SPECTRAL PROPERTIES OF MULTIPLE MYOELECTRIC SIGNALS 1

Prévia do material em texto

Neuroscience 355 (2017) 22–35
SPECTRAL PROPERTIES OF MULTIPLE MYOELECTRIC SIGNALS:
NEW INSIGHTS INTO THE NEURAL ORIGIN OF MUSCLE SYNERGIES
JULIEN FRÈRE *
University of Lorraine, Laboratory ‘‘Development, Adaption
and Disability” (EA 3450), Faculty of Sports Sciences, 30 rue
du Jardin Botanique, CS 30156, F-54603 Villers-lès-Nancy, France
Abstract—It is still unclear if muscle synergies reflect neural
strategies or mirror the underlying mechanical constraints.
Therefore, this study aimed to verify the consistency of
muscle groupings between the synergies based on the
linear envelope (LE) of muscle activities and those incorpo-
rating the time–frequency (TF) features of the electromyo-
graphic (EMG) signals. Twelve healthy participants
performed six 20-m walking trials at a comfort and fast
self-selected speed, while the activity of eleven lower limb
muscles was recorded by means of surface EMG. Wavelet-
transformed EMG was used to obtain the TF pattern and
muscle synergies were extracted by non-negative matrix
factorization. When five muscle synergies were extracted,
both methods defined similar muscle groupings whatever
the walking speed. When accounting the reconstruction
level of the initial dataset, a new TF synergy emerged. This
new synergy dissociated the activity of the rectus femoris
from those of the vastii muscles (synergy #1) and from the
one of the tensor fascia latae (synergy #5). Overall, extract-
ing TF muscle synergies supports the neural origin of mus-
cle synergies and provides an opportunity to distinguish
between prescriptive and descriptive muscle synergies.
� 2017 IBRO. Published by Elsevier Ltd. All rights reserved.
Key words: locomotion, motor module, neural control, non-
negative matrix factorization, wavelet analysis.
INTRODUCTION
Low-dimensional motor modules formed by muscles
activated simultaneously, named muscle synergies,
have been proposed to simplify the construction of
motor behaviors (Ivanenko et al., 2003; d’Avella and
Bizzi, 2005; Ting and McKay, 2007; Torres-Oviedo and
Ting, 2007; Ting and Chvatal, 2010). To face the great
http://dx.doi.org/10.1016/j.neuroscience.2017.04.039
0306-4522/� 2017 IBRO. Published by Elsevier Ltd. All rights reserved.
*Fax: +33 372 746 712.
E-mail address: julien.frere@univ-lorraine.fr
Abbreviations: BF, biceps femoris; CNS, central nervous system; CV,
coefficient of variation; EMG, electromyographic; Gmax, gluteus
maximus; LE, linear envelope; LG, gastrocnemius lateraleris; MG,
gastrocnemius medialis; NMF, non-negative matrix factorization; RF,
rectus femoris; SOL, soleus; ST, semitendinosus; TA, tibialis anterior;
TF, time–frequency; TFL, tensor fascia latae; VAF, variance accounted
for; VL, vastus lateralis; VM, vastus medialis; wt, wavelet.
22
amount of degrees of freedom of the human body and
for a given motor task, the synchronous muscle synergies
allow a decrease in the number of variables controlled by
the central nervous system (CNS). Thus, rather than indi-
vidual muscles, it seems that the primary neural element
to produce movement is muscle synergy, which is itself
controlled by a higher neural command that functionally
modulates the pattern of activation of multiple muscles
(Rana et al., 2015). In human locomotion, it has been
found that a set of a limited number of muscle synergies
(four to five) explain the multi-muscle activation and are
found to represent functional subtasks of the gait cycle
(Ivanenko et al., 2004; Neptune et al., 2009; Chvatal
and Ting, 2012). Across a variety of constraints, it has
been suggested that the time-varying modulation of simi-
lar muscle groupings (i.e., motor modules) may represent
the integration of sensory inflows (Cheung et al., 2005;
Hug et al., 2011; Safavynia and Ting, 2012, 2013; van
den Hoorn et al., 2015). Also, the number of muscle syn-
ergies extracted for a given task has been suggested to
express the complexity of the neuromuscular control of
the motor behavior (Clark et al., 2010). Therefore, muscle
synergies may be an integrative, useful tool to analyze the
neural structures (spinal cord, brainstem, motor cortex)
underlying motor behaviors and to quantify changes
related to motor deficit or to the efficiency of any given
therapy or rehabilitation program (Safavynia et al., 2011;
Ting et al., 2012, 2015; Routson et al., 2013; Roemmich
et al., 2014; Wenger et al., 2016).
However, the neural origin of muscle synergies is still
a matter of debate within the current literature. It is
unclear if muscle synergies effectively reflect the CNS
strategies (Bizzi and Cheung, 2013) or simply mirror the
underlying mechanical constraints (i.e., descriptive syn-
ergies) (Kutch and Valero-Cuevas, 2012; de Rugy et al.,
2013). For instance, muscle synergies may be
movement-related since non-neural constraints, such as
a low-dimensional space of muscle–tendon length
change, may explain the dimensionality reduction of
multi-muscle activations (Kutch and Valero-Cuevas,
2012). According to Valero-Cuevas (2016) ‘‘The question
then is, how can one infer prescriptive synergies (i.e., the
existence of synergies of neural origin) from experimental
data that naturally exhibit descriptive synergies? This is
the heart of the debate in this area at the moment.” Using
the spectral properties of the surface electromyographic
(EMG) signals has been found to be another approach
to determine the neural structures underlying the muscle
activation. More specifically, frequency bands from
http://dx.doi.org/10.1016/j.neuroscience.2017.04.039
mailto:julien.frere@univ-lorraine.fr
http://dx.doi.org/10.1016/j.neuroscience.2017.04.039
J. Frère / Neuroscience 355 (2017) 22–35 23
EMG-EMG coherence might reflect subcortical [�10 Hz;
Grosse and Brown (2003), Boonstra et al. (2009)] or cor-
tical [20–60 Hz; Grosse et al. (2002)] pathways. For
instance, during a postural task Danna-Dos-Santos
et al. (2014, 2015) found significant peaks of intermuscu-
lar coherence within the low-frequency bands (0–5 and 5–
20 Hz) among muscles grouped in functional synergies.
These results corroborated the neural origin hypothesis
of muscle synergies to lower the dimensionality of the
neuromuscular control. In combination with the extraction
of muscle synergies during a pedaling task, De Marchis
et al. (2015) determined that solely the knee extensors
muscle synergy had a significant peak of EMG-EMG
coherence within the 30–60-Hz frequency band, likely
reflecting a cortically mediated muscle synergy to produce
power during the descending phase of the pedaling cycle.
This result also suggested that the other muscle syn-
ergies would be descriptive of the mechanical constraints
of the pedaling task.
Therefore, coupling intermuscular coherence analysis
with the extraction of muscle synergies might be a
promising approach to discriminate prescriptive muscle
synergies from descriptive ones. However, such a
method does not allow investigating any change in
frequency as function of time of activation. Indeed, the
extraction of muscle synergies is a time-domain analysis
while the EMG-EMG coherence provides correlates
solely in the frequency-domain. Moreover, it has been
showed that a similar EMG envelope could be explained
by different underlying time–frequency patterns
(Wakeling, 2004; Hodson-Tole and Wakeling, 2007;
Frère et al., 2012a). Consequently, a method able to
extract synergies composed of muscles sharing similar
time–frequency features would provide new evidences
relative to their potential neural origin.
