Biomedical Signal Processing and Control 9 (2014) 45–55

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

journal homepage: www.elsevier.com/locate/bspc

Synchronizing chaotification with support vector machine and wolf

pack search algorithm for estimation of peripheral vascular occlusion
in diabetes mellitus
Chien-Ming Li a,b , Yi-Chun Du c , Jian-Xing Wu a , Chia-Hung Lin d,∗ , Yueh-Ren Ho e ,
Ying-Jr Lin a , Tainsong Chen a,1
a
Department of Biomedical Engineering, National Cheng Kung University, Tainan City, Taiwan
b
Division of Infectious Diseases, Department of Medicine of Chi Mei Medical Center, Tainan City, Taiwan
c
Department of Electrical Engineering, Southern Taiwan University of Science and Technology, Tainan City, Taiwan
d
Department of Electrical Engineering, Kao-Yuan University, Kaohsiung City, Taiwan
e
Department of Biochemistry, Medical College of National Cheng Kung University, Tainan City, Taiwan

a r t i c l e i n f o a b s t r a c t

Article history: This study proposes a method for the estimation of peripheral vascular occlusion (PVO) in diabetic foot
Received 17 January 2013 using a support vector machine (SVM) classifier with the wolf pack search (WPS) algorithm. The long-
Received in revised form term presence of elevated blood sugar levels commonly results in peripheral neuropathy, peripheral
29 September 2013
vascular disease, nephropathy, and retinopathy in patients with Type 2 diabetes mellitus. Patients with
Accepted 5 October 2013
PVO disease have decreased walking capability and life quality in diabetes mellitus and poor peripheral
Available online 27 October 2013
circulation of PVO causes morbidity like infection and amputation of the legs or feet of diabetics. This
progressively vascular occlusion is often ignored by the patients and primary care physicians in early
Keywords:
Peripheral vascular occlusion (PVO) stage. Therefore, a reliable method of diagnostic assistance is crucial for early diagnosis and monitor-
Support vector machine (SVM) ing of PVO and prevention of amputation. Photoplethysmography (PPG) is a non-invasive technique for
Wolf pack search (WPS) detecting blood volume changes in peripheral vascular bed. Literature indicates that the pulse transit
Photoplethysmography (PPG) time increases and waveform shape changes increase in PPG of the vascular occlusion. PPG pulses of
Synchronizing chaotification feet gradually become asynchronous due to the different speed of deteriorating patency and collateral
circulation in the peripheral arteries. We utilized synchronizing chaotification to compare the bilateral
similarity and asymmetry of PPG signals, and applied SVM to estimate three degrees of PVO. Among 33
subjects tested, this classification technique could recognize various butterfly motion patterns represent-
ing severities successfully including normal condition, lower-degree disease, and higher-degree disease.
The proposed method has potential for providing diagnostic assistance for PVO of diabetics and other
high-risk populations, with efficiency and higher accuracy.
© 2013 Elsevier Ltd. All rights reserved.

1. Introduction sedentary life style. Diabetes causes complications like vasculopa-

Diabetes mellitus is a complex metabolic disorder characterized
by hyperglycemia. This global chronic disease is caused by insulin
deficiency or resistance. There are two main types of diabetes melli-
tus. Type 1 diabetes often inflicts younger patients and results from
autoimmune destruction of insulin-producing cells in pancreas;
while type 2 results from both impairment of insulin secretion
and action, with a rising prevalence in people of overweight and
∗ Corresponding author at: Department of Electrical Engineering, Kao-Yuan Uni-
versity, No. 1821, Jhongshan Road, Lujhu, Kaohsiung City 821, Taiwan.
Tel.: +886 7 6077011.
E-mail addresses: eechl53@gmail.com (C.-H. Lin), chents@mail.ncku.edu.tw
(T. Chen).
1
Department of Biomedical Engineering, National Cheng Kung University, Tainan
City 701, Taiwan.
thy, retinopathy, neuropathy and nephropathy. Diabetic patients
are a high-risk population for development of peripheral vascular
occlusion (PVO). This atherosclerotic change of vessels also afflicts
hypertensive and elderly people, mostly affecting lower limbs. Fur-
thermore, diabetic neuropathy leaves injuries to feet unnoticed,
causing limb-threatening morbidities such as infection and gan-
grene. The risk for amputation can be reduced in diabetes, with
proper intervention and prevention.
Photoplethysmography (PPG) is an optical technique that can be
used to non-invasively detect blood volume changes in the vascular
bed of tissue. The advantages of this technique include compact-
ness, speed, and low cost, with a potential as a portable instrument
for screening PVO in early stage of the disease. Typical PPG sig-
nals consist of alternating current (AC) and direct current (DC)
components, revealing physiological information, e.g., stroke vol-
ume of heartbeat, vasomotor activity, and physiological arousal of

1746-8094/$ – see front matter © 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.bspc.2013.10.001
46 C.-M. Li et al. / Biomedical Signal Processing and Control 9 (2014) 45–55

