Artificial Organs
23(5):443–446, Blackwell Science, Inc.
© 1999 International Society for Artificial Organs
Adaptive Restriction Rules Provide Functional and Safe
Stimulation Pattern for Foot Drop Correction
*Aleksandar Kostov, †Morten Hansen, †Morten Haugland, and †Thomas Sinkjær
*Faculty of Rehabilitation Medicine, University of Alberta and Glenrose Rehabilitation Hospital, Edmonton, Alberta,
Canada; and †Center for Sensory-Motor Interaction, Aalborg University, Aalborg, Denmark
Abstract: We report on our advances in sensory feedback
data processing and control system design for functional
electrical stimulation (FES) assisted correction of foot
drop. We have applied 2 methods of signal purification on
the bin integrated electroneurogram (i.e., optimized low
pass filtering and wavelet denoising) before training adap-
tive logic networks (ALN). ALN generated stimulation
control pulses, which correspond to the swing phase of the
impaired leg when dorsal flexion of the foot is necessary to
provide safe ground clearance. However, the obtained
control signal contained sporadic stimulation spikes in the
Functional electrical stimulation (FES) can assist
hemiplegic patients by stimulating the peroneal
nerve in the swing phase of the affected leg and
thereby providing dorsiflexion of the foot. Detection
of the heel strike and the impaired leg’s foot lift are
prerequisites for successful FES use in foot drop cor-
rection. Systems based on surface stimulation and
controlled by a heel switch located in the shoes have
been most successful even if the performance has not
been optimal. Major obstacles in better acceptance
of these systems include the following. First is the
complicated daily application and removal of the
FES system. Second is the need to position elec-
trodes precisely to produce correct movement of the
foot. Third is that the surface stimulation can be un-
comfortable at required stimulation levels and may
produce skin irritation. Fourth is that the sensor de-
tecting heel strike which controls the stimulator is
Received December 1998.
Presented in part at the 6th Vienna International Workshop on
Functional Electrostimulation, held September 22–24, 1998, in Vi-
enna, Austria.
Address correspondence and reprint requests to Dr. Aleksan-
dar Kostov, Faculty of Rehabilitation Medicine, 3-48 Corbett
Hall, University of Alberta, Edmonton, AB T6G 2G4 Canada.
E-mail: Aleks.Kostov@ualberta.ca
443
stance phase, which can collapse the subject, and infre-
quent broken stimulation pulses in the swing phase, which
can result in unpredictable consequences. In this study, we
have introduced adaptive restriction rules (ARR), which
are initially used as previously reported and then dynami-
cally adapted during the use of the system. Our results
suggest that ARR provide a safer and more reliable stimu-
lation pattern than fixed restriction rules. Key Words:
Foot drop—Functional electrical stimulation—Control
strategies—Neural recordings—Sensory feedback—
Adaptive logic network.
cosmetically undesirable and is exposed to adverse
conditions, causing frequent failures. Fully im-
planted stimulators can resolve the first 2 problems;
however, once the stimulator is implanted it would
be desirable to have sensory feedback generated
from implanted or natural sensory sources. It has
been suggested that afferent neural signals, such as
those recorded from the sural or calcaneal nerves,
which carry mainly cutaneous information, can be
used as sensory feedback signals in fully implanted
FES systems (1,3). Heel strike, which is important
for determination of the stimulus duration, is easy to
detect in the rectified and bin integrated electroneu-
rogram (RBI ENG). At the same time, detection of
foot lift, which is important for the start of the stimu-
lation, is more difficult to detect because the slower
change of skin stretch reduces the firing rate of cu-
taneous receptors and second because of the larger
influence of the unwanted nerve signal resulting
from the activation of cutaneous receptors by socks,
footwear, and bending in the metatarsal joints. In
our previous work, we demonstrated the potential of
using adaptive logic networks (ALN) and restriction
rules (RR) for gait event detection and FES con-
trol (2). In this study, we introduced dynamic
444
adaptation of the restriction rules, which provides a
wider range of operating conditions for safe use of
the FES system.
