Version 5

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING 1
Electromyographic Signal ltering based on

Independent Component Analysis and Empirical
Mode Decomposition for Estimating the Motor
Activation Pattern in Human Walking
Claudio Tapia, Omar Daud, Member, IEEE, Mauricio Delgado, Leonidas Arias, and Javier Ruiz-del-Solar, Senior
Member, IEEE
AbstractA new method for determining the pattern of motor
activation during walking in healthy human subjects, based
on the use of Independent Component Analysis (ICA) and
Empirical Mode Decomposition (EMD) over electromyographic
(EMG) signals, is proposed. The acquisition of the EMG signals
corresponds to 16 muscles from both lower limbs during the
walking of 11 healthy women. The signals are rst pre-processed
by Principal Component Analysis (PCA), then analyzed by
ICA and nally ltered by EMD. This permits reconstructing
the original source signals, thus allowing the ltering of the
components that do not have a muscular origin. The data is then
segmented and further processed using the Hilbert Transform
in order to obtain the representation of a gait cycle from all
records. The results demonstrate that the proposed approach
allows obtaining the sequence of motor activation during human
walking. These results were also compared to the ones obtained
by using two other methods, namely a conventional method and
another based only in EMD ltering. The proposed method was
validated not only by the analysis of the EMG signals, but also by
comparing it to the corresponding biomechanical behavior, that
in turn, was reconstructed by using the kinematic data of the
hip, knee and ankle of every participant. Such activity is closely
related to the biomechanical behavior during the movement of
the lower extremities.
Index TermsGait, Principal Component Analysis, Indepen-
dent Component Analysis, Empirical Mode Decomposition, Sur-
face Electromyography, Pattern of Motor Activation.
I. INTRODUCTION
T
HE most widely technique used for acquiring and eval-
uating the electrical activity produced by the muscles
is electromyography (EMG). An electromyographic signal
detects the electrical potential generated by muscle cells
C. Tapia, is with Facultad de Ciencias de la Rehabilitaci on, Universidad
Nacional Andr es Bello, Fern andez Concha 700, Las Condes, Santiago, Chile
& with the Department of Electrical Engineering Universidad de Chile, Av.
Tupper 2007, 837-0451 Santiago, Chile. E-mail: ctmalebran@gmail.com
O. Daud and J. Ruiz-del-Solar are with the Department of Electrical En-
gineering & the Advanced Mining Technology Center, Universidad de Chile,
Av. Tupper 2007, 837-0451 Santiago, Chile. E-mail:omar.daud@amtc.cl;
jruizdelsolar@amtc.cl
Mauricio Delgado is with Carrera de Kinesiologa, UDA Cs. Salud, Facultad
Medicina, Ponticia Universidad Cat olica de Chile, Vicu na Mackenna 4860,
Santiago, Chile. E-mail: m.delgadobravo@gmail.com
Leonidas Arias is with Facultad de Ciencias de la Rehabilitaci on, Univer-
sidad Nacional Andr es Bello, Fern andez Concha 700, Las Condes, Santiago,
Chile & with the Magster de Kinesiologa y Biomec anica de la Universidad
Metropolitana de Ciencias de la Educaci on, Luis Bisquert 2765, Santiago,
Chile. E-mail: arias07eduardo@gmail.com
when these cells are electrically or neurologically activated.
The signals can be analyzed to detect medical abnormalities,
activation level, recruitment order or to analyze the biome-
chanics of human movement. In this sense, the sequence
of muscle activation is important as it has different clinical
applications. The evaluation of the muscle activation has been
demonstrated to be very useful in orthopedics [1], for the
treatment of the cerebral palsy [2], the assessment of spastic
paresis [3], in the rehabilitation of post-stroke patients [4],
and in sports medicine [5]. Nevertheless, not only for people
with musculoskeletal disorders but also for healthy people, gait
patterns have also been researched and several studies have
been published [6], [7], [8], [9], [10]. However, in all these
publications there is an issue that has still not been achieved,
and deals with the (most appropriate) method for processing
and analyzing EMG signals. As a matter of fact, in clinical
practice, the interpretation of motor activation is determined
by the analysis made by the clinician, resulting in an apparent
dependence on the evaluator.
When a motor unit potential res, it is usually accompa-
nied by an instantaneous increase in signal amplitude and
frequency. Traditional methods for processing raw electromyo-
graphic signals involve rectication and subsequent low-pass
ltering [9], for obtaining the envelope of the corresponding
electromyographic signal. Such transformation produces a loss
of sharpness in the processed signals [8], [11], [12], [13].
An important issue dealing with ltering, is the fact that the
phases of muscle activity become smoother in amplitude and
longer in duration as the cut-off frequency of the low pass
lter is reduced. Therefore, ltering affects the recognition
and estimation of the muscle activity through EMG, basically
the onset time detection. In order to overcome the problem of
time activation, several algorithms have been proposed [14],
[15], [16]. Besides, for avoiding the overestimation of the
duration of the activity phase, in [17] is proposed a cut-
off frequency around 9 Hz in natural walking by healthy
