Vous êtes sur la page 1sur 4

Application of Feedforward Backpropagation Neural Network to Center of

Mass Estimation for Use in a Clinical Environment


A. L. Betker1 , T. Szturm2 , Z. Moussavi1
1

Department of Electrical Engineering, University of Manitoba, Manitoba, Canada


2
Department of Rehabilitation, University of Manitoba, Manitoba, Canada

Abstract In this paper, a feedforward backpropagation


neural network model is developed to estimate the resultant
center of mass (COM) trajectory in the sagittal plane. The
COM trajectory is one of the primary outputs of the human
postural control system, and is indicative of the systems
stability. However, currently available systems that calculate
the COM are not clinically available, making it difficult to
widely assess balance problems. The inputs to the neural
network model developed in this paper are obtained using
equipment that is inexpensive, easy to use and portable. The
results indicate that neural network models show promising
results for obtaining COM estimates that have clinical
applications.
Keywordsaccelerometer, center of mass, clinical, neural
network

I. INTRODUCTION
Balance of the human multi-segmental system is
maintained by an integrated network, controlled by the
central nervous system (CNS). The system consists of
visual, vestibular and somatosensory inputs [1], used to
maintain the equilibrium of the bodys center of mass
(COM) in relation to its base of support. When the COM is
projected outside the base of support and does not have the
required momentum, in the appropriate direction, to re-enter
the base of support area, instability occurs. Thus, the
trajectory of the COM provides us with a measure of
stability and is a primary output of the control system.
Although there exists variability [2] in which
equilibrium is maintained, postural strategies or patterns
emerge [1,3,4]. During quiet stance, an ankle strategy is
generally adopted [5]. As the frequency of oscillation
becomes larger, the velocity of the COM increases.
Therefore larger and quicker movements are required to
maintain the COM within the base of support. Thus a hip
strategy emerges, in which the body can be modeled as a
two-segment inverted pendulum consisting of the trunk and
limbs [1,3,4].
Models have been developed to predict state estimates
during quiet stance using optimal estimation control theory,
which attempts to optimize the systems state vectors based
on a cost function for a single-segment [6] and a threesegment model [7]. However, these systems can not be
practically applied to a clinical situation as the data
collection equipment is expensive and not portable, and the
models are fit to specific subjects.
Current movement trajectories are collected using
motion detection systems and force plates. The equipment is

0-7803-7789-3/03/$17.00 2003 IEEE

expensive, requires specialized training and is not portable,


and ergo not readily available to the clinical world. In
addition, most motion detection systems are not real-time
and the data requires time -consuming off-line processing.
Conversely, accelerometers are inexpensive, portable and
provide real-time data. Sophisticated multi-degree of
freedom moving platforms are available that can create
sinusoidal translations at different frequencies, and can
perform sway-referencing. This effect can also be achieved
by asking a subject to stand on a dense piece of foam, and
having the subjects induce voluntary sway in either
anterior-posterior (AP) or medial-lateral (ML) directions.
While neural networks have been widely applied to
motor control of robots and some studies have applied
neural networks to a specific task or aspect of motor control,
few studies have been applied to overall postural control
analysis in humans. In [8], sway length and an index to the
proximity of stability were simulated based on
environmental and task conditions. However, sway can
easily be measured through the use of accelerometers;
equipment that is clinically available.
This paper proposes the application of feedforward
backpropagation neural networks to estimate the COM
trajectory for a two-segment inverted pendulum, using
clinically available information.

II. M ETHODOLOGY
Two healthy subjects, a male aged 39 and a female aged
27, volunteered and gave informed consent. Ethics approval
was granted prior to recruiting subjects by The University of
Manitoba, Faculty of Medicine, Ethics Committee.
A. Experimental Setup
An AMTI force plate, model OR-6, was used to obtain
forces and moments in the sagittal, frontal and vertical
planes. This data was then calibrated and the center of
pressure (COP) in the frontal and sagittal planes were
calculated. Two ADXL202EB accelerometers (Analog
Devices) were affixed to the subject; one representing the
trunk segment placed on the neck, and the second
representing limb segment placed on the shank below the
knee joint. The sway in the AP and ML planes were
recorded by each accelerometer. The sway signals were
filtered using a 2nd order Butterworth bandpass filter, with
cutoff frequencies of 0.01 Hz and 100 Hz. The force plate

