REDUCING THE NUMBER OF CHANNELS AND
SIGNAL-FEATURES FOR AN ACCURATE CLASSIFICATION IN AN
EMG PATTERN RECOGNITION TASK
Iker Mesa, Angel Rubio, Javier Diaz, Jon Legarda and Beatriz Sedano
CEIT, Paseo Mikeletegi, N 48, 20018 San Sebastian, Spain
Keywords:
EMG, sEMG, mRMR, SVM, Pattern recognition, Variable selection, Feature selection.
Abstract:
In this work 32 surface Electromyography (sEMG) electrode locations and 41 signal-features are evaluated
in order to achieve an accurate classification rate in a static-hand gesture classification task. A novel imple-
mentation of the minimal Redundancy Maximal Relevance (mRMR) Variable Selection algorithm is proposed
with the aim of selecting the most informative and least redundant combination of sEMG channels and signal
features. The performance of the new algorithm and of the selected set of channels and signal-features are
tested with a Support Vector Machine classifier.
1 INTRODUCTION
Surface Electromyography (sEMG) is a noninvasive
technique to measure the electrical activity on the skin
produced by the muscles beneath it. With disposable
adhesive electrodes (channels) the sum of multiple
Motor Unit Action Potentials (MUAP) can be easily
captured. The interpretation of this electrophysiolog-
ical activity leads to diverse applications. The most
common one is to diagnose neuromuscular disorders
like dystrophy, tremor or nocturnal bruxism (Merletti
and Parker, 2004) (Fuglsang-Frederiksen and Pug-
dahl, 2010). In addition, sEMG is being largely used
in many other different research areas such as ex-
oskeletons (Khokhar et al., 2010) or powered upper-
limb prostheses (Tenore et al., 2007), (Kuiken et al.,
2007).
The aforementioned applications are based on the
following scheme: A) recording the sEMG signals
from several channels; B) extraction of features from
the signals to represent the recordings using a vec-
tor of variables; C) application of a dimensionality
reduction technique to the vector; and finally D) clas-
sification of the vector of variables in one of K pos-
sible classes. The meaning of each of these classes
depends on the application. For example, the classes
could be the different gestures that an amputee wants
to do with his prosthesis (Tenore et al., 2009); or the
kind of tremor that a patient suffers (Palmes et al.,
2010).
The classification of step (D) is done by machine
learning algorithms such as Neural Networks (Tenore
et al., 2009) or Support Vector Machines (Palmes
et al., 2010). Naturally, supervised training is needed;
using training samples (examples of vectors) the clas-
sifier learns how to classify them into the different
classes. Next, the generalization properties of the
classifier can be tested with validation samples (cross-
validation).
In this work, the classes represent different hand
static-gestures among K different ones. In addition,
we use the term trial for the set of sEMG signals
recorded during a period of time in which the user
performed a static-gesture k; the term feature for the
parameter obtained after a given signal-processing is
applied to a signal; and the term variable, denoted as
z, for the pair (channel, signal-feature). This means
that a variable refers to a specific signal-feature ex-
tracted from a specific channel. In addition, we de-
note to the set of all possible variables as Z. After the
feature extraction step (B), a trial is represented by a
vector of V variables and it is denoted as
z
j
V
j=1
.
The dimensionality reduction step (C) is recom-
mended since the higher the dimension V of the vec-
tor of variables
z
j
V
j=1
, the more training samples
are needed in order to sufficiently train the classifier.
This is known as the curse of dimensionality (Bell-
man, 1961) or as the Hughes phenomenon (Hughes,
1968). However, diminishing the number of variables
may lead to a loss in the discrimination power, and
hence a worse classification performance (Jain and
38
Mesa I., Rubio A., Diaz J., Legarda J. and Sedano B..
REDUCING THE NUMBER OF CHANNELS AND SIGNAL-FEATURES FOR AN ACCURATE CLASSIFICATION IN AN EMG PATTERN RECOGNITION
TASK.
DOI: 10.5220/0003767300380048
In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2012), pages 38-48
ISBN: 978-989-8425-89-8
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
Duin, 2000). In practice, if we have a limited training
set it is better to select only a low number of powerful
variables. Variable z
j
is more powerful than variable
z
h
, with z
j
,z
h
Z, if the patterns from the different
classes can be classified easier with z
j
than with z
h
.
There are two main dimension reduction tech-
niques: Variable Selection and Variable Projection.
On the one hand, the first method aims to determine
a subset of variables that brings together the most im-
portant information of the whole set. On the other
hand, the projection techniques transforms the data to
a space of fewer dimensions by linear or non-linear
combinations of the original variables. As the Vari-
able Selection techniques do not alter the original rep-
resentation of the variables, but merely select a subset
of them, they provide a direct interpretation possibil-
ity. Specially, the Variable Selection filter methods
rank the variables according to an evaluation function
that relies solely on properties of the data. There is
active research in this area related to the sEMG clas-
sification problem; (Zardoshti-Kermani et al., 1995)
and (Boostani and Moradi, 2003) rank several signal-
features according to Davies-Bouldin or Scattering
criteria without taking into consideration the relation
among the signal-features or the channels. However,
it is known in the Variable Selection field that the m
best independent variables are not the best m variables
(Cover, 1974), (Jain and Duin, 2000). The reason for
this is the redundancy. The combination of two uncor-
related variables could be more informative than the
combination of two correlated variables with the best
individual properties.
The objective of this research is to reduce the
necessary number of the sEMG channels and signal-
features present in an sEMG pattern recognition sys-
tem. There are two main reasons for this reduction.
Firstly, if we reduce the number of physical sEMG
channels we make the sEMG-recording device sim-
pler, thus cheaper. Secondly, if we diminish the num-
ber of features and channels we reduce the dimension
V of the vector
z
j
V
j=1
presented to the classifier,
hence, less training examples are needed and, further-
more, the classifier requires less memory and compu-
tational power.
In this work we propose a novel multivariate filter
method, based on the minimal Redundancy maximal
Relevance algorithm (mRMR) (Peng et al., 2005), to
rank the variables with the aim of finding the most
informative channels and signal-features and, at the
same time, the least redundant. In addition to the
aforementioned advantages, the reduction of the re-
quired number of sEMG channels can also be seen
as an analysis of the best electrode placement for this
kind of applications.
