IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 15, NO.
4, DECEMBER 2007 535
Real-Time Classification of Forearm
Electromyographic Signals Corresponding
to User-Selected Intentional Movements
for Multifunction Prosthesis Control
Kaveh Momen, Sridhar Krishnan, Senior Member, IEEE, and Tom Chau, Senior Member, IEEE
Abstract—Pattern recognition-based multifunction prosthesis
control strategies have largely been demonstrated with subsets
of typical able-bodied hand movements. These movements are
often unnatural to the amputee, necessitating significant user
training and do not maximally exploit the potential of residual
muscle activity. This paper presents a real-time electromyography
(EMG) classifier of user-selected intentional movements rather
than an imposed subset of standard movements. EMG signals
were recorded from the forearm extensor and flexor muscles of
seven able-bodied participants and one congenital amputee. Par-
ticipants freely selected and labeled their own muscle contractions
through a unique training protocol. Signals were parameterized
by the natural logarithm of root mean square values, calculated
within 0.2 s sliding and non overlapping windows. The feature
space was segmented using fuzzy C-means clustering. With only 2
min of training data from each user, the classifier discriminated
four different movements with an average accuracy of 92.7%
3.2%. This accuracy could be further increased with additional
training data and improved user proficiency that comes with
practice. The proposed method may facilitate the development
of dynamic upper extremity prosthesis control strategies using
arbitrary, user-preferred muscle contractions.
Index Terms—Adaptive, classification, clustering, electromyo-
graphy (EMG), fuzzy c-means, prosthesis, real-time, root mean
square (rms).
I. INTRODUCTION
YOELECTRIC prosthetic hands are controlled using
M electromyography (EMG) signals. These signals origi-
nate from the depolarization and repolarization of the muscle
membrane during voluntary contractions and can be measured
at the skin surface using either dry or wet-type electrodes. The
EMG control signal can be derived from a single site [1]–[3] or
Manuscript received February 23, 2007; revised July 29, 2007; accepted Au-
gust 22, 2007. This work was supported in part by the Canada Research Chairs
program, in part by the REMAD Foundation, in part by Ryerson University, in
part by the Natural Sciences and Engineering Research Council of Canada, in
part by Bloorview Children’s Hospital Foundation, and in part by the Myoelec-
tric service at Bloorview Kids Rehab.
K. Momen is with the Toronto Rehabilitation Institution, Toronto, M5G 2A2
ON, Canada.
S. Krishnan is with Ryerson University, Toronto, M5B 2K3 ON, Canada.
T. Chau is with the Bloorview Kids Rehab, Toronto, M4G 1R8 ON, Canada
(e-mail: tom.chau@utoronto.ca).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TNSRE.2007.908376
1534-4320/$25.00 © 2007 IEEE
from multiple sites on the forearm to control a powered pros-
thesis. Past studies have employed two [4]–[7], three [8]–[10],
four [6], [11]–[16], and up to eight [17], [18] recording sites
with varying levels of success.
To control multiple functions, it is necessary to map EMG sig-
nals corresponding to different muscle contractions to a variety
of prosthetic functions. This mapping has been demonstrated
usually through offline pattern recognition [1], [4], [7], [12],
[17]–[19] and in a few instances, with real-time classifiers [5],
[6], [11], [20]. For the purposes of this paper, we focus only on
the latter. For offline approaches, the interested reader is referred
to EMG prosthesis control schemes based on linear classifiers
[2], [6], [11], [21], artificial neural networks (ANN) [1], [3]–[5],
[8]–[10], [12], [14], [15], fuzzy and neuro-fuzzy classifiers [1],
[22], [23], and K-nearest neighbour classifiers [7], [9].
A. Real-Time EMG Classifiers
A small number of real-time EMG classifiers have appeared
in the literature. Focussing on the transient EMG, Hudgins et
al. [4], used simple time domain features and a neural network
classifier to achieve 90% accuracy in a four-class problem. The
main drawback was that contractions had to be exclusively ini-
tiated from rest. This restriction prohibited the user from intu-
itively switching between classes and impeded the coordination
of complex tasks involving multiple degrees-of-freedom [12].
Moving away from the transient signal, Englehart et al. devel-
oped a “continuous classifier” using wavelet analysis [6] and
subsequently with time domain features [11] to process four
channels of steady-state EMG signal. They succeeded in classi-
fying four classes of motion with an average of 0.5% error, and
six classes with 2% error.
