26th Iranian Conference on Electrical Engineering (ICEE2018)

Neurodegenerative Disease Classification Using

Nonlinear Gait Signal Analysis, Genetic Algorithm
And Ensemble Classifier
Bashir Najafabadian Abolfazl Sheibani
Department of Biomedical Engineering, Science and Department of Biomedical Engineering, Science and
Research Branch, Islamic Azad University, Tehran, Iran Research Branch, Islamic Azad University, Tehran, Iran
Email: [email protected] Email: [email protected]

Hamidreza Jalali
Department of Biomedical Engineering, Science and Keivan Maghooli*
Research Branch, Islamic Azad University, Tehran, Iran Department of Biomedical Engineering, Science and
Email: [email protected] Research Branch, Islamic Azad University, Tehran, Iran
Email: [email protected]

Abstract— Neurodegenerative diseases refer to the dysfunction of affecting the basal ganglia that lead to problems with
muscular and nervous system control. Gait analysis is an effective movement timing and rhythm. The data from gait analysis are
tool for identification and classification of the severity and the widely used for kinematic studies in order to investigate motor
types of these diseases. In this paper, a novel method for diseases and disorders. Gait signal analysis is used for better
neurodegenerative disease classification using nonlinear gait recognition of mechanisms of motor disorders. It is a high-
signal analysis is presented. For this purpose, the physionet potential non-invasive method based on walking dynamics for
database, which contains four sets of records for Huntington, diagnosis of various types of neurodegenerative disorders
Amyotrophic lateral sclerosis, Parkinson's and the control group, (NDD). Gait dynamics is controlled by mixed nerves. It
was used. The signal is decomposed by independent components
appears that the feedback loops that control the physiological
analysis after pre-processing using wavelet transform and noise
reduction. A set of chaotic and fractal features were extracted.
systems should receive a set of spatial and temporal scales to
Optimal features were selected using the genetic algorithm. The adapt to any environmental changes. Therefore, walk/stop/turn
set of optimal features were considered as the input of AdaBoost time series affect the spatial and temporal scales.
structure. The results of the simulation showed an accuracy of Extensive studies have been conducted in recent years with
92.34±2.1% in classification of neurodegenerative diseases. the help of computer tools to measure the parameters of
walking distance in healthy adult human models and
Keywords- Neurodegenerative; ICA; Genetic algorithm;
characteristics of the gait signal in neurodegenerative diseases.
AdaBoost; fractal
Scafetta et al. [1] used a central pattern generator model to
I. INTRODUCTION simulate the dynamics of human walking. He also discussed
random and fractal characteristics of walking signals in three
Neurodegenerative diseases refer to the dysfunction of categories of the disease. Liao et al. [2] proposed a method for
muscular and nervous system control. Since lower extremity measuring the signal of gait asymmetry to diagnose NDD using
flexion and extension movements are controlled by the central the multi-resolution entropy analysis of stance time
nervous system, the gait is always abnormal in a patients with fluctuations. The results of this study showed that gait
neurodegenerative disease. This is caused by the loss of motor symmetry signal in subjects with NDD, especially with ALS, is
neurons. These diseases include Parkinson's disease, significantly disturbed.
amyotrophic lateral sclerosis, and Huntington's disease. Since
neurons of the basal ganglia are involved in neuromuscular Daliri [3] presented a method for diagnosing
control (e.g. balance and sequential movements), neurodegenerative diseases based on gait analysis. He used the
neurodegenerative diseases are expected to disturb gait cycles data from a time series of stride intervals, swing intervals,
and dynamics. stance intervals and double support intervals of stride-to-stride
measures of footfall contact times. The support vector
Amyotrophic lateral sclerosis (ALS) is a primary machines using different kernels were examined for the
neurodegenerative disease characterized by progressive diagnosis.
muscular paralysis reflecting degeneration of motor neurons in
the primary motor cortex, brainstem, and spinal cord. Aziz and Arif [4] converted stride time series to a symbolic
Parkinson’s and Huntington's diseases are common disorders sequence and proposed a threshold based on the symbolic