MATERIALS AND METHODS
The experimental setup for this study is schemati-
cally illustrated in Fig. 1. In addition to a traditional
FES system for foot drop correction (solid lines), we
recorded and processed a sensory nerve signal elec-
troneurogram (ENG) from the sural nerve in the
affected leg (dashed lines). The experimental part of
the study was performed at the Center for Sensory-
Motor Interaction, Aalborg University, Denmark
while data processing took place in the Laboratory
for Advanced Assistive Technology, University of
Alberta, Canada. The data acquisition has been
thoroughly described (1). The implant and recording
experiments are therefore only briefly described
here. A 35-year-old male subject with spastic
hemiplegia was implanted with a 30 mm long tripolar
whole nerve cuff on the sural nerve approximately 7
cm proximal and 3 cm posterior to the lateral mal-
leolus. The 3 wires were led subcutaneously to an
exit point approximately 25 cm proximal to the lat-
eral malleolus.
The neural signal was amplified with a total gain
of approximately 100 dB, band pass filtered between
800 Hz and 3.6 kHz and rectified and bin integrated
(RBI) into 10 ms bins (top trace in Fig. 2). The data
used in this study were recorded during FES assisted
walking with surface stimulation of the deep pero-
neal nerve, which caused dorsiflexion of the ankle
during the swing phase. For ALN training purposes,
we manually created a target FES control signal
FIG. 1. The experimental setup is for the development of a con-
trol system for correction of foot drop problem based on sensory
neural feedback.
Artif Organs, Vol. 23, No. 5, 1999
A. KOSTOV ET AL.
FIG. 2. Shown is an example of RBI ENG recorded from the
sural nerve and a manually created target signal used for ALN
training. Transitions between the stance and swing phases are
referred to as heel strike (A) and foot lift (B).
(bottom trace in Fig. 2). The heel strike information
was clearly indicated in the ENG signal by a burst of
neural activity. Foot lift was indicated by a smaller
burst followed by reduced activity in the ENG. A
heel switch was used to confirm heel strike timing;
however, it was not very useful in determining foot
lift. Two representative data sets were selected for
this study, recorded on Day 38 after implantation
with the subject walking barefoot on the floor in the
lab area. The length of each data set was 60 s,
equivalent to approximately 32 gait cycles.
To create a control algorithm for FES assisted
walking, we used ALN, a machine learning method
for supervised learning based on a multilayer per-
ceptron. In all initial ALN evaluation experiments,
we calculated the performance as a percentile of cor-
rectly predicted samples in the ALN output signal
(4). Implementation of ALN involves splitting the
data set into training and test data sets, and the ALN
performance was measured on both data sets, the
first one representing the quality of the training, and
the second one representing its generalization. Due
to the binary character of the output signal, obtain-
ing an error close to or larger than 50% was consid-
ered a random ALN performance.
The optimized feature extraction and ALN train-
ing parameters were as follows. First, to increase the
signal-to-noise ratio of the RBI ENG, 2 methods
were compared, low pass filtering and wavelet de-
noising. Although wavelet denoising has the advan-
tage of selectively removing low energy signal com-
ponents regardless of their frequency, low pass
filtering, which has lower computational demands in
real-time implementation, was selected for signal
preprocessing. In the rest of this study, a first order
Butterworth filter with a 6.5 Hz cut off frequency
was used. Second, expanding the ALN training do-
main from using only original ENG signal samples to
original plus 2 previous samples (5th and 10th) re-
 ADAPTIVE RULES FOR SAFE STIMULATION
sulted in the best ALN training performance. Third,
a minimum of 10 gait cycles was necessary for suc-
cessful training; we used 15. Fourth, the ALN train-
ing parameters were 5 layers, 0.15 learning rate, and
15 epochs.
A new measure of evaluating the ALN output was
introduced. In addition to calculating the actual per-
centile error, which does not reveal information
about the type of error, the output signals were as-
sessed by calculating the frequency distribution of
the transition duration (TD). This distribution was
calculated by counting the number of samples be-
tween each transition. This approach was selected
due to the periodic nature of the gait. Comparing the
TD distribution of the ALN output to the distribu-
tion of the target signal revealed information of its
functionality (Fig. 3).
For the elimination of functional errors (2,5), dif-
ferent methods using restriction rules were investi-
gated. In the first method, we eliminated transitions
shorter than a specified duration (blanking) while in
the second method, we applied a moving average
filter followed by a threshold (smoothing). Both
methods eliminated short duration transitions; how-
ever, they introduced a small delay in the output
signal. This delay was considered to be negligible for
foot drop correction. Each method was tested indi-
vidually as well as combined with the other one.
The restriction rules used in this study were a
FIG. 3. Shown are the distribution plots of the swing and stance phase transition durations for the target control signal (A), ALN output
(B), and ALN output processed by adaptive restriction rules (C).