sympathetic nervous system [1]. PPG signals can be recorded from of synchronous detection. Chaos synchronization provides a new
earlobe, index finger, thumb, and big toe, bilaterally or unilater- approach for deeper understanding of non-linear physical systems
ally. Most current PPG applications focus on cardiac and vascular and physical informatics. In general, a chaotic system consists of
diseases [2–4] and supported by recent development in computer- a master system and a slave system, mimicking the behavior of
based digital signal processing and analysis algorithms. another chaotic system; such a design is called chaos synchroniza-
Both the time-domain and the frequency-domain methods have tion (CS). A typical CS system can be described as:
been applied to analyze plethysmograms of PVO disease [5–7].
Master system : Ẋ = A(X)X (1)
In the time-domain, the PPG signals recorded from both body
Slave System : Ẏ = A(Y )Y − U (2)
sites are described according to timing and shape features. It
was found that the differences of transit time between left and where X = [x1 , x2 , x3 ] and Y = [y1 , y2 , y3 ] are state variables, A
right signals increase as the disease severity exacerbates, and the is called coefficient matrix, and U is control input. The family
calibrated amplitudes decrease [5,6]. These findings reflected asyn- of Lorenz chaotic systems contained at least thirty members of
chronous occlusion and collateral circulation of peripheral arteries three-dimensional autonomous system widely applied to control-
in both lower limbs. It also indicates that PPG signals have compo- ling mechanical, electrical and biological systems like arrhythmias
nents of both low frequency and high frequency that were related and epileptics [11–16]. One of them called the Chen–Lee sys-
to synchronous cardiac changes in vascular blood volume with tem, with known and unknown, can generate two-scroll attractors
each heartbeat [8]. On the other hand, frequency domain analy- and showed good symmetrical behavior, with wide applications
sis showed the bilateral similarity with characteristics of higher [20,21]. This chaotic system provided a simplified mathematical
frequency in healthy subjects, and verifies the bilateral asymmetry model for generation of attractors, as described in the following
with characteristics of lower frequency in diabetic patients [9,10]. non-linear differential Eq. [19]:
⎡ ⎡
⎤ ⎤
In recent years, synchronizing chaotification has been exten- ⎤ ⎡ x1 ⎤
ẋ1 −x2 x3 ⎡
sively applied to adaptive control systems, biomedical signal a 0 0
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
processing, secure communication, and information processing Master System (MS) : ⎣ ẋ2 ⎦ = ⎣0 b 0 ⎦ ⎣ x2 ⎦ + ⎢ x1 x3 ⎥
[11–16]. In this study we proposed to extract a broad range of PPG ⎣ ⎦
0 0 c 1
signals in time domain, and utilized a chaotic non-linear dynamical ẋ3 x3 x1 x2
3
system to construct a mathematical model of PPG that character- (3)
ized by random signals. This system would have broad spectrums of ⎡ ⎤
⎤ ⎡ y1 ⎤ ⎡
Fourier transform, and fractal properties of motion in phase space. ẏ1 a 0 0
⎢ ⎥ ⎢ ⎥
During the development of synchronizing chaotification, a Slave System (SS) : ⎣ ẏ2 ⎦ = ⎣ 0 b 0 ⎦ ⎣ y2 ⎦
Chen–Lee system was designed as a discrete-time chaotic system,
ẏ3
0 0 c y3
which would extract the features from PPG signals. The architec-
ture consisting of a master system and a slave system enabled the ⎡ ⎤ ⎡ ⎤
−y2 y3 u1
response of the slave system to automatically track the response
⎢y y ⎥ ⎢ ⎥
of the master system [17–19]. The dynamic error equation was ⎢ ⎥
+
⎦ + ⎣ u2 ⎦
1 3 (4)
defined as the difference between the slave system and the mas- ⎣
1
ter system, which was used to construct various butterfly patterns y1 y2 u3
for estimation of PVO severity. With a swarm-intelligence algo- 3
rithm for searching optimal solution in high-dimensional problem where a, b, and c are system parameters, and U = [u1 , u2 , u3 ] is
space [20,21], a SVM was constructed to judge the three degrees non-linear controller for tracking the differences of right-to-left
or categories of PVO: normal (Nor) condition, lower-degree (LG) sided PPG signals. For the asymptotically stable, x1 ≈ y1 , x2 ≈ y2 ,
disease, and higher-degree (HG) disease. Eventually, a wolf pack and x3 ≈ y3 , we hypothesized x1 x2 − y1 y2 ≈ (x1 − y1 )(x2 − y2 ),
search (WPS) algorithm was implemented to automatically esti- x1 x3 − y1 y3 ≈ (x1 − y1 )(x3 − y3 ), and x2 x3 − y2 y3 ≈ (x2 − y2 )(x3 − y3 ),
mate the best network parameter and adjust the desired targets and defined the error states e = [e1 , e2 , e3 ]T as Eq. (5).
with training patterns. This proposed classifier was then tested in e1 = x1 − y1 , e2 = x2 − y2 , e3 = x3 − y3 (5)
a dynamical modeling environment to improve classification accu-
Then, subtract Eq. (4) derived from Eq. (3), and define
racy. This paper is organized as follows: Section 2 addresses the
the terms, (x1 − y1 )(x2 − y2 ) = e1 e2 , (x1 − y1 )(x3 − y3 ) = e1 e3 , and
problem formulation, Section 3 describes how the pattern classifier
(x2 − y2 )(x3 − y3 ) = e2 e3 , the dynamics of the error system become
was designed, and Sections 4–6 explains the classifier training, test-
as Eq. (6).
ing results, and the final section provides a conclusion to summarize ⎡ ⎤
the efficiency of the proposed model.
⎡ ⎤ ⎡ ⎤ ⎡ x1 − y1 ⎤ −x2 x3 + y2 y3
ė1
a 0 0
⎢ ⎥ ⎣ ⎢ ⎥ ⎢ ⎥
2. Problem formulation ⎣ ė2 ⎦ = 0 b 0 ⎦ ⎣ x2 − y2 ⎦ + ⎢
⎣
x1 x3 − y1 y3 ⎥
⎦
ė3 0 0 c x3 − y3
1
(x1 x2 − y1 y2 )
2.1. The model design of chaos synchronization (CS) 3
⎡ ⎤⎡
⎡ ⎤ ⎡ e1 ⎤ 0 −e3 0 e1
⎤
The present diagnosis of lower-limb PVO relies on the ankle a 0 0
⎢ ⎥ ⎢ ⎥
0⎥⎢e ⎥
brachial pressure index as a gold standard that is insensitive ≈ ⎣ 0 b 0 ⎦ ⎣ e2 ⎦ + ⎢ e3 0
for hardened arteries. Vascular occlusion can also be diagnosed ⎣ ⎦⎣ 2 ⎦
0 0 c 1
with angiography that is invasive. Modern medicine is still lack- e3 e2 0 0 e3
3
ing a proper tool of diagnostic aid for PVO. As the occlusion of
⎡ ⎤⎡ ⎡ ⎤
peripheral arteries exacerbates, the differences of PPG transit time a −e3 0 ⎤ ⎡ ⎤ a −e3 0
e1 e1
between right and left side would increase significantly, so did ⎢ ⎥ ⎢ ⎥
= ⎢ e3 b 0⎥⎢e ⎥ = ⎢ (e3 )2 ⎥⎢ ⎥
the differences of bilateral amplitudes, referring to bilateral asyn- ⎣ ⎦⎣ 2 ⎦ ⎢⎣
0 b+
a
0 ⎥ ⎣ e2 ⎦
⎦
chronous. By contrast, both discrepancies of bilateral signals were 1
e2 0 c e3 e3
insignificant in healthy subjects, referring to bilateral approximate 3 0 0 c
synchronization. These phenomena can be described as a problem (6)
 C.-M. Li et al. / Biomedical Signal Processing and Control 9 (2014) 45–55 47

Based on previous reports on chaos control, the system (6) was

able to generate a chaotic attractor if these satisfied the following
condition [19–21]:

(A) = ( − a)( − b)( − c) = 0 ⇒ 

1 1

2
= a,  = (b + c) ± (b + c) − 4c (7)
2 2

Subject to a > 0, b < 0, c < 0, and 0 < a < −(b + c)