The aim of this study was to propose a new method of
muscle synergies extraction that incorporates the spectral
properties (i.e., time–frequency domain) of multiple
muscle activities and to verify the consistency of muscle
groupings with muscle synergies based on the global
muscle activities (i.e., time domain) during human gait.
In considering that the muscle synergies are of neural
origin, it was hypothesizedthat the muscle vectors (i.e.,
motor modules) were similar across the two methods of
extraction, whatever the walking velocity. In case of
discrepancy between the methods of muscle synergy
extraction, one might consider the time–frequency
muscle synergies as a new tool to distinguish
prescriptive from descriptive muscle synergies.
EXPERIMENTAL PROCEDURES
Participants
Twelve volunteers (10 men and 2 women, age: 31.9
± 9.3 years, height: 178 ± 8 cm, body mass: 77
± 10.8 kg) participated in this study. They were
informed of the purpose of the study and methods used
before providing written consent. The experimental
procedure was carried out in accordance with the
principles of the Declaration of Helsinki.
Protocol
Participants were asked to walk overground within a
corridor at a self-selected speed. Two walking self-
selected speeds were assessed: a comfort condition
and a fast condition. For each condition, each
participant walked at least 20 m three times, in order to
assess 10 walking cycles per trial (the first and last
walking cycles were not retained). At least, 30 walking
cycles were recorded and analyzed for each condition.
All participants began with a comfort trial but the order
of the five following trials was randomized. A walking
cycle was defined as the time between two consecutive
heel strikes of the same foot.
Materials and data collection
The 20-m walking distance was materialized by means of
three pairs of ground cone markers, each 10-m apart. The
20-m walking time was manually recorded with a digital
chronometer between two instants: when the foot left
the ground at the walking initiation and when the
participant crossed the last pair of ground markers.
The activity of eleven muscles of the right side of the
body was simultaneously recorded: tibialis anterior (TA),
soleus (SOL), gastrocnemius lateraleris (LG),
gastrocnemius medialis (MG), vastus lateralis (VL),
rectus femoris (RF), vastus medialis (VM), biceps
femoris (long head, BF), semitendinosus (ST), tensor
fascia latae (TFL), and gluteus maximus (Gmax). The
EMG activity was recorded using wireless electrodes
(Delsys TrignoTM Wireless System, Boston, MA, USA)
with an inter-electrode distance of 10 mm. The
electrodes were placed longitudinally with respect to the
underlying muscle fiber arrangement (de Luca, 1997)
and were located according to recommendations of Sur-
face EMG for Non-Invasive Assessment of Muscles
[SENIAM, Hermens et al. (2000)]. Before applying elec-
trodes, the skin was shaved and cleaned with alcohol to
minimize impedance. Raw EMG signals were preampli-
fied (gain 300, bandwidth 20–450 Hz) at a sample rate
of 2000 Hz. Two triaxial accelerometers (Delsys TrignoTM
Wireless System, Boston, MA) were placed at the level
of the third metatarsal and of the heel (sampling rate
148.18 Hz). All the raw signals (EMG and accelerations)
were synchronized and stored in digital format using
EMGworks� Acquisition software (Delsys, Boston, MA).
Data were then processed offline with custom-built Mat-
lab� scripts (The Mathworks, Natick, Massachussetts,
USA).
Data analysis
From the 20-m walking time, the mean walking velocity (in
m.s�1) was calculated. The longitudinal acceleration (x-
axis) of the heel accelerometer and the frontal
acceleration (z-axis) of the metatarsal accelerometer
were used to determine each heel contact with the
ground. Both signals of acceleration were rectified and
smoothed with a zero lag low-pass filter (5 Hz,
Butterworth filter, 2nd order). Both signals had common
peaks which corresponded to the heel strikes and thus
24 J. Frère / Neuroscience 355 (2017) 22–35
were used to define the walking cycles. For each trial, the
total time spent to perform the 10 walking cycles was
retained to compute the mean cycle frequency (cycle.
s�1). The mean cycle length (m.cycle�1) was computed
from the mean walking velocity divided by the mean
cycle frequency. The average value across the three
trials per condition was finally calculated for each
variable (mean velocity, cycle frequency and cycle
length) and for each participant.
Raw EMG signals were band-pass filtered (25–
450 Hz, Butterworth filter, 4th order) and then processed
across a wavelet filter bank with a nonlinear scale
function (von Tscharner, 2000; Frère et al., 2012b). A
set of nine wavelets was used with center frequencies,
cf, ranging from 38�Hz (wt #1) to 395�Hz (wt #9). The
wavelet intensity corresponded to the time-varying power
of the signal resolved for one wavelet only (Fig. 1B). The
wavelet transformation of the EMG signal resulted on a
time–frequency map of nine wavelet intensities with equal
length (i.e., the number of time points).
Muscle synergies were extracted in two different
ways: (i) linear envelope (LE) muscle synergies as
Fig. 1. Representative individual example of surface EMG signals proces
extraction. (A) Illustrative sample of raw signal of the 11 lower limb muscles
frame), the signal was transformed (B) into its TF domain by means of a wave
all the wavelets, that each owns a center frequency. The TF map is normalize
between 0 and 1 (right panel) for each wavelet domain (wt #1 to wt #9). (C) T
time point providing the envelope. In the building of the initial data matrix for L
and normalized in time and in amplitude. (D) After the time normalization o
intensities which then allows the building of the initial data matrix for TF mus
classically done in the literature and (ii) time–frequency
(TF) muscle synergies. For both processes of muscle
synergy extraction, inter-cycle variability was taken into
account (Clark et al., 2010; Oliveira et al., 2014) as the ini-
tial data matrix compiled three sets of 10 consecutive
walking cycles (Fig. 1A) as previously prescribed (Hug
et al., 2011; Frère and Hug, 2012; van den Hoorn et al.,
2015). For the LE muscle synergies, the initial data matrix
was computed as follows: the envelope for each walking
cycle was determined from the total intensity which corre-
sponded to the sum of all the wavelet intensities. Then,
the total intensity was smoothed with a zero lag low-
pass filter (9 Hz, Butterworth filter, 4th order), time-
normalized to obtain 100 data points for each walking
cycle and normalized to its peak value (Fig. 1C). The ini-
tial data matrix was thus an 11-row (number of muscles)
and 3000-column (30 walking cycles of 100 time points)
matrix. For the TF muscle synergies, the initial data matrix
was computed as follows: the whole TF map of each mus-
cle and walking cycle was interpolated to 100 time points
and normalized to its peak value. Then, each normalized
TF map was reshaped into a long TF row vector by con-
sing in the building of the initial data matrix for muscle synergies
during 30 walking cycles. For each muscle and walking cycle (gray
let bank. The TF map (left panel) represents the changes in intensity of
d to its peak value providing values of intensity constrained in a range
he total intensity corresponds to the sum of wavelet intensities in each
E muscle synergy extraction, the envelope of each cycle is smoothed
f the TF map, a TF row vector is built by concatenating the wavelet
cle synergy extraction.