For the accomplishment of the objectives, the fol-
lowing procedure is used. Firstly, sEMG data is
obtained from a static-hand gesture experiment per-
formed by 6 volunteers. After the correspondingdata-
preprocessing and feature-extraction steps, the vari-
ables are ranked with three methods: A) a simple
univariate filter method, the F-statistic, B) the regular
mRMR algorithm, C) a novel implementation of the
mRMR to penalize even more the redundancy. There-
fore, three different ranking lists of variables are ob-
tained for each user.
Secondly, a Support Vector Machine (SVM) clas-
sifier is used to classify the hand gestures with the
top-ranked variables of each list. This way, the per-
formance of each ranking method can be analyzed.
An SVM is used since it performs well in many dif-
ferent fields and it is insensitive to overtraining (Jain
and Duin, 2000) (i.e. a degradation in generaliza-
tion properties due to an excessive number of training
samples).
Thirdly, a study of the top-rankedvariables is done
to select the best combinations of features and chan-
nels. The aim, as mentioned before, is to reduce the
number of variables. Finally, the selected variables
are tested with the SVM classifier and the results are
discussed.
2 METHODS
2.1 Data Acquisition
Six healthy normal-limbed subjects volunteered for
the experiment. None of them reported any muscular
or skin disorders. The following protocol was used
during each session.
After a careful skin preparation, 32 Ag/AgCl dis-
posable electrodes (diameter 10 mm) are equidis-
tantly placed on each subject’s right forearm in 4 rows
of 8 electrodes each. The reference and ground elec-
trodes are placed on the shoulder and wrist respec-
tively. The layout of the electrodes is described in
Figure 1. To ensure an adequate position, a flexible
armband with 32 small guide points is used.
The user sits comfortably in front of a computer
screen and a webcam. With the elbow resting on a
table the subject can perform hand movements with-
out constraints. For the experiment, K = 15 static-
gestures were selected from the Spanish Sign Lan-
guage Alphabet (letters A, B, C, D, E, F, I, K, L, M
N, O, P, Q and U) as they involve a wide variability of
wrist and finger positions.
There are two computers involved in the acquisi-
tion, namely, one for guiding the user and watching
REDUCING THE NUMBER OF CHANNELS AND SIGNAL-FEATURES FOR AN ACCURATE CLASSIFICATION
IN AN EMG PATTERN RECOGNITION TASK
39
Figure 1: The electrode naming and positioning scheme.
the correct execution of the gestures; and another one
for recording the sEMG signals. Each participant per-
forms between 3 and 5 runs of 50 gestures each. A
computer program in the first machine is in charge of:
A) asking the subject to perform a specific gesture by
displaying an aleatory gesture on the screen, B) en-
suring with the webcam that the user is maintaining a
steady gesture for 3 seconds, C) informing the sec-
ond PC about the gesture that the user is asked to do
and if the subject has performed it correctly (steady)
or not (not steady).
The signals are recorded with a sampling fre-
quency of 1000 Hz by a BrainAmp Standard amplifier
connected to the second PC.
2.2 Data Preprocessing
The signals are high-pass filtered (10 Hz) to remove
movement artifacts, and a notch filter is used to re-
move the 50 Hz band (power-line noise). These oper-
ations are performed by 4
th
order Butterworth filters.
Next, the correctly performed gestures of each
subject are arbitrarily divided into 3 datasets:
Training-dataset, Validation-dataset and Test-dataset.
The first dataset is used for training the classifier as
it is explained in Section 2.5, the Validation-dataset is
used in the same section to compare the different Vari-
able Selection techniques, and finally, the Test-dataset
is used in Section 4 to check the performance of the
final selection of variables.
Each of these performed gestures has a duration of
3 seconds. To increase the size of the datasets they
are segmented into trials of 256 ms with an overlap-
ping of 50% (Englehart et al., 2003). Consequently,
the number of available trials in each dataset is de-
scribed in Table 1. Each of these trials contains the
32 sEMG signals (one per channel) recorded during
a 256 ms period. The static-gesture that the user
was performing during the recording of each trial is
known.
2.3 Feature Extraction
Let x
i
be the i-th sample of sEMG signal x, and let N
Table 1: The number of trials in each dataset. Each trial
contains the 32 sEMG signals of 256ms long. The static-
gesture corresponding to each trial is known.
Subject Train-dataset Validation-dataset Test-dataset
S1 4929 1232 0
S2 6382 1595 0
S3 4787 1115 655
S4 5625 1446 964
S5 2218 0 555
S6 3871 968 0
be the signal length. From the sEMG signals of each
of the trials obtained in Section 2.2, the following 41
features are extracted. Note that each trial has 32
sEMG signals, hence, a vector of V = 1312 variables,
z
j
1312
j=1
, is obtained for each trial.
Mean Absolute Value (MAV). It is a very common
sEMG amplitude indicator which provides informa-
tion about the muscle contraction level.
MAV =
1
N
N
i=1
|x
i
| (1)
Median Absolute Value (MedAV). Like MAV, the
MedAV is an indicator of the contraction level. How-
ever, it is more robust against outliers.
MedAV = median
i
|x
i
| (2)
Variance (VAR). As the mean value of the sEMG sig-
nal is 0 the variance of the signal can also be seen as a
measurement of the average signal power (Zardoshti-
Kermani et al., 1995), (Boostani and Moradi, 2003).
VAR =
1
N 1
N
i=1
x
2
i
(3)
Waveform Length (WL). It is the cumulative length
of the waveform over the segment. The resultant
value is an indicator of the waveform’s amplitude, fre-
quency, and duration (Hudgins et al., 1993).
WL =
N1
i=1
i
(4)
Where
i
= |x
i
x
i+1
|
Mean Absolute Difference Value (MADV). The
mean value of the difference in amplitude between ad-
jacent samples.
MADV =
1
N 1
N1
i=1
i
(5)
Note that, as N is constant the MADV provides ex-
actly the same information as WL. We compute both
features to test the different Variable Selection meth-
ods, since a good method has to be able to find the
BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing
40
correlation between both features and discard one of
them. As it is explained in Section 3.1, there are some
methods that are unable to do this.
Zero Crossing (ZC). It counts the number of times
that the signal crosses the zero amplitude axis. To
avoid signal crossing counts due to low-level noise,
a threshold ε is included. In our case it is set to ε =
10µV according to (Hudgins et al., 1993).
ZC =
N1
i=1
φ(
i
,x
i
,x
i+1
) (6)
Where:
φ(
i
,x
i
,x
i+1
) =
1 x
i
· x
i+1
< 0,
i
> ε
0 otherwise
Number of Turns (NT). This feature counts the num-
ber of times that the slope of the waveform changes
in sign. This means that it counts the number of lo-
cal maxima and minima in the sEMG signal. To re-
duce noise effects, a slope change is only taken into
account if the difference in amplitude with adjacent
slope changes is at least ε = 30µV (Hudgins et al.,
1993).