Restriction condition (7) was exponentially stable with the system’s

parameters for the error system (6). The negative eigenvalues (A)
could be easily determined, and then the robust asymptotic stabil-
ity of system (6) was guaranteed. These scaling factors a, b, and c,
are nonzero constants controlling the chaotic motion of the system
(6), with phase trajectories indicating a controllable chaotic motion.
When this construction applied to bilateral plethysmograms, tim-
ing parameters or time-delay parameters can be used to examine
various dynamic behaviors, such as random or aperiodic signals in
time domain.
For computer implementation, discrete CS systems were used
to describe the dynamic rules and time is represented by the
integer n. The formula of system (6) can be used to track the sim-
ilarity or asymmetry of bilateral PPG signals, on a cycle by cycle
basis. Let the error states be e1 [i] = x[i] − y[i], e2 [i] = x[i + 1] − y[i + 1],
e3 [i] = x[i + 2] − y[i + 2], i = 1, 2, 3, . . ., n − 2, and the initial condi-
tions e1 [0] = e2 [0] = e3 [0] = 0. Then the dynamic error equation was
defined as Eq. (8).
ıy
˚1i = ae1 [i] − e3 [i]e2 [i]
ıx
1 2 (8)
˚2i = be2 [i] + e2 [i](e3 [i])
a erally. Technically, pulse foot (PF)–pulse foot (PF) intervals were
˚3i = ce3 [i]
where x[i], x[i + 1], and x[i + 2] are the data sequence of the right-side
PPG signals, y[i], y[i + 1], and y[i + 2] are those of the left side, and n
is the number of sampling points. Thus, Chen–Lee based CS detec-
tor for PVO estimation can be used to illustrate that the tracking
dynamic error would be zero or bounded ranges.
2.2. Chaotic attractor of bilateral PPG signals
Chronic disease prevention heavily relies on early diagnosis and
risk factor intervention. Lower-limb PVO estimation was a poten-
tial tool of diagnostic aid for providing information to prevent limb
amputation resulting from peripheral vascular disease, presenting
with intermittent claudication, resting leg pain, and gangrene of
toes or feet of diabetes. When PVO disease gradually deteriorates,
asymmetry of bilateral PPG pulses is observed [1,6], because of
concurrent presentation of bilateral diabetic foot is seldom seen
in clinical practice. With the peak of R wave of electrocardiogram
(ECG) as timing reference, several discrepancies of bilateral transit-
timing parameters were defined, such as PTTf (R-peak to pulse
foot), PTTp (R-peak to pulse peak), and RT (foot-to-pulse peak rise
time), as shown in Fig. 1. When the vascular stenosis or the periph-
eral vascular resistance increases, comparison of bilateral PPG was
characteristic of delayed transit time, altered reduced amplitudes
(AMP), and distorted shape of waveforms. More specifically, these
right-to-left side differences of signals, consisting of PTTf , PTTp ,
RT, and AMP, differ more significantly as this vascular dis-
ease worsening. Comparing healthy subjects to diabetic patients,
or other patients with peripheral vascular disease, these timing
parameters would gradually increase and their changes are highly
correlated with PVO severity. In this study, Chen–Lee system was
Fig. 1. ECG and PPG pulses and pulse landmarks for a diabetic patient with PVO
disease.
adopted to examine various dynamic behaviors of bilateral PPG sig-
nals. The aim of the study is to develop a Chen–Lee based chaos
synchronization (CS) detector to estimate the degree of PVO dis-
ease.
Fig. 1 shows PPG pulses and parameters of transit time bilat-
located with n sampling data obtained within an acquiring win-
dow. A total of 800 sampling data (National Instruments DAQ card,
analog-to-digital converter, 16channels, 1 kHz sampling rate and
1.25 ms/s) were used for auto input to a computer via photo-sensor.
The dynamic errors 1i , 2i , and 3i , i = 1, 2, 3, . . ., n − 2, could be
calculated using Eq. (8). The results revealed that these parameters
were specified as a = 2, b = −4, and c = −3, following the specific con-
ditions of a > 0, b < 0, c < 0, and 0 < a < −(b + c), and the dynamic error
system exhibited a unique attractor. The resultant butterfly pat-
terns lied in the first and third quadrant between the systolic rising
edge and dicrotic notch edge, fulfilling the meaning of “synchroniz-
ing chaotification”. The choice of parameters a, b, and c was subject
to Eq. (7) and controlled of the chaotic motions. The origin point
was asymptotically stable, and each trajectory was bounded inside
circular curves. Dynamic errors increased as the disease severity
exacerbates. Most importantly, the feature of dynamic error 2 was
more distinct than 1 and 3 , as shown in Fig. 2. Therefore, non-
linear transformation function, 2 , could be used for enhancing
the bilateral PPG discrepancies in different degrees of PVO dis-
eases. Accordingly, we used 2 to construct various patterns for
PVO estimation.
3. Experimental setup
An optical device was mounted on bilateral big toes of the
subjects to acquire PPG signals operating in reflectance mode.
Components of the device included a light source (940 nm near
infrared, spectral bandwidth: 45 nm, forward voltage: 1.2 V),
photo-detector, trans-impedance amplifier, filter circuit, and inter-
face. In principle, the light source was directed down into the
peripheral vascular bed of the toes, and was backscattered from
the skin adjacent to the photo-detector. The detector converts
48 C.-M. Li et al. / Biomedical Signal Processing and Control 9 (2014) 45–55

Fig. 2. Butterfly motions of Chen–Lee chaotic attractor.

Table 1 differences, supported by clinical presentations like foot gangrene

Preliminary PVO disease assessment with 33 subjects aged 24–65 years.
and findings of angiography.
Parameter Subject

Normal subjects (11) Diabetic patients (12) 4. Pattern classification using wolf pack search algorithm
PTTf (ms)
Mean 2.58 9.2 29.2 4.1. Support vector machine (SVM)
Max–min 0.3–7.4 5.1–23.7 23.6–34.8