J. Frère / Neuroscience 355 (2017) 22–35 25
catenating the wavelet intensities one next to the other
(Fig. 1D). Therefore, the initial data matrix was an 11-
row (number of muscles) and 27,000-column (30 walking
cycles of nine wavelet intensities of 100 time points)
matrix. Non-negative matrix factorization (NMF) was per-
formed from these two types of datasets (LE vs. TF initial
data matrix). For this purpose, the multiplicative update
rules algorithm (Lee and Seung, 2001) was used to
extract muscle synergies (Matlab nnmf function;
option = ‘mult’), as follows:
E ¼ WCþ e ð1Þ
where E is an i-by-j initial matrix (i= number of muscles
and j= number of time points), W is an i-by-n matrix
(n= number of synergies), C is a n-by-j matrix and e is
ani-by-j matrix. W represents the muscle synergy
vectors matrix (also called motor modules) which
represents the relative weighting of each muscle within
each synergy, C is the synergy activation coefficients
matrix (also called motor primitives) which represents the
recruitment of the muscle synergy over time and e is the
residual error matrix. The algorithm is based on iterative
updates of an initial random guess of W and C that
converge to a local optimal matrix factorization [see Lee
and Seung (2001) for more details]. To avoid local minima,
the algorithm was repeated 20 times for each participant.
The lowest cost solution was kept (i.e., minimized squared
error between original and reconstructed EMG patterns).
One assumption of the NMF algorithm is that the
number of muscle synergies (n) to be extracted should
be given a priori. As previously done in the literature,
the determination of n was classically performed from
the changes in total Variance Accounted For [VAF;
Torres-Oviedo et al. (2006), Oliveira et al. (2014)] as func-
tion of the number of muscle synergies. One method con-
sisted in varying n between 1 and 11 and then selecting
the least number of synergies that accounted for >90%
of VAF or until adding an additional synergy did not
increase VAF by >5% (Clark et al., 2010). Although often
used, this method is largely dependent to the absolute
VAF values. This might be an issue for the current study,
since it has been shown that the VAF values also depend
on the dimensionality of the initial data matrix (Oliveira
et al., 2014). Therefore, one might expect a bias in the
determination of n between LE and TF muscle synergies
due to the large difference in the dimensionality of the
respective initial data matrix. To take this effect into
account, the proposed method by Cheung et al. [named
knee point method herein; Cheung et al. (2009)] might
be more appropriate since it focused on the changes in
slope of the VAF-number of synergies curve rather than
using the absolute VAF values. Briefly, the VAF-number
of synergies curve was constructed from both the original
EMG dataset and an unstructured EMG dataset gener-
ated by randomly shuffling the original dataset across
time and muscles. n was then defined as the point beyond
which the original-slope drops below 75% of the
surrogate-slope. It corresponds to the number beyond
which any further increase in the number of extracted syn-
ergies yields a VAF increase smaller than 75% of that
expected from chance. Consequently, two analyses were
conducted: (i) because most previous studies found that
five muscle synergies accounted for muscle activity dur-
ing human walking (Ivanenko et al., 2004; Cappellini
et al., 2006; Oliveira et al., 2014; van den Hoorn et al.,
2015), a value of n= 5 was selected for both methods
(LE and TF muscle synergies) and both conditions (com-
fort and fast walking) and (ii) n was determined from the
knee point method for each participant, conditions and
methods.
In addition to these comparisons between both
methods, some complementary analyses were
conducted to identify if the dimensionality of the TF
initial data matrix and the amount of intensity among the
wavelet bands could influence the composition of the
muscle synergies. Indeed, Oliveira et al. (2014) deter-
mined that increasing the dimensionality of the initial data
matrix would decrease the reconstruction quality (VAF).
Therefore, the TF muscle synergy vectors were extracted
from the TF maps of solely four walking cycles in order to
obtain a TF initial data matrix of 3600 time points (closed
to the 3000 time points of the LE initial data matrix). To
create this reduced TF initial data matrix, the four walking
cycles were randomly selected among the 30 cycles and
the procedure was iterated 20 times. This led to obtain, for
each participant and each walking condition, 20 different
reduced TF initial data matrices, each subjected to a mus-
cle synergy extraction by the NMF algorithm. Finally, the
average over the 20 muscle synergy vectors was retained
for each participant and walking condition. Moreover, the
amount of intensity among the wavelets could bias the TF
muscle synergy extraction. Indeed, by normalizing the TF
map of each cycle by its peak value, some wavelets had
much more intensity than others. As the pattern of activa-
tion of these high intensity frequency bands were strongly
similar to the linear envelope of the EMG signal (Hodson-
Tole and Wakeling, 2007, 2009), it is likely that the NMF
algorithm replicated the LE method for muscle synergy
extraction. Consequently, the TF map of each walking
cycle was normalized in amplitude with each wavelet
band normalized by its peak value. Such normalization
led the NMF algorithm to extract muscle synergies that
fully took into account the activation of all the wavelets,
since each frequency band had an equal power within
the TF map. Also, the muscle synergy extraction was car-
ried out for each wavelet separately, to obtain the motor
modules related to one frequency band. This would bring
an additional opportunity to verify if the muscles involved
in one muscle synergy shared common time–frequency
features or if there were some discrepancies among the
wavelets.
Statistical analysis
Spatiotemporal data (mean velocity, cycle frequency and
length) for both walking conditions were found to be
normally distributed (Shapiro–Wilk tests) with
homogenous variances (two-sample F-tests). Therefore,
a paired student t-test was used to compare the
spatiotemporal data between the comfort and fast
conditions. A two-way ANOVA for repeated measures
(factors: extraction methods and walking conditions)
was used to assess the effect of these two factors on
26 J. Frère / Neuroscience 355 (2017) 22–35
the number of muscle synergies to be extracted. The
scalar product normalized by the product of the norms
of each vector [q; Oliveira et al. (2014)] was used as sim-
ilarity criterion for the muscle synergy vector matrix
between two methods of extraction for each participant.
A pair of muscle synergy vectors was considered to be
similar if q � 0.80 (Oliveira et al., 2014). When five muscle
synergies were extracted systematically, the similarity of
the muscle synergy vector matrix was carried out
between the reference method (i.e., LE method) and the
four TF datasets [TF maps from 30 and 4 walking cycles
normalized by their peak values (TF30p and TF4p, respec-
tively); TF maps from 30 and 4 walking cycles normalized
with the peak value of each wavelet (TF30w and TF4w,
respectively)] and between the reference method and
one of the nine muscle synergy vector matrices (wt#1 to
wt#9). When the number of muscle synergies was
according to the knee point method, the reference method
for assessing the similarity of the muscle synergy vector
matrices was the TF30p. A two-way ANOVA for repeated
measures was used to assess the effect of walking condi-
tions (Comfort vs. Fast), and of muscle synergy extraction
methods (LE, TF30p, TF4p, TF30w and TF4w) on the recon-
struction quality (VAF). A two-way ANOVA for repeated
measures was used to assess the effect of walking condi-
tions (Comfort vs. Fast), and of wavelet bands (wt#1 to
wt#9) on the reconstruction quality (VAF). A four-way
ANOVA for repeated measures was used to assess
potential changes in the similarity of muscle synergies
(q-values), according to the effect of the walking condition
(Comfort vs. Fast), the TF maps normalization (TF map
peak vs. wavelet peak), the number of walking cycles
(30 vs. 4) and the synergy number (W#1 to W#5). A
three-way ANOVA for repeated measures was used to
assess potential changes in the similarity of muscle syn-
ergies (q-values), according to the effect of the walking
condition (Comfort vs. Fast), the wavelet bands (wt#1 to
wt#9) and the synergy number (W#1 to W#5). As previ-
ously done (Ivanenko et al., 2004; Cappellini etal.,
2006; Turpin et al., 2011), q-statistics were based on Z-
transformed values. Post hoc analyses were made with
Scheffe’s tests. The level of significance was p= 0.05.