2714

EMBC 2003

and accelerometers data were recorded using a custom


LabView VI at a sampling rate of 500 Hz.
Kinematic data was obtained using a Sony Hi 8mm
video camera, model CCD-V101, connected to a Panasonic
SVHS video recorder model A7350. The reflective markers
were placed on the end points of each segment as described
in [4]. The data was sampled at 60 Hz, then filtered using a
4th order lowpass Butterworth filter, with a cutoff frequency
of 5 Hz. The data was digitized using the Peak 2D Video
Motion Analysis System (Peak Performance Technologies
Inc). The x and y (horizontal and vertical respectively)
coordinates for each reflective markers centroid was
determined in relation to an earth-fixed marker. For
calculation of the COM, we used a six-segment model
consisting of the head, arm, trunk-neck-pelvis, thigh, shank
and foot. The location of the COM x and y coordinates, and
resultant trajectory in the sagittal plane was calculated using
anthropometric data and methods as described in [9] for
each frame. The kinematic data was then interpolated to a
sampling rate of 500 Hz; i.e., to that of the kinetic and
accelerometer data.
B. Protocol
The subjects stood with their feet apart at their preferred
normal position; this position was recorded on the first trial
and kept constant for the remaining trials. The subjects kept
their arms crossed in front of their chest to avoid obstructing
the markers. In the three trials performed, the subjects were
instructed to induce voluntary sinusoidal movement of 0.5
Hz kept via a metronome. The duration of each trial was 10
seconds with a 2 minute rest period between each trial. In
the first trial, the subjects stood on the force platform, with
eyes open. In the second and third trials, the subject stood on
a dense piece of foam with a board placed on top to evenly
distribute their mass. This was performed for eyes open and
eyes closed conditions. The sponge introduces uncertainty
into the system which damps the signals that the cutaneous
sensors of the feet receive. In effect, the trials attempt
eliminate or distort visual cues, somatosensory cues or both.
C. Feedforward Backpropagation Neural Network Model
A feedforward backpropagation neural network consists
of two layers. The first layer, or hidden layer, has a tansigmoid (tan-sig) activation function, and the second layer,
or output layer, has a linear activation function. Thus, the
first layer limits the output to a narrow range, from which
the linear layer can produce all values. The output of each
layer can be represented by
YNx1 = f(W NxM XM,1 + b N,1 ),
(1)
where Y is a vector containing the output from each of the N
neurons in a given layer, W is a matrix containing the
weights for each of the M inputs for all N neurons, X is a
vector containing the inputs, b is a vector containing the

biases and f() is the activation function [10,11]. The


network was created using the neural network toolbox from
Matlab 6.0 release 13 (The MathWorks, Natick, Mass.,
USA).
The architecture of the neural network used to estimate
the resultant COM trajectory in the sagittal plane consists of
a single feedforward backpropagation neural network. The
inputs to the network are the resultant sway trajectory of the
trunk calculated from accelerometer data, the sway
trajectory of the shank in the AP direction provided by
accelerometer data, visual sensory input, 0 or 1 for eyes
closed and eyes open respectively, and somatosensory input,
0 or 1 for sponge and normal surface respectively.
In a backpropagation network, there are two steps
during training that are used alternately. The
backpropagation step calculates the error in the gradient
descent and propagates it backwards to each neuron in the
output layer, then hidden layer. In the second step, the
weights and biases are then recomputed, and the output from
the activated neurons is then propagated forward from the
hidden layer to the output layer.
The network is initialized with random weights and
biases, and was then trained using the Levinson-Marquardt
algorithm [11]. The weights and biases are updated
according to
Dn+1 = Dn [JT J + I] -1 JTe,
(2)
where Dn is a matrix containing the current weights and
biases, Dn+1 is a matrix containing the new weights and
biases, e is the network error, J is a Jacobean matrix
containing the 1st derivative of e with respect to the current
weights and biases, I is the identity matrix and is a
variable that increases or decreases based on the
performance function. The gradient of the error surface, g, is
equal to JTe [11].
The training data consisted of every third sample from
the first 6 seconds of data, for each of the three trials of each
subject; i.e. the network was trained for all conditions at
once. The length of the training data was 5000 points. The
network contained 50 neurons, and was trained until an
acceptable percentage error was achieved. The test data
consisted of the remaining 4000 samples from each trial for
each subject; 2000 samples from the first 6 seconds, and
2000 samples from the last 4 seconds.
The target vectors were obtained from the acquired
kinematic and kinetic data. The input and target vectors are
normalized by the mean plus standard deviation (SD) of the
training data. When simulating the network, the test vectors
are first normalized to lie within the range 0 to 1; i.e., the
range for which the tan-sig function is most sensitive [11].
Once simulated, the resultant vectors are then transformed to
their original means and SDs. The resulting estimations of
the resultant COM trajectory is then lowpass filtered using a
4th order Butterworth filter, with a cutoff frequency of 4 Hz.
The target vectors and estimated resultant COM
trajectory are then normalized to lie within the range of 0 to

