Perturbation Experiment

NIH Public Access
Author Manuscript
IEEE Int Conf Rehabil Robot. Author manuscript; available in PMC 2013 April 10.
Published in final edited form as:
NIH-PA Author Manuscript
IEEE Int Conf Rehabil Robot. 2011 ; 2011: 5975474. doi:10.1109/ICORR.2011.5975474.
Model-Based Estimation of Active Knee Stiffness

Serge Pfeifer*, Michael Hardegger*, Heike Vallery*, Renate List†, Mauro Foresti‡, Robert
Riener*, and Eric J. Perreault§,¶
*Sensory-Motor Systems Lab, Institute of Robotics and Intelligent Systems, ETH Zurich, Zurich,
Switzerland & Spinal Cord Injury Center, Medical Faculty, Balgrist University Hospital, Zurich,
Switzerland [email protected], [email protected], [email protected] †Institute for
Biomechanics, ETH Zurich, Zurich, Switzerland ‡Institute for Surgical Technology &
Biomechanics, University of Bern, Bern, Switzerland §Department of Biomed. Eng. and
Department of Phys. Med. and Rehab., Northwestern University, Chicago, IL, USA ¶Sensory
Motor Performance Program, Rehabilitation Institute of Chicago, Chicago, IL, USA e-
[email protected]
Abstract
Knee joint impedance varies substantially during physiological gait. Quantifying this modulation
is critical for the design of transfemoral prostheses that aim to mimic physiological limb behavior.
Conventional methods for quantifying joint impedance typically involve perturbing the joint in a
controlled manner, and describing impedance as the dynamic relationship between applied
perturbations and corresponding joint torques. These experimental techniques, however, are
difficult to apply during locomotion without impeding natural movements. In this paper, we
propose a method to estimate the elastic component of knee joint impedance that depends on
muscle activation, often referred to as active knee stiffness. The method estimates stiffness using a
musculoskeletal model of the leg and a model for activation-dependent short-range muscle
stiffness. Muscle forces are estimated from measurements including limb kinematics, kinetics and
muscle electromyograms. For isometric validation, we compare model estimates to measurements
involving joint perturbations; measured stiffness is 17% lower than model estimates for extension,
and 42% lower for flexion torques. We show that sensitivity of stiffness estimates to common
approaches for estimating muscle force is small in isometric conditions. We also make initial
estimates of how knee stiffness is modulated during gait, illustrating how this approach may be
used to obtain parameters relevant to the design of transfemoral prostheses.
I. Introduction
The mechanical properties of the human leg can be tuned to maximize performance over a
wide range of behaviors, and even during different phases of a single behavior. For example,
during locomotion, the leg must be sufficiently stiff to prevent buckling at heel strike, yet
also sufficiently compliant to allow for an effortless swing phase. Understanding how leg
mechanics are modulated across different tasks, including locomotion, is an important
problem in human motor control. It is also critical for understanding how to design artificial
legs that can begin to replicate the capabilities of the unimpaired human lower limb. Current
approaches for regulating knee stiffness in artificial devices are largely heuristic [1], since
quantitative estimates of how knee stiffness is modulated during locomotion are not
available. Here we propose a method for obtaining those estimates in a manner that does not
impede natural locomotor behaviors.
The mechanical properties of a limb and the joints within that limb can be characterized by
their mechanical impedance. The impedance of a joint describes the dynamic relationship
Pfeifer et al. Page 2
between an externally induced movement of a joint and the torques required to effect that
movement [2]. For small displacements, impedance can be described by the inertial, viscous
and elastic properties of the joint [3]. The viscous and elastic properties of a joint arise from
the passive properties of the anatomical structures within and surrounding the joint, as well
as from the activation-dependent properties of the muscles spanning the joint [4], [5]. Joint
viscoelasticity can be changed through changes in muscle activation, as determined from
previous experiments [6], [7]. The activation-dependent elasticity, or stiffness, of the
muscles spanning the joint is thought to be closely linked to the short-range stiffness