subjects. However, this frequency must be raised up if there
are pathological synchronization in patients with tremor or
clonus. Anyhow, choosing the suitable parameters for smooth-
ing ltering is still very challenging. According to [18], the
proposed cut-off frequency for ltering the signal was found
to be around 3 Hz. In [19], the power spectrum of the rectied
electromyographic signal was investigated during gait activity.
Such study was conducted in order to achieve an optimal cut-
off frequency before calculating the electromyographic signal
envelope. According to this study, the frequency content of the
rectied electromyographic signals was found to be different
for healthy subjects compared to patients. Moreover, con-
sidering the gastrocnemius lateralis muscle in young healthy
subjects, 95% of the power was found to be below 3.8 Hz,
whereas the 95% limit was 5.6 Hz for elderly subjects, 13.1
Hz for parkinsonians, and 11.5 Hz for patients with cerebellar
ataxia. For the hamstring lateralis muscle, the 95% limit was
4.5 Hz for the young healthy subjects, 12.8 Hz for the elderly
subjects, 33.8 Hz for the parkinsonians, and 26.0 Hz for
the cerebellar ataxia patients. These results indicate that an
incorrect election of the low-pass cut-off frequencies generally
affect (negatively) the analysis of normal and pathological gait,
as useful information could be lost.
This study proposes a new method for determining the
sequence of motor activation of the lower extremities by
using signal processing over EMG signals, acquired during
walking in healthy individuals. The method is based on the
assumption that the available records may contain information
from undesirable sources; therefore, the acquired signals are
rst preprocessed by PCA, analyzed through ICA, and nally
ltered by EMD. The main contribution of the this manuscript
is based on a new approach for processing of the EMG
signals as a blind separation problem, which avoids the loss
of important signal components due to the low-pass ltering
of the signals used in traditional approaches. In order to
demonstrate such difference, the proposed method is compared
to other two methods: one is based on a traditional method,
and so low pass ltering, while the second is merely based on
EMD ltering. Besides, the proposed method is proven by the
biomechanical behavior of the corresponding motor activation
patterns.
This manuscript is organized as follows. In Section II,
methods are introduced with a description of the participants
and the procedures. Signal processing of the raw data along
with the different methods that were used in this manuscript
are explained in Section III. The results are presented in
Section IV, while the discussion of the results in Section V.
Finally, conclusions are drawn in Section VI.
II. METHODS
A. Participants
Our subjects were 11 women aged between 21 and 27
(average age of 24.2 with a standard deviation of 1.75), with-
out musculoskeletal pathology. Participants were contacted
directly. The experiments were taken at the Biomechanical
Laboratory of the Ponticia Universidad Cat olica de Chile,
Santiago, Chile. The protocol was approved by the ethics
committee of the Ponticia Universidad Cat olica de Chile,
while each participant gave their informed consent before
running the experiments.
B. Procedures
In order to capture the 3D movement, an optic system
(Vicon Motion System, Oxford, UK), consisting in 8 infrared
cameras (Bonita model), with 100 Hz of sampling rate and a
spatial resolution of 1 mm, where connected via MX Giganet
to a PC (Dell, T1650 model, Intel Xeon CPU E3-1290 V2,
3.70 Hz). Thirty nine reector markers were used, each one of
14 mm diameter. The marking protocol that was used belongs
to the Plug In Gait (Vicon Nexus, Oxford, UK) system for the
complete body.
For surface electromyographic records, a 16 channel elec-
tromyograph, Delsys R product, and a Trigno
TM
Wireless Sys-
tem, were used for the acquisition. The sampling frequency
used for the experiments was 1 KHz.
The synchronization signals were obtained by two
FlexiForce R pressure sensors, which allowed determining the
start and ending of each gait cycle by means of impulse
functions generated at the time of heel contact during the gait.
The muscular signals were obtained by recording differential
pre-amplied electrodes. The SENIAM protocol was used for
localization of the electrodes, selecting the back of the hand
as the reference signal, as is shown in Figure 1.
Sixteen EMG electrodes were positioned on the surface over
the following muscles: gluteus maximus (GM), gluteus medius
(Gmed), rectus femoris (RF), vastus lateralis (Vlat), biceps
femoris (BF), tibialis anterior (TA), lateral gastrocnemius
(LG), soleus (SOL). Each muscle was recorded in both lower
extremities. There were four records for each participant; the
rst recording was considered to be a test and its purpose was
to allow habituation to the conditions under which the actual
tests would be conducted. The remaining three records were
considered valid and subsequently analyzed.
The Vicon Nexus 1.8.5 software, under the Windows 7 oper-
ating system, was used for synchronizing the signal acquisition
from all the devices. This software used also a 4
th
order
Butterworth low pass lter, with a 6 Hz cut-off frequency,
to ltered out the trajectories traced by the markers.
III. SIGNAL PROCESSING
In clinical gait analysis, determining in a correct way the