Fig. 2. Changes in Variance Accounted For (VAF in %) as a function
of the number of muscle synergies extracted from the LE initial data
matrix during the comfort (A) and the fast (B) walking conditions and
from the TF initial data matrix during the comfort (C) and the fast (D)
walking conditions. For each graph, data are group mean values
(±SD) for both VAF calculated from the initial data matrix (bold solid
line) and from the unstructured initial data matrix generated by
randomly shuffling the original data matrix across time and muscles
(thin dotted line). The arrow denotes the number of muscle synergies
to be extracted according to the knee point method.
RESULTS
Spatiotemporal walking data
Excellent to good repeatability (assessed by means of
coefficient of variation, CV) was found across the trials
of the comfort and fast walking conditions for mean
velocity (CV range: 1.0–8.2% and 0.6–8.5%,
respectively), mean cycle frequency (CV range: 0.5–
3.6% and 0.5–6.8%, respectively), and length (CV
range: 0.4–5.4% and 0.1–4.0%, respectively). Mean
walking velocity was significantly lower (p< 0.001)
during the comfort condition in comparison with the fast
condition (1.28 ± 0.13 and 2.03 ± 0.12 m.s�1,
respectively). Mean cycle frequency (0.94 ± 0.06 and
1.20 ± 0.07 cycle.s�1, for comfort and fast conditions,
respectively) and length (1.35 ± 0.10 and 1.69
± 0.09 m.cycle�1, for comfort and fast conditions,
respectively) were also significantly different (p< 0.001).
Number of muscle synergies
As explained above (Data analysis in the Methods
section), the muscle synergies were extracted according
to (i) the literature (n= 5) and (ii) the knee point
method. In the first case, extracting five muscle
synergies during the comfort walking condition
accounted for a mean VAF of 90.1 ± 1.2% (range:
88.5–92.1%) and of 71.9 ± 1.2% (range: 70.0–73.1%)
for LE and TF methods, respectively (Fig. 2A, C).
During the fast walking condition, five muscle synergies
accounted for a mean VAF of 90.2 ± 1.5% (range:
87.8–92.8%) and of 72.6 ± 1.8% (range: 69.1–75.5%)
for LE and TF methods, respectively (Fig. 2B, D). In the
second case, according to the knee point method, a
mean n value of 4.9 ± 0.9 and 6.0 ± 0.7 was found for
LE and TF methods, respectively, for the comfort
condition. A mean n value of 4.8 ± 0.7 and 5.9 ± 1.1
for LE and TF methods, respectively, was found for the
fast condition. A two-way ANOVA for repeated
measures (factors: extraction methods and walking
conditions) solely found a main significant effect for the
J. Frère / Neuroscience 355 (2017) 22–35 27
factor methods (p= 0.028). This result confirmed that,
whatever the locomotion speed, five and six muscles
synergies accounted for the lower limb muscle
coordination during human gait when considering the
global activity (LE) and the spectral properties (TF) from
multiple myoelectric signals, respectively. Therefore,
according to the knee point method, six TF muscles
synergies were extracted and accounted for a mean
VAF of 78.1 ± 1.4% (range: 76.3–80.5%) and of 78.4
± 1.5% (range: 75.4–80.3%) for comfort and fast
walking conditions, respectively (Fig. 2C, D).
LE vs. TF muscles synergies
The extracted muscle synergies between both methods
(LE vs. TF) are depicted in Fig. 3, which showed the
communalities in both spatial (W) and time (C) domains.
The similarity of muscle synergy vectors between both
methods of extraction was assessed by the scalar
product. High to very high values were found for the first
to fourth muscle synergy vectors while the fifth
presented lower values (Table 1). For each muscle
Fig. 3. Mean values of the five muscle synergies extracted during the comfo
(bold line) from the five muscle synergies (C#1 to C#5) extracted from the LE
panel: Mean (±SD) LE (black bars) and TF (white bars) muscle synergy vect
comfort walking condition. Right panel: Mean TF synergy activation coefficien
method during a walking cycle. Time-frequency synergy activation coefficient
with high intensities denoted by dark shading.
synergy (#1 to #5), no significant difference in q-values
(p> 0.99) was found between the comfort and fast
walking condition. Therefore, between both walking
conditions and on considering the threshold q-value of
0.80 for similarity, four pairwise comparisons (out of 24
possibilities, i.e., 17%) were considered as different for
synergy #1, as well as 0 (0%) for synergy #2, 3 (13%)
for synergy #3, 1 (4%) for synergy #4, and 9 (38%) for
synergy #5. According to these distributions of dissimilar
muscle synergy vectors, one can note that the
synergies #1 and #5 were not systematically consistent
among methods of synergy extraction, walking
conditions and participants. This highlights a
discrepancy in the composition of these two muscle
synergies and might suggest a non-physiological source.
The knee point method confirmed that five muscle
synergies accounted for the global activity of the eleven
lower limb muscles (LE initial data matrix), while six
muscle synergies were necessary to account for the TF
domain of these same muscles during human gait. A
priori, this seemed to confirm the discrepancy previously
found between both methods of synergy extraction,
rt walking condition. Left panel: Mean synergy activation coefficients
method during a walking cycle. Gray areas represent ±1 SD. Middle
ors from the five muscle synergies (W#1 to W#5) extracted during the
ts from the five muscle synergies (C#1 to C#5) extracted from the TF
s are shown as a function of time (% of walking cycle) and frequency,
Table 1. Similarity of the muscle synergy vectors (W) between both methods of extraction (LE vs. TF) for each condition of walking
Comfort condition Fast condition
Mean (S.D.) Range Mean (SD) Range
Scalar product (q) W #1 0.89 (0.12) 0.69–1.00 0.92 (0.06) 0.83–0.98
W #2 0.99 (0.00) 0.99–1.00 0.98 (0.03) 0.91–1.00
W #3 0.94 (0.08) 0.74–1.00 0.89 (0.15) 0.54–1.00
W #4 0.98 (0.01) 0.95–0.99 0.94 (0.10) 0.68–1.00
W #5 0.76 (0.33) 0.21–1.00 0.67 (0.37) 0.03–1.00
28 J. Frère / Neuroscience 355 (2017) 22–35
especially for synergies #1 and #5. It appeared that the
first four muscle synergies remained quite similar. The
RF muscle, which was implied in both synergies #1 and
#5 when extracting five muscle synergies, composed
alone the synergy #5 when six muscle synergies were
extracted. When five muscle synergies were extracted,
the TFL muscle composed with the RF muscle the hip
flexor muscle synergy (#5), but constituted alone the
muscle synergy #6 when extracting six muscle synergies.
