectional actuators (treated as “muscles” in the inverse dynamics analysis) were
introduced to generate a normal force perpendicular to the laboratory floor and static friction forces (friction coefficient of 0.1
[41]) in the mediolateral and anteroposterior directions. In addition, the transition from no contact to full contact was smoothed
similarly to the method described in Skals et al. [34] using the
velocity of the nodes. One was aligned with GRFv and the other
two pairs represented the mediolateral and anteroposterior components, GRFML and GRFAP , respectively (Fig. 2C). Additionally,
55 muscles with constant strength were added to each leg and
Please cite this article as: M. Eltoukhy et al., Prediction of ground reaction forces for Parkinson’s disease patients using a kinect-driven
musculoskeletal gait analysis model, Medical Engineering and Physics (2017),
[m5G;October 24, 2017;23:55]
M. Eltoukhy et al. / Medical Engineering and Physics 000 (2017) 1–8
Table 1
Comparison between measured and calculated kinetics variables during the stance phase.
Vertical Force
Braking vertical
force (GRFV1 )
vertical force
(GRFV2 )
horizontal force
horizontal force
Force Impulses
Index (SIAP )
Mean ± SD
Mean ± SD
1.02 ± 0.03
1.02 ± 0.03
0.04 ± 0.02
−0.01 ± 0.05
1.01 ± 0.05
1.10 ± 0.04
0.09 ± 0.06
−0.09 ± 0.06
−0.10 ± 0.03
−0.13 ± 0.03
0.03 ± 0.02
0.12 ± 0.04
0.16 ± 0.04
−0.86 ± 0.25
−0.91 ± 0.25
(± 95%CI)
(± 95%CI)
0.03 ± 0.02
0.04 ± 0.02
−0.04 ± 0.03
0.11 ± 0.08
0.05 ± 0.13
Limits of Agreement
indicates a significant between system difference (p ≤ 0.05).
indicates a significant correlation (p ≤ 0.05).
muscle recruitment under-determinacy issue during the doublestance phase was solved using the muscle recruitment method by
Damsgaard et al. [42].
2.6. Data analysis
The joints’ coordinate system for the Kinect was kept consistent
with the International Society of Biomechanics recommendations
(X-axis is the mediolateral, Y-axis is the vertical, and the Z-axis
is the anteroposterior direction) [43]. Data were analyzed during
the stance phase of the gait cycle when the foot was in contact
with the force plate. The stance phase was defined in the Kinect
using the initial contact (IC) and toe off (TO) events, which were
expressed according to the method described Zeni Jr. et al. [44].
Specifically, the IC was defined as the point at which the maximum
anteroposterior distance between the ankle and mid-PSIS occurred;
while TO was defined as the point at which the minimum distance
2.6.1. Ground reaction forces
Ground reaction forces were normalized to the subjects’ body
weight (BW). The calculated peak vertical and horizontal ground
reaction forces, GRFV and GRFAP respectively, were compared to
the corresponding measured values. Specifically, the braking and
propulsive peak points commonly used in automatic Parkinsonian
gait detection [16,17] were determined. These vertical and horizontal force values are GRFV1 , GRFV2 , GRFAP1 and GRFAP2 (Fig. 3A,B).
All calculations were performed using the original data sets without any up- or down-sampling to prevent any introduction of noise
or loss of resolution of the data.
The measured and calculated GRF in the vertical, anteroposterior, and mediolateral direction were graphed across the gait cycle
with associated 90% confidence intervals (CI90 ). Due to the difference in the sampling rates, and to normalize the time to 100% of
the stance phase for graphing purposes, a custom LabVIEW code
was used to resample the measured GRF data set to obtain an
identical number of data points for any given gait cycle while
maintaining the signal integrity in the time and frequency domains [45]. The measured and calculated GRF data sets were synchronized by the automatically identified and logged reference gait
events (initial contact and toe off), using a 10 N threshold on the
vertical ground reaction forces [46].
Anteroposterior impulse, which was shown to be the main indicator of the lack of appropriate lower limb rate of force generation necessary during gait termination in patients with PD [13],
was used to further investigate the validity of the calculated GRF
in obtaining similar impulse of forces. Symmetry Index (SIAP ), defined as the ratio between the braking (negative) and the propulsive (positive) impulse of the GRF in the anteroposterior direction
(IB and IP respectively), was determined using the calculated and
measured GRFAP curves (Fig. 3B) [28].