In a subsequent work, Englehart et al. [12] further optimized
the continuous classifier. They processed four channels of EMG
signal, with the task of discriminating six classes of limb move-
ment. They obtained an impressive 93.25% accuracy using a
time-domain feature set, with a multilayer perceptron (MLP)
classifier. The classes of motion were constrained to six stan-
dard able-bodied movements.
In a separate study, Nishikawa et al. [5] proposed a novel
online learning method for discriminating six different prede-
termined hand movements on the basis of two EMG signals
harnessed from the palmar and dorsal sides of the wrist. Using
a variant of the adaptive heuristic critic (AHC) algorithm,
536 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 15, NO. 4, DECEMBER 2007

Nishikwa et al. discriminated six forearm movements with

89.9% accuracy.
More recently, Ajiboye and Weir [20] used heuristic fuzzy
logic to classify up to four forearm movements (i.e., wrist exten-
sion and flexion, ulnar deviation and finger flexion). A trained
prosthetist selected four recording sites for able-bodied partic-
ipants and three sites for amputees. One channel of EMG was
assigned to one expected movement. A simple amplitude-driven
inference rule base, constructed using fuzzy C-means (FCMs),
achieved real-time classification rates from 94% to 99%, during
both steady state motions and transitions between movements.
However, amputees did not participate in real-time testing and
the possibility that they might have been more proficient at gen-
erating other “atypical” movements was not explored.

B. Outstanding Challenges
Based on the above appraisal of literature, some outstanding
challenges in real-time EMG classification can be identified.
1) Dependence on thresholds. EMG classifiers have typi-
cally required the determination of some threshold. With
time domain features [4], [6], [11], [12], [19], thresholds
have been required in the calculation of zero-crossings
and amplitude inflexion points [4]. Thresholds have also
been defined in the estimation of contraction onset [1],
[4] and termination [5]. A very low threshold augments
the system’s susceptibility to noise, while an excessively
large threshold reduces responsiveness to intentional
muscle activity. As muscle and skin conditions change,
Fig. 1. Protocol for system and user training to user-selected movements.
for example, due to perspiration and fatigue, thresholds
need to be adjusted accordingly, a nontrivial real-time
task. However, one could circumvent the onset threshold
by aptly defining a rest class within a pattern recognition the discrimination rate of many EMG classifiers would
framework, as demonstrated in [24]. Alternatively, the drop to about 60% over the course of actual usage [5]. It
effectiveness of a prescribed threshold may be improved is not clear whether or not previously reported classifiers
by amplitude normalization, where the normalization are conducive to real-time adaptation to daily variations
constants are derived from the amplitude statistics of a in EMG signals.
sufficiently-sized training set.
2) Adherence to subsets of admissible movements. Generally C. Overview of the Proposed Method
for prosthesis control, participants must adhere to a subset We propose a new approach to map EMG signals to different
of typical contractions associated with normative hand functions in real-time, whereby the participants freely select
movements of the able-bodied limb (e.g., wrist flexion/ex- and label their own movements. Unlike previous research
tension, finger flexion/extension, pronation/supination) [4]–[6], [11], [12], [19], [25], there are no preset contraction
[3], [5], [6], [11], [12], [17]–[19], [22], [25]. In part, this lengths, amplitude thresholds, or movement sequences for
may be due to the prevalent application of supervised training the system. The user freely performs forearm move-
classifiers, which require accurately labeled movements ments of choice while the machine automatically determines
for correct partitioning of the feature space. Obtaining the most discernible and repeatable muscle signals. Based
such labeled movements necessitates a strict experimental on this subset of usable signals, the user can then choose the
protocol where users can not freely generate contractions preferred assignment of muscle activity to functional output. In
of choice, but must adhere to a predetermined subset of this way, the user’s unique abilities are accommodated and user
normative movements. However, these normative move- training might be minimized. Since, the number and types of
ments may be meaningless and unnatural to the congenital movements are not known a priori, we adopt an unsupervised
amputee, who may be more proficient at generating other approach to classifier construction.
individual-specific contractions. Hence, systems tested
successfully with able-bodied participants may not be
II. PROTOCOL FOR SYSTEM AND USER TRAINING TO
suitable for amputees, as reported in [20], potentially
USER-SELECTED MOVEMENTS
demanding significant user training.
