Abstract
In this paper we proposed a method of stance
phase detection for portable sensor system. The sensor system is a part of the
feedback system in forthcoming FES gait re-education orthosis. It consisted of
two goniometers, two accelerometers and a gyroscope mounted on an aluminum
plate. Three healthy subjects and an incomplete spinal cord injured patient participated
in the experiment. They were asked to walk with normal walking speed, then faster
and afterwards slower. The Artificial Neural
Networks were used to recognize stance phase from the sensor signals. We
concluded that the sensor system together with signal processing composed of the
Artificial Neural Networks could
detect the beginning of the stance phase with high probability.
1. Introduction
Functional electrical stimulation
training of paralyzed muscles is effective in the great majority of incomplete
spinal cord injured (SCI) patients in the early period of the rehabilitation process
[1]. The aim of an FES rehabilitative system for re-education of walking is not
only to deliver electrical stimulation to the paralyzed muscles, but also to
assess the sensor information from the paralyzed limb [2], [3]. The sensory
information is fed back to the patient and not to the stimulator control unit. The
rehabilitation systems are intended to be used in incomplete SCI patients soon
after the accident or onset of disease. These systems are to be used within
rehabilitation centers and applied by therapists.
The aim of this study is to develop the FES
rehabilitative system for re-education of walking. In order to deliver the sensory
information about walking to the patients in the gait re-education system, the
necessary sensor data during the stance phase should be determined. We have developed
a stance phase detection algorithm that is based on the Artificial Neural Network (ANN) for the proposed sensor system [3]. In
this paper we examined the ability of the ANN
to detect the moment of initial stance and initial swing using the
sensor signals as inputs [4], [5].
2. Methods
2.1 Sensor system
We developed a sensor system device,
equipped with two accelerometers (ACCESS), two goniometers (Penny & Gilles)
and a gyroscope (Murata)[3]. The sensor system was attached to the lower leg
(Fig.1). The gyroscope measured one-degree
angular velocity of leg movement during the swing phase. The accelerometers measured
tangential and radial component of the leg acceleration. The goniometers were
attached to the ankle and the knee joint. As a reference measurement system the
optical contactless measuring system OPTOTRAK (Northern Digital) and force
plate (Advanced Mechanical Technology) were used. Personal computer using a
data acquisition board and Simulink software (MathWorks), the Matlab toolbox
software, were employed.
Fig.1
Portable sensor system was attached to the shank. Optical markers and force
plate were also used.
Fig.2 Structure of the artificial neural network used in
detection of initial stance and initial swing.
2.2 Structure
of artificial neural network
The moments of initial
stance (beginning of stance phase) and initial swing (end of stance phase) were
detected with ANNs. The structure of the ANNs was three layers feed-forward
style (input: 24, hidden: 6, output: 1) as shown in Fig.2. All the sensory signals
except data of the ankle joint angle were used as an input to the ANN. The
information of the ankle joint angle was omitted from the input of the ANN
because the recognition rate of the ANN decreased by using the information of
the ankle joint to the input.
The ANN output signal was computed at the
given time using the past 6 input samples and provided a normalized numerical
value (0 to 1.0).
During learning of the ANN a teacher signal ON (numerical
value was 0.99) was given at the moment of initial stance or swing, otherwise the
teacher signal was set to OFF (numerical value was 0.01). The learning of the
ANN was carried out by means of the Back propagation algorithm for each subject.
Two preliminary measurements were selected as learning data.
2.3 Experimental
procedure
Three healthy subjects (24-27 years old, male) and an incomplete SCI patient
(27 years old, male) participated in the
experiment. Their tasks were normal, faster
and slower walking in the testing environments. The patient
had incomplete C6-7 spinal cord injury. He had to use crutches bilaterally
during walking. He could hardly vary the walking speed. The recording tracks
were 7 to 18 seconds long and our subjects were able to do approximately 5
steps.
3. Results
3.1 Measurement data
Measurement data are shown in Fig.3, together
with ANN detection results of the moments of the initial stance and swing. The
represented data were measured at normal speed walking of the healthy subject.
The angular velocity measured by the gyroscope resulted in a large positive
wave, which described the swing phase of walking. The tangential acceleration
(At) and the radial acceleration (Ar) had a tendency to vary more intensively
during the swing phase than during the stance phase. In addition to this
tendency, the Ar varied intensively up and down at initial foot contact.
