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





j=1

The dimensionality reduction step (C) is recom-

mended since the higher the dimension V of the vec-

tor of variables





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

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

 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

is more powerful than variable

, with z

∈ Z, if the patterns from the different

classes can be classiﬁed easier with z

than with z

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





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

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

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

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,





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 =

∑

i=1

| (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

| (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 =

N − 1

∑

i=1

(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 =

N−1

∑

i=1

∆

(4)

Where ∆

= |x

− x

i+1

Mean Absolute Difference Value (MADV). The

mean value of the difference in amplitude between ad-

jacent samples.

MADV =

N − 1

N−1

∑

i=1

∆

(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

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 =

N−1

∑

i=1

φ(∆

i+1

) (6)

Where:

φ(∆

i+1

) =



1 x

· x

i+1

< 0, ∆

> 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 =

N−1

∑

i=1

ψ(∆

) (7)

Where ψ(∆

) = 1 if ∆

> 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

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.

∑

r=1

i−r

+ ε

(8)

Where the a

are the auto-regressive coefﬁcients and

the error term. The a

coefﬁcients are computed

minimizing the squared error between the original

signal and its approximation. Each coefﬁcient a

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

are

obtained from the Auto-Regressive coefﬁcients a

= −a

−

∑

r−1

n=1



1−



r−n

(9)

Where a

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

and the

Auto-Regressive one a

. 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

be the b-th frequency of the sig-

nal’s spectrum and P( f

) the power of f

. The mean

frequency is computed as follows:

Fmean =

∑

b=1

P( f

)

∑

b=1

P( f

)

(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





1312

j=1

and it belongs to a known class k (static-gesture). We

can measure the discriminant power of each variable

using the F-statistic as follows:

REDUCING THE NUMBER OF CHANNELS AND SIGNAL-FEATURES FOR AN ACCURATE CLASSIFICATION

IN AN EMG PATTERN RECOGNITION TASK

F(z

) =

∑

k=1

(k)

− z

)

/(K − 1)







∑

k=1

(k)

∑

t=1

(k)

− z

(k)

)







/(n

− K)

(11)

Where:

K : Number of possible classes.

: Number of samples of variable z

(k)

: Number of samples of z

within the

k-th class.

: Mean value of z

(k)

: Mean value of z

within the k-th class.

(k)

: Sample t-th of z

within the k-th class.

F-statistic only gives a measure of the classiﬁca-

tion power of each variable z

by itself. However,

the aim is to select a group S of m variables z

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

(S), D

|S|

∑

∈S

F(z

)

(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

(S), R

|S|

∑

∈S

|c(z

(13)

The Pearson correlation coefﬁcient c(z

) is

computed as:

c(z

) =

∑

t=1

− z

)(z

− z

)

∑

t=1

− z

)

∑

t=1

− z

)

(14)

where z

is the sample t-th of z

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

) (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

). 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

variable from

the set {Z − S

m−1

} (i.e. all the variables except those

already selected) that combines the following condi-

tions:

max

∈Z−S

m−1

F(z

) (16)

min

∈Z−S

m−1

m− 1

∑

∈S

m−1

|c(z

)| (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

variable sat-

isﬁes the following condition known as FCQ: F-test

Correlation Quotient (Ding and Peng, 2005):

max

∈Z−S

m−1

F(z

)

m−1

∑

∈S

m−1

|c(z

(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−S

m−1

F(z

) ·



1− max

∈S

m−1

|c(z



(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

for the set S if its F-value is

high and if the averaged correlation with previously

selected variables (set S

m−1

) is low. However, if there

is a variable in S

m−1

that has a high correlation with

it may go unnoticed. On the other hand, the FCO

(19) selects a variable z

whose individual F-value is

high, but whose maximum individual correlation with

the variables of S

m−1

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

ﬁcation rate that is obtained when we only maintain

in the vector





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

to a variable

that is top-ranked in the ranking list MET. If we only

maintain in





1312

j=1

the top-P variables of the rank-

ing list MET the new vector is denoted as



MET



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

FCQ

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

of the ranking list MET (i.e. now

each trial is represented by a vector of 5 variables,



MET



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



MET



p=1

of the Training-dataset.

4. The vectors



MET



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

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

P: Nº of top−ranked variables

Classification Rate [%]

(a) Subject S1

0 50 100 150 200

P: Nº of top−ranked variables

Classification Rate [%]

(b) Subject S2

0 50 100 150 200

P: Nº of top−ranked variables

Classification Rate [%]

0 50 100 150 200

P: Nº of top−ranked variables

Classification Rate [%]

(d) Subject S4

0 50 100 150 200

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





1312

j=1

is reduced to



MET



p=1

, where z

MET

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



MET



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



MET



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



MET



p=1

is between 5 and 200). There-

fore, we could infer that maintaining the whole vector





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

ject S3 the difference in performance between using



FCO



p=1



FCO



200

p=1

is insigniﬁcant, whereas if

we are using the ranking list Base the difference be-

tween



Base



p=1



Base



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



Base



p=1

FCQ

p=1

have to have a high amount of variables to approx-

imate the rate achieved with



FCO



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

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

1 0 0 2 1 1 6 1 2

3 2 4 4 3 6 7 6 12

WAMP

3 4 3 4 6 4 5 9 8

1 0 0 2 1 0 3 2 0

1 0 0 1 0 0 3 1 0

0 0 0 0 0 0 1 0 0

0 0 0 1 0 0 2 1 0

0 0 0 0 0 0 2 1 0

0 0 0 0 0 0 2 0 0

0 0 0 1 0 0 1 0 0

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

0 3 4 0 4 6 0 9 11

1 0 0 1 0 0 2 1 0

1 0 0 1 0 0 1 1 0

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

0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 1 0 0

0 0 0 0 0 0 2 0 0

0 1 0 1 1 0 3 2 1

1 0 0 1 1 0 3 3 1

0 1 0 1 2 0 2 3 2

2 1 0 3 2 0 6 2 2

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

4 5 5 6 7 8 8 8 10

30 4 6 7 8 10 11 11 12 14

5 7 9 13 16 18 19 20 26

70 5 8 11 16 19 22 25 27 34

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

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

2 2 2 2 1 1 1 3 2

2 2 4 1 1 1 1 2 1

WAMP

2 2 2 2 1 1 1 0 0

1 2 1 3 0 0 1 0 0 0

0 0 1 1 0 0 0 0 1 0

0 0 0 0 0 0 0 0 0 0

0 1 0 0 0 0 0 0 0 0

1 1 1 0 1 0 0 0 0 0

1 1 1 1 0 0 0 0 0 0

0 1 0 0 1 0 0 0 0 0

0 1 1 0 0 0 0 0 1 0

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

1 1 0 2 0 0 0 0 1 1

1 1 1 0 1 0 1 1 0 0

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

0 0 0 0 0 0 0 0 0 0

1 1 0 0 0 1 0 0 1 1

1 0 0 1 0 0 0 0 0 0

1 0 0 1 1 1 0 0 1 1

1 1 1 1 1 1 1 2 1

1 1 1 0 1 1 1 2 0

1 1 1 0 2 1 1 2 1

3 2 0 1 2 2 2 2 2

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

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%

83% 80% 84%

- - 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

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