Wilson Amplitude (WAMP). The Wilson amplitude
counts the number of times that the difference in am-
plitude for any two consecutive samples exceeds a
certain threshold ε. This ε is set to 50µV according
to (Philipson, L, Larsson, P, 1988).
WAMP =
N1
i=1
ψ(
i
) (7)
Where ψ(
i
) = 1 if
i
> ε; and 0 otherwise.
Amplitude Histogram (A). This feature counts the
number of times that the sEMG signal reaches 9 dif-
ferent amplitude levels. In other words, it is the
voltage histogram of the signal into 9 equal bins
(Zardoshti-Kermani et al., 1995). Therefore, 9 fea-
tures are obtained per signal, named A1-A9.
Auto-Regressive Coefficients (AR). The AR model
of the signal approximates the sample x
i
as a linear
combination of earlier samples plus an independent
error term. The number of samples taken into account
defines the order of the model. As well as (Tkach
et al., 2010) and (Boostani and Moradi, 2003) an or-
der of 4 is set.
x
i
=
4
r=1
a
r
x
ir
+ ε
i
(8)
Where the a
r
are the auto-regressive coefficients and
ε
i
the error term. The a
r
coefficients are computed
minimizing the squared error between the original
signal and its approximation. Each coefficient a
r
is
used as a different feature, hence, 4 different features
are obtained, denoted as AR1-AR4, for each signal.
Cepstral Coefficients (C). Cepstral analysis has been
mainly applied to speech recognition but it has been
also shown to be suitable for sEMG classification
(Kang et al., 1995). The cepstral coefficients c
r
are
obtained from the Auto-Regressive coefficients a
r
.
c
1
= a
1
c
r
= a
r
r1
n=1
1
n
r
a
n
c
rn
(9)
Where a
r
is the r-th auto-regressive coefficient. Con-
sequently, there are 4 cepstral features per signal and
they are denoted as C1-C4.
As in the case of WL and MADV, there is no dif-
ference between the cepstral coefficient c
1
and the
Auto-Regressive one a
1
. Likewise, this fact plays an
importantrole in the comparison of Variable Selection
techniques as it is explained in Section 3.1.
Mean Frequency (Fmean). This feature measures
the mean frequency of the signal’s power spectrum.
Assuming the frequency spectrum is divided into M
frequencies, let f
b
be the b-th frequency of the sig-
nal’s spectrum and P( f
b
) the power of f
b
. The mean
frequency is computed as follows:
Fmean =
M
b=1
f
b
P( f
b
)
M
b=1
P( f
b
)
(10)
Quantiles (Q). The quantiles mark the boundaries be-
tween specific consecutive subsets of the signal spec-
trum, e.g. the Qy quantile is the frequency that marks
the upper boundary of the lower y% of the spectrum’s
power. The following frequency spectrum quantiles
are considered: Q10, Q30, Q50, Q60, Q75, Q90.
Therefore, 6 features are obtained per signal.
Frequency Histogram (F). The frequency spectrum
is divided into 9 equal-size segments and the percent-
ages of power in each of them are taken as features.
Hence, 9 features are obtained per signal and are de-
noted as F1-F9.
2.4 Variable Ranking: mRMR
Algorithm
After the feature extraction step (Section 2.3), each
trial is represented by a vector of variables
z
j
1312
j=1
,
and it belongs to a known class k (static-gesture). We
can measure the discriminant power of each variable
z
j
using the F-statistic as follows:
REDUCING THE NUMBER OF CHANNELS AND SIGNAL-FEATURES FOR AN ACCURATE CLASSIFICATION
IN AN EMG PATTERN RECOGNITION TASK
41
F(z
j
) =
"
K
k=1
n
(k)
j
(z
(k)
j
z
j
)
2
#
/(K 1)
K
k=1
n
(k)
j
t=1
(z
(k)
j
t
z
(k)
j
)
2
/(n
j
K)
(11)
Where:
K : Number of possible classes.
n
j
: Number of samples of variable z
j
.
n
(k)
j
: Number of samples of z
j
within the
k-th class.
z
j
: Mean value of z
j
.
z
(k)
j
: Mean value of z
j
within the k-th class.
z
(k)
j
t
: Sample t-th of z
j
within the k-th class.
F-statistic only gives a measure of the classifica-
tion power of each variable z
j
by itself. However,
the aim is to select a group S of m variables z
j
which
jointly have the largest relevance with the classifica-
tion task. This is called maximal relevance criterion
(Ding and Peng, 2005), and it has the following form:
max D
F
(S), D
F
=
1
|S|
z
j
S
F(z
j
)
(12)
However, we also aim to ensure that the redun-
dancy among the variables is as low as possible.
Therefore, it is necessary to introduce the minimal
redundancy condition. The minimal redundancy cri-
terion based on the Pearson correlation coefficient
(Ding and Peng, 2005) has the following form:
min R
c
(S), R
c
=
1
|S|
2
z
j
,z
h
S
|c(z
j
,z
h
)|
(13)
The Pearson correlation coefficient c(z
j
,z
h
) is
computed as:
c(z
j
,z
h
) =
n
j
t=1
(z
j
t
z
j
)(z
h
t
z
h
)
s
n
j
t=1
(z
j
t
z
j
)
2
s
n
j
t=1
(z
h
t
z
h
)
2
(14)
where z
j
t
is the sample t-th of z
j
.