properties of muscle [8]. In many cases, there appears to be a close relationship between
muscle or joint stiffness and the corresponding viscosity [3], suggesting that if one can be
estimated, the other can be at least approximately inferred.
Experiments to determine joint stiffness typically involve perturbing the joint in a controlled
manner and measuring the corresponding motions and torques. These methods have been
applied to assess passive knee joint stiffness (i.e. when muscles are not activated) [2], [9],
[10]; research on active joint stiffness (i.e. joint stiffness attributed to activation-dependent
muscle elasticity) for the lower limb mainly focused on the ankle joint, for example
investigating dependence of stiffness on ankle angle [11], ankle torque [6], [12], influence of
co-contraction [13] or stiffness variation due to non-isometric contractions [14]. Fewer
studies have been published on active knee joint stiffness. Those that have been completed
considered only isometric conditions, and reported similar findings to those observed at the
ankle [7], [15].
Direct estimates of knee stiffness during locomotion would require a device that can perturb
the knee without impeding natural movements, which is at best difficult. Therefore, an
approach that does not hinder natural movements is needed. A few such methods have been
suggested for the upper limb, though these require subject-specific calibration, often
together with experimental measurements of stiffness [16], [17].
In this paper, we propose a method to estimate active knee joint stiffness using a detailed
musculoskeletal model of the leg [18] and a parameterized model for short-range stiffness
[19]. We also present initial results validating model performance during isometric
conditions, and evaluating its sensitivity to the assumptions within the model. We also make
initial estimates of how knee stiffness is modulated during gait, illustrating how this
approach may eventually be used to obtain parameters relevant to the design of transfemoral
prostheses for the restoration of locomotion.
II. Model
We propose a model-based method to estimate active knee joint stiffness without the need to
apply perturbations to the joint (Fig. 1). A key requirement of the approach is the estimation
of individual muscle forces. Many different combinations of muscle forces can produce the
same torque, as the knee joint is spanned by many muscles. Two different methods are
commonly used in the literature. Muscle forces are either estimated by distributing the joint
torque among the muscles spanning a joint using optimization techniques [20], which is also
called load sharing. Joint torque is usually found by inverse dynamics. The other commonly
used method involves EMG recordings, which yield estimates of muscle activity from which
muscle force can be computed. We evaluate both methods, as both have their advantages
and disadvantages; load sharing typically fails to predict any co-contraction, and EMG-
based estimates usually fail to estimate muscle forces that result in the measured joint
torque.
A. Musculoskeletal Model
We used the lower-limb model developed by Arnold et al. [18], which features seven
degrees of freedom and 43 muscles. We only considered the twelve muscles spanning the
knee joint (listed in Section II-D). Parameters of the model like maximal isometric force,
pennation angle, and optimal fiber length are based on a parameter study examining 35
cadavers [21]. It is implemented in OPENSIM, an open-source software for musculoskeletal
modeling and simulations [22]. The input of the model is the knee flexion angle, the outputs
are the moment arms of each muscle.
B. Musculotendon-Stiffness Model
The musculotendon stiffness was modeled as muscle stiffness in series with tendon stiffness,
as commonly done in the literature (e.g. by Morgan [23]). Tendon stiffness was modeled
based on a dimensionless force-strain curve as proposed by Zajac [24] and implemented by
Delp et al. [25]. The parameters were taken from the lower-limb model [18]. As described in
the introduction, we modeled the muscle’s short-range stiffness, which has been shown to be
proportional to the muscle force and inversely proportional to the optimal fiber length [19],
[26]. We used the proportionality constant identified by Cui et al. based on measurements
with feline muscles [19] which has already been successfully applied to the human arm [27].
C. Muscle Force Estimation Based on Load Sharing

