Parkinson Disease (PD) is the degenerative illness of the brain that affects
patients ability to perform a well-learned motor task such as walking,
body balancing as well as writing appropriately (Morris
and Lansek, 1996). Basically, there are two types of PD known as motor and
non-motor. Four primary symptoms categorized in motor PD includes tremor or
also known as hand shaking; a characteristic appearance of pill-rolling
movement between thumb and forefinger. Next is bradykinesia due to slowness
of movements and loss of involuntary movement. Thirdly is rigidity or limbs
stiffness and finally is postural instability that caused patients to fall easily
(Bock and Nausieda, 1999). On the other hand, the symptoms
of non-motor PD are neuropsychiatric disturbance that is a common symptom which
includes depression, anxiety and apathy, sleep disturbance, autonomic disturbance
such as constipation, dizziness, drooling of saliva, sexual dysfunction and
bladder problem. Apart of it, the gastrointestinal tract problem also occur
in non-motor PD patients that includes nausea, choking and sensation of incomplete
voiding of bowels, difficulty swallowing, severe constipation and inability
to taste as well as sensory symptom. Numerous researches have been reported
related to motor PD. For instance, Morris et al.
(1999) has studied the constraints on the kinetic, kinematic and spatiotemporal
parameter of gait in a 71 year old woman with PD. Results showed that PD has
greatly influence the subjects gait by reducing the speed, stride length,
total range of movement during walking and power generation at push off in ankle
power profile. Moreover, statistical test conducted by Sofuwa
et al. (2005) and Roiz et al. (2010)
also proven that PD did affect patients gait pattern in terms of basic
spatiotemporal, kinematic and kinetic parameter. Due to these findings by various
researches that stated motor PD caused noticeable deviation of patients
gait, therefore this study will deem further the application of Artificial Neural
Network (ANN) and Support Vector Machine (SVM) as classifiers to recognize gait
pattern of PD patients.
Conversely, the use of ANN in human gait analysis is widely explored. Firstly,
Gioftsos and Grieve (1995) had successfully recognized
98% of the walking speed in a given condition and 99% for the vice versa case.
Lafunte et al. (1997) used ANN to discriminate
between healthy and lower limb arthritis patients recorded 80% accuracy. Further,
Begg and Kamruzzaman (2006) applied ANN for detecting
and classifying walking pattern changes due to ageing obtained maximum of 83.3%.
Holzreiter and Kohle (1993) implemented ANN in order
to distinguish the gait pattern between normal and pathological
with accuracy approximately 95%. Inspired by Holzreiter
and Kohle (1993) and Barton and Lee (1997) used
the approach of kinematic analysis to differentiate the gait pattern for three
conditions, specifically normal gait, simulations of leg length difference along
with simulation of leg weight difference. From this study, 83.33% of unknown
gait pattern were successfully assigned into the right category and proven that
ANN is capable to be applied in clinical practice for automated diagnosis of
On the other hand, SVM is another popular classifier too. For instance, Singla
et al. (2011) concluded the superior performance of SVM over ANN
classifer in detecting eye event from electroencephalographic (EEG) signal.
Results obtained showed that SVM successfully classified 90.8% correct EEG pattern
whilst ANN only gained 86.8% accuracy. Also, Wu and Liu
(2011) implemented SVM and NN in finger-vein pattern identification and
SVM showed good performance in recognizing finger-vein features. Another application
of SVM in pattern classification was used by Roy and Bhattacharya
(2005) for iris recognition and the performance of SVM as a classifier was
proven far better than ANN. Thus, by referring to the performance of these two
types of machine classifiers as reported by the previous researches, the classification
ability of these classifiers will be evaluated to determine the most suitable
classifier for PD gait pattern recognition.
MATERIALS AND METHODS
Gait data acquisition and gait features: Twelve PD patients and twenty
healthy volunteers without any other diseases that could contribute to the additional
effect on their body balance such as post-stroke or any additional muscular
disease participated in this study. It is also confirmed that all subjects must
also be able to walk freely without any cane or mechanical aid device during
walking to eliminate the influence of aid device on the data acquired. Next,
thirty seven reflective markers are adhered to the subjects skin before
the walking trials are preceded. This marker will be traced by infrared camera
for kinematic analysis purpose. In addition, subjects are instructed to walk
freely at their comfortable speed on two embedded force plate with sampling
rate of 200 Hz. By walking on the force plate, Ground Reaction Forces (GRF)
of the subjects are recorded. The three categories of GRF are lateral GRF (Fx),
horizontal GRF (Fy) and vertical GRF (Fz). In this study,
two types of significant GRF are chosen as the kinetic features specifically
the horizontal and vertical GRF. The extraction of these two GRF based on the
five significant walking events are as shown in Fig. 1.