The minimal redundancy - maximal relevance cri-
terion (mRMR) (Ding and Peng, 2005) (Peng et al.,
2005) combines (12) and (13). The optimization of
these two conditions can be done in several different
ways. One commonly used combination criteria is the
following:
max (D
F
/R
c
) (15)
A near optimal solution of the mRMR is obtained
with the following iterative method (Ding and Peng,
2005). The first variable is selected according to (12),
i.e. the variable with the highest F(z
j
). Next, the rest
of variables are added to the set S in the following in-
cremental way. Assuming that we have already cho-
sen m 1 variables, we select the m
th
variable from
the set {Z S
m1
} (i.e. all the variables except those
already selected) that combines the following condi-
tions:
max
z
j
ZS
m1
F(z
j
) (16)
min
z
j
ZS
m1
1
m 1
z
h
S
m1
|c(z
j
,z
h
)| (17)
The equation (16) is equivalent to (12), while (17)
is an approximation of (13). Therefore, if we use the
combination criteria (15) the desired m
th
variable sat-
isfies the following condition known as FCQ: F-test
Correlation Quotient (Ding and Peng, 2005):
max
z
j
ZS
m1
F(z
j
)
1
m1
z
h
S
m1
|c(z
j
,z
h
)|
(18)
Moreover, in this work we introduce a new com-
bination criterion that gives more weight to the redun-
dancy factor. We call it FCO: F-test Correlation Out:
max
z
j
ZS
m1
F(z
j
) ·
1 max
z
h
S
m1
|c(z
j
,z
h
)|
(19)
The reason for this new scheme is that we dis-
covered that the FCQ criterion is unable to separate
highly correlated variables in the rank lists. The FCQ
(18) selects a variable z
j
for the set S if its F-value is
high and if the averaged correlation with previously
selected variables (set S
m1
) is low. However, if there
is a variable in S
m1
that has a high correlation with
z
j
it may go unnoticed. On the other hand, the FCO
(19) selects a variable z
j
whose individual F-value is
high, but whose maximum individual correlation with
the variables of S
m1
is the lowest possible. As it is
shown in Section 3.1, FCO outperforms FCQ in the
sEMG classification task.
2.4.1 Application
For each subject, three different ranking lists of
variables are computed from the subject’s Training-
dataset. The first ranking list is obtained using only
a F-statistic (11) to rank the variables; this method is
called Base ranking. The second and third ranking
lists are obtained with the FCQ and FCO algorithms
respectively.
2.5 Performance of the Variable
Selection Algorithms
In order to evaluate the performance of a Variable Se-
lection’s ranking list we have to compute the classi-
BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing
42
fication rate that is obtained when we only maintain
in the vector
z
j
1312
j=1
variables that are top-ranked in
the ranking list. Given a Variable Selection method
MET (i.e. a ranking list of variables) with MET
{Base,FCQ, FCO}, we denote as z
MET
p
to a variable
that is top-ranked in the ranking list MET. If we only
maintain in
z
j
1312
j=1
the top-P variables of the rank-
ing list MET the new vector is denoted as
z
MET
p
P
p=1
(e.g. if we want to represent the trials with the Top-
5 variables of the FCQ list the vector is denoted as
n
z
FCQ
p
o
5
p=1
).
The following procedure is followed with a Sup-
port Vector Machine (SVM) classifier and the data of
subjects {S1, S2, S3, S4, S6} to evaluate the perfor-
mance of the Base, FCQ and FCO algorithms. Sub-
ject S5 is excluded as according to Table 1 no trials
are present in that subject’s Validation-dataset. Given
a subject and a Variable Selection method MET we
have: A) the user’s Training-dataset and Validation-
dataset, in which the trials are represented by vectors
of 1312 variables (Section 2.3); and B) a ranking list
of the variables (Section 2.4.1). The evaluation pro-
cedure of the ranking list is as follows:
1. The vectors of variables from the Training-dataset
and Validation-dataset are reduced to the Top-5
variables z
p
of the ranking list MET (i.e. now
each trial is represented by a vector of 5 variables,
z
MET
p
5
p=1
).
2. A grid search using a 5-fold cross-validation with
the Training-dataset is used to find the best ker-
nel’s parameters of the SVM (radial basis function
as kernel)(Chang and Lin, 2011).
3. The SVM classifier is trained with the vectors
z
MET
p
5
p=1
of the Training-dataset.
4. The vectors
z
MET
p
5
p=1
of the Validation-dataset
are classified, hence, the classification rate is ob-
tained.
The results are shown in Figure 2 for subjects
{S1, S2, S3, S4, S6} and Variable Selection methods
{Base, FCQ, FCO}. In addition to the Top-5 variables
z
p
of the ranking lists (step 1 of the above procedure),
the Top-{10, 20, 30, 50, 90, 100 and 200} variables of
each ranking list are also evaluated (i.e. P goes from
P = 5 to P = 200 for each method).
0 50 100 150 200
20
30
40
50
60
70
80
90
P: Nº of top−ranked variables
Classification Rate [%]
(a) Subject S1
0 50 100 150 200
20
30
40
50
60
70
80
90
P: Nº of top−ranked variables
Classification Rate [%]
(b) Subject S2
0 50 100 150 200
20
30
40
50
60
70
80
90
P: Nº of top−ranked variables
Classification Rate [%]
(c) Subject S3
0 50 100 150 200
20
30
40
50
60
70
80
90
P: Nº of top−ranked variables
Classification Rate [%]
(d) Subject S4
0 50 100 150 200
20
30
40
50
60
70
80
90
P: Nº of top−ranked variables
Classification Rate [%]
(e) Subject S6
Figure 2: Classification rate for the Validation-dataset
of each user when the vector
z
j
1312
j=1
is reduced to
z
MET
p
P
p=1
, where z
MET
p
are top-ranked variables of the
ranking list MET, with MET {Base,FCQ,FCO} and
P = {5,10,20,30,50,90,100,200}.
3 RESULTS
3.1 Analysis of the Performance of the
Variable Selection Algorithms
Figure 2 shows that the more variables the vector
z
MET
p
P
p=1
has, the better the classification rate. This
is due to the fact that in our experiment the num-
ber of available training samples is much higher than
the dimension P of the vector
z
MET
p
P
p=1
(e.g. for
subject S1 there are 4929 training samples accord-
ing to Table 1, but the number of variables of the
vector
z
MET
p
P
p=1
is between 5 and 200). There-
fore, we could infer that maintaining the whole vector
z
j
1312
j=1
would be the best option. Nevertheless, as
it is explained in Section 1, it would require more
training samples and the classifier would be much
more complex (Jain and Duin, 2000). In addition,
note that the advantage of having more variables be-
comes insignificant at some point depending on what
variables we are maintaining. For example, for sub-
REDUCING THE NUMBER OF CHANNELS AND SIGNAL-FEATURES FOR AN ACCURATE CLASSIFICATION
IN AN EMG PATTERN RECOGNITION TASK
43
ject S3 the difference in performance between using
z
FCO
p
50
p=1
or
z
FCO
p
200
p=1
is insignificant, whereas if
we are using the ranking list Base the difference be-
tween
z
Base
p
50
p=1
or
z
Base
p
200
p=1
is high. This high-
lights the importance of the adequate selection of vari-
ables to have a set of variables as reduced as possi-
ble but obtaining, at the same time, a high classifica-
tion rate with it; the set of variables capable of this is
defined as a compact-set and the following facts re-
veal FCO as an excellent algorithm to search for this
compact-set.