2.7. Statistical analysis
Mean differences between the measured and calculated ground
reaction forces and symmetry variables were compared using
paired samples t-tests in order to establish any significant difference between measurement approaches. Absolute error between
these measurement approaches was calculated as an indicator of
error magnitude independent of directionality of error between approaches (Eq. 1).
Absolute Error =|GRFmeasured − GRFcalculated |
Relative error and associated limits of agreement were calculated (Eq. 2) to describe the magnitude and directionality of error
associated with calculated ground reaction forces based on directly
measured variables with consideration for the scale of measurement for each variable.
Relative Error = GRFmeasured − GRFcalculated
The absolute agreement (ICC3,1 ) and relative consistency (ICC2,1 )
between measurement approaches for all variables were assessed
Please cite this article as: M. Eltoukhy et al., Prediction of ground reaction forces for Parkinson’s disease patients using a kinect-driven
musculoskeletal gait analysis model, Medical Engineering and Physics (2017),
[m5G;October 24, 2017;23:55]
M. Eltoukhy et al. / Medical Engineering and Physics 000 (2017) 1–8
using interclass correlations coefficients [47]. Absolute agreement
considers the within subject agreement between systems without
consideration of systematic error. Relative consistency takes this
factor into account when estimating between system consistencies.
ICCs were interpreted as poor ( < 0.4), fair (0.4 - < 0.6), good (0.6
- < 0.75), and excellent ( ≥ 0.75) [48]. All statistical analyses were
completed using SPSS version 20.0 (IBM, Chicago Illinois). Alpha
level was set a-priori as P ≤ 0.05.
Comparison of ensemble curves and associated 90% confidence intervals (CI90 ) of the vertical, anteroposterior, and mediolateral ground reaction forces was performed to investigate if
the KinectAB model is able to consistently and accurately predict
ground reaction forces throughout the gait cycle. Significant differences between systems were established when the CI90 for each
system did not overlap consecutively for at least 3% of the gait cycle [49]. Ensemble curve analyses were completed using Microsoft
Excel (version 2010, Microsoft Corp. Redmond, WA).
3. Results
Calculated propulsive vertical force (p = .01), braking horizontal force (p = .004), and propulsive horizontal force (p = .01) were
significantly greater when compared to the directly measured approach during the gait cycle (Table 1). For the vertical component, both braking and propulsive force exhibited poor agreement
and consistency between the two approaches while the horizontal component exhibited good agreement (ICC3, 1 = 0.63−0.75) and
excellent consistency (ICC2, 1 = 0.84−0.89, Table 1). In addition, anteroposterior impulse symmetry index exhibited excellent agreement (ICC3, 1 = 0.93) and consistency (ICC2, 1 = 0.93, Table 1) between measurement approaches. When measurement approaches
were compared across the gait cycle, calculated vGRF significantly
differed from measured vGRF during the mid-stance phase (10–
35% of the stance phase) while both the anteroposterior and mediolateral GRF significantly differed from the directly measured approach during the loading response (0- 12%) of the stance phase
immediately following initial contact (Fig. 4).
4. Discussion
Clinical evaluation of gait abnormalities experienced by individuals with Parkinson’s disease is commonly a subjective process
that is largely focused on comparison of kinematic patterns with
those expected of an elderly individual without neurodegenerative
disease. While this has clinical benefit, it remains difficult to quantify the progress of disease or improvements due to clinical interventions due to the lack of sensitivity of subjective assessment as
well as the inability to include kinetic evaluation in this style of assessment. In this study, we attempted to validate a musculoskeletal
gait analysis model that was driven by the Kinect v2 sensor for application in individuals with PD. The goal of this approach was to
enhance the quality of clinical gait assessment in this population
while ensuring cost effectiveness and feasibility of implementation
in a clinical setting. The ability to accurately assess GRF parameters using low cost technology that does not require the outfitting
of participants, especially those at high risk of fall such as individuals with PD, with cumbersome equipment that may interfere
with normal gait patterns represents a meaningful advance in this
area of research. Based on our findings, it appears that the Kinectdriven musculoskeletal model shows promise for the estimation of
unilateral ground reaction force components as well as indices of
ground reaction force symmetry in the PD population. However,
it should be noted that the observed findings must be replicated
and model reliability established using more advanced analysis approaches and larger, more diverse patient populations prior to clinical integration.