3) Lack of real-time adaptation. It has been argued that The proposed method is outlined in Fig. 1 and each block is
without the capability to dynamically update over time, discussed below.
MOMEN et al.: REAL-TIME CLASSIFICATION OF FOREARM ELECTROMYOGRAPHIC SIGNALS
Fig. 2. Example of electrode placement on the forearm flexors of the amputee
participant.
Fig. 3. Extensor versus flexor feature space. Each point is calculated from 200
ms of data. (a) Feature space based on rms values. (b) Features space based on
natural log of rms values.
A. Acquisition of Self-Selected Contractions
As most below-elbow myoelectric prosthesis users control
their prosthetic appliances via the forearm flexors and extensors,
the raw EMG signals were recorded from these two sites, using
537
pregelled Ag/AgCl GRASS F-E8SD disposable electrodes. At
each recording site, the electrodes were arranged in a differential
configuration over the muscle bulk to harness the highest ampli-
tude signals [26], as illustrated in Fig. 2. The electrode locations
were photographed for future reference. Wires emanating from
the electrodes were secured to the participant’s forearm via tape.
Likewise, wires running along the table surface to the acquisi-
tion equipment were also secured via tape.
The raw signals were amplified (GRASS Telefactor 15A54)
5000 times, bandpass filtered in hardware with a passband of
30–1000 Hz and then sampled at 2 kHz. This bandpass range en-
compasses the useful EMG bandwidth previously reported [26],
[27]. The low cutoff frequency of the filter was chosen to avoid
motion artifact [26]. Additionally, a 60-Hz notch filter was used
to eliminate line noise. The signals were acquired via a personal
computer (P4- 2.8 GHz, 1 GB RAM) equipped with a National
Instruments, 16 bit data acquisition board, PCI-6014. The nec-
essary software for data acquisition was written in MATLAB
7.1.
To establish a well-defined and repeatable resting position
for each individual, each participant was asked to sit on a chair
and to rest his or her arm in a comfortable position. There was
no visual feedback provided during training. Rather, partici-
pants were instructed to attend to their forearm movements.
When prompted by the computer, the participants produced as
many different forearm movements as they could naturally and
repeatedly create. Movements were generated at the particant’s
own pace without any further prompting, at any desired contrac-
tion level and without reference to a corresponding prosthetic
function. Participants were not specifically instructed to avoid
forearm motion. The bandpass filter mentioned above along
with the securing of wires to the participant and to the table
mitigated contamination due to motion artifacts. Note that
movements did not have to be punctuated with rest states nor
were any specific motions or movement sequences imposed. A
2 min, continous stream of data was collected in this manner
for each channel.
B. System Training
Training of the system involved the partitioning of the feature
space to represent different classes of separable motions.
1) Feature Extraction: The root mean square (rms) value of
each channel was calculated to create a 2-D feature vector. It
has been argued that the response time of the control system
should not introduce a perceivable delay, generally regarded to
be roughly 300 ms [6], [11]. With this general prescription in
mind, we calculated the rms features of each channel within
a 200 ms window in real-time. Additionally, to spread the
concentrated data points while condensing the highly scattered
points, we invoked a nonlinear transformation, in this case the
natural logarithm. The log-transformed feature space, shown
in Fig. 3(b), demonstrates a more uniform scattering of points
compared to the untransformed rms features of an able-bodied
participant shown in Fig. 3(a). Additionally, the fea-
ture is computationally simple and by the monotonicity of the
logarithmic transformation, nonlinearly related to contraction
force [28]. The feature extraction was peformed in
real-time throughout the protocol.
538 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 15, NO. 4, DECEMBER 2007
Fig. 4. Cluster labeling. (a)–(f) Movements chosen by the participant. (g) Corresponding bar graph labels selected by the participant. (h) Cluster labels superim-
posed on the clustered feature space. (a) Rest. (b) Casi rest. (c) Supinated punio. (d) Mano estirada adajo. (e) Puino arriba. (f) Mano estirada right.
2) Clustering: The fuzzy C-means clustering algorithm [29]
was then used to segment the feature space into regions,
where the number of movements was specified by the user.