The mRMR Variable Selection algorithms (FCO,
FCQ) perform better than the classic univariate
method of ranking by F-statistic. Figure 2 shows
that when the variables are selected according to
the Base rank the number of variables has to be
higher to achieve an acceptable classification rate.
It is true that the top-ranked variablesof Base have
large discriminant power individually, but they are
redundant and when too few are used the classi-
fier is unable to extract enough information from
them. However, when the mRMR lists (FCO,
FCQ) are used the classifier achieves better clas-
sification rate with less variables. Moreover, the
fact that the FCO’s classification rates are better
than the FCQ’s ones illustrates even more that re-
ducing redundancy is a key factor.
FCO algorithm shows better performance dealing
with highly correlated variables. On the one hand,
if we count the number of times that each fea-
ture appears in the ranking lists (Table 2), we can
see that there are groups of features that appear
more than others. It is clearly visible that the first
AR coefficient (AR1) is a strong candidate for the
compact-set as it appears many times (i.e. it ex-
tracts important information from many channels)
and the three methods rank it very high. On the
other hand, the first cepstral coefficient (Ceps1)
also appears many times in the FCQ and Base
ranks, but neverin the FCO one. The reason is that
if we look at (9) we see that if features AR1 and
C1 are computed in the same channel, the coeffi-
cients are identical (with opposite signs). Unsur-
prisingly, we verified that the FCQ and Base top-
ranked variables whose feature-term was AR1 and
C1 had the same channel-term; this means that
while FCO finds the correlation, FCQ and Base
do not. In addition, the same behaviour was found
in the WL (4) and MADV (5) features. Therefore,
we can state that the F-test and the FCQ criteria
fail to rank appropriately highly correlated vari-
ables.
FCO algorithm finds key variables. Figure 2
shows that when the Base or FCQ lists are used,
the respective vectors
z
Base
p
P
p=1
or
n
z
FCQ
p
o
P
p=1
have to have a high amount of variables to approx-
imate the rate achieved with
z
FCO
p
P
p=1
. Based
on Table 2 we could make the assumption that the
classifier reaches the desired rate when features
like MAV, MedAV or MADV are computed. This
is supported by the fact that these variables are not
in the first tops (Top-{5, 10, 20}) but they increase
their presence as we take larger tops (Top-{30, 50,
100}). However, two facts reject this hypothesis.
Firstly, these features are utterly discarded by the
FCO algorithm (Table 2), they are not even in the
FCO-Top-200. Secondly, at the end, the FCQ and
Base lists include the best FCO variables. This
is explained if the number of coincidences among
the lists is analyzed. Table 3 shows how many
variables from the FCO-Top-P are in the Top-
{5,10,...,200} of the FCQ list. The following ex-
amples are highlighted in bold. Table 3 shows that
only 8 variables from the FCO-Top-30 are in the
FCQ-Top-30, but if we look at the FCQ-Top-200
we see that it has 26 variables from the FCO-Top-
30. On the other hand, only 14 variables from the
FCQ-Top-30 appear in the FCO-Top-200. This is
even more impressive if we check higher tops of
the lists; while the whole FCO-Top-5 and Top-10
appear in the the FCQ-Top-200 list, just 3 and 6
variables of the FCQ-Top-{5, 10} respectivelyare
in the FCO-Top-200. That means that even the
most important variables of the FCQ are not that
informative according to the FCO method. The
same trends were found for the case of Base-FCO.
From these facts and linked with the performance
data, we can infer that there is a group of key variables
that hold the important information, and they are well
ranked by the FCO algorithm. The FCQ and Base
lists do not achieve good classification rates due to the
high number of variables, but rather at the end these
lists include the best variables of the FCO list; the
key variables. As explained in Section 1, a variable
is a combination of a channel and the signal-feature
extracted from it, hence, the task now is to analyze
these variables and detect which are the best features
and channels.
3.2 Selection of Features and Channels
The search for the variables of the compact-set en-
tails a selection of features and channels among all the
possible combinations. As it has been shown in Sec-
tion 2.5 the FCO algorithm ranks these combinations
and it highlights some of them as the most important
BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing
44
Table 2: The number of times that a feature appears in
the Top-{30, 50, 100} of the FCO, FCQ and Base rank-
ing lists (e.g. AR1 appears 7 times in the FCO-Top-30, that
means that the information that AR1 extracts from 7 differ-
ent channels is very important). The median across subjects
is given. In gray the values greater than zero.
Top-30 Top-50 Top-100
FCO FCQ Base FCO FCQ Base FCO FCQ Base
MAV 0 1 1 0 2 2 0 4 5
MedAV
0 0 0 0 1 1 0 3 3
VAR
0 0 0 0 0 0 0 1 1
MADV
0 2 2 0 3 4 0 7 8
WL
0 2 2 0 3 4 0 7 8
ZC
1 0 0 2 1 1 6 1 2
NT
3 2 4 4 3 6 7 6 12
WAMP
3 4 3 4 6 4 5 9 8
A1
1 0 0 2 1 0 3 2 0
A2
1 0 0 1 0 0 3 1 0
A3
0 0 0 0 0 0 1 0 0
A4
0 0 0 0 0 0 1 0 0
A5
0 0 0 1 0 0 2 1 0
A6
0 0 0 0 0 0 2 1 0
A7
0 0 0 0 0 0 2 0 0
A8
0 0 0 1 0 0 1 0 0
A9
1 1 0 2 2 1 3 3 2
AR1
7 3 4 8 6 6 10 9 11
AR2
1 0 0 3 1 1 6 3 2
AR3
0 0 0 1 0 0 2 0 0
AR4
0 0 0 0 0 0 0 0 0
C1
0 3 4 0 4 6 0 9 11
C2
1 0 0 1 0 0 2 1 0
C3
1 0 0 1 0 0 1 1 0
C4
0 0 0 0 0 0 2 0 0
Fmean
2 1 1 2 2 2 3 4 3
Q10
0 0 0 0 0 0 1 0 0
Q30
0 0 0 0 0 0 1 0 0
Q50
0 0 0 0 1 0 2 1 1
Q60
0 1 0 0 1 1 1 3 2
Q75
0 1 1 1 2 1 3 4 3
Q90
2 1 1 2 2 2 3 5 4
F1
0 0 0 0 0 0 0 0 0
F2
0 0 0 0 0 0 1 0 0
F3
0 0 0 0 0 0 1 0 0
F4
0 0 0 0 0 0 2 0 0
F5
0 1 0 1 1 0 3 2 1
F6
1 0 0 1 1 0 3 3 1
F7
0 1 0 1 2 0 2 3 2
F8
2 1 0 3 2 0 6 2 2
F9
1 0 0 2 1 0 3 1 0
Table 3: The number of coincidences between the FCO-
Top-P and the FCQ-Top-P (e.g. 26 variables from the FCO-
Top-30 are in the FCQ-Top-200). The median across sub-
jects is given. In bold the examples explained in the text.