timing of muscle activation (on-off) is a key issue [20].
The evaluation of this on-off pattern, particularly when it
comes along with kinematics, and eventually kinetics (not
reported in this manuscript) studies, gives a better understand-
ing on how the performance of the muscles and their role
are accomplished by a motor task [21]. The importance of
determining accurately the muscle activation pattern is based
on its usefulness in the treatment of cerebral palsy, sports
medicine and orthopedics. Moreover, this on-off pattern
constitutes the only reliable information that can come from
the raw EMG signals acquired during dynamic contractions.
As the amplitude of the EMG signal is not only inuenced
by the actual electrical activation level of the muscle, but
also by the electrical characteristics of the tissue between the
active muscle bers and the surface electrodes, the relationship
between electrical activity level and muscle force is non-linear.
Therefore, a suitable signal processing of the EMG signals,
particularly those acquired in dynamic conditions, allows a
correct interpretation. The main components for processing
the raw data are shown in the block diagram of Fig. 2.
(a) Frontal view. (b) Side view.
(c) General view of the setup.
Fig. 1. SENIAM protocol used for localization of the electrodes along with
the distribution of the 39 reector markers.
A. Principal Component Analysis
For rst, Principal Component Analysis (PCA) was applied.
Note that the PCA algorithm was not used for dimensionality
reduction, but for ensuring convergence in the blind source
separation process. Therefore, the projection of the signals was
made considering all the components. PCA is based on the
generation of a new set of uncorrelated variables. The new
variables are linear combinations of the above, and they are
built according to the order of importance in terms of the total
variability of collected samples.
In a nutshell, PCA can be described as follows. Let us
consider a nonlinear system modeled by:
x = f( x, x), x(t) =
_
x
1
(t) x
2
(t) . . . x
m
(t)
T
The response matrix of such a system is given by:
X =
_
x(t
1
) x(t
2
) . . . x(t
n
)
=
_
_
x
1
(t
1
) x
1
(t
2
) . . . x
1
(t
n
)
x
2
(t
1
) x
2
(t
2
) . . . x
2
(t
n
)
.
.
.
.
.
.
.
.
.
.
.
.
x
m
(t
1
) x
m
(t
2
) . . . x
m
(t
n
)
_
_
(1)
If the mean values of each of the row vector are not zero,
then the Eq. (1) is modied and becomes:
X =
_
x(t
1
) x x(t
2
) x . . . x(t
n
) x
=
_
_
x
1
(t
1
) x x
1
(t
2
) x . . . x
1
(t
n
) x
x
2
(t
1
) x x
2
(t
2
) x . . . x
2
(t
n
) x
.
.
.
.
.
.
.
.
.
.
.
.
x
m
(t
1
) x x
m
(t
2
) x . . . x
m
(t
n
) x
_
_
(2)
where x is the mean value of the data set.
Hence, the covariance matrix of the matrix X in Eq. (2) is
given by,
C
X
=
1
n
XX
T
The C
X
matrix (m x m) represents the correlations be-
tween all possible pairs of measurements. The diagonal and
non-diagonal terms are called the variance and covariance,
respectively. Therefore, the main scope here is to maximize
the variance and minimize the covariance. Such operation can
be achieved by performing a linear transformation in order to
diagonalize C
X
, and nd the eigenvalues and eigenvectors of
the matrix C
X
.
C
X
U = U
Thus, seeking the values in the following form:
C
X
= UU
T
in which,
U =
_
U
1
U
2
. . . U
m
=
_
1
0 . . . 0
0
2
. . . 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 . . .
m
_
_
Once eigenvectors and eigenvalues are found from the
covariance matrix, the next step is to order the eigenvectors by
eigenvalues, from the highest to lowest ones. This operation
classies the components in order of signicance. Then it
turns out that the eigenvector with the highest eigenvalue is
the principle component of the data set. The next step is to
construct a feature vector by choosing the eigenvectors from
the list of the eigenvectors (PC
x(t)
).
In this particular case, all the principal components were
chosen, since PCA (as mentioned above) was not used for
dimensionality reduction, but for ensuring convergence in the
blind source separation process during the analysis of the
independent components.
B. Independent Component Analysis
Independent Component Analysis (ICA) is an iterative
technique that estimates the statistical independence of signals
from a given set of its linear combinations [22]. ICA seeks to
nd a linear separation of multiple sources without having
Motor Activation
Pattern
Gait Kinematic
Analysis
Hilbert Transform
& Envelope
Signal
Reconstruction
x
r
(t) = As s = Wx(t)
ICA
PC
x(t)
PCA
x(t)
Raw EMG
Signals
EMD
with
Soft-thresholding
Fig. 2. Block Diagram for Signal Processing.
information from the original signals or the weights of the
mixture [23], [24]. More formally, ICA tackles the problem
of blind separation of sources, in which it is considered that
the signals received by the sensors are mixtures derived from
various independent sources. The goal of the method is to
analyze these mixtures and to obtain the original signals from
them [25]. In a nutshell, ICA can be computed as follows:
Given measurement, x(t), which in this particular case
is PC
x(t)
, nd the independent components, s, and the
associated mixing matrix, A, such that:
x(t) = As
Find w
j
that maximizes non-Gaussianity of w
T
j
x(t).
Independent components s is then found from:
s = Wx
where,
A
1
= W =
_
w
1
w
2
. . . w
m
According to the procedure adopted for the data analysis