Several different cluster validity algorithms, including the Xie-
Beni index [30] and silhouettes [31], were initially used to deter-
mine the target number of clusters . However, preliminary ex-
periments indicated that these methods yielded cluster numbers
that were too often incongruent with the user’s perception of
the number of achievable movements. Therefore, we elected to
use the participant-specified number of movements as the target
cluster number. In the next section, this choice of cluster number
is validated by the user. The fuzzy C-means procedure mini-
mized the generalized sum-of-squares criterion function
(1)
where is the number of points (feature vectors) in the training
sample, is the fuzziness index, , is the 2-D
feature vector, is the degree of membership of observation
in the cluster , is the center of the th cluster, and is the
Euclidean norm measuring the similarity between an empirical
data point and a cluster center [29]. Pal and Bezdek [32] showed
the best choice for for a large class of problems lies in the in-
terval [1.5, 2.5]. The midpoint, , has often been the pre-
ferred choice for many applications. Based on preliminary em-
pirical investigations, was fixed at 1.8 in this study. Iterative
optimization of the objective function (1), was achieved through
successive updates of the membership and the cluster cen-
tres as prescribed by [29].
C. Exploration and Validation
After system training, the user was given 2 min to explore the
relationship between the discovered clusters and forearm move-
ments. To facilitate this exploration, the system classified new
movements into one of the learned clusters. First, the
features were extracted from the corresponding EMG signals
in 200 ms intervals. Subsequently, the membership of a feature
vector , was estimated by [29]
(2)
where as before is the fuzziness index, is the number of
clusters and is the center of the th cluster. Each vector was
labeled according to the cluster with the highest membership.
For visual feedback during the exploration stage, the different
clusters were represented as vertical bars. Once a movement was
classified as belonging to a certain cluster, the corresponding
vertical bar would be highlighted. If the user judged the system
to be unresponsive to his or her movements after 2 min of explo-
ration, the system would be retrained. As shown in Fig. 1, this
would mean reiterating data collection, system training and ex-
ploration until the user deemed the system responsive. Respon-
siveness was judged subjectively, but was generally interpreted
by users as the provision of correct classifications within a tol-
erable delay.
D. Cluster Labeling and User Training
Once satisfied with the system’s classification of movements,
the participants were asked to label the vertical bars, which cor-
responded to their forearm movements. For this phase of the
protocol, the participants generated each of their selected move-
ments in turn and observed which vertical bar was highlighted.
The bars were labeled in any language or with any words which
would help participants to recall the movements. An example
of six different movements and their corresponding labeled ver-
tical bars are illustrated in Fig. 4. This labeling step was nec-
essary for subsequent system testing. After all the movements
were named, the participants were given one minute to become
accustomed at generating the labeled movements.
MOMEN et al.: REAL-TIME CLASSIFICATION OF FOREARM ELECTROMYOGRAPHIC SIGNALS
TABLE I
AVERAGE CLASSIFICATION ACCURACIES
Subject 2 repeated the test due to misunderstanding of the protocol. Results from his second trial are presented.
Subject 8 (amputee) was asked to participate in the study twice to investigate the possible improvement from the first to the second session.
III. EMPIRICAL EVALUATION
The protocol was approved by the Research Ethics Board of
Bloorview Kids Rehab (Toronto, ON, Canada). All participants
provided informed written consent. We recruited seven able-
bodied adults and one below-elbow adult congenital amputee.
Able-bodied participants consisted of a convenience sample of
adults aged 20–30 years with no upper limb deficiencies. A con-
genital amputee without previous experience with myoelectric
control was recruited from the general population. It was partic-
ularly desirable to choose an amputee who was not a powered
prosthesis user since he or she would most likely be unfamiliar
with forearm muscle control. It is generally recognized that re-
liable EMG control necessitates user training (see, for example,
[33]). Therefore, the lack of prior training in our selected par-
ticipant might present a nontrivial challenge to the automatic
separation of EMG signals.
To estimate the real-time classification error, the user was re-
peatedly cued by the computer to perform the labeled move-
ments in random sequence. The computer randomly selected
one of the movements and highlighted the corresponding ver-
tical bar in yellow. The user was given 3 s from the time of
the cue to generate the corresponding movement. Within the
random sequence of cues, each movement was requested 10
times. For the purposes of system evaluation, a movement by
the participant would be considered intentional if it generated
five consecutive feature vectors with the same class label. This
condition helped to filter out abrupt spurious movements.
The performance index for the movement was given
by
(3)
where is the number of times movement was correctly clas-
sified, is the total number of different movements, and
is the number of repetitions of each movement. In the present
study, . The accuracy for each participant was simply
calculated by averaging the performance indices over all move-
ments, i.e., . data collection and 1 min of formal user training.