Fig. 4. The ensemble curves (mean and 90% confidence intervals) of the measured
(grey lines) and calculated (black lines) vertical (A), horizontal (B), and mediolateral
(C) ground reaction forces. IC: initial contact and TO: toe off.
Vertical ground reaction force is commonly utilized as a measure of limb loading during dynamic tasks among populations with
neurodegenerative disease and is a measure sensitive to progressive changes in gait patterns that are often a hallmark of disease
progression such as reduction in peak values. Recent investigations have indicated that emerging technologies such as wireless
Please cite this article as: M. Eltoukhy et al., Prediction of ground reaction forces for Parkinson’s disease patients using a kinect-driven
musculoskeletal gait analysis model, Medical Engineering and Physics (2017),
[m5G;October 24, 2017;23:55]
M. Eltoukhy et al. / Medical Engineering and Physics 000 (2017) 1–8
accelerometers and inertial measurement units are able to validly
and reliably measure differences in kinematic and kinetic characteristic among individuals with PD [50,51]. These advances have
highlighted the need for continued development of realistic and
cost-effective technology that may be able to facilitate clinical
integration of motion analysis in order to aid in the diagnosis
and monitoring of disease progression among individuals with PD.
Based on our findings, the ability of the KinectAB model to effectively estimate peak GRF is highly dependent on the component
force being evaluated. In the case of GRFV , despite small magnitude
absolute and relative differences between assessment techniques
(force plates and KinectAB model), the agreement and consistency
between systems was poor (Table 1). Conversely, for GRFML and
GRFAP agreement ranged from good to excellent despite significant
differences being present between systems for the GRFML in the
braking and propulsive phases. Interestingly, the strongest agreement and consistency within our outcome measures was found
in the symmetry-based measure, SIAP , which may highlight a potential area for future investigation in terms of improving the
measurement agreement for GRFV and GRFML . Furthermore, the
greater discrepancies in the calculated GRFs during the period of
the stance phase between the loading response and mid-stance
phases is in agreement with the observations reported by Eltoukhy
et al. [30] for the sagittal plane kinematics obtained by the Kinect
v2 during treadmill waking; since in this study the Kinect v2 indicated a stiffer knee joint during this same period of the gait cycle.
Despite the observed overestimation in the calculated peak GRFAP
values, GRFAP1 and GRFAP2 ; the braking and propulsive impulses of
the GRFAP , resulted in small overestimation in the estimated symmetry index of 5.9% (Table 1).
The measurement of ground reaction force at discrete points
(i.e. peak GRF) during the gait cycle is the most common approach
to kinetic assessment where gait abnormalities in PD patients are
characterized by a reduction in the second peak of the GRFv ; therefore, it is important to have a clear understanding of the ability of
the KinectAB model to effectively measure ground reaction force
throughout the gait cycle. An approach that allows the most comprehensive assessment of limb loading and continuous comparison
can enable the detection of more subtle changes in gait that occur
due to slow incremental progression of neurodegenerative conditions such as Parkinson’s disease [52]. As shown in Fig. 4, the calculated GRF plots obtained by the KinectAB model showed similar
patterns to the measured values, especially the vertical and horizontal GRFs (Fig. 4A,B). A maximum difference between the calculated and measured average peak vertical GRFs of 8.3% was obtained, yet, higher overestimations in the peak GRFAP values were
observed. On the other hand, the calculated GRFML showed less
consistent pattern as compared to the measured values (Fig. 4C);
this was especially apparent during the loading response phase of
the gait cycle. This lack of GRFML prediction accuracy in general
can be explained by the absence of the eversion/inversion kinematics driver in the foot model used in the KinectAB model. While
this represents a key limitation of this model, the integration of
richer kinematic source data for model development in future iterations of the model may help in improving the measurement
agreement especially in the cases of GRFML and GRFAP within the
first 10% of the stance phase.
Finally, according to the data shown in table 1, the difference
between the measured and calculated GRFV ranged between 0.02
and 0.15 N/BW. To the best of our knowledge, no minimal clinically accepted error in the ground reaction forces produced during gait for PD elderly populations was established. Thus, and to
relate our findings to potential clinical applications, we first calculated the average difference between the maximum GRFV produced by healthy and PD elderly populations during gait. This was
done by using the public gait database [53], which consists of GRFV
data records of 93 PD and 73 healthy elderly subjects. The database
contained the subjects’ measurements of the GRFV in Newtons as
they walked at their self-selected pace on level ground as well
as his/her demographic information. The calculated average normalized GRFV difference between the healthy and PD groups was
0.51 N/BW. As shown, the maximum error in the calculated GRFV
was more than three-folds less than the average GRFV difference
between the pathological PD gait and the normal gait values produced by age-matched healthy population which indicates that the
proposed Kinect-based approach can detect the up-normal changes
in the clinically-relevant GRFV values.