978-1-5386-4916-9/18/$31.00 ©2018 IEEE

1482
26th Iranian Conference on Electrical Engineering (ICEE2018)

entropy method for analysis of the complexity of gait signal. processing features, space reduction methods, feature selection
They showed that normalized corrected Shannon entropy and classification structures are discussed in the section of
(NCSE) of walking decreases in various values of small techniques. The accuracy, sensitivity and specificity parameters
threshold of Shannon entropy in ALS cases. Wu and Shi [5] of used methods are calculated and compared in the results
proposed an analytical statistic to classify healthy and ALS section. In conclusion, an optimal and superior technique is
cases. In the former study, the probability density functions introduced and several recommendations are given for its use
(PDFs) were estimated using the nonparametric Parzen- in future studies.
window approach. The Kullback–Leibler divergence (MKLD)
was then extracted. This method can effectively separate the II. TECHNIQUE AND METHODS
stride patterns between the groups of healthy controls and ALS A. DATASET DESCRIPTIONS
patients with an overall accuracy rate of 82.2%.
This paper used Physionet database [9, 12], which included
In a similar study [6], it was assumed that the number of gait signal recording in four groups: ALS, Huntington,
swing intervals in patients with ALS may be different from Parkinson's, and control (healthy people). This database has 13
healthy controls. Patterns of walking in ALS patients were records from ALS patients, 20 from Huntington's patients, 15
classified using this parameter and linear and nonlinear from Parkinson's patients, and 20 from the control group
structures. The results of this study showed that this parameter (healthy state). The data were measured using force-sensitive
along with the mean intervals of strides can be used to resistance sensors placed under the foot that measured the force
differentiate walking patterns in healthy and ALS cases with an according to the steps of the right and left legs. Each record
accuracy of 89.66%. contains two sample columns, each of which is dedicated to the
Wavelet-Based Multi-Scale Entropy and non-static analysis sensor output embedded below the right and left foot. The
based on statistical methods and frequency-time analysis were recorder was a 12-bit converter with a sampling frequency of
proposed in various papers to investigate the relationship 300 Hz, with a record length of 5 minutes and a total of 90,000
between walking signal and NDD. Sarbaz et al. [7] used samples.
spectral analysis to classify the patients with Huntington’s B. PREPROCESSING
disease and normal cases. Sang-hong et al. [8] presented a A discrete wavelet transform (DWT) was used for
method to classify the patients with Parkinson's disease using preprocessing raw data. Wavelet transform is a time-domain
wavelet transform based feature extraction of gait frequency converter that besides maintaining frequency
characteristics. Hausdroff et al. [9] compared the stride components it is capable of retaining the time components.
intervals in walking signal between healthy individuals and the Wavelet transform utilizes a sliding window scaled according
patients with Huntington’s disease. They calculated the to the signal changes. In other words, while preserving
correlation between stride intervals in comparison to previous temporal resolution, it provides a proper resolution in
and next intervals of walking. The results of this study showed frequency by applying comparative scales in the mother
that walking pattern in Huntington's cases is more random than wavelet.
healthy individuals. Zheng [10] categorized healthy individuals
from patients with Parkinson's disease using a statistical Wavelet transform utilizes a filter bank structure (FBS) and
method. analyzes the signal based on FBS and the number of user-
specified levels. At each level of decomposition, signal sample
Sugavaneswaran et al. [11] offered a method for classifying is reduced, and the sparse output is considered as high-pass and
healthy individuals from ALS cases based on time-frequency low-pass filter input. High-pass filter output is called details,
analysis. Baratian et al. presented a diagnostic support method which contains components of a higher-frequency signal. The
for classifying healthy individuals and ALS cases based on an low-pass filter output is called generalities containing
unbiased cross-validation strategy. components of a lower-frequency signal. The output of each
This paper presents a six-step algorithm for classifying filtering level based on filter impulse response in a discrete
neurodegenerative diseases. In the first step, a set of gait state is defined as follows:
signals is collected from the physionet database [9, 12]. In the
‫ݕ‬௛௜௚௛ ሾ݇ሿ=σ௡ ‫ݔ‬ሾ݊ሿǤ ݄ሾʹ݇ െ ݊ሿ (1)
second step, a wavelet-based waveform pre-processing
algorithm and sub-band analysis are used. In the third step, ‫ݕ‬௟௢௪ ሾ݇ሿ=σ௡ ‫ݔ‬ሾ݊ሿǤ ݃ሾʹ݇ െ ݊ሿ (2)
independent component analysis is performed to extract the
optimal band signal. In the fourth step, chaotic and fractal
nonlinear features of each sub-band are extracted. In the fifth Where h [n] is the high-pass filter (HPF) and g [n] is the
step genetic algorithms are used to reduce space and select the low pass filter (LPF), y high is the output of HPF y low is the
best features. In the sixth step, structures of linear and output of LPF and x [n] is the input signal. The output of HPF
nonlinear classification are used to diagnose and distinguish is again redirected to the sample decimation block and high-
healthy individuals from patients with neurodegenerative pass and low-pass filter to form FBS based on the number of
diseases. user-defined levels. This paper used wavelet-transform to
The structure of the presentation of these concepts in this remove the signal noise by matching mother wavelet with
article is as follows. Data collection methods and used daughter signal. For this purpose, the default coefficients of
databases, the methods used to extract preprocessing and