FCO-Top-P variables
5 10 20 30 50 70 90 100 200
FCQ-Top-P variables
5 1 2 2 2 2 2 2 2 3
10 3 3 4 4 4 4 4 4 6
20
4 5 5 6 7 8 8 8 10
30 4 6 7 8 10 11 11 12 14
50
5 7 9 13 16 18 19 20 26
70 5 8 11 16 19 22 25 27 34
90
5 9 12 17 22 26 31 32 41
100
5 10 14 18 22 26 32 33 46
200
5 10 18 26 37 47 53 56 85
ones for achieving a high classification rate. However,
not only the best variables are the ones that maximize
the classification rate, but also the ones that minimize
the number of different channels and signal-features
needed. Therefore, the search has to be focused on:
A) features that appear many times in the ranking
list and that share the same channels, since this re-
veals features with complementary information; and
B) channels that appear many times in the highest po-
sitions of the FCO ranking list, because the algorithm
has found them to be the most informative ones.
3.2.1 Features
We analyze the variables of the FCO-Top-100 by
looking for variables whose channel-term is the same
one. In other words, we count the number of times
that any two features share the same channel in the
FCO-Top-100 list. Table 4 shows the best features in
the horizontal axis. We can see that the group of fea-
tures formed by the first and second Auto-Regressive
coefficients (AR1,AR2), the Zero Crossing (ZC), the
Number of Turns (NT), and the Wilson Amplitude
(WAMP) appear together many times. Furthermore,
they also appear many times in the best positions of
the FCO ranking list (Table 2), hence we can state
that they interact perfectly well among themselves
and that they are of great importance for identify-
ing the different kinds of EMG spike burst activity.
In addition, they seem to be able to share a channel
with almost any other feature of the list. Note that
there is no interaction with the features {MAV, Me-
dAV, Var, MADV, WL} because these are not present
in the FCO-Top-100 (Table 2). This is due to the fact
that some of these features do not provide any valu-
able information or are highly correlated with the first
group. Therefore, we can discard the features {MAV,
MedAV, Var, MADV, WL}; and consider the {AR1,
AR2, WAMP, ZC, NT} as the initial members of the
compact-set.
Another group of features that shows great inter-
action is the frequency histogram features (F1-F9).
However, a selection becomes necessary as there are
important differences among them:
On the one hand, we select for the compact-set
the coefficients corresponding to high parts of the
spectrum ({F5-F9}). These coefficients represent
the percentage of power that resides in the 222-
500 Hz band (divided in 5 segments). It is very
interesting to see in Table 4 that {F5-F9} fit with
NT, ZC and the AR coefficients AR1 and AR2.
The reasons are the following. With the number
of turns (NT) we measure the number of spikes
that are generated in the muscle fibers; therefore,
NT grows with this number. However, if the fre-
quency of these spikes is too high they are hardly
distinguishable in the resulting sEMG wave; the
NT does not count those slope turns and the coef-
ficient does not grow. Therefore, the {F5-F9} co-
REDUCING THE NUMBER OF CHANNELS AND SIGNAL-FEATURES FOR AN ACCURATE CLASSIFICATION
IN AN EMG PATTERN RECOGNITION TASK
45
efficients combine perfectly with NT as they pro-
vide the missing information. Similarly, ZC can
be complemented by {F5-F9} as a high number
of spikes may hinder the crossings in the zero am-
plitude axis. Finally, the FCO algorithm does not
find high correlations between the {F5-F9} and
{AR1, AR2}, hence both groups of coefficients
can share the same channels. The interaction of
{F5-F9} with the WAMP feature is not so strong
as some redundancy is found, but given their high
positions in the FCO ranking we decide to main-
tain both {F5-F9} and WAMP in the compact-set.
On the other hand, the information of the {F1-F4}
coefficients is not relevant as it is embedded in
features like NT, ZC and WAMP.
The amplitude histogram features (A1-A9) tend
to appear together and with the {AR1, AR2, WAMP,
ZC, NT } group. However, their interaction with other
features is insignificant and their presence in the dif-
ferent FCO-Tops (Table 2) is not large enough. More-
over, establishing the size of the different amplitude
levels is not straightforward, as it is different for each
user and requires some tuning. Based on these facts,
we can reject them for the compact-set.
The auto-regressive coefficients AR3 and AR4,
and the four Cepstral coefficients (C1-C4) are dis-
carded because they are are not top-ranked by FCO.
For identical reasons we discard also the quantiles
{Q10-Q90}. The information of {Q10, Q30} is al-
ready provided by the group {AR1, AR2, WAMP, ZC,
NT }, and {Q50-Q90} are not necessary as the fre-
quency histogram features {F5-F9} are found to be
more informative.
Finally, the mean frequency (Fmean) is discarded
because even though it is well ranked in the FCO-Top-
30 (it appears twice), its presence in FCO-Top-50 and
FCO-Top-100 is almost the same. This means that its
contribution is not determinant for the classification
task, the FCO algorithm finds that any of the features
{AR1, AR2, WAMP, ZC, NT, F5-F9} provides better
information that Fmean.
In summary, the selected signal-features for the
compact-set are {AR1, AR2, WAMP, ZC, NT, F5-
F9}.
3.2.2 Channels
The selection of the best channels is straightforward.
We search in the different FCO-tops of each user the
most repeated channels and their evolution. Thanks to
the mRMR-FCO algorithm we know that those chan-
nels will carry useful and non-redundant information.
Table 5 shows the number of times that each chan-
nel appears in the FCO-Top-{50,100} (the median
Table 4: The number of times that, in the FCO-Top-100
list, two given features share the same channel, e.g. the ZC
shares two channels with NT, 2 with WAMP, 1 with A1, etc.
The median across subjects is given. We only show in the
horizontal axis the best features. The values greater than
zero are in gray.