PTTP (ms) SVM is a statistical learning method that separates high-

Mean 7.3 22.6
Max–min 0.4–22.3 14.3–56.5
RT (ms)
Mean 6.84 14.6
Max–min 1.3–15.6 3.4–32.3
Ankle Brachial Pressure ≥0.9 <0.9
Index (ABPI)
Note: The values of PTTf , PTTp , and RT are absolute values. (1)
PTTf = |PTTRf − PTTLf |, (2) PTTp = |PTTRp − PTTLp |, and (3) RT = |RTR − RTL |, where
suffix words R and L are defined the right and left legs.
light energy into an electrical voltage and connected to a trans-
impedance amplifier and filter circuits. DC component of electrical
voltage was reduced by high-pass Butterworth filter. In this study,
signals were captured at a sampling rate of 1 kHz, and a NI DAQ card
acts as an analog-to-digital (A/D) converter. The advantages of this
optical instrument as a diagnostic aid of PVO were long operating
life, low cost, and real-time acquisition.
Total 33 subjects were recruited that included 16 healthy sub-
jects and 17 diabetic patients, ranging in age from 24 to 65 years,
who were inpatients of division of infectious diseases medical cen-
ter in Chi Mei hospital, Tainan, Taiwan. The study was approved
by the hospital research ethics committee. All subjects also gave
written informed consent. Twenty-three data sets were randomly
selected as training data for training a SVM, and the others were
used as testing data. Subjects in the prospective study were divided
into three groups: normal (Nor) subjects, diabetic patients with
lower grade (LG) disease, and higher-grade (HG) disease. Table 1
shows the preliminary PVO disease estimation based on the Ankle
Brachial Pressure Index (ABPI) and the ranges of bilateral-timing
52 dimensional data or variables by finding a hyper-plane. With
46.2–57.8 complicated data, a SVM can map to a high-dimensional feature
space by the non-linear mapping of the input vector and target
23.4 vector. This study applied SVM to group the dataset of PVO into
11.5–35.3 three classes: normal, LG, and HG. For a classification problem, an
<0.5
individual classifier with fusion of training patterns increases train-
ing time and decreases adaptive capability. Therefore, input pattern
is labeled as one class or none, turning the SVM to be a linear
machine with one output that dealt with linearly and non-linearly,
which is separable for two-class training patterns [22–25]. Binary
classification is then achieved using this optimal hyper-plane.
Given the input vector  = [ϕ1 , ϕ2 , . . ., ϕi , . . ., ϕn−2 ] and target vec-
tor f(), the hyper-plane is determined by an orthogonal vector W
and a bias, which satisfy WT  + bias = 0. By finding a hyper-plane
that maximizes the margin to the two groups of data, “+1” for class
I and “−” for class II (linearly separable application), it is reasonably
expected that this classifier will have better generalization ability.
The SVM has the following linear decision function:
fj (˚) = sign(W T  + bias)
+1, W T  + bias > +1, for Class I
= (9)
−1, W T  + bias < −1, for Class II
where (W, bias) are chosen to correctly classify training data and
maximize the margin of hyper-plane, with a set of K training-data
pairs (input vector–target vector).
Most of the data, which are geometrically represented in input
space, are not linearly separable. SVM can find an optimal hyper-
plane for linearly separable patterns; otherwise, it also can transfer
 C.-M. Li et al. / Biomedical Signal Processing and Control 9 (2014) 45–55 49

original data into higher dimensional space. In this study, a specific

function called “kernel-based transformation” transfers the input
space into higher dimensional space by non-linear transformation.
Let support vector k has (n − 2)-dimension, k = 1, 2, 3, . . ., K, and
target vector wjk , j = 1, 2, 3, . . ., m, the hyper-plane of jth output
node for non-linearly separable data is defined as [23]

K
fj (˚) = ˛jk wjk R(˚k , ) + biasj (10)
k=1

⎡   2 ⎤
1  − k 
R(˚k , ) = exp ⎣− ⎦ (11)
2 

where parameter ˛jk , 0 ≤ ˛jk ≤ C, is the nonnegative multiplier, C is

a constant parameter (C = 1 in this study) with flexibility allowing
for separation of the classes, R(k ,) is a Gaussian function,  is
the width of the Gaussian function, K is the number of support vec-
tors (training patterns), and m is the number of output nodes. The
desired target vector Wk = [w1k , w2k , w3k ] = [Nor, LG, HG], m = 3, are
encoded as binary values with signal “1” or “0” for the three disease
severity of PVO.
The algorithm was first designed for SVM mapping the input Fig. 3. Architecture of the SVM classifier with wolf pack search algorithm.
vectors into a high-dimensional feature space. Based on the least
mean square method, let erjk = fjk − fj (k ), and find the optimal
densest odor. When approaching the targeted prey, they can sepa-
parameters ˛j = ˛j1 = ˛j2 = ˛j3 = · · · = ˛jk , , and biasj . This implemen-
rate into several subgroups, as shown in Fig. 4. Each subgroup will
tation is sufficient to minimize the object function, which is defined
chase the quarry in different routes wolf, which can be expressed
as the mean squared error function (MSEF)
as
⎛ ⎞
1

m K
p
MSEF = (erjk )2 pmax − p wolfbest − wolfg
= wolfg + ⎝ ⎠
p+1 p
K (12) wolfg )( 
j=1 k=1 pmax 2 p 2
Subject to 0 < aj ≤ 1, j = 1, 2, 3, . . ., m (wolfbest ) + (wolfg )
p p
where ˛j , , and biasj are regularization parameters, controlling a = wolfg + pwolfg (14)
compromise between maximizing the margin and minimizing the
where wolfbest is the best search route, wolfg p the search route, g
MSEF. With the non-linear SVM mechanism, a non-linear decision
the population size, p the number of running steps, pmax the max-
function is used for extending the non-linear boundary to separate
imum running step, and wolfg p the change of search route. With
the high dimensional and the training data that are non-linearly
time-varying running steps (TVRS), wolves are allowed to move
separable. Given the final outputs of node fj (), j = 1, 2, 3, . . ., m,
large distance p around the search space at the beginning stage.
fj () ∈ {0,1}, the decision function is defined as Eq. (13).
 K Each subgroup will communicate the other wolf groups and jointly

1, ≥
fj () = sign ˛jk wjk R(k , ) + biasj = sign() =
k=1 0, <
(13)

= ˛jk wjk R(k , ) + biasj

k=1

Eq. (13) is a label that determines the class of input vector . Once
a binary classifier is trained with the known training patterns and
the corresponding target associations, the targets fjk , k = 1, 2, 3, . . .,
K, are encoded as binary values with signal “1” denoting the three
disease severity, and the rest of the weights are zero. In our design,
three SVM classifiers are built for estimation of the three degrees
of PVO disease. Fig. 3 illustrates these output nodes (m = 3).