1483
26th Iranian Conference on Electrical Engineering (ICEE2018)

MATLAB software were used for FBS using Daubechies4 ݀‫்ݔ‬ ݀‫ି்ݔ݀ ்ݔ‬ଵ ݀‫ݔ‬ଵ (7)
mother wavelet. ൌ ‫ڮ‬
݀‫ݔ‬଴ ݀‫ି்ݔ‬ଵ ݀‫ି்ݔ‬ଶ ݀‫ݔ‬଴
C. PROCESSING
‫ܯ‬ᇱ ሺ‫ି்ݔ‬ଵ ሻ‫ܯ‬ᇱ ሺ‫ି்ݔ‬ଶ ሻ ‫ܯ ڮ‬ᇱ ሺ‫ݔ‬଴ ሻ (8)
Independent component analysis was used to process noise-
reduced data. In this method, it is assumed that a signal is a
linear combination of it sources and statistically independent Thus, we will have:
்ିଵ (9)
(ICA). Thus, this method can be used to decompose the input ͳ
signal and extraction of independent sources whose numbers ݄ ൌ Ž‹ ෍ ݈݊ȁ‫ܯ‬ᇱ ሺ‫ݔ‬௡ ሻȁ
்՜ஶ ܶ
are equal to the signal dimension. If we assume that a signal is ௡ୀ଴
a linear combination of its sources, we can write:
Given the natural boundary discussion for a set, h can be also
ܵ ൌ ܹܺ (3) shown as follows:
(10)
Where x is a multi-dimensional signal and s is the ݄ ൌ න ݈݊ȁ‫ܯ‬ᇱ ሺ‫ݔ‬௡ ሻȁ݀ߤሺ‫ݔ‬ሻ
independent sources.
In most applications, calculating only the largest Lyapunov
Kurtosis is used to extract the best independent component, coefficient is enough. Lyapunov exponent of an n-dimensional
which is defined as follows. system is as Ȝ1•Ȝ2•…• Ȝn, where Ȝ1 is the largest Lyapunov
‫ݐݎݑܭ‬ሺ‫ݏ‬ሻ ൌ ‫ܧ‬ሼ‫ ݏ‬ସ ሽ െ ͵ሺ‫ܧ‬ሼ‫ ݏ‬ଶ ሽሻଶ (4) coefficient and considered as a feature in this study.