Effect of the characteristics of the TF initial data
matrix
When five muscle synergies were extracted, the muscle
synergy vectors were similar between the LE methods
and the four types of TF datasets (TF30p, TF4p, TF30w,
and TF4w). Indeed, there was no main effect of all the
factors assessed (Fig. 4A–D): walking condition
(Condition vs. Fast, p= 0.999), TF maps normalization
(TF map peak vs. wavelet peak, p> 0.999), number of
walking cycles (30 vs. 4, p= 0.999) and synergy
number (W#1 to W#5, p> 0.999). Relative to the
reconstruction quality (Fig. 4E), there was a significant
main effect of the methods of synergy extraction on the
VAF values (p< 0.001) but no main effect of walking
condition (p= 0.424). Post-hoc tests showed that the
LE method provided the highest VAF values and that
TF30w and TF4w methods improved the reconstruction
quality in comparison with the TF maps normalized by
their peak value (TF30p and TF4p). When five muscle
synergies were extracted from the nine datasets, each
composed of the intensity patterns of one wavelet band,
the muscle synergy vectors were similar with those
extracted from the LE method (Fig. 5A–D). No main
effect of the dataset type (LE vs. wt#1 to wt#9,
p> 0.999),of the walking condition (Comfort vs. Fast,
p= 0.999) and of the synergy number (W#1 to W#5,
p> 0.999) was found on the muscle synergy
similarities. Also, there was a main effect of the wavelet
bands (p< 0.001) but not of the walking condition
(p= 0.860) on the reconstruction quality (Fig. 5E). An
interaction effect was found (p= 0.013) which
underlined that VAF values were higher in the fast
walking condition for the lowest frequency bands (wt#2
to wt#5), while the VAF values were higher in the
comfort condition for the highest frequency bands (wt#7
to wt#9).
When six muscle synergies were extracted (Fig. 6),
the muscle synergy vectors were similar between the
TF30p methods and the three other types of TF datasets
(TF4p, TF30w, and TF4w). Statistically, the results were
identical to those obtained when five muscle synergies
were extracted. Indeed, there was no main effect of
walking condition (Condition vs. Fast, p> 0.999),
synergy extraction method (TF4p, TF30w, and TF4w,
p> 0.999), and of the synergy number (W#1 to W#6,
p> 0.999). This was also the case for the
reconstruction quality (Fig. 6E), there was a significant
main effect of the methods of synergy extraction on the
VAF values (p< 0.001) but no main effect of walking
condition (p= 0.806). Post-hoc tests showed that the
TF30p method provided the lowest VAF values and that
TF30w and TF4w methods provided higher values in
comparison with TF30p and TF4p methods. When six
muscle synergies were extracted from the nine
datasets, each composed of the intensity patterns of
one wavelet band, the muscle synergy vectors were
similar with those extracted from the TF30p method
(Fig. 7A–D). No main effect of the dataset type (TF30p
vs. wt#1 to wt#9, p> 0.999), of the walking condition
(Comfort vs. Fast, p= 0.998) and of the synergy
number (W#1 to W#5, p> 0.999) was found on the
muscle synergy similarities. Also, there was a main
effect of the wavelet bands (p< 0.001) but not of the
walking condition (p= 0.988) on the reconstruction
quality (Fig. 7E). An interaction effect was found
(p< 0.001) which underlined that VAF values were
higher in the fast walking condition for the lowest
frequency bands (wt#3 to wt#5), while the VAF values
were higher in the comfort condition for the highest
frequency bands (wt#6 to wt#9).
DISCUSSION
The aim of this study was to verify the consistency of the
muscle groupings between two methods of muscle
synergy extraction (LE vs. TF muscle synergies) during
human gait. When five muscle synergies were
extracted, both methods provided similar muscle
groupings whatever the walking condition. Otherwise, it
appeared that an additional muscle synergy emerged
when accounted for the spectral features of the initial
dataset. This main result might suggest that extracting
time–frequency muscle synergies could be an
opportunity to distinguish between prescriptive and
descriptive muscle synergies.
Neurophysiological interpretations
In the current study, both LE and TF muscle synergies
were extracted during two walking conditions: a comfort
and a fast walking condition, each at a self-selected
Fig. 4. Mean (±SD) LE muscle synergy vectors (black bars) and those obtained from four TF datasets (dark gray to white bars for TF30p, TF4p,
TF30w and TF4w, respectively) when five muscle synergies (W#1 toW#5) were extracted during the comfort (A) and fast (B) walking condition. Mean
(±SD) values of similarity between the LE muscle synergy vectors and those extracted from each of the four TF datasets (dark gray to white bars
for TF30p, TF4p, TF30w and TF4w, respectively) during the comfort (C) and fast (D) walking condition. (E) Total variance accounted for (VAF, in%)
according to the method of muscle synergy extraction and walking condition. *Significant difference with p< 0.001.
J. Frère / Neuroscience 355 (2017) 22–35 29
speed. Previous results within the literature indicated that
under different mechanical constraints, consistency in the
composition of LE muscle synergies was found in a
variety of postural and locomotor behaviors (Ivanenko
et al., 2004; Cappellini et al., 2006; Clark et al., 2010;
Torres-Oviedo and Ting, 2010; Hug et al., 2011; Turpin
et al., 2011; Hagio et al., 2015). Therefore, it was consid-
ered of interest to check if a difference could emerge
between the LE and TF muscle synergies among these
two conditions of walking. The results did not present
any effect of walking speed in the similarity between LE
and TF muscle synergies. According to the recurring
argument that consistency in muscle groupings across
mechanical constraints supports the hypothesis of neural
drive that selects and activates muscle synergies
(Neptune et al., 2009; Monaco et al., 2010; Torres-
Oviedo and Ting, 2010; Hug et al., 2011; Turpin et al.,
2011; Oliveira et al., 2013; Hagio et al., 2015; Martino
et al., 2015), the current data provided further evidences
of the decrease in dimensionality in the neural control of
locomotor behaviors through a low number of hard-
wired motor modules into the motoneuronal network.
Fig. 5. Mean (±SD) LE muscle synergy vectors (black bars) and those obtained from each wavelet band separately (dark gray to white bars wt#1
to wt#9, respectively) when five muscle synergies (W#1 to W#5) were extracted during the comfort (A) and fast (B) walking condition. Mean (±SD)
values of similarity between the LE muscle synergy vectors and those extracted from each wavelet band (dark gray to white bars wt#1 to wt#9,
respectively) during the comfort (C) and fast (D) walking condition. (E) Total variance accounted for (VAF, in%) according to the wavelet band for
muscle synergy extraction and walking condition. *Significant difference with p< 0.05.
30 J. Frère / Neuroscience 355 (2017) 22–35
Indeed, in addition to be activated in synchrony, the pro-
posed method herein emphasized that all the muscles
of each muscle synergy possessed substantial common-
alities within the time–frequency features of their respec-
tive EMG signals. This was true whatever the
normalization process of the time–frequency maps
(Fig. 4) and also when extracting muscle synergies from
each wavelet band separately (Fig. 5). In agreement with
previous results on EMG-EMG coherence (Danna-Dos-
Santos et al., 2014, 2015; De Marchis et al., 2015), these
results suggested that the CNS might select and activate
each muscle synergy through a common neural drive to
synchronize multiple muscles. However, the coherence
analyses of frequency bands shared between muscles
to infer the neural pathways (subcortical or cortical) gov-
erning these locomotor muscle synergies were not inves-
tigated in this study. The first reason was that this study
dealt with the comparison of muscle groupings between
two methods of muscle synergy extraction. Therefore,
the frequency bands, in which correlated muscle activities
belong, were not further analyzed. The second reason
was that low frequency band [wavelet #1 in the original
wavelet bank of von Tscharner (2000)] was removed to
prevent from movement artifacts due to muscle vibra-
Fig. 6. Mean (±SD) TF30p muscle synergy vectors (black bars) and those obtained from the three other TF datasets (dark gray to white bars for
TF4p, TF30w and TF4w, respectively) when six muscle synergies (W#1 to W#6) were extracted during the comfort (A) and fast (B) walking condition.
