Escolar Documentos
Profissional Documentos
Cultura Documentos
By
ANIRBAN DUTTA
August, 2008
1
To the Almighty, Baba, Ma, Kiran, and Rachna
2
TABLE OF CONTENTS
page
LIST OF TABLES...........................................................................................................................6
Introduction....................................................................................................................................13
Abstract...................................................................................................................................20
Introduction.............................................................................................................................21
Methods ..................................................................................................................................22
Subjects............................................................................................................................22
Test of Controllability .....................................................................................................23
Test of Discriminability...................................................................................................24
Statistical Analysis ..........................................................................................................28
Results.....................................................................................................................................29
Results from the Test of Controllability..........................................................................29
Results from the Test of Discriminability .......................................................................30
Discussion...............................................................................................................................32
Conclusion ..............................................................................................................................33
References...............................................................................................................................34
Figures ....................................................................................................................................37
Tables......................................................................................................................................46
Abstract...................................................................................................................................50
Introduction.............................................................................................................................50
Methods ..................................................................................................................................52
Subjects............................................................................................................................52
Data Acquisition and Processing.....................................................................................53
Muscle Selection .............................................................................................................55
Classifier Development and Offline Testing...................................................................56
Classifier Testing During FES-assisted Ambulation.......................................................58
Results.....................................................................................................................................59
3
Classifier Performance ....................................................................................................59
Repeatability of the Classifier Performance....................................................................60
Discussion...............................................................................................................................60
Conclusion ..............................................................................................................................62
References...............................................................................................................................63
Figures ....................................................................................................................................66
Abstract...................................................................................................................................73
Introduction.............................................................................................................................74
Methods ..................................................................................................................................75
Subjects............................................................................................................................75
Gait Data Acquisition ......................................................................................................76
Gait Parameters ...............................................................................................................78
Statistical Analysis ..........................................................................................................79
Results.....................................................................................................................................80
Discussion...............................................................................................................................81
Conclusion ..............................................................................................................................83
References...............................................................................................................................84
Figures ....................................................................................................................................85
Tables......................................................................................................................................90
Abstract...................................................................................................................................92
Introduction.............................................................................................................................92
Methods ..................................................................................................................................95
Subjects............................................................................................................................95
Gait Data Acquisition ......................................................................................................95
Coordination and Stability Analysis of Gait initiation ....................................................98
Results...................................................................................................................................103
Linear regression model for gait initiation ....................................................................103
Coordination and stability during FES-assisted gait initiation......................................104
Discussion.............................................................................................................................105
Conclusions...........................................................................................................................109
References.............................................................................................................................110
Figures ..................................................................................................................................113
Abstract.................................................................................................................................123
Introduction...........................................................................................................................124
Methods ................................................................................................................................126
4
Subjects..........................................................................................................................126
Command source selection............................................................................................127
Implantation of intramuscular EMG electrode..............................................................129
Classifier development for iEMG-triggered FES-assisted stepping .............................130
Online testing of the classifier in the laboratory ...........................................................134
Results...................................................................................................................................136
Muscles and location selection for intramuscular EMG ...............................................136
Classifier development and online performance ...........................................................137
Discussion.............................................................................................................................142
Conclusions...........................................................................................................................145
References.............................................................................................................................146
Figures ..................................................................................................................................149
Tables....................................................................................................................................163
5
LIST OF TABLES
Table page
Table 2.1: The mean, the minimum, and the maximum average absolute tracking error in
%MVC during the four parts (0-25 sec, 25-50 sec, 50-75 sec, 75-100 sec) of the Test
for Controllability. The p-value from the one-way two-tailed ANOVA test for the
average tracking error over the whole trial (100 sec) was not ‘statistically significant’
( ≥ 0.01)...............................................................................................................................46
Table 2.2: The results from the Test of Discriminability for the muscles Gluteus Medius
(GM), Biceps Femoris (BF), Medial Gastrocnemius (MG), Rectus Femoris (RF),
Tibialis Anterior (TA), and Erector Spinae (ES at T9) are presented for the able-
bodied subjects. The Wilcoxon statistic (W) was similar in magnitude to the
corresponding Discriminability Index (DI). Similarly the Standard Deviation (SD) of
the DI over 10 random partitions (i.e., 10-fold cross-validation) was similar in
magnitude to the Standard Error (SE) found for the Wilcoxon statistic (W). There
were ‘statistically significant’ (p ≥ 0.05) differences in the means of DI due to the
muscle type as well as the classifier type...........................................................................47
Table 2.3a: The results from the Test of Discriminability of iSCI-1 for the left step classifier.
The Wilcoxon statistic (W) was similar in magnitude to the corresponding value of
the Discriminability Index (DI). Similarly the Standard Deviation (SD) of the DI was
similar in magnitude to the Standard Error (SE) found for the Wilcoxon statistic (W).
There were ‘statistically significant’ (p ≥ 0.05) differences in the means of DI due to
the muscle type as well as the classifier type.....................................................................48
Table 2.3b: The results from the Test of Discriminability of iSCI-1 for the right step. The
Wilcoxon statistic (W) was similar in magnitude to the corresponding value of the
Discriminability Index (DI). Similarly the Standard Deviation (SD) of the DI was
similar in magnitude to the Standard Error (SE) found for the Wilcoxon statistic (W).
There were ‘statistically significant’ (p ≥ 0.05) differences in the means of DI due to
the muscle type as well as the classifier type.....................................................................48
Table 2.4a: The results from the Test of Discriminability of iSCI-2 for the left step. The
Wilcoxon statistic (W) and the corresponding value of the Discriminability Index
(DI) were similar. The Standard Deviation (SD) of the DI and the Standard Error
(SE) found for the Wilcoxon statistic (W) were similar. There were ‘statistically
significant’ (p ≥ 0.05) differences in the means of DI due to the muscle type as well
as the classifier type. ..........................................................................................................49
Table 2.4 b: The results from the Test of Discriminability of iSCI-2 for the right step
classifier. The Wilcoxon statistic (W) and the corresponding value of the
Discriminability Index (DI) were similar. The Standard Deviation (SD) of the DI and
the Standard Error (SE) found for the Wilcoxon statistic (W) were similar. There
were ‘statistically significant’ (p ≥ 0.05) differences in the means of DI due to the
muscle type as well as the classifier type...........................................................................49
6
Table 4.1: The Mean, Standard Deviation (S.D), coefficient of variation (C.V.), 95%
confidence interval (95% C.I.) over 10 trials (N=10) of the EMG-triggered and
switch-triggered gait parameters – gait speed (m/s), left step length (m), right step
length (m), left double support duration (s), right double support duration (s), left
swing phase duration (s), right swing phase duration (s) for the subject iSCI 1. [
‘statistically significant’ (p<0.05) difference between the command sources are
shaded] ...............................................................................................................................90
Table 4.2: The Mean, Standard Deviation (S.D), coefficient of variation (C.V.), 95%
confidence interval (95% C.I.) over 10 trials (N=10) of the EMG-triggered and
switch-triggered gait parameters – gait speed (m/s), left step length (m), right step
length (m), left double support duration (s), right double support duration (s), left
swing phase duration (s), right swing phase duration (s) for the subject iSCI 2. [
‘statistically significant’ (p<0.05) difference between the command sources are
shaded] ...............................................................................................................................91
Table 6.2: The performance of the pattern recognition classifier for iSCI-2...............................163
Table 6.3: The minimum achievable sample time in µs for the classifier algorithm in
different hardware that can serve as single board computer............................................164
7
LIST OF FIGURES
Figure page
Figure 2.1: Experimental setup for the Test of Controllability of the surface EMG from
Rectus Femoris using visual pursuit tasks while the knee is fixed in a dynamometer. .....37
Figure 2.2: Experimental setup for surface EMG data collection with switch-triggered FES-
assisted overground walking..............................................................................................38
Figure 2.3: Experimental protocol for surface EMG data collection during overground
walking, where the subject had to start from standing and achieve a self-selected gait
speed within 5m. ................................................................................................................39
Figure 2.4: The left column shows the cumulative distribution function for the three cases,
0 < DI ≤ 0.5, 0.5 ≤ DI < 1, DI = 1 and the right column shows the corresponding
Receiver Operating Characteristics curve..........................................................................40
Figure 2.5: TRACKING (broken black line) and TARGET (solid black line) signals during
visual pursuit task for the Test of Controllability. The boxes at each data point show
the lower quartile and upper quartile values of the TRACKING signal. Whiskers
extending at the top and bottom of the boxes show the range of the TRACKING
signal. The top panel presents the results for iSCI-1 and the bottom panel for iSCI-2.
The left panel presents the results for the left Rectus Femoris and the right panel
presents the results for the right Rectus Femoris. ..............................................................41
Figure 2.6: TRACKING (broken black line) and TARGET (solid black line) signals during
visual pursuit task for the Test of Controllability with able-bodied subjects. The
boxes at each data point show the lower quartile and upper quartile values of the
TRACKING signal. Whiskers extending at the top and bottom of the boxes show the
range of the TRACKING signal. .......................................................................................42
Figure 2.7: Top panel shows the results from the post hoc analysis of the Discriminability
Index with their critical values from Scheffe’s S procedure for the muscles – Gluteus
Medius (GM), Biceps Femoris (BF), Medial Gastrocnemius (MG), Rectus Femoris
(RF), Tibialis Anterior (TA), and Erector Spinae (ES at T9) obtained from the Test
of Discriminability with able-bodied subjects. The bottom panel shows the results
from the post hoc analysis of the Discriminability Index with their critical values
from Scheffe’s S procedure for different classifiers – Pattern Recognition Classifier
(PRC) and Threshold-based Classifier (TC) obtained from the Test of
Discriminability with able-bodied subjects. ......................................................................43
Figure 2.8: Top panel shows the results from the post hoc analysis of the Discriminability
Index with their critical values from Scheffe’s S procedure for the muscles – Gluteus
Medius (GM), Biceps Femoris (BF), Medial Gastrocnemius (MG), Rectus Femoris
(RF), Tibialis Anterior (TA), and Erector Spinae (ES at T9) obtained from the Test
of Discriminability of the left and right step classifiers of iSCI-1. The bottom panel
8
shows the results from the post hoc analysis of the Discriminability Index with their
critical values from Scheffe’s S procedure for different classifiers – Pattern
Recognition Classifier (PRC) and Threshold-based Classifier (TC) obtained from the
Test of Discriminability of the left and the right step classifiers of iSCI-1.......................44
Figure 2.9: Top panel shows the results from the post hoc analysis of the Discriminability
Index with their critical values from Scheffe’s S procedure for the muscles – Gluteus
Medius (GM), Biceps Femoris (BF), Medial Gastrocnemius (MG), Rectus Femoris
(RF), Tibialis Anterior (TA), and Erector Spinae (ES at T9) obtained from the Test
of Discriminability of the left and right step classifiers of iSCI-2. The bottom panel
shows the results from the post hoc analysis of the Discriminability Index with their
critical values from Scheffe’s S procedure for different classifiers – Pattern
Recognition Classifier (PRC) and Threshold-based Classifier (TC) obtained from the
Test of Discriminability of the left and the right step classifiers of iSCI-2.......................45
Figure 3.1: a) X-ray of the iSCI subject implanted with implantable receiver-stimulator
(IRS-8) b) iSCI subject stepping with the switch-triggered FES system ..........................66
Figure 3.2: Experimental setup for testing EMG-triggered FES-assisted walking with the
block-diagram for the EMG-triggered FES-system (ECU: external control unit, LE:
linear envelope)..................................................................................................................67
Figure 3.3: Processing of the sampled EMG from Erector Spinae for training the classifier a)
rectified and reconstructed EMG signal b) linear envelope found from processed
EMG signal ........................................................................................................................68
Figure 3.4: Muscle selection for the classifier using receiver operating characteristics curve
from switch-triggered FES-assisted gait data (FS: Foot-Strike, FO: Foot-Off) a)
linear envelope (LE) indicating class ‘True’ b) linear envelope (LE) indicating class
‘False’ ................................................................................................................................69
Figure 3.5: Receiver operating characteristics curve of the classifiers using the test data ............70
Figure 3.7: Offline testing of the classifier using receiver operating characteristics curve a)
time-error (negative means prediction) in detection of foot-off by the classifier b)
duration of the gait phases (Left DS: double support phase following left swing
phase, Right DS: double support phase following right swing phase, SW: swing
phase) .................................................................................................................................72
Figure 4.1: Experimental setup for testing EMG-triggered FES-assisted walking with the
block-diagram for the EMG-triggered FES-system (ECU: external control unit). ...........85
Figure 4.2: EMG-based gait event detector for triggering FES-assisted steps. ............................86
Figure 4.3: Plot of the Root Mean Square Error (RMSE) between the low-pass filtered and
unfiltered foot progression in sagittal plane with cut-off frequencies to find the
9
optimum cut-off frequency for low-pass filtering the gait kinematics data. Optimum
cut-off frequency was found to be 3.5 Hz for iSCI data....................................................87
Figure 4.4: Gait data collection protocol in laboratory conditions where the subject had to
start from standing and achieve a self-selected gait speed within ViconTM volume of
data capture (~5m). ............................................................................................................88
Figure 4.5: Boxplot of average body weight support provided by the walker during EMG-
triggered (N=10 trials) and switch-triggered (N=10 trials) gait normalized by the
mean of that during EMG-triggered trials of iSCI 2. The box shows the lower
quartile, median, and upper quartile with whiskers extending at each end showing the
range of the data. The notches around the median show the estimate of the
uncertainty. The boxes whose notches don’t overlap indicate that their medians differ
at 5% significance level. ....................................................................................................89
Figure 5.1: Laboratory setup for EMG-triggered FES-assisted walking shown with a
flowchart for the EMG-based gait event detector for triggering FES-assisted steps.......113
Figure 5.2: Top panel: Selection of optimum cut-off frequencies for low-pass filtering the
kinematic data. Bottom panel: Most of power content in the signals was below the
optimum cut-off frequency, which were 6 Hz for able-bodied and 3.5 Hz for iSCI
data...................................................................................................................................114
Figure 5.4: Typical pelvis motion in the direction of progression during gait initiation.............116
Figure 5.5: Euclidean distance from the origin of the perturbation of the 36 states during gait
initiation at the maximum left knee flexion. Left panel: able-bodied data (4 subjects).
Middle panel: iSCI data (subject C1). Right panel: iSCI data (subject C2).
[Normative: 4 subjects, 10 trials each; iSCI EMG-trigger: 2 subjects, 10 trials each;
iSCI Switch-trigger: 2 subjects, 10 trials each; iSCI Auto-trigger: 2 subjects, 10 trials
each].................................................................................................................................117
Figure 5.6: Percent Variance Accounted For (%VAF) by the Principal Components (PC).
Top panel: able-bodied data. Middle panel: iSCI-1 walking with EMG, switch, and
auto triggered FES. Bottom panel: iSCI-2 walking with EMG, switch, and auto
triggered FES. All the plots show the data averaged over 6 gait events..........................118
Figure 5.7: Typical loading of the first 3 Principal Components (PCs) on the joint angles
(HA: Hip Angle, KA: Knee Angle, AA: Ankle Angle) found from the weight
matrix W of the subject Able1. The prefix ‘l’ indicates the left side and ‘r’ indicates
the right side. The suffix ‘x’ denotes sagittal plane, ‘y’ denotes frontal plane, and ‘z’
denotes transverse plane for the joint angles. ..................................................................119
Figure 5.8: Euclidean distance from the origin of the perturbation of the 5 principal
components at maximum left knee flexion Left panel: able-bodied (4 subjects).
Middle panel: iSCI-1 subject C1. Right panel: iSCI-2 subject C2. [Normative: 4
10
subjects, 10 trials each; iSCI EMG-trigger: 2 subjects, 10 trials each; iSCI Switch-
trigger: 2 subjects, 10 trials each; iSCI Auto-trigger: 2 subjects, 10 trials each] ............120
Figure 5.9: Top panel: Scatter plot of QoF and Av. Eig. at 6 gait events for the groups; the 4
able-bodied subjects: Able1, Able2, Able3, Able4, and the 2 iSCI subjects with
different trigger modes: EMG1, EMG2, SW1, SW2, Auto1, Auto2. Bottom panel:
MANOVA cluster dendrogram plot of the groups ..........................................................121
Figure 5.10: Mahalanobis distances matrix between each pair of group means .........................122
Figure 6.1: Experimental setup for data collection during FES-assisted walking with the
block-diagram for the FES-system (ECU: external control unit, LE: linear envelope)...149
Figure 6.2: Processing of the sampled surface EMG a) rectified and reconstructed sEMG
signal b) linear envelope found from processed sEMG signal. .......................................150
Figure 6.3: Experimental protocol for the collection of EMG data during over-ground
walking in the laboratory. ................................................................................................151
Figure 6.4: Multi-electrode matrix for simultaneous collection of the surface EMG from
multiple locations on the muscle belly.............................................................................152
Figure 6.5: The steps during the implantation of intramuscular EMG electrode a) insertion of
probe, b) deployment of peelable sheath over probe, c) insertion of the iEMG
electrode through the peelable sheath, d) peeling off of the polymer sheath leaving
the iEMG electrode in place. ...........................................................................................153
Figure 6.6: Pulse-width map of the stimulation patterns used for walking shown as an
example. ...........................................................................................................................154
Figure 6.7: The real-time cycle in IST with 50 ms time period for stimulation frequency of
20 Hz................................................................................................................................155
Figure 6.8: Parameters for the iEMG classifier computed from the training data that was
collected with the switch-triggered FES system..............................................................156
Figure 6.9: The flow chart of the iEMG-based two-stage classifier for triggering FES for
walking.............................................................................................................................157
Figure 6.10: Usability Rating Scale to find the user perspective on ease/difficulty of using
the classifier [6.29]...........................................................................................................158
Figure 6.11: Best location found from the surface EMG for implanting intramuscular EMG
electrodes a) Left gastrocnemius and right erector spinae b) Left and right
gastrocnemius. .................................................................................................................159
Figure 6.12: a) Discriminability Index (DI) of left medial gastrocnemius (MG) for the swing
phase (SW) and double support phase (DS) during over-ground walking for the
11
subject iSCI-1 at each data point of the gait cycle b) Discriminability Index (DI) of
right erector spinae (ES) for the swing phase (SW) and double support phase (DS)
during over-ground walking for the subject iSCI-1 at each data point of the gait
cycle. ................................................................................................................................160
Figure 6.13: Inhibition of iEMG from right erector spinae during right swing phase (SW) as
shown in the top panel due to electrical stimulation of the same muscle when
compared to that in absence of electrical stimulation shown in the bottom panel of
the subject iSCI-1.............................................................................................................161
Figure 6.14b: a) Discriminability Index (DI) of right medial gastrocnemius (MG) for the
swing phase (SW) and double support phase (DS) during over-ground walking for
the subject iSCI-2 at each data point of the gait cycle b) Discriminability Index (DI)
of left medial gastrocnemius (MG) for the swing phase (SW) and double support
phase (DS) during over-ground walking for the subject iSCI-2 at each data point of
the gait cycle. ...................................................................................................................162
12
CHAPTER 1
INTRODUCTION
Paralysis can be caused by an injury to the spinal cord that may partially or completely
damage the communication between the brain and the muscles. The spinal cord injury (SCI) can
between the brain and the lower motor neurons below the level of injury. There are
approximately 250,000 people living with SCI in USA and about 11,000 new cases each year
[1.1]. If the paralyzed muscles below the level of injury remain innervated after the injury then
they can be electrically activated by applying a series of electrical current pulses. Functional
neuromuscular function after paralysis. FES was first used by Liberson for actuating paralyzed
limbs [1.2]. FES has been successful in providing walking function to spinal cord injured
individuals with limited or no walking abilities [1.3]. Most of the commercially available FES-
systems as well as the one that is currently used by our group needs user input to select menu
options and to trigger FES-assisted stepping action. The current command interface for our FES-
system is a push-button, which can be mounted on the walker or worn on a finger [1.4-1.7]. The
push-button as a command interface is plausible for selecting menu options during standing but it
is an impediment when it has to be actuated with fingers during walking to trigger every step.
Some individuals with limited finger and hand function find it difficult to press push-buttons,
more so while trying to maintain balance during ambulation. This particular function of the push-
button as a trigger for stepping action can be replaced by a gait event detector. The gait event
detector can identify the event (appropriate time during a gait cycle) to activate the required
pattern of electrical stimulation. Some of the gait event detectors investigated in past by other
13
researchers are based on foot-switches, accelerometers, gyroscopes, and the electromyogram
since it temporally precedes the joint kinetics and kinematics (electromechanical delay about 100
ms [1.17]) and may be feasible as a control source even for individuals with incomplete spinal
cord injury (iSCI), who may have lost their ability to move but may still have volitionally
controllable EMG activity [1.18]. The natural latency between electrophysiological and
biomechanical events provides time to detect the intent and then assist the intended movement
with FES. EMG-based triggering of FES patterns should integrate the FES-generated movement
seamlessly with the volitional effort that is necessary in the case of iSCI individuals who have
The gait is roughly a cyclic process which can be divided into stepping of one side
followed by the other. A step defines the phase of the gait between foot-off that is the instant
when foot loses contact with the ground to the foot-off of the contralateral limb. Gait is
nevertheless a dynamic process where the steps dynamics are not isolated but one step leads to
the other steps in terms of the dynamics of the locomotor system. The transition between the
steps involves energy injection through push-off that generates a burst of energy causing the foot
to plantarflex and shifts the body towards the contralateral limb and subsequently allowing the
limb to swing forward. The push-off correlates with a burst in the muscle activity over multiple
synergist muscles, mainly the ankle plantar-flexors. Electromyogram (EMG) is the time history
of electrical activity in the muscle that can be used to find the activation of the muscles. The
burst in the muscle acitivity during the push-off produces burst in the volitional EMG of all the
synergist muscles which have a pattern of activation during the transition phase of gait (i.e. left
to right step and right to left step transitions). This synergistic modulation of the volitional EMG,
14
if present in partially paralyzed muscles, can be used as a feature template to identify the intent
to transition from the left to right step and right to left step even when partially paralyzed
muscles are too weak to produce enough moment at the joint to produce effective push-off.
Prior work has shown that the EMG synergies found by principal component analysis can
provide information related to gait events and also gait-speed [1.19]. Transition specific EMG
features can be identified using principal component analysis which can then be used to identify
the transition phase of the gait. A binary classifier to trigger the transition from left to the right
step and vice versa can be trained with the parameters from correlation analysis of the EMG
pattern with transition specific EMG feature template. The correlation coefficients of the features
associated with these transitions are postulated to be clustered in the feature space. During online
operation, the classifier will have to identify the cluster from windowed EMG using cross-
correlation with the specified features and determine the ‘intended’ transition. This method can
be conceptually extended to identify the transitions to other tasks like side-stepping, stair-
climbing, different gait-speeds etc. It is postulated that EMG-triggered FES-controller will have
an impact on the coordination of the FES-assisted iSCI gait. Seamlessly integrating the FES-
generated movement with the volitional movement should significantly enhance the transitions
There are challenges associated with the implementation of this method. The nature of
motor deficits in incomplete SCI population is very heterogeneous. Some individuals can walk to
a certain extent with upper-body support, some can stand using the extensor tone and some are
such that the volitional EMG from those muscles had enough information to identify the gait
phase transitions. The EMG had to be blanked during the stimulation to remove stimulation
15
artifact that reduced the information content in the EMG. This may produce overlapping clusters
in EMG feature space that will be difficult to classify with low false positive rate. More EMG
channels (more than preferred two) may be needed in order to reduce the false positive rate in
that case. In this study, the enhanced coordination during ambulation was investigated by
dynamical systems tools like return map analysis [1.20]. Subjective impressions of the two
The overall goal was to develop and evaluate an EMG-based trigger for the FES-controller
which can assist volitional motor function synergistically with electrical stimulation during gait.
Aim 1 - Muscle selection for EMG-based trigger: Select a set of two partially paralyzed
muscles in individuals with iSCI that yield consistent and reliable command information for
FES-assisted gait.
Hypothesis 1: The two partially paralyzed muscles will have volitionally controllable
1. The iSCI subjects have volitional control over the surface EMG from the partially
2. The iSCI subjects have EMG pattern in 2 partially paralyzed muscles with enough
and online-testing of a FES-controller for ambulation with a surface EMG-based classifier for
Hypothesis 2: It was hypothesized that two muscles can be used to detect intention for
foot-off with false-positive rate less that 2 % and true-positive rate greater than 85 %.