Where E {} is Operator of expected value and s is the 2. Detrended fluctuation analysis (DFA)
source component. DFA measures the intra-signal correlation, which is not only
If the kurtosis approaches zero, data distribution approaches resistant to non-stationary, but can also systematically
Gaussian, otherwise it distances from normal distribution. eliminate various timescale deviations caused by noise from
Therefore, in order to extract the optimal independent time series. The main idea of DFA to calculate the root mean
component in the gate signal, selection criterion is the largest square is the time series fluctuations. This correlation for the
kurtosis value between the components. This paper used eight different time scales is extracted as follows:
independent components for each sensor of the left and right ୩
feet and for each sensor. Eight component with the largest ሺሻ ൌ ෍ ഥሻ ǡ  ൌ ͳǡ ǥ ǡ 
ሺ୨ െ  (11)
kurtosis value was selected. It appeared that the use of ICA is ୨ୀଵ
effective in signal improvement and noise reduction.
ഥ
 represents the mean of time samples. The summed up
D. FEATURE EXTRACTION series is then divided into parts of equal length n. In each
Non-linear dynamic features based on chaos concepts were section, the data are estimated with the least square error line
used to model the nonlinear behavior of the gait signal. The that displays trend in that window. Yn (k) shows these
signal analysis seems to reflect the valuable information of the regression lines. The summed up series Y (k) is then detrended
signal in the form of the feature in the case where its stable in each window. Thus, the square root of the mean square
behavior is important. However, the use of non-linear features oscillation is summed up and Yn (k) is calculated. This
paves the ground for making feature extraction process calculation is repeated on all time scales (window sizes) to
resistant to noise (it compensates for the defect of time domain obtain the relation between F(n) and the size of window n. F(n)
features). We dealt with the features extracted in this paper as increases with window size.
outlined below.
ͳ ୒
1. Lyapunov exponent ሺሻ ൌ ඨ ෍ ሺ›ሺሻ െ ›୬ ሺሻሻଶ (12)
 ୏ୀଵ
One index of chaotic systems to show the sensitivity of the
system response to small stimuli was evaluated through this The oscillation in small windows is related to oscillations
quantity. For a one-dimensional map, this quantity measures that can be expressed by auto-similarity factor, and Į, which is
the divergence rate of the response to the initial conditions called scaling exponent, is the line slope of log f (n) relative to
close to each other. This means that if we have two initial log (n). In general, Į can be considered as a sign of the
unevenness of the original time series.
conditions close to each other, such as‫ݔ‬଴ ǡ ‫ݔ‬଴ ൅ ݀‫ݔ‬଴ , and apply
map M to them after n consecutive times, we will have: 3. Time reversibility (TR)
݀‫ݔ‬௡ ൎ ݁ ௛௡ ݀‫ݔ‬଴ (5) A time series is called TR only if its possible characteristics
are constant with TR. A test was suggested for zero assumption
h is Lyapunov exponent for this map, so one can write: that a time series is reversible. The rejection of the zero
ͳ ݀‫்ݔ‬ (6) hypothesis shows that the time series cannot be described by a
݄ ൌ Ž‹ ݈݊ ฬ ฬ random Gaussian linear process. Thus, TR can be used as a
்՜ஶ ܶ ݀‫ݔ‬଴
strong reason for non-linearity. In this study we used a simple
method that calculates TR for signal s:

1484
26th Iranian Conference on Electrical Engineering (ICEE2018)

ே optimization algorithm that can be used for optimal feature

ͳ selection.
ܶ‫ݎ‬ሺ߬ሻ ൌ ෍ ሺܵ௡ െ ܵ௡ିఛ ሻଷ (13)
ܰെ߬ Each gene from a chromosome assumes zero and one,
௡ୀఛାଵ
which shows the value one in existence of this feature in the
optimal feature space and zero in its absence.
Where N is the length of the signal and this study used IJ = 1.
Thus, the position of a feature in the selected space is coded
with zero (non-existence) and one (existence) as binary.
Step 1: Entering feature matrix
4. Fractal Dimension (FD) Step 2: Generation of the initial population: to generate a
This method divides the waveform into N0 = N-1 separate population, several binary chromosomes whose number of
sections, where N is the length of the signal and then calculates genes is equal to the number of features are randomly created.
the fractal with respect to the Kutsz equation below: Step 3: The fitness function is determined by increasing the
accuracy in AdaBoost Algorithm.
Ž‘‰ଵ଴ ሺ ᇱ ሻ Step 4: Given the fitness level of each chromosome, a new
ൌ ୢ (14) population is created
Ž‘‰ଵ଴ ቀ ቁ ൅ Ž‘‰ଵ଴ ሺ ᇱ ሻ
୐
A. crossover: Two parent chromosomes are crossover by
XOR function and a new population is created.
Where d is the flatness of the curve and L is the curve B) Mutation: A random number is selected at the length of
length, both of which are defined as: the chromosome (feature) and the value of that gene is
† ൌ ƒšሺ†‹•–ƒ…‡ሺͳǡ ‹ሻሻ mutated.
(15) Step 5: Selecting optimal population according to fitness
function