2715

1, and the normalized percentage error (NPE) is calculated.


The mean of the normalized percent error is also reported.

III. RESULTS
The network was simulated with data from each of the
trials, for each subject. The results for each of the different
test conditions for subjects 1 and 2 are given in Fig. 1 and
Fig. 2 respectively. The first 2000 samples consist of the
remaining samples from the first 6 seconds of the trial that
were not used for training. Thus testing the networks
reliability. The remaining 2000 points consist of the data
from the last 4 seconds of the trial. Thus testing the
networks response to new data that is independent from the
training data in that there is no overlap between them as
with the first 2000 points.
The true COM trajectory is indicated with a solid line,
and the estimated COM trajectory is indicated with a dashed
line. The mean and maximum values for the NPE for each
subject, during each condition, are given in Table 1.
IV. DISCUSSION
Control models have been developed that estimate state
variables by incorporating the actual biomechanics of the
system [6,7]. Optimal estimation control theory is used,
which optimizes the state variables according to a cost
function. A benefit of this model is that sensory noise is
incorporated. A Kalman filter is used to provide an optimal
estimate of the state variables, given delayed sensor plus
noise signals and delayed control signals. As Kalman filters
operate on linear data, and the relationship between sensory
and state information is not always linear, optimal estimates
are then passed through a non-linear predictor.

Fig. 2. Normalized resultant COM trajectory; subject 2 results for a)


normal surface, eyes open, b) sponge, eyes open and c) sponge, eyes
closed.

The model in [6] was for a single-link inverted


pendulum during quiet stance, and had to be fit to each
subject. The model in [7] was not tested with subjects, but
rather a three-link model system, during quiet stance. As the
degrees of freedom are increased, so are the dynamics of the
system and complexity of the model. While these models
give an understanding to sensory integration, they can not be
practically applied in a clinical environment.
The model developed in this paper is not attempting to
model system mechanics. Instead, it attempts to determine a
relationship between sway information, during different
conditions, and the corresponding motion of the COM,
when two-segment inverted pendulum dynamics are
considered.
The subjects induced voluntary, periodic, sinusoidal
motion, which was reflected in the resulting trajectory of the
COM during the normal surface, eyes open trial; however,
shifts in the baseline of the signal occurred as errors
accumulated in the system. The estimated COM trajectory
closely followed the true COM trajectory, and was able to
identify these baseline shifts.
TABLE I
Mean and Maximum Normalized Percentage Error
Subject 1
Trial

Fig. 1. Normalized resultant COM trajectory; subject 1 results for a)


normal surface, eyes open, b) sponge, eyes open and c) sponge, eyes
closed.

2716

Subject 2

Mean
NPE (%)

Maximum
NPE (%)

Mean
NPE (%)

Maximum
NPE (%)

