Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
TOWARD OPTIMIZING ELECTRODE CONFIGURATIONS TO IMPROVE
MYOELECTRIC PATTERN RECOGNITION Aaron Young
1,2, Levi Hargrove
2,3
1. Dept. of Biomedical Engineering, Northwestern University, Evanston, IL, USA
2. Center for Bionic Medicine, Rehabilitation Institute of Chicago, Chicago, IL, USA
3. Dept. of Physical Medicine and Rehabilitation, Northwestern University, Chicago, IL, USA
ABSTRACT
Standard myoelectric control systems use carefully
placed bipolar electrode pairs to provide independent
myoelectric signals (MESs) for prosthesis control. Because
myoelectric pattern recognition systems do not require
isolated MESs, the two electrode poles used for each MES
channel may be placed longitudinally along individual
muscles or transversely across multiple muscles. In
addition, each electrode pole can be combined with a
number of additional poles to form multiple channels.
However, practical issues limit the number of poles that can
be used in clinical settings. In this study, we investigated
classification error reduction and controllability
improvements provided by a combination of transverse and
longitudinal MES channels in two conditions: (1) a constant
number of electrode poles, and (2) a constant number of
MES channels. In both cases, we also investigated
performance when the electrodes were slightly shifted from
their original positions to evaluate sensitivity to electrode
shift. We found that a combination of two transverse and
two longitudinal electrode channels constructed from four
poles significantly outperformed the individual
performances of either two transverse or two longitudinal
channels each constructed from four poles (p<0.01). Using
eight poles, we found that the best channel subset was
always comprised of a combination of transverse and
longitudinal channels. These results are important because
the number and arrangement of poles and channels is a
practical consideration for successful clinical
implementation of myoelectric pattern recognition control.
INTRODUCTION
Myoelectric pattern recognition systems show promise
for intuitive control of prostheses with multiple degrees of
freedom [1]. Despite two decades of extensive research,
these systems have yet to be clinically implemented.
Typically, studies of pattern recognition systems have
focused on signal processing aspects including data
windowing [2], feature extraction [3], classification [4], and
post-processing [5]. However, a key component of the
system is the placement and configuration of the electrode
poles. This is true in two contexts: first, the information
content of each MES channel is affected by the
configuration of the two electrode poles, and second, use of
more electrode channels increases computational and
financial costs and use of more electrode poles poses
technical difficulties in embedding electrodes into prosthetic
sockets.
The information content of an MES is determined by the
electrode detection volume, which defines the selectivity of
the electrode. The primary factor affecting selectivity is the
interelectrode distance: a rough estimate of detection
volume is given by a sphere with radius equal to the
interelectrode distance [6]. Most experiments have been
conducted with interelectrode distances of approximately 2
cm, resulting in selective recordings. However, in pattern
recognition systems, nonselective MES recordings may
provide a different set of information that is complementary
to the selective information. In addition, nonselective MES
recordings may be less sensitive to changes in the location
of the recording electrodes [7], such as those that result from
donning and doffing the prosthesis or socket shift during
use. Electrode shift is a potential problem in clinical
applications because data presented to the classifier when
electrodes are shifted are different from training data [8].
The number of channels used for classification of MESs
varies between studies based on the classification problem
and available recording equipment. Four channels are often
used in studies of myoelectric pattern recognition control by
transradial amputees [9]. One study [4] showed that for a
myoelectric control of 10 motion classes using forearm
muscles, four channels provided a sufficient amount of
information for classification, and additional channels did
not increase classification accuracy. In fact, the study
showed a small decrease in accuracy as the number of
channels was increased to 16. Other studies have shown
similar findings with a plateau effect in classification
accuracy with increasing numbers of channels [10, 11]. The
number of channels is an important property of the MES
detection system, and this study examines this property in
terms of robustness to electrode shift.
We compared pattern recognition system performance
when using combinations of selective and nonselective
recordings by measuring classification error and
controllability testing. Selective recordings were obtained
from bipolar electrode pairs with small interelectrode
distances aligned with the underlying muscle fibers.
Nonselective recordings were obtained from bipolar
electrode pairs aligned in a transverse orientation to
underlying muscle fibers and spanning muscle groups: one
electrode pole was placed on the wrist flexor muscle group
and one electrode pole is placed on the wrist extensor
From "MEC 11 Raising the Standard," Proceedings of the 2011 MyoElectric Controls/Powered Prosthetics Symposium Fredericton, New Brunswick, Canada: August 14-19, 2011. Copyright University of New Brunswick.
Distributed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 United States License by UNB and the Institute of Biomedical Engineering, through a partnership with Duke University and the Open Prosthetics Project.
muscle group. This configuration has a large interelectrode
distance and records a global signal from multiple muscles
[3].