ேᇱ
Step 6: Checking the termination terms
 ൌ ෍ ݀݅‫ݐݏ‬ሺ݅ǡ ݅ ൅ ͳሻ (16) A. Not satisfied: return to step 4
௜ୀଵ B. Satisfied: The announcement of the optimal
chromosome
And Dist (i, j) is the distance between the points i and j from End
the curve. Several authors used this algorithm to analyze F. CLASSIFICATION AND DIAGNNOSIS
biological signals. FDs show the geometric properties of the
absorbent bed, but have a high computational speed.
AdaBoost was used to classify the extracted features. The
input of classification structure is extracted, considered once
5. Petrosian fractal dimension (PFD) after the application of reducing feature space, and once after
PFD method is a very simple and fast way to calculate the the reduction of feature space as the input of the classification
dimension. This dimension is calculated as follows: structure. The output of the classification structure is the label
of disease class.
Ž‘‰ଵ଴ ሺܰሻ
ൌ ே (17)
Ž‘‰ଵ଴ ሺܰሻ ൅ Ž‘‰ଵ଴ ሺ ሻ
ேା଴Ǥସேο III. RESULT
Where N is the length of the time series and Nǻ is the In the first step of simulation, raw data is collected from the
number of changes in the signal in the derivative of the signal. physionet database for neurodegenerative diseases. Each data
Since the signal is discrete, the sequential difference of the time consists of two columns, one for the right and the other for the
series elements is calculated. left foot. Data was stored in four separate folders, and each
record was about five minutes, and was used for this article.
In comparing FDs, it should be noted that the correlation
method requires a large number of samples to calculate higher In the second step, 1D wavelet transform toolbox was used
dimensions over a time series. This method is not suitable for to reduce the noise of raw signals. To accomplish this, the
signals that do not have a stationary condition over time. ddencmp MATLAB function was used to achieve this goal.
This command eliminates the signal noise with the default
E. FEATURE SPACE REDUCTION coefficient while maintaining the approximation coefficients
Genetic Algorithm has been used to improve the accuracy and reduce the noise. In step 3, the independent components
of the classification and selection of optimal features. This is a- were extracted from the clean signal that was extracted after

1485
26th Iranian Conference on Electrical Engineering (ICEE2018)

