Comparison of Three Different Types of Wrist Pulse Signals by Their Physical Meanings and Diagnosis Performance

Abstract

Increasing interest has been focused on computational pulse diagnosis where sensors are developed to acquire pulse signals and machine learning techniques are exploited to analyze health conditions based on the acquired pulse signals. By far, a number of sensors have been employed for pulse signal acquisition, which can be grouped into three major categories, i.e., pressure, photoelectric, and ultrasonic sensors. To guide the sensor selection for computational pulse diagnosis, in this paper, we analyze the physical meanings and sensitivities of signals acquired by these three types of sensors. The dependency and complementarity of the different sensors are discussed from both the perspective of cardiovascular fluid dynamics and comparative experiments by evaluating disease classification performance. Experimental results indicate that, each sensor is more appropriate for the diagnosis of some specific disease that the changes of physiological factors can be effectively reflected by the sensor, e.g., ultrasonic sensor for diabetes and pressure sensor for arteriosclerosis, and improved diagnosis performance can be obtained by combining three types of signals.

Full text

1
 Abstract— Increasing interest has been focused on
computational pulse diagnosis where sensors are developed to
acquire pulse signals and machine learning techniques are
exploited to analyze health conditions based on the acquired pulse
signals. By far, a number of sensors have been employed for pulse
signal acquisition, which can be grouped into three major
categories, i.e., pressure, photoelectric, and ultrasonic sensors. To
guide the sensor selection for computational pulse diagnosis, in this
paper, we analyze the physical meanings and sensitivities of signals
acquired by these three types of sensors. The dependency and
complementarity of the different sensors are discussed from both
the perspective of cardiovascular fluid dynamics and comparative
experiments by evaluating disease classification performance.
Experimental results indicate that, each sensor is more
appropriate for the diagnosis of some specific disease that the
changes of physiological factors can be effectively reflected by the
sensor, e.g., ultrasonic sensor for diabetes and pressure sensor for
arteriosclerosis, and improved diagnosis performance can be
obtained by combining three types of signals.
Index Terms— computational pulse diagnosis; pulse signal
acquisition; pressure sensor; photoelectric sensor; ultrasonic
sensor.
I. INTRODUCTION
ULSE diagnosis has played an important role in traditional
Chinese medicine (TCM) and traditional Ayurvedic
medicine (TAM) for thousands of years [1-3]. Generally, the
wrist pulse signals are mainly produced by cardiac contraction
and relaxation and are also affected by the movement of blood
and changes in the vessel diameter, making them effective for
analyzing both cardiac and non-cardiac diseases.
However, pulse diagnosis is a subjective skill which needs
years of training and practice to master [4]. Moreover, the
diagnosis result relies on the personal experience of the
practitioner. With different practitioners, the diagnosis results
may then be inconsistent. To overcome these limitations,
computational pulse diagnosis has been recently studied to
objectify and quantify pulse diagnosis, and researchers have
Manuscript received June 14, 2014. This work is partially supported by the
General Research Fund from the Hong Kong SAR Government, in part by the
central fund from the Hong Kong Polytechnic University, National Natural
Science Foundation of China under Grant Nos. 61271093, 61332011,
61020106004, 61272292, and 61271344, Shenzhen Fundamental Research
Fund (JCYJ20130401152508661), and Key Laboratory of Network Oriented
Intelligent Computation, Shenzhen, China.
verified the connection of pulse signals with several certain
diseases [5-12].
During the development of computational pulse diagnosis, a
number of sensors and systems have been developed for
acquiring pulse signals. Sorvoja et al. [13] reported a pressure
pulse sensor based on electromechanical film. Kaniusas et al.
[14] used a magnetoelastic skin curvature sensor to design a
mechanical electrocardiography system for non-disturbing
measurement of blood pressure signals. Chen et al. [15]
presented a liquid sensor system that measures pulse signals.
Wu et al. [16] proposed an air pressure system that measures
pulse signals. Renevey et al. [17] proposed an infrared (IR)
pulse detection system. Wang et al. [18] proposed a multi-
channel pressure pulse signal acquisition system with a linear
sub sensor array. Hu et al. [19] proposed a pulse measurement
system based on a polyvinylidene fluoride (PVDF) pressure
sensor array. Zhang and Wang [20] proposed a photoelectric
system that measures pulse signals on fingers.
Among these systems, the three major types of sensors for
pulse signal acquisition are: pressure, photoelectric, and
ultrasonic sensors. The pressure sensor is adopted in pulse
diagnosis to imitate the TCM procedure of pulse taking [18],
the photoelectric sensor is mainly adopted because it is
inexpensive and easy to make[21] , and the ultrasonic sensor is
usually adopted for its robustness to interference [6]. As for
pulse diagnosis, pressure signals have been investigated for
pulse waveform classification and the diagnosis of cholecystitis,
nephrotic syndrome, and diabetes [18, 22-24]. Lee et al. found
that the photoplethysmogram (PPG) variability is related to
sympathetic vasomotor activity, and photoelectric signal (i.e.,
PPG) had been combined with routine cardiovascular
measurements (i.e., heart rate and mean arterial pressure) for
the diagnosis of low systemic vascular resistance (SVR) [25].
Finally, ultrasonic signals have been investigated for the
diagnosis of arteriosclerosis, pancreatitis, duodenal bulb ulcers,
cholecystitis, and nephritis [5, 9, 26].
In this paper, by conducting a comparative study, we analyze
the physical meanings, correlations, sensitivities to
physiological and pathological factors, and diagnosis
W. Zuo, P. Wang and D. Zhang are with the Computational Perception and
Cognition Centre, School of Computer Science and Technology, Harbin
Institute of Technology, Harbin 150001, China (e-mail:
dualblacklagrange@gmail.com, wmzuo@hit.edu.cn).
D. Zhang is also with the Biometrics Research Center, Department of
Computing, the Hong Kong Polytechnic University, Kowloon, Hong Kong (e-
mail: csdzhang@comp.polyu.edu.hk).
Comparison of Three Different Types of Wrist
Pulse Signals by Their Physical Meanings and
Diagnosis Performance
Wangmeng Zuo, Member, IEEE, Peng Wang, and David Zhang, Fellow, IEEE
P

2
performance of pulse signals acquired by these three types of
sensors. With these studies, we intend to reveal the relative
advantages of each type of pulse signal, which can guide us to
choose a proper sensor for the diagnosis of specific diseases and
to combine different types of pulse signals for improved
diagnosis accuracy.
The remainder of the paper is organized as follows. Section
2 is a discussion on the acquisition method and physical
meaning of the signals sampled by the three types of sensors.
Section 3 provides an analysis on the relationship between
different pulse signals and sensitivities of these pulse signals
with respect to different physiological and pathological factors.
Section 4 provides the experimental results that demonstrate the
relative advantages of different types of pulse signals and
improved performance by combining different sensors. Finally,
Section 5 gives several concluding remarks.
II. M
EASUREMENT
M
ECHANISM
In this section, we introduce the measurement mechanism of
the three major types of sensors to reveal the physical meaning
of the acquired pulse signals. As shown in Fig. 1, one typical
pulse acquisition hardware system usually involves three parts:
the sensor, amplifier and digitizer units, and the major
difference between these pulse systems is the sensor unit. Thus,
by analyzing the sensor units, we discuss the measurement
mechanism of different pulse signal acquisition systems.
A. Measurement Mechanism of Pressure Sensors
As shown in Fig. 2, pressure sensor is designed to measure
the transmural pressure at certain positions of the blood vessel.
Pulse waves are generated by the expulsion of blood with heart
contraction into the aorta, resulting in the dilatation of the vessel
[27]. Blood flow takes place in a closed system of vessels, and
any generated pressure affects the entire system. The wrist
radial artery is close to the skin surface and thus changes in
pressure can be non-invasively measured.
The measured pressure p
m
is composed of the counterforce
of the hold down pressure p
c
and transmural pressure from
blood vessel p, i.e., p
m
= p
c
+ p. Usually, p is smaller than its
true value due to the damping of the skin and tissue. Since the
radial artery is close to the surface of the skin, the damping
usually is slight and that is why a TCM practitioner chooses the
wrist as the position for pulse diagnosis.
B. Measurement Mechanism of Photoelectric Sensors
As shown in Fig. 3, photoelectric sensor is designed to
measure the blood volume at certain area of the blood vessel.
The intensity of the reflected light is in proportion with the
volume of the vessel. When the blood volume in the vessel
changes with the heartbeat, the reflected light will change
accordingly and thus the volume variation can be recorded by
measuring the intensity of the reflected light from the vessel.
Infrared light is usually employed in photoelectric sensor
because it can penetrate deeper into the vessel than visible light
while being absorbed/reflected less by epidermal melanin [28].
The measured volume signal V
m
is composed of the light
reflected from tissue V
0
and the reflected light from vessel V
s
,
i.e., V
m
=V
s
+V
0
, where V
0
is almost constant for the same person
and V
s
is time dependent and changes with the vessel volume.
As shown in Fig. 3, for photoelectric sensor the phototransistor
can only receive the reflected infrared light from a certain area.
The measured volume is the integral of the cross-section area
over the length l determined by the sensor size. Since l is
constant, the measured volume by photoelectric sensor would
depend on the change of cross-section area within the measure
area.
C. Measurement Mechanism of Ultrasonic Sensors
Ultrasonic sensor is employed to measure the blood velocity
at certain positions along the blood vessel. As shown in Fig. 4,
velocity information can be obtained by measuring the
frequency-shifting between the ultrasonic wave emitted by the
transmitter (T) and that returned to the receiver (R). The
ultrasonic signals reflect the velocities of red blood cells in the
Fig. 1. Pulse signal acquisition framework
Fig. 2. Measurement mechanism of a pressure sensor.
Fig. 3. Measurement mechanism of photoelectric sensor

3
vessel, where the relationship of the frequency-shift and
velocity can be formulated as:
0
0
()
2cos
r
cf f
uf


 (1)
where
0
f
is the emitted frequency,
r
f
is the reflected
frequency, is the flow velocity, and is the speed of sound
in soft tissue (about 1,540 m/s) [29]. Usually, the angle

should
be between 30º and 60º[27]. In this work, we assume that the
speed is the speed of blood at the center of the vessel because
this scenario can be applicable by locating the position with
maximum blood speed u
max
and this strategy is commonly used
in practice[29].
III. D
EPENDENCY AND
C
OMPLEMENTARITY OF
M
EASURED
S
IGNALS
In this section, we first analyze the dependence between the
different types of pulse signals. The analysis and discussion are
begun with some ideal simplifying assumptions to deduce the
basic facts of the relationship between signals acquired by
different types of sensors. Then, the diagnostic factors that
affect each type of measured signal will be discussed. Finally,
we will consider their complementarities.
A. Assumptions
The arterial system is a complex nonlinear anisotropic and
viscoelastic system, which comprises tapered, curved, and
branching tubes. In order to obtain some basic facts about
arterial pulse characteristics, we put forth some assumptions on
both the arterial system and the sensors to simplify the analysis.
First, we assume that the blood composition, blood density,
and the elasticity of the vessel wall are uniform and that the
flow in the sampling window is laminar as indicated by most
physiological fluid dynamic theories. We also assume that the
vessel is a straight cylindrical tube and has a circular cross-
section and the distortion of the vessel is minimal. Therefore,
for photoelectric signals we have:
V
m
=V
s
+V
0
=Al+V
0
=lπR
2
+V
0
(2)
where R is the radius of the vessel. The length is fixed and thus
photoelectric signals are in proportion with the square of the
radius under this assumption. Table I summarizes the physical
meaning of measured signals. For simplicity purposes, the p
c
in
the measured pressure signals and V
0
in the measured
photoelectric signals are not taken into consideration. Since
these signals were continuously recorded, both the values and
time derivatives of these signals can be easily obtained.
B. Relationship among Blood Velocity, Radius, and Pressure
in Steady Laminar Flow
In this section, we analyze the relationship of the blood
velocity, area (radius) and pressure in steady laminar and
pulsatile flows to reveal their relationship.
We first provide a cylindrical coordinate system (r, θ, z)
which will be used in this section. As shown in Fig. 5, the
coordinate is set along the vessel, the z-axis is in the center of
the vessel and toward the direction of the blood flow, and the r-
axis is perpendicular to the skin.
In order to determine the basic arterial pulse characteristics,
we first consider the simplest model: steady laminar flow in
which the radius R, density ρ, viscosity η, velocity u, volume V
and pressure p, are all constant. The pressure gradient is also
uniform. The general time-dependent governing equations can
be given by the continuity and the Navier-Stokes equations. For
steady laminar flow, the solution is [30, 31]:
22
4
Rrp
uz



(3)
The pressure gradient along the z-axis is hard to measure with
only one sensor, considering that the pulse wave is a wave that
is travelling without distortion with velocity c. Then the
pressure will have the form of [32]:
0
()
z
pp ftc
 
(4)
By differentiating Eqn. (4) with respect to t and z, we get:
1pp
zct



(5)
uc
u
Fig. 5. Cylindrical coordinates and flow in the vessel
Fig. 4. Measurement mechanism of ultrasonic sensor
TABLE I
PHYSICAL MEANING OF THE MEASURED SIGNALS
Sensor Physical meaning
Pressure p
Photoelectric A ( or R
2
)
Ultrasonic u
max

4
Therefore, a good approximation to the pressure gradient along
the z-axis is the time derivative of the measured pressure signal.
The measured ultrasonic signals are the velocity at the center
of the vessel and thus r=0. By inserting r=0 and Eqn. (5) into
Eqn. (3), we obtain:
2
max 1
4
p
uR
ct


  (6)
One can see that the measured velocity umax, radius R, and
time derivative of the measured pressure p are highly correlated
with each other. Under the simplifying assumptions described
in Section III.A, if we have parameters

and c, the continuous
pressure and photoelectric signals respectively, we can get the
velocity by using Eqn. (6). If we have three types of measured
signals, we can estimate the parameter

c.
The discussion above is based on several strong assumptions
in order to determine the basic relationship between the three
types of measured signals. If we consider more realistic
conditions, the relationship will become more complicated. For
example, if we let the pressure gradient be in a pulsatile form:
it
pae
z


 (7)
the velocity can be estimated by the model given by Womersley
[32],
3
2
0
3
2
0
(i)
1
(i)
it Jr
ae
uiJR



 









(8)
where J0(xi3/2) is a Bessel function of the order zero with a
complex argument. If we insert r=0 and Eqn. (5) into Eqns. (7)
and (8), we can obtain a more complicated model on the
measured blood velocity, volume, and pressure. If we let the
radius R be a time-dependent variable, the model would be
more difficult to obtain. Moreover, for real pulse signals, the
parameters

and

are also time-dependent and all of these
parameters would vary with individual, health conditions, and
many other factors. Thus, although the three types of signals are
closely related, it is difficult to obtain an explicit model that
uses two signals to estimate the other while using multiple types
of signals to estimate some of the circulatory parameters that
are still available which we will discuss in the next section.
C. Influence of Physiological and Pathological Factors
Different types of sensors have different physical meanings
and thus would be influenced by different circulatory
parameters. In this section, we analyze the influence of
physiological and pathological factors on the different types of
pulse signals.
Pressure signals are associated with the elasticity of the
vessel wall, and the radius. From the definition of incremental
elastic modulus:
inc
inc
inc
Rp R
Eht Rt





(9)
we can get
2
p
Eh R
tRt



(10)
where is the incremental elastic modulus, is the
incremental stress, is the incremental strain, R is the mean
radius of the blood vessel, and h is the thickness of the blood
wall.
From these equations, one can see that pressure signals are
sensitive to changes in the radius, the elastic property and the
thickness of the blood vessel.
Photoelectric signals can be used to measure the volume
changes of the blood in the vessel and are primarily sensitive to
radius changes. Moreover, physically the photoelectric signals
are measurements of the intensity of reflected infrared light, and
influenced by blood composition. The infrared absorption
spectra of blood elements are also different, i.e., water,
oxyhemoglobin and deoxyhemoglobin exhibit different
absorption spectra, and thus the composition ratio may
influence the infrared absorption rate [33]. Actually, the blood
oxygen monitor is designed by using this principle to measure
blood oxygen saturation.
Ultrasonic signals represent the velocity of the blood flow.
From Eqns. (6) and (8), one can see that the ultrasonic signals
are associated with the pressure gradient, blood density and
viscosity. The velocity also reflects the flow statement. For
example, the Reynolds number [30] :
2uR
Re


 (11)
is a measure of the tendency for turbulence to occur. The
viscosity of blood is normally about 1/30 poise, and the density
is only slightly greater than 1. When the Reynolds number rises
above 200 to 400, turbulent flow will occur in some branches
of vessels; when the Reynolds number rises above
approximately 2000, turbulence will usually occur even in
straight and smooth vessels [27].
Table II summarizes the sensitiveness of different types of
signals to changes in physiological parameters. Pressure signals
mainly represent the transmural pressure and are sensitive to the
parameter changes of the vessel wall, such as the wall elastic
modulus and its thickness. Photoelectric signals represent the
volume information of the vessel and are sensitive to the area
of the cross section. Moreover, volume information is measured
by the intensity of the reflected infrared light and thus
photoelectric signals are also sensitive to blood composition.
Ultrasonic signals represent the blood velocity and are sensitive
to the parameters associated with blood flow, such as viscosity
and blood flow state. All three signals are sensitive to the
changes of the vessel radius. As discussed above, the complex
inc
Einc

inc

TABLE II
INFLUENCE FROM CIRCULATORY PARAMETERS
Pressure Photoelectric Ultrasonic
Radius of vessel ○ ○ ○
Wall elastic property ○
Wall thickness ○
Blood composition ○
Blood flow status ○
Blood viscosity ○

5
nonlinear anisotropic and viscoelastic properties of the arterial
system and the relationship between different parameters are
mostly nonlinear. Thus, the analysis results in Table II are
obtained based on some common assumptions used in most
physiological fluid dynamic theories. Except for these basic
circulatory parameters, some other important diagnosis
parameters are also associated with pressure, volume and
velocity, such as the blood flow, and vascular compliance and
resistance. The diagnostic validity of these parameters reveals
the complementarity of the different measured signals.
Blood flow is the quantity of blood that passes through a
given point in circulation in a given period. Blood flow can
be calculated by . Since photoelectric signals are in
proportion to area , blood flow is related to both ultrasonic
and photoelectric signals.
Vascular compliance is of particular significance in
cardiovascular physiology and has been reported to be sensitive
to hypertension, congestive heart failure and aging [34, 35].
Vascular compliance is defined as
V
Cp

 (12)
which is a measure of the tendency of the arteries to stretch in
response to pressure. Blood vessels with a higher compliance
deform easier than those of lower compliance under the same
pressure and volume conditions. From its definition, ∆V has a
fixed length, and we can replace ∆V with ∆A and Eqn. (18)
becomes:
A
A
pAp
Cc
tz tt
p
 
 
 
 (13)
From Eqns. (12) and (13), one can see that vascular compliance
is related to photoelectric and pressure signals, and can be
calculated by using their time derivatives.
Vascular resistance is defined as:
1pp
Res QuAt

 (14)
Several documented conditions are associated with low
vascular resistance, such as sepsis, pancreatitis, cirrhosis, and
so on [36]. From the definitions, we can see that the resistance
parameter is related to all three types of signals.
D. Summary
Signals measured by three popular sensors are closely related,
but have different physical meanings and are sensitive to
different circulatory parameters. By combining different types
of signals, some useful diagnostic parameters, such as blood
flow, and vascular compliance and resistance can be inferred,
which demonstrate the complementarity of different signals.
Since they have different sensitivity parameters to diseases that
may be associated with a certain parameter, the use of a sensor
which is sensitive to that certain parameter may achieve better
diagnosis performance.
From Table II, one can see that all type of signals are
sensitive to the radius of the vessel and thus a disease that is
associated with radius change may be detected by these sensors.
As pressure signals are more sensitive to changes in the elastic
property and the thickness of the vessel wall, they should be
more effective than signals from the other types of sensors in
the diagnosis of the related disease (e.g., arteriosclerosis). With
a similar rationale, photoelectric sensors should be more
effective than the other sensors in the diagnosis of diseases
related to the area of the vessel cross section and blood
composition, and ultrasonic sensors should be more effective
than the other sensors in the diagnosis of the blood flow-related
disease, e.g., diabetes which has been reported to be related to
viscosity. Moreover, since the different signals are
complementary, the combination of these signals may further
reveal other diagnostic parameters, like vascular compliance
and resistance which are related to many kinds of diseases. Thus,
the combination of all of these signals may further increase
diagnosis performance.
IV. CASE STUDIES
In the case studies, the pulse signals are acquired by using
the pressure and photoelectric systems designed by our lab [18],
and the CBS 2000 ultrasonic system provided by EDAN
Instruments, Inc. Fig. 6 shows the pulse signals sampled from a
healthy volunteer using different systems, where Fig. 6(a)
shows the pressure signals, Fig. 6(b) the photoelectric signals,
and Fig. 6(c), the ultrasonic signals. From these figures one can
see that different types of pulse signals are similar in terms of
waveform, which demonstrated that these signals are dependent.
However their difference are also obvious this may because
their different physical attributes and influential factors. Thus,
it is natural to suppose that the diagnosis of different disease
using the three types of pulse signals may also have different
performance.
Diabetes and arteriosclerosis diagnoses are used as two
examples in the experiments because diabetes is reported to be
associated with blood viscosity [37] and arteriosclerosis refers
to the thickening and hardening of the arteries [38]. Using these
two diseases, experiments were conducted to compare the
diagnosis performance of the three types of pulse sensors and
to validate the effectiveness of combining three types of
measured signals.
A. Method
In this subsection we introduced the preprocessing, feature
extraction, and classification methods. In pre-processing, the
high frequency noise and baseline drift coupled with pulse
signal was removed. These two interferences are mainly
introduced by the power line interference and breathing,
respectively. The wavelet denoising method was adopted to
remove the noise and the wavelet-based cascaded adaptive filter
[39] was adopted to correct the baseline drift.
In feature extraction, three kinds of features were extracted
to characterize the pulse signal. First, the time domain fiducial-
point based feature was extracted. Fiducial-point based feature
describes the shape of an average pulse cycle, it includes the
position of the primary peak, dicrotic notch and secondary peak
which is the most common feature used in pulse signal
classification [7, 40-44]. Second, the multi-scale sample
entropy of the pulse signal was calculated. Sample entropy can
be used to measure the unpredictability of pulse signal and
multi-scale sample entropy can reveal unpredictability with
long-range correlations on multiple spatial and temporal scales
Q
QuA
A

6
which had been successful applied for pulse signal
classification [45-47]. The last feature was the TWED feature.
TWED is an elastic metric to measure the distance between time
series, which is reported effective and efficient in pulse signal
classification [7, 48]
In the classification we adopted the composite kernel
learning (CKL) method to combine these features since it is
more flexible than SVM in combining features from different
sources [49]. For each type of pulse signal, we extract three
kinds of features, i.e., time domain fiducial point-based feature,
multi-scale sample entropy (SampEn), and TWED feature.
Thus, nine basis kernels K1,1~K3,3 are built that correspond to
three type of signals and three kinds of features, where K1,m,
K2,m, K3,m are the basis kernels for pressure, photoelectric, and
ultrasonic signals, respectively, Kl,1, Kl,2, and Kl,3 are the basis
kernels for time domain fiducial point-based feature, SampEn,
and TWED feature, respectively.
To combine the basis kernels, we adopt the composite kernel
learning (CKL) model [7] which defines a tree structure to
guide the selection and removal of kernels. In pulse
classification, as shown in Fig. 7, the basis kernels are
structured into three groups based on the types of sensors, i.e.,
G1 = {K1,1, K1,2, K1,3}, G2 = {K2,1, K2,2, K2,3}, and G3 = {K3,1,
K3,2, K3,3}. In CKL [7], the combination of the kernels is defined
by: (,) (, )
llmlm
lm
KKx
y
x
y

 (15)
where

lm and

l are the non-negative coefficients for kernel Klm
and group Gl, respectively. For kernel selection/removal, the
following constraints are imposed on

lm and

l,
2/ 1, 0
p
ll l
ld


 (16)
2/ 1, 0
q
lm lm
lm


 (17)
where p and q are two hyper-parameters for tuning sparsity
within or between groups, and dl is the number of basis kernels
in group Gl. Let
p
q
lm l lm


 and () ( ,)
lm i lm i
i
fKxxx

,
the CKL model [7] is formulated as,


1( )
1/
min ( ) s.t. 1, 0
pq
q
pq
llm lm
lm
Jd
γγ


 (18)
with

2
,,
11
() min
2
s.t. ( ) 1 , 0
lm
lm
lm i
lm i
fb lm
ilmi ii
lm
JfC
yf b
ξ
γ
x




 




(19)
As introduced in [7], the model above can be efficiently
solved by iterating between (i) updating flm with the standard
SVM solver, and (ii) updating  with the fixed point algorithm.
In this work we set p = q = 0.5 and set C = 100.
The 10-fold cross-validation method is used to evaluate the
classification performance. Three performance indicators, i.e.,
classification accuracy, sensitivity, and specificity, are adopted
for quantitative evaluation. The accuracy is defined as the
percentage of all the correctly identified samples, the sensitivity
is defined as the percentage of patients who are correctly
identified as sick and specificity was defined as the percentage
of healthy people who are correctly identified as healthy [50].
B. Diabetes Experiment
In the diabetes experiment we constructed a dataset of 392
volunteers, including 191 healthy and 201 diabetic volunteers
by collaborating with the Yao Chung Kit Diabetes Assessment
Centre in Hong Kong. Each volunteer provided three different
types of sample data for the dataset. To avoid the potential
influence of biological factors, we also ensured that the
distributions of gender and age of volunteers with diabetes are
similar to those of the healthy volunteers. Table III lists the
basic information of the dataset.
Fig. 6. Three types of pulse signals from a healthy volunteer: (a) pressure
signals, (b) photoelectric signals, and (c) ultrasonic signals.
01000 2000 3000 4000 5000 6000 7000 8000
Time(ms)
Ultrasonic Signal
01000 2000 3000 4000 5000 6000 7000 8000
Time(ms)
Pressure Signal
01000 2000 3000 4000 5000 6000 7000 8000
Time(ms)
Photoelectric Signal
TABLE III
SUMMARY OF DIABETES DATASETS
Age Distribution Gender
Distribution
1 ~ 40 40 ~ 50 50 ~ 60 > 60 Male Female
Healthy 9 41 69 72 119 72
Diabetes 3 35 67 96 131 70
Fig. 7. A tree that depicts groups of kernels for pulse signal classification

7
In the experiment, we used the preprocessing, feature
extraction and classification methods described in Section IV.A.
In the classification of signals from single source, we have only
one group in CKL model and in the classification of pulse
signals from multiple sources we have three groups in CKL
model as shown in Fig. 7. The classification results of the
signals from the different types of sensors are listed in Table IV.
One can see that the ultrasonic pulse signal is more effective in
the diagnosis of diabetes than the other types of pulse signals.
The result is consistent with a previous study in which diabetes
is reported to be associated with viscosity [37]. Note that the
ultrasonic sensor is more sensitive to the changes in the
viscosity. The results also show that the combination of the
three types of signals can obtain improved classification
accuracy, which indicates that the three types of signals contain
some complementary information for the diagnosis of diabetes.
The result of the McNemar’s test [51] shows that in the diabetes
classification, the performance difference between ultrasonic
and other types of signals are statistically significant at

= 0.05.
The performance difference between the group of combined
signals and any single signal group are also statistically
significant at

= 0.05.
C. Arteriosclerosis Experiment
In the arteriosclerosis experiment, we constructed a dataset
of 184 volunteers, including 95 healthy and 89 arteriosclerosis
volunteers by collaborating with the Guangzhou Hospital of
TCM. The principle of the constructing the dataset and the
format of the samples are the same as those of the diabetes
dataset. Table V lists the basic information of the
arteriosclerosis data set.
The results are listed in Table VI. From Table VI, it can be
observed that the pressure pulse signal outperforms the other
types of pulse signal in the arteriosclerosis experiment. This
may because it is more sensitive to the changes in vessel
hardness and thickness which are related to arteriosclerosis [38].
The ultrasonic group also obtained good performance because
the hardness of the vessel will also influence the blood speed.
Moreover, in the arteriosclerosis experiment, the classification
performance can also be improved by combining all three types
of signals. The result of the McNemar’s test [51] shows that in
the arteriosclerosis classification, the performance difference
between the pressure and photoelectric signals is statistically
significant at

= 0.1. The performance difference between the
group of combined signals and any single signal group are also
statistically significant at

= 0.1. The higher

value can be
partially explained by that the dataset size in the arteriosclerosis
experiment is lower than that in the diabetes experiments.
V. CONCLUSION
In this paper, we study the dependency and complementarity
among three major types of sensors, i.e., pressure, photoelectric,
and ultrasonic sensors, for pulse signal diagnosis. Our analysis
on their physical meanings, relationships and sensitivity factors
shows that: (i) the changes in the elastic property and thickness
of the vessel wall can be more readily detected using pressure
signals; (ii) the changes in the area of the cross section and
blood composition can be more readily captured using
photoelectric signals; (iii) the changes in the blood viscosity and
the blood flow state can be more effectively characterized using
ultrasonic signals. Thus, we state that each sensor is more
appropriate for the diagnosis of some specific disease that the
changes of physiological factors can be effectively reflected by
the sensor, and different types of signals are complementary.
Case studies have been conducted to validate these
statements. The experimental results show that, in terms of
accuracy, sensitivity and specificity, the ultrasonic sensor is
superior to the others for the diagnosis of diabetes while the
pressure sensor outperforms the others for the diagnosis of
arteriosclerosis. These results can be explained by that the onset
of diabetes is usually accompanied by the changes in blood flow
viscosity [37] which can be well depicted by ultrasonic signals,
arteriosclerosis usually results in the changes in the hardness
and thickness of wrist radial artery [38] which can be well
characterized by pressure signals. Moreover, the combination
of the three types of signals further improves the diagnosis
performance, which can be explained by the complementarity
among sensors.
ACKNOWLEDGMENT
The authors would like to thank the associate editor and the
anonymous reviewers for their constructive suggestions.
REFERENCES
[1] S. Walsh and E. King, Pulse Diagnosis: A Clinical Guide. Sydney
Australia: Elsevier, 2008.
[2] V. D. Lad, Secrets of the Pulse. Albuquerque, New Mexico: The
Ayurvedic Press, 1996.
[3] E. Hsu, Pulse Diagnosis in Early Chinese Medicine. New York,
American: Cambridge University Press, 2010.
[4] R. Amber and B. Brooke, Pulse Diagnosis: Detailed Interpretations
For Eastern & Western Holistic Treatments. Santa Fe, New Mexico:
Aurora Press, 1993.
TABLE IV
DIABETICS DIAGNOSIS PERFORMANCE
Accuracy Sensitivity Specificity
Pressure 84.2% 86.1% 82.2%
Photoelectric 79.8% 82.6% 77.0%
Ultrasonic 87.0% 88.6% 85.3%
Combination of three types of signals 91.6% 92.5% 90.6%
TABLE V
SUMMARY OF ARTERIOSCLEROSIS DATASETS
Age Distribution Gender Distribution
1 ~ 40 40 ~ 50 50 ~ 60 > 60 Male Female
Healthy 19 22 23 31 60 35
Arteriosclerosis 15 18 24 32 59 30
TABLE VI
ARTERIOSCLEROSIS DIAGNOSIS PERFORMANCE
Accuracy Sensitivity Specificity
Pressure 86.4% 87.6% 85.2%
Photoelectric 79.3% 83.1% 75.8%
Ultrasonic 83.7% 85.4% 82.1%
Combination of three types of signals 89.7% 91.0% 88.4%

8
[5] Y. Chen, L. Zhang, D. Zhang, and D. Zhang, "Wrist pulse signal
diagnosis using modified Gaussian Models and Fuzzy C-Means
classification," Medical Engineering & Physics, vol. 31, pp. 1283-
1289, Dec 2009.
[6] Y. Chen, L. Zhang, D. Zhang, and D. Zhang, "Computerized wrist
pulse signal diagnosis using modified auto-regressive models,"
Journal of Medical Systems, vol. 35, pp. 321-328, Jun 2011.
[7] L. Liu, W. Zuo, D. Zhang, N. Li, and H. Zhang, "Combination of
heterogeneous features for wrist pulse blood flow signal diagnosis
via multiple kernel learning," IEEE Transactions on Information
Technology in Biomedicine, vol. 16, pp. 599-607, Jul 2012.
[8] L. Liu, W. Zuo, D. Zhang, N. Li, and H. Zhang, "Classification of
wrist pulse blood flow signal using time warp edit distance,"
Medical Biometrics, vol. 6165, pp. 137-144, 2010.
[9] D. Y. Zhang, W. M. Zuo, D. Zhang, H. Z. Zhang, and N. M. Li,
"Wrist blood flow signal-based computerized pulse diagnosis using
spatial and spectrum features," Journal of Biomedical Science and
Engineering, vol. 3, pp. 361-366, 2010.
[10] Q. L. Guo, K. Q. Wang, D. Y. Zhang, and N. M. Li, "A wