Medical Engineering & Physics: Jie Liu
Medical Engineering & Physics: Jie Liu
Medical Engineering & Physics: Jie Liu
Technical note
a r t i c l e
i n f o
Article history:
Received 27 August 2014
Accepted 13 September 2014
Keywords:
Myoelectric pattern recognition
Surface EMG
Feature selection
Feature projection
a b s t r a c t
This study investigates the effect of the feature dimensionality reduction strategies on the classication
of surface electromyography (EMG) signals toward developing a practical myoelectric control system.
Two dimensionality reduction strategies, feature selection and feature projection, were tested on both
EMG feature sets, respectively. A feature selection based myoelectric pattern recognition system was
introduced to select the features by eliminating the redundant features of EMG recordings instead of
directly choosing a subset of EMG channels. The Markov random eld (MRF) method and a forward
orthogonal search algorithm were employed to evaluate the contribution of each individual feature to
the classication, respectively. Our results from 15 healthy subjects indicate that, with a feature selection analysis, independent of the type of feature set, across all subjects high overall accuracies can be
achieved in classication of seven different forearm motions with a small number of top ranked original
EMG features obtained from the forearm muscles (average overall classication accuracy >95% with 12
selected EMG features). Compared to various feature dimensionality reduction techniques in myoelectric
pattern recognition, the proposed lter-based feature selection approach is independent of the type of
classication algorithms and features, which can effectively reduce the redundant information not only
across different channels, but also cross different features in the same channel. This may enable robust
EMG feature dimensionality reduction without needing to change ongoing, practical use of classication
algorithms, an important step toward clinical utility.
2014 IPEM. Published by Elsevier Ltd. All rights reserved.
1. Introduction
Robotic devices are considered as the leading interactive rehabilitation systems available. Robotic rehabilitation is an effective
platform for sensorimotor training in people with different neurological injuries such as hemiparetic stroke, cerebral palsy, multiple
sclerosis, Parkinsons disease or spinal cord injury [17]. In contrast, advances have been made to build lighter, stronger and more
versatile upper-limb powered robotic devices [813], relatively
little progress has been made on improving the intuitive control of robotic exoskeletons. A neural control interface is crucial
to providing accurate, intuitive control of upper limb exoskeletons. Among the potential biological signals for humanmachine
interaction (brain, nerve, and muscle signals), electromyography
(EMG), directly reect the human motion intention, may be the only
1717
in the measurement space are transformed into the dimensionreduced space utilizing specied transformation, such as principal
components analysis (PCA) [25] and uncorrelated linear discriminant analysis (ULDA) [32]. In contrast, feature selection attempts
to determine the best subset of the original feature set. Feature selection reduces the dimension of features, computation
load and improves classication performance. In general, feature
selection approaches can be grouped into two categories: lter
methods and wrapper methods. The lter methods select features from the original feature space with a given evaluation
criterion which is independent of the type of classier, such as distance measures which reect how well the classes separate from
each other. While, the wrapper methods choose a feature subset with high classication performance estimated by a specied
classier.
The objective of this study was to preliminarily assess the
effect of dimensionality reduction strategies on the surface EMG
classication performance. Two dimensionality reduction strategies, feature selection and feature projection, were preliminarily
tested on classical EMG feature sets, respectively. This study
focused on the exploration of various feature dimensionality
reduction techniques with an emphasis on feature selection
techniques.
2. Methods
In contrast to previous studies [17,2931], except for using feature projection for dimension reduction, feature selection was also
applied to reduce the complexity of the signal processing. Previous
studies have shown that classication accuracy is affected more by
the choice of feature set than by the choice of classier [27,29].
To solve the general feature subset selection problem in myoelectric control, two relatively new lter methods the Markov random
eld (MRF) based FisherMarkov selector [38] and the forward
orthogonal search algorithm, FOS-MOD [39] were examined in this
study to select the most appropriate original surface EMG features,
respectively. Because the lter methods estimate the classication
performance by some indirect assessments which reect how well
the classes separate from each other, making the approach more
exible due to the fact that the selected features are useful for
general classiers.