Knee joint torque was determined using conventional inverse dynamics. To estimate
individual musculotendon force we used a static optimization based on the min-max
objective function, which in the case of a single degree of freedom joint reduces to an
analytical solution: All muscles which cannot actively contribute to the observed torque
(antagonists) will be set to zero; the active muscles (agonists), which are either the four
extensors or the eight flexors in the model, are equally activated [28]. We compared the
results to results obtained by minimizing the sum of squared muscle forces [29]. In both
cases we used the normalization factor for each musculotendon unit i, where is its
maximum isometric force, and αi is its pennation angle.
We also estimated the muscle activity aLS corresponding to the forces described above so
that the results of the load sharing algorithm could be compared to the experimentally
measured EMG, which is our proxy for muscle activity. This was done simply by dividing
the estimated muscle force by the maximum force attainable by that muscle, as defined in
our model [18]. This normalized measure of muscle activation was then compared to the
EMG-based estimate, described below.
D. Muscle Force Estimation Based on EMG

As an alternative to load sharing (Section II-C) an EMG-based approach was used to

estimate muscle forces. In contrast to load sharing, which relies on an inverse dynamic
model to obtain joint torque, the EMG-based approach directly operates on the level of
muscle activations. The downside of the latter is large measurement noise. Transcutaneous
EMG of seven easily accessible muscles (rectus femoris (rf), vastus lateralis (vl), vastus
medialis (vm), semitendinosus (st), biceps femoris long head (blh), gastrocnemius medialis
(gm), gastrocnemius lateralis (gl)) was recorded. The remaining five muscles in the model
spanning the knee joint, which are less easy to access, were estimated similar to Barrett et al.
[30]: the vastus intermedius (vi), the biceps femoris short head (bsh), the semimembranosus
(sm), the gracilis (gr) and the sartorius (sr). The respective activations a were: avi = 0.5 ·
(avm + avl), absh = ablh, asm = ast, agr = asr = 0.5 · (ablh + ast). EMGs were sampled at 1200
Hz after an analog bandpass filter between 5 Hz and 500 Hz. The recorded EMG was
rectified and filtered using an RMS filter with window size 200 ms. A delay of 50 ms
accounted for the delay between EMG signal and muscle force. The signals were normalized
to values obtained during maximum voluntary contractions (MVC), yielding the EMG-based
estimate of muscle activity aEMG. These estimates for each muscle i were multiplied by
to obtain the muscle force; is the muscle’s maximum isometric force and αi is
its pennation angle (values from the literature [21]).
III. Evaluation Methods

A. Load-Sharing Accuracy
As an indication of the accuracy with which muscle activity can be estimated, we quantified
the discrepancy between the two fundamentally different procedures, load sharing (Section
II-C) and EMG measurements (Section II-D).
Both computational procedures were evaluated on a single experimental data set. In these
experiments six subjects performed isometric contractions while seated in a device that
allows measurement of knee torque (described previously [31]); their foot was fixed and the
knee angle was between 80° and 86°. Subjects were required to maintain constant knee
torques of either 15% or 30% of their maximum voluntary contraction in extension and
flexion. Contractions at each level were maintained for six seconds; two measures were
obtained at each level. The first and last second of each trial were discarded and only the
four middle seconds were analyzed.
B. Model Validation under Isometric Conditions

To validate our model-based estimates of knee stiffness, we also obtained perturbation-
based estimates under isometric measurements. Data were collected from four subjects. All
procedures were approved by the Institutional Review Board at Northwestern University.
Perturbations were applied by a rotary torque motor configured as a rigid position servo.
Subjects were attached to the motor using a custom-made thermoplastic cast extending from
the toes to just below the knee. The knee was maintained at a flexion angle of θ = 60°, and
stochastic perturbations with a bandwidth of 7 Hz and a standard deviation of 0.5° were
applied for the purpose of stiffness estimation. Experimental trials lasted for 60 seconds,
during which subjects were instructed to produce constant knee flexion torques ranging from
−40 Nm up to 40 Nm in steps of 10 Nm.
Mechanical impedance was estimated nonparametrically [3]. The responses, which were
second-order, were then parameterized by the following equation.
The parameter I is inertia, b viscosity, and the static gain k corresponds to elasticity or
stiffness, which we compared to model estimates. The average stiffness when no torque was
exerted by the subjects, the passive joint stiffness kp, was subtracted from all measurements,
because the model only predicts active joint stiffness ka.
The experimentally identified values of the stiffness k = ka + kp were compared to model-

based estimates in matched conditions (fixed flexion angle θ = 60° at different flexion and
extension torques). Only muscle forces obtained through load sharing were considered, since
EMGs were not available for this data set. Specifically, the min-max criterion described in
Section II-C was used. Muscle parameters were taken from our model described above [18],
not specified for each subject in our experiments.
C. Stiffness during Gait

We also made initial attempts to estimate stiffness during gait. This was computed over five
gait cycles for one healthy male subject (28 years old, 70 kg, 180 cm). The data used for
these estimates was obtained during level-ground walking. It contained kinematic data (hip,
knee and ankle angles) obtained with an optical tracking system (Vicon), kinetic data
obtained using inverse dynamics and force platform measurements (Kistler), and EMG-
measurements from six muscles; the same muscles as described in Section II-D, but without
the rectus femoris. The activity of the rectus femoris was assumed to be equal to the mean of
vastus medialis and vastus lateralis, analog to the method already applied for the hardly
accessible muscles [30]. Resulting stiffness based on load sharing (Section II-C) and based
on EMG (Section II-D) was compared.
IV. Results
A. Load-Sharing Accuracy
There were substantial differences between the muscle activations estimated from the load-
sharing and EMG algorithms (Fig. 2). This was quantified by the standard deviation of these
differences across all muscles and subjects. This measure was high and increased slightly
with increasing activation. For the min-max criterion the standard deviation was 6.9%MVC
(mean value 1.4%MVC) for the tasks where subjects had to exert 15% of their maximal
torque capacity, and 9.4%MVC (mean value 3.7%MVC) for the 30%-torque tasks. Results
for load sharing based on the sum of normalized squared muscle forces were slightly worse
(mean 3.6%MVC, std. dev. 8.5%MVC for 15% torque, mean 7.8%MVC, std. dev.
12.7%MVC for 30% torque). Based on these results we performed a sensitivity analysis to
investigate influence of these different estimates of muscle force distribution on knee
stiffness, as described in the following section.
B. Sensitivity of Stiffness to Load-Sharing Accuracy

A Monte-Carlo analysis was used to investigate the sensitivity of joint stiffness estimates to
the distribution of muscle forces obtained using the EMG-based approach and the min-max
load-sharing algorithm. Model-based estimates of stiffness were determined for a torque
range from −50 Nm to 50 Nm flexion torque in steps of 2 Nm; the knee angle was fixed at
60° for all simulations. Muscle activations were estimated using load sharing with the min-
max criterion, which sets the antagonistic muscles to zero. Random errors were added to the
activations of the agonist muscles to simulate the range of differences that could be obtained
from kinetic and EMG-based estimates of muscle force. The standard deviation of the
activation errors was obtained from the results presented in Fig. 2 as follows. Because we
observed increasing errors with increasing activity, a linear model of the standard deviation
of aEMG – aLS was fit to the data. This model of the error standard deviation was used to
generate simulated muscle activations with a normal distribution about the min-max
estimate; the resulting muscle activations were not constrained to produce the original knee
torque. We performed 1000 simulations using these random activation errors.
Overall, the joint stiffness estimates were relatively insensitive to changes in muscle
activation. The greatest variation in predicted knee stiffness was at low force levels. This is
likely due to the nature of our Monte Carlo simulations, which incorporated a model of
activation uncertainty that was high even at low muscle forces. The standard deviation of the
estimated knee stiffness at 10 Nm of knee flexion was 32.6 Nm/rad, which is 41.7% of the
corresponding min-max estimate (Fig. 3). In contrast, the standard deviation of the modeled
activation error was 117.8% of the min-max activation estimate at this same knee torque.
This corresponds to a sensitivity of approximately 37%. At 50 Nm of knee flexion, the
standard deviation of the stiffness estimates was 11.3%, and that of the muscle activation
was 34.7% of the activation level. The sensitivity at these higher forces is similar (33%), yet
the accuracy of the stiffness estimation is greater due to the smaller range of relative muscle
activations.
C. Model Validation under Isometric Conditions