16
Aim 3 - Evaluation of EMG-triggered FES-assisted gait: Compare the FES-assisted gait
with the surface EMG-triggered FES-controller with the switch-triggered one with dynamical
systems tools like return map analysis and subjective tools like Usability Rating Scale to evaluate
from partially paralyzed muscles as a command source for triggering FES-assisted steps during
walking.
steps with surface EMG-based classifier running in real-time during over-ground ambulation.
switch-triggered stepping. Chapter 4 discusses the gait parameters during over-ground walking in
the laboratory. Chapter 5 discusses the coordination and stability during stand-to-walk transition
in the laboratory.
in the earlier chapters for triggering FES-assisted steps with a fully implantable FES system.
References
1.1. SCIIN, Spinal cord injury: facts and figures at a glance - June 2005. 2005, Spinal Cord
Injury Information Network.
17
1.2. W. T. Liberson, H. J. Holmquest, D. Scott, M.Dow, “Functional electrotherapy: stimulation
of the peroneal nerve synchronized with the swing phase of the gait of hemiplegic
patients,” Arch Phys Med Rehabil, vol. 42, 1961, pp. 101-105.
1.4. B. Smith, Z. Tang, M.W. Johnson, S. Pourmehdi, M.M. Gazdik, J.R. Buckett, and P.H.
Peckham, “An externally powered, multichannel, implantable stimulator-telemeter for
control of paralyzed muscle,” IEEE Trans Biomed Eng., vol. 45, no. 4, 1998, pp. 463-475.
1.5. Z. Tang, B. Smith, J.H. Schild, and P.H. Peckham, “Data transmission from an implantable
biotelemeter by load-shift keying using circuit configuration modulator,” IEEE Trans
Biomed Eng., vol. 42, no. 5, 1995, pp. 525-528.
1.6. N. Bhadra, K.L. Kilgore, and P.H. Peckham, “Implanted stimulators for restoration of
function in spinal cord injury,” Med. Eng. Phys., vol. 23, 2001, pp. 19-28.
1.7. J. Knutson, M. Audu, and R. Triolo, “Interventions for mobility and manipulation after
spinal cord injury: a review of orthotic and neuroprosthetic options,” Topics in Spinal Cord
Rehab, in press.
1.8. J. R.W. Morris, “Accelerometry a technique for the measurement of human body
movements,” J Biomech., vol. 6, 1973, pp. 729–36.
1.9. I. P. Pappas, M. R. Popovic, T. Keller, V. Dietz, and M. Morari, “A reliable gait phase
detection system,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 9, no. 2, Jun. 2001, pp.
113-125.
1.10. A. Mansfield, and G. M. Lyons, “The use of accelerometry to detect heel contact events for
use as a sensor in FES assisted walking,” Med. Eng. Phys., vol. 25, no. 10, Dec. 2003, pp.
879-885.
1.11. R. Williamson, and B. J. Andrews, “Gait event detection for FES using accelerometers and
supervised machine learning,” IEEE Transactions on Rehabilitation Engineering, vol. 8,
2000, pp. 312–319.
1.13. R. T. Lauer, R. T. Smith, and R. R. Betz, “Application of a neuro-fuzzy network for gait
event detection using electromyography in the child with cerebral palsy,” IEEE Trans.
Rehabil. Eng., vol. 52, no. 9, Sep. 2005, pp. 1532–1540.
18
1.14. D. Graupe, and H. Kordylewski, “Artificial neural network control of FES in paraplegics
for patient responsive ambulation,” IEEE Trans. Biomed Eng., vol. 42, no. 7, Jul. 1995, pp.
699-707.
1.16. A. Dutta, R. Kobetic, and R. J. Triolo, “EMG based triggering and modulation of
stimulation patterns for FES assisted ambulation – a conceptual study,” presented at XXth
Congress of the International Society of Biomechanics, Cleveland, OH, Aug. 2005.
1.18. A. Dutta, and R. J. Triolo, “Volitional surface EMG based control of FES-assisted gait after
incomplete spinal cord injury – a single case feasibility study,” presented at NIH Neural
Interfaces Workshop, Bethesda, MD, Sep. 2005.
1.20. Y. Hurmuzlu, and C. Basdogan, “On the measurement of stability in human locomotion,”
ASME Journal of Biomechanical Engineering, vol. 116, 1994, pp. 30-36.
19
CHAPTER 2
EVALUATION OF SURFACE ELECTROMYOGRAM FROM PARTIALLY PARALYZED
MUSCLES AS A COMMAND SOURCE FOR FUNCTIONAL ELECTRICAL
STIMULATION
Abstract
electrically activating the muscles of the lower extremities by exciting the peripheral motor
nerves. The FES-assisted stepping can be triggered by a manual switch or by a gait event
detector (GED). The objective of this study was to evaluate the performance of the surface
electromyogram (EMG) from partially paralyzed muscles for detecting the intent to step during
level over-ground walking. Two subjects with incomplete spinal cord injuries (iSCI) and four
able-bodied subjects volunteered for this study. Subject iSCI-1 (age 23 years, C6 ASIA C) was
non-ambulatory without the assistance of FES. Subject iSCI-2 (age 34 years, T1 ASIA D) could
walk only short distances without FES. The four able-bodied subjects, Able-1 (age 26 years),
Able-2 (age 25 years), Able-3 (age 25 years) and Able-4 (age 54 years) had no known injury or
pathology to either lower extremity during the study. Partially paralyzed muscles showed
pursuit tasks involving contraction of the muscle in a controlled fashion. Ipsilateral erector
spinae and ipsilateral medial gastrconemius consistently performed well to identify the intent to
step in both able-bodied subjects and subjects with iSCI. Left erector spinae with a mean
Discriminability Index (DI) of 0.87 for left step trigger and right erector spinae with a mean DI
of 0.83 for the right step trigger were best command sources for iSCI-1. Left erector spinae with
a mean DI of 0.93 for left step trigger and right medial gastrocnemius with a mean DI of 0.88 for
the right step trigger were the best command sources for iSCI-2.
20
Introduction
to wheelchair dependent individuals with incomplete spinal cord injuries (iSCI). FES systems
can electrically activate a customized set of muscles selected to address individual gait deficits
extremities for ambulation [2.1], [2.2]. Users normally use a switch to manually trigger each step
and progress through the customized pattern of stimulation to achieve walking function. In this
study we evaluated the controllability (the ability to volitionally modulate the surface
electromyogram (EMG) in a visual pursuit task) and discriminability (the ability to determine the
intent to step during level overground walking) of the surface EMG from both able-bodied
volunteers and individuals with iSCI. Our goal was to specify a process and criterion for
selecting two muscles for a new command and control interface that can be implemented with
two channels of implanted EMG recording electrodes with our next family of implantable
stimulator-telemeters (IST) [2.3-2.6]. This report summarizes the evaluation of the surface EMG
from partially paralyzed muscles of two subjects with iSCI and its comparison with normative
While gait event detection is possible with physical sensors such as force sensitive
resistors, accelerometers, gyroscopes [2.7], [2.8], biopotentials such as EMG can also provide
useful and reliable information [2.9-2.11]. The EMG temporally precedes the generation of force
in a muscle and the resulting movement of a joint. This makes EMG an attractive signal for
detection of intent and can allow the desired movement to be assisted by FES. Graupe and
Kordylewski presented a neural network based classifier with on-line learning capabilities for
individuals with complete paraplegia [2.11], [2.12]. Thorsen et al. showed improved wrist
extension with stimulation controlled by surface EMG from partially paralyzed wrist extensors
21
[2.13]. Futami et al. showed the feasibility of proportional control of FES with the surface EMG
from the same muscle (partially paralyzed knee extensors) in incomplete hemiplegia [2.14]. Our
preliminary study demonstrated the feasibility of FES-assisted walking triggered by the surface
EMG during double-support phase of gait (when both the feet are on ground) [2.5]. A
quantitative method is presented in this paper to evaluate the electromyogram from partially
paralyzed muscles as a command source for triggering FES-assisted steps during ambulation.
Methods
Subjects
Two male subjects with incomplete spinal cord injury (iSCI) volunteered for this study.
iSCI-1 was a 23 years old male with C7 motor and C6 sensory incomplete spinal cord injury
(ASIA C) who could stand but could not initiate a step without the assistance from FES. iSCI-2
was a 34 years old male with T1 motor and C6 sensory incomplete spinal cord injury (ASIA D)
who could walk only short distances without the assistance from FES. They each received an 8
channel Implantable Receiver Stimulator (IRS-8) and eight surgically implanted intramuscular
electrodes in a related study designed to facilitate household and limited community ambulation
[2.15]. The four able-bodied subjects, Able-1 (age 26 years), Able-2 (age 25 years), Able-3 (age
25 years) and Able-4 (age 54 years) provided the normative data for comparison. They had no
known injury or pathology to either lower extremity during the course of the study.
iliopsoas, vastus intermedius and lateralis, tensor fasciae latae, tibialis anterior, and peroneus
longus muscles. The subject iSCI-2 received stimulation electrodes only on his left side
recruiting iliopsoas, vastus intermedius and lateralis, tensor fasciae latae, gluteus medius, gluteus
maximus, posterior portion of adductor magnus, and tibialis anterior (2 electrodes). Temporal
patterns of stimulation to activate the muscles were customized for their particular gait deficits
22
according to established tuning procedures in order to achieve forward stepping in a rolling
walker [2.16], [2.17]. The subjects completed 6 weeks of over-ground gait training (2 hour
sessions, 3 times per week) with a physical therapist using the implanted FES system. After
discharge from rehabilitation, they volunteered for the studies using the myoelectric control of
Informed consent was obtained from all the subjects before their participation and all study
related procedures were approved by the Institutional Review Board of the Louis Stokes
Test of Controllability
Controllability was defined as the ability to modulate the EMG activity from one level to
another in a finite time during a visual pursuit task. The experimental setup for evaluating the
controllability of a muscle with biofeedback is shown in Figure 2.1. The surface EMG was
collected from the rectus femoris while the subject was asked to track the absolute value of a
sinusoid of amplitude 0.7 and frequency 0.01 Hz over one time-period (i.e., the TARGET signal)
during a trial. The rectus femoris was maintained in an isometric condition by Biodex System3
(Biodex Medical Systems, USA) dynamometer as shown in Figure 2.1. The EMG was pre-
preamplifier (Cambridge Electronic Design, England) before being sampled at 2200 Hz by the
The data processing and graphical display (GUI) were performed using Matlab R13 (The
MathWorks, Inc., USA) in the same PC. The EMG sampled by the data-acquisition card was
band-pass filtered (5th order zero-lag Butterworth, 20-500 Hz), de-trended and rectified before
being evaluated as a command signal (i.e., the TRACKING signal). The average EMG during
two seconds of maximum voluntary isometric contraction (MVC) was used for normalization.
23
The average magnitude of the EMG over two seconds while the subject was asked to relax the
muscle provided an estimate of the baseline activity. During visual pursuit, the estimated
baseline was subtracted from the EMG and then it was normalized by the MVC. The normalized
EMG was then divided into bins, each holding 0.1 sec of data. The TRACKING signal (i.e., the
processed EMG) pursuing the TARGET signal was updated every 0.1 sec with the average value
of the data in the latest bin only if the mean was greater by twice the standard deviation, or less
Both TARGET and TRACKING signals were projected on the wall in front of the subject
seated in the dynamometer. A set of five trials with a minimum 5 minutes of rest in between the
trials were conducted on the left and right rectus femoris of the subjects with iSCI. A set of five
trials were conducted only on the right rectus femoris of the right-handed able-bodied subjects.
The absolute value of the difference between the TARGET and TRACKING signals, the
tracking error signal ( ERROR signal = TRACKING signal - TARGET signal ), was ensemble averaged
over the set of five trials. The trial period of 100 sec was divided into four parts of 25 sec each.
The first (0-25 sec) and the third (50-75 sec) parts were the periods during which the subject was
trying to contract the muscle to catch-up with the TARGET signal. The second (25-50 sec) and
the fourth parts (75-100 sec) were the periods when the subject was trying to relax the muscle.
The mean of the absolute tracking error was computed for each of these four parts for
comparison.
Test of Discriminability
Discriminability was defined as the ability to detect the intent to step using the surface
EMG during the double-support phase of gait when both the feet are in contact with the ground.
Surface EMG signals were collected from gluteus medius (GM), biceps femoris (BF), medial
24
gastrocnemius (MG), rectus femoris (RF), tibialis anterior (TA), and erector spinae (ES at T9)
bilaterally. In case of iSCI subjects, the surface EMG was collected during switch-triggered FES-
assisted gait when each step was initiated by depression of ring-mounted finger switch. The
experimental setup is shown in Figure 2.2 where subject is walking with an implanted switch-
triggered FES-system based on an IRS-8 implanted pulse generator under the control of an
external control unit (ECU). Surface EMG was collected using Ag/AgCl electrodes with 2 cm.
inter-electrode distance following the SENIAM guidelines [2.18]. The EMG signals were
amplified and low-pass filtered (anti-aliasing, frequencycutoff=1000 Hz) by CED 1902 amplifiers
National Instruments, USA) in the host personal computer (PC). The CED 1902 amplifier has a
switching circuit (clamp) which was activated by a trigger pulse that disconnected the electrode
inputs from the amplifier and connected them to the common electrode just before the start of the
stimulation pulse. The input channels of CED 1902 were clamped this way when stimulation
pulses were applied to the muscles to prevent stimulation artifact. The gain of each channel was
set separately in the CED 1902 amplifiers to prevent saturation at the maximum muscle activity
during the gait-cycle. The implanted FES system (i.e., IRS-8) delivered electrical pulses at a
frequency of 20 Hz, so the sampled EMG was divided into bins of 50ms duration. In each bin,
30ms following the start of the stimulation pulse was blanked to remove the residual stimulation
artifact and M-wave, thus leaving signal related to voluntary muscle activity. The remaining 20
ms of data in each bin was detrended, band-pass filtered (5th order zero-lag Butterworth, 20-500
Hz), and rectified. The blanked portion of the EMG was reconstructed with the average value of
the EMG in the preceding and succeeding blocks [2.19]. Then the whole EMG pattern was low
pass filtered (5th order zero-lag Butterworth, frequencycutoff=3 Hz) to get the linear envelope. The
25
EMG pattern for each muscle was normalized by the maximum value of the EMG linear
envelope (LE) during a gait cycle. The normalized LEs during a gait cycle were then divided into
double-support and swing phase of gait based on the occurrence of foot-strike and foot-off. The
foot and ground contact sequences were determined from the insole foot switches (B&L
Engineering, USA) placed bilaterally at the medial and lateral heel, first and fifth metatarsal, and
big toe. The intent to step can be detected based on the magnitude of the LE when it crosses a
The subjects were asked to start walking after standing for 3 sec and reach a self-selected
speed within 5m from the start position. After reaching the self-selected speed the subjects had to
decelerate and return to standing. The experimental protocol is shown in Figure 2.3. The subjects
were asked to wait in terminal stance for 3 sec. The normalized LEs of each muscle were divided
into two classes: the class ‘True’ was comprised of LEs (~ 150) during double-support phase
prior to foot-off and the class ‘False’ consisted of the LEs (~150) during terminal stance and
initial standing. Half of the data were randomly allocated to training and used to find a
characteristic pattern of activation by ensemble averaging the LEs. The characteristic pattern
found for the class ‘True’ was cross-correlated with the LEs from the other half of the data (test
data) for the classes – ‘True’ and ‘False’. A Receiver Operating Characteristics (ROC) curve
shows the tradeoff between sensitivity (True Positive Rate) and 1 – specificity (False Positive
Rate) of a binary classifier [2.20]. The ROC curve was computed from the cross-correlation
coefficient (i.e., PRC for the pattern-recognition classifier) and the amplitude (i.e., TC for the
threshold-based classifier) of the LEs as the decision threshold was varied over the range of data
in the two classes, ‘True’ and ‘False’. The LEs from all the able-bodied subjects were pooled
26
together. In case of able-bodied data, the left and the right sides were considered similar and the
performance of the PRC and TC was evaluated only for the right side. The ipsilateral muscles are
the muscles of the right side and contralateral muscles are the muscles of the left side for the
classifiers (PRC and TC) trying to detect the ‘intent to step’ on the right side.
Discriminability Index (DIPRC and DITC) was defined as the area under the ROC curve
(AUC) which gave a measure of performance for the binary classifiers, PRC and TC. Bradley
showed that AUC exhibits a number of desirable properties when compared to overall accuracy
of the classifiers like increased sensitivity in Analysis of Variance (ANOVA) tests – standard
error decreased as both AUC and the number of test samples increased. AUC is also decision
threshold independent and it is invariant to a priori class probabilities [2.21]. The area under the
ROC curve was numerically computed with trapezoidal integration. Figure 2.4 illustrates the
three cases, where 0 < DI ≤ 0.5, 0.5 ≤ DI < 1, DI = 1 . We are interested in 0.5 ≤ DI < 1 such that the
mean of the ‘True’ data is greater than or equal to the mean of the ‘False’ data and the values
The data were randomly partitioned ten times into training and test data-sets for a 10-fold
cross-validation. For consistency the same training and test data-sets were used by both the
classifiers (PRC and TC) for the computation of the ROC curves in a paired experimental design.
Therefore, 10 ROC curves for each classifier were generated by randomly pooling the LEs into
training and test data-sets. The DI was computed for each ROC curve and then averaged to find
the mean (DIPRC and DITC) and standard deviation (SD(DIPRC) and SD(DITC)) for each classifier
(DIPRC for pattern-recognition classifier and DITC for the threshold-based classifier) [2.21].
Wilcoxon statistic (W) was computed as an alias of DI (i.e., a performance measure of the
classifier) to compare the two for robustness. The standard error (SE(WPRC) and SE(WTC)) was
27
computed from an approximation of the Wilcoxon statistic (WPRC and WTC) which assumes
exponential distribution of the data in the classes, ‘True’ and ‘False’. SE (W) has been shown to
W (1 − W ) + (C p − 1)(Q1 − W 2 ) + (C n − 1)(Q2 − W 2 )
SE (W ) =
C p Cn
2
W 2W
Q1 = ; Q2 =
(2 − W ) (1 + W )
Where Cp and Cn are the number of data points in the classes, ‘True’ and ‘False’
respectively.
Statistical Analysis
Inc., USA) was performed on the absolute tracking error that was obtained from the Test of
Controllability. All observations were considered to be mutually independent for the ANOVA
test. The p-value was computed for the null hypothesis that the absolute tracking error parameter
has the same mean for all the cases. If the p-value was close to zero (<0.01) then the null
hypothesis was rejected and the result was considered statistically significant.
Two-way two-tailed analysis of variance (‘anova2 in MatlabTM R14, The MathWorks, Inc.,
USA) was performed on Discriminability Index (DI) – the performance measure for the classifier
and muscle, that was obtained from the Test of Discriminability. All observations were
considered to be mutually independent for the ANOVA test. The p-value was computed for the
null hypotheses: 1 – the performance measure for the pattern-recognition classifier (PRC) and
threshold-based classifier (TC) have equal means, 2 – the performance measure for all the
muscles have equal means, 3 – there is no interactions between the classifier type and muscle
type. If the p-value was close to zero (<0.05) then that null hypothesis was rejected and the result
28
was considered statistically significant. To find which pairs were significantly different, post hoc
tests were performed with the critical values found from Scheffe’s S procedure.
ANOVA is insensitive to departures from the assumption of equal variances when the
sample sizes are equal, as in our case. Moreover, prior work has shown that ANOVA is “robust”
Results
The TRACKING signal (broken black line) using the surface EMG of rectus femoris and
the TARGET signal (solid black line) that was the absolute value of a 0.01 Hz sinusoid of
amplitude 0.7 during visual pursuit over 100 ms is shown in Figure 2.5 and 2.6. The top panel of
Figure 2.5 shows the results for the rectus femoris of the left and the right sides of iSCI-1 and the
bottom panel shows the same for iSCI-2. Figure 2.6 shows the results for surface EMG from
rectus femoris of the right side for the able-bodied subjects. The solid black line is the TARGET
signal and the broken black line shows the TRACKING signal that was ensemble averaged over
5 trials. The boxes at each data point show the lower quartile and upper quartile values of the
TRACKING signal. Whiskers extending at the top and bottom of the boxes show the range of
the TRACKING signal. Table 2.1 presents the mean, the minimum, and the maximum average
absolute tracking error during four parts (0-25 sec, 25-50 sec, 50-75 sec, 75-100 sec) of the trial.
The p-value from the one-way two-tailed ANOVA test for the average tracking error over the
whole trial (100 sec) was not statistically significant ( ≥ 0.01). This shows that all the subjects
(iSCI and able-bodied) performed similarly in the visual pursuit task for the Test of
Controllability. Individuals with iSCI were able to control the contraction of their muscles
29
The average absolute tracking error was smallest (mean = 5.48) in the first part (0-25 sec)
of the trial period, for the subjects with iSCI, corresponding to the initial period of increasing
isometric contraction. There was a slight deterioration in the performance of the iSCI subjects in
the third part of the trial, corresponding to the second period of increasing contraction (50-75
sec, mean=7.96) when compared to the first part (0-25 sec, mean=5.48). The subjects with iSCI
performed worse in the second (25-50 sec, mean=9.11) and fourth (75-100 sec, mean=10.27)
parts of the trial period, which required relaxing the muscle in a controlled fashion.
Table 2.2 shows the results from the Test of Discriminability for the muscles gluteus
medius (GM), biceps femoris (BF), medial gastrocnemius (MG), rectus femoris (RF), tibialis
anterior (TA), and erector spinae (ES at T9) for the able-bodied subjects. The Wilcoxon statistic
(W) was similar in magnitude to the corresponding Discriminability Index (DI). Similarly the
Standard Deviation (SD) of the DI over 10 random partitions (i.e., 10-fold cross-validation) was
similar in magnitude to the Standard Error (SE) found for the Wilcoxon statistic (W). There were
statistically significant (p ≥ 0.05) differences in the means of DI due to the muscle type as well as
the classifier type. The results from the post hoc analysis are presented in Figure 2.7. The top
panel of Figure 2.7 shows that the Ipsilateral MG, Ipsilateral ES, Contralateral BF, Contralateral
GM, and Contralateral TA in black markers that have performed the best (mean DI=1) as a
command source in the Test of Discriminability. The bottom panel of Figure 2.7 shows that the
Pattern Recognition Classifier (mean DIPRC=0.7586) performed much better than the Threshold-
Table 2.3a and b show the results from the Test of Discriminability of iSCI-1 for the left
step and right step classifiers respectively. The Wilcoxon statistic (W) was similar in magnitude
to the corresponding value of the Discriminability Index (DI). Similarly the Standard Deviation
30
(SD) of the DI was similar in magnitude to the Standard Error (SE) found for the Wilcoxon
statistic (W). There were statistically significant (p ≥ 0.05) differences in the means of DI due to
the muscle type as well as the classifier type. The results from post hoc analysis are presented in
Figure 2.8. The top panel first column of Figure 2.8 shows that Left ES (mean DI=0.8705) in
black marker performed the best followed by Left MG (mean DI=0.7881) in gray marker in the
Test of Discriminability for the left step. The bottom panel first column of Figure 2.8 shows that
Pattern Recognition Classifier (mean DIPRC=0.6486) in black marker performed slightly better
than the Threshold-based Classifier (mean DITC=0.6071) in gray marker in the Test of
Discriminability for the left step. The top panel second column of Figure 2.8 shows that Right ES
(mean DI=0.8299) performed the best followed by Right MG (mean DI=0.7989) in the Test of
Discriminability for the right step. The bottom panel first column of Figure 2.8 shows that
Pattern Recognition Classifier (mean DIPRC=0.6559) in black marker performed slightly better
than the Threshold-based Classifier (mean DITC=0.5886) in gray marker in the Test of
Table 2.4a and b show the results from the Test of Discriminability of iSCI-2 for the left
step and right step classifiers respectively. The Wilcoxon statistic (W) and the corresponding
value of the Discriminability Index (DI) were similar. The Standard Deviation (SD) of the DI
and the Standard Error (SE) found for the Wilcoxon statistic (W) were similar. There were
statistically significant (p ≥ 0.05) differences in the means of DI due to the muscle type as well as
the classifier type. The results from post hoc analysis are presented in Figure 2.9. The top panel
first column of Figure 2.9 shows that Left ES (mean DI=0.9293) in black marker performed the
best followed by Left MG (mean DI=0.8736) and Right RF (mean DI=0.8536 in the Test of
Discriminability for the left step. The bottom panel first column of Figure 2.9 shows that Pattern
31
Recognition Classifier (mean DIPRC=0.71856) in black marker performed much better than the
for the left step. The top panel second column of Figure 2.9 shows that Right MG (mean
DI=0.8805) performed the best followed by Right ES (mean DI=0.853) and Left RF (mean
DI=0.8484) in the Test of Discriminability for the right step. The bottom panel first column of
Figure 2.9 shows that Pattern Recognition Classifier (mean DIPRC=0.6853) in black marker
performed slightly better than the Threshold-based Classifier (mean DITC=0.54613) in black
Discussion
The controllability of the surface EMG from partially paralyzed rectus femoris of subjects
with incomplete SCI was evaluated in a visual pursuit task and found to have performance
similar (p<0.01) to able-bodied subjects (Figure 2.5 & 2.6, Table 2.1). The average absolute
tracking error was the performance measure during the visual pursuit task which was smallest in
the first part (0-25 sec) of the trial period for the subjects with iSCI. The subjects with iSCI
found it difficult to relax the muscle in a controlled fashion after the first part (0-25 sec), which
deteriorated the performance in the second part (25-50 sec) of the trial period. Subject iSCI-1
found it difficult to reach the baseline at the end of second part (25-50 sec) which deteriorated
the performance during the muscle contraction task in the third part (50-75 sec). Subject iSCI-2
could reach the baseline at the end of the second part (25-50 sec) with his right rectus femoris but
not with his left. Subject iSCI-2 had more volitional control over his right side than his left.