539
TABLE II
PERFORMANCE RESULTS FROM PARTICIPANT 8 (THE AMPUTEE)-TRIAL #2
IV. RESULTS
Table I summarizes the average classification accuracies and
the number of discernible movements from all the participants in
this study. Note that the quantity of movements discernible with
greater than 80% accuracy (fourth column) is either slightly
less than or equal to the number specified by the user (second
column). We refer to this subset of movements as the discernible
subset. This suggests that users were reasonably good at esti-
mating the number of distinct movements they created. Partic-
ipants 1, 5, and 8 grossly overestimated the number of move-
ments, i.e., by three or more. In these instances, we note that
while the raw accuracies are low, the accuracies of the corre-
sponding discernible subset of movements are much higher.
The individual movement accuracies for the congenital
amputee and an able-bodied participant are exemplified in
Tables II and III, respectively. Each row reports the accuracy
of one movement. The left column is the user-assigned labels,
based on his or her own perception of the movement. Since
participants trained the system based on their self-selected
intentional movements, the number, naming and nature of
movements differed among participants. In this example, it
can be seen that the able-bodied participant created 6 different
movements, discernible by the system with 96.7% 4.7%
accuracy. The amputee participant, who had never used a pow-
ered prosthetic hand, created four movements, recognizable by
the system with 87.5% 13% accuracy, after only 2 min of
540 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 15, NO. 4, DECEMBER 2007
TABLE III
PERFORMANCE RESULTS FROM PARTICIPANT 6
V. DISCUSSION
A. Classification of Multiple Movements
From the second and third columns of Table I, it can be seen
that all participants except Participant #5 could produce between
four and six movements that were discernible by machine with
an average accuracy of 75% or better. Although Participant #5
perceived nine different movements during the initial data col-
lection, he failed to reproduce the same type of movements later
in the testing session. Participant #5 later revealed that a subset
of the nine movements produced during the initial data collec-
tion did not feel natural. Therefore, during the testing session,
he could not remember how to generate those movements. The
algorithm suggested that he could perform only four repeatable
self-selected movements with more than 92% 4.3% accuracy.
The important point here is that multimovement discrimination
was facilitated with all participants and that the system was able
to recommend the most discernible movements to the user.
Since participants selected and labeled their own movements,
it was expected that they could reliably recreate the movements
as required. A few participants however, expressed nervousness
in responding to the computer prompt. When asked by the com-
puter to produce a movement during the test session, sometimes
users reported that they performed the wrong motion. These user
errors would diminish the accuracy of the system.
The current one second delay (0.2 s five consecutive
windows of uniform classification) for classifying a movement
was imposed for testing purposes only. Preliminary tests sug-
gest that it is possible to decrease the 200-ms feature extractor
window down to 40 ms without compromising classification
performance.
B. Focus on Natural Movements
The training did not impose any restriction on the partici-
pant’s choice of movements. Indeed, the system learned from
the participant’s self-selected, natural movements. Unlike pre-
vious studies [12], there was no need to train the system with
only the transient or steady-state portions of the signal. Both
signal components were captured by the cluster representation.
All the participants in this study encountered the system for
the first time and trained the system for only a mere 2 min.
It is expected that the accuracy of the system would increase
further as the user becomes more proficient at generating indi-
vidual movements. The surprisingly high accuracy for the lone
amputee participant in this study (Participant #8-Trial 2) sug-
gests that the system may be trainable with self-selected in-
tentional movements from individuals with unique musculature
and aytpical soft tissue configurations. This result emphasizes
the importance of accomodating movements natural to an am-
putee (see movement labels in Table II), which may be very
different from those of typical able-bodied individuals. Inci-
dentally, the amputee participant expressed extreme satisfac-
tion with the system responsiveness. Further studies with larger
sample sizes are required to validate these observations.
C. Prosthesis Control
The proposed algorithm could potentially be used to control
current powered prosthetic hands, which usually only open and
close. The additional movement classes can be mapped to give
the prosthesis enhanced functionality. For example, tight fist and
soft fist movements expressed by Participant #6, might control
the rate of opening or closing the powered prosthetic hand, sim-
ilar to functionality available in conventional myoelectric pros-
theses.
Previous studies argued that a participant’s EMG corre-
sponding to a specific movement changes over time [5]. Using
well-established dynamic fuzzy clustering principles [34], it
is conceivable that the proposed algorithm might be rendered
adaptive. However, this remains to be demonstrated and valida-
tion of the adaptation process, i.e., that there is indeed improved
classification due to dynamic updating, would be a nontrivial
issue.