In this paper, we seek to provide clinical
recommendations for electrode placement and number of
electrodes and recording channels based on offline
classification error, real-time controllability scores, and
robustness to electrode shift.
METHODS
Experiment 1: Effect of Pole Number and Placement
Seven able-bodied subjects participated in the study,
which was approved by the Northwestern University
Institutional Review Board. Two control sites on the
forearm were used: one on the flexor muscle group and one
on the extensor muscle group. At each control site, two
surface electrodes were placed longitudinal to the direction
of the underlying muscle fibers. A ground electrode was
placed on a bony region near the elbow away from the
muscles of interest. Four bipolar MES channels were
formed from these four electrode pole locations (
). Two channels were longitudinal (spanning each
individual control site), and two were transverse (one pole
on flexors and one on extensors).
For each subject, the classifier was trained and tested
offline with electrodes located at the nominal (or no-shift)
location. The electrodes were manually shifted 1 and 2 cm
from the nominal position in the direction parallel to the
underlying muscle fibers (distal to the subject) and 1 and 2
cm from the nominal position perpendicular to the
underlying muscles (clockwise from the subject’s
perspective). Testing data were recorded at each of these
four shift locations. Seven motion classes were recorded:
wrist flexion, wrist extension, forearm pronation, forearm
supination, hand close, hand open, and no movement.
A pattern recognition system similar to that used
previously [1] was used to discriminate motion classes.
MESs were sampled at 1 kHz and high-pass filtered at 20
Hz. Data were windowed in 250 ms intervals with 50 ms
overlap [2]. Time domain features [3] were classified using
linear discriminant analysis [1].
The performance of the classifier with three different
channel combinations using the same four electrode pole
locations were tested: (1) using two longitudinal channels,
(2) using two transverse channels, and (3) using a
combination of two longitudinal and two transverse
channels. We also compared our results to those achieved
when data collected from the shifted locations were
incorporated into the training data (referred to as
displacement training) in order to reduce classifier
sensitivity to electrode shift [12].
A Target Achievement Control (TAC) test [5] was used
to evaluate controllability. Subjects controlled a virtual
prosthesis to achieve target postures shown on a screen (see
[2, 5] for more details on TAC testing). Only one motion
class (e.g. wrist flexion and extension) was required per
trial. If mistakes were made, for example, activation of a
different motion class or overshooting the target, the subject
had to make a corrective activation. The subject had 17 s to
complete each trial. A test consisted of two trials of each
motion class. Performance was assessed by failure rate: the
percentage of trials that the subject did not complete during
a test. TAC tests were completed at the no-shift and 2 cm
shift (both parallel and perpendicular) locations. At each
location, one test was completed for two longitudinal
channels, and one test was completed using two longitudinal
and two transverse channels.
Experiment 2: Effect of Number of Channels
This experiment was similar to the first except four
control sites spaced equally around the circumference of the
forearm were used. Eight bipolar MES channels were
formed from eight electrode pole locations (Figure 2). Four
channels were longitudinal and four were transverse in an
arrangement similar to experiment 1.
Medial View Lateral View
Figure 1: Electrode placement for experiment 1.
Electrode poles 1 and 2 are located on wrist flexors and
3 and 4 on wrist extensor muscles. The two longitudinal
channels were the bipolar pairs of 1 & 2 and 3 & 4. The
transverse channels were 1 & 3 and 2 & 4.
From "MEC 11 Raising the Standard," Proceedings of the 2011 MyoElectric Controls/Powered Prosthetics Symposium Fredericton, New Brunswick, Canada: August 14-19, 2011. Copyright University of New Brunswick.
Distributed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 United States License by UNB and the Institute of Biomedical Engineering, through a partnership with Duke University and the Open Prosthetics Project.
Medial View Lateral View
Figure 2: Electrode placement for experiment 2.
Longitudinal channels were formed between poles 1 & 2, 3
& 4, 5 & 6, and 7 & 8. Transverse channels were formed
between poles 1 & 3, 2 & 4, 5 & 7, and 6 & 8.
The same motions and classification techniques as
described for experiment 1 were used. The classifier was
trained and tested at the no-shift location. In this
experiment, electrodes were only shifted 1 and 2 cm
perpendicular to the no-shift location, because results from
experiment 1 showed that the classifier was more sensitive
to perpendicular shifts than to parallel shifts. Classifier
performance resulting from use of from one to eight
channels was evaluated and the best channel subset for each
number of electrodes was determined based on the error at
the no shift, 1 cm shift, and 2 cm shift locations using the
following weighted error formula:
Weighted Error = 2*Error(No Shift) + 1.5*Error(1 cm Shift)
+Error(2 cm shift).