4.2. Wolf pack search (WPS) algorithm

Wolf pack search algorithm (WPS) is a relatively new swarm

intelligence algorithm, which simulates the hunting behavior of
wolf packs to solve optimization problems [26,27]. In nature, a pack
of wolves can work together to increase the change of catching a
prey. This hunting behavior can be simplified as the following pro-
cess. The wolves initially roam around. They begin to search for the
quarry odor left as coat and egesta. Then the arrest activity begins
as the wolf pack tracking the herd of preys in the direction of the Fig. 4. Searching concept with wolf group in a solution space by WPS algorithm.
50 C.-M. Li et al. / Biomedical Signal Processing and Control 9 (2014) 45–55
Table 2
Quantitative results of the WPS algorithm.
Task Method
WPS with TVRS
Training data 23
Population size G = 20–40
Convergent condition Maximum number of iterations pmax = 25–50
Parameter assignment ˛j (0 ≤ ˛j ≤ 1), , and biasj = 0
Note:(1) TVRS: time-varying running step; (2) CRS: constant running step.
attack the targeted quarry. By reducing the walk p, a small step
change allows the wolves to chase down the quarry (optimal solu-
tion) at the end of the continuous search. Based on this metaphoric
behavior, the WPS is summarized as follows:
Step 1: Randomly generate a pack of wolves G, g = 1, 2, 3, . . ., G,
and set the maximum number of running steps pmax .
Step 2: Find the best route, wolfbest , defined as optimal solution
and calculate the object function MSEF.
Step 3: Update the running route wolfg p of all the wolves using Eq.
(14).
Step 4: Circulate and update the running routes. If the maximum
number of running step is achieved or the MSEF is less than the
pre-specified value, then stop the procedure, or else, go back to
Step 2.
In the WPS algorithm, when a new solution is more optimal than
the best solution so far, it can be saved as the best one. Therefore,
the solution sequence is monotone decreasing. In this study, sup-
pose constant biasj = 0, adjusting parameters ˛ (˛=˛1 = ˛2 = · · · = ˛m )
and  would refine the discrimination accuracy in a dynamic envi-
ronment. The optimum parameters are intended to minimize the
MSEF, which are given by
⎛ ⎞
˛best − ˛jg p
˛jg p+1 = ˛jg p + p ⎝  ⎠ (15)
(˛best )2 + (˛jg p )2
a simplified concept to easily control parameters and implement
best − g p
g p+1 p
= g + p 
(best )2 + (g p )2
Subject to 0 < aj ≤ 1, j = 1, 2, 3, . . ., m
There are two convergent conditions to terminate the WPS algo-
rithm: (a) the objective function MSEF is less than the pre-specified
value ε; (b) the number of running steps achieves the maximum
allowable number pmax . Thus, the optimum parameters can be
obtained by minimizing the MSEF.
4.3. SVM classifier training results
PPG pulses were simultaneously measured from both toes of
the tested subjects. Then each PF–PF interval of the signals were
located, and 800 sampling data were acquired within an sampling
window, as shown in Fig. 5(a), where the horizontal axis is the
sampling number, and the vertical axis is the normalized ampli-
tude. In this figure, the rising edge left to the vertical dashed line
exhibits a systolic phase of cardiac cycle, and the dicrotic notch edge
on the other side can be seen during diastole in healthy compliant
arteries. In case that resistance or obstruction of peripheral vessels
increases, bilateral PPG pulses become asynchronous, resulting to
proportional increase of transit time delay. According to Eq. (8), the
dynamic errors 2i , i = 1, 2, 3, . . ., n − 2 (n = 800), were calculated
and shown in Fig. 5(b). Let a = 2, b = −4, and c = −3, the dynamic
WPS with CRS
23
G = 20–40
Maximum number of iterations pmax = 25–50
˛j (0 ≤ ˛j ≤ 1), , biasj = 0, and p = 0–1
errors 1 , 2 , and 3 reflect the characteristic chaotic behaviors
for the three disease degrees, with butterfly patterns lying in the
first and third quadrant. As a chaotic system displays, the choice
of parameters a, b, and c controls governs the chaotic motion,
and its trajectory diverges exponentially as these parameters
increase.
Parts of these butterfly patterns were used to train SVM classi-
fier. We have 23 input–output pairs of training data (K = 23), each
number of averaged patterns from the same category are 11-, 6-,
and 6-set data, respectively. These datasets are computed from a
total of 20 good-quality PPG signals from consecutive recording. We
proposed an SVM classifier with 798 input nodes, 3 decision nodes,
and 3 output nodes (m = 3). The desired target vectors were encoded
as binary values, following the pattern [w1k , w2k , w3k ] = [Nor, LG,
HG] = [1/0, 1/0, 1/0], k = 1, 2, 3, . . ., 23, for 3 degrees, and the vectors
were encoded as Nor: [1 0 0], LG: [0 1 0], and HG: [0 0 1]. A thresh-
old value  = 0.5 was set to confirm the decision boundary using Eq.
(13), which can be defined as

1, ≥
fj () = sign() = , j = 1, 2, 3 (17)
0, <
where the index fj () is binary value “1” referring to a possible
degree.
The quantitative results of the WPS algorithm are shown in
Table 2. In optimization research, WPS can find a good solution
using three control factors, such as population size, number of run-
ning step, and moving step. A WPS with constant running step is
in engineering problems. In this study the WPS with an invari-
(16) able velocity was given by population size G = 20, 30, and 40, walk
p = 0.2, 0.5, and 0.7 for each step, with maximum allowable run-
ning step pmax = 25, and pre-specified value ε ≤ 0.05. The walk p
was determined following a trial-and-error selection; as a result,
the fittest parameters were obtained from the training dataset and
shown in Table 3. One of its advantages is high search speed for
a good solution, requiring less than 5 running steps, as shown in
Fig. 6(a). However, when confronting with inequality constraint
0 < ˛ ≤ 1, special attention must be paid to avoid being trapped
on local minima and resulting premature convergence. Therefore,
global searching of the WPS was improved with varying running
step given by the population size G = 20–40 for each running step,
and maximum allowable running step pmax = 25 and 50. For exper-
imental analysis, only slight improvement of optimal parameters
was obtained by increasing population size and running step. In
addition, the whole process had a large evolutionary computation,
and average CPU time also increased from 5.131 to 10.194 s. It could
converge within less than 10 running steps to classify the 23-pairs
of training data. As a result, the fittest parameters  best = 0.0879
and ˛best = 0.8917 can minimize the MSEF with G = 30, pmax = 25,
and the pre-specified value ε ≤ 0.03. Fig. 6(b) illustrates the solution
list that is monotone decreasing. Thus, WPS with time-varying run-
ning step provides a stable mode for SVM learning and construction
procedure.
 C.-M. Li et al. / Biomedical Signal Processing and Control 9 (2014) 45–55 51

Fig. 5. (a) PPG pulses at the right and left big toes from normal subject and diabetic patients with PVO disease. (b) The dynamic errors 2 for normal subject and diabetic
patients with PVO disease.

Table 3
Results of the parameter estimation.