Overall, subjects with iSCI performed better in muscle contraction tasks (0-25 sec and 50-75 sec)
than the muscle relaxation tasks (25-50 sec and 75-100 sec). The muscles that are under
contraction and increased muscle activity during the double-support phase of the gait would be
more robust for triggering the subsequent step. The partially paralyzed muscles were evaluated
32
for classification in the Test of Discriminability, where the muscles which had 0.5 ≤ DI < 1 were
considered, such that the mean of the ‘True’ data is greater than or equal to the mean of the
‘False’ data and the values greater than the discrimination threshold were classified as ‘True’.
well in both able-bodied subjects and the subjects with iSCI. The Discriminability Index (DI)
was computed from the Area under the ROC curve which was similar to the Wilcoxon
Statististic (W) which assumes that the data in the ‘True’ and ‘False’ classes have exponential
distributions. The variability associated with DI was estimated from the Standard Error (SE)
associated with W which was similar in magnitude to the Standard Deviation (SD) computed
from 10-fold cross-validation partitions. Also, the Pattern Recognition Classifier (PRC)
consistently performed better than the Threshold-based Classifier (TC) in both the cases – able-
bodied subjects and subjects with iSCI. The basic PRC used an ensemble average of the LEs in
the ‘True’ class as the feature for pattern recognition. The PRC performance can improve with
better feature extraction using principal component analysis [2.5]. The two best muscles for the
command source and the type of the classifier were selected with post hoc analysis using the
critical values found from Scheffe’s S procedure after two-way ANOVA of the Discriminability
Index. Left ES with mean DI=0.87 for left step PRC and Right ES with mean DI=0.83 for the
right step PRC were the best command sources for iSCI-1. Left ES with mean DI=0.93 for left
step PRC and Right MG with mean DI=0.88 for the right step PRC were the best command
sources for iSCI-2. It was found that the distal muscle like MG was more susceptible to muscle
spasms in iSCI-1.
Conclusion
The goal was to define a procedure to select EMG command sources from partially
paralyzed muscles and evaluate the classifier performance prior to surgical installation of an
33
implanted stimulator-telemeter that acquire and transmit EMG information from implanted EMG
electrodes. The new family of implantable stimulator-telemeter (IST-12) has only 2 implanted
EMG channels and the capability to perform the signal processing required by the classifier
[2.6]. Partially paralyzed muscles showed controllability similar (p<0.01) to able-bodied muscles
in visual pursuit tasks involving contraction of the muscle in a controlled fashion during the Test
of Controllability. Test of discriminability was used to select two best muscles for classification
to identify the ‘intent to step.’ Bilateral ES were the best command sources for iSCI-1 and Left
ES & Right MG were the best command sources for iSCI-2. Ipsilateral ES and MG performed
the best in the Test of Discriminability for able-bodied as well as iSCI subjects.
References
2.1. Graupe, D. and K. Kohn, Functional Electrical Stimulation for Ambulation by Paraplegics.
1994: Krieger.
2.2. Kralj, A. and T. Bajd, Functional Electrical Stimulation: Standing and Walking After
Spinal Cord Injury. 1989: CRC Press.
2.4. B. Smith, Z. Tang, M.W. Johnson, S. Pourmehdi, M.M. Gazdik, J.R. Buckett, and P.H.
Peckham , “An externally powered, multichannel, implantable stimulator-telemeter for
control of paralyzed muscle,” IEEE Trans Biomed Eng., vol. 45, no. 4, 1998, pp. 463-475.
2.5. A. Dutta, R. Kobetic, and R. Triolo, “Ambulation after incomplete spinal cord injury with
EMG-triggered Functional Electrical Stimulation,” IEEE Transactions on Biomedical
Engineering, 55:2, February 2008.
2.6. N. Bhadra, K.L. Kilgore, and P.H. Peckham, “Implanted stimulators for restoration of
function in spinal cord injury,” Med. Eng. Phys., vol. 23, 2001, pp. 19-28.
2.7. I. P. Pappas, M. R. Popovic, T. Keller, V. Dietz, and M. Morari, “A reliable gait phase
detection system,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 9, no. 2, Jun. 2001, pp.
113-125.
34
2.8. A. Mansfield, and G. M. Lyons, “The use of accelerometry to detect heel contact events for
use as a sensor in FES assisted walking,” Med. Eng. Phys., vol. 25, no. 10, Dec. 2003, pp.
879-885.
2.10. R. T. Lauer, R. T. Smith, and R. R. Betz, “Application of a neuro-fuzzy network for gait
event detection using electromyography in the child with cerebral palsy,” IEEE Trans.
Rehabil. Eng., vol. 52, no. 9, Sep. 2005, pp. 1532–1540.
2.11. D. Graupe, and H. Kordylewski, “Artificial neural network control of FES in paraplegics
for patient responsive ambulation,” IEEE Trans. Biomed Eng., vol. 42, no. 7, Jul. 1995, pp.
699-707.
2.13. R. Thorsen, R. Spadone, and M. Ferrarin, “A pilot study of myoelectrically controlled FES
of upper extremity,” IEEE Trans. Rehabil. Eng., vol. 9, no. 2, June 2001, pp. 161–167.
2.14. R.Futami, K.Seki, T.Kawanishi, T.Sugiyama, I.Cikajlo, and Y.Handa, “Application of local
EMG-driven FES to incompletely paralyzed lower extremities,” presented at 10th Annual
Conference of the International FES Society, Montreal, Canada, July 2005.
2.16. R. Kobetic, and E.B. Marsolais, “Synthesis of paraplegic gait with multichannel functional
neuromuscular stimulation,” IEEE Trans Rehab Eng., vol. 2, no. 2, 1994, pp. 66-79.
2.17. R. Kobetic, R. J. Triolo, and E. B. Marsolais, “Muscle selection and walking performance
of multichannel FES systems for ambulation in paraplegia,” IEEE Trans. Rehabil. Eng.,
vol. 5, no. 1, Mar. 1997, pp. 23–29.
2.20. T. D. Dickens, Elementary Signal Detection Theory. US: Oxford University Press, 2001,
pp. 66, 121.
35
2.21. A. P. Bradley, "The use of the area under the ROC curve in the evaluation of machine
learning algorithms,"Pattern Recognition, vol. 30, no. 7, 1997, pp. 1145-1159.
2.22. J. A. Hanley and B. J. McNeil, "The meaning and use of the area under a receiver operating
characteristic (ROC) curve," Radiology, vol. 143, 1982, pp. 29-36.
2.23. R. E. Walpole and R. H. Myers, Probability and Statistic for Engineers and Scientists.
Macmillan, New York, 1990
36
Figures
Figure 2.1: Experimental setup for the Test of Controllability of the surface EMG from Rectus
Femoris using visual pursuit tasks while the knee is fixed in a dynamometer.
37
Figure 2.2: Experimental setup for surface EMG data collection with switch-triggered FES-
assisted overground walking.
38
Figure 2.3: Experimental protocol for surface EMG data collection during overground walking,
where the subject had to start from standing and achieve a self-selected gait speed
within 5m.
39
Figure 2.4: The left column shows the cumulative distribution function for the three cases,
0 < DI ≤ 0.5, 0.5 ≤ DI < 1, DI = 1 and the right column shows the corresponding Receiver
Operating Characteristics curve.
40
Figure 2.5: TRACKING (broken black line) and TARGET (solid black line) signals during
visual pursuit task for the Test of Controllability. The boxes at each data point show
the lower quartile and upper quartile values of the TRACKING signal. Whiskers
extending at the top and bottom of the boxes show the range of the TRACKING
signal. The top panel presents the results for iSCI-1 and the bottom panel for iSCI-2.
The left panel presents the results for the left Rectus Femoris and the right panel
presents the results for the right Rectus Femoris.
41
Figure 2.6: TRACKING (broken black line) and TARGET (solid black line) signals during
visual pursuit task for the Test of Controllability with able-bodied subjects. The boxes
at each data point show the lower quartile and upper quartile values of the
TRACKING signal. Whiskers extending at the top and bottom of the boxes show the
range of the TRACKING signal.
42
Figure 2.7: Top panel shows the results from the post hoc analysis of the Discriminability Index
with their critical values from Scheffe’s S procedure for the muscles – Gluteus
Medius (GM), Biceps Femoris (BF), Medial Gastrocnemius (MG), Rectus Femoris
(RF), Tibialis Anterior (TA), and Erector Spinae (ES at T9) obtained from the Test of
Discriminability with able-bodied subjects. The bottom panel shows the results from
the post hoc analysis of the Discriminability Index with their critical values from
Scheffe’s S procedure for different classifiers – Pattern Recognition Classifier (PRC)
and Threshold-based Classifier (TC) obtained from the Test of Discriminability with
able-bodied subjects.
43
Figure 2.8: Top panel shows the results from the post hoc analysis of the Discriminability Index
with their critical values from Scheffe’s S procedure for the muscles – Gluteus
Medius (GM), Biceps Femoris (BF), Medial Gastrocnemius (MG), Rectus Femoris
(RF), Tibialis Anterior (TA), and Erector Spinae (ES at T9) obtained from the Test of
Discriminability of the left and right step classifiers of iSCI-1. The bottom panel
shows the results from the post hoc analysis of the Discriminability Index with their
critical values from Scheffe’s S procedure for different classifiers – Pattern
Recognition Classifier (PRC) and Threshold-based Classifier (TC) obtained from the
Test of Discriminability of the left and the right step classifiers of iSCI-1.
44
Figure 2.9: Top panel shows the results from the post hoc analysis of the Discriminability Index
with their critical values from Scheffe’s S procedure for the muscles – Gluteus
Medius (GM), Biceps Femoris (BF), Medial Gastrocnemius (MG), Rectus Femoris
(RF), Tibialis Anterior (TA), and Erector Spinae (ES at T9) obtained from the Test of
Discriminability of the left and right step classifiers of iSCI-2. The bottom panel
shows the results from the post hoc analysis of the Discriminability Index with their
critical values from Scheffe’s S procedure for different classifiers – Pattern
Recognition Classifier (PRC) and Threshold-based Classifier (TC) obtained from the
Test of Discriminability of the left and the right step classifiers of iSCI-2.
45
Tables
Table 2.1: The mean, the minimum, and the maximum average absolute tracking error in %MVC
during the four parts (0-25 sec, 25-50 sec, 50-75 sec, 75-100 sec) of the Test for
Controllability. The p-value from the one-way two-tailed ANOVA test for the
average tracking error over the whole trial (100 sec) was not ‘statistically significant’
( ≥ 0.01).
46
Table 2.2: The results from the Test of Discriminability for the muscles Gluteus Medius (GM),
Biceps Femoris (BF), Medial Gastrocnemius (MG), Rectus Femoris (RF), Tibialis
Anterior (TA), and Erector Spinae (ES at T9) are presented for the able-bodied
subjects. The Wilcoxon statistic (W) was similar in magnitude to the corresponding
Discriminability Index (DI). Similarly the Standard Deviation (SD) of the DI over 10
random partitions (i.e., 10-fold cross-validation) was similar in magnitude to the
Standard Error (SE) found for the Wilcoxon statistic (W). There were ‘statistically
significant’ (p ≥ 0.05) differences in the means of DI due to the muscle type as well as
the classifier type.
47
Table 2.3a: The results from the Test of Discriminability of iSCI-1 for the left step classifier. The
Wilcoxon statistic (W) was similar in magnitude to the corresponding value of the
Discriminability Index (DI). Similarly the Standard Deviation (SD) of the DI was
similar in magnitude to the Standard Error (SE) found for the Wilcoxon statistic (W).
There were ‘statistically significant’ (p ≥ 0.05) differences in the means of DI due to
the muscle type as well as the classifier type.
Muscles for left step SD SE SD SE
classifier of iSCI-1 DIPRC (DIPRC) WPRC (WPRC) DITC (DITC) WTC (WTC)
Left GM 0.77 0.06 0.78 0.06 0.66 0.06 0.67 0.06
Left BF 0.56 0.03 0.56 0.03 0.57 0.05 0.58 0.05
Left MG 0.81 0.02 0.81 0.02 0.77 0.06 0.77 0.06
Left RF 0.72 0.04 0.73 0.04 0.57 0.05 0.57 0.05
Left TA 0.63 0.08 0.64 0.08 0.58 0.06 0.58 0.06
Left ES 0.96 0.04 0.96 0.038 0.78 0.06 0.79 0.06
Right GM 0.5 0.04 0.5 0.039 0.53 0.03 0.55 0.03
Right BF 0.56 0.04 0.57 0.036 0.56 0.05 0.56 0.05
Right MG 0.58 0.04 0.59 0.039 0.56 0.06 0.56 0.06
Right RF 0.65 0.08 0.67 0.075 0.53 0.06 0.54 0.05
Right TA 0.56 0.05 0.56 0.045 0.55 0.06 0.56 0.06
Right ES 0.51 0.06 0.51 0.057 0.60 0.05 0.61 0.05
Table 2.3b: The results from the Test of Discriminability of iSCI-1 for the right step. The
Wilcoxon statistic (W) was similar in magnitude to the corresponding value of the
Discriminability Index (DI). Similarly the Standard Deviation (SD) of the DI was
similar in magnitude to the Standard Error (SE) found for the Wilcoxon statistic (W).
There were ‘statistically significant’ (p ≥ 0.05) differences in the means of DI due to
the muscle type as well as the classifier type.
48
Table 2.4a: The results from the Test of Discriminability of iSCI-2 for the left step. The
Wilcoxon statistic (W) and the corresponding value of the Discriminability Index
(DI) were similar. The Standard Deviation (SD) of the DI and the Standard Error (SE)
found for the Wilcoxon statistic (W) were similar. There were ‘statistically
significant’ (p ≥ 0.05) differences in the means of DI due to the muscle type as well as
the classifier type.
Table 2.4 b: The results from the Test of Discriminability of iSCI-2 for the right step classifier.
The Wilcoxon statistic (W) and the corresponding value of the Discriminability Index
(DI) were similar. The Standard Deviation (SD) of the DI and the Standard Error (SE)
found for the Wilcoxon statistic (W) were similar. There were ‘statistically
significant’ (p ≥ 0.05) differences in the means of DI due to the muscle type as well as
the classifier type.
49
CHAPTER 3
FEASIBILITY ANALYSIS OF SURFACE EMG-TRIGGERED FES-ASSISTED
AMBULATION AFTER INCOMPLETE SPINAL CORD INJURY
A part of this chapter was published in IEEE Trans Biomed Eng. 2008 Feb; 55(2).
Abstract
Ambulation after spinal cord injury is possible with the aid of functional electrical
stimulation (FES). Individuals with incomplete spinal cord injury (iSCI) retain partial volitional
control of muscles below the level of injury, necessitating careful integration of FES with intact
voluntary motor function for efficient walking. In this study, the surface electromyogram (EMG)
of the volitionally controlled erector spinae was used to detect the intent to step and trigger FES-
assisted walking in a volunteer with iSCI via an 8-channel implanted stimulation system. The
classifier was able to trigger the FES-assisted swing-phase of gait with a false positive rate less
than 1% and true positive rate greater than 82% during over-ground ambulation on a level
surface. The performance of the EMG classifier highlights its potential as a natural command
interface to better coordinate stimulated and volitional muscle activities than conventional
Introduction
to wheelchair dependent individuals with incomplete spinal cord injuries (iSCI). The implanted
FES systems can electrically activate a customized set of muscles selected to address individual
gait deficits with pre-programmed patterns of stimulation to produce cyclic movement of the
lower extremities for ambulation. Our 8-channel implantable receiver-stimulator (IRS-8) delivers
stimulation via implanted electrodes to the targeted motor nerves activating the muscles required
to produce stepping motions [3.1]. Power and stimulus control information are transmitted to the
implanted receiver stimulator through the skin via an inductive link by a wearable external
50
control unit (ECU). Implant recipients normally use a ring-mounted thumb switch connected to
the ECU to manually trigger each step and progress through the customized pattern of
stimulation to achieve walking function. This study was undertaken to evaluate the potential for
better coordinating the actions of the stimulator with remaining volitional movements through a
The objective of this study was to evaluate the feasibility of detecting the intent to take a
step using the surface electromyogram (EMG) in an implant recipient with iSCI and eliminate
the need for manual triggering during FES-assisted ambulation. Our long-term goal is to specify
a new command and control interface that can be implemented with two channels of implanted
EMG recording electrodes with our next family of multichannel implantable stimulator-
telemeters (IST) [3.2-3.4]. This report summarizes the development and testing of a new
command structure for FES-assisted ambulation that integrates stimulated and voluntary muscle
activity in a method suitable for eventual realization in a fully implantable neuroprosthesis for
walking after iSCI. While gait event detection is possible with physical sensors such as force
electroneurogram [3.5-3.7], the EMG can provide useful and reliable information prior to the gait
events during ambulation [3.8]. The electromyogram temporally precedes the generation of force
in a muscle (electromechanical delay) and resulting movement of a joint. This makes the EMG
an attractive signal for detection of intent and then the desired movement can be assisted by FES.
Graupe et al. proposed EMG based control of FES with time-series model of EMG [3.9].
Auto-regressive (AR) time-series models of EMG can give information related to limb function
[3.10] and can provide patient-responsive control of FES-assisted walking [3.11]. The AR
parameters determined offline from the time-series model of the above-lesion upper-trunk EMG
51
were used for function identification [3.11]. Recently, Graupe and Kordylewski presented a
neural network based classifier for complete paraplegics with on-line learning capabilities [3.12,
3.13]. Our preliminary studies indicate that the EMG from partially paralyzed muscles in
incomplete paraplegics can also provide significant information related to the volitional activity
of the muscle and can be used for gait-event detection [3.14, and 3.15]. Thorsen et al. have
showed improved wrist extension with stimulation controlled by EMG from partially paralyzed
wrist extensor [3.16]. Futami et al. showed the feasibility of proportional control of FES with the
EMG from the same muscle (partially paralyzed knee extensors) in incomplete hemiplegia
[3.17]. The ratio of the volitional EMG with the intended effort changes with the intensity of the
stimulation and the time since the stimulation pulse, when the EMG from the same muscle is the
This paper presents a linear classifier for foot-off intent detection with EMG from two
partially paralyzed muscles below the level of injury. The classifier detects the intent to initiate
swing-phase and integrates that information in the FES controller to trigger FES-assisted swing
of the limb.
Methods
Subjects
The subject iSCI-1 was a 23 years old male with C7 motor and C6 sensory incomplete
spinal cord injury (ASIA C) who could stand but could not initiate a step without the assistance
of FES. About eight months before the start of this study, he received an IRS-8 and eight
and community ambulation [3.19]. Electrodes were implanted bilaterally at the lumbar spinal
roots to activate iliopsoas for hip flexion, in the tensor fasciae latae for hip flexion and abduction,
vastus intermedius/lateralis for knee extension, and tibialis anterior and peroneous longus for
52
ankle dorsiflexion. Temporal patterns of stimulation to activate those muscles were customized
for his particular gait deficits according to established tuning procedures [3.20, and 3.21] in order
to achieve forward stepping in a rolling walker. Each step was triggered by depression of a
manual switch to progress through the stimulation pattern and initiate FES-assisted swing phase,
as shown in Figure 3.1. The limited range of motion of his fingers made it difficult to operate the
standard manual ring-mounted thumb switch. Switches were adapted and mounted to the frame
The subject completed 6 weeks of over-ground gait training (2 hour sessions 3 times per
week) with a physical therapist using the implanted FES system. After discharge from
subject volunteered for studies related to the myoelectric control of the FES system.
Informed consent was obtained from the subject before his participation and all study
related procedures were approved by the Institutional Review Board of the Louis Stokes
The experimental set-up for EMG-triggered FES-assisted walking is shown in Figure 3.2.
Surface EMG signals were collected from gluteus medius, biceps femoris, medial gastrocnemius,
rectus femoris, tibialis anterior, and erector spinae (at T9) bilaterally during manually-triggered
FES-assisted gait. Surface EMG was collected using Ag/AgCl electrodes with 2 cm. inter-
electrode distance following the SENIAM guidelines [3.22]. The EMG signals were amplified
and low-pass (fcutoff=1000 Hz) filtered by CED 1902 amplifiers (Cambridge Electronic Design,
England) before being sampled at 2400 Hz (AT-MIO-64F-5, National Instruments, USA) in the
host personal computer. The CED 1902 amplifier has a switching circuit (clamp) which is
activated by a pulse that disconnects the electrode inputs from the amplifier and connects it to the
53
common electrode to prevent stimulation artifact. The input channels of CED 1902 were
clamped when stimulation pulses were applied to the muscles to prevent stimulation artifact as
described below and illustrated in Figure 3.3. The gain of each channel was set separately in the
CED 1902 amplifiers to prevent saturation at the maximum muscle activity during the gait-cycle.
Baseline EMG data were collected during trials of quiet standing for 3 seconds.
Retro-reflective markers were placed on the body segments according to the ‘plug-in’ gait
marker set in the Vicon WorkstationTM (Vicon Peak, USA) software to acquire lower-body
kinematics data using a seven camera ViconTM (Vicon Peak, USA) motion capture system. The
fifteen markers were placed on the anterior superior iliac spine, thigh (on a wand at 1/3rd distance
between the greater trochanter and lateral femoral epicondyle), lateral femoral epicondyle, shank
(on a wand at 1/3rd distance between the lateral femoral epicondyle and lateral malleolus), lateral
malleolus, toe (dorsum of the foot over second metatarsal head), heel (inline with the toe marker
on the long axis of the foot) bilaterally, and at the sacrum (midway between the posterior
superior iliac spines). The measurement volume of the motion capture system was approximately
5m from the starting position on the walkway, ensuring collection of steady state walking data.
The marker trajectories were captured at 60 Hz for the computation of the joint angles. Gait
events (foot-strike and foot-off) were derived from foot-floor contact patterns obtained from
insole-mounted foot switches placed bilaterally at medial and lateral heel, first and fifth
metatarsal, and big toe, and confirmed with the kinematics data acquired.