(see Fig. 2), ICA was applied to the principal components
obtained from PCA. After obtaining the independent compo-
nents, for each gait sequence and for each different subject,
the components with higher variance were set to zero. This
allowed reconstructing the input signal, now without the
unwanted components:
x
r
(t) = As
PCA and ICA computations were done by using the Fas-
tICA toolbox, which is based on the model developed by Aapo
Hyv arinen and Erkki Oja [26].
C. Empirical Mode Decomposition with soft-thresholding
The Empirical Mode Decomposition (EMD) is based on
the Hilbert-Huang transform (HHT) [27]. Basically, the HHT
is carried out in two steps (see Fig. 3). The rst step uses
the EMD algorithm in order to get a nite set of functions.
Herein, the EMD is a method for decomposing a given signal
into a nite and often small number of functions, that are
called Intrinsic Mode Functions (IMFs). At the second step,
the instantaneous frequency spectrum of the given signal
is obtained by applying the Hilbert Transform [28] to the
IMFs. Therefore, applying the HHT allows the computation of
the instantaneous frequency spectrum of non-linear and non-
stationary signals.
Given
Signal
Sifting
Process
Intrinsic Mode
Function
Stopping
Criterion
EMPIRICAL MODE DECOMPOSITION
YES
NO
Marginal
Spectrum
Hilbert
Spectrum
Hilbert
Transform
Time Integral
Fig. 3. Empirical Mode Decomposition and Hilbert-Huang Transform.
1) Frequency Implicit Mode Function: In [27], an oscil-
lating wave is dened as an IMF, if it satises these two
conditions:
1) In the whole data set, the number of extrema and the
number of zero crossings must either equal or differ at
most by one.
2) At any point, the mean value of the envelope dened by
the local maxima and the envelope dened by the local
minima is zero.
2) Sifting Extracting Implicit Mode Function: Given a data
set x(t):
1) Identify all upper extrema of x(t).
2) Interpolate the local maxima to form an upper envelope
u(x).
3) Identify all lower extrema of x(t).
4) Interpolate the local minima to form an lower envelope
l(x).
5) Calculate the mean envelope: m(t) = (u(x) +l(x))/2.
6) Extract the mean from the signal: h(t) = x(t) m(t).
7) Check whether h(t) satises the IMF condition:
YES: h(t) is an IMF, stop sifting.
NO: let x(t) = h(t), keep sifting.
A stopping criterion for the sifting process can be deter-
mined by a bounded value of the standard deviation (SD) from
two consecutive sifting results.
SD =
T
t=0
_
|(h
1(k1)
(t) h
1k
(t))|
2
h
2
1(k1)
(t)
_
h
1k
: residue after the k
t
h iteration of the 1st IMF.
As stated in [27], a typical value for SD can be set between
0.2 and 0.3.
IMF
1
IMF
2
.
.
.
IMF
N
tIMF
1
tIMF
2
.
.
.
tIMF
N
Soft-
thresholding
Summation
Filtered Signal
EMD
Fig. 4. Soft-thresholding process of the IMFs.
3) Soft-thresholding: However, a slight modication over
the EMD method is proposed, as shown in Fig. 4. Hence,
after decomposing the signal into IMFs, they are denoised
by applying a threshold to the estimated IMFs, and nally
the signal is reconstructed by summing all the denoised IMFs.
This modication is a practical procedure due to the empirical
nature of the EMD method, as it doesnt make any assumption
about the input time-series [29].
First, as stated in Fig. 3, the EMD method is used for de-
composing the input signal into IMFs. Then soft-thresholding
is applied to each IMFs, yielding to tIMF
1
, tIMF
2
, . . . , tIMF
N
,
which are the denoised versions of the original components,
as shown in Eq. 3.
tIMF
N
= sgn(IMF
N
)(|IMF
N
| t
n
) (3)
The threshold t
n
is estimated by using a window of noise
from the original signal, in which the boundaries of this
window are used to extract a region of noise. Then, the
standard deviation of each region is estimated and used as
the threshold (t
1
, . . . , t
N
).
Finally, the ltered signal is obtained by linear summation
of the denoised IMFs.
D. The Hilbert Transform & Envelope
The Hilbert transform of a given signal x(t) is dened as a
function of time, as follows:
x(t) = H[x(t)] =
1
_
+
x()
(t )
d
It can be obtained by applying the convolution between the
given signal x(t) and 1/t. It is the response of a linear time-
invariant lter, as follows:
x(t) =
1
t
x(t) (4)
Applying the Fourier transform to Eq. 4:
F{ x(t)} =
1
F
_
1
t
_
F{x(t)} (5)
Since,
F
_
1
t
_
=
_
+
1
x
e
j2fxdx
= jsgn(f)
With,
jsgn(f) =
_
_
_
j, if f > 0
0, if f = 0
j, if f < 0
Then, the Fourier transform of the Hilbert transform dened
in Eq. 5 can be obtained by the following expression:
F{ x(t)} = jsgn(f)F{x(t)} (6)
The Hilbert transform is easier to understand in the fre-
quency domain than in the time domain: the Hilbert transform
does not change the magnitude of F{x(t)}, it only changes
the phase. For negative frequencies, the spectrum of x(t) is
multiplied by j (corresponding to a phase change of /2),
while for positive frequencies, such spectrum is multiplied by j
(corresponding to a phase change of /2). However, to obtain
the result in time domain it is necessary to perform the inverse
Fourier transform.
By denition, the analytic signal can be written explicitly
as:
x
a
(t) = x(t) +j x(t) (7)
The Fourier transform of the analytic signal (Eq. 7), can be
determined by the following expression:
F{x
a
(t)} = F{x(t)} +jF{ x(t)} (8)
Substituting Eq. 6 into Eq. 8, the above expression becomes:
F{x
a
(t)} = F{x(t)} +j[jsgn(f)]F{x(t)}
With,
F{x
a
(t)} =
_
_
_
2F{x(t)}, if f > 0
F{x(t)}, if f = 0
0, if f < 0
To obtain the envelope of the original signal x(t) it is
necessary to take the absolute value of the analytic signal
x
a
(t):
B(t) =
_
x
2
(t) + x
2
(t) (9)
IV. RESULTS
In order to highlight the results of the approach described in
section III, these has also been compared to the ones obtained
using other two different approaches: a conventional one and a
variation (simplication) of the one described above in section
III.
In order to compare the results, the raw data were normal-
ized. Particularly, gait kinematic data are commonly normal-
ized to a certain percentage of the stride, with 0% being heel
strike and 100% being the next successive heel strike of the
same foot. Herein, data were normalized to 100 percent of
one stride for treadmill. Afterwords, normalization by Z-score
0 10 20 30 40 50 60 70 80 90 100
2
0
2
GM
0 10 20 30 40 50 60 70 80 90 100
5
0
5
Gmed
0 10 20 30 40 50 60 70 80 90 100
5
0
5
RF
0 10 20 30 40 50 60 70 80 90 100
2
0
2
Vlat
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
0 10 20 30 40 50 60 70 80 90 100
2
0
2
BF
0 10 20 30 40 50 60 70 80 90 100
2
0
2
TA
0 10 20 30 40 50 60 70 80 90 100
5
0
5
LG
0 10 20 30 40 50 60 70 80 90 100
2
0
2
SOL
100% Stride Normalization
Fig. 5. Normalized raw data of the right/left (blue/red) lower extremities.
0 10 20 30 40 50 60 70 80 90 100
2
0
2
GM
0 10 20 30 40 50 60 70 80 90 100
5
0
5
Gmed
0 10 20 30 40 50 60 70 80 90 100
5
0
5
RF