The knee stiffness measured during isometric conditions increased similar to that predicted
by the model, though the prediction accuracies differed between flexion and extension. Only
active stiffnesses were compared, since our model does not predict passive joint stiffness.
The experimentally measured passive stiffness ranged from 37 – 57 Nm/rad across the four
tested subjects. Model estimates reproduced experimentally identified stiffness better for
extension torques than for flexion torques (Fig. 4); on average, excluding the zero torque
conditions, experimentally measured stiffness was 17% (std. dev. 9%) lower than model
estimates for extension torques, and 42% (std. dev. 13%) lower for flexion torques.
D. Stiffness during Gait

The magnitudes of the active knee stiffness during gait estimated based on load sharing and
based on EMG were remarkably similar with a maximum of approximately 600 Nm/rad
(Fig. 5). This result is consistent with our finding that stiffness magnitude is relatively
insensitive to changes in muscle load sharing (Fig. 3). However, there was a substantial
difference in timing; stiffness based on EMG increased prior heel strike, most likely due to
preparatory co-contraction, while stiffness based on load sharing did not increase until after
heel strike, due to its inability to estimate muscle co-contraction. This is probably also the
reason why the minimal values of stiffness estimates based on load sharing were
substantially lower than EMG-based estimates.
V. Discussion
The objective of this study was to develop a model-based method for estimating knee
stiffness during locomotion. We have developed such a model, which predicts active
stiffness based on kinematic parameters of the knee and the force-dependent short-range
stiffness of the muscles acting about the knee. Our initial results demonstrate that model
predictions in isometric conditions are robust with respect to different approaches for
estimating muscle force, and that these predictions are consistent with many features of
stiffness measurements involving joint perturbations. These results also highlighted areas for
improvement, which are described below.
In our isometric validation, estimated values of stiffness for knee extension torques were
very close to experimentally determined values (Fig. 4). However, the model overestimated
stiffness for knee flexion torques. This could be partly explained by errors in load sharing,
though stiffness estimates during flexion fall outside of our confidence intervals related to
load-sharing accuracy (Section IV-B). Previous experimental findings have suggested that
active knee stiffness changes substantially with joint angle [7]. Our previous work has
demonstrated that model-based estimates of arm stiffness are also sensitive to changes in
joint angles and to muscle moment arms [27]. We have yet to complete similar sensitivity
studies for the knee, though we would expect similar findings. This would suggest that
precise angle measurements and subject-specific model parameters could lead to improved
accuracy.
Our model-based and experimental results were similar to the two previous studies reported
in the literature. Zhang et al. [7] demonstrated a similar torque-dependent increase in knee
stiffness to that shown in this manuscript, though stiffnesses reported by Zhang were
approximately 50% smaller than our estimates. The discrepancy is likely due to the
difference in perturbation amplitudes used in the two studies, since muscle short-range
stiffness is known to decrease with increasing perturbation amplitude [8]; the amplitudes
used by Zhang et al. were 140% larger than ours. These amplitude-dependent effects on
stiffness estimates have also been reported at the knee [15]. Larger amplitudes also may
evoke substantial reflexes, which are not explicitly represented in our model. However, if
such reflexes did contribute to knee stiffness during the continuous perturbations used to
estimate stiffness, they would do so via a continuous change in muscle activation and force.
Such steady-state changes would be captured in our current modeling approach.
Nevertheless, the role of transient reflexes could be quite important during locomotion, and
is something that will be considered in future validations of model performance during
locomotion.
The stiffness estimates during gait are remarkably similar in magnitude for both the EMG-
based approach and the load-sharing approach. The difference in timing, however, was
substantial. It is much bigger than what could be expected based on our sensitivity analysis
of stiffness to load-sharing errors in isometric conditions. This highlights the need to
validate how well these approaches predict muscle force over time, which we have not
attempted to incorporate in our current version of the model. EMG-based estimates may be
improved by incorporating a more complex EMG-to-force processing [32]; a velocity-
dependent component [33] might further improve estimates, as could a calibrated delay
between EMG signal and force onset, which has been shown to be influenced by many
parameters [34]. EMG-based estimates also can be improved by combining with load
sharing estimates to ensure that the muscle activations and joint torques are matched
throughout the gait cycle [35]. We are currently evaluating such approaches to obtain
reliable estimates of stiffness during gait, with the aim to incorporate these in variable-
impedance control of transfemoral prostheses. In this context, the role of joint viscosity will
also have to be addressed, especially its relative importance compared to joint elasticity.
VI. Conclusion
We have developed a model-based approach that allows quantitative assessment of active
knee joint stiffness without the need for applying perturbations to the joint. Our initial work
has focused on quantifying the accuracy of the model-based estimates and identifying how
they may be improved. Though there is still work to be done, the model can provide the first
estimates of knee stiffness during locomotion. These estimates are essential for
understanding how the mechanics of the leg contribute to physiological gait, and for
understanding how to design robotic prostheses that can begin to replicate that gait in
amputees.
Acknowledgments
This work was supported by the ETH Research Grant ETHIIRA, the Gottfried und Julia Bangerter-Rhyner Stiftung
and the Swiss National Science Foundation through the National Centre of Competence in Research Robotics. We
also thank Timothy M. Haswell, Robert Nguyen and Thomas Ukelo for their help.
References
[1]. Sup F, Varol H, Mitchell J, Withrow T, Goldfarb M. Preliminary evaluations of a self-contained
anthropomorphic transfemoral prosthesis. Mechatronics, IEEE/ASME Transactions on. Dec.2009
vol. 14(no. 6):667–676.
[2]. Crowninshield R, Pope M, Johnson R, Miller R. The impedance of the human knee. Journal of
Biomechanics. 1976; vol. 9(no. 8):529–530. IN1–IN3, 531–535.
[3]. Kearney R, Hunter I. System identification of human joint dynamics. Critical reviews in
biomedical engineering. 1990; vol. 18(no. 1):55–87. [PubMed: 2204515]

[4]. Johns R, Wright V. Relative importance of various tissues in joint stiffness. Journal of Applied
Physiology. 1962; vol. 17(no. 5):824–828.
[5]. Winters J, Stark L, Seif-Naraghi A-H. An analysis of the sources of musculoskeletal system
impedance. Journal of Biomechanics. 1988; vol. 21(no. 12):1011–1025. [PubMed: 2577948]
[6]. Hunter I, Kearney R. Dynamics of human ankle stiffness: Variation with mean ankle torque.
Journal of Biomechanics. 1982; vol. 15(no. 10):747–752. [PubMed: 7153227]
[7]. Zhang L-Q, Nuber G, Butler J, Bowen M, Rymer WZ. In vivo human knee joint dynamic
properties as functions of muscle contraction and joint position. Journal of Biomechanics. 1998;
vol. 31(no. 1):71–76. [PubMed: 9596540]
[8]. Rack PMH, Westbury DR. The short range stiffness of active mammalian muscle and its effect on
mechanical properties. The Journal of Physiology. 1974; vol. 240(no. 2):331–350. [PubMed:
4424163]
[9]. Such CH, Unsworth A, Wright V, Dowson D. Quantitative study of stiffness in the knee joint.
Annals of the Rheumatic Diseases. 1975; vol. 34(no. 4):286–291. [PubMed: 1190848]
[10]. Riener R, Edrich T. Identification of passive elastic joint moments in the lower extremities.
Journal of Biomechanics. 1999; vol. 32(no. 5):539–544. [PubMed: 10327008]
[11]. Gottlieb GL, Agarwal GC. Dependence of human ankle compliance on joint angle. Journal of
Biomechanics. 1978; vol. 11(no. 4):177–181. [PubMed: 690152]