Mean (±SD) values of similarity between the TF30p muscle synergy vectors and those extracted from each of the three other TF datasets (dark gray
to white bars for TF4p, TF30w and TF4w, respectively) during the comfort (C) and fast (D) walking condition. (E) Total variance accounted for (VAF, in
%) according to the method of muscle synergy extraction and walking condition. *Significant difference with p< 0.001.
J. Frère / Neuroscience 355 (2017) 22–35 31
tions. As this low-frequency band in EMG-EMG coher-
ence analysis might infer to subcortical pathways
(Grosse and Brown,2003; Boonstra et al., 2009), data
of interest were not available in the present study. Conse-
quently, further studies should be conducted with appro-
priate signal processing based on wavelet coherence
(Charissou et al., 2016) or on rhythmicity within the
time–frequency pattern (von Tscharner et al., 2011;
Maurer et al., 2013) in addition with the proposed method
herein. Such complementary analyses would finally pro-
vide further insights into the hierarchy of the neural net-
works (at the spinal, supraspinal or cortical levels)
underlying the selection, activation and combination of
muscle synergies.
Fig. 7. Mean (±SD) TF30p muscle synergy vectors (black bars) and those obtained from each wavelet band separately (dark gray to white bars
wt#1 to wt#9, respectively) when six muscle synergies (W#1 to W#6) were extracted during the comfort (A) and fast (B) walking condition. Mean
( ± SD) values of similarity between the TF30p muscle synergy vectors and those extracted from each wavelet band (dark gray to white bars wt#1 to
wt#9, respectively) during the comfort (C) and fast (D) walking condition. (E) Total variance accounted for (VAF, in %) according to the wavelet band
for muscle synergy extraction and walking condition. *Significant difference with p< 0.05.
32 J. Frère / Neuroscience 355 (2017) 22–35
Prescriptive vs. descriptive muscle synergies
Whatever the walking condition, the composition of the
muscle synergies was consistent between both methods
of extraction when a number of five muscle synergies
was used. However, a higher number of cases of
dissimilar muscle synergy vectors (q< 0.80) between
both methods was found for the muscle synergy #1 and
#5 in comparison to the three other muscle synergies.
When defining the number of muscle synergies based
on the VAF curve, it appeared that six muscle synergies
accounted for the time–frequency dataset. Taken
together, these results suggested a likely non-
physiological source of muscle grouping for the muscle
synergy #1 and #5 when extracting the muscle
synergies solely from the global patterns of muscle
activities (i.e., linear envelope). More specifically, the
J. Frère / Neuroscience 355 (2017) 22–35 33
muscle synergy #1 was mainly constituted by the knee
extensor muscles (VL, RF, and VM) but solely both
vastii of the quadriceps muscle remained in the time–
frequency muscle synergy #1, while the RF muscle
constituted alone a new one (time–frequency muscle
synergy #5). The RF muscle also dissociated with the
TFL muscle (muscle synergy #5) that finally constituted
alone the time–frequency muscle synergy #6. With
regards to the Valero-Cuevas’s (2016) query, these differ-
ences between both methods of extraction (LE vs. TF
muscle synergies) might reveal how the proposed method
herein could differentiate prescriptive muscle synergies to
descriptive ones. Indeed, during walking, the RF muscle
had one main peak of activity relative to the control of
the knee angle during the weight acceptance in the early
stance and a second (lower) one relative to the hip flexion
at the initiation of the swing phase. According to previous
findings (Neptune et al., 2009; Clark et al., 2010; Ting
et al., 2015), this unique pattern of RF muscle activity dur-
ing walking expressed its contribution to two muscle syn-
ergies (LE muscle synergy #1 and #5, herein) and
suggested that the CNS effectively reduced the muscu-
loskeletal redundancy in the construction of motor behav-
iors (d’Avella and Bizzi, 2005). A similar phenomenon of
RF muscle activity expressing two muscle synergies (with
VM and VL during the knee extension and with TFL during
the hip flexion) has been found during pedaling (Hug
et al., 2011), thus reinforcing the hypothesis that across
motor tasks, the CNS used and flexibly combined a low
number of muscle synergies. But none of these previous
studies took into account spectral features of the EMG
signals. Therefore, there was no possibility to determine
if the RF pattern originated from a structural factor [e.g.,
change in muscle–tendon unit length; Kutch and Valero-
Cuevas (2012)] or actually expressed underlying muscle
synergies of neural origin. In that way, the time–frequency
muscle synergies differed from the classical muscle syn-
ergies and might underline a neural specificity of this bi-
articular muscle in comparison with the mono-articular
muscles for knee extension (VL and VM muscles) and
hip flexion (TFL muscle). Such assumption remains con-
ceivable in light of previous reports relative to the neural
control of the different heads of the quadriceps muscle.
For instance, an alternate muscle activity between RF
and either VL or VM muscles has been found to attenuate
fatigue (Kouzaki et al., 2002; Kouzaki and Shinohara,
2006). Also, Place et al. (2006) found no difference in
the EMG activity between RF, VL and VM muscles before
a fatiguing task, but at task failure, RF muscle activity was
systematically below those of the vastii muscles. This
suggested that the CNS was not able to maintain the orig-
inal descending drive to face the higher fatigability of the
RF muscle. Other evidences of a dissociation of neural
drive between RF and vastii muscles were provided dur-
ing pain experiments (Hug et al., 2014). Finally, it has
been shown that both vastiimuscles were highly synergist
by sharing most of their synaptic input (Mellor and
Hodges, 2005; Laine et al., 2015). Taking all into account,
one might assume that the changes in muscle synergies
found in the current study would rely on distinct neural
drives between these muscles and that using time–fre-
quency features of multiple muscles allowed to discrimi-
nate prescriptive muscle synergies to descriptive ones.
Methodological considerations
One main issue of this study might come from the lower
values of VAF provided by the time–frequency muscle
synergies in comparison to values provided by classical
ones. This finding was true whatever the dimensionality
and normalization of the time–frequency dataset
(Figs. 4E-7E). As found by Oliveira et al. (2014), decreas-
ing the dimensionality of the initial dataset increase the
reconstruction quality, but the results herein never
reached the 90% of VAF threshold. One might thus con-
sider these low values as an inability of time–frequency
method to sufficiently reconstruct the initial dataset. But,
the high level of similarity of the muscle synergies what-
ever the dataset (Figs. 4–7) provided strong evidence that
muscles composing one motor module shared spectral
communalities in addition to regularities in the time
domain. Nevertheless, one has to acknowledge that the
NMF algorithm only captured motor modules from the
intensity pattern of several wavelet bands simultaneously
(Figs. 4 and 6) or separately (Figs. 5 and 7). In that way,
the magnitude of the real coupling between two frequency
bands was not taken into account as it could be obtained
from coherence analyses. Overall, these methodological
considerations highlighted the pressing need to further
investigate the potentiality of the proposed method to dis-
tinguish between prescriptive and descriptive muscle
synergies.
CONCLUSIONS
Overall, extracting time–frequency muscle synergies
could be viewed as a great opportunity to discriminate
prescriptive muscle synergies to descriptive ones and
supports the hypothesis of the neural origin of the
muscle synergies. This method offers further
possibilities to investigate the underlying mechanisms
for the production of motor behaviors. Additional studies
should however be conducted using other cyclic tasks
with a wider variety of mechanical constraints to confirm
the potential use of time–frequency muscle synergies to
define prescriptive motor modules.
Acknowledgments—I thank Nicolas TURPIN for writing the first
version of the script relative to the Cheung’s method to determine
the number of muscle synergies to extract. My warmestgratitude
to Brian EASTON for his skillful editing of the English manuscript.
No conflicts of interest, financial or otherwise, are declared by the
author.
REFERENCES
Bizzi E, Cheung VCK (2013) The neural origin of muscle synergies.
Front Comput Neurosci 7:51.
Boonstra TW, Daffertshofer A, Roerdink M, Flipse I, Groenewoud K,
Beek PJ (2009) Bilateral motor unit synchronization of leg
muscles during a simple dynamic balance task. Eur J Neurosci
29:613–622.
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0005
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0005
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0010
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0010
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0010
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0010
34 J. Frère / Neuroscience 355 (2017) 22–35
Cappellini G, Ivanenko YP, Poppele RE, Lacquaniti F (2006) Motor
patterns in human walking and running. J Neurophysiol
95:3426–3437.
Charissou C, Vigouroux L, Berton E, Amarantini D (2016) Fatigue-
and training-related changes in ‘‘beta” intermuscular interactions
between agonist muscles. J Electromyogr Kinesiol 27:52–59.
Cheung VCK, d’Avella A, Tresch MC, Bizzi E (2005) Central and
sensory contributions to the activation and organization of muscle
synergies during natural motor behaviors. J Neurosci
25:6419–6434.
Cheung VCK, Piron L, Agostini M, Silvoni S, Turolla A, Bizzi E
(2009) Stability of muscle synergies for voluntary actions after
cortical stroke in humans. Proc Natl Acad Sci U S A
106:19563–19568.
Chvatal SA, Ting LH (2012) Voluntary and reactive recruitment of
locomotor muscle synergies during perturbed walking. J Neurosci
32:12237–12250.
Clark DJ, Ting LH, Zajac FE, Neptune RR, Kautz SA (2010) Merging
of healthy motor modules predicts reduced locomotor
performance and muscle coordination complexity post-stroke. J
Neurophysiol 103:844–857.
d’Avella A, Bizzi E (2005) Shared and specific muscle synergies in
natural motor behaviors. Proc Natl Acad Sci U S A
102:3076–3081.
Danna-Dos-Santos A, Boonstra TW, Degani AM, Cardoso VS,
Magalhaes AT, Mochizuki L, Leonard CT (2014) Multi-muscle
control during bipedal stance: an EMG-EMG analysis approach.
Exp Brain Res 232:75–87.
Danna-Dos-Santos A, Degani AM, Boonstra TW, Mochizuki L,
Harney AM, Schmeckpeper MM, Tabor LC, Leonard CT (2015)
The influence of visual information on multi-muscle control during
quiet stance: a spectral analysis approach. Exp Brain Res
233:657–669.
de Luca CJ (1997) The use of surface electromyography in
biomechanics. J Appl Biomech 13:135–163.
De Marchis C, Severini G, Castronovo AM, Schmid M, Conforto S
(2015) Intermuscular coherence contributions in synergistic
muscles during pedaling. Exp Brain Res 233:1907–1919.
de Rugy A, Loeb GE, Carroll TJ (2013) Are muscle synergies useful
for neural control? Front Comput Neurosci 7:19.
Frère J, Göpfert B, Slawinski J, Tourny-Chollet C (2012a) Shoulder
muscles recruitment during a power backward giant swing on high
bar: a wavelet-EMG-analysis. Hum Mov Sci 31:472–485.
Frère J, Göpfert B, Slawinski J, Tourny-Chollet C (2012b) Influence of
the scale function on wavelet transformation of the surface
electromyographic signal. Comput Methods Biomech Biomed
Eng 15:111–120.
Frère J, Hug F (2012) Between-subject variability of muscle synergies
during a complex motor skill. Front Comput Neurosci 6:99.
Grosse P, Brown P (2003) Acoustic startle evokes bilaterally
synchronous oscillatory EMG activity in the healthy human. J
Neurophysiol 90:1654–1661.
Grosse P, Cassidy MJ, Brown P (2002) EEG-EMG, MEG-EMG and
EMG-EMG frequency analysis: physiological principles and
clinical applications. Clin Neurophysiol 113:1523–1531.
Hagio S, Fukuda M, Kouzaki M (2015) Identification of muscle
synergies associated with gait transition in humans. Front Hum
Neurosci 9:48.
Hermens HJ, Freriks B, Disselhorst-Klug C, Rau G (2000)
Development of recommendations for SEMG sensors and
sensor placement procedures. J Electromyogr Kinesiol
10:361–374.
Hodson-Tole EF, Wakeling JM (2007) Variations in motor unit
recruitment patterns occur within and between muscles in the
running rat (Rattus norvegicus). J Exp Biol 210:2333–2345.
Hodson-Tole EF, Wakeling JM (2009) Motor unit recruitment for
dynamic tasks: current understanding and future directions. J
Comp Physiol 179:57–66.
Hug F, Hodges PW, van den Hoorn W, Tucker K (2014) Between-
muscle differences in the adaptation to experimental pain. J Appl
Physiol 117:1132–1140.
Hug F, Turpin NA, Couturier A, Dorel S (2011) Consistency of muscle
synergies during pedaling across different mechanical
constraints. J Neurophysiol 106:91–103.
Ivanenko YP, Grasso R, Zago M, Molinari M, Scivoletto G, Castellano
V, Macellari V, Lacquaniti F (2003) Temporal components of the
motor patterns expressed by the human spinal cord reflect foot
kinematics. J Neurophysiol 90:3555–3565.
Ivanenko YP, Poppele RE, Lacquaniti F (2004) Five basic muscle
activation patterns account for muscle activity during human
locomotion. J Physiol 556:267–282.
Kouzaki M, Shinohara M (2006) The frequency of alternate muscle
activity is associated with the attenuation in muscle fatigue. J Appl
Physiol 101:715–720.
Kouzaki M, Shinohara M, Masani K, Kanehisa H, Fukunaga T (2002)
Alternate muscle activity observed between knee extensor
synergists during low-level sustained contractions. J Appl
Physiol 93:675–684.
Kutch JJ, Valero-Cuevas FJ (2012) Challenges and new approaches
to proving the existence of muscle synergies of neural origin.
PLoS Comput Biol 8:e1002434.
Laine CM, Martinez-Valdes E, Falla D, Mayer F, Farina D (2015)
Motor Neuron Pools of Synergistic Thigh Muscles Share Most of
Their Synaptic Input. J Neurosci 35:12207–12216.
Lee D, Seung H (2001) Algorithms for nonnegative matrix
factorization. Adv Neural Inf Process Syst 13:556–562.
Martino G, Ivanenko YP, d’Avella A, Serrao M, Ranavolo A, Draicchio
F, Cappellini G, Casali C, Lacquaniti F (2015) Neuromuscular
adjustments of gait associated with unstable conditions. J
Neurophysiol 114:2867–2882.
Maurer C, von Tscharner V, Nigg BM (2013) Speed-dependent
variation in the Piper rhythm. J Electromyogr Kinesiol
23:673–678.
Mellor R, Hodges P (2005) Motor unit synchronization between
medial and lateral vasti muscles. Clin Neurophysiol
116:1585–1595.