445
modified version of the rules described previously
(5). The following equation illustrates the concept.
共Pst − 2 · ␴st兲 ⬍ Pst ⬍ 共Pst + 2 · ␴st兲 (1)
Pbs ⬎ Pbs − 2 · ␴bs
where Pst is the stimulation period (swing) and Pbs is
the period between stimulations (stance).
We developed adaptive restriction rules (ARR)
by calculating the mean and standard deviations for
the stance and swing phases and using them to elimi-
nate control signal transitions considered to be out
of range for the particular user. The modifications
consisted of continuous adaptation of the rules to
current gait as the ALN were evaluated on new data.
A buffer containing transition duration information
on stance and swing phases during the previous 15
gait cycles was used to store and to calculate new
mean and standard deviation values for the restric-
tion rules. The buffer was updated following each
transition in the output from the restriction rules.
RESULTS AND DISCUSSION
Figure 3 illustrates the TD distributions of the tar-
get FES control signal, ALN output, and the ALN
output processed by the adaptive restriction rules.
Most of the swing phases were detected correctly by
ALN while most of the errors occur in the stance
phase (Fig. 3B). Almost none of the complete stance
Artif Organs, Vol. 23, No. 5, 1999
446 A. KOSTOV ET AL.
FIG. 4. The example is of the target FES control signal, its pre-
diction by ALN, and the ALN output processed by adaptive re-
striction rules.
phases were detected. For both the stance and swing
phases, a large number of short duration transitions
(i.e., 1–3 samples) were present (not shown). This
indicated that additional processing of the ALN out-
put was necessary before the signal could be used to
control a stimulator. The blanking method was the
most successful in eliminating short transitions as
well as in restoring complete stance phases prior to
applying statistical restriction rules.
Applying adaptive restriction rules removed short
duration transitions from both the stance and swing
phases and altered the TD distribution to resemble
the distribution of the target signal (Fig. 3C). Dashed
lines illustrate initial acceptance margins calculated
from the training set using Eq. 1. If the restriction
rules were not gradually adapted to the test data,
almost half of the swing phase pulses (i.e., stimula-
tion pulses) would be shortened to the upper limit of
the acceptance range. Some of the transition dura-
Artif Organs, Vol. 23, No. 5, 1999
tions were clipped during the adaptation period,
which resulted in their higher accumulation at the
initial upper limit for the swing phase.
Results from applying ALN and adaptive restric-
tion rules are illustrated by an example of an off-line
test in Fig. 4. All of the swing and stance phases in
the restricted ALN output were complete and of
similar duration to the original, which demonstrates
the potential of this approach for the safe operation
of the FES control system. As a consequence of us-
ing blanking for short transition elimination prior to
applying the restriction rules, all transitions are de-
layed. At a sampling rate of 100 Hz, a 20–30 ms delay
is considered to be negligible for the overall perfor-
mance of the foot drop correction system.
Acknowledgments: Funding for this research was pro-
vided by the Danish Research Council, the Danish Na-
tional Research Foundation, and the Natural Sciences and
Engineering Research Council of Canada (NSERC).
REFERENCES
1. Haugland, M, Sinkjær T. Cutaneous whole nerve recordings
used for correction of footdrop in hemiplegic man. IEEE
Trans Rehab Eng 1995;3:307–17.
2. Kostov A, Sinkjær T, Upshaw B. Gait event discrimination
using ALNs for control of FES in foot drop problem. Pro-
ceedings of the 18th IEEE-EMBS Conference. (CD-ROM).
Nijmegen, The Netherlands: Soetekow Videoproductions,
1996:1042–3.
3. Sinkjær T, Haugland M, Haase J. The use of natural sensory
nerve signals as an advanced heel-switch in drop-foot patients.
Proceedings of the 4th Vienna International Workshop on
Functional Electrostimulation. Vienna: Department of Bio-
medical Engineering and Physics, University of Vienna, 1992:
134–7.
4. Kostov A, Andrews BJ, Popovic DB, Stein RB, Armstrong
WW. Machine learning in control of functional electrical
stimulation systems for locomotion. IEEE Trans Biomed Eng
1995;42(6):541–51.
5. Kostov A. Functional error assessment in gait event discrimi-
nation for FES-assisted locomotion. Proceedings of the 18th
IEEE-EMBS Conference. (CD-ROM). Nijmegen, The Neth-
erlands: Soetekow Videoproductions, 1996;969–70.