[12]. Weiss P, Hunter I, Kearney R. Human ankle joint stiffness over the full range of muscle
activation levels. Journal of Biomechanics. 1988; vol. 21(no. 7):539–544. [PubMed: 3410857]
[13]. Nielsen J, Sinkjær T, Toft E, Kagamihara Y. Segmental reflexes and ankle joint stiffness during
co-contraction of antagonistic ankle muscles in man. Experimental Brain Research. 1994; vol.
102:350–358.
[14]. MacNeil J, Kearney R, Hunter I. Identification of time-varying biological systems from ensemble
data (joint dynamics application). Biomedical Engineering, IEEE Transactions on. Dec.1992 vol.
39(no. 12):1213–1225.
[15]. Tai C, Robinson C. Knee elasticity influenced by joint angle and perturbation intensity.
Rehabilitation Engineering, IEEE Transactions on. Mar.1999 vol. 7(no. 1):111–115.
[16]. Osu R, Gomi H. Multijoint muscle regulation mechanisms examined by measured human arm
stiffness and EMG signals. Journal of Neurophysiology. 1999; vol. 81(no. 4):1458–1468.
[PubMed: 10200182]
[17]. Shin D, Kim J, Koike Y. A myokinetic arm model for estimating joint torque and stiffness from
EMG signals during maintained posture. Journal of Neurophysiology. 2009; vol. 101(no. 1):387–
401. [PubMed: 19005007]
[18]. Arnold E, Ward S, Lieber R, Delp S. A model of the lower limb for analysis of human
movement. Annals of Biomedical Engineering. Feb.2010 vol. 38(no. 2):269–279. [PubMed:

19957039]
[19]. Cui L, Perreault EJ, Maas H, Sandercock TG. Modeling short-range stiffness of feline lower
hindlimb muscles. Journal of Biomechanics. 2008; vol. 41(no. 9):1945–1952. [PubMed:
18499113]
[20]. Erdemir A, McLean S, Herzog W, van den Bogert AJ. Model-based estimation of muscle forces
exerted during movements. Clinical Biomechanics. 2007; vol. 22(no. 2):131–154. [PubMed:
17070969]
[21]. Ward S, Eng C, Smallwood L, Lieber R. Are current measurements of lower extremity muscle
architecture accurate? Clinical Orthopaedics and Related Research. 2009; vol. 467:1074–1082.
[PubMed: 18972175]
[22]. Delp S, Anderson F, Arnold A, Loan P, Habib A, John C, Guendelman E, Thelen D. Opensim:
Open-source software to create and analyze dynamic simulations of movement. Biomedical
Engineering, IEEE Transactions on. Nov.2007 vol. 54(no. 11):1940–1950.
[23]. Morgan DL. Separation of active and passive components of short-range stiffness of muscle. Am
J Physiol Cell Physiol. 1977; vol. 232(no. 1):C45–49.
[24]. Zajac F. Muscle and tendon: properties, models, scaling, and application to biomechanics and
motor control. Crit Rev Biomed Eng. 1989; vol. 17(no. 4):359–411. [PubMed: 2676342]
[25]. Delp S, Loan J, Hoy M, Zajac F, Topp E, Rosen J. An interactive graphics-based model of the
lower extremity to study orthopaedic surgical procedures. Biomedical Engineering, IEEE
Transactions on. 1990; vol. 37(no. 8):757–767.
[26]. Walmsley B, Proske U. Comparison of stiffness of soleus and medial gastrocnemius muscles in
cats. J Neurophysiol. 1981; vol. 46(no. 2):250–259. [PubMed: 7264713]
[27]. Hu X, Murray WM, Perreault EJ. Modeling muscle contributions to multijoint mechanics.
American Society for Biomechanics Annual Conference. 2009
[28]. Rasmussen J, Damsgaard M, Voigt M. Muscle recruitment by the min/max criterion – a
comparative numerical study. Journal of Biomechanics. 2001; vol. 34(no. 3):409–415. [PubMed:
11182135]
[29]. Challis JH. Producing physiologically realistic individual muscle force estimations by imposing
constraints when using optimization techniques. Medical Engineering & Physics. 1997; vol.
19(no. 3):253–261. [PubMed: 9239644]
[30]. Barrett RS, Besier TF, Lloyd DG. Individual muscle contributions to the swing phase of gait: An
EMG-based forward dynamics modelling approach. Simulation Modelling Practice and Theory.
2007; vol. 15(no. 9):1146–1155.
[31]. Riener, R.; Angeli, M.; Fuhr, T.; Donaldson, N.; Wood, D. A new design of a multi-moment-
chair system; Proc. 6th IFESS Conference; Cleveland, USA. 2001. p. 288-290.
[32]. Lloyd DG, Besier TF. An EMG-driven musculoskeletal model to estimate muscle forces and
knee joint moments in vivo. Journal of Biomechanics. 2003; vol. 36(no. 6):765–776. [PubMed:
12742444]
[33]. Amarantini D, Rao G, Berton E. A two-step emg-and-optimization process to estimate muscle
force during dynamic movement. Journal of Biomechanics. 2010; vol. 43(no. 9):1827–1830.
[PubMed: 20206935]
[34]. Corcos DM, Gottlieb GL, Latash ML, Almeida GL, Agarwal GC. Electromechanical delay: An
experimental artifact. Journal of Electromyography and Kinesiology. 1992; vol. 2(no. 2):59–68.
[PubMed: 20719599]
[35]. Cholewicki J, McGill SM. EMG assisted optimization: A hybrid approach for estimating muscle
forces in an indeterminate biomechanical model. Journal of Biomechanics. 1994; vol. 27(no. 10):
1287–1289. [PubMed: 7962016]
Fig. 1.
Overview of method to determine active knee joint stiffness ka. Two different methods are
used for muscle force estimation: one is based on load sharing (Section II-C) and one is
based on EMG (Section II-D). Inputs of the algorithm are knee flexion angle θ and force
sensor or force-plate signals F , or EMG signals. In the static case the inverse-dynamics
block reduces to a simple kinematic transformation. Intermediate variables are: torque τ,
musculotendon forces fMT, musculotendon stiffnesses kMT. The vector of moment arms r is
used for muscle force estimation based on load sharing, and for the kinematic mapping from
musculotendon stiffness to joint stiffness.
Fig. 2.
Load-sharing accuracy for the min-max criterion and the cost function based on the sum of
squared normalized muscle forces (min (f)2). The difference between the activation obtained
using load sharing (aLS) and EMG-based activation (aEMG) is shown separately for the tasks
using 15% torque level and 30% torque level (Section III-A).
Fig. 3.
Active knee joint stiffness at knee flexion angle of θ = 60° for different knee torque levels
evaluated using different muscle activations. Stiffness resulting from load sharing based on
the min-max criterion and load sharing based on the sum of squared normalized muscle
forces (min (f)2) is shown, together with mean and standard deviation of the Monte-Carlo
analysis described in Section IV-B.
Fig. 4.
Active knee joint stiffness at knee flexion angle of θ = 60° for different knee torque levels.
The continuous line represents model estimates, circles represent stiffness identified from
perturbation experiments from four subjects. Average passive stiffness kp was