4.3. Limitations
Several key limitations associated with this investigation do exist and should be taken into account when considering the results.
The relatively small sample size of patients with PD (N = 9), and
the single camera-based motion analysis system may have limited
the ability of the authors to establish a stable estimate of Kinect v2
measurement accuracy that could be applied across different walking speeds. Also, the GaitFullBody model adopted in this study did
not include an eversion/inversion kinematics driver. Lastly, future
studies should consider the use of two Kinect v2 sensors to improve the tracking accuracy of the sagittal plane kinematics of the
ankle joint, which could in return enhance the prediction outcome
of the GRFAP in general and during the loading response phase in
5. Conclusions
Continued evolution and improvement of clinical gait analysis
techniques for individuals with neurodegenerative diseases is essential in order for more comprehensive evaluations of the progression of these diseases and for rapidly advancing treatments.
While reliability and responsiveness of this assessment approach
have yet to be established, based on this initial step, the Kinect v2
sensor-driven kinetic model shows promise for estimating ground
reaction forces and kinetic symmetry, which are key indicators of
gait dysfunction in PD patients. Continued investigation to further
develop this modeling approach and improve the scope of its application, while addressing barriers to clinical implementation, are
essential prior to integration into the clinical setting.
Ethical approval
The study was approved by the University’s Internal Review
Board for the Use and Protection of Human Subjects and all participants provided written informed consent prior to enrollment.
Conflict of interest statement
All authors declare that there are no relevant conflicts for any
of the authors included on this manuscript.
The authors would like to thank the participant who gave selflessly their time and effort to allow the completion of this research.
[1] Understanding Parkinson’s Disease, Retrieved from,, 2014.
[2] Halliday SE, Winter DA, Frank JS, Patla AE, Prince F. The initiation of gait in
young, elderly, and Parkinson’s disease subjects. Gait Posture 1998;8(1):8–14.
[3] King LA, Horak FB. Delaying mobility disability in people with Parkinson disease using a sensorimotor agility exercise program. Phys Ther 2009;89(4):384.
Please cite this article as: M. Eltoukhy et al., Prediction of ground reaction forces for Parkinson’s disease patients using a kinect-driven
musculoskeletal gait analysis model, Medical Engineering and Physics (2017),
[m5G;October 24, 2017;23:55]
M. Eltoukhy et al. / Medical Engineering and Physics 000 (2017) 1–8
[4] Olanow CW, Tatton WG. Etiology and pathogenesis of Parkinson’s disease. Ann
Rev Neurosci 1999;22(1):123–44.
[5] Sofuwa O, Nieuwboer A, Desloovere K, Willems AM, Chavret F, Jonkers I. Quantitative gait analysis in Parkinson’s disease: comparison with a healthy control
group. Arch Phys Med Rehabil 20 05;86(5):10 07–13.
[6] Morris ME, McGinley J, Huxham F, Collier J, Iansek R. Constraints on the kinetic, kinematic and spatiotemporal parameters of gait in Parkinson’s disease.
Hum Mov Sci 1999;18(2):461–83.
[7] Toro B, Nester C, Farren P. A review of observational gait assessment in clinical
practice. Phys Theory Pract 2003;19:137–49.
[8] Coutts F. Gait analysis in the therapeutic environment. Manu Ther
[9] Studenski S, Duncan PW, Chandler J. Postural responses and effector factors in
persons with unexplained falls: results and methodologic issues. J Am Geriatr
Soc 1991;39(3):229–34.
[10] Koozekanani SH, Balmaseda MT Jr, Fatehi MT, Lowney ED. Ground reaction
forces during ambulation in Parkinsonism: pilot study. Arch Phys Med Rehabil 1987;68(1):28–30.
[11] Ueno E, Yanagisawa N, Takami M. Gait disorders in Parkinsonism: a study with
floor reaction forces and EMG. Adv Neurol 1993;60:414–18.