As the subjects walk on the force plates, six infrared cameras are activated
to record the walking event. These infrared cameras have the ability to trace
the coordinate of reflective markers that is adhered to each subjects
||The five kinetic features extracted from GRF reading
|| Selected feature to be used as classifier inputs
From the recording of walking event, all the relevant basic spatiotemporal
gait features namely walking time, speed, cadence and step length can be calculated.
In addition, the angle of significant walking event also can be obtained from
the coordinate of the reflective markers. All the input features from three
gait parameter specifically basic, kinematic and kinetic with description are
as listed in Table 1. Moreover, all subjects participated
are instructed to complete three gait trials and the basic, kinetic and kinematic
parameters are calculated based on the mean of the three trials. As a precaution,
if the observer found that any subject purposely extended or shortened their
normal stride in order to ensure contact with the force plates as required or
their foot overlapped the edge of the other foot, the trial is eliminated and
need to be repeated.
Normalization method: The use of data normalization often provided great improvement in pattern recognition. In this paper, there are three normalization methods implemented prior to classification which are:
GRF data normalization: Normalization of GRF data is obtained with GRF data acquired as dividend and body weight as divisor. This is to eliminate the influence of subjects body weight to GRF reading which can be differ for each person regardless their health status:
Intra group data normalization: Here, all input data recorded during gait recording experiments are mathematically normalized by dividing all data acquired with the maximum value of the respective group of both normal and PD patients:
Inter group data normalization: As for inter group normalization, input data are divided by the highest value amongst both normal and PD group:
Artificial Neural Network (ANN): The architecture of the ANN used in
this study is the multilayer feed-forward network with Back Propagation (BP)
algorithm. BP algorithm is a generalization of the least mean squared algorithm
that minimizes the mean squared error between the desired and actual output
of the network by modifying its network weight (Mehrotra et
al., 1997). As BP algorithm used supervised learning, the network is
trained using the known data of inputs and desired output. After the data is
trained, the network weights are frozen to be used later for new input samples
to compute an output values (Mehrotra et al., 1997).
The proposed ANN configuration in this study consisted of an input layer with
number of input neurons corresponding to the gait feature vectors that acted
as inputs to the ANN. Additionally, one hidden layer with eleven neurons and
one output layer with a single neuron indicate normal and abnormal gait pattern.
Weight are adjusted by employing Lavenberg-Marquardt as the training algorithm
and the learning rate are set at 0.3. Lavenberg-Marquardt is chosen as the training
algorithm due to fast training and its ability to minimize the error function
(Abraham, 2005). Then, tanh-sigmoidal activation function
(tansig) is applied for both hidden and output layer.
Support Vector Machine (SVM): SVM which has been introduced by Vapnik
is an excellent learning technique that is developed based on the framework
of structural risk minimization (SRM) and the VC bounds theory cited by Kumar
and Gopal (2011). In a binary classification task, SVM will map the input
space to a high-dimensional feature space based on the selected kernel function.
Then, an Optimal Separating Hyperplane (OSH) that separates the two classes
by maximizing the margin between the closest point of the classes are constructed,
as illustrated in Fig. 2. In this study, three type of kernel
function are evaluated during classification namely linear, radial basis function
||Example of SVM binary classification with OSH and maximum
||Three SVM kernel functions used in classification task and
its mathematical formula
The mathematical formula for the entire kernel function are as given in Table
2 (Kumar and Gopal, 2011).
Cross-validation and classification method: Cross-validation is a statistical
method of dividing data into two segments, which are training and testing set.
Since the PD subjects in this study is relatively small, cross-validation method
are employed to overcome the constraint, inspired from technique used by Barton
and Lee (1997). Data set obtained from both category that generated total
of 32 subjects are divided into 4 subsets. Initially, one subset will be used
as training set, while the remainder are assigned as unseen or testing data.
The process repeated until all subsets are utilized as testing data.