experimentally determined at zero knee torque and subtracted from all measurements in
order to compare to the model estimates of active knee joint stiffness (ka = k – kp).
Fig. 5.
Active knee joint stiffness for a healthy subject during level-ground walking evaluated using
conventional load sharing (Section II-C) and EMG-based muscle force estimation (Section
II-D). Mean and standard deviation of five gait cycles (from heel strike to heel strike) are
depicted.

Perturbation Experiment

Uploaded by

Copyright:

Available Formats

Perturbation Experiment

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Perturbation Experiment

Uploaded by

Copyright:

Available Formats

NIH Public Access

IEEE Int Conf Rehabil Robot. 2011 ; 2011: 5975474. doi:10.1109/ICORR.2011.5975474.

Model-Based Estimation of Active Knee Stiffness

ankle [7], [15].

C. Muscle Force Estimation Based on Load Sharing

D. Muscle Force Estimation Based on EMG

As an alternative to load sharing (Section II-C) an EMG-based approach was used to

III. Evaluation Methods

four middle seconds were analyzed.

B. Model Validation under Isometric Conditions

The experimentally identified values of the stiffness k = ka + kp were compared to model-

C. Stiffness during Gait

B. Sensitivity of Stiffness to Load-Sharing Accuracy

C. Model Validation under Isometric Conditions

D. Stiffness during Gait

biomedical engineering. 1990; vol. 18(no. 1):55–87. [PubMed: 2204515]

Biomechanics. 1978; vol. 11(no. 4):177–181. [PubMed: 690152]

movement. Annals of Biomedical Engineering. Feb.2010 vol. 38(no. 2):269–279. [PubMed:

perturbation experiments from four subjects. Average passive stiffness kp was

You might also like