[12] Bishop MD, Brunt D, Kukulka C, Tillman MD, Pathare N. Braking impulse and
muscle activation during unplanned gait termination in human subjects with
Parkinsonism. Neurosci Lett 2003;348(2):89–92.
[13] Manap HH, Tahir NM, Yassin AIM. Statistical analysis of Parkinson disease gait
classification using artificial neural network. In: Proceedings of IEEE international symposium on signal processing and information technology (ISSPIT);
2011. p. 60–5.
[14] Manap HH, Tahir NM. Detection of Parkinson gait pattern based on vertical
ground reaction force. In: Proceedings of IEEE international conference on control system, computing and engineering (ICCSCE); 2013. p. 631–6.
[15] Dubey M, Wadhwani AK, Wadhwani S. Gait-based vertical ground reaction
force analysis for Parkinson’s disease diagnosis using self organizing map. Int J
Adv Biol Biomed Res 2013;1(6):624–36.
[16] Zhang Y, Ogunbona PO, Li W, Munro B, Wallace GG. Pathological gait detection
of Parkinson’s disease using sparse representation. In: Proceedings of international conference on digital image computing: techniques and applications
(DICTA); 2013. p. 1–8.
[17] Movement Disorder Society Task Force on Rating Scales for Parkinson’s Disease
The unified Parkinson’s disease rating scale (UPDRS): status and recommendations. Mov Disord Off J Mov Disord Soc 2003;18(7):738.
[18] Frenkel-Toledo S, Giladi N, Peretz C, Herman T, Gruendlinger L, Hausdorff JM.
Effect of gait speed on gait rhythmicity in Parkinson’s disease: variability of stride time and swing time respond differently. J Neuroeng Rehabil
[19] Chen PH, Wang RL, Liou DJ, Shaw JS. Gait disorders in Parkinson’s disease:
assessment and management. Int J Gerontol 2013;7(4):189–93.
[20] Whittle MW. Clinical gait analysis: a review. Hum Mov Sci. 1996;15:369–87.
[21] Clark RA, Pua Y-H, Fortin K, Ritchie C, Webster KE, Denehy L. Validity
of the Microsoft kinect for assessment of postural control. Gait Posture
[22] Doherty C, Bleakley CM, Hertel J, Caulfield B, Ryan J, Delahunt E. Laboratory
measures of postural control during the star excursion balance test after acute
first-time lateral ankle sprain. J Athl Train 2015;50(6):651–64.
[23] Seel T, Raisch J, Schauer T. IMU-based joint angle measurement for gait analysis. Sensors (Basel, Switzerland) 2014;14(4):6891–909.
[24] Bergmann JH, Mayagoitia RE, Smith IC. A portable system for collecting
anatomical joint angles during stair ascent: a comparison with an optical
tracking device. Dyn Med : DM 2009;8:3.
[25] Olney SJ, Richards C. Hemiparetic gait following stroke. Part I: characteristics.
Gait Posture 1996;4:136–48.
[26] Eltoukhy M, Oh J, Kuenze C, Signorile J. Improved kinect-based spatiotemporal
and kinematic treadmill gait assessment. Gait Posture 2017;51:77–83.
[27] Geerse DJ, Coolen BH, Roerdink M. Kinematic validation of a multi-kinect v2
instrumented 10-meter walkway for quantitative gait assessments. PloS one
[28] Vaverka F, Elfmark M, Svoboda Z, Janura M. System of gait analysis based on
ground reaction force assessment. Acta Gymnica 2015;45(4):187–93.
[29] Hoehn M, Yahr M. Parkinsonism: onset, progression and mortality. Neurology
[30] Folstein MF, Robins LN, Helzer JE. The mini-mental state examination. Arch
Gen Psychiatr 1983;40(7):812.
[31] Vineetha GR, Sreeji C, Lentin J. Face expression detection using Microsoft
kinect with the help of artificial neural network. Trends Innov Comput 2012
- Intell Syst Des 2012;2014:176–80 12/15/2014.
[32] Ganapathi V, Plagemann C, Koller D, Thrun S. Real time motion capture using
a single time-of-flight camera. In: Proceedings of IEEE conference on computer
vision and pattern recognition (CVPR), Stanford University, Computer Science
Department, Stanford, CA, USA; 2010. p. 755–62. 07/13-18.