The subject was asked to make multiple passes across the straight level walkway. Each
pass constituted one trial and multiple trials were collected during a session. The EMG data
collection was evenly spread over multiple sessions over a month to capture the day-to-day
variability. 150 steps for each side (total 300 steps) were captured over multiple sessions during
54
this period. The sampled EMG and joint kinematics were processed in MATLAB TM R14 (The
The implanted FES system delivered electrical pulses at a frequency of 20 Hz, so the
sampled EMG was divided into bins of 50ms duration in the host personal computer. In each bin,
30ms following the start of the stimulation pulse was blanked to remove residual stimulation
artifact and M-wave. The remaining 20 ms of data in each bin was detrended, band-pass filtered
(5th order zero-lag Butterworth, 20-500 Hz), and rectified. The blanked portion of the EMG was
reconstructed with the average value of the EMG in the pre- and post- blocks [3.23] as shown in
Figure 3.3a. Then the whole EMG pattern was low pass filtered (5th order zero-lag Butterworth,
fcutoff=3 Hz) to get the linear envelope (Figure 3.3b). The EMG linear envelopes during a gait
cycle were then divided into double-support and swing phases of left and the right side based on
the occurrence of foot-strike and foot-off as determined from the insole foot switch data. The
EMG pattern for each muscle was normalized by the maximum value of its linear envelope.
Muscle Selection
The next family of implantable stimulator-telemeter (IST) developed and being tested
clinically at our center has 2 channels of implanted EMG signal acquisition [3.4]. Therefore, the
number of muscles was limited to 2 for this feasibility study. The muscles were selected based
on the difference between their EMG linear envelopes during double-support phase when
The processed linear envelopes (LE) for the EMG from each muscle pair were divided into
two classes: the class ‘True’ was comprised of LEs during double-support prior to foot-off (150
EMG patterns) and the class ‘False’ consisted of the LE during terminal stance (150 EMG
patterns). The muscle selection was performed separately, prior to classifier development to limit
the size of the data (each muscle has two classes and each class has 150 EMG patterns). The
55
characteristic pattern of activation for each class was found by ensemble averaging the LE,
shown in Figures 3.4a, b. The characteristic pattern found for class ‘True’ was cross-correlated
with the LE patterns for classes ‘True’ and ‘False’. The cross-correlation coefficients for both
the classes had a normal distribution and passed the Lilliefors test for normality (α=0.05;
MatlabTM function ‘lillietest’) [3.24]. A normal distribution function was estimated for each class
using minimum variance unbiased estimator (MatlabTM ‘normfit’ function). The Receiver
Operating Characteristics (ROC) curve shows the trade-off between selectivity (i.e. false positive
rate) and sensitivity (i.e. true positive rate) as the threshold is varied from one extreme to the
other to cover the probability density function of both (‘True’ and ‘False) the classes [3.25]. A
statistic, the Discriminability Index (DI), was computed from the area under the ROC curve
[3.25]. DI equal to 1 (maximum) is the best case when the probability density function for the
two classes has no overlap. The bilateral Erector Spinae (ES) exhibited the highest DI (=0.875)
The LE of EMG from bilateral ES was selected as inputs to the classifiers to identify the
intent to initiate swing-phase of the ipsilateral limb during FES-assisted gait. The LE of EMG
from each ES for the classes ‘True’ (150 patterns) and ‘False’ (150 patterns) were pooled
together in a single dataset. The first step was to find a minimum set of uncorrelated patterns
(vectors) such that their linear combinations accounted for most of the variance in the mean-
adjusted dataset. This was achieved by principal component analysis which seeks such
orthonormal vectors (principal components) and their associated eigen-values from the
covariance matrix of the mean-adjusted data [3.26]. The first 4 principal components (PCs)
ranked by their eigen-values accounted for more than 90% of the variance in the data. The
56
orthogonal rotation (MATLABTM ‘varimax’) was applied to minimize the number of factors (i.e.
PCs after rotation) and increase the loading on fewer factors for the classes ‘True’ and ‘False’
[3.26]. Loadings found for each LE of EMG represent a point in the space defined by the four
4
th
e.g. for the i Linear Envelope, LEi = ∑ loadingi , j × factorj ,
j =1
loadingi , j (j=1,2,3, and 4) serve as the coordinates of the LEi (i=1,2,3 ... 300)
in the feature space defined by the vectors, factorj (j=1,2,3, and 4)
The loadings on the factors after the ‘varimax’ rotation created separate clusters of points
for the classes in the feature space. The loadings found for each LE of EMG were normalized by
the square-root of the sum of the squared loadings to define a unit vector from the origin to each
point. The mean of all the points in the cluster was computed for each class and called the
centroid of that class. The classifier estimated the factor loadings for LE of a candidate EMG to
detect the intent (i.e. case ‘True’). The factor loading was estimated for each feature from the dot
product between the factor (i.e. feature) and the mean-adjusted LE of the candidate EMG. The
Euclidean distance between the normalized loading and the centroid of the case ‘True’ was used
to detect the intent. An appropriate threshold was selected for this distance measure based on the
ROC plot shown in Figure 3.5. The classification threshold was specified such that the false
positive rate was below a reasonable value of 2% found based on our preliminary study [3.14].
The true positive rate was around 82% for the same threshold, as seen in Figure 3.5.
The classifier was implemented in SimulinkTM (The MathWorks, Inc., USA) and
incorporated into the real-time FES control system. The ECU was controlled in real-time by a
dedicated personal computer running xPC-targetTM (The MathWorks, Inc., USA) that executed
the control algorithm to deliver pre-programmed stimulation patterns to the muscles via the IRS-
8. The classifier started scanning the LE of EMG from ES after the end of the stimulation pattern
57
(swing-phase) of the contralateral limb to detect the intent (foot-off) to initiate swing of the
ipsilateral limb. The classifier was first evaluated offline using test data to find the error in
timing of the FES-assisted foot-off with EMG control compared to that found from the insole
foot switches during manual switch-triggering. For safety, the classifier was allowed to trigger
the FES-assisted swing-phase of the ipsilateral limb 1.5 sec after the end of the swing-phase of
the contralateral limb. A minimum duration of 1.5 seconds for the double-support phase was
found necessary for the subject to maintain balance The duration of swing phase was determined
by the pre-programmed temporal pattern of the stimulation and therefore was highly repeatable
while the duration of the double-support phase was determined by the subject with the manual
The EMG classifier was integrated into the FES controller to trigger pre-programmed
stimulation patterns during FES-assisted walker-aided ambulation. The subject was first asked to
walk a few steps using the manual switch-triggered FES system for calibration purposes. During
calibration the classifier found the maximum muscle activity in ES, which was used to normalize
The subject practiced with the EMG-based FES-controller for about 9 days (2 hours each
day) over 3 weeks to get accustomed to the classifier before initial testing. During testing, the
subject was asked to walk across a straight walkway with the EMG-triggered FES-controller.
Multiple passes were made across the walkway, each pass constituted a trial. The subject
triggered the first step manually from the ‘stand’ state using a start switch after which the
subsequent steps were triggered by the EMG classifier. The FES controller returned to the
‘stand’ state when the classifier failed to trigger the next step within 3 seconds of entering the
double support phase. The subject used a switch to manually stop the FES controller if it
58
triggered a step when none was intended (false positive). If the FES controller stopped and failed
to trigger an intended step (false negative) then the subject was able to over ride the controller to
manually trigger the step, after which the EMG classifier would resume operation. The state
During each trial the subject tried to reach a self-selected steady gait from stand position.
The subject maintained a double-support phase after reaching the end of the walkway and waited
for 3 sec. to return the FES controller to the ‘stand’ state. The performance of the classifier was
evaluated over 60 steps (each side) on level ground in a single session (1 day) of data collection.
The repeatability of system performance was evaluated with randomized trials of EMG-triggered
and switch-triggered FES-systems performed over 2 more sessions distributed over 2 additional
days of testing.
Results
Classifier Performance
Off-line timing analysis indicated that with the selected threshold, the classifier output
preceded actual foot-off (a negative timing difference) when predicting both left and right swing
as shown in Figure 3.7a. The duration of the double-support (DS) and swing (SW) gait phases is
shown in Figure 3.7b. The classifier successfully minimized the possibility of triggering a step
unexpectedly. During the first session of testing, the false positive rate was limited to 1.66%
while the true positive rate was close to 80%. The one false positive observed for the right leg
over 60 steps was during terminal stance. The false negatives were mostly observed during the
first 3 steps (each side) when the gait is transitioning from stand position to a steady state. The
user learned to convert the last step of the walking pattern into a short step to decelerate, by not
moving his body forward with the walker and then stop the classifier by delaying the shift of the
59
body weight to the contralateral support limb by 3 seconds after the end of its swing phase. This
helped to prevent false positives during terminal stance in the subsequent trials.
Repeatability of system performance was assessed with data collected during the two
additional sessions in which switch- and EMG-triggered controllers were presented randomly to
the subject. No false positives were reported by the subject during more than 50 steps taken at
these 2 sessions. When combined with data from the initial testing day, the system exhibited
overall false positive rate of less than 1% and true positive rate of 82% for left foot-off and 83%
for right foot-off over total of 110 steps taken (total 3 sessions over 3 days).
Discussion
in the laboratory. The classifier developed during current study detected unloading of the limb
and shift of body-weight to the contralateral limb that took place during double-support phase
preceding the swing-phase of the ipsilateral side. The user of the EMG-triggered FES-system
exercised the voluntary muscles in conjunction with the stimulated muscles. The EMG-triggered
implanted FES system reduced the need to manipulate manual switches which triggered each
step with current FES systems, or make special adaptations to them for the specific motor
deficits of individuals with cervical level injuries. The subject easily learned to use the system
and their practice with EMG-triggered system improved the switch-triggered walking. This
interesting result associated with learning was a reduction in the duration of the double-support
phase with switch-triggered walking from around 4 seconds (Figure 3.6b) during training data
collection to around 2 seconds during online testing sessions when the switch-triggered system
was presented in a random order with the EMG-triggered system. The gait speed was limited by
the ability to move the body forward using upper-body support while maintaining balance. The
60
gait-speed was similar for EMG-triggered and switch-triggered walking. The gait speed that was
close to 0.1 m/sec was significantly lower than able-bodied slow-cadence gait, in part due to the
subject’s cervical level injury which limited his ability to move the body forward with his upper
extremities. This necessitated a minimum duration of 1.5 sec for double-support phase to prepare
Proximal muscles were found to be more suitable as command sources than the distal
muscles. Proximal muscles were also less susceptible to muscle spasms. The classifier was
considered to be more robust to muscle spasms and fatigue since it was based on pattern
recognition instead of EMG amplitude threshold-based. The characteristic EMG pattern from the
ES that was used by the classifier for event detection was similar to normative EMG pattern
during late stance and early swing of the ipsilateral limb [3.27].
The classifier was not tested specifically for robustness to muscle spasms and fatigue. The
number of trials during a session was decided by the subject. The erector spinae was free from
muscle spasms during all the trials. In case of muscle spasms and fatigue, the user could
manually actuate the stop switch to return the controller to the ‘stand’ state. This would switch-
off the classifier and prevent false triggers. The potential exists to make the classifier adaptive by
updating it during operation (online training) using the EMG patterns related to false positives
and false negatives, thus having it learn to accommodate for fatigue or other time varying factors
The classifier needed to be customized which involved individualized muscle selection and
feature extraction from EMG training data of the main muscle groups (extensors and flexors)
involved with gait. Training data was collected during switch-triggered FES-assisted walking in
the laboratory utilizing surface stimulation and EMG recording, or with the EMG added to the
61
systems of other recipients of our older 8-channel stimulation-only implanted neuroprostheses.
The processes of muscle selection and feature extraction also lend themselves to automation,
thus eliminating much of the clinical decision making currently required. One limitation of
selecting muscles based on the EMG and surface stimulation is that the optimal set of command
sources and motor targets will be determined only from the most superficial muscles accessible
from the skin’s surface. Deeper muscles can be accessed with intramuscular stimulating or
Conclusion
The feasibility of using EMG for triggering FES-assisted ambulation on level ground for
individuals with iSCI was demonstrated. An EMG-based trigger for swing-phase significantly
reduced the need to rely on manual switch during FES-assisted iSCI gait, which was the typical
command source in the currently available implanted neuroprostheses for ambulation. The
classifier did not completely obviate the need for the manual trigger, since about 18% of the false
negatives produced by the classifier still needed to be over-ridden with the manual switch during
testing sessions. The false positives were successfully reduced to 1% minimizing the likelihood
of potentially more disruptive event of initiating a step unexpectedly. Two channels of surface
EMG were shown to be sufficient for triggering steps during FES-assisted walking. The subject
learned to effectively command the EMG-based FES-controller within about 18 hours of use
over 9 days. The FES-system used by the subject for this feasibility study was a laboratory based
tethered system. The new family of implantable stimulator-telemeters (IST-12) has 2 EMG
channels and the capability to perform the signal processing required by the classifier. This study
provided the basis for the development of a fully portable FES-system based on IST-12. The next
phase of this study will focus on the development of a classifier and the FES–controller that
could be implemented in the current ECU with the IST family of implants for a fully portable
62
FES-system for outdoor use. It is postulated that incomplete SCI subjects with reduced hand and
finger function will prefer the reduced dependence on manual switch and accept automatic FES-
References
3.2. B. Smith, Z. Tang, M.W. Johnson, S. Pourmehdi, M.M. Gazdik, J.R. Buckett, and P.H.
Peckham , “An externally powered, multichannel, implantable stimulator-telemeter for
control of paralyzed muscle,” IEEE Trans Biomed Eng., vol. 45, no. 4, 1998, pp. 463-475.
3.3. Z. Tang, B. Smith, J.H. Schild, and P.H. Peckham, “Data transmission from an implantable
biotelemeter by load-shift keying using circuit configuration modulator,” IEEE Trans
Biomed Eng., vol. 42, no. 5, 1995, pp. 525-528.
3.4. N. Bhadra, K.L. Kilgore, and P.H. Peckham, “Implanted stimulators for restoration of
function in spinal cord injury,” Med. Eng. Phys., vol. 23, 2001, pp. 19-28.
3.5. I. P. Pappas, M. R. Popovic, T. Keller, V. Dietz, and M. Morari, “A reliable gait phase
detection system,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 9, no. 2, Jun. 2001, pp.
113-125.
3.6. A. Mansfield, and G. M. Lyons, “The use of accelerometry to detect heel contact events for
use as a sensor in FES assisted walking,” Med. Eng. Phys., vol. 25, no. 10, Dec. 2003, pp.
879-885.
3.8. R. T. Lauer, R. T. Smith, and R. R. Betz, “Application of a neuro-fuzzy network for gait
event detection using electromyography in the child with cerebral palsy,” IEEE Trans.
Rehabil. Eng., vol. 52, no. 9, Sep. 2005, pp. 1532–1540.
63
3.11. D. Graupe, “EMG pattern analysis for patient-responsive control of FES in paraplegics for
walker-supported walking,” IEEE Trans. Biomed. Eng., vol. 36, no. 7, July 1989, pp. 711-
719.
3.12. D. Graupe, and H. Kordylewski, “Artificial neural network control of FES in paraplegics
for patient responsive ambulation,” IEEE Trans. Biomed Eng., vol. 42, no. 7, Jul. 1995, pp.
699-707.
3.14. A. Dutta, R. Kobetic, and R. J. Triolo, “EMG based triggering and modulation of
stimulation patterns for FES assisted ambulation – a conceptual study,” presented at XXth
Congress of the International Society of Biomechanics, Cleveland, OH, Aug. 2005.
3.15. A. Dutta, and R. J. Triolo, “Volitional surface EMG based control of FES-assisted gait after
incomplete spinal cord injury – a single case feasibility study,” presented at NIH Neural
Interfaces Workshop, Bethesda, MD, Sep. 2005.
3.16. R. Thorsen, R. Spadone, and M. Ferrarin, “A pilot study of myoelectrically controlled FES
of upper extremity,” IEEE Trans. Rehabil. Eng., vol. 9, no. 2, June 2001, pp. 161–167.
3.17. R.Futami, K.Seki, T.Kawanishi, T.Sugiyama, I.Cikajlo, and Y.Handa, “Application of local
EMG-driven FES to incompletely paralyzed lower extremities,” presented at 10th Annual
Conference of the International FES Society, Montreal, Canada, July 2005.
3.18. P. N. Taylor, and P. H. Chappell, “Variation in system gain when using voluntary EMG to
control electrical stimulation of the same muscle,” presented at 9th Annual Conference of
the International FES Society, Bournemouth, U.K., 2004.
3.20. R. Kobetic, and E.B. Marsolais, “Synthesis of paraplegic gait with multichannel functional
neuromuscular stimulation,” IEEE Trans Rehab Eng., vol. 2, no. 2, 1994, pp. 66-79.
3.21. R. Kobetic, R. J. Triolo, and E. B. Marsolais, “Muscle selection and walking performance
of multichannel FES systems for ambulation in paraplegia,” IEEE Trans. Rehabil. Eng.,
vol. 5, no. 1, Mar. 1997, pp. 23–29.
64
3.23. A. E. Hines, P. E. Crago, G. J. Chapman, and C. Billian, “Stimulus artifact removal in
EMG from muscles adjacent to stimulated muscles,” J. Neurosci. Methods, vol. 64, no. 1,
Jan. 1996, pp. 55-62.
3.24. H. W. Lilliefors, “On the Kolmogorov-Smirnov test for normality with mean and variance
unknown,” J. of Am. Stat. Assoc., vol. 62, 1967, pp. 339-402.
3.25. T. D. Dickens, Elementary Signal Detection Theory. US: Oxford University Press, 2001,
pp. 66, 121.
3.26. I. T. Jolliffe, Principal Component Analysis, New York: Springer, 1986, pp. 44-48, 269-
296.
3.27. D. A. Winter, Biomechanics and Motor Control of Human Gait: Normal, Elderly and
Pathological. Waterloo Biomechanics, 1991, ch.1.
3.28. J. B. Dingwell, B. L. Davis, and D. M. Frazier, “Use of an instrumented treadmill for real-
time gait symmetry evaluation and feedback,” Prosthet. Orthot. Int., vol. 20, no. 2, Aug.
1996, pp. 101-110.
65
Figures
a b
Figure 3.1: a) X-ray of the iSCI subject implanted with implantable receiver-stimulator (IRS-8)
b) iSCI subject stepping with the switch-triggered FES system
66
Figure 3.2: Experimental setup for testing EMG-triggered FES-assisted walking with the block-
diagram for the EMG-triggered FES-system (ECU: external control unit, LE: linear
envelope)
67
a
Figure 3.3: Processing of the sampled EMG from Erector Spinae for training the classifier a)
rectified and reconstructed EMG signal b) linear envelope found from processed
EMG signal
68
a
Figure 3.4: Muscle selection for the classifier using receiver operating characteristics curve from
switch-triggered FES-assisted gait data (FS: Foot-Strike, FO: Foot-Off) a) linear
envelope (LE) indicating class ‘True’ b) linear envelope (LE) indicating class ‘False’
69
Figure 3.5: Receiver operating characteristics curve of the classifiers using the test data
70
EMG during Feature Set
double-support & Threshold
Loadings
Find the distance from centroid
of case ‘T’ in feature space
Euclidean distance
Distance Y Trigger
< swing-
threshold phase
STOP
Double- Y Wait for
support manual
> 3 sec? trigger
N
Get next
EMG sample
71
a
Figure 3.7: Offline testing of the classifier using receiver operating characteristics curve a) time-
error (negative means prediction) in detection of foot-off by the classifier b) duration
of the gait phases (Left DS: double support phase following left swing phase, Right
DS: double support phase following right swing phase, SW: swing phase)
72
CHAPTER 4
SURFACE EMG-TRIGGERED FES-ASSISTED GAIT PARAMETERS DURING OVER-
GROUND WALKING IN THE LABORATORY
Abstract
activating the muscles of the lower extremities. The FES-assisted stepping can either be triggered
trigger). The two command sources were presented in random order to two subjects with
incomplete spinal cord injuries (iSCI) during ten trials over two alternate days. Subject iSCI-1
(C6 ASIA C) was non-ambulatory without the assistance of FES and could stand but not initiate
a step volitionally. Subject iSCI-2 (T1 ASIA D) could walk only short distances with great
difficulty without FES. Gait kinematics were captured during FES-assisted over-ground walking
with a rolling walker under laboratory conditions. Basic parameters of gait like speed, left and
right step length, left and right double support duration, left and right swing phase duration were
extracted from the kinematics data. Mean, standard deviation, coefficient of variation, and 95%
confidence interval were computed for each gait parameter under each triggering condition.
Average body weight support provided by an instrumented walker was also recorded for iSCI-2.
One way analysis of variance (ANOVA) was performed on each gait parameter to determine
whether significant differences existed between command sources. The left and right double
support duration was significantly (p<0.05) lower during EMG-triggered gait than switch-
triggered for iSCI-1. The average body weight support from the walker was significantly
(p<0.05) higher during switch-triggered gait than EMG-triggered one for iSCI-2
73
Introduction
individuals with spinal cord injuries (SCI) to achieve brace-free ambulation. Implanted FES
systems have allowed individuals with either complete or partial paralysis to stand and step via
activation of a customized set of muscles selected to address individual gait deficits with pre-
programmed patterns of stimulation to produce cyclic movement of the lower extremities [4.1],
[4.2]. The 8-channel implantable receiver-stimulator (IRS-8) utilized in these systems delivered
stimulation via implanted electrodes to the targeted motor nerves activating the muscles required
for walking [4.3]. Implant recipients used a ring-mounted thumb switch connected to a wearable
external control unit to manually trigger each step and activate a customized pattern of
stimulation to achieve reciprocal walking. Our studies have indicated that surface EMG from
partially paralyzed muscles in people with paraplegia due to incomplete spinal cord injuries
(iSCI) can provide significant information related to the volitional activity of the muscle and can
be used for gait-event detection [4.4]. While the electromyogram showed promise of being a
feasible and more natural command source than manual switches for triggering the FES-assisted
stepping, its benefits for improving walking function remained to be investigated [4.4].
The objective of the study was to evaluate the relative benefits of switch-triggered and
parameters. The basic gait parameters that were investigated were gait speed, left and right step
length, left and right double support duration, left and right swing phase duration [4.5]. In
addition, upper extremity exertion in terms of the body weight placed on the rolling walker was
also measured. One way analysis of variance (ANOVA) was performed on each gait parameter
to find if significant differences exist between the EMG-triggered and switch-triggered over-
74
Methods
Subjects
Two male subjects with incomplete spinal cord injury volunteered for this study. The
subject iSCI-1 was a 23 years old male with C7 motor and C6 sensory incomplete spinal cord
injury (ASIA C) who could stand but could not initiate a step without the assistance from FES.
The subject iSCI-2 was a 34 years old male with T1 motor and C6 sensory incomplete spinal
cord injury (ASIA grade D) who could walk only short distances without the assistance from
FES. They each received an eight channel Implantable Receiver Stimulator (IRS-8) and eight
The subject iSCI-1 presented with bilateral hip and ankle weakness and received
latae, and tibialis anterior/peroneus longus muscles. FES assisted limb advancement during
swing phase and knee stability during single limb stance, while the subject achieved forward
progression through the voluntary contractions of his hip extensor musculature. Subject iSCI-2
presented with unilateral weakness and received stimulating electrodes only on his left side to
recruit iliopsoas, vastus intermedius/ lateralis, tensor fasciae latae, gluteus medius, gluteus
maximus, posterior portion of adductor magnus, and tibialis anterior (2 electrodes). Temporal
patterns of stimulation to activate the muscles were customized for each subject’s individual gait
deficits according to established tuning procedures [4.7], [4.8] in order to achieve forward
stepping in a rolling walker. The subjects completed 6 weeks of over-ground gait training (2
hour sessions, 3 times per week) with a physical therapist using the implanted FES system. After
discharge from rehabilitation, they volunteered for the studies using the myoelectric control of
75
Informed consent was obtained from both the subjects before their participation and the
Institutional Review Board of the Louis Stokes Cleveland Department of Veterans Affairs
The laboratory set-up for EMG-triggered FES-assisted walking is shown in Figure 4.1.