VI. LIMITATIONS
As the system uses wet-type electrodes and performs amplifi-
cation distally, the system is susceptible to motion artifact. Such
artifacts can be mitigated by embedding both the amplifier and
the filters inside the electrode at the skin surface, as in conven-
tional Otto Bock electrodes.
The system appeared to be sensitive to variations in the
resting state. For example, resting one’s forearm with the
hand closed is interpreted by the system as a different state than
resting with the hand open. From the participant’s point of view,
both motions represent the resting state. This discrepancy in
interpretations inflated within cluster variability and accounted
for errors relating to the classification of “Rest” signals. Future
studies may investigate the possibility of employing adaptive
clustering [34] to address this limitation through the automatic
merging of multiple rest classes.
Participants 5, 6, and 7 judged the system to be nonrespon-
sive during the exploration and validation stage. This may have
been the result of the user generating movements which he or
she could not easily replicate or the fuzzy clustering algorithm
settling into a suboptimal local minimum. Subsequently, these
participants retrained the system a second time. While this re-
training extends the overall classifier training time, it is some-
what comforting that only a maximum of one retraining was
required in the present experiments.
VII. CONCLUSION
We have described a new approach to real-time classification
of forearm muscle signals corresponding to self-selected inten-
tional movements. Unlike previous work, our method admits
MOMEN et al.: REAL-TIME CLASSIFICATION OF FOREARM ELECTROMYOGRAPHIC SIGNALS
any muscle contractions preferred by the user and can classify
contractions continuously, without the need of a rest period be-
tween successive contractions. The method exploits a low di-
mensional feature space comprised of the natural log of rms
signal values from two sites and deploys fuzzy C-means clus-
tering to partition the feature space. Using a maximum member-
ship decision criterion, the classifier recognized four and five
different movements with average accuracies of 92.7 3.2%
and 79.9 16.8%, respectively, based on minimal user training.
ACKNOWLEDGMENT
The authors would like to thank the research participants who
eagerly gave of their time
REFERENCES
[1] F. H. Y. Chan, Y.-S. Yang, F. K. Lam, Y.-T. Zhang, and P. A. Parker,
“Fuzzy EMG classification for prosthesis control,” IEEE Trans. Re-
habil. Eng., vol. 8, no. 3, pp. 305–311, Sep. 2000.
[2] S. M. ElBasiouny, A. M. El-Bialy, M. F. Taher, A. H. Kandil, and M.
E. Rasmy, “A myoelectric prosthesis controller,” in Proc. IEEE 29th
Annu. Bioeng. Conf., Mar. 2003, pp. 140–141.
[3] B. Karlik, M. O. Tokhi, and M. Alci, “A fuzzy clustering neural
network architecture for multifunction upper-limb prosthesis,” IEEE
Trans. Biomed. Eng., vol. 50, no. 11, pp. 1255–1261, Nov. 2003.
[4] B. Hudgins, P. Parker, and R. Scott, “A new strategy for multifunction
myoelectric control,” IEEE Trans. Biomed. Eng., vol. 40, no. 1, pp.
82–94, Jan. 1993.
[5] D. Nishikawa, W. Yu, H. Yokoi, and Y. Kakazu, “On-line learning
method for EMG prosthetic hand control,” Electron. Commun. Japan,
vol. 84, no. 10, pt. 3, pp. 1510–1519, Nov. 2001.
[6] K. Englehart, B. Hudgins, and P. A. Parker, “A wavelet-based con-
tinuous classification scheme for multifunction myoelectric control,”
IEEE Trans. Biomed. Eng., vol. 48, no. 3, pp. 302–311, Mar. 2001.
[7] D. Peleg, E. Braiman, E. Yom-Tov, and G. F. Inbar, “Classification
of finger activation for use in a robotic prosthesis arm,” IEEE Trans.
Neural Syst. Rehabil. Eng., vol. 10, no. 4, pp. 290–293, Dec. 2002.
[8] N. Ma, D. K. Kumar, and N. Pah, “Classification of hand direction
using multi-channel electromyography by neural network,” in The 7th
Australian New Zealand 2001 Intelligent Inf. Syst. Conf., Perth, WA,
Nov. 2001, pp. 405–410.