Population Maximum Walk p Parameter MSEF ≤ 0.03

size G running estimation
step pmax
 ˛

20 25 TVRS 0.4730 0.8225 0.0271

30 0.0879 0.8917 0.0247
40 0.1834 0.8448 0.0265

20 50 TVRS 0.6444 0.7471 0.0260

30 0.1578 0.8599 0.0292
40 0.4808 0.8323 0.0253

20 25 CRS p = 0.2 0.2286 0.8620 0.0523

30 0.1876 0.8529 0.1274
40 0.2248 0.9089 0.0916

20 25 CRS p = 0.5 0.5274 0.8689 0.0517

30 0.5131 0.8982 0.0688
40 0.5374 0.7843 0.0422

20 25 CRS p = 0.7 0.7504 0.7871 0.0305

30 0.5814 0.8934 0.1115
40 0.6523 0.8432 0.1897
52 C.-M. Li et al. / Biomedical Signal Processing and Control 9 (2014) 45–55
Fig. 6. (a) Mean squared errors versus the number of running steps for WPS with
CRS. (b) Mean squared errors versus the number of running steps for WPS with TVRS.
5. Experimental results and discussion
The proposed method was developed on a PC Pentium-IV
3.0 GHz with 480 MB RAM and MATLAB software (Mathwork, Nat-
ick, MA). Portable optical measurement was employed to acquire
the biomedical signals of the tested persons. The experiment was
proceeded as the follows: (1) record PPG signals; (2) track the differ-
ence of the timing parameters using the dynamic error equation;
and (3) estimate PVO degree with the SVM classifier. To demon-
strate the effectiveness of the proposed method, ten subjects were
tested as shown in Table 4. With the reference to ABPI and clinical
presentations like findings of angiography, there are five records
of the normal subjects (#1–5), three of cases of low-degree (LG)
disease (#6–8), and two of high-degree (HG) (#9–10). Numerical
experiments are presented as detailed below to verify the proposed
method.
5.1. PVO disease estimation
Before PPG measurement, subjects were asked to relax and
lie supine on a couch for 10 min. The laboratory tempera-
ture was 25 ± 1 ◦ C, while the hospital room temperature was
set at 23 ± 1 ◦ C. The examination lasted for 10 min, with two
PPG probes mounted on both big toes of subjects as Fig. 7
illustrated. In Fig. 8, a representative result is displayed, con-
taining 10 time-domain signals, dynamic errors, and butterfly
motions for LG disease (#6) and HG disease (#9). Our test
results showed that the normal range of timing parameters
PTTf , PTTp , and RT were: 0.3 ms–7.4 ms, 0.4 ms–22.3 ms, and
1.3 ms–15.6 ms, respectively, as listed in Table 1. According to
the bilateral-timing differences and the ABPI, except two sub-
jects (#6: PTTf = 15.321, PTTp = 17.286, RT = 1.964, and #9:
PTTf = 33.072, PTTp = 48.500, RT = 15.428) were rightly clas-
sified to the output space of PVO diseases.
These timing differences of bilateral plethysmograms are highly
correlated with different severity of PVO in a retrospective study
Fig. 7. Practical measurements in the medical center.
conducted by John Allen et al. [5]. Two subjects were randomly
chosen to validate the diagnostic value of this method. Initially,
each peak-foot interval was identified with digital signal processing
algorism, followed by the computation of dynamic errors. When
the trajectories approached a stable point, the origin point was
seen as asymptotically stable for normal condition, as shown in
Fig. 8. For the bilateral synchronous signals (normal subject), the
trajectories of the butterfly patterns were bounded in red dash-line
portion (1: ±0.10 and 2: ±0.10). By contrast, at the phase plane
of 1 versus 2 , their trajectories diverged in higher degree of PVO
disease; accordingly, the resultant tracking dynamic errors can be
utilized to estimate the degree of the PVO disease, as illustrated in
Fig. 8. With the 100 consecutive PPG signals, the diagnostic results
of the ten subjects in Table 4 demonstrate that the outputs of the
SVM classifier could be computed using Eq. (17), with correct esti-
mation in the respective output space of the Nor, LG, and HG. The
maximum one is greater than the threshold  = 0.5, and sign() = 1,
 ≥ , indicating the possible degree. This result confirms the major
degrees are “LG disease” and “HG disease”, as shown in Table 4.
These test results imply that the proposed method has the poten-
tial to be a diagnostic tool for estimating the degree of PVO disease
and monitoring the progression of this vascular disorder.
5.2. Classifier performance comparison
In literature many classification algorithms, including artificial
neural network [28,29], the Fuzzy method, and adaptive network-
based fuzzy inference system (ANFIS) [30], have been proposed for
PVO disease estimation. By comparison, Fuzzy method has three
input variables for PTTf , PTTp , and RT, each has three mem-
bership functions for various ranges of bilateral timing differences;
further, each output variable needs three membership functions for
individual group of Nor, LG, and HG, as shown in Table 1. The seven
linguistic Fuzzy rules for the three degrees are based on the ankle
brachial index [14].
The Fuzzy decision representing the output is defuzzified using
the center-of-area method, which converts linguistic information
to numerical center values. For the 10 subjects tested, the Fuzzy
method had 80% accuracy with two failures, as depicted in Table 4.
 C.-M. Li et al. / Biomedical Signal Processing and Control 9 (2014) 45–55 53

Fig. 8. (a) PPG signals in the time-domain, dynamic error 2 , and butterfly motions for normal subject, (b) PPG signals in the time-domain, dynamic error 2 , and butterfly
motions for diabetic patient with LG disease and (c) PPG signals in the time-domain, dynamic error 2 , and butterfly motions for diabetic patient with HG disease.

Measurement of parameters PTTf , PTTp , and RT might be affected by
cardiac cycle, asynchronous measurement during extraction of sig-
nals, and quantification error, leading to incorrect the PVO disease
estimation. ANFIS describes the linguistic variables and memorizes
some assigned weighted values, 1, 2, and 3, for the three degrees
of the vascular disease. It is a feed-forward multi-layer network,
integrating the basic principle of fuzzy system into a connectionist
structure. Table 5 summarizes the quantitative results of compar-
isons among these methods. The results show that ANFIS can be
trained by numerical data and linguistic information expressed
by if-then rules, and can automatically construct itself from
training data [30]. The back-propagation algorithm is used to find
the optimum parameters under mean square error criterion, with
an enhanceable accuracy of estimation.
A neural network classifier is an adaptive mechanism that has
been widely applied in continuous modeling system (network
topology 799 × 23 × 4 × 3), which is suitable for tasks such as tuning
network parameters and automatic target adjustment. By contrast,
particle swarm optimization algorithm is an optimization tech-
nique that avoids the drawbacks of back-propagation algorithm,
regardless of gradient or steepest descent method. Briefly, these
algorithms have difficulty in obtaining the first and second partial

Table 4
The results of PVO disease estimation in ten subjects.