0 10 20 30 40 50 60 70 80 90 100
2
0
2
Vlat
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
0 10 20 30 40 50 60 70 80 90 100
2
0
2
BF
0 10 20 30 40 50 60 70 80 90 100
2
0
2
TA
0 10 20 30 40 50 60 70 80 90 100
5
0
5
LG
0 10 20 30 40 50 60 70 80 90 100
2
0
2
SOL
Fig. 6. Normalized raw data of the right/left (blue/red) lower extremities with
a half cycle shift.
were applied to the EMG amplitudes. These results can be
shown in Fig. 5.
If these results are shifted in a half cycle, it can clearly
be seen how the activation patterns are the same for both
extremities, as shown in Fig. 6.
As a conventional approach, it is understood that the follow-
ing operations are computed: the normalized EMG raw data
were ltered, using a rst order low pass Butterworth lter
(10 Hz cutoff frequency), and full-wave rectied into a linear
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
GM
0 10 20 30 40 50 60 70 80 90 100
0
1
2
Gmed
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
RF
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
Vlat
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
BF
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
TA
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
LG
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
SOL
Fig. 7. Lowpass ltering and rectication of the normalized raw data of the
right/left (blue/red) lower extremities.
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
GM
0 10 20 30 40 50 60 70 80 90 100
0
1
2
Gmed
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
RF
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
Vlat
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
BF
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
TA
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
LG
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
SOL
Fig. 8. Lowpass ltering and rectication of the normalized raw data of the
right/left (blue/red) lower extremities with a half cycle shift.
envelope, as shown in Fig. 7.
Again, if these results are shifted in a half cycle, it can
clearly be seen how the activation patterns are the same for
both extremities, as shown in Fig. 8.
The other method to be compared with is indeed a variation
of the method described in section III, and it consists in not
considering either PCA or ICA, but only EMD. Therefore, the
normalized raw data is only passed through the EMD lter.
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
GM
0 10 20 30 40 50 60 70 80 90 100
0
2
4
Gmed
0 10 20 30 40 50 60 70 80 90 100
0
1
2
RF
0 10 20 30 40 50 60 70 80 90 100
0
2
4
Vlat
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
0 10 20 30 40 50 60 70 80 90 100
0
1
2
BF
0 10 20 30 40 50 60 70 80 90 100
0
1
2
TA
0 10 20 30 40 50 60 70 80 90 100
0
2
4
LG
0 10 20 30 40 50 60 70 80 90 100
0
1
2
SOL
Fig. 9. EMD ltering and thresholding of the normalized raw data of the
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
GM
0 10 20 30 40 50 60 70 80 90 100
0
2
4
Gmed
0 10 20 30 40 50 60 70 80 90 100
0
1
2
RF
0 10 20 30 40 50 60 70 80 90 100
0
2
4
Vlat
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
0 10 20 30 40 50 60 70 80 90 100
0
1
2
BF
0 10 20 30 40 50 60 70 80 90 100
0
1
2
TA
0 10 20 30 40 50 60 70 80 90 100
0
2
4
LG
0 10 20 30 40 50 60 70 80 90 100
0
1
2
SOL
Fig. 10. EMD ltering and thresholding of the normalized raw data of the
These results are shown in Fig. 9.
Once again, if these results are shifted in a half cycle, it
can clearly be seen how the activation patterns are the same
for both extremities, as shown in Fig. 10.
The results of the method proposed in section III, are shown
in Fig. 11. Once again, and in order to show how the activation
patterns are closely related in both extremities, these results
are reported with a shift of half a cycle, as shown in Fig 12.
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
GM
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
Gmed
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
RF
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
Vlat
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
0 10 20 30 40 50 60 70 80 90 100
0
0.05
0.1
BF
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
TA
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
LG
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
SOL
Fig. 11. Proposed method for processing the normalized raw data of the
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
GM
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
Gmed
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
RF
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
Vlat
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
0 10 20 30 40 50 60 70 80 90 100
0
0.05
0.1
BF
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
TA
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
LG
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
SOL
Fig. 12. Proposed method for processing the normalized raw data of the
More results are reported in Fig. 13, Fig. 14 and Fig.
15. These results highlight the differences among these three
methods. In order to have a closer look, the activation of the
RF, TA, LG and SOL are separately reported in Fig. 16.
In order to see the correlation between the activation pat-
terns and the kinematics of the movements, the kinematic
information traduced in the movement of the hip, knee and
ankle, are also reported in Fig 20.
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
GM
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
0 10 20 30 40 50 60 70 80 90 100
0
1
2
Gmed
0 10 20 30 40 50 60 70 80 90 100
0
2
4
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
RF
0 10 20 30 40 50 60 70 80 90 100
0
1
2
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
Vlat
0 10 20 30 40 50 60 70 80 90 100
0
2
4
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
BF
0 10 20 30 40 50 60 70 80 90 100
0
1
2
0 10 20 30 40 50 60 70 80 90 100
0
0.2
0.4
TA
0 10 20 30 40 50 60 70 80 90 100
0
1
2
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
LG
0 10 20 30 40 50 60 70 80 90 100
0
2
4
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
SOL
0 10 20 30 40 50 60 70 80 90 100
0
1
2
Fig. 13. Lowpass ltering and rectication of the normalized raw data of
the right (blue) lower extremity versus EMD ltering and thresholding of the
normalized raw data of the right (green) lower extremity.
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
GM
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
0 10 20 30 40 50 60 70 80 90 100
0
1
2
Gmed
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
RF
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
Vlat
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
BF
0 10 20 30 40 50 60 70 80 90 100
0
0.05
0.1
0 10 20 30 40 50 60 70 80 90 100
0
0.2
0.4
TA
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
LG
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