[9] H.-P. Huang, Y.-H. Liu, and C.-S. Wong, “Automatic EMG feature
evaluation for controlling a prosthetic hand using supervised feature
mining method: An intelligent approach,” in Proc. 2003 IEEE Int. Conf.
Robotics Autom., Taipei, Taiwan, Sep. 2003, vol. 1, pp. 220–225.
[10] H.-P. Huang, Y.-H. Liu, L.-W. Liu, and C.-S. Wong, “EMG classi-
fication for prehensile postures using cascaded architecture of neural
networks with self-organizing maps,” in Proc. 2003 IEEE Int. Conf.
Robotics Autom., Taipei, Taiwan, Sep. 2003, vol. 1, pp. 1497–1502.
[11] K. Englehart and B. Hudgins, “A robust, real-time control scheme for
multifunction myoelectric control,” IEEE Trans. Biomed. Eng., vol. 50,
no. 7, pp. 848–854, Jul. 2003.
[12] K. Englehart, B. Hudgins, and A. D. C. Chan, “Continuous multifunc-
tion myoelectric control using pattern recognition,” Technol. Disability,
vol. 15, pp. 95–103, 2003.
[13] M. Vuskovic and S. Du, “Classification of prehensile EMG patterns
with simplified fuzzy ARTMAP networks,” in Proc. 2002 Int. Joint
Conf. Neural Networks, IJCNN ’02, Honolulu, HI, May 2002, vol. 3,
pp. 2539–2544.
[14] Y. Yazama, M. Fukumi, Y. Mitsukura, and N. Akamatsu, “Feature
analysis for the EMG signals based on the class distance,” in Proc.
2003 IEEE Int. Symp. Computational Intell. Robotics Autom., Kobe,
Japan, Jul. 2003, vol. 2, pp. 860–863.
[15] Y. Matsumura, Y. Mitsukura, M. Fukumi, and N. Akamatsu, “Recog-
nition of EMG signal patterns by neural networks,” in Proc. 9th Int.
Conf. Neural Inf. Process., Nov. 2002, vol. 2, pp. 750–754.
541
[16] M. Santa-Cruz, R. Riso, and F. Sepulveda, “Optimal selection of time
series coefficients for wrist myoelectric control based on intramuscular
recordings,” in Proc. 23rd Annu. Int. Conf. IEEE Eng. Medicine Biol.
Soc., Oct. 2001, vol. 2, pp. 1384–1387.
[17] R. Boostani and M. H. Moradi, “Evaluation of the forearm EMG signal
features for the control of a prosthetic hand,” Physiol. Meas., vol. 24,
no. 2, pp. 309–319, May 2003.
[18] F. Sebelius, L. Eriksson, H. Holmberg, A. Levinsson, G. Lundborg, N.
Danielsen, J. Schouenborg, C. Balkenius, T. Laurell, and L. Montelius,
“Classification of motor commands using a modified self-organising
feature map,” Med. Eng. Phys., vol. 27, pp. 403–413, 2005.
[19] Y. Huang, K. B. Englehart, B. Hudgins, and A. D. C. Chan, “A gaussian
mixture model based classification scheme for myoelectric control of
powered upper limb prostheses,” IEEE Trans. Biomed. Eng., vol. 52,
no. 11, pp. 1801–1811, Nov. 2005.
[20] A. B. Ajiboye and R. F Weir, “A heuristic fuzzy logic approach to EMG
pattern recognition for multifunctional prosthesis control,” IEEE Trans.
Neural Syst. Rehabil. Eng., vol. 13, no. 3, pp. 280–291, Sep. 2005.
[21] K. Englehart, B. Hudgins, P. Parker, and M. Stevenson, “Classification
of the myoelectric signal using time-frequency based representations,”
Med. Eng. Phys., vol. 21, no. 6–7, pp. 431–438, Jul. 1999.
[22] X. Zhang, Y. Yang, X. Xu, and M. Zhang, “Wavelet based neuro-fuzzy
classification for EMG control,” in 2002 Int. Conf. Commun., Circuits
Syst. West Sino Exposition Proc., Chengdu, China, Jul. 2002, vol. 2, pp.
1087–1089.
[23] S. E. Hussein and M. H. Granat, “Intention detection using a neuro-
fuzzy EMG classifier,” IEEE Eng. Med. Biol. Mag., vol. 21, no. 6, pp.
123–129, Nov.-Dec. 2002.