Subject no. Mean values of bilateral differences Professional Proposed ANN decision Fuzzy decision ANFIS
physicians method (average (center value) decision
decision decision output) (center value)
(sign() = 1,
PTTf (ms) PTTp (ms) RT (ms)  ≥ 0.5)

1 2.7330 1.9970 3.1330 Normal Normal [1 0 0] Normal [1 0 0] Normal (0.1750) Normal

(1.0000)
2 1.1730 5.8700 4.6950 Normal Normal [1 0 0] Normal [1 0 0] Normal (0.1620) Normal
(1.0000)
3 2.3330 3.6660 1.3330 Normal Normal [1 0 0] Normal [1 0 0] Failure Normal
(1.0000)
4 3.7910 4.2500 8.0420 Normal Normal [1 0 0] Normal [1 0 0] Normal (0.1480) Normal
(1.0000)
5 6.1670 4.9170 11.084 Normal Normal [1 0 0] Normal [1 0 0] Normal (0.1730) Normal
(1.0000)
*6 15.321 17.286 1.9640 LG diabetic LG diabetic LG diabetic LG diabetic LG diabetic
[0 1 0] [0 1 0] (0.4000) (2.0000)
7 6.3850 16.923 17.556 LG diabetic LG diabetic LG diabetic LG diabetic LG diabetic
[0 1 0] [0 1 0] (0.3830) (1.9996)
8 23.667 33.333 32.833 LG diabetic LG diabetic LG diabetic LG diabetic LG diabetic
[0 1 0] [0 1 0] (0.4020) (2.0000)
*9 33.072 48.500 15.428 HG diabetic HG diabetic HG diabetic HG diabetic HG diabetic
[0 0 1] [0 0 1] (0.8000) (2.9996)
10 23.606 57.700 35.000 HG diabetic HG diabetic HG diabetic Failure HG diabetic
[0 0 1] [0 0 1] (2.9995)

Note: (1) ANFIS method: 3 singleton membership functions, and the center values of the functions are 1, 2, and 3 for Nor, LG, and HG, (2) Fuzzy method: 1 output variable
with 3 triangular membership functions, and the center values of the functions are 0.0, 0.4, and 0.8 for Nor, LG, and HG.
54 C.-M. Li et al. / Biomedical Signal Processing and Control 9 (2014) 45–55

Table 5
Related data of the proposed method and other methods.

Method Related parameter and architecture

Proposed method CS detector Chen–Lee chaotic system

Network architecture • Input layer: 798 nodes
• Decision layer: 3 decision functions
• Output layer: 3 nodes
Learning algorithm WPS with TVRS
Parameter assignment • Maximum number of running steps
• Population size in wolf pack
• Convergent condition

ANN method CS detector Duffing–Holmes chaotic system [30]

Network architecture • Input layer: 799 nodes
• Hidden layer: 23 nodes, nodes
• Summation layer: 4 nodes
• Output layer: 3 nodes
Learning algorithm PSO algorithm
Parameter assignment • Maximum number of iterations
• Population size in particle swarm
• Initial value and final value (range assignment) of the
acceleration coefficients
• Inertia weight coefficient (range assignment)
• Convergent condition

Fuzzy method Architecture • 3 Input variables, each variable having 3 triangular

membership functions
• Defuzzifier
• 1 Output variable with 3 triangular membership
functions
Rules 7 Linguistic fuzzy rules

ANFIS method [31] Network architecture • Input layer: 3 linguistic nodes

• Fuzzification Layer: 9 nodes
• Fuzzy rule layer: 7 nodes
• Defuzzifier: 4 nodes
• Output layer: 3 nodes
Learning algorithm BP algorithm
Parameter assignment • According to Table 1, triangular membership functions
for linguistic nodes are defined by the mean values and
max–min boundary values.
• 3 Singleton membership functions
Rules 7 Linguistic fuzzy rules