SOL
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
Fig. 14. Lowpass ltering and rectication of the normalized raw data of
the right (blue) lower extremity versus proposed method for processing the
V. DISCUSSION
Conventional processing methods based on the use of low-
pass lters eliminate frequency components which can be of
interest in studying the muscle functions. This issue diminishes
the information being processed and provides an incomplete
basis for determining the onset of motor activity. The proposed
approach based on PCA, ICA and EMD ltering avoids this
fact, while also decreases the effect of cross-talk which is
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
GM
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
0 10 20 30 40 50 60 70 80 90 100
0
2
4
Gmed
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
0 10 20 30 40 50 60 70 80 90 100
0
1
2
RF
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
0 10 20 30 40 50 60 70 80 90 100
0
2
4
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
Vlat
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
0 10 20 30 40 50 60 70 80 90 100
0
1
2
BF
0 10 20 30 40 50 60 70 80 90 100
0
0.05
0.1
0 10 20 30 40 50 60 70 80 90 100
0
1
2
TA
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
0 10 20 30 40 50 60 70 80 90 100
0
2
4
LG
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
0 10 20 30 40 50 60 70 80 90 100
0
1
2
SOL
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
Fig. 15. EMD ltering and thresholding of the normalized raw data of
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
(a)
0 10 20 30 40 50 60 70 80 90 100
0
1
2
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
(b)
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
0 10 20 30 40 50 60 70 80 90 100
0
1
2
(c)
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
Fig. 16. Comparison among the three methods over the RF. (a) Lowpass
ltering and rectication of the normalized raw data of the right (blue) lower
extremity versus EMD ltering and thresholding of the normalized raw data
of the right (green) lower extremity. (b) Lowpass ltering and rectication of
the normalized raw data of the right (blue) lower extremity versus proposed
method for processing the normalized raw data of the right (green) lower
extremity. (c) EMD ltering and thresholding of the normalized raw data of
present in the records. Similarly, the dissipation of the muscle
signal through the tissue, due to volume conduction, makes
it difcult to identify individual patterns of activation by
0 10 20 30 40 50 60 70 80 90 100
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
(a)
0 10 20 30 40 50 60 70 80 90 100
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
0.3
0.4
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
(b)
0 10 20 30 40 50 60 70 80 90 100
0
0.05
0.1
0.15
0.2
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
0 10 20 30 40 50 60 70 80 90 100
0
1
2
(c)
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
Fig. 17. Comparison among the three methods over the TA. (a) Lowpass
conventional methods. Herein, the proposed approach is better
than the traditional methods, as it can efciently separate by
blind source separation, the signals corresponding the muscle
activation and others that belong to background noise. More-
over, the proposed approach reduces the undesirable effects of
these factors, thus constituting an important tool in processing
EMG signals. In addition, computing the envelopes by using
the Hilbert transform, enables the chance of considering all the
components of the signals, preventing the loss of information
due to low pass ltering (and/or rectication) of the signal,
which is a common practice in conventional methods.
According to the results, the proposed approach for process-
ing the EMG signals, however, gave no further insights about
the activity of the eight muscles that were studied. Indeed,
although the proposed approach has a major difference com-
pared to the conventional approaches, particularly in aspects
regarding low pass ltering and envelope reconstruction, such
method gave similar results compared to, not only the other
two approaches studied in this manuscript, but also to previous
studies that can be found in literature.
Therefore, it can be established that for healthy subjects,
conventional approaches can be considered reliable, and along
with the proposed approach, they can allow obtaining signals
entirely consistent with the biomechanical gait, showing a
high degree of neuromotor organization, which is reected
in the observed phase shift between the left and right lower
extremities. This situation is based on the activation of central
pattern generators and their complex organization. Such high
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
(a)
0 10 20 30 40 50 60 70 80 90 100
0
2
4
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
(b)
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
0 10 20 30 40 50 60 70 80 90 100
0
2
4
(c)
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
Fig. 18. Comparison among the three methods over the LG. (a) Lowpass
degree of neuromotor is so evident that can clearly be seen for
the three methods reported in this manuscript: in Fig. 7 and
Fig. 8 for the conventional approach, in Fig. 9 and Fig. 10
for the approach using only EMD ltering, and in Fig. 11 and
Fig. 12 for the proposed approach. Moreover, even without
any kind of signal processing (only normalization) applied to
the EMG signals, this issue can still be noticed clearly (see
Fig. 5 and 6).
Nevertheless, there is a great advantage on using the pro-
posed approach (based on PCA, ICA and EMD) in patients
with neurological pathologies, or with other types of diseases
dealing with motor impairments, as if compared to conven-
tional approaches, it is not necessary to know a priori the
cut-off frequency to be used for low pass ltering, or band
pass ltering as well. Therefore, this is a key issue, as for
different neurological diseases, different cut-off frequencies
are required, otherwise the resulting motor pattern activations
can be biased.
However, conventional methods have their own drawbacks,
since they cannot allow the user to set independently the
detection, creating false alarm probabilities. This issue can
clearly be seen in Fig. 17 for the TA muscle, in which the
traditional method reports always activation during the stride,
in circumstances that there is activation only at the beginning,
at the middle, and at the end of the stride.
In order to achieve early activations of the pattern of
movements, a higher quality is required. This implies higher
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
(a)
0 10 20 30 40 50 60 70 80 90 100
0
1
2
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
(b)
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
Z
s
c
o
r
e