Conversely, input data from each gait category, namely basic spatiotemporal, kinematic and kinetic are classified solely and the classifier performance are recorded. Then, gait features from two parameters are fused or combined to evaluate the effectiveness of these features in classifying gait of PD. This classification task are repeated for both ANN and SVM classifier based on intra group and inter group normalization method too. By performing these two classification mode, potential of each gait parameter can be evaluated in identifying the most significant gait features that contributed for recognition of PD gait.
RESULT AND DISCUSSION
The overall performances of both classifiers based on the two technique of data normalization are as tabulated in Table 3 and 4. Additionally, three statistical indices are computed specifically accuracy, sensitivity and specificity using the computed results from True Acceptance Rate (TAR), False Rejection Rate (FRR), True Rejection Rate (TRR) and False Acceptance Rate (FAR).
The detail of TAR, FRR, TRR and FAR are as below:
||TAR: Classifiers identify PD as PD
||FRR: Classifiers identify PD as normal
||TRR: Classifiers identify normal as normal
||FAR: Classifiers identify normal as PD
||Sensitivity: Classifiers ability to identify PD as PD
||Specificity: Classifiers ability to identify normal as normal
As observed in Table 3, based on intra group normalization technique, both ANN as well as SVM attained perfect accuracy rate based on four features that is extracted from the basic spatiotemporal category namely stride time, cadence, step length and walking speed. Moreover, basic spatiotemporal data also recorded perfect sensitivity and specificity rate that indicated the classifiers ability to identify PD subjects accurately. On the other hand, it is realized that kinematic parameter that comprised of twelve feature vectors extracted from hip angle, knee angle as well as ankle angle during heel strike, toe off, along with maximum extension and maximum flexion contributed worst recognition rate based on both ANN and SVM as classifiers. It is also observed that the sensitiveness of both classifiers deteriorated to correctly detect PD gait features. In addition, results attained also verified that kinematic features either solely as inputs or based on data fusion are unsuitable as classification inputs due to insignificant or less deviation of joint angle parameters during walking between the PD subjects and normal group. Apart from basic spatiotemporal and kinematic features, kinetic gait parameter is also vital in gait recognition. ANN classifier obtained classification rate of 90.6% for kinetic parameter as feature inputs as in Table 3. On the whole, recognition accuracy as well as sensitivity rate with kinetic parameters as feature inputs is higher as compared to kinematic parameters for both ANN and SVM classifiers. Hence, it is proven that basic spatiotemporal solely acted as the most suitable gait features for classification of PD gait. From the findings of this study, it is also suggested for future researches to only consider basic and kinetic gait parameter as input features for recognition of PD gait.
Nevertheless, it is observed too that intra group normalization technique contributed
higher accuracy rate. For instance, SVM with polynomial kernel contributed as
the most significant kernel amongst the three kernels as well as ANN as classifier
with 95.8% accuracy rate for kinetic parameters based on intra group normalization
and the accuracy rate reduced for inter group normalization category. As in
Table 4, the classification rates of both classifiers worsened
and similar achievement of recognition rate is observed too irrespective of
the feature inputs. Therefore, this finding confirmed that inter group normalization
technique is unsuitable for PD gait pattern recognition conducted in this research.
|| Results based on intra group normalization data
|| Results based on inter group normalization data
SVM with RBF as kernel performed well in distinguishing gait pattern based on data fusion of two gait parameters as showed in Table 3 and 4 with fusion of basic spatiotemporal and kinetic gait parameter achieved highest accuracy that is 98.2%. The increased in accuracy rate for SVM upon fusion of input features indicated the incapability of SVM as classifier based on small number of input features. This confirmed the lower accuracy of SVM using basic, kinematic or kinetic as inputs solely. Overall, for data fusion, the classifiers performances are enhanced based on fusion of three parameters as compared to data fusion of two parameters. Finally, SVM with polynomial kernel attained highest specificity that confirmed the classifer ability not to generate false detection (normal healthy gait pattern) rather than the ability to accurately recognize the pathological gait pattern.
In conclusion, SVM as classifier performed better as compared to ANN specifically for data fusion of gait parameters. However, both classifiers attained higher classification rate based on intra- group normalization method instead of inter group normalization irrespective of data fusion or otherwise. Finally, initial results proven that basic spatiotemporal contributed as the best feature inputs based on perfect accuracy, specificity and sensitivity rate attained and the feature vectors can be utilized to classify pathological gait of PD.
This study was supported by Ministry of Science, Technology and Innovative (MOSTI) Malaysia under the E-Science Grant No: 06-01-01-SF0314.