ZC NT WAMP AR1 AR2 F5 F6 F7 F8 F9
MAV 0 0 0 0 0 0 0 0 0 0
MedAV
0 0 0 0 0 0 0 0 0 0
VAR
0 0 0 0 0 0 0 0 0 0
MADV
0 0 0 0 0 0 0 0 0 0
WL
0 0 0 0 0 0 0 0 0 0
ZC
2 2 2 2 1 1 1 3 2
NT
2 2 4 1 1 1 1 2 1
WAMP
2 2 2 2 1 1 1 0 0
A1
1 2 1 3 0 0 1 0 0 0
A2
0 0 1 1 0 0 0 0 1 0
A3
0 0 0 0 0 0 0 0 0 0
A4
0 1 0 0 0 0 0 0 0 0
A5
1 1 1 0 1 0 0 0 0 0
A6
1 1 1 1 0 0 0 0 0 0
A7
0 1 0 0 1 0 0 0 0 0
A8
0 1 1 0 0 0 0 0 1 0
A9
1 1 0 1 1 1 0 0 1 1
AR1
2 4 2 0 1 0 0 1 1
AR2
2 1 2 0 1 1 2 2 1
AR3
1 0 1 1 0 0 0 0 1 0
AR4
0 0 0 0 0 0 0 0 0 0
C1
0 0 0 0 0 0 0 0 0 0
C2
1 1 0 2 0 0 0 0 1 1
C3
1 1 1 0 1 0 1 1 0 0
C4
2 1 1 0 1 0 0 0 1 0
Fmean
0 1 1 0 1 0 0 1 1 1
Q10
0 0 1 1 0 0 0 0 0 0
Q30
0 0 0 1 0 0 0 0 0 0
Q50
1 1 1 1 1 1 1 1 1 0
Q60
1 1 0 1 0 0 0 0 1 0
Q75
1 0 0 1 1 0 1 0 1 0
Q90
2 1 0 1 1 0 1 1 1 0
F1
0 0 0 0 0 0 0 0 0 0
F2
1 1 0 0 0 1 0 0 1 1
F3
1 0 0 1 0 0 0 0 0 0
F4
1 0 0 1 1 1 0 0 1 1
F5
1 1 1 1 1 1 1 2 1
F6
1 1 1 0 1 1 1 2 0
F7
1 1 1 0 2 1 1 2 1
F8
3 2 0 1 2 2 2 2 2
F9
2 1 0 1 1 1 0 1 2
across subjects is shown). There are two important
groups clearly visible: {Ch05, Ch06, Ch13, Ch14,
Ch21, Ch22, Ch29, Ch30} which are above the ex-
tensor carpi radialis and extensor digitorum commu-
nis muscles; and {Ch25, Ch26} which are above
the flexor digitorum sublimis and flexor carpi ulnaris
muscles.
In (Hargrove et al., 2007) they found the same re-
sults with an EMG pattern recognition task with simi-
lar gestures and a wrapper Variable Selection method.
The results in the present study are further confirma-
tion of the importance of these muscles and the fact
that the mRMR-FCO algorithm is an excellent tool
for Variable Selection.
4 FINAL TEST
The selected 10 features {AR1, AR2, WAMP, ZC,
BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing
46
Table 5: The number of times that a channel appears in the
FCO-Top-{50,100}.The median across subjects is given.
The selected channels are in bold.
FCO-Top-50
Ch01-08 0 0 1 1 3 2 1 1
Ch09-16 1 2 0 0 3 3 1 0
Ch17-24 1 1 0 0 4 5 2 0
Ch25-32 3 3 0 0 3 6 1 0
FCO-Top-100
Ch01-08 1 1 2 2 5 5 4 1
Ch09-16 1 2 0 1 4 8 2 1
Ch17-24 1 3 0 2 6 10 2 0
Ch25-32 4 4 1 1 6 9 2 2
NT, F5-F9} and the selected 10 channels {Ch05,
Ch06, Ch13, Ch14, Ch21, Ch22, Ch25, Ch26, Ch29,
Ch30} form 100 combinations (i.e. a compact-set
of 100 variables). To test the hypothesis that this
compact-set holds most of the key information, the
following procedure is applied to the data of subjects
S3, S4 and S5. Firstly, from the trials of the Train,
Validation and Test datasets the selected variables
from the compact-set are extracted, which means us-
ing only the selected channels and computing the se-
lected features from them. Secondly, an SVM classi-
fier is trained with the Train-dataset. Finally the Vali-
dation and Test datasets are passed to the classifier for
classification. The results are shown in Table 6. Note
that as subject S5 has no Validation-dataset (Table 1)
no classification rate is computed in that case.
On the one hand, with the classification rate of the
Validation-datasets we can make a comparison of the
performance between using the Top-100 variables of
the FCO ranking list or using the 100 variables of the
compact-set. As it is shown in Table 6, the differences
in performance are very low. For example, for user
S3, using the best 100 combinations of channels and
features of that specific user (according to FCO) we
only achieve a 4% extra classification rate than when
using the selected 100 combinations of the compact-
set. On the other hand, Table 7 shows that the FCO-
Top-100 list of user S3 demands 25 different sEMG
channels in the system, whereas the compact-set list
only demands 10 channels. Moreover, the compact-
set has the same channels and signal-features for ev-
ery user, unlike the FCO-Top-100 list, which has the
best ones depending on the user.
Finally, as the channels and features of the
compact-set are selected knowing that they ensure a
high classification rate of the Validation-dataset’s tri-
als (Figure 2), it is imperative to compute the classi-
fication rate for non-tested trials; the Test-dataset. As
it is shown in Table 6 the information extracted from
the features {AR1, AR2, WAMP, ZC, NT, F5-F9} and
the channels {Ch05, Ch06, Ch13, Ch14, Ch21, Ch22,
Ch25, Ch26, Ch29, Ch30} is enough for an accurate
classification in any user.
Table 6: Classification rate of the Validation and Test
datasets when the compact-set of variables is used.
Validation-dataset Test-dataset
Subject FCO-Top-100 Compact-set Compact-set
S3 88% 84% 77%
S4
83% 80% 84%
S5
- - 80%
Table 7: Comparison of the number of different channels
and signal features in the Top-100 variables of the FCO,
FCQ and Base ranking lists. The compact-set has also 100
variables as all the combinations among the selected chan-
nels and the selected features are included.