Every combination of channels was tested and a
weighted error was assigned to each combination. The
optimal combination was that which had the lowest
weighted error compared to others with the same number of
channels.
RESULTS
Experiment 1: Effect of Pole Number and Placement
In the first experiment, we evaluated the best electrode
channel configuration for a constant number of electrode
poles. Two longitudinal and two transverse channels were
formed from the same four electrode poles in different
configurations. The combined selective and nonselective
information from both longitudinal and transverse channels
performed the best with and without shift (Figure 3)
compared to only transverse channels (p<0.01) and only
longitudinal channels (p<0.05).
0
10
20
30
40
50
60
No shift 1cm || 2cm || 1cm ⊥ 2cm ⊥
Cla
ssif
ica
tio
n E
rro
r
Trasverse Only
Longitudinal Only
Longitudinal Only with Displacement Training
Longitudinal and Transverse
Longitudinal and Transverse with Displacement Training
Figure 3: Classification error of electrode configurations
using four electrode poles. Displacement training results are
also displayed for two of the configurations. Error bars
show one standard error of the mean. || refers to parallel
shifts and ⊥ refers to perpendicular shifts.
Training with displacement data increased error at the
no-shift location in both cases, but also decreased sensitivity
to electrode shift (Figure 3). Displacement training did not
reduce sensitivity to shift for the longitudinal channels as
much as adding the transverse channel did. However, by
adding the transverse channels and incorporating
displacement training data, sensitivity to shift was greatly
reduced at all shift locations (p<0.05) compared to a
combination of longitudinal and transverse channels without
displacement training.
TAC test controllability results demonstrated a trend
similar to that of the classification error: a combination of
longitudinal and transverse channels outperformed
longitudinal channels alone (Figure 4), especially when
electrodes were shifted. The controllability test was not hard
enough to separate out the performance of the two
classifiers at the no shift location, as both had very low
failure rates.
From "MEC 11 Raising the Standard," Proceedings of the 2011 MyoElectric Controls/Powered Prosthetics Symposium Fredericton, New Brunswick, Canada: August 14-19, 2011. Copyright University of New Brunswick.
Distributed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 United States License by UNB and the Institute of Biomedical Engineering, through a partnership with Duke University and the Open Prosthetics Project.
0
10
20
30
40
50
60
70
80
No Shift Parallel Shift Perpendicular Shift
Failu
re R
ate
(%)
Longitudinal and Transverse
Longitudinal Only
Figure 4: TAC test failure rates for two electrode
configurations with and without 2 cm shifts. Error bars
show one standard error of the mean.
The longitudinal channels alone performed well without
shift, but had high failure rate with shift in either direction.
By adding transverse channels, the no-shift failure rate
dropped slightly and the failure rate was substantially
reduced for both shift directions. In particular, using both
longitudinal and transverse channels and a 2 cm parallel
shift the failure rate was lower than the longitudinal
channels without shift.
Experiment 2: Effect of Number of Channels
First, the best subset of channels for each number of
channels (between one and eight) was determined based on
their weighted classification errors. Only the unweighted
classification errors are displayed here (Figure 5).
Interestingly, combinations with at least one longitudinal
and one transverse channel were always in the best subset
when more than one channel was used.
The use of four to six channels had the lowest
classification error across the three shift conditions. An
important result was that errors were below 15% at the
nonshifted and 1 cm shift locations when using more than
two recording channels. The 2 cm shift location had high
classification error regardless of the number of recording
channels.
0
5
10
15
20
25
30
35
40
45
50
1 2 3 4 5 6 7 8
Cla
ssif
icat
ion
Err
or
(%)
Number of Channels
no shift
1 cm shift
2cm shift
Figure 5: Effect of the number of recording channels on
classification error. Error bars show one standard error of
the mean.
DISCUSSION
The first goal of this study was to find the electrode
configuration that gave the highest performance given a
limited number of electrode pole locations. Performance
was measured in terms of classification error,
controllability, and robustness to electrode shift. We found
that a combined configuration that included both
longitudinal and transverse channels was the highest
performing configuration of those that were tested. This
configuration had lower error and better controllability
without shift and at every shift location tested compared to
using only longitudinal or transverse channels.
Previous investigators have considered including
displacement location data in the training data in order to
train a more robust pattern recognition system [12]. We
repeated this analysis for the configuration with longitudinal
channels and the configuration with combined longitudinal
and transverse channels. We found that displacement
training increased classification error at the no-shift
location, but helped to reduce sensitivity at shift locations.
The combination of displacement location with longitudinal
and transverse channels performed especially well with less
than 15% classification error at all tested locations. This is a
clinically useful result as with only four channels and four
pole locations, low classification errors were achieved with
high robustness across shift conditions.
The second goal of this study was to analyze the
number of recording channels necessary for sufficiently
high classification and to determine the optimal composition
of channel subsets. Based on weighted averages, it was
found that a combination of longitudinal and transverse
channels was always optimal regardless of the number of
channels. The combination of selective and nonselective
From "MEC 11 Raising the Standard," Proceedings of the 2011 MyoElectric Controls/Powered Prosthetics Symposium Fredericton, New Brunswick, Canada: August 14-19, 2011. Copyright University of New Brunswick.
Distributed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 United States License by UNB and the Institute of Biomedical Engineering, through a partnership with Duke University and the Open Prosthetics Project.
information decreased error with and without shift
compared to configurations with only one type of electrode
configuration. Experiment 2 showed that there was little or
no improvement in terms of classification error when using
more than six channels. For the transradial case, this study
demonstrates that four to six channels are sufficient to
obtain classification errors of less than 15% both without
electrode shift and with shifts up to 1 cm.
ACKNOWLEDGEMENTS
The authors would like to acknowledge Ann Barlow and
Aimee Schultz for helping to revise and edit the manuscript,
Annie Simon for providing valuable contributions, and
Todd Kuiken for providing overall project guidance for this
study.
REFERENCES
[1] T. A. Kuiken, G. Li, B. A. Lock, R. D. Lipschutz, L. A. Miller,
K. A. Stubblefield, and K. B. Englehart, "Targeted muscle reinnervation for
real-time myoelectric control of multifunction artificial arms," JAMA, vol. 301, pp. 619-28, 2009.
[2] L. H. Smith, L. Hargrove, B. A. Lock, and T. Kuiken,
"Determining the optimal window length for pattern recognition-based myoelectric control: balancing the competing effects of classification error
and controller delay," IEEE Transactions on Neural Systems and
Rehabilitation Engineering, vol. In Press, 2010. [3] B. Hudgins, P. Parker, and R. N. Scott, "A new strategy for
multifunction myoelectric control," IEEE Transactions on Biomedical
Engineering, vol. 40, pp. 82-94, 1993. [4] L. Hargrove, K. Englehart, and B. Hudgins, "A Comparison of
Surface and Intramuscular Myoelectric Signal Classification," IEEE
Transactions on Biomedical Engineering, 2007.
[5] A. Simon, L. Hargrove, B. A. Lock, and T. Kuiken, "A
decision-based velocity ramp for minimizing the effect of
misclassifications during real-time pattern recognition control," IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol.
Accepted, 2010.
[6] G. Kamen and D. Gabriel, Essentials of Electromyography. Champaign: Human Kinetics Publishers, 2009.
[7] A. Young, L. Hargrove, and T. Kuiken, "The Effects of
Electrode Size and Orientation on the Sensitivity of Myoelectric Pattern Recognition Systems to Electrode Shift," IEEE Transactions on
Biomedical Engineering, under minor revisions. [8] L. Hargrove, K. Englehart, and B. Hudgins, "The effect of
electrode displacements on pattern recognition based myoelectric control,"
in Proceedings of the 28th IEEE EMBS Annual International Conference, New York City, 2006, pp. 2203-2206.
[9] K. Englehart and B. Hudgins, "A robust, real-time control
scheme for multifunction myoelectric control," IEEE Transactions on Biomedical Engineering, vol. 50, pp. 848-54, 2003.
[10] H. Huang, P. Zhou, G. Li, and T. A. Kuiken, "An Analysis of
EMG Electrode Configuration for Targeted Muscle Reinnervation Based Neural Machine Interface," IEEE Transactions on Neural Systems and
Rehabilitation Engineering, vol. 16, pp. 37-45, 2008.
[11] G. Li, A. E. Schultz, and T. A. Kuiken, "Quantifying pattern recognition-based myoelectric control of multifunctional transradial
prostheses," IEEE Transactions on Neural Systems and Rehabilitation
Engineering, vol. 18, pp. 185-92, 2010. [12] L. Hargrove, K. Englehart, and B. Hudgins, "A training strategy
to reduce classification degradation due to electrode displacements in
pattern recognition based myoelectric control," Biomedical Signal Processing and Control, vol. 3, pp. 175-180, 2008.
From "MEC 11 Raising the Standard," Proceedings of the 2011 MyoElectric Controls/Powered Prosthetics Symposium Fredericton, New Brunswick, Canada: August 14-19, 2011. Copyright University of New Brunswick.
Distributed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 United States License by UNB and the Institute of Biomedical Engineering, through a partnership with Duke University and the Open Prosthetics Project.