the wavelet reconstruction. For this purpose, the Fast ICA was
used to compute the independent components. The number of IV. CONCLUSION
independent components was considered 8, and the critter of
kurtosis was used to select the independent components for the In this paper, a novel method for diagnosis of
recorded data. The component with the highest value of neurodegenerative diseases based on the extraction of chaotic
kurtosis was selected as an independent component. Thus, for and fractal features is presented. The results of the simulation
each record two independent components were calculated one showed that the fractal features provide a basis for classifying
for the right foot and the other for the left foot different disease types. Also, the results showed that the
application of the genetic algorithm in the selection of features
In the fifth step, a set of features that were addressed in the
technique and methods section, was extracted from optimal could improve the classification indexes. However, this
component. These features are calculated for each registry for technique is still a challenge in terms of time to solve the
two independent components. In the sixth step, the optimal optimization problems.
features of the extracted features were selected using the
genetic algorithm.
The results of simulation showed that DFA, FD PFD are
the best features that are selected by genetic algorithm. Thus REFERENCES
the feature space is reduced from 10 to 6 features. The
attributes before and after feature selection are considered as [1] N. Scafetta, D. Marchi, and B. J. West, “Understanding the
inputs of the AdaBoost classification structure. Cross complexity of human gait dynamics,” Chaos An Interdiscip. J. Nonlinear Sci.,
vol. 19, no. 2, p. 26108, 2009.
validation 5 is used to test the classification structure. The
[2] F. Liao, J. Wang, and P. He, “Multi-resolution entropy analysis of
results of the simulations are presented in Table 1. These gait symmetry in neurological degenerative diseases and amyotrophic lateral
results showed that the use of the genetic algorithm in the sclerosis,” Med. Eng. Phys., vol. 30, no. 3, pp. 299–310, 2008.
selection of optimal features improved the classification [3] M. R. Daliri, “Automatic diagnosis of neuro-degenerative diseases
indexes. The Figure 1. Shows the graphic Compression of the using gait dynamics,” Measurement, vol. 45, no. 7, pp. 1729–1734, 2012.
accuracy of the proposed method with another studies. The [4] W. Aziz and M. Arif, “Complexity analysis of stride interval time
result shows that the use of genetic algorithm and adaboost series by threshold dependent symbolic entropy,” Eur. J. Appl. Physiol., vol.
classifier has improved the accuracy of neurodegenerative 98, no. 1, pp. 30–40, 2006.
disease Classification. [5] Y. Wu and L. Shi, “Analysis of altered gait cycle duration in
amyotrophic lateral sclerosis based on nonparametric probability density
function estimation,” Med. Eng. Phys., vol. 33, no. 3, pp. 347–355, 2011.
TABLE I. CLASSIFICATION EFFICIENCY EVALUATION
[6] Y. Wu and S. Krishnan, “Computer-aided analysis of gait rhythm
Features type sensitivity accuracy specificity fluctuations in amyotrophic lateral sclerosis,” Med. Biol. Eng. Comput., vol.
All features 86.54±3.7% 87.27±3.2% 88.38±3.3%
47, no. 11, p. 1165, 2009.
Selected features 93.34±1.7% 92.34±2.1% 91.34±1.1%
[7] Y. Sarbaz, M. Banaie, M. Pooyan, S. Gharibzadeh, F.
Towhidkhah, and A. Jafari, “Modeling the gait of normal and Parkinsonian
persons for improving the diagnosis,” Neurosci. Lett., vol. 509, no. 2, pp. 72–
75, 2012.
[8] S.-H. Lee and J. S. Lim, “Parkinson’s disease classification using
gait characteristics and wavelet-based feature extraction,” Expert Syst. Appl.,
92.34%

vol. 39, no. 8, pp. 7338–7344, 2012.

89.66%

[9] J. M. Hausdorff, S. L. Mitchell, R. Firtion, C.-K. Peng, M. E.

Cudkowicz, J. Y. Wei, and A. L. Goldberger, “Altered fractal dynamics of
gait: reduced stride-interval correlations with aging and Huntington’s
disease,” J. Appl. Physiol., vol. 82, no. 1, pp. 262–269, 1997.
[10] H. Zheng, M. Yang, H. Wang, and S. McClean, “Machine learning
and statistical approaches to support the discrimination of neuro-degenerative
diseases based on gait analysis,” in Intelligent patient management, Springer,
82.20%

2009, pp. 57–70.

[11] L. Sugavaneswaran, K. Umapathy, and S. Krishnan, “Ambiguity
domain-based identification of altered gait pattern in ALS disorder,” J. Neural
Eng., vol. 9, no. 4, p. 46004, 2012.
[12] J. M. Hausdorff, A. Lertratanakul, M. E. Cudkowicz, A. L.
Peterson, D. Kaliton, and A. L. Goldberger, “Dynamic markers of altered gait
rhythm in amyotrophic lateral sclerosis,” J. Appl. Physiol., vol. 88, no. 6, pp.
2045–2053,2000.

[5] [6] PROPOSED

Figure 1. compression of acuracy of the proposed method with another studies.

1486