[33] Microsoft, Kinect for Windows features, Available at:, https://developer., 2014, Accessed June 14,
[34] Skals S, Rasmussen KP, Bendtsen KM, Yang J, Andersen MS. A musculoskeletal model driven by dual Microsoft kinect sensor data. Multibody Syst Dyn
[35] Andersen MS, Damsgaard M, Rasmussen J. Kinematic analysis of over-determinate biomechanical systems. Comput Methods Biomech Biomed Eng
[36] Andersen MS, Damsgaard M, MacWilliams B, Rasmussen J. A computationally
efficient optimisation-based method for parameter identification of kinematically determinate and over-determinate biomechanical systems. Comput Methods Biomech Biomed Eng 2010;13(2):171–83.
[37] Andersen MS, Yang J, de Zee M, Zhou L, Bai S, Rasmussen J. Full-body musculoskeletal modeling using dual microsoft kinect sensors and the anybody
modeling system. In: Proceedings of 14th international symposium on computer simulation in biomechanics; 2013. p. 23–4.
[38] Rasmussen J, Zee MD, Damsgaard M, Christensen ST, Marek C, Siebertz K. A
general method for scaling musculo-skeletal models. In: Proceedings of international symposium on computer simulation in biomechanics; 2005.
[39] Winter DA. Biomechanics and motor control of human movement. Hoboken:
John Wiley & Sons; 2005.
[40] Fluit R, Andersen MS, Kolk S, Verdonschot N, Koopman HFJM. Prediction of
ground reaction forces and moments during various activities of daily living. J
Biomech 2014;47(10):2321–9.
[41] Kleiner A, Galli M, Franceschini M, De Pandis MF, Stocchi F, Albertini G,
et al. The coefficient of friction in Parkinson’s disease gait. Funct Neurol
[42] Damsgaard M, Rasmussen J, Christensen ST, Surma E, De Zee M. Analysis of
musculoskeletal systems in the anybody modeling system. Simul Model Pract
Theory 20 06;14(8):110 0–11.
[43] Wu G, Cavanagh PR. ISB recommendations for standardization in the reporting
of kinematic data. J Biomech 1995;28(10):1257–61.
[44] Zeni JA, Richards JG, Higginson JS. Two simple methods for determining gait
events during treadmill and overground walking using kinematic data. Gait
Posture 2008;27(4):710–14.
[45] Thiebaut C, Roques S. Time-scale and time-frequency analyses of irregularly sampled astronomical time series. EURASIP J Appl Signal Process
20 05;20 05:2486–99.
[46] Lescano CN, Rodrigo SE, Christian DA. A possible parameter for gait clinimetric
evaluation in Parkinson’s disease patients. In: Proceedings of conference series
on physics, 705. IOP Publishing; 2016.
[47] Shrout PE, Fleiss JL. Intraclass correlations: uses in assessing rater reliability.
Psychol Bull 1979;86(2):420–8.
[48] Cicchetti V. Guidelines, criteria, and rules of thumb for evaluating normed
and standardized assessment instruments in psychology. Psychol Assess
[49] Chinn L, Dicharry J, Hertel J. Ankle kinematics of individuals with chronic ankle
instability while walking and jogging on a treadmill in shoes. Phys Ther Sport
[50] Sejdic E, Lowry KA, Bellanca J, Redfern MS, Brach JS. A comprehensive assessment of gait accelerometry signals in time, frequency and time-frequency domains. IEEE Trans Neural Syst Rehabil Eng 2014;22(3):603–12.
[51] Coste CA, Sijobert B, Pissard-Gibollet R, Pasquier M, Espiau B, Geny C. Detection
of freezing of gait in Parkinson disease: preliminary results. Sensors (Basel).
[52] Muniz A, Liu H, Lyons KE, Pahwa R, Liu W, Nadal J. Evaluation of the effects of
deep brain stimulation on ground reaction force in Parkinson disease patients
using principal component analysis. Braz J Biomech 2013;13(26):29–34.
[53] Hausdorff J, et al. Gait in parkinson’s disease. National Institute of Health, National Parkinson Foundation, and the Parkinson’s Disease Foundation; 2005.
Retrieved from: Accessed
15 06 17].
Please cite this article as: M. Eltoukhy et al., Prediction of ground reaction forces for Parkinson’s disease patients using a kinect-driven
musculoskeletal gait analysis model, Medical Engineering and Physics (2017),
Без категории
Размер файла
1 693 Кб
medengphy, 2017, 004
Пожаловаться на содержимое документа