The mathematical details of the MRF and the FOS-MOD feature selection methods were described in [38,39]. In brief, the
MRF method aims to select the best subset of candidate features that maximizes the between-class distance while minimizing
the within-class distance in a higher dimensional kernel space.
The MRF method efciently selects the globally optimal subset
of features, which are the most useful in characterizing differences among the possible classes [38]. The FOS-MOD algorithm
[39] is a forward orthogonal search feature selection algorithm by
maximizing the overall dependency to nd signicant variables,
which also provides a rank list of selected features ordered according to the percentage contribution for representing the overall
structures.
The data used in this work are identical to the data in a previous
study [33]. Eight-channel EMG data were collected from forearm
muscles and the biceps muscle of 15 healthy subjects using AgAgCl electrodes with a reference electrode placed on the wrist. The
surface EMG signals were sampled at 3000 Hz per channel. Subjects
performed seven forearm motions: hand open, hand close, supination, pronation, wrist exion, wrist extension and rest. EMG data
used in this study were collected in ve trials nished by each subject. Each motion was repeated four times with duration of three
seconds within each data collection trial. The order of these motions
was randomized.
2.2. Feature extraction and pattern recognition
The EMG recordings for each distinct motion were composed
of four active segments corresponding to four repetitions of the
motion for each data collection trial. By visual inspection, active
segments were extracted from EMG data in between manually
located onset and offset times, which were the same for all the
eight channels. A sliding window scheme was used to facilitate
both utilization of limited data stream and continuity of decision output by a classier. Specically, for each active segment,
eight-channel EMG data were further divided into a series of sliding analysis windows (window length: 256 ms, overlapping step:
32 ms).
For each analysis window, a set of features was extracted to
characterize the EMG data for classication of the forearm motions.
The feature set was extracted on each of the eight EMG channels
and then concatenated into a feature vector. In order to examine
if the performance of the feature selection technique would not
depend on the feature set, two classical feature sets were investigated in this study including the time domain (TD) feature set [28],
which comprised of four time domain statistics including the mean
absolute value (MAV), the number of zero crossings (ZC), the waveform length (WL), and the number of slope sign changes (SSC); the
combination of autoregressive (AR) model (six-order as suggested
by Farina and Merletti [34]) coefcients and the root mean square
(RMS) of the signal as a feature set (AR + RMS) [35]. These feature
sets have been shown to be effective signal representations for EMG
pattern recognition with relatively low computational complexity
[1619,30,31,36].
Three classiersthe linear discriminant analysis (LDA) classier [16], the k-nearest neighbor (KNN) classier [37] and the
support vector machine (SVM) [18] were used respectively in the
current study.
2.3. Feature projection
100
100
90
90
Classification accuracy(%)
Classification accuracy(%)
1718
80
70
ULDA TD
PCA TD
ULDA AR+RMS
PCA AR+RMS
60
50
40
30
80
70
MRF TD
FOS-MOD TD
MRF AR+RMS
FOS-MOD AR+RMS
60
50
40
30
5
10
15
Number of feature dimensions
20
0
20
Fig. 1. The effect of feature set dimension upon the classication accuracy, when
using ULDA and PCA dimensionality reduction, respectively. The accuracy, averaged
across 15 subjects, is shown for the TD feature set and AR + RMS feature set. Results
are given for the LDA classier. Note that ULDA can yield six projected features for
seven classes at most in this study.
all the analysis windows in the testing datasets across all motion
patterns.
5
10
15
Number of selected features
20
Fig. 2. The change in average classication performance with the number of applied
original features in a feature set (TD and AR + RMS feature set) selected using MRF
and FOS-MOD feature selection, respectively.
Pre =
TP
(TP+FP)
(2)
A comparison was made between feature projection and selection on the dataset. Because ULDA is more efcient than PCA for
feature projection (Fig. 1) and the MRF method achieved similar
performance in term of feature selection compared to the FOS-MOD
method (Fig. 2), but the FOS-MOD algorithm only ranks features
that are most representative of the feature data. The MRF algorithm
was tested against ULDA. The comparison between the two methods reveals that the MRF method achieves better performance than
ULDA, as shown in Fig. 3. It is noted that ULDA has higher accuracies
when the dimension of feature subset is less than eight, because the
selected projected features are obtained by ULDA from all the original features. It cannot be argued that projected features by ULDA
are more concentrated with relevant information than the original
features obtained from the MRF selector.