Monaco V, Ghionzoli A, Micera S (2010) Age-related modifications of
muscle synergies and spinal cord activity during locomotion. J
Neurophysiol 104:2092–2102.
Neptune RR, Clark DJ, Kautz SA (2009) Modular control of human
walking: a simulation study. J Biomech 42:1282–1287.
Oliveira AS, Gizzi L, Farina D, Kersting UG (2014) Motor modules of
human locomotion: influence of EMG averaging, concatenation,
and number of step cycles. Front Hum Neurosci 8:335.
Oliveira AS, Silva PB, Lund ME, Kersting UG, Farina D (2013) Fast
changes in direction during human locomotion are executed by
impulsive activation of motor modules. Neuroscience
228:283–293.
Place N, Matkowski B, Martin A, Lepers R (2006) Synergists
activation pattern of the quadriceps muscle differs when
performing sustained isometric contractions with different EMG
biofeedback. Exp Brain Res 174:595–603.
Rana M, Yani MS, Asavasopon S, Fisher BE, Kutch JJ (2015) Brain
Connectivity Associated with Muscle Synergies in Humans. J
Neurosci 35:14708–14716.
Roemmich RT, Fregly BJ, Hass CJ (2014) Neuromuscular complexity
during gait is not responsive to medication in persons with
Parkinson’s disease. Ann Biomed Eng 42:1901–1912.
Routson RL, Clark DJ, Bowden MG, Kautz SA, Neptune RR (2013)
The influence of locomotor rehabilitation on module quality and
post-stroke hemiparetic walking performance. Gait Posture
38:511–517.
Safavynia SA, Ting LH (2012) Task-level feedback can explain
temporal recruitment of spatially fixed muscle synergies
throughout postural perturbations. J Neurophysiol 107:159–177.
Safavynia SA,Ting LH (2013) Sensorimotor feedback based on task-
relevant error robustly predicts temporal recruitment and
multidirectional tuning of muscle synergies. J Neurophysiol
109:31–45.
Safavynia SA, Torres-Oviedo G, Ting LH (2011) Muscle Synergies:
Implications for Clinical Evaluation and Rehabilitation of
Movement. Top Spinal Cord Inj Rehabil 17:16–24.
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0015
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0015
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0015
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0020
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0020
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0020
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0020
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0025
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0025
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0025
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0025
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0030
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0030
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0030
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0030
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0035
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0035
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0035
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0040
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0040
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0040
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0040
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0045
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0045
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0045
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0050
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0050
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0050
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0050
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0055
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0055
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0055
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0055
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0055
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0060
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0060
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0065
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0065
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0065
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0070
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0070
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0075
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0075
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0075
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0080
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0080
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0080
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0080
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0085
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0085
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0090
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0090
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0090
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0095
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0095
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0095
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0100
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0100
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0100
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0105
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0105
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0105
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0105
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0110
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0110
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0110
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0115
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0115
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0115
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0120
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0120
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0120
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0125
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0125
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0125
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0130
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0130
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0130
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0130
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0135
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0135
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0135
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0140
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0140
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0140
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0145
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0145
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0145
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0145
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0150
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0150
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0150
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0155
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0155
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0155
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0160
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0160
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0165
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0165
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0165
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0165
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0170
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0170
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0170
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0175
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0175
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0175
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0180
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0180
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0180
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0185
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0185
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0190
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0190
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0190
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0195
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0195
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0195
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0195
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0200
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0200
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0200
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0200
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0205
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0205
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0205
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0210
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0210
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0210
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0215
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0215
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0215
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0215
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0220
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0220
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0220
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0225
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0225
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0225
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0225http://refhub.elsevier.com/S0306-4522(17)30297-X/h0230
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0230
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0230
J. Frère / Neuroscience 355 (2017) 22–35 35
Ting LH, Chiel HJ, Trumbower RD, Allen JL, McKay JL, Hackney ME,
Kesar TM (2015) Neuromechanical principles underlying
movement modularity and their implications for rehabilitation.
Neuron 86:38–54.
Ting LH, Chvatal SA (2010) Decomposing muscle activity in motor
tasks: methods and interpretation. In: Danion F, Latash ML,
editors. Motor control: theories, experiments, and
applications. Oxford: Oxford University Press. p. 102–138.
Ting LH, Chvatal SA, Safavynia SA, McKay JL (2012) Review and
perspective: neuromechanical considerations for predicting
muscle activation patterns for movement. Int J Numer Methods
Biomed Eng 28:1003–1014.
Ting LH, McKay JL (2007) Neuromechanics of muscle synergies for
posture and movement. Curr Opin Neurobiol 17:622–628.
Torres-Oviedo G, Macpherson JM, Ting LH (2006) Muscle synergy
organization is robust across a variety of postural perturbations. J
Neurophysiol 96:1530–1546.
Torres-Oviedo G, Ting LH (2007) Muscle synergies characterizing
human postural responses. J Neurophysiol 98:2144–2156.
Torres-Oviedo G, Ting LH (2010) Subject-specific muscle synergies
in human balance control are consistent across different
biomechanical contexts. J Neurophysiol 103:3084–3098.
Turpin NA, Guével A, Durand S, Hug F (2011) Effect of power output
on muscle coordination during rowing. Eur J Appl Physiol
111:3017–3029.
Valero-Cuevas FJ (2016) The Nature and Structure of Feasible Sets.
In: Fundamentals of neuromechanics, biosystems &
biorobotics. London: Springer. p. 135–157.
van den Hoorn W, Hodges PW, van Dieën JH, Hug F (2015) Effect of
acute noxious stimulation to the leg or back on muscle synergies
during walking. J Neurophysiol 113:244–254.
von Tscharner V (2000) Intensity analysis in time-frequency space of
surface myoelectric signals by wavelets of specified resolution. J
Electromyogr Kinesiol 10:433–445.
von Tscharner V, Barandun M, Stirling LM (2011) Piper rhythm of the
electromyograms of the abductor pollicis brevis muscle during
isometric contractions. J Electromyogr Kinesiol 21:184–189.
Wakeling JM (2004) Motor units are recruited in a task-dependent
fashion during locomotion. J Exp Biol 207:3883–3890.
Wenger N et al (2016) Spatiotemporal neuromodulation therapies
engaging muscle synergies improve motor control after spinal
cord injury. Nat Med 22:138–145.
(Received 6 December 2016, Accepted 26 April 2017)
(Available online 05 May 2017)
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0235
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0235
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0235
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0235
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0240
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0240
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0240
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0240
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0245
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0245
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0245
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0245
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0250
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0250
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0255
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0255
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0255
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0260
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0260
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0265
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0265
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0265
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0270
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0270
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0270
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0275
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0275
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0275
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0280
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0280
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0280
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0285
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0285
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0285
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0290
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0290
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0290
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0295
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0295
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0300
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0300
http://refhub.elsevier.com/S0306-4522(17)30297-X/h0300
	Spectral properties of multiple myoelectric signals: �New insights into the neural origin of muscle synergies
	Introduction
	Experimental procedures
	Participants
	Protocol
	Materials and data collection
	Data analysis
	Statistical analysis
	Results
	Spatiotemporal walking data
	Number of muscle synergies
	LE vs. TF muscles synergies
	Effect of the characteristics of the TF initial data matrix
	Discussion
	Neurophysiological interpretations
	Prescriptive vs descriptive muscle synergies
	Methodological considerations
	Conclusions
	Acknowledgments
	References