Subject S3 Channels Features
FCO-Top-100 25 26
FCQ-Top-100 21 18
Base-Top-100 24 16
Compact-set 10 10
Subject S4 Channels Features
FCO-Top-100 22 31
FCQ-Top-100 20 32
Base-Top-100 16 24
Compact-set 10 10
Subject S5 Channels Features
FCO-Top-100 25 24
FCQ-Top-100 26 22
Base-Top-100 28 15
Compact-set 10 10
5 CONCLUSIONS
In the present study 32 sEMG electrode locations and
41 signal-features are tested with the aim of reducing
the necessary number to obtain an accurate classifi-
cation rate in a 15 static-hand gesture classification
task. A novel implementation of the mRMR Vari-
able Selection algorithm is introduced to highlight the
most informative but least redundant combinations of
sEMG channels and signal-features.
The results show that the electrodes above the ex-
tensor carpi radialis, extensor digitorum, flexor dig-
itorum sublimis and flexor carpi ulnaris muscles are
the best ones for this kind of application. Moreover,
there is a group of signal-features that have high dis-
criminant power individuallyand that can extract non-
redundant information from each of these channels.
REDUCING THE NUMBER OF CHANNELS AND SIGNAL-FEATURES FOR AN ACCURATE CLASSIFICATION
IN AN EMG PATTERN RECOGNITION TASK
47
The signal-features are the first and second Auto-
Regressive coefficients, the Zero Crossing, the Num-
ber of Turns, the Wilson Amplitude and the frequency
histogram coefficients of the 225-500 Hz band (di-
vided in 5 segments).
We believe that the proposed methodology for
channel and feature selection makes a significant im-
provement with respect to the current ones. This
novel methodology may not only be appropriate for
the particular application presented in this paper (i.e.
channel and feature selection for hand gesture detec-
tion based on sEMG signals) but also in other case
scenarios, such as gene selection for classification of
phenotypes based on microarray data.
ACKNOWLEDGEMENTS
The authors would like to thank the department of Ed-
ucation of the Basque Government for their support.
REFERENCES
Bellman, R. (1961). Adaptive Control Processes: A Guided
Tour, volume 19 of A Rand Corporation Research
Study Series. Princeton University Press.
Boostani, R. and Moradi, M. H. (2003). Evaluation of the
forearm EMG signal features for the control of a pros-
thetic hand. Physiological measurement, 24(2).
Chang, C.-c. and Lin, C.-j. (2011). LIBSVM : a Library for
Support Vector Machines.
Cover, T. T. (1974). The Best Two Independent Measure-
ments are not the Two Best. Systems, Man and Cyber-
netics, IEEE Transactions, 1(JANUARY):116–117.
Ding, C. and Peng, H. (2005). Minimum redundancy fea-
ture selection from microarray gene expression data.
Journal of bioinformatics and computational biology.
Englehart, K., Hudgins, B., and Member, S. (2003). A Ro-
bust , Real-Time Control Scheme for Multifunction
Myoelectric Control. Engineering In Medicine And
Biology, 50(7):848–854.
Fuglsang-Frederiksen, A. and Pugdahl, K. (2010). Current
status on electrodiagnostic standards and guidelines
in neuromuscular disorders. Clinical neurophysiol-
ogy : official journal of the International Federation
of Clinical Neurophysiology, 122(3):440–455.
Hargrove, L. J., Englehart, K., and Hudgins, B. (2007).
A comparison of surface and intramuscular myoelec-
tric signal classification. IEEE transactions on bio-
medical engineering, 54(5):847–53.
Hudgins, B., Parker, P., and Scott, R. N. (1993). A new
strategy for multifunction myoelectric control. IEEE
transactions on bio-medical engineering, 40.
Hughes, G. (1968). On the mean accuracy of statistical pat-
tern recognizers. IEEE Transactions on Information
Theory, 14(1):55–63.
Jain, A. and Duin, R. (2000). Statistical pattern recognition:
a review. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 22(1):4–37.
Kang, W. J., Shiu, J. R., Cheng, C. K., Lai, J. S., Tsao,
H. W., and Kuo, T. S. (1995). The application of cep-
stral coefficients and maximum likelihood method in
EMG pattern recognition. IEEE transactions on bio-
medical engineering, 42(8):777–85.
Khokhar, Z. O., Xiao, Z. G., and Menon, C. (2010). Sur-
face EMG pattern recognition for real-time control of
a wrist exoskeleton. Biomedical engineering online.
Kuiken, T. a., Miller, L. a., Lipschutz, R. D., Lock, B. a.,
Stubblefield, K., Marasco, P. D., Zhou, P., and Du-
manian, G. a. (2007). Targeted reinnervation for en-
hanced prosthetic arm function in a woman with a
proximal amputation: a case study. Lancet.
Merletti, R. and Parker, P. (2004). Electromyography: Phys-
iology, engineering, and noninvasive applications. En-
gineering In Medicine And Biology.
Palmes, P., Ang, W. T., Widjaja, F., Tan, L. C. S., and Au,
W. L. (2010). Pattern mining of multichannel sEMG
for tremor classification. IEEE transactions on bio-
medical engineering, 57(12):2795–805.
Peng, H., Long, F., and Ding, C. (2005). Feature selec-
tion based on mutual information: criteria of max-
dependency, max-relevance, and min-redundancy.
IEEE transactions on pattern analysis and machine
intelligence, 27(8):1226–38.
Philipson, L, Larsson, P, G. (1988). The electromyographic
signal as a measure of muscular force: a compari-
son of detection and quantification techniques. Elec-
tromyography and clinical neurophysiology.
Tenore, F., Ramos, A., Fahmy, A., Acharya, S., Etienne-
Cummings, R., and Thakor, N. V. (2007). Towards the
control of individual fingers of a prosthetic hand using
surface EMG signals. Annual International Confer-
ence of the IEEE Engineering in Medicine and Biol-
ogy Society.
Tenore, F. V. G., Ramos, A., Fahmy, A., Acharya, S.,
Etienne-Cummings, R., and Thakor, N. V. (2009). De-
coding of individuated finger movements using sur-
face electromyography. IEEE transactions on bio-
medical engineering, 56(5):1427–34.
Tkach, D., Huang, H., and Kuiken, T. a. (2010). Study
of stability of time-domain features for electromyo-
graphic pattern recognition. Journal of neuroengi-
neering and rehabilitation, 7:21.
Zardoshti-Kermani, M., Wheeler, B., Badie, K., and
Hashemi, R. (1995). EMG feature evaluation for
movement control of upper extremity prostheses.
IEEE Transactions on Rehabilitation Engineering.
BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing
48