Sen =
TP
(TP+FN)
(3)
(1)
3. Results
3.1. The effect of feature projection on performance
Fig. 1 shows the effect of the number of feature dimensions on
the classication performance averaged across all subjects, using
a combination of the AR + RMS or TD feature set and the LDA classier, where the feature dimensions were reduced via the ULDA
and PCA methods, respectively. The comparison between the two
methods indicates that ULDA is more efcient than PCA in reducing
feature dimensions. The projected features obtained by ULDA are
more concentrated with relevant information than the projected
features obtained via PCA.
3.2. The effect of feature selection on performance
The MRF and FOS-MOD feature selection algorithms were tested
on the AR + RMS and TD feature sets, respectively. Fig. 2 demonstrates the average classication accuracy as a function of the
number of selected original features. It was observed that as features are added, the classication accuracy increases, independent
of the type of feature set. However, the increase is not linear, which
quickly reaches a high value as the rst few features are added then
slowly approaches a maximum. The results indicated that it is feasible to greatly reduce the number of features while maintaining high
classication accuracy. The two methods achieved similar results.
100
90
Classification accuracy(%)
CA =
80
70
MRF TD
ULDA TD
MRF AR+RMS
ULDA AR+RMS
60
50
40
30
10
15
Number of features
20
1719
Fig. 4. Class-to-class confusion matrices derived from the EMG data after using the MRF criterion for feature selection. Class 1 (hand open), Class 2 (hand close), Class 3
(supination), Class 4 (pronation), Class 5 (wrist exion), Class 6 (wrist extension), Class 7 (rest). The precision (Pre), the sensitivity (Sen) for each class and the overall accuracy
for each class are also reported, respectively.
LDA
KNN
SVM
TD
AR + RMS
95.7 6.7
95.6 4.1
99.2 1.9
97.2 3.4
98.1 2.5
97.5 4.2
1720
[12] Kim H, Miller LM, Byl N, Abrams GM, Rosen J. Redundancy resolution of
the human arm and an upper limb exoskeleton. IEEE Trans Biomed Eng
2012;59:17709.
[13] Ren YP, Kang SH, Park HS, Wu YN, Zhang LQ. Developing a multi-joint upper
limb exoskeleton robot for diagnosis, therapy, and outcome evaluation in neurorehabilitation. IEEE Trans Neural Syst Rehabil Eng 2013;21:4909.
[14] Jiang N, Falla D, dAvella A, Graimann B, Farina D. Myoelectric control in neurorehabilitation. Crit Rev Biomed Eng 2010;38:38191.
[15] Scheme E, Englehart K. Electromyogram pattern recognition for control of powered upper-limb prostheses: state of the art and challenges for clinical use. J
Rehabil Res Dev 2011;48:64359.
[16] Englehart K, Hudgins B. A robust, real-time control scheme for multifunction
myoelectric control. IEEE Trans Biomed Eng 2003;50:84854.
[17] Zhou P, Lowery MM, Englehart KB, Huang H, Li G, Hargrove L, et al. Decoding
a new neural-machine interface for control of articial limbs. J Neurophysiol
2007;98:297482.
[18] Oskoei MA, Hu H. Support vector machine-based classication scheme
for myoelectric control applied to upper limb. IEEE Trans Biomed Eng
2008;55:195665.
[19] Hargrove LJ, Scheme EJ, Englehart KB, Hudgins BS. Multiple binary classications via linear discriminant analysis for improved controllability of a powered
prosthesis. IEEE Trans Neural Syst Rehabil Eng 2010;18:4957.
[20] Boostani R, Moradi MH. Evaluation of the forearm EMG signal features for the
control of a prosthetic hand. Physiol Meas 2003;24:30919.