derivatives of non-linear mean square error function. Furthermore,
learning performance is heavily relied on the choices of the initial
condition, learning rate, and convergent condition. Under high-
dimensional pattern space, the training process and classification
efficiency become a main problem when handling huge train-
ing data. Training neural network is time-consuming, without a
guaranteed global minimum. However, the swarm intelligence can
guide searches using high-dimensional search space or multiple
particles rather than individuals. Individuals in a swarm approach
the optimum through the present solution, previous experience,
and other individuals’ experiences, and can avoid of trapping at
a local minimum. A good solution is sensitive in consideration of
certain control factors, such as the number of iterations and pop-
ulation sizes, time-varying acceleration coefficients, and inertia
weight coefficient as a whole. Suitable control of those param-
eters is important to find the optimum solution accurately and
efficiently. Our comparison also indicates that neural network using
particle swarm optimization algorithm also has high accuracy for
PVO estimation.
Another finding of the present study is the proposed SVM-
based classifier could be built with a small network architecture
(network topology 798 × 3 × 3) for high-dimensional pattern clas-
sification. A simple WPS algorithm was written to search for a
good solution nonlinearly, with training data and less parame-
ter assignment. The WPS with time-varying running step would
lower the possibility of premature convergence in the early
search stage, and improve convergence at the end of optimiza-
tion stage. So this machine learning provides a potential technique
for classifier design with relatively short design cycle. Through the
experimental tests, a Chen–Lee based chaos synchronization detec-
tor was able to track synchronous and asynchronous indices from
the bilateral PPG signals. Eventually, the dynamic errors and chao-
tification patterns were successfully applied to estimate the degree
of PVO disease.
6. Conclusion
In this study synchronizing chaotification for feature extrac-
tion was designed using Chen–Lee chaotic system to estimate
disease severity of peripheral vascular occlusion. This technique
was developed to track dynamic errors and construct butterfly
motion patterns from the difference of the timing parameters
between bilateral PPG signals. Due to the increases of differences,
the results show that dynamic errors increase as asynchronous
phenomena gradually become serious, and motion trajectories are
also increasingly revealed in the diabetic group. And then the
classifier of support vector machine was applied to estimate the dis-
ease severity. In addition, numerical experiments are presented to
demonstrate and verify the performance of the proposed method.
This optical-based technology without extra ECG measurement
could be a portable diagnostic aid for PVO as it is a noninvasive, low
cost, and simple-to-use method. However, automatic processing of
PPG waveform and timing parameters still deserves further atten-
tion. In order to develop an automatic screening tool for PVO, the
proposed method can be further integrated in computer-based
digital signal processing and telemedicine for home healthcare
applications.
 C.-M. Li et al. / Biomedical Signal Processing and Control 9 (2014) 45–55
Acknowledgments
This work is supported in part by the National Science Council
of Taiwan, under contract number: NSC 101-2221-E-244-001 and
NSC 102-2218-E-218-006, duration August 1, 2012–July 31, 2014.
The Institutional Review Board (IRB) of the Division of
Infectious Diseases of Chi Mei Medical Center approved this
study. The research study was also approved by the hospital
research ethics committee, correspondence: Dr. Chien-Ming Li,
yrho@mail.ncku.edu.tw.
References
[1] J. Allenc, Photoplethysmography and its application in clinical physiological
measurement, Physiological Measurement 28 ((3) February) (2007) 1–39.
[2] M. Nitzan, A. Babchenko, B. Khonokh, D. Landau, The variability of the pho-
toplethysmographic signal—a potential method for the evaluation of the
automatic nervous system, Physiological Measurement 19 (1998) 93–102.
[3] M. Nitzan, A. Babchenko, B. Khonokh, Very low frequency variability in arterial
blood pressure and blood volume pulse, Medical & Biological Engineering &
Computing 37 (1) (1999) 54–58.
[4] D.W. Kim, S.W. Kim, S.C. Kim, K.C. Nam, E.S. Kang, J.J. Im, Detection of dia-
betic neuropathy using blood volume ratio of finger and toe by PPG, in:
Proceedings of the 29th Annual International Conference of the IEEE Engineer-
ing in Medicine and Biology Society, August, Lyon, France, vol. 23–26, 2007, pp.
2211–2214.
[5] M. Nitzan, B. Khonokh, Y. Slovik, The difference in pulse transit time to the toe
and figure measured by photo-plethysmography, Physiological Measurement
23 (2002) 85–93.
[6] J. Allen, C.P. Oates, T.A. Lees, A. Murray, Photoplethysmography detection of
lower limb peripheral arterial occlusive disease: a comparison of pulse tim-
ing, amplitude and shape characteristics, Physiological Measurement 26 (2005)
811–821.
[7] H. Hsiu, C.-L. Hsu, T.-L. Wu, A preliminary study on the correlation of frequency
components between finger PPG and radial arterial BP waveforms, in: Inter-
national Conference on Biomedical and Pharmaceutical Engineering, 2009, pp.
1–4.
[8] W.B. Murray, P.A. Foster, The peripheral pulse wave: information overlooked,
Journal of Clinical Monitoring and Computing 12 (5) (1996) 365–377.
[9] A. Buchs, Y. Slovik, M. Rapoport, C. Rosenfeld, B. Khanokh, M. Nitzan, Right-left
correlation of the sympathetically induced fluctuations of the photoplethys-
mographic signal in diabetic and non-diabetic subjects, Medical & Biological
Engineering & Computing 43 (2) (2005) 252–257.
[10] L. Gbaoui, E. Kaniusas, Decomposition of photoplethysmographical arterial
pulse waves by independent component analysis: possibilities and limitations,
Lecture Notes in Electrical Engineering 55 (2010) 1100–1119.
[11] C.D. Wagner, R. Mrowka, B. Nafz, P.B. Persson, Complexity and chaos in blood
pressure after baroreceptor denervation of conscious dogs, American Physio-
logical Society 269 (5) (1996) 1760–1766.
[12] J.S. Lin, J.-J. Yan, T.L. Liao, Chaotic synchronization via adaptive sliding mode
observers subject to input nonlinearity, Chaos, Solitons & Fractals 24 (1) (2005)
371–381.
55
[13] J.-S. Lin, J.-J. Yan, T.-L. Liao, Robust control of chaos in lorenz systems subject to
mismatch uncertainties, Chaos, Solitons & Fractals 27 (2) (2006) 501–510.
[14] X. Zhang, A. Khadra, D. Yang, D. Li, Unified impulsive fuzzy-model-based
controllers for chaotic systems with parameter uncertainties via LMI, Com-
munication in Nonlinear Science and Numerical Simulation 15 (1) (2010) 105–
144.
[15] A. Bogris, D. Kanakidis, A. Argyris, D. Syvridis, Performance characterization
of a closed-loop chaotic communication system including fiber transmission
in dispersion shifted fibers, IEEE Journal of Quantum Electronics 40 (9) (2004)
1326–1336.
[16] Y. Hong, K.A. Shore, Power loss resilience in laser diode-based optical chaos
communications systems, Journal of Light wave Technology 28 (3) (2010)
270–276.
[17] H.K. Chen, C.-I. Lee, Anti-control of chaos in rigid body motion, Chaos, Solitons
& Fractals 21 (4) (2004) 957–965.
[18] H.K. Chen, Global chaos synchronization of new chaotic systems via nonlinear
control, Chaos, Solitons & Fractals 23 (4) (2005) 1245–1251.
[19] H.K. Chen, Synchronization of two different chaotic systems: a new system and
each of the dynamical systems Lorenz, Chaos, Solitons & Fractals 23 (5) (2005)
1245–1251.
[20] J.H. Chen, Controlling chaos and chaotification in the Chen–Lee system
by multiple time delays, Chaos, Solitons & Fractals 36 (4) (2008) 843–
852.
[21] L.M. Tama, J.-H. Chen, H.-K. Chen, W.-M. Si Tou, Generation of hyperchaos from
the Chen–Lee system via sinusoidal perturbation, Chaos, Solitons & Fractals 38
(3) (2005) 826–839.
[22] E. Pranckeviciene, R. Somorjai, R. Baumgartner, M.-G. Jeon, Identification of sig-
natures in biomedical spectra using domain knowledge, Artificial Intelligence
in Medicine 35 ((3) November) (2005) 215–226.
[23] J. Kamruzzaman, R.K. Begg, Support vector machines and other pattern recog-
nition approaches to the diagnosis of cerebral palsy gait, IEEE Transactions on
Biomedical Engineering 53 (12) (2006) 2479–2490.
[24] M.E. Mavroforakis, H.V. Georgiou, N. Dimitropoulos, D. avourasionisis, S.
Theodoridis, Mammographic masses characterization based on localized tex-
ture and dataset fractal analysis using linear, neural and support vector
machine classifiers, Artificial Intelligence in Medicine 37 ((2) June) (2006)
145–162.
[25] Y. Peng, Y. Zhang, L. Wang, Artificial intelligence in biomedical engineering and
informatics: an introduction and review, Artificial Intelligence in Medicine 48
((2/3) February/March) (2010) 71–73.
[26] H.S. Chang, Converging marriage in honey-bees optimization and application to
stochastic dynamic programming, Journal of Global Optimization 35 (3) (2006)
423–441.
[27] C. Yang, X. Tu, J. Chen, Algorithm of marriage in honey bees optimization based
on the wolf pack search, in: International Conference on Intelligent Pervasive
Computing, Jeju Island, Korea, 2007, pp. 462–467.
[28] C.-H. Lin, Y.-F. Chen, Y.-C. Du, J.-X. Wu, T. Chen, Chaos synchronization detec-
tor combining radial basis network for estimation of lower limb peripheral
vascular occlusive disease, Lecture Notes in Computer Science 6165 (2010)
126–136.
[29] H.-T. Yau, C.-L. Kuo, J.-J. Yan, Fuzzy sliding mode control for a class of chaos
synchronization with uncertainties, International Journal of Nonlinear Sciences
and Numerical Simulation 7 (3) (2006) 333–338.
[30] Y.-C. Du, C.-H. Lin, Adaptive network-based fuzzy inference system for assess-
ment of lower limb peripheral vascular occlusive disease, Journal of Medical
System 36 (2012) 301–310.