N
o
r
m
a
l
i
z
a
t
i
o
n
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
1.5
2
(c)
0 10 20 30 40 50 60 70 80 90 100
0
0.05
0.1
0.15
0.2
Fig. 19. Comparison among the three methods over the SOL. (a) Lowpass
0 10 20 30 40 50 60 70 80 90 100
10
0
10
20
30
40
A
n
g
l
e

[
D
e
g
r
e
e
]
HIP
0 10 20 30 40 50 60 70 80 90 100
0
10
20
30
40
50
60
A
n
g
l
e

[
D
e
g
r
e
e
]
KNEE
0 10 20 30 40 50 60 70 80 90 100
10
5
0
5
10
15
A
n
g
l
e

[
D
e
g
r
e
e
]
ANKLE
Fig. 20. Angular position of the hip, knee and ankle during the cycle gait.
degree of resolution in processing EMG signals. Herein, if
more resolution is requested, conventional approaches can
fail or cannot achieve such requirement, while the proposed
method can t perfectly, as low pass lters are neither used
nor required. This can be demonstrated in Fig. 16, Fig. 17, 18
and 19, in which the performance of the proposed method is
comparatively higher than the other two, particularly against
the conventional approach.
It is also important to highlight the fact that the three
method, and especially the one proposed in this manuscript,
have a strong correlation with the kinematic activity expressed
in terms of angular displacement of the hip, knee and ankle
(see Fig. 20). However, such correlation is not fully achieved
in the case of the TA muscle, in which either the conventional
method or the method based on only EMD couldnt accom-
plish it, meanwhile the proposed method (based on PCA, ICA
and EMD) could afford it.
VI. CONCLUSION
The use of the proposed method for processing EMG
signals during human walking was effective for determining
the muscle activation patterns. These results are conrmed to
be coherent and are aligned with the ones found in literature,
but not only, these are consistent with the two other methods
for which they were compared.
The results show and conrm the robustness of the proposed
method, as the obtained motor activation sequences are closely
related to the functional mechanical behavior of the lower
extremities during walking.
These results provide an important basis for future research
in the eld of biological signal processing, opening the possi-
bility for implementing more accurate models, both in process-
ing and analyzing EMG signals. In addition, the improvement
in signal processing techniques can affect the applications that
can be covered, expanding the eld of application. Herein, this
could be useful not only for the diagnosis and treatment of
several clinical pathologies, but also for the development of
advanced human machine interfaces.
REFERENCES
[1] S. Huang and D. Ferris, Muscle activation patterns during walking
from transtibial amputees recorded within the residual limb-prosthetic
interface, Journal of NeuroEngineering and Rehabilitation, vol. 9,
no. 55, pp. 116, 2012.
[2] R. T. Lauer, C. Stackhouse, P. A. Shewokis, B. T. Smith, M. Orlin,
and J. J. McCarthy, Assessment of wavelet analysis of gait in children
with typical development and cerebral palsy, Journal of Biomechanics,
vol. 38, no. 6, pp. 13511357, 2005.
[3] F. Manganelli, C. Pisciotta, R. Dubbioso, R. Iodice, C. Criscuolo,
L. Ruggiero, G. D. Michele, and L. Santoro, Electrophysiological
characterisation in hereditary spastic paraplegia type 5, Clinical Neu-
rophysiology, vol. 122, no. 4, pp. 819 822, 2011.
[4] P. Tropea, V. Monaco, M. Coscia, F. Posteraro, and S. Micera, Effects
of early and intensive neuro-rehabilitative treatment on muscle synergies
in acute post-stroke patients: a pilot study, Journal of NeuroEngineering
and Rehabilitation, vol. 10, no. 1, pp. 115, 2013.
[5] V. Santilli, M. A. Frascarelli, M. Paoloni, F. Frascarelli, F. Camerota,
L. De Natale, and F. De Santis, Peroneus longus muscle activation pat-
tern during gait cycle in athletes affected by functional ankle instability:
A surface electromyographic study, The American Journal of Sports
Medicine, vol. 33, no. 8, pp. 11831187, 2005.
[6] J. Perry, Gait analysis: Normal and pathological function, 2nd ed. Slack
Incorporated, 2010.
[7] A. Pedotti, A study of motor coordination and neuromuscular activities
in human locomotion, Biological Cybernetics, vol. 26, no. 1, pp. 5362,
1977.
[8] C. Frigo and R. Shiavi, Applications in movement and gait analysis,
in Electromyography: physiology, engineering, and noninvasive appli-
cations, R. Merletti and P. J. Parker, Eds. Wiley-IEEE Press, 2004, pp.
381401.
[9] D. Winter and H. Yack, EMG proles during normal human walking:
stride-to-stride and inter-subject variability, Electroencephalography
and Clinical Neurophysiology, vol. 67, no. 5, pp. 402 411, 1987.