The walking experiments in incomplete SCI
patient were carried out and analyzed in a similar
way. According to the results obtained by all sensory signals, there were oscillations
in the early period of the stance phase when the foot wasn’t firmly contacted
the floor. Both joint angle trajectories of the incomplete SCI patient proved
much different than those of the healthy subject. The differences between the patient
and the healthy subject were observed by comparing the outputs of all sensors.
3.2 Detection results
of the moments of the initial stance and
the initial swing
When the
moments of the initial stance and the initial swing were detected, the ANN
outputted a numerical value. Table 1 shows the detection results of the initial stance and initial swing excluding data used
in the process of ANN learning. The successful detection was determined by output value greater than or equal 0.5. The
table shows the detected and expected number of initial stance or initial swing.
The notation of “initial stance: #” in column of “detection
of initial swing” is the number that the ANN wrongly detects the initial swing
at the initial stance. And the notation of “error: #” shows the number that the
ANN wrongly detects the initial stance or swing at other time.
From the results it is evident that proposed
ANN could recognize the initial stance with high probability (95% in average).
Although the ANN could recognize the initial swing with high probability (99%
in average) in healthy subjects, the recognition rate decreased in the incomplete
SCI patient. The ANN, which was thought to detect the initial swing, occasionally
failed at the initial stance. However, the miss-detection could be distinguished
from the other one, by making a comparison with the ANN output that detected
the initial stance. It was also possible to detect the stance phase during
various walking speeds, since two speeds learning data were used.
Table 1
Detection results of the moments of the initial stance and
the initial swing.
|
detection of initial stance |
detection of initial swing |
|
normal
A |
24/26 |
30/30 initial stance:3 |
|
normal
B |
35/37 error:2 |
34/35 initial stance:1 |
|
normal
C |
20/21 |
16/16 |
|
incomplete SCI patient |
16/16 |
9/13 initial stance:6 error:3 |
4. Discussion
The results detected in the incomplete SCI
patient walking were comparable with those of the healthy subjects. The moment
of the initial swing during walking of the patient could not be detected
satisfactory. On the other hand, the moment of the initial stance could be detected
with high probability. Therefore, the method of detecting stance phase by the
ANN can be implemented in the FES rehabilitative system.
The force plate was used to recognize the
event of the foot contact. However, the teacher signals used in the learning of
the ANN cannot be obtained for every step with the force plate. The output of the
gyroscope was therefore used as the teacher signal in the learning. However, we
have to perform new walking experiments to obtain more precise teacher signal,
because the present teacher signal is not showing the moments of the initial stance
and the initial swing satisfactory.
Acknowledgment
This study was partly supported by the
Ministry of Education, Culture, Sports, Science and Technology of Japan, Japan
society for the promotion of science under the Bilateral Program, and Takayanagi
Foundation for Electronics Science and Technology.
References
[1]
T.Bajd, A.Kralj, M.Štefančič
and N.Lavrač, “Use of Functional
Electrical Stimulation in the Lower Extremities of Incomplete Spinal Cord
Injured Patients,” Artificial Organs, Vol.23, No.5, pp. 403-409, 1999.
[2] T.Bajd,
I.Cikajlo, R.Šavrin, R.Erzin
and F.Gider, “FES Rehabilitative Systems for Re-Education of Walking in
Incomplete Spinal Cord Injured Persons,” Neuromodulation, Vol.3, No.3, pp.
167-174, 2000.
[3]
N.Furuse, I.Cikajlo and T.Bajd,
“Training of foot contact phase during FES assisted walking,” Proc. of the
International Federation for Medical & Biological Engineering, pp.686-689,
2001.
[4]
N.Furuse, T.Watanabe,
S.Ohba, R.Futami, N.Hoshimiya and Y.Handa, “Control- Command Detection for FES
using Residual Specific Movements,” Proc. of the 4th Annual
Conference of the International Functional Electrical Stimulation Society, pp.
319-322, 1999.
[5]
T.Watanabe, S.Yamagishi,
H.Murakami, N.Furuse, N.Hoshimiya and Y.Handa, “Recognition of Lower Limb
Movements by Artificial Neural Network for Restoring Gait of Hemiplegic
Patients by Functional Electrical Stimulation,” Proc. of the 23rd IEEE EMBS
Conference, 2001.