[24] L. Hargrove, K. Englehart, and B. Hudgins, “A comparison of sur-
face and intramuscular myoelectric signal classification,” IEEE Trans.
Biomed. Eng., vol. 54, no. 5, pp. 847–853, May 2007.
[25] A. D. Chan and K. Englehart, “Continuous classification of myoelec-
tric signals for powered prostheses using gaussian mixture models,” in
Proc. 25th Annu. Int. Conf. IEEE Eng. Medicine Biol. Soc., Cancun,
Mexico, Sep. 2003, vol. 3, pp. 2841–2844.
[26] C. J. DeLuca, “The use of surface electromyography in biomechanics,”
J. Appl. Biomech., vol. 13, pp. 135–163, 1997.
[27] J. V. Basmajian and C. J. Deluca, Muscles Alive, Their Functions Re-
vealed by Electromyography, 5th ed. Baltimore, MD: Williams &
Wilkins, 1985.
[28] J. Weir, L. Wagner, and T. Housh, “Linearity and reliability of the in-
tegrated EMG versus torque relationship for the forearm flexors and
leg extensors,” Amer. J. Phys. Medicine Rehabil., vol. 71, no. 5, pp.
283–287, 1992.
[29] J. C. Bezdek, Pattern Recognition with Fuzzy Objective Function Al-
gorithms. New York: Plenum, 1981.
[30] X. L. Xie and G. Beni, “A validity measure for fuzzy clustering,” IEEE
Trans. Pattern Anal. Mach. Intell., vol. 3, no. 8, pp. 841–846, Aug.
1991.
[31] P. J. Rousseeuw, “Silhouettes: A graphical aid to the interpretation and
validation of cluster analysis,” J. Computational Appl. Math., vol. 20,
pp. 53–65, 1987.
[32] N. R. Pal and J. C. Bezdek, “On cluster validity for the fuzzy c-means
model,” IEEE Trans. Fuzzy Syst., vol. 3, no. 3, pp. 370–379, Aug. 1995.
[33] J. Pons, R. Cere, E. Rocon, S. Levin, I. Markovitz, B. Saro, D. Rey-
naerts, W. Moorleghem, and L. Bueno, “Virtual reality training and
EMG control of the MANUS hand prosthesis,” Robotica, vol. 23, no.
3, pp. 311–317, 2005.
[34] L. Angstenberger, Dynamic Fuzzy Pattern Recognition. Boston, MA:
Kluwer Academic, 2001.
Kaveh Momen received the B.A.Sc. degree in electronics engineering from
Azad University, Tehran, Iran, in 1995, and the M.A.Sc. degree in electrical en-
gineering and computer science from Ryerson University, Toronto, ON, Canada,
in 2005. He is currently working toward the Ph.D. degree in biomedical engi-
neering at University of Toronto, Toronto, ON, Canada.
His research interests include biomedical signal processing, pattern recogni-
tion, and stochastic processes.
542 IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, VOL. 15, NO. 4, DECEMBER 2007
Sridhar Krishnan received the B.E. degree in elec-
tronics and communication engineering from Anna
University, Madras, India, in 1993, and the M.Sc. and
Ph.D. degrees in electrical and computer engineering
from the University of Calgary, Calgary, AB, Canada,
in 1996 and 1999, respectively.
He joined the Department of Electrical and Com-
puter Engineering, Ryerson University, Toronto, ON,
Canada, in July 1999, and currently he is an Asso-
ciate Professor and Chairman of the Department. He
holds the Canada Research Chair position in Biomed-
ical Signal Analysis.
Dr. Krishnan is a recipient of the 2007 Young Engineer Achievement Award
from Engineers’ Canada.
Tom Chau (S’92–M’97–SM’03) received the BASc.
degree in engineering science in 1992 and the MASc.
degree in electrical and computer engineering in 1994
from the University of Toronto. He received the Ph.D.
degree in systems design engineering in 1998 from
the University of Waterloo.
After spending some years in industry, he joined
the staff of Bloorview Kids Rehab, where he is cur-
rently a scientist in the Bloorview Research Institute.
He is also an Assistant Professor in the Institute of
Biomaterials and Biomedical Engineering at the Uni-
versity of Toronto, where he directs the Clinical Engineering Program. He holds
a Canada Research Chair in Pediatric Rehabilitation Engineering. His current
research interests revolve around innovative access technologies for children
and youth with severe motor impairments who are nonverbal.