Eyes Open,
Normal Surface

7.3

23.6

5.4

15.0

Eyes Open,
Sponge

8.6

24.7

6.5

21.3

Eyes Closed,
Sponge

12.7

41.8

10.5

41.4

When the information provided by sensory cues is


altered, it introduces uncertainties in the system. In
particular, the foam causes damping of the cutaneous
information provided, which in turn produces large errors in
the postural systems predictions. In the trial where the
somatosensory information is altered, but visual information
is still available, there are small deviations from a sinusoid,
and peaks where the subject had difficulty maintaining
balance. As there is still predictive elements to the signal,
the estimated COM trajectory was able to follow the true
COM trajectory, however, not as closely as in the normal
surface, eyes open trial. When visual cues are also removed,
the postural system begins to accumulate large errors and
the sway increases. The true COM trajectory had larger
deviations from a sinusoid, again with peaks where balance
was difficult to maintain. There are few predictive ele ments
to the signal as large uncertainty is added to the signal via
the removal of visual and somatosensory information. Due
to this lack of sensory information, our own balance systems
cannot predict the disturbances. This is reflected in the fact
that the estimated COM trajectory had difficulty following
the true path in this case. Therefore, the model requires
additional information to be able to more accurately
estimate the COM trajectory.
In order to improve the performance of the network, and
begin to incorporate the biomechanics of the system,
additional clinically available information will need to be
incorporated into the model. Also, training and test data
from a larger subject group would aid in training the
network with additional balance patterns.
The linear envelope of the electromyogram (EMG)
signal provides information regarding preparatory or
reactive movements, onsets of corrective actions and the
selected strategy via which muscle groups become active
and the order in which they are activated. In addition, post
hoc testing reveals that the incorporation of COP
coordinates into the model would be beneficial.
The UltraThin OrthoTest Mat (Vista Medical Ltd, 3-55
Henlow Bay, Winnipeg, MB, CAN) is a new device that
contains piezo resistive sensors, used to calculate the
vertical COP. The OrthoTest Mat provides real-time data via
a small, light-weight interface unit, that connects to a
computer via a serial port. The dimensions of the OrthoTest
Mat are 53 cm x 53 cm x 0.036 cm, meaning it is easily
portable. As the OrthoTest Mat is also relatively
inexpensive, it provides a clinically available alternative
from which the COP data can be obtained.

information as traditional systems. It is believed that with


additional complexity, neural network models can be
developed to provide a clinician with the movement
trajectories required to assess balance.

V. CONCLUSION
The motivation behind this research is the development
of a clinically available system, that provides the same

2717

REFERENCES
[1] J. J. Buchanan and F. B. Horak, Emergence of Postural Patterns
as a function of vision and translation frequency Journal of
Neurophysiology, vol. 6, pp. 2325-2339, 1999.
[2] M. Schieppati, A. Giordano and A. Nardone, Variability in a
dynamic postural t ask attests ample flexibility in balance control
mechanisms Experimental Brain Research, vol. 144, pp. 200210, 2002.
[3] Y. G. Ko, J. H. Challis and K. M. Newell, Postural coordination
patterns as a function of dynamics of the support surface
Human Mo vement Science, vol. 20, no. 6, pp. 737-764, 2001.
[4] T. Szturm and B. Fallang, Effects of varying acceleration of
platform translation and toes-up rotations on the pattern and
magnitude of balance reactions in humans Journal of Vestibular
Research, vol. 8, no. 5, pp. 381-397, 1998.
[5] P. Gatev, S. Thomas, T. Kepple and M. Hallet, Feedforward
ankle strategy of balance during quiet stance in adults Journal
of Physiology, 514.3, pp. 915-928, 1999.
[6] H. van der Kooij, R. Jacobs, B. Koopman and H. Grootenboer,
A multisensory integration model of human stance control
Biological Cybernetics, vol. 80, pp. 299-308, 1999.
[7] T. Kiemel, K. S. Oie and J. J. Jeka, Multisensory fusion and the
stochastic structure of postural sway Biological Cybernetics,
vol. 87, pp. 262-277, 2002.
[8] W. Wang and A. Bhattacharya, A back-propagation neural
network model for prediction of loss of balance, IEEE
Proceedings (peer-reviewed) of the 15th Southern Biomedical
Engineering Conference, Pub. # 0-7803-3131-1/96, pp. 85-88,
1996.
[9] D. A. Winter, Biomechanics and motor control of human
movement. John Wiley & Sons Canada, LTD., 1990, pp. 5172.
[10] J. A. Freeman and D. M. Skapura, Neural networks: algorithms,
applications and programming techniques. Addison Wesley
Longman, 1991, pp. 89-105.
[11] Neural Networks Toolbox (4.0) Users Guide, The MathWorks,
Natick, MA, 2000, pp. 2-22-22, 5-22-56.