[21] Delis AL, Carvalho JLA, da Rocha AF, Ferreira RU, Rodrigues SS, Borges GA. Estimation of the knee joint angle from surface electromyographic signals for active
control of leg prostheses. Physiol Meas 2009;30:93146.
[22] Nazarpour K, Sharafat AR, Firoozabadi SM. Application of higher order statistics to surface electromyogram signal classication. IEEE Trans Biomed Eng
2007;54:17629.
[23] Chan ADC, Englehart KB. Continuous myoelectric control for powered prostheses using hidden Markov models. IEEE Trans Biomed Eng 2005;52:1214.
[24] Englehart K, Hudgins B, Parker PA, Stevenson M. Classication of the myoelectric signal using time-frequency based representations. Med Eng Phys
1999;21:4318.
[25] Hargrove LJ, Li G, Englehart KB, Hudgins BS. Principal components analysis
preprocessing for improved classication accuracies in pattern-recognitionbased myoelectric control. IEEE Trans Biomed Eng 2009;56:140714.
[26] Huang H, Kuiken TA, Lipschutz RD. A strategy for identifying locomotion modes
using surface electromyography. IEEE Trans Biomed Eng 2009;56:6573.
[27] Huang Y, Englehart KB, Hudgins B, Chan ADC. A Gaussian mixture model based
classication scheme for myoelectric control of powered upper limb prostheses. IEEE Trans Biomed Eng 2005;52:180111.
[28] Hudgins B, Parker P, Scott RN. A new strategy for multifunction myoelectric
control. IEEE Trans Biomed Eng 1993;40:8294.
[29] Hargrove LJ, Englehart K, Hudgins B. A comparison of surface and intramuscular
myoelectric signal classication. IEEE Trans Biomed Eng 2007;54:84753.
[30] Liu J, Zhou P. A novel myoelectric pattern recognition strategy for hand function
restoration after incomplete cervical spinal cord injury. IEEE Trans Neural Syst
Rehabil Eng 2013;21:96103.
[31] Zhang X, Zhou P. High-density myoelectric pattern recognition toward
improved stroke rehabilitation. IEEE Trans Biomed Eng 2012;59:164957.
[32] Ye J, Li T, Xiong T, Janardan R. Using uncorrelated discriminant analysis for
tissue classication with gene expression data. IEEE/ACM Trans Comput Biol
Bioinform 2004;1:18190.
[33] Chan ADC, Green GC. Myoelectric control development toolbox. In: Proceedings
of the 30th conference of the Canadian medical & biological engineering society; 2007.
[34] Farina D, Merletti R. Comparison of algorithms for estimation of EMG
variables during voluntary isometric contractions. J Electromyogr Kinesiol
2000;10:33749.
[35] Graupe D, Salahi J, Kohn KH. Multifunctional prosthesis and orthosis control
via microcomputer identication of temporal pattern differences in single-site
myoelectric signals. J Biomed Eng 1982;4:1722.
[36] Lee SW, Wilson KM, Lock BA, Kamper DG. Subject-specic myoelectric pattern
classication of functional hand movements for stroke survivors. IEEE Trans
Neural Syst Rehabil Eng 2011;19:55866.
[37] Cover TM, Hart PE. Nearest neighbor pattern classication. IEEE Trans Inf Theory
1967;13:217.
[38] Cheng Q, Zhou H, Cheng J. The FisherMarkov selector: fast selecting maximally
separable feature subset for multiclass classication with applications to highdimensional data. IEEE Trans Pattern Anal Mach Intell 2011;33:121733.
[39] Wei HL, Billings SA. Feature subset selection and ranking for data dimensionality reduction. IEEE Trans Pattern Anal Mach Intell 2007;29:1626.
[40] Rojas-Martinez M, Mananas MA, Alonso JF, Merletti R. Identication of isometric contractions based on high density EMG maps. J Electromyogr Kinesiol
2013;23:3342.
[41] Haufe S, Meinecke F, Gorgen K, Dahne S, Haynes JD, Blankertz B, et al. On the
interpretation of weight vectors of linear models in multivariate neuroimaging.
Neuroimage 2014;87:96110.