[10] D. A. Winter, Biomechanics and motor control of human movement.
John Wiley & Sons, Inc., 2009, pp. 4581.
[11] S.-S. Chang, C. De Luca, and S. Nawab, Aliasing rejection in precision
decomposition of emg signals, in Engineering in Medicine and Biology
Society, 2008. EMBS 2008. 30th Annual International Conference of the
IEEE, 2008, pp. 49724975.
[12] D. Winter, A. Fuglevand, and S. Archer, Crosstalk in surface elec-
tromyography: Theoretical and practical estimates, Journal of Elec-
tromyography and Kinesiology, vol. 4, no. 1, pp. 1526, 1994.
[13] C. Frigo and P. Crenna, Multichannel SEMG in clinical gait analysis:
A review and state-of-the-art, Clinical Biomechanics, vol. 24, no. 3,
pp. 236245, 2009.
[14] P. Bonato, T. DAlessio, and M. Knaitz, A statistical method for the
measurement of muscle activation intervals from surface myoelectric
signal during gait, Biomedical Engineering, IEEE Transactions on,
vol. 45, no. 3, pp. 287299, 1998.
[15] G. Staude, C. Flachenecker, M. Daumer, and W. Wolf, Onset detection
in surface electromyographic signals: A systematic comparison of meth-
ods, EURASIP Journal on Advances in Signal Processing, vol. 2001,
no. 2, pp. 6781, 2001.
[16] X. Li and A. Aruin, Muscle activity onset time detection using teager-
kaiser energy operator, in Engineering in Medicine and Biology Society,
2005. IEEE-EMBS 2005. 27th Annual International Conference of the,
2005, pp. 75497552.
[17] R. Shiavi, C. Frigo, and A. Pedotti, Electromyographic signals during
gait: Criteria for envelope ltering and number of strides, Medical and
Biological Engineering and Computing, vol. 36, no. 2, pp. 171178,
1998.
[18] S. J. Olney and D. A. Winter, Predictions of knee and ankle moments
of force in walking from EMG and kinematic data, Journal of Biome-
chanics, vol. 18, no. 1, pp. 9 20, 1985.
[19] J. Nielsen, L. Arendt-Nielsen, and A. Pedotti, Power spectrum analysis
of the rectied electromyogram during gait for normals and patients,
Journal of Electromyography and Kinesiology, vol. 4, no. 2, pp. 105
115, 1994.
[20] C. J. De Luca, The use of surface electromyography in biomechanics,
Journal of applied biomechanics, vol. 13, no. 2, pp. 135163, 1997.
[21] M. G. Benedetti, P. Bonato, F. Catani, T. DAlessio, M. Knaitz,
M. Marcacci, and L. Simoncini, Myoelectric activation pattern during
gait in total knee replacement: relationship with kinematics, kinetics,
and clinical outcome, Rehabilitation Engineering, IEEE Transactions
on, vol. 7, no. 2, pp. 140149, 1999.
[22] G. R. Naik, D. K. Kumar, V. P. Singh, and M. Palaniswami, Hand
gestures for HCI using ICA of EMG, in Proceedings of the HCSNet
workshop on Use of vision in human-computer interaction-Volume 56.
Australian Computer Society, Inc., 2006, pp. 6772.
[23] X. Ren, Z. Yan, Z. Wang, and X. Hu, Noise reduction based on ICA
decomposition and wavelet transform for the extraction of motor unit
action potentials, Journal of Neuroscience Methods, vol. 158, no. 2, pp.
313 322, 2006.
[24] A. Delorme, T. Sejnowski, and S. Makeig, Enhanced detection of
artifacts in EEG data using higher-order statistics and independent
component analysis, NeuroImage, vol. 34, no. 4, pp. 14431449, 2007.
[25] D. A. Alvarez and E. Giraldo, ICA aplicado a la extracci on de
caractersticas en im agenes, Scientia Et Technica, vol. XIV, pp. 43
48, 2008.
[26] A. Hyv arinen and E. Oja, Independent component analysis: algorithms
and applications, Neural networks, vol. 13, no. 4, pp. 411430, 2000.
[27] N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Q. Zheng, N.-
C. Yen, C. C. Tung, and H. H. Liu, The empirical mode decomposition
and the hilbert spectrum for nonlinear and non-stationary time series
analysis, Proceedings of the Royal Society of London. Series A:
Mathematical, Physical and Engineering Sciences, vol. 454, no. 1971,
pp. 903995, 1998.
[28] F. E. Hern andez-Montero and M. Guti errez-Garca, Enfoques del
an alisis de envolvente al procesamiento de vibraciones para el di-
agn ostico de maquinarias, Ingeniera Mec anica, vol. 13, no. 1, pp.
3140, 2010.
[29] A. O. Andrade, S. Nasuto, P. Kyberd, C. M. Sweeney-Reed, and
F. Van Kanijn, EMG signal ltering based on empirical mode decom-
position, Biomedical Signal Processing and Control, vol. 1, no. 1, pp.
4455, 2006.

Version 5

Diunggah oleh

Informasi Dokumen

Deskripsi Asli:

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Version 5

Diunggah oleh

Hak Cipta:

Format Tersedia

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING 1

Electromyographic Signal ltering based on

According to the procedure adopted for the data analysis

Anda mungkin juga menyukai