Surface electromyogram (EMG) was collected according to SENIAM guidelines [4.10]. The
EMG signals were pre-amplified (gain: 100), and low-pass (fcutoff=1000 Hz) filtered by CED
1902 (Cambridge Electronic Design, England) amplifier (gain: 330 or 990) before being sampled
at 2400 Hz (AT-MIO-64F-5, National Instruments, USA) in the host personal computer. The
sampled data were detrended, band-pass filtered (5th order zero-lag Butterworth, 20-500 Hz),
and rectified before being low pass filtered (5th order zero-lag Butterworth, fcutoff=3 Hz) to get
the linear envelope (LE) of the signal. Surface EMG signals from bilateral erector spinae (at T9)
were used by the gait event detector (GED) for triggering FES-assisted left and right steps of
iSCI-1. The GED for iSCI-2 used EMG from left gastrocnemius and right tibialis anterior for
triggering only the FES-assisted left step during EMG-triggered walking. A pattern recognition
algorithm (i.e., the EMG-classifier) based on feature templates derived from the first three
principal components of the time course of the LEs during double support was trained for each
subject to detect the intention to take a step. Details of the development, testing and technical
performance of the EMG-classifier GED have been presented in Chapter 3 [4.4]. A flow chart
summarizing the algorithm is shown in Figure 4.2. The EMG-classifier was trained to look for a
set of features (templates found from the training data) in the EMG LEs that indicated the
intention to trigger the next step. The features were detected using correlation analysis of the
windowed EMG LE, which was represented as a weighted (weights = correlation coefficients)
mean of those features. If the weights (or loadings) of those features fell within a certain
76
Euclidean distance (a threshold found from training data) of the centroid for the class associated
with initiating a step, then the next FES-assisted step was triggered. The EMG-classifier kept
looking for those features for a given time (typically 3 seconds or less) while the user was in the
double support phase of gait. If the EMG-classifier couldn’t find those features in the given time
then it stopped scanning the EMG LEs and waited for the user to trigger the next step with a
manual switch. The EMG-classifier and stimulation control system operated in real time on xPC
Retro-reflective markers were placed on the body segments according to the ‘plug-in’ gait
marker set in the Vicon WorkstationTM (Vicon Peak, USA) software to acquire lower-body
kinematics data using a seven camera ViconTM 360 motion capture system. The fifteen markers
were placed on the anterior superior iliac spine (ASIS), thigh (on a wand at 1/3rd the distance
between the greater trochanter and lateral femoral epicondyle), lateral femoral epicondyle, shank
(on a wand at 1/3rd the distance between the lateral femoral epicondyle and lateral malleolus),
lateral malleolus, toe (dorsum of the foot over second metatarsal head), and heel (in line with the
toe marker on the long axis of the foot) bilaterally. The final marker was placed at the sacrum
(midway between the posterior superior iliac spines). The measurement volume of the motion
capture system was approximately 5m from the position from where the subjects started walking
on the walkway. The marker trajectories were captured at 60 frames/sec. The optimum cut-off
frequency (5th order zero-lag Butterworth) for low-pass filtering the kinematic data was found to
balance the Root Mean Square Error (RMSE) due to noise and RMSE due to the attenuation of
the true signal [4.5]. Foot progression in the sagittal plane was selected to find the optimum low
pass cut-off frequency since it has the highest frequency content [4.9]. The Figure 4.3 shows the
RMSE with the optimum cut-off frequency, which was found to be 3.5 Hz for the iSCI data. Gait
77
events related to foot-floor contact (foot-strike – when any portion of the foot first touched the
ground, and foot-off – when the whole foot was off the ground) were derived from foot insole-
mounted foot switches (B&L Engineering, USA) placed bilaterally at the medial and lateral heel,
first and fifth metatarsal, and big toe, and confirmed with the kinematics data acquired.
The FES-assisted stepping in the iSCI subjects was triggered with one of the two command
sources – manually triggered with a switch (switch-trigger) or triggered with the EMG-based
GED (EMG-trigger). The two command sources were presented in a random order to the
subjects. The iSCI subjects initiated the gait with a switch-triggered left step and then walked in
a rolling walker with FES-assisted steps triggered by one of the two command sources. All the
subjects were asked to make multiple passes across the straight level walkway with a period of 5-
10 minutes of rest period in between the passes. Each pass constituted one trial and 10 trials were
collected during a session. Sessions were repeated on two separate days for a total of 20 trials,
representing 10 trials consisting of two passes of the walkway for each command source. During
each trial the subjects were instructed to reach a self-selected steady-state gait at a comfortable
velocity from the standing position. The ViconTM motion capture system recorded marker
trajectories for four or more strides starting after the first (switch initiated) step, as illustrated in
Figure 4.4. The joint kinematics data were analyzed using custom programs developed in
MATLABTM R14 (The MathWorks, Inc., USA). The trials were considered independent for
statistical analysis, to compute and compare the basic spatio-temporal gait parameters.
Gait Parameters
The basic gait parameters were computed from the kinematics data where the definitions
for the gait events were modified to account for the gait characteristics of iSCI subjects. It was
often difficult to identify heel strike and toe off during iSCI walking with FES, so these events
were replaced with initial contact/foot-strike (when one or more of the foot switches first
78
registered contact with the ground) and foot-off (when all the foot switches registered loss of
contact). Gait speed was defined as the average velocity of progression over a stride period
which was computed from the velocity of the sacrum marker and reported in meters per second
(m/s). Left and right step lengths were determined by the horizontal distance covered along the
plane of progression during a step, which were measured from the contact position of the
contralateral foot to the subsequent contact position of the ipsilateral foot and reported in meters
(m). The step period (left and right) was the duration of one step, which was measured from the
contact time of the contralateral foot to the subsequent contact time of the ipsilateral foot and
reported in seconds (sec). The double support period was the time in seconds (sec) when both the
feet were on the ground as measured from the foot-strike of the contralateral foot to the
ipsilateral foot-off. The swing period was the duration of time in seconds (sec) when only one
foot was in contact with the ground and was measured from ipsilateral foot-off to subsequent
ipsilateral foot-strike. During walking, swing period is exactly equal to the single support period
of the contralateral limb. The stance period in seconds (sec) was defined as the time in seconds
(sec) when the foot was in contact with the ground, and was sum of the double support period
The average body weight support provided by the rolling walker over a stride was also
measured for iSCI-2. For comparison, the average body weight supported during switch and
EMG triggered gait were each normalized by the mean of that observed during EMG-triggered
gait.
Statistical Analysis
One way analysis of variance (‘anova1’ in MATLABTM R14, The MathWorks, Inc., USA)
was performed on each gait parameter to determine if significant differences existed between the
79
considered to be mutually independent for the ANOVA test. The p-value was computed for the
null hypothesis that a particular EMG-triggered gait parameter and switch-triggered gait
parameter had the same mean. If the p-value was close to zero (<0.05) then the null hypothesis
was rejected and the difference between the mean of that gait parameter for EMG-trigger and
Results
Subject iSCI-1 could trigger FES-assisted steps with the EMG-based GED with a false
positive rate lower than 1% and a true positive rate higher than 82% for the left step and 83% for
the right step. Subject iSCI-2 could trigger a FES-assisted left step with the EMG-based GED
with a false positive rate lower than 1% and a true positive rate higher than 96%. The mean,
standard deviation (S.D. ), coefficient of variation (C.V. ), and 95% confidence interval (C.I.) of
the EMG-triggered and switch-triggered gait parameters over 10 trials (N=10) for subjects iSCI-
1 and iSCI-2 are summarized in Table 4.1 and Table 4.2, respectively. The gait parameters that
show statistically significant (p<0.05) difference are shaded. The subject iSCI-1 had a
statistically significant decrease in double support duration bilaterally with the EMG-triggered
FES-system, indicating a more dynamic walking pattern than with switch triggering. Gait speed
increased slightly due to the decrease in the double support duration, but failed to reach
significance (p=0.092). Boxplot of the normalized average body weight supported by the walker
for iSCI-2 during EMG-triggered (N=10 trials) and switch-triggered (N=10 trials) gait is shown
in Figure 4.5. The average body weight support over a stride during the switch-triggered gait was
almost 1.4 times that of EMG-triggered gait, which was statistically significant (p<0.05) and
indicated a more erect posture and less upper extremity effort for stability and forward
propulsion.
80
Discussion
for most of the spatio-temporal parameters measured for both subjects in spite of the differences
in their injury levels, degree of preserved volitional motor control, and muscle set chosen for
stimulation. Due to individual variations in their FES systems and levels of paralysis, EMG-
triggering demonstrated significant benefits over standard switch triggering for different aspects
of gait in each individual subjects. Both subjects could increase or decrease walking velocity by
either advancing or delaying depressions of the command switch between successive steps, or
modifying the timing of the voluntary contraction of the muscles used as inputs to the EMG
classifier.
Because iSCI-1 could not initiate a step without FES-assistance and required bilateral
stimulation to advance his swing limb, the duration of swing phase was determined by the timing
of the pre-programmed pattern of stimulation for both switch- and EMG-triggered systems. The
spatiotemporal parameters of the swing phase were therefore similar for both the command
sources for this subject. The only method for modulating gait speed available to iSCI-1,
therefore, involved altering the time spent in stance phase. He exhibited a statistically significant
decrease in double support duration bilaterally with automatic triggering of stimulation with the
EMG-based GED, which tended to increase his walking velocity. These results have functional
implications and indicate improved confidence and dynamic stability with EMG triggering.
Subject iSCI-2, in contrast, exhibited a thoracic level motor injury and could walk short
distances without FES-assistance, albeit with significant effort. Because his FES system was
implemented unilaterally, he was able to modulate his walking speed by adjusting not only the
time spent in double support, but also by the timing of his volitional right step. He walked
slightly faster with switch-trigger (0.215 ± 0.112 m/s) than with EMG-trigger (0.202 ± 0.063 m/s)
81
though this minor difference was not statistically significant. In terms of all the spatio-temporal
gait parameters measured, iSCI-2 could walk as well with the EMG-trigger as with the switch-
trigger. A significant difference was found in the body weight placed on the rolling walker,
which was significantly higher for switch-triggered gait. EMG-triggering enabled the subject to
place more of his body weight on his lower extremities, better balance himself in swing and
These results imply that EMG-triggering may have the potential to lower energy
expenditure due to reduced double support phase and decreased upper extremity exertion on the
walker during long distance walking. Because of the laboratory nature of this preliminary
feasibility study, the impact of EMG control on walking distance, energy expenditure and
Subject iSCI-1 exhibited difficulty operating the standard manual finger switch due to
reduced hand and finger function. Customized reed switches were interfaced to the external
controller of the neuroprosthesis and mounted on his walker to allow him to manually trigger the
system via gross extension movements of his wrist or metacarpal joints. Such specialized
adaptations were not necessary with EMG triggering. Other individuals with incomplete SCI
with reduced hand and finger function may benefit from the reduced dependence on manual
The goal of this work is to facilitate limited community ambulation by giving the FES
system users command sources that are natural, intuitive and easily incorporated into their
walking patterns. It demonstrates the initial feasibility of one method of automatically triggering
FES-assisted stepping to reduce dependence on the manual switch and better integrate
stimulation with the preserved voluntary movements associated with walking. The EMG-based
82
GED used in this feasibility study was implemented in a PC which was tethered to the subject,
and is therefore obviously unsuitable for community use. The new family of implantable
stimulator-telemeters (IST) developed at Case Western Reserve University and the Louis Stokes
Cleveland Department of Veterans Affairs Medical Center is capable of processing two channels
of EMG signals from implanted recording electrodes and deriving the commands required to
control 12 channels of stimulation [4.3]. The next step in the development of EMG-controlled
assisted ambulation with a completely implantable technology. Implanting the control and
activation sources in a single system will enable the evaluation of the benefits of EMG control in
the home and community environments, during activities of daily living and over long,
Conclusion
The gait performances of individuals with partial paralysis due to incomplete SCI with
gait were found to be at least as good as those exhibited by switch-triggered FES systems FES-
assisted stepping with EMG-triggering can reduce the duration of double support and thus
improve dynamic stability. Stepping with EMG-triggered stimulation also reduced the body
weight exerted on the rolling walker or other assistive device when compared to switch-triggered
gait. Specialized adaptive equipment and customizations for individuals with upper extremity
impairments were minimized or eliminated through application of EMG control. The benefits
derived from EMG triggering depend on the extent of paralysis, level of impairment and
individual configuration of the FES system. This study presented the results from indoor walking
under laboratory conditions. It is hypothesized that EMG-triggered FES-system will have more
83
significant benefit during gait transitions and dynamic stability during that period than that
during steady state gait since it allows the user to transition more fluidly and stably from one
References
4.2. B. Smith, Z. Tang, M.W. Johnson, S. Pourmehdi, M.M. Gazdik, J.R. Buckett, and P.H.
Peckham , “An externally powered, multichannel, implantable stimulator-telemeter for
control of paralyzed muscle,” IEEE Trans Biomed Eng., vol. 45, no. 4, 1998, pp. 463-475.
4.3. N. Bhadra, K.L. Kilgore, and P.H. Peckham, “Implanted stimulators for restoration of
function in spinal cord injury,” Med. Eng. Phys., vol. 23, 2001, pp. 19-28.
4.4. A. Dutta, R. Kobetic, and R. Triolo, “Ambulation after incomplete spinal cord injury with
EMG-triggered Functional Electrical Stimulation,” IEEE Transactions on Biomedical
Engineering, 55:2, February 2008.
4.5. D. A. Winter, The Biomechanics and Motor Control of Human Movement, 2nd edition,
New York: Wiley, 1991.
4.7. R. Kobetic, and E.B. Marsolais, “Synthesis of paraplegic gait with multichannel functional
neuromuscular stimulation,” IEEE Trans Rehab Eng., vol. 2, no. 2, 1994, pp. 66-79.
4.8. R. Kobetic, R. J. Triolo, and E. B. Marsolais, “Muscle selection and walking performance
of multichannel FES systems for ambulation in paraplegia,” IEEE Trans. Rehabil. Eng.,
vol. 5, no. 1, Mar. 1997, pp. 23–29.
4.9. C. Angeloni, P. O. Riley, and E. D. Krebs, “Frequency Content of Whole Body Gait
Kinematic Data,” IEEE Trans. Rehabilitation Engineering, vol. 2, no. 1, 1994, pp. 40-46.
84
Figures
Figure 4.1: Experimental setup for testing EMG-triggered FES-assisted walking with the block-
diagram for the EMG-triggered FES-system (ECU: external control unit).
85
EMG during Feature Set
double-support & Threshold
Loadings
Find the distance from centroid
of case ‘T’ in feature space
Euclidean distance
Distance Y Trigger
< swing-
threshold phase
STOP
Double- Y Wait for
support manual
> 3 sec? trigger
N
Get next
EMG sample
Figure 4.2: EMG-based gait event detector for triggering FES-assisted steps.
86
Figure 4.3: Plot of the Root Mean Square Error (RMSE) between the low-pass filtered and
unfiltered foot progression in sagittal plane with cut-off frequencies to find the
optimum cut-off frequency for low-pass filtering the gait kinematics data. Optimum
cut-off frequency was found to be 3.5 Hz for iSCI data.
87
Figure 4.4: Gait data collection protocol in laboratory conditions where the subject had to start
from standing and achieve a self-selected gait speed within ViconTM volume of data
capture (~5m).
88
Figure 4.5: Boxplot of average body weight support provided by the walker during EMG-
triggered (N=10 trials) and switch-triggered (N=10 trials) gait normalized by the
mean of that during EMG-triggered trials of iSCI 2. The box shows the lower
quartile, median, and upper quartile with whiskers extending at each end showing the
range of the data. The notches around the median show the estimate of the
uncertainty. The boxes whose notches don’t overlap indicate that their medians differ
at 5% significance level.
89
Tables
Table 4.1: The Mean, Standard Deviation (S.D), coefficient of variation (C.V.), 95% confidence
interval (95% C.I.) over 10 trials (N=10) of the EMG-triggered and switch-triggered
gait parameters – gait speed (m/s), left step length (m), right step length (m), left
double support duration (s), right double support duration (s), left swing phase
duration (s), right swing phase duration (s) for the subject iSCI 1. [ ‘statistically
significant’ (p<0.05) difference between the command sources are shaded]
Left Step Length (m) EMG-trigger 0.381 0.107 0.280 0.315 - 0.447
Switch-trigger 0.337 0.066 0.194 0.297 - 0.378
Right Step Length (m) EMG-trigger 0.422 0.167 0.395 0.319 - 0.526
Switch-trigger 0.428 0.075 0.174 0.382 - 0.475
Left Swing Phase Duration (s) EMG-trigger 0.697 0.123 0.176 0.621 - 0.773
Switch-trigger 0.642 0.128 0.199 0.563 - 0.721
Right Swing Phase Duration (s) EMG-trigger 0.635 0.133 0.209 0.553 - 0.717
Switch-trigger 0.635 0.120 0.189 0.560 - 0.709
90
Table 4.2: The Mean, Standard Deviation (S.D), coefficient of variation (C.V.), 95% confidence
interval (95% C.I.) over 10 trials (N=10) of the EMG-triggered and switch-triggered
gait parameters – gait speed (m/s), left step length (m), right step length (m), left
double support duration (s), right double support duration (s), left swing phase
duration (s), right swing phase duration (s) for the subject iSCI 2. [ ‘statistically
significant’ (p<0.05) difference between the command sources are shaded]
Left Step Length (m) EMG-trigger 0.359 0.009 0.026 0.353 - 0.364
Switch-trigger 0.346 0.069 0.199 0.304 - 0.389
Right Step Length (m) EMG-trigger 0.450 0.138 0.307 0.365 - 0.536
Switch-trigger 0.521 0.176 0.338 0.412 - 0.630
Left Double Support Duration (s) EMG-trigger 1.192 0.171 0.144 1.086 - 1.299
Switch-trigger 1.060 0.294 0.278 0.878 - 1.243
Right Double Support Duration (s) EMG-trigger 1.165 0.059 0.051 1.128 - 1.201
Switch-trigger 1.148 0.076 0.066 1.101 - 1.196
Left Swing Phase Duration (s) EMG-trigger 1.255 0.062 0.050 1.216 - 1.293
Switch-trigger 1.271 0.093 0.073 1.213 - 1.328
Right Swing Phase Duration (s) EMG-trigger 1.043 0.058 0.056 1.007 - 1.079
Switch-trigger 1.054 0.065 0.062 1.013 - 1.094
91
CHAPTER 5
COORDINATION AND STABILITY OF SURFACE EMG-TRIGGERED FES-ASSISTED
OVERGROUND WALKING IN THE LABORATORY
A manuscript based on this chapter was accepted for publication in ASME Journal of
Biomechanical Engineering.
Abstract
activating the muscles of the lower extremities. Individuals with incomplete spinal cord injury
(iSCI) retain partial volitional control of muscles below the level of injury, necessitating careful
integration of FES with intact voluntary motor function for efficient walking. The FES-assisted
stepping can be triggered automatically at a fixed rate (auto-trigger), by a manual switch (switch-
postulated that EMG being a more natural command source than manual switches will enable
better coordination of stimulated and volitional motor function necessary during gait. In this
study, the above stated hypothesis was investigated in two volunteers with iSCI during the over-
ground FES-assisted gait initiation. Four able-bodied volunteers provided the normative data for
comparison. The EMG-triggered FES-assisted gait initiation was found to be more coordinated
and dynamically more stable than auto-triggered and switch-triggered cases. This highlighted the
potential of surface EMG as a natural command interface to better coordinate stimulated and
Introduction
to wheelchair dependent individuals with spinal cord injuries. Implanted FES systems have
allowed individuals with either complete or partial paralysis to stand and step via activation of a
customized set of muscles selected to address individual gait deficits with pre-programmed
92
patterns of stimulation to produce cyclic movement of the lower extremities [5.1][5.2]. The 8-
channel implantable receiver-stimulator (IRS-8) utilized in these systems delivers stimulation via
implanted electrodes to the targeted motor nerves activating the muscles required for walking.
Implant recipients have used a ring-mounted thumb switch connected to a wearable external
control unit (ECU) to manually trigger each step activated by the customized pattern of
stimulation to achieve walking. This pattern of stimulation can also be cycled automatically at a
fixed rate with the user initiating and terminating the cyclic pattern with the manual switch.
Graupe et al. proposed an alternative EMG-based control of FES based on time-series models of
the surface electromyogram (EMG) [5.3]. Auto-regressive (AR) time-series models of EMG
from the intact musculature above the level of lesion can give information related to limb
function [5.4] and can provide patient-responsive control of FES-assisted walking [5.5].
Recently, Graupe and Kordylewski presented a neural network based classifier that has online
learning capabilities for individuals with complete paraplegia due to spinal cord injuries [5.6],
[5.7]. Our studies indicate that EMG from partially paralyzed muscles in people with paraplegia
due to incomplete spinal cord injuries can provide significant information related to the
volitional activity of the muscle and can be used for gait-event detection [5.8], [5.9]. While the
electromyogram showed promise of being a more natural command signal than manual switches
but the postulated improvement in function in terms of better coordination and dynamic stability
Gait initiation refers to the transient state between standing and steady state walking.
Analysis of gait initiation is an important diagnostic tool to study pathologic gait [5.22]. In
steady state normal gait, the net mechanical work over one stride is zero [5.23]. So enough
momentum is generated during gait initiation for the body to reach a steady state velocity [5.22].
93
The first step is used to propel the body into a dynamic state, and then in the next step a major
energy input takes place to raise the energy level of the body [5.22]. This energy input is a
disturbance to the locomotor system to propel the body forward. In the subsequent steps (3rd step
and onwards), this disturbance decays and the body readjusts to attain a steady gait speed [5.22].
Hurmuzlu and Basdogan presented a method to investigate orbital dynamic stability of human
gait in terms of its response to disturbances [5.10]. Based on their methods, Hurmuzlu et al.
found that gait in post-polio patients is less stable than able-bodied controls [5.11]. Recently
Dingwell et al. found that neuropathic patients do not improve their orbital dynamic gait stability
by lowering their gait speed and are also not less stable than able-bodied controls in spite of their
The current study used the method of gait analysis introduced by Hurmuzlu and Basdogan
to evaluate dynamic stability of walking with FES under a variety of control schemes [5.10]. It is
desirable that the user with incomplete spinal cord injury (iSCI) fluidly and stably transition from
a static ‘stand’ state to a dynamic ‘walk’ state during the FES-assisted gait initiation. The goal of
this study was to determine the effects of the command input to the FES system on dynamic
stability and convergence to a normal gait pattern. The working hypothesis of the investigation
was that coordinating the actions of the stimulator with the remaining volitional movements with
a more natural command interface like EMG would result in improvements in gait transitions
from standing to walking. An EMG-based gait-event-detector (GED) was used to trigger the
FES-assisted stepping [5.9]. The linear classifier in the GED detected the intent to initiate swing-
phase from the EMG of two partially paralyzed muscles and used that information to trigger
94
Methods
Subjects
Two subjects with incomplete spinal cord injury (iSCI) volunteered for this study. The first
subject (iSCI-1) was a 23 years old male with C7 motor and C6 sensory incomplete spinal cord
injury (ASIA C) who could stand but could not initiate a step without the assistance from FES.
The second subject (ISCI-2) was a 34 years old male with T1 motor and C6 sensory incomplete
spinal cord injury (ASIA D) that resulted in left hemiplegia (unilateral paralysis) who could walk
only short distances with great difficulty without the assistance from FES. Each subject received
an IRS-8 and eight surgically implanted intramuscular electrodes in a related study designed to
facilitate household and community ambulation [5.1], [5.2]. Temporal patterns of stimulation to
activate the muscles were customized for their particular gait deficits according to established
tuning procedures [5.13], [5.14] in order to achieve forward stepping in a rolling walker. The
subjects completed 6 weeks of over-ground gait training (2 hour sessions 3 times per week) with
a physical therapist using the implanted FES system. After discharge from rehabilitation, they
volunteered for the studies related to the myoelectric control of the FES system. Four able-
bodied volunteers (Able1: 25 years, Able2: 52 years, Able3: 26 years, Able4: 54 years), who had
no known injury or pathology during the study, provided the normative gait data.
Informed consent was obtained from the subjects before their participation and the
Institutional Review Board of the Louis Stokes Cleveland Department of Veterans Affairs
The laboratory set-up for EMG-triggered FES-assisted walking is shown in Figure 5.1a.
Two channels of surface EMG (EMG) from bilateral Erector Spinae (at T9) for iSCI-1 and from
left gastrocnemius & right tibialis anterior for iSCI-2 were used by the gait-event-detector (GED)
95
for EMG-triggered FES-assisted gait. Details of the development, testing and technical
performance of the EMG-classifier GED have been published elsewhere [5.9]. A flow chart
summarizing the algorithm is shown in Figure 5.1b. The EMG-classifier was trained to look for a
set of features (or pattern of muscle activity) in the EMG linear envelopes (LEs) that indicated
the intention to trigger the next step. The features were detected using correlation analysis of the
windowed EMG LE, which was represented as a weighted (weights = correlation coefficients)
mean of those features. If the weights (or loadings) of those features lie within a certain
Euclidean distance (i.e. a threshold) of the centroid for the class associated with initiating a step
found from the training data, then the next FES-assisted step was triggered. The EMG-classifier
kept looking for those features for a given time while the user was in the double support phase of
gait. If the EMG-classifier couldn’t find those features in the given time then it stopped scanning
the EMG LEs and waited for the user to trigger the next step with a manual switch. The EMG-
classifier and stimulation control system was implemented in a real time using xPC TargetTM
Retro-reflective markers were placed on the body segments according to the ‘plug-in’ gait
marker set in the Vicon WorkstationTM (Vicon Peak, USA) software to acquire lower-body
kinematic data using a seven camera ViconTM (Vicon Peak, USA) motion capture system. The
fifteen markers were placed on the anterior superior iliac spine (ASIS), thigh (on a wand at 1/3rd
distance between the greater trochanter and lateral femoral epicondyle), lateral femoral
epicondyle, shank (on a wand at 1/3rd distance between the lateral femoral epicondyle and lateral
malleolus), lateral malleolus, toe (dorsum of the foot over second metatarsal head), heel (inline
with the toe marker on the long axis of the foot) bilaterally, and at the sacrum (midway between
the posterior superior iliac spines). The measurement volume of the motion capture system was
96
approximately 5m from the position from where the subjects started walking on the walkway.
The marker trajectories were captured at 60 Hz for the computation of the joint angles.
The joint angles were low-pass filtered before Euler differentiation to compute the joint
velocity. Numerical differentiation amplifies the noise so the noise in the joint angles data was
reduced with low pass filtering. The optimum cut-off frequency (5th order zero-lag Butterworth)
for the able-bodied and iSCI data was found to balance the Root Mean Square Error (RMSE) due
to noise and RMSE due to the attenuation of the true signal [5.24]. The foot kinematic data were
selected to find the optimum low pass cut-off frequency since the high frequency noise induced
by the foot-strike (i.e. collision) during walking affects the foot kinematic data the most (highest
frequency noise expected) and then the physiological damping of high-frequency content occurs
caudal to the head [5.25]. The Figure 5.2a shows the RMSE with the optimum cut-off
frequencies, which were 6 Hz for the able-bodied and 3.5 Hz for the iSCI kinematic data. The
RMSE value was similar for the able-bodied and iSCI data after low-pass filtering (Figure 5.2a).
The Figure 5.2b shows that most of the power-content of the signal was below their respective
optimum cut-off frequency for low pass filtering. Gait events (foot-strike and foot-off) were
derived from foot-floor contact patterns obtained from insole-mounted foot switches (B&L
Engineering) placed bilaterally at the medial and lateral heel, first and fifth metatarsal, and big
The FES-assisted swing-phase in iSCI subjects could be triggered with three command
interfaces – manually triggered with a switch that is mounted on the walker (switch-trigger),
without user-intervention (auto-trigger). The three trigger modes were presented randomly to the
iSCI subjects during a trial and the iSCI subjects initiated the gait with their left leg (i.e. the
97
initiator limb). All the subjects were asked to make ten passes across the straight level walkway.
Each pass constituted one trial and 10 trials were collected during a session of data capture.
During each trial the subjects tried to reach a self-selected steady gait from standing. ViconTM
(Vicon Peak, USA) motion capture system recorded marker trajectories for four or more strides
starting after the first (initiator) step, as illustrated in Figure 5.3. Able-bodied subjects usually
reach steady state gait within three to four steps from standing [5.22], [5.11].
The time-series analysis used to evaluate the stand-to-walk transition required that the
subjects complete the transition from standing to a walking state. The time-series model
evaluated how well this transition happened based on a linear model. Hence, the time-series
analysis was performed only on the trials in which the pelvis position estimated from the
centroid of the ASIS markers and the sacrum marker had a positive progression velocity after the
first step from standing during the whole trial, as shown in Figure 5.4 The time-series analysis
was performed to compare the stability and coordination of the FES-assisted iSCI gait initiation
(with the three trigger modes) with the able-body normative data, as discussed below. The joint
kinematic data were analyzed using custom programs developed in MATLABTM R14 (The
rotations at the ankle, knee, and hip) and the corresponding 18 joint velocities (for a total of 36
states) represented the human locomotor system during gait initiation. The swing and stance
phases were identified from the foot-floor contact patterns recorded with insole-mounted foot
switches. The time-series analysis was primarily based on the method presented by Hurmuzlu
and Basdogan [5.10]. The asymptotic behavior of the dynamical locomotor system during the
gait initiation was assumed to be a bounded set which would be called the steady-state gait. The
98
convergence of the gait during the gait initiation towards a nearly periodic cyclic steady state
was investigated [5.22]. The cyclic time-series of the joint kinematics was probed at a particular
gait event from gait cycle to cycle. Any gait event can be selected for this analysis. Theoretically
the measure should be invariant to the selection of the gait event although this invariance
property may not hold for the simplified model of the gait dynamics [5.11]. To test this
invariance, we selected six gait-events to do the same analysis. The gait events were bilateral
A mapping function was then found that returns the cross-section of the flow (or
trajectory) to itself (called a return map) from gait cycle to cycle at the gait event. If the
disturbance is not large enough then a linear return map can be found which can then be
subjected to further analysis for stability using methods for linear systems analysis [5.10]. Our
preliminary studies showed that this mapping function during gait initiation was mostly linear for
able-bodied subjects [5.15]. A more rigorous analysis of this linearity was performed for
coordination of gait initiation. A linearized mapping function was approximated from the joint
kinematic data with linear regression techniques. The Jacobian of the linear model provided an
estimate of the rate of convergence during the first few steps taken from standing even if the gait
did not reach a steady state in the volume of motion capture. The shorter the gait initiation phase,
the more stable the gait. A rate parameter found from the Jacobian gave a measure of stability, as
discussed below.
Let Xn be the ‘36 x 1’ dimensional vector of the 36 states at the nth gait event i.e. at the nth
gait cycle (GC) during the gait initiation ( n = 1K m ) of a trial. The last GC (i.e. the mth gait
event) of the trial was considered to have reached a steady state. Let a ‘36 x 1’ dimensional
99
vector X* represent the 36 states at the mth (i.e. the last) gait event of the trial. The discrete
mapping of the states of the system at the nth gait event to the (n+1)th gait event (called the
‘returns’) was represented by an equation, Xn+1 =f (Xn), where f is the (return) mapping function.
(X n +1
− X* = ) df
dx
(X n
− X* )
x= X Equation 1
*
δ X n +1 = J • δ X n
( 36×1) ( 36×36 ) ( 36×1)
A multivariate linear regression model of the gait kinematics over ‘p’ gait initiations (i.e. p
trials) gave an estimate of the Jacobian J . The 36 perturbations at the first (m-1) gait events
( i.e. n = 1K m − 1 ) served as the predictor variables δ Xn . The average of the magnitude of the
eigenvalues of the transposed estimated Jacobian matrix J T has been proposed to be an overall
stability measure (α-measure of stability) by Hurmuzlu and Basdogan [5.10]. (1-α) gives an
estimate of the rate of convergence during the gait initiation and smaller values of α-measure
will denote a faster gait initiation and a gait more resilient to disturbances.
For multivariate linear regression analysis, the 36-predictor variables at (m-1) gait events
during a trial were stacked by trials in a ( ( m − 1) p × 36 ) dimensional design matrix δXn . The
response variables were the corresponding ‘returns’ of the 36 perturbations ( i.e. n = 2 K m ) during
δX n+1 = δX n • J + ε Equation 2
( ( m −1) p ×36 ) ( ( m −1) p × 36 ) ( 36× 36 ) ( ( m −1) p ×36 )
100
The residuals ε , and the parameter matrix J , were the unknowns [5.16]. The 36 columns of
the residuals ε are expected to have zero mean since the perturbations should ideally decay to
The data collected from 4 able-bodied subjects showed multi-colinearity for the ill-
conditioned matrix ( δXn T • δXn ) , which was expected since joint angles co-vary during human
gait [5.17]. Principal Component Analysis (PCA) was performed to address the problem of
multi-colinearity (redundancy) of the predictor variables and to improve the condition number
[5.18]. PCA determines fewer linear combinations of the 36 predictor variables that can explain
most of the variance-covariance structure. Principal Components (PCs) being orthogonal to each
other will solve the problem of multi-collinearity of the predictor variables. The dimension
reduction with PCA was performed on the 36 standardized perturbation states, which were
stacked in a ( mp × 36 ) dimensional matrix δZ . The weight matrix W from PCA was used to
project the 36 standardized response ( δZn+1 ) and predictor variables ( δZn ) on to new basis
vectors [5.18].
δPn+1 = δZ n +1 • W
( ( m −1) p×36) ( ( m −1) p×36) ( 36×36)
δPn = δZ n • W
( ( m −1) p×36) ( ( m −1) p×36) ( 36×36)
Equation 4
The multivariate linear regression model in terms of the new response δPn+1 and predictor
101
The adequate number of principal components ( η ) was determined from the cumulative
percent variance accounted for (%VAF) by the PCs such that %VAF ≥ 90% . If η is less than 36
(dimension of the original model) then there is a reduction in the dimension of the model, which
will improve the confidence on the elements of J * for the same number of observations ( m − 1) p .
For a multivariate linear regression model with the first η PCs as the predictor and response
variables, we will define a quality of fit ( QoFη ) for the reduced dimension linear regression
model:
or trace SS totη = trace SS predη + trace SS resη
(η ×η ) (η ×η ) (η ×η )
QoFη = 1 −
( )
trace SS resη
trace ( SS ) tot
η
Since the columns of δPn+1 and δPn are linearly independent so SS totη and SS predη will be diagonal
Also trace(X ) is the sum of all the diagonal elements of the matrix X
Equation 6
The QoFη , which should lie between 0 and 1 gives a measure of the quality of the linear
regression model for cycle-to-cycle coordination during the gait initiation whiles the eigenvalues
of J*η T provide a measure of rate of convergence to a steady state. The measures, QoFη and the
average eigenvalue of J *η T (called Av. Eig.) were used to compare between able-bodied and iSCI
gait initiation.
102
Results
The Euclidean distance of the perturbation of 36 states to the origin (18 joint angles and 18
joint velocities) at a gait event (maximum left knee flexion) during gait initiation is shown in
Figure 5.5. Among all the gait events, the knee flexion was found to give results with lowest
variance especially with pathological data [5.11]. The multivariate linear regression model fitted
the convergence in a 36-D space while the Figure 5.5 shows the convergence lumped in
Euclidean distance for ease of illustration. The perturbation as seen from the Euclidean distance
converged towards zero for able-bodied gait initiation, as shown in Figure 5.5a. The iSCI
subjects made, on an average five strides within the volume of motion capture due to their
shorter stride length while the able-bodied could take only four strides.
This convergence of 36-dimensional perturbation during gait initiation was analyzed with a
linear regression model (Equation 2, 3). The condition number for δX n T • δX n (in Equation 3) was
found to be very high (~1020) which indicated numerical instability during the computation of its
inverse.
Principal component analysis (PCA) was performed to reduce the dimensionality of the linear
regression model and to improve the condition number. Five principal components (PCs)
accounted for greater than 90% variance in the data, as shown in Figure 5.6. Thus the number of
principal components (denoted byη ) selected for the reduced dimension linear regression model
(Equation 6) was 5 (i.e. η = 5 ). The rest of the variance (<10% VAF) was accounted for by the
PCs from 6 – 36. These small amplitude movements were not included in the overall
coordination and stability analysis. The condition number of δX nT • δX n improved after PCA,
which was of the order of 10. The first three principal components accounted for greater than
103
70% variance in the data, as shown in Figure 5.6. In Figure 5.7, the sagittal plane angles (suffix
‘x’) have the highest loading for the first two principal components (PCs) while the 3rd PC also
accounted for the pronation/supination at the ankle. The first three PCs mostly accounted for the
kinematic coupling in the sagittal plane where most of the limb movements occur during gait.
Lilliefors test for goodness of fit to a normal distribution (‘lillietest’ MatlabTM) couldn’t reject
the hypothesis that each of those error vectors ε*η in the reduced dimension linear regression
model (Equation 6, η = 5 ) had a normal distribution with significance level of 5%. The
perturbation of 5 Principal Components (PC) to the origin as seen from the Euclidean distance
for the ease of illustration is shown in Figure 5.8. After dimension reduction of the linear
regression model, the perturbation as seen from the Euclidean distance still converged towards
zero for able-bodied normative gait. The convergence of the perturbation during EMG-triggered
iSCI gait initiation as seen with the Euclidean distance was more similar to normative gait for the
iSCI subjects, iSCI-1 and iSCI-C2. The multivariate linear regression model will provide the
The scatter plot of QoF5 and average eigenvalue magnitude (Av. Eig.) at the six gait events
is shown in Figure 5.9a for each of the able-bodied and iSCI subjects with different trigger
modes. There are six data points corresponding to 6 gait events for each group in Figure 5.9a.
The QoF and Av. Eig. varied substantially based on the selection of the gait event, as shown in
Figure 5.9a. Three clusters shown with ellipses on Figure 5.9a correspond to the able-bodied
data, EMG and switch – triggered iSCI data, and the auto-triggered iSCI data. One-way
if there is a significant difference in the entire set of means (of QoF5 and Av. Eig.) from group to
104
group. It was found that the means differ and the difference lies on a plane (dimensionality=2,
p=0.0002). Figure 5.9b shows the dendogram plot (‘manovacluster’ MatlabTM), which is the
hierarchical binary cluster tree formed using the average distance between all pairs of data-points
in the two clusters. The height of each of the U-shaped lines connecting the two objects in the
binary tree shows the average distance between them. Closest are Able1 and Able3, Able2 and
Able4, EMG1 and EMG2, SW1 and SW2, Auto1 and Auto2. Then at the next level, all the
groups with able-bodied data are close, and so are the groups with EMG-triggered and switch-
triggered data. Then farthest are the groups with iSCI data and those with the able-bodied data.
The able-bodied gait initiation (based on QoF and Av. Eig.) was found to be significantly
(p=0.0002) different from the pathological iSCI gait initiation. Figure 5.10 shows the
Mahalanobis distance between each pair of group means, which is a dissimilarity measure
(higher the distance, more is the dissimilarity). Unlike the Euclidean distance, the Mahalanobis
distance takes into account the covariance and it is scale-invariant. In spite of the statistically
significant difference between the able-bodied and iSCI gait initiation, the EMG-triggered FES-
assisted iSCI gait initiation was closest to the able-bodied normative clusters with lowest
Mahalanobis distance (darker on the color scale) when compared to other trigger modes, as
Discussion
researchers in past but the resulting improvement in the gait dynamics was never investigated in
depth. Gait parameters like gait-speed, step length and time, stride length and time, and duration
of single and double support phases averaged over multiple gait cycles provide a static picture of
the gait dynamics, ignoring the gait time-series. EMG-triggered FES was postulated to provide
better coordination between the stimulated and volitional motor activity and the resulting
105
functional gain should be seen in terms of kinematic coordination [5.9]. A time-series model is
required to investigate kinematic coordination and stability of a dynamic task like gait. Carletti
et al. presented a univariate linear regression model using stride-to-stride interval for non-
invasive diagnosis of neurodegenerative diseases [5.19]. They could differentiate between the
Sclerosis, and healthy controls with a rate parameter and the variance of the error term of the
linear regression model. Our pathological gait analysis is based on a similar idea but with a
multivariate linear regression model using the kinematic data. We also looked at a rate parameter
(i.e. Av. Eig.) computed from the Jacobian of the linear model and a normalized stochastic
amplitude parameter (i.e., QoF ) computed from the sum of the squares of the error or residuals.
The EMG classifier detected the unloading of the limb and shift of body-weight to the
contralateral limb that takes place during double-support phase preceding the swing-phase of the
ipsilateral side [5.9]. It then triggered the FES-assisted stepping of the ipsilateral limb. It is
necessary that FES-assisted gait initiation be coordinated and stable. It should be noted that
push-off during iSCI gait wasn’t assisted by FES so the energy input during FES-assisted gait
initiation mostly came from the upper-body support on the walker. The upper-body movements
necessary to push the rolling walker and then pull the body forward, needed to be coordinated
with the FES-assisted lower limb movements during gait initiation. Faster transition and better
coordination during this transition phase will mean a gait more resilient to disturbances, which
should make the EMG-trigger a better and preferred command source for outdoor ambulation.
The Euclidean distance of the perturbation to the origin computed from the 36 states (as
shown in Figure 5.5) for iSCI subjects didn’t converge steadily towards zero, except for the
subject iSCI-1 walking with EMG-trigger. This is because most of the frontal plane movements
106
during walker-supported iSCI gait didn’t converge to a steady state. The FES-assisted
movements during iSCI gait were mostly in the sagittal plane and the frontal plane movements
were mostly volitional as a result of upper extremity interactions with the walker. More
prominent frontal plane movements of the partially paralyzed right limb of the subject iSCI-2
were due to weak hip abductors since only his left limb musculature were activated by FES.
model (Equation 3) was too high (~1020) for reliable computation of its inverse. This was due to
synergistic movement of the joints, which make joint angles covary during gait [5.17]. Principal
component analysis decreased the dimension of the linear regression model to five, which
accounted for greater than 90% of the variance in the original model (Figure 5.6). Mah et al. has
showed that gait kinematics has much lower dimensionality because of these movement
synergies. The dimension reduction to five improved the condition number (~10) of δX nT • δX n
by several orders of magnitude from 1020 to 10. The first 3 principal components accounted for
more than 70% of the variance observed, and primarily represented the movements in the sagittal
plane during able-bodied gait initiation (Figure 5.6 and 5.7). The goal of the reduced-dimension
model was to capture the convergence of the major joint trajectories to a steady state during gait
initiation. The convergence of the Euclidean distance of the perturbations towards the origin
computed from the first five principal components for EMG-triggered iSCI gait initiation looked
more similar to able-bodied data in both the iSCI subjects (Figure 5.8). It should be noted that
Figure 5.5 and 5.8 give a marginal view of the convergence since it only shows the Euclidean
distance of the perturbations from origin whereas the multivariate linear regression model
107
The linearity of the reduced-dimension model was chosen as one of the features of
normative gait initiation. The Quality of Fit ( QoF ) of the reduced-dimension linear regression
model was used to measure the linearity of the return-map data. The rate of convergence was
determined from the Jacobian of the linear model, which gave an estimate of the stability since
faster convergence from disturbances means a more stable gait. A linear return-map model with
high rate of convergence is desirable since that will ensure a stable gait initiation.
Able-bodied gait initiation was found to be more stable than FES-assisted iSCI gait
initiation (Figure 5.9). Interestingly, the dendrogram plot of Figure 5.9b showed that the cluster
for Able1 (age: 25 years) was closer to Able3 (age: 26 years) and the cluster for Able2 (age: 52
years) was closer to Able4 (age: 54 years) possibly showing the effect of age on those
parameters. Also the iSCI data clusters of same trigger mode were closer to each other. The
average steady-state gait speed reached by the 4 able-bodied subjects was much higher (1.12
m/sec) as compared to 0.16 m/sec for the 2 iSCI subjects. Based on the results from Dingwell et
al. it was assumed that the orbital gait stability was not greatly mediated by gait speed [5.12].
Nevertheless, gait speed is one of the main predictors of independent community ambulation and
the improvement in gait speed is necessary along with better coordination and stability [5.20].
Both the EMG-triggered and switch-triggered iSCI gait initiations were more similar to the
able-bodied data in terms of coordination and stability than auto-triggered iSCI data (Figure 5.9).
This is because auto-triggered iSCI gait had a fixed step frequency that could not be modulated
as needed during the gait initiation. Both the switch and the EMG trigger gave an opportunity to
modulate the step frequency that was necessary to fluidly transition from stationary standing to a
steady gait. Overall, EMG-triggered FES-assisted gait initiation was found to be closest to the
able-bodied data based on Mahalonobis distance between the clusters (Figure 5.10). This
108
proximity can also be seen in Figure 5.9 where EMG-triggered and switch-triggered iSCI data
have comparable QoF5 but the EMG-triggered iSCI data has lower Av. Eig than switch-triggered
iSCI data. This means faster convergence to steady state and better stability for EMG-triggered
FES-assisted iSCI gait. Also, the user of the EMG-triggered FES-system had to exercise the
voluntary muscles in conjunction with the stimulated muscles to trigger the steps which may
Conclusions
The EMG-triggered FES-assisted iSCI gait was shown to be more normative in terms of
coordination and rate of convergence during gait transition from standing to walking. This paper
developed a method based on prior work on time-series analysis of dynamical systems and
applied that to evaluate gait transitions using a well accepted tool called return-map analysis. The
results presented in this paper indicated that the user of the EMG-triggered FES-system
successfully exercised the voluntary muscles in conjunction with the stimulated muscles to
synergistically generate the volitional EMG to trigger the FES-assisted steps. This enhanced the
coordination between the voluntary and stimulated muscles, as presented in the results. The
EMG-triggered FES-assisted iSCI gait was therefore more normative than switch- and auto-
triggered iSCI gait during gait transitions which is needed for outdoor walking. The feasibility
study and gait evaluation were performed with a tethered system in the laboratory which needs to
be translated to a portable system for outdoor use. The newer family of implantable stimulator-
telemeters (IST-12) has 2 EMG channels and the capability to perform the signal processing
required by the GED. The GED and the FES–controller can be implemented in the current
portable External Control Unit (ECU) with the IST family of implants. Based on the results
presented in this paper from a laboratory based evaluation, it is hypothesized that EMG-triggered
109
portable FES-system will allow the user to transition more fluidly and stably from one speed to
another by modulating the step frequency, which is required for better outdoor ambulation.
References
5.5. D. Graupe, “EMG pattern analysis for patient-responsive control of FES in paraplegics for
walker-supported walking,” IEEE Trans. Biomed. Eng., vol. 36, no. 7, July 1989, pp. 711-
719.
5.6. D. Graupe, and H. Kordylewski, “Artificial neural network control of FES in paraplegics
for patient responsive ambulation,” IEEE Trans. Biomed Eng., vol. 42, no. 7, Jul. 1995, pp.
699-707.
5.8. A. Dutta, R. Kobetic, and R. J. Triolo, “EMG based triggering and modulation of
stimulation patterns for FES assisted ambulation – a conceptual study,” presented at XXth
Congress of the International Society of Biomechanics, Cleveland, OH, Aug. 2005.
5.9. A. Dutta, R. Kobetic, and R. Triolo, “Ambulation after incomplete spinal cord injury with
EMG-triggered Functional Electrical Stimulation,” IEEE Transactions on Biomedical
Engineering, 55:2, February 2008.
5.10. Y. Hurmuzlu, and C. Basdogan, “On the measurement of dynamic stability of human
locomotion,” Journal of Biomechanical Engineering, vol. 116, 1994, pp. 30-36.
110
5.11. Y. Hurmuzlu, C. Basdogan, and D. Stoianovici, “Kinematics dynamic stability of the
locomotion of post-polio patients,” Journal of Biomechanical Engineering, vol. 118, no. 3,
1996, pp. 405–411.
5.12. J. B. Dingwell, K. H. Gu, and L. C. Marin, “The effects of sensory loss and walking speed
on the orbital dynamic stability of human walking,” Journal of Biomechanics, vol. 40,
2007, pp. 1723-1730.
5.13. R. Kobetic, and E.B. Marsolais, “Synthesis of paraplegic gait with multichannel functional
neuromuscular stimulation,” IEEE Trans Rehab Eng., vol. 2, no. 2, 1994, pp. 66-79.
5.14. R. Kobetic, R. J. Triolo, and E. B. Marsolais, “Muscle selection and walking performance
of multichannel FES systems for ambulation in paraplegia,” IEEE Trans. Rehabil. Eng.,
vol. 5, no. 1, Mar. 1997, pp. 23–29.
5.15. A. Dutta, and R. Triolo, “Stability analysis of Functional Electrical Stimulation (FES)-
assisted overground gait in an incomplete spinal cord injured subject,” Journal of
Biomechanics, vol. 39, 2006, pp. S113.
5.16. R. A. Johnson, and D. W. Wichern, Applied Multivariate Statistical Analysis. Prentice Hall,
2002.
5.18. I. T. Jolliffe, Principal Component Analysis, New York: Springer, 1986, pp. 44-48, 269-
296.
5.19. T. Carletti, D. Fanelli, and A. Guarino, “A new route to non invasive diagnosis in
neurodegenerative diseases,” Neuroscience Letters, vol. 394, 2006, pp. 252-255.
5.20. P. J. Friedman, D. E. Richmond, and J. J. Baskett, “A prospective trial of serial gait speed
as a measure of rehabilitation in the elderly,” Age and Aging, vol. 17, no. 4, 1988, pp. 227-
235.
5.22. C. A. Miller, and M. C. Verstraete, “Determination of the step duration of gait initiation
using mechanical energy analysis,” Journal of Biomechanics, vol. 29, no. 9, 1996, pp.
1195-1199.
111
5.24. D. A. Winter, The Biomechanics and Motor Control of Human Movement, 2nd edition, New
York: Wiley, 1991.
5.25. C. Angeloni, P. O. Riley, and E. D. Krebs, “Frequency Content of Whole Body Gait
Kinematic Data,” IEEE Trans. Rehabilitation Engineering, vol. 2, no. 1, 1994, pp. 40-46.
112
Figures
Loadings
Find the distance from centroid
of case ‘T’ in feature space
Euclidean distance
Distance Y Trigger
< swing-
threshold phase
STOP
Double- Y Wait for
support manual
> 3 sec? trigger
N
Get next
EMG sample
Figure 5.1: Laboratory setup for EMG-triggered FES-assisted walking shown with a flowchart
for the EMG-based gait event detector for triggering FES-assisted steps.
113
Figure 5.2: Top panel: Selection of optimum cut-off frequencies for low-pass filtering the
kinematic data. Bottom panel: Most of power content in the signals was below the
optimum cut-off frequency, which were 6 Hz for able-bodied and 3.5 Hz for iSCI
data.
114
volume of marker data capture (~4-5 steps)
115
Figure 5.4: Typical pelvis motion in the direction of progression during gait initiation
116
Figure 5.5: Euclidean distance from the origin of the perturbation of the 36 states during gait
initiation at the maximum left knee flexion. Left panel: able-bodied data (4 subjects).
Middle panel: iSCI data (subject C1). Right panel: iSCI data (subject C2).
[Normative: 4 subjects, 10 trials each; iSCI EMG-trigger: 2 subjects, 10 trials each;
iSCI Switch-trigger: 2 subjects, 10 trials each; iSCI Auto-trigger: 2 subjects, 10 trials
each]
117
Figure 5.6: Percent Variance Accounted For (%VAF) by the Principal Components (PC). Top
panel: able-bodied data. Middle panel: iSCI-1 walking with EMG, switch, and auto
triggered FES. Bottom panel: iSCI-2 walking with EMG, switch, and auto triggered
FES. All the plots show the data averaged over 6 gait events.
118
Figure 5.7: Typical loading of the first 3 Principal Components (PCs) on the joint angles (HA:
Hip Angle, KA: Knee Angle, AA: Ankle Angle) found from the weight matrix W of
the subject Able1. The prefix ‘l’ indicates the left side and ‘r’ indicates the right side.
The suffix ‘x’ denotes sagittal plane, ‘y’ denotes frontal plane, and ‘z’ denotes
transverse plane for the joint angles.
119
Figure 5.8: Euclidean distance from the origin of the perturbation of the 5 principal components
at maximum left knee flexion Left panel: able-bodied (4 subjects). Middle panel:
iSCI-1 subject C1. Right panel: iSCI-2 subject C2. [Normative: 4 subjects, 10 trials
each; iSCI EMG-trigger: 2 subjects, 10 trials each; iSCI Switch-trigger: 2 subjects, 10
trials each; iSCI Auto-trigger: 2 subjects, 10 trials each]
120
Figure 5.9: Top panel: Scatter plot of QoF and Av. Eig. at 6 gait events for the groups; the 4
able-bodied subjects: Able1, Able2, Able3, Able4, and the 2 iSCI subjects with
different trigger modes: EMG1, EMG2, SW1, SW2, Auto1, Auto2. Bottom panel:
MANOVA cluster dendrogram plot of the groups
121
Figure 5.10: Mahalanobis distances matrix between each pair of group means
122
CHAPTER 6
DEVELOPMENT OF AN IMPLANTED INTRAMUSCULAR EMG-TRIGGERED FES-
SYSTEM FOR AMBULATION AFTER INCOMPLETE SPINAL CORD INJURY
Abstract
Ambulation after spinal cord injury is possible with the aid of neuroprosthesis employing
functional electrical stimulation (FES). Individuals with incomplete spinal cord injury (iSCI)
retain partial volitional control of muscles below the level of injury, necessitating careful
integration of FES with intact voluntary motor function for efficient walking. In this study, the
intramuscular electromyogram (iEMG) was used to detect the intent to step and trigger FES-
assisted walking in two volunteers with iSCI via an implanted neuroprosthesis consisting of two
channels of bipolar iEMG signal acquisition and 12 independent channels of stimulation. The
detection was performed with two types of classifiers – a threshold-based classifier that
compared the running mean of the iEMG with a discrimination threshold to generate the trigger
and a pattern recognition classifier that compared the time-history of the iEMG with a specified
template of activity to generate the trigger whenever the cross-correlation coefficient exceeded a
discrimination threshold. The two classifiers were presented in a random order to one of the iSCI
volunteers for evaluation under indoor laboratory conditions. The pattern recognition classifier
False Positive triggers. The overall True Positive rates for the threshold-based classifier were
61.6% and 87.2% for the right and left steps with overall False Positive rates of 38.4% and
33.3%. The overall True Positive rates for the left and right step with the pattern recognition
classifier were 57.2% and 93.3% and the overall False Positive rates were 11.9% and 24.4%.
Subjects showed no preference for either the threshold- or pattern recognition-based classifier as
determined by the Usability Rating Scale (URS) score collected after each trial. The average
123
URS score was 2 out of a 7 point scale for both the classifiers, indicating that both the classifiers
Introduction
electrically activating a customized set of muscles selected to address individual gait deficits
lower extremities [6.1]. Users can trigger each step with a manual switch and progress through
the customized pattern of stimulation to achieve walking function. The potential for triggering
FES from the electromyographic (EMG) activity of muscles which remain under volitional
control after partial paralysis to coordinate the actions of the stimulated muscles with voluntary
movement has previously been evaluated with signals acquired from the surface of the skin [6.2].
The main objective of this study was to develop a method to identify optimal command sources
and evaluate the feasibility of detecting the intent to take a step using the intramuscular
partially paralyzed muscles in two implant recipients with incomplete spinal cord injury (iSCI).
The overall aim was to specify the development of a new command and control interface to
trigger FES-assisted stepping that can be implemented with two channels of intramuscular EMG
electrodes with a multichannel implantable stimulator-telemeter (IST) [6.3], [6.4]. This report
ambulation that integrated stimulated and voluntary muscle activity in a fully implantable
neuroprosthesis.
Prior work has shown that gait event detection for the control of electrical stimulation is
possible with physical sensors such as force sensitive resistors, accelerometers, gyroscopes, as
124
well as with biopotentials such as the electroneurogram and electromyogram [6.5-6.8]. The
electromyogram temporally precedes the generation of force in the muscle and resulting
movement of a joint. This makes the EMG an attractive signal for detection of intent so that the
desired movement can then be assisted with FES. Graupe et al. proposed EMG based control of
FES with time-series model of surface EMG (sEMG) [6.9]. Auto-regressive (AR) time-series
models of four channels of sEMG has been shown to give information related to limb function
[6.10] and can provide patient-responsive control of FES-assisted walking [6.11]. The AR
parameters determined offline from the time-series model of the above-lesion upper-trunk
surface EMG were used to identify standing and walking functions [6.11]. Recently, Graupe and
Kordylewski presented a neural network based EMG classifier with on-line learning capabilities
to control stimulation for achieving stepping motions in individuals with complete paraplegia
[6.12, 6.13]. Our preliminary studies indicate that the surface EMG from partially paralyzed
muscles of individuals with incomplete paraplegia can provide significant information related to
the volitional activity of the muscle and can be used for gait-event detection [6.14, and 6.15].
Thorsen et al. have showed improved wrist extension with stimulation controlled by surface
EMG from partially paralyzed wrist extensors [6.16]. Futami et al. showed the feasibility of
proportional control of FES with the sEMG from the same muscle (partially paralyzed knee
This chapter outlines the development of a binary classifier as a command interface for
foot-off intent detection with iEMG from partially paralyzed muscles via implanted recording
electrodes in two subjects wit hiSCI. The classifier detected the intent to initiate stepping and
integrated that information in the FES controller to trigger FES-assisted stepping with that limb.
The real-time operation of the iEMG–based classifier was evaluated by one iSCI volunteer for
125
FES-assisted walking under laboratory conditions. The technical challenges and preliminary
results from initial attempts to implement iEMG-based control systems in real-time with
Methods
Subjects
Two male subjects with incomplete spinal cord injury volunteered for this study. Subject
iSCI-1 was a male volunteer with C6 incomplete spinal cord injury (ASIA C). Subject iSCI-2
was a male volunteer with T1 motor and C6 sensory incomplete spinal cord injury (ASIA D)
who could walk only short distances with great difficulty without the assistance from FES. They
stimulating electrodes [6.3] and two implanted intramuscular recording electrodes as part of a
Subject iSCI-1 presented with more extensive paralysis on his right side. In a single
posterior adductor magnus, tensor fasciae latae, and tibialis anterior muscles on the right side
only. Subject iSCI-2 presented with more extensive weakness primarily on his left side. He had
stimulating electrodes only on his left side to recruit iliopsoas, vastus intermedius/lateralis, tensor
fasciae latae, gluteus medius, gluteus maximus, posterior portion of adductor magnus, and
tibialis anterior (2 electrodes). He was re-evaluated for the 12-channel EMG-controlled system
and upgraded to the IST-12 for iEMG triggered walking. The original IRS-8 was disconnected
from the stimulating electrodes and replaced with the IST-12. At the time of his revision
126
surgery, additional intramuscular stimulating electrodes were inserted into the bilateral erector
spinae, right iliopsoas, right gluteus maximus, and right tibialis anterior to complete the 12-
channel system. Temporal patterns of stimulation to activate the muscles were customized for
each subject’s individual gait deficits according to established tuning procedures [6.7], [6.8] in
Informed consent was obtained from both the subjects before their participation and the
Institutional Review Board of the Louis Stokes Cleveland Department of Veterans Affairs
intramuscular EMG recording electrodes [6.2-6.4]. Therefore two muscles and the best location
within those muscles had to be selected for implanting intramuscular recording electrode based
on surface EMG (sEMG) data. Subjects were asked to walk either volitionally or with surface
FES for the pre-surgery data collection. The experimental setup for collecting sEMG data during
walking is shown in Figure 6.1. Surface EMG signals were collected from gluteus medius (GM),
biceps femoris (BF), medial gastrocnemius (MG), rectus femoris (RF), tibialis anterior (TA), and
The sEMG was collected using Ag/AgCl electrodes with 2 cm. inter-electrode distance
following the SENIAM guidelines [6.22]. The sEMG signals were amplified and low-pass
(fcutoff=1000 Hz) filtered by CED 1902 amplifiers (Cambridge Electronic Design, England)
before being sampled at 2400 Hz (AT-MIO-64F-5, National Instruments, USA) in the host
personal computer. The CED 1902 amplifier has a switching circuit (clamp) which was activated
by a square pulse that disconnected the electrode inputs from the amplifier and connected it to
the common electrode to prevent stimulation artifact. The input channels of CED 1902 were
127
clamped when the stimulation pulses were applied to the muscles to prevent stimulation artifact.
The blanked portion of the sEMG was reconstructed with the average value of the sEMG in the
pre- and post- blanking periods [6.23] as shown in Figure 6.2a. Then the whole sEMG pattern
was low pass filtered (5th order zero-lag Butterworth, fcutoff=3 Hz) to get the linear envelope
(Figure 6.2b). The gain of each channel was set separately in the CED 1902 amplifiers to prevent
Baseline sEMG data were collected during 3 seconds of initial standing before the start of
each trial. During each trial, the subjects were asked to reach a self selected speed within ~5m of
the start position and then decelerate to come to rest at the end of the walkway. They had to wait
in the terminal stance for 3 seconds at the end of every trial. The experimental protocol is shown
in Figure 6.3. The subjects made multiple passes across the straight level walkway. Gait events
(foot-strike and foot-off) were derived from foot-floor contact patterns obtained from insole-
mounted foot switches placed bilaterally at medial and lateral heel, first and fifth metatarsal, and
big toe, and confirmed with the kinematic data that were acquired simultaneously.
Each pass constituted one trial and multiple trials were collected during a session. The
sEMG data were sampled and processed in MATLAB TM R14 (The MathWorks, Inc., USA). The
sEMG linear envelopes (LEs) during a gait cycle were then divided into double-support (DS)
phase – when both the feet were in contact with the ground, and swing (SW) phase – when the
foot was not in contact with the ground [6.27]. The sEMG LE during each trial was normalized
The normalized LEs of each muscle were divided into two classes: the class ‘True’ was
comprised of LEs during double-support phase prior to foot-off and the class ‘False’ consisted of
the LEs during terminal stance and initial standing. Half of the data were randomly allocated to
128
training and used to find a characteristic pattern of activation by ensemble averaging the LEs.
The characteristic pattern found for the class ‘True’ was cross-correlated with the LEs from the
other half of the data (test data) for the classes ‘True’ and ‘False’. A Receiver Operating
Characteristics (ROC) curve showed the tradeoff between sensitivity (i.e. True Positive rate) and
1 – specificity (i.e. False Positive rate) of the binary classifier [6.24]. A Discriminability Index
(DI) was defined as the area under the ROC curve (AUC) which gave a measure of performance
for the binary classifier [6.25]. The data were randomly partitioned ten times into training and
test data-sets for a 10-fold cross-validation. Therefore, 10 ROC curves for the pattern recognition
classifier were generated by randomly pooling the LEs into training and test data-sets. The DI
was computed for each ROC curve and then averaged to find the mean DI and standard deviation
(SD(DI)) for the pattern recognition classifier [6.25]. This identified the muscles yielding the
best separation of classes, and hence the primary targets for the implanted recording electrodes.
The best location for implantation of the intramuscular recording electrode was estimated
based on a similar analysis of the DI form the sEMG acquired from different locations on the
bellies of the target muscles. A matrix of sEMG electrodes was placed to cover the whole muscle
belly, as shown in Figure 6.4. Multiple bipolar sEMG recordings were made along the length of
the muscle. The sEMG data were collected with the same experimental protocol and analyzed
similarly to find the best location on the muscle belly that had the highest DI. The best location
was noted with respect to the anatomical landmarks for easy identification during surgery when
The implantation procedure for the iEMG electrode is shown in Figure 6.5. First a
stimulation probe was inserted subcutaneously to the best location that was identified based on
anatomical landmarks, as shown in Figure 6.5a. Stimulation was applied by clipping a cable to
129
the probe to ensure that the tip of the probe is at the desired site with viable muscle fibers. A
peelable polymer sheath was then inserted on the top of the probe with the help of the markings
such that its tip approximately coincided with the probe tip, as shown in Figure 6.5b. The probe
was then removed and the intramuscular EMG electrode was inserted in the place of the probe
with the help of the lead carrier. The cannula-like lead carrier held the iEMG electrode lead in
place during insertion through the sheath. The lead carrier was then removed which left the
iEMG electrode lead in the peelable sheath, as shown in Figure 6.5c. The polymer sheath was
then gently peeled off leaving the iEMG electrode at the selected location, as shown in Figure
6.5d.
customized for each subject [6.20, and 6.21] in order to achieve forward stepping in a rolling
walker, as shown in Figure 6.6. The stimulation current was fixed at 20 mA and the stimulation
frequency was fixed at 20 Hz (50 ms inter-pulse-interval). Each step could be triggered either by
the depression of a manual switch or with an iEMG-based classifier to progress through the
stimulation pattern.
Two kinds of iEMG-based classifier – one based on thresholding and the other based on a
pattern recognition algorithm [6.2] were developed. The left foot-off (i.e., the intent to initiate
left swing) and right foot-off (i.e., the intent to initiate right swing) were detected sequentially
and independently by the classifiers to trigger FES-assisted left and right steps respectively. The
20 Hz sampling rate for the iEMG signal corresponded with the stimulation frequency. The
sampling period was divided between the two iEMG channels which were multiplexed in the
IST-12 to use the same gain and integration circuitry. The iEMG from each channel was pre-
amplified, rectified, and blanked in the hardware during the electrical stimulation before being
130
integrated for 10 ms. The iEMG of the each channel was sampled and integrated for 10ms
sequentially one after the other every 100 ms. The IST-12 therefore telemetered back to the
External Control Unit (ECU) the integrated iEMG of each channel at 10 Hz.
The real-time operation cycle in the IST is illustrated in Figure 6.7. The real-time cycle of
iEMG channel (i.e., 1 or 2) to be sampled was set after 4ms into the real-time cycle with the gain
set to zero to minimize switching transients. After 6ms into the real-time cycle, the gain was then
set to an appropriate level such that it did not cause saturation. After 13ms into the real-time
cycle, the integrator started integrating the iEMG for 10ms. The integrated iEMG was read after
23ms into the real-time cycle. After 25ms into the real-time cycle, the iEMG channel gain was
reset back to zero so that 12-channels of stimulation can be applied to the muscles from 28ms
onwards. About 10ms of time was needed at the end of the real-time cycle to update the events
and do housekeeping tasks. The data from each iEMG channel were reconstructed during the
next 50ms time step and held at that value when the other channel was sampled. A 10-point
moving-average smoothing filter was applied to the iEMG signal in the ECU to reduce random
noise.
The threshold-based binary classifier started computing the running mean of the relevant
iEMG signal after waiting a fixed duration (twait) from the start of the stimulation pattern of a
step. The ‘twait’ was found from the training data. The wait duration was long enough such that
the iEMG activity processed by the classifier represented the unloading of the contralateral limb
in preparation for the swing phase of that side. When the running mean exceeded a selected
threshold (thrTC), the stimulation pattern advanced and stepping of the contralateral limb was
triggered. The pattern recognition classifier started processing the iEMG time history at the same
131
time as the threshold-based classifier started processing the iEMG level. The pattern recognition
classifier cross-correlated a windowed portion (of length twin) of the relevant iEMG signals to
detect feature templates required for triggering the contralateral step. A trigger was then
Classifier parameters such as twait, thrTC, twin are illustrated in Figure 6.8. The start of the
stimulation is shown to coincide with foot-off in Figure 6.8 to coincide just for illustration
purposes, and such timing may not necessarily be true in reality. The foot-off may precede the
start of stimulation pattern for stepping in subjects who have enough volitional strength or there
Figure 6.8 also shows trigdelay which is the delay in generating the trigger by the threshold-
based classifier primarily due to the 10-point moving average FIR filter. The group delay of the
10-point moving average FIR filter is (10-1)/2=4.5 samples long for each channel of iEMG (i.e.
225 msec for a nominal stimulating and sampling rate of 20 Hz). By definition, it takes 9
samples i.e. 450 msec to reach the unit amplitude after the unit step is applied to the 10-point
moving average filter. These delays were largely unavoidable due to the processing embedded
Most of the parameters for the pattern recognition and the threshold-based classifiers were
found from the training data collected with the IST-12 during FES-assisted switch-triggered
walking in the laboratory. Training data collection followed the same experimental protocol
explained earlier in Chapter 4 and shown in Figure 6.3. The iEMG patterns during walking were
normalized by the average of the baseline iEMG data collected during 3 seconds standing. The
iEMG patterns during steps taken while walking constituted the class ‘True’ and the iEMG
patterns during initial standing and terminal step constituted the class ‘False’ of the binary
132
classifier for that side. All the iEMG patterns in both the classes were time normalized such that
they each contained 50 data points. This allowed comparison of the iEMG patterns and iEMG
levels at different time points to be selected for classification by the Discriminability Index (DI).
All patterns were randomly partitioned 10 times into training and test data-sets for a 10-
fold cross-validation. The data sets were used to generate separate ROC curves for each of the
classifiers. Ten ROC curves for each point in the gait cycle were generated for the threshold-
based classifier by randomly pooling data from different iEMG patterns in the training and test
data-sets. The pattern recognition classifier works with a pattern of iEMG activity. Therefore, 10
ROC curves for each phase of the gait cycle, such as left and right swing (SW) and double
support (DS), were generated from the same training and test data-sets as used to quantify the
performance of the threshold-based classifier The DI was computed from each ROC curve and
then averaged to find the mean DI and standard deviation [6.25]. The classification region was
defined as the largest set of contiguous data points with a high (>0.7) mean DI where the
decision has to be made whether to trigger the FES-assisted contralateral step or not. The
classification regions were identified from the results from the threshold-based classifier and
used to define the wait time (twait), which was equal to the elapsed time from the start of the FES-
For the pattern recognition classifier, the iEMG patterns from the class ‘True’ in the
classification region were ensemble averaged and served as the feature for classification. One
feature served as the template for one muscle for identifying left foot-off and another feature
served as the template for another muscle for right foot-off. The window-size parameter twin was
equal to the average duration of that classification region for all the steps in the training data. The
feature was cross-correlated with all the iEMG patterns in the class ‘True’ and the average cross-
133
correlation coefficient served as the thrPRC. The parameter thrTC was equal to the average running
mean of the iEMG signal over the classification region for all the iEMG patterns in the class
‘True’.
The pattern recognition and threshold-based classifiers were tested separately during online
The two classifiers were developed as separate models in the Simulink (The MathWorks,
Inc., USA) and compiled via C-code to run standalone in the portable ECU that also executed the
control algorithm to deliver pre-programmed stimulation patterns to the muscles via the IST-12.
The IST-12 telemetered back the iEMG signal that was used by the classifier to trigger the
The standalone FES controller occasionally experienced overrun errors which occurred
when the ECU couldn’t finish all the required tasks in the 50ms time window for real-time
operation. The stack size for the FES controller running in ECU was set at 4096 which was the
space needed for the arguments, the stack variables, and all the variables. The computational
demand of the classification algorithm alone was investigated with a Matlab (The MathWorks,
Inc., USA) utility called ‘XPCBENCH’. The Simulink model for the classification algorithm
consisting only of the pattern recognition or the threshold-based components was made to run in
a PC environment with synthesized EMG signals as the inputs and a binary output as the trigger
signal. The pattern recognition algorithm mostly experienced overrun error and was simplified
from the one published earlier [6.2] such that only one feature was cross-correlated with iEMG
from one muscle for classification. The minimum achievable sample time for running only the
simplified pattern recognition classifier model (i.e., no FES stimulation) was 3800 µs on an Intel
134
486DX 40 MHz. The 50ms real time window was sufficient for running the pattern recognition
classification algorithm in isolation. The overrun error issue during online operation even with
the simplified pattern recognition classifier was possibly due to the hardware events required to
integrate the iEMG signal with implanted stimulation (like changing the stimulus, the iEMG
gain, the iEMG channel, etc), which took up most of the cycle time. Figure 6.7 shows the
hardware events taking place in the 50 ms sampling window, which provided only about 9 ms of
free cycle time prior to the initiation of the next cycle, which was probably insufficient for robust
The stimulation pattern for the first step was triggered with a manual switch. The classifier
started scanning the iEMG after waiting ‘twait’ duration from the start of stimulation pattern of
that step to detect the intent to initiate stepping with the contralateral limb. For safety, the
classifier was allowed to trigger the stimulation pattern for stepping of the contralateral limb
within 1 sec after entering the double support phase. If the user was unable to trigger the
stimulation pattern with iEMG (i.e., a “false negative”) within 1 sec after entering the double
support phase then the manual switch could be used as a trigger, and the classifier then resumed
processing the iEMG to determine the intent to trigger subsequent steps. When the user was
unable to stop with the classifier (i.e., “false positive”) then a manual switch could be used to
Subject iSCI-2 evaluated walking and stopping with both the classifiers while the true
positive (1-false negative) and false positive rates were recorded as measures of performance.
The subject walked with the iEMG triggered FES assisted stepping on a straight walkway across
the gait laboratory. The iEMG classifier was started with a manual switch during standing to
trigger the first step. After that the iEMG triggered the steps during ambulation across
135
approximately 8m before stopping with the iEMG classifier. The delay (trigdelay) in triggering
the FES-assisted step from the instant the desired muscle activity was detected was also recorded
to further compare the technical performance of the classifiers. The pattern recognition and the
threshold-based classifiers were presented in a random order during two days of evaluation.
Total eight trials (39 left steps and 39 right steps) for the threshold-based classifier and nine trials
for the pattern recognition classifier (45 left steps and 42 right steps) were captured.
After each trial, the subjective assessments of perceived ease of use of each classifier
during FES-assisted stepping in real-time were evaluated with Usability Rating Scale (URS)
[6.29]. The URS, depicted in Figure 6.10 is a 7 point scale originally developed to determine
differences in the perceived ease of use. After each walking trial, the subject was asked to rate
the EMG-triggered system as “Difficult”, “Moderate” or “Easy” to use. After selecting one of
those three choices, the subject was asked to refine his selection as “Very”, Moderately” or
“Barely”. Each final subjective rating was assigned a numerical value from -3 (very difficult) to
+3 (very easy). Each walking trial was evaluated independently so the subject only rated his
most recent walking experience and was not required to compare his current perception to a prior
walking trial.
Results
The left medial gastrocnemius (MG) and the right erector spinae (ES) were selected as the
command sources for iSCI-1. The best location for intramuscular EMG was estimated based on
the DI from the surface EMG from the Left MG and Right ES of iSCI-1, which are shown with
color scale in Figure 6.11a. The best location on Left MG was 6 cm medial and 18 cm distal
from the popliteal crease line. The best location found on Right ES was 2.75 cm lateral to the
spinous process of L2 and 2 cm superior to the L2 level. The Right MG and Left MG were
136
selected as the command sources for iSCI-2. The best location for intramuscular EMG that was
estimated based on the DI from the surface EMG from the Left MG and Right MG of iSCI-2 are
shown with color scale in Figure 6.11b. The best location found on the Left and Right MG was 6
The iEMG was collected and processed to determine the parameters for each classifier. The
iEMG pattern during walking was divided into swing phase (SW) and double support (DS)
phases of gait based on the foot-floor contact patterns obtained from the foot switches. Figure
6.12a and Figure 6.12b show the Discriminability Index (DI) at each time normalized data point
from the ROCs obtained with the threshold-based classifier during the gait cycle that was used to
identify the classification region. Figure 6.12a shows that the iEMG pattern of the left medial
gastrocnemius had DI close to 0.5 at the end of left SW phase, a DI close to 0.8 at the end of left
DS phase, a DI close to 0.95 at the end of right SW phase, and a DI close to 1 which means very
good discriminability during the right DS phase. Figure 6.12b shows that the iEMG pattern of
right ES had low DI at the end of the left SW phase of gait which was close to 0.5, a poor
discriminability with DI less than 0.5 during left DS, a good discriminability with DI close to 1
during right SW and again a good discriminability with DI close to 0.9 during right DS.
The gait phases of interest for the classifier are the left and right DS phases, which should
be a subset of the classification region when the decision has to be made to trigger the FES-
assisted SW phase of the contralateral limb. For the right ES, left DS was not good for
classification since the DI was less than 0.5. Figure 6.12b shows that the right ES had a
classification region that started at data point 25 in the right swing phase (SW) and covered the
137
The iEMG training data were collected in absence of right ES stimulation; however the
electrical stimulation of the right ES during FES-assisted walking created an artifact in the iEMG
from right ES during the right SW that is shown in Figure 6.13. Taylor and Chappel have shown
that the ratio of the volitional EMG with the intended effort changes with the intensity of the
stimulation and the time since the stimulation pulse as in this case when the EMG from the same
muscle is the control signal for stimulation [6.18]. Because of this inhibitory action of the
stimulation on the iEMG from the same muscle during right SW (right ES was only stimulated
during right SW), the classification region was started after the end of electrical stimulation to
assist right ES i.e., at the data point 1 in the right DS. Right ES was used by the classifier to
trigger the left step in this classification region which started at 1.8 seconds into the right step
and was 0.8 seconds long, as found from the foot switch data. The right step determined from the
foot switch data started on an average 0.2 seconds before the start of its stimulation pattern so
twait=1.6 seconds and twin=0.6 seconds for the left step classifier. The thrTC in the classification
region was about 1.1 times the baseline level of the iEMG during quiet standing.
Since right ES was found suitable only to trigger the left step, left MG was selected to
trigger the right step. The classification region started at data point 2 in the left DS and covered
the whole left DS of the gait. Left MG was used to trigger the right step during this classification
region which started 1.45 seconds after the start of the left step and was 0.9 seconds long, as
found from the foot switch data. The left step found from the foot switch data started 0.15
seconds before the start of the stimulation pattern so twait=1.3 seconds and twin=0.75 seconds for
the right step classifier. The thrTC during the classification region was equal to 0.45 times the
baseline level of iEMG during quiet standing. The pattern recognition classifier used a
monotonically increasing feature – [0.2418 0.3406 0.3984 0.5056 0.5231] with a cross-
138
correlation threshold, thrPRC=0.6. The monotonically increasing nature of the feature is necessary
since the iEMG pattern in that case will cross the threshold only once in the classification region.
The subject iSCI-1 is currently using the iEMG classifier in the laboratory during his
rehabilitation phase. No performance data have been collected yet for the subject iSCI-1 since he
Figure 6.14a and Figure 6.14b show the Discriminability Index (DI) at each time
normalized data point during the gait cycle that was used to identify the classification region of
iSCI-2. The Figure 6.14a shows that the iEMG pattern of right MG had moderate DI at the end
of the left SW phase of gait (~0.8), a good DI (~0.8) as the beginning of left DS and a poor DI at
the end (~0.5), a poor discriminability with DI close to 0.5 during right SW and a poor
discriminability during the right DS phase with DI close to 0.6. The Figure 6.14b shows a
moderate DI at the start which deteriorated to a low DI by the end of left SW phase of gait
(~0.5), a low DI close to 0.5 during left DS, a moderate discriminability with DI close to 0.7 at
the end of right SW, a moderate discriminability during the right DS phase with DI close to 0.7.
Figure 6.14a shows that the right MG had a classification region starting at data point 35 in
the left SW and extended till data point 2 in the left DS of the gait. Right MG was used by the
classifier to trigger the right step during this classification region which started at 1.5 seconds
into the left step and was 0.85 seconds long, as found from the foot switch data. The left step
found from the foot switch data started 0.3 seconds before the start of its stimulation pattern so
twait=1.2 seconds and twin=0.55 seconds for the right step classifier. The thrTC was equal to 1
times the baseline magnitude of iEMG during quiet standing. The pattern recognition classifier
used a monotonically increasing feature – [1.0436 1.0651 1.0900 1.1423 1.2358] with a
139
Figure 6.14b shows that the left MG had a classification region starting at data point 30 in
the right SW and extended till the end of the right DS of the gait. Left MG was used by the
classifier to trigger the left step in this classification region which started at 1.55 seconds into the
right step and was 0.85 seconds long, as found from the foot switch data. The right step
determined from the foot switch data started 0.5 seconds before the start of the stimulation
pattern so twait=1.05 seconds and twin=0.35 seconds for the left step classifier. The thrTC was equal
to 1.1 times the baseline magnitude of iEMG during quiet standing. The pattern recognition
Both the the pattern recognition and the threshold-based classifier incurred similar delays
(trigdelay) in triggering the FES-assisted step. The elapsed time from the instant of the desired
muscle activity to initiation of stimulation was 0.53 ± 0.12 seconds. Table 6.1 shows the
performance of the threshold-based classifier for iSCI-2. The threshold-based classifier for
triggering the right step had a false positive rate of 38.4% and a true positive rate of 61.6%. The
threshold-based classifier for triggering the left step had a false positive rate of 33.3% and a true
positive rate of 87.2%. Table 6.2 shows the performance of the pattern recognition classifier for
iSCI-2. The pattern recognition classifier for triggering the right step had a false positive rate of
11.9% and a true positive rate of 57.2%. The pattern recognition classifier for triggering the left
step had a false positive rate of 24.4% and a true positive rate of 93.3%. The standard deviation
of the false positives and false negatives from trial-to-trial as seen in Tables 6.1and 6.2 indicated
that the threshold-based classifier was more sensitive to natural variations in the EMG activity
from the left side of iSCI-2 which had more extensive weaknesses.
140
The average Usability Rating Scale (URS) score was found to be 2 in a 7 point scale for
both the classifiers, indicating that both the classifiers were moderately easy to use. The subject
didn’t show any preference for a particular classifier as determined from URS score after each
trial.
Because the threshold-based classifier was found to be susceptible to false positives while
the pattern recognition classifier was found to be susceptible to false negatives. Therefore after
testing each classifier separately in the gait laboratory, the final implementation in the ECU for
use by the subject had the two classifiers integrated in to a two-stage classifier to reduce the
overall false positives, as illustrated in Figure 6.9. Moreover, it was found during the testing that
the classification threshold (thrTC) for the threshold-based subsystem was sensitive to natural
variation in the EMG activity over time. Therefore a provision was made in the final
implementation in ECU to retrain only the classification threshold (thrTC) for the threshold-based
subsystem which could be adjusted during online operation using binary search. In the two-stage
implementation, the threshold-based subsystem served as the primary classifier while the pattern
recognition subsystem could be turned on or off for a given subject as needed. False positives
(i.e., unintended initiations of swing) were considered to be more disruptive and potentially
unsafe during walking than false negatives (i.e. failure to initiate an intended swing) because of
the risk of falling during unanticipated unloading of one of the limbs during double support. The
iEMG-based two-stage classifier triggered the stimulation pattern of the contralateral step when
both the threshold-based and pattern recognition subsystems agreed on initiating swing. The two-
stage classifier was computationally more demanding and so worsened the failure rate of the
ECU hardware.
141
Discussion
This work presented a method for selecting command sources for iEMG-triggered FES-
assisted ambulation, and evaluated two kinds of classifiers based on either the pattern recognition
or simple thresholding during online operation under laboratory conditions. Two muscles and the
location for implantation of intramuscular EMG electrode in those two muscles determined prior
to implantation, and two kinds of iEMG-based classifiers were implemented in the external
control unit (ECU) that works with the current family of ISTs.
The two classifiers were made computationally simple to work with the limited resources
available with the portable ECU based on Motorola HC12 16-bit family of microcontrollers.
Only 10 ms of integrated iEMG data sampled every 100 ms from each iEMG channel was
available to the classifiers. In order to limit the computational load on the ECU that was causing
overrun errors when the ECU was not able to finish all the computations in the allocated 50ms
real-time cycle time, the pattern recognition classifier used only one feature template to identify
the iEMG pattern as compared to three that were presented in our prior work in Chapter 4 [6.2].
The iEMG-triggered FES-assisted ambulation with both the classifiers was evaluated in the
The interference with the iEMG due to electrical stimulation of the same muscle was
observed with right erector spinae of the subject iSCI-1 (Figure 6.13). The electrical stimulation
of the same muscle inhibited the EMG activity in the right erector spinae muscle. This has
previously been observed by other researchers as well [6.18]. Therefore the electrical stimulation
of the same muscle during EMG recording was avoided in this study.
The performance of the threshold-based classifier that was based on the iEMG from the
left side with more extensive weaknesses of the subject iSCI-2 varied from trial-to-trial. This
variation was due to the sensitivity of the threshold-based classifier to spasms and other natural
142
variation of EMG activity like fatigue. In the final implementation of the classifier as a two-stage
system with a threshold-based and a pattern recognition subsystems, the threshold of the
threshold-based subsystem (thrTC) was made adaptive by updating it during operation (online
training) with a binary-search using the iEMG levels related to false positives and false
negatives, thus having it learn to accommodate for fatigue or other day-to-day time varying
factors. In view of the online training that was needed during real-time operation of the
classifiers, the extensiveness of the offline training to find the threshold (thrTC) and other
parameters could be reduced significantly. The two-stage classifier (Figure 6.9) was
computationally more demanding and so worsened the failure rate of the ECU hardware.
The processing power of the ECU limited the performance since the amount of
computations required for pattern recognition algorithm that was developed in prior work
presented in Chapter 4 could not be implemented [6.2]. The pattern recognition algorithm
presented in this study was based on pattern matching of the iEMG activity with just one feature
as compared to three in the prior work. Moreover, a sampling frequency of only 10 Hz was
possible with IST as compared to 2400 Hz in an xPC TargetTM (The Mathworks, Inc., USA) that
was presented in the prior work which reduced the SNR of the iEMG signal. Most of the time in
the 50ms real-time cycle was taken up by the hardware events. The classifier algorithm had
limited time to process the iEMG data which caused frequent overrun errors, especially with the
two-stage classifier. A faster processing unit will decrease the processing time for a more
complex classifier algorithm like the one presented in Chapter 4 and should help in eliminating
overrun errors.
A way to provide the processing power needed by the controller algorithm is to implement
the controller on a more powerful but compact PC/104 (size 3.6 by 3.8 inches) single board
143
computer (SBC) running xPC targetTM(The MathWorks, Inc., USA) [6.31]. The xPC targetTM
SBC can serve as the main processor for implementing the classifier and other software control
algorithms while ECU can handle the RF power and command link with the IST for stimulus
control, stimulus current regulation, iEMG data control, stimulus blanking control, and system
control. The compact PC/104 SBC can be made to fit inside the enclosure of ECU and the
PC/104 architecture supports self stacking of different modules like data acquisition board that
can provide extra analog and digital input/output channels for interfacing with more sensors.
The quality of the iEMG signal can be significantly affected by the location of the
intramuscular electrode in the partially paralyzed muscle. The optimal set of command sources
and motor targets was determined only from the superficial muscles accessible from the skin’s
surface. It was challenging to find the optimal location for implanting the intramuscular electrode
in a partially paralyzed muscle based on EMG recordings from the surface of the skin. Deeper
implanted for acute testing. Advanced source localization techniques can be applied to multi-
electrode surface EMG recordings to determine an optimal command sources. It was not possible
to record from the muscle during implantation of the iEMG electrode since the subject was under
general anesthesia so the best location had to be found before the surgery.
Additional physical sensors in parallel to the iEMG may be able to enhance the
performance of the classifiers. Sensors like linear accelerometers and gyroscopes can be placed
in the ECU that is worn externally close to the center of mass (COM) of the body. Gyroscopes
may be able to capture the orientation and the accelerometers the acceleration of body segments.
The tilt and acceleration of the upper body during a rolling walker aided ambulation will have
significant information related to gait cycle. If the sensors can be placed in the shoe then it has
144
been shown that foot linear acceleration and foot angular velocity can be used to correctly
Conclusions
This study presented a selection criterion to identify the command sources for an iEMG-
based classifier to trigger FES-assisted gait. The selection is based on surface EMG recordings
from multiple muscles that are associated with ambulation to identify two best muscles to trigger
FES-assisted left and right step. The location for implanting the intramuscular electrode in those
two muscles were also found using the same selection criteria that was then applied on multi-
based on iEMG for triggering FES-assisted steps was demonstrated during real-time operation in
one subject. The pattern recognition classifier generally outperformed the threshold-based
classifier, particularly with respect to minimizing false triggers. Subject showed no preference
for either the threshold- or pattern-recognition based classifier as determined by the Usability
the iEMG from the side with more extensive weakness. In the final implementation of the
provision was made to retrain the threshold of the threshold-based subsystem with binary search
The pattern recognition classifier algorithm was simplified significantly to run with the
limited processing power available in ECU. This significantly deteriorated the performance when
compared to the results presented in Chapter 4. The reliability of the final implementation of the
145
classifier as a two-stage adaptive system was compromised because of hardware instability due
More research needs to be done in evaluating the optimality of the implantation site for the
intramuscular EMG electrode found from non-invasive surface EMG recordings from the muscle
belly. The location of the implantation site is critical in order to record good iEMG from a
partially paralyzed muscle. Additional sensors should also be explored in conjunction to two
References
6.2. A. Dutta, R. Kobetic, and R. Triolo, “Ambulation after incomplete spinal cord injury with
EMG-triggered Functional Electrical Stimulation,” IEEE Transactions on Biomedical
Engineering, 55:2, February 2008.
6.3. B. Smith, Z. Tang, M.W. Johnson, S. Pourmehdi, M.M. Gazdik, J.R. Buckett, and P.H.
Peckham , “An externally powered, multichannel, implantable stimulator-telemeter for
control of paralyzed muscle,” IEEE Trans Biomed Eng., vol. 45, no. 4, 1998, pp. 463-475.
6.4. N. Bhadra, K.L. Kilgore, and P.H. Peckham, “Implanted stimulators for restoration of
function in spinal cord injury,” Med. Eng. Phys., vol. 23, 2001, pp. 19-28.
6.5. I. P. Pappas, M. R. Popovic, T. Keller, V. Dietz, and M. Morari, “A reliable gait phase
detection system,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 9, no. 2, Jun. 2001, pp.
113-125.
6.6. A. Mansfield, and G. M. Lyons, “The use of accelerometry to detect heel contact events for
use as a sensor in FES assisted walking,” Med. Eng. Phys., vol. 25, no. 10, Dec. 2003, pp.
879-885.
6.8. R. T. Lauer, R. T. Smith, and R. R. Betz, “Application of a neuro-fuzzy network for gait
event detection using electromyography in the child with cerebral palsy,” IEEE Trans.
Rehabil. Eng., vol. 52, no. 9, Sep. 2005, pp. 1532–1540.
146
6.9. D. Graupe, K. H. Kohn, S. Basseas, and E. Naccarato, “Patient controlled electrical
stimulation via EMG signature discrimination for providing certain paraplegics with
primitive walking functions,” J. Biomed. Eng., vol. 5, 1983, pp. 220-226.
6.11. D. Graupe, “EMG pattern analysis for patient-responsive control of FES in paraplegics for
walker-supported walking,” IEEE Trans. Biomed. Eng., vol. 36, no. 7, July 1989, pp. 711-
719.
6.12. D. Graupe, and H. Kordylewski, “Artificial neural network control of FES in paraplegics
for patient responsive ambulation,” IEEE Trans. Biomed Eng., vol. 42, no. 7, Jul. 1995, pp.
699-707.
6.14. A. Dutta, R. Kobetic, and R. J. Triolo, “EMG based triggering and modulation of
stimulation patterns for FES assisted ambulation – a conceptual study,” presented at XXth
Congress of the International Society of Biomechanics, Cleveland, OH, Aug. 2005.
6.15. A. Dutta, and R. J. Triolo, “Volitional surface EMG based control of FES-assisted gait after
incomplete spinal cord injury – a single case feasibility study,” presented at NIH Neural
Interfaces Workshop, Bethesda, MD, Sep. 2005.
6.16. R. Thorsen, R. Spadone, and M. Ferrarin, “A pilot study of myoelectrically controlled FES
of upper extremity,” IEEE Trans. Rehabil. Eng., vol. 9, no. 2, June 2001, pp. 161–167.
6.17. R.Futami, K.Seki, T.Kawanishi, T.Sugiyama, I.Cikajlo, and Y.Handa, “Application of local
EMG-driven FES to incompletely paralyzed lower extremities,” presented at 10th Annual
Conference of the International FES Society, Montreal, Canada, July 2005.
6.18. P. N. Taylor, and P. H. Chappell, “Variation in system gain when using voluntary EMG to
control electrical stimulation of the same muscle,” presented at 9th Annual Conference of
the International FES Society, Bournemouth, U.K., 2004.
6.20. R. Kobetic, and E.B. Marsolais, “Synthesis of paraplegic gait with multichannel functional
neuromuscular stimulation,” IEEE Trans Rehab Eng., vol. 2, no. 2, 1994, pp. 66-79.
147
6.21. R. Kobetic, R. J. Triolo, and E. B. Marsolais, “Muscle selection and walking performance
of multichannel FES systems for ambulation in paraplegia,” IEEE Trans. Rehabil. Eng.,
vol. 5, no. 1, Mar. 1997, pp. 23–29.
6.24. T. D. Dickens, Elementary Signal Detection Theory. US: Oxford University Press, 2001,
pp. 66, 121.
6.25. A. P. Bradley, "The use of the area under the ROC curve in the evaluation of machine
learning algorithms," Pattern Recognition, vol. 30, no. 7, 1997, pp. 1145-1159.
6.26. I. T. Jolliffe, Principal Component Analysis, New York: Springer, 1986, pp. 44-48, 269-
296.
6.27. D. A. Winter, Biomechanics and Motor Control of Human Gait: Normal, Elderly and
Pathological. Waterloo Biomechanics, 1991, ch.1.
6.31. http://www.pc104.org/pc104_specs.php
148
Figures
Figure 6.1: Experimental setup for data collection during FES-assisted walking with the block-
diagram for the FES-system (ECU: external control unit, LE: linear envelope).
149
a
Figure 6.2: Processing of the sampled surface EMG a) rectified and reconstructed sEMG signal
b) linear envelope found from processed sEMG signal.
150
Figure 6.3: Experimental protocol for the collection of EMG data during over-ground walking in
the laboratory.
151
T9 level
L2 level
crease line
Figure 6.4: Multi-electrode matrix for simultaneous collection of the surface EMG from multiple
locations on the muscle belly.
152
a b
c c d
Figure 6.5: The steps during the implantation of intramuscular EMG electrode a) insertion of
probe, b) deployment of peelable sheath over probe, c) insertion of the iEMG
electrode through the peelable sheath, d) peeling off of the polymer sheath leaving the
iEMG electrode in place.
153
Figure 6.6: Pulse-width map of the stimulation patterns used for walking shown as an example.
154
Figure 6.7: The real-time cycle in IST with 50 ms time period for stimulation frequency of 20
Hz.
155
Figure 6.8: Parameters for the iEMG classifier computed from the training data that was
collected with the switch-triggered FES system.
156
Figure 6.9: The flow chart of the iEMG-based two-stage classifier for triggering FES for
walking.
157
Figure 6.10: Usability Rating Scale to find the user perspective on ease/difficulty of using the
classifier [6.29].
158
a
Figure 6.11: Best location found from the surface EMG for implanting intramuscular EMG
electrodes a) Left gastrocnemius and right erector spinae b) Left and right
gastrocnemius.
159
a) Left MG
b) Right ES
Figure 6.12: a) Discriminability Index (DI) of left medial gastrocnemius (MG) for the swing
phase (SW) and double support phase (DS) during over-ground walking for the
subject iSCI-1 at each data point of the gait cycle b) Discriminability Index (DI) of
right erector spinae (ES) for the swing phase (SW) and double support phase (DS)
during over-ground walking for the subject iSCI-1 at each data point of the gait cycle.
[shaded portion is the classification region]
160
Figure 6.13: Inhibition of iEMG from right erector spinae during right swing phase (SW) as
shown in the top panel due to electrical stimulation of the same muscle when
compared to that in absence of electrical stimulation shown in the bottom panel of the
subject iSCI-1.
161
a) Right MG
b) Left MG
Figure 6.14b: a) Discriminability Index (DI) of right medial gastrocnemius (MG) for the swing
phase (SW) and double support phase (DS) during over-ground walking for the
subject iSCI-2 at each data point of the gait cycle b) Discriminability Index (DI) of
left medial gastrocnemius (MG) for the swing phase (SW) and double support phase
(DS) during over-ground walking for the subject iSCI-2 at each data point of the gait
cycle. [shaded portion is the classification region]
162
Tables
FP & FN
0.384 0.128 0.384 0.333
rates
Table 6.2: The performance of the pattern recognition classifier for iSCI-2.
163
Intel PIII 600MHz 97
Intel PII 400MHz 135
AMD K6-2 400MHz 159
Intel PI 166MHz 567
Intel PI 90MHz 1400
Intel 486DX 40MHz 3800
Table 6.3: The minimum achievable sample time in µs for the classifier algorithm in different
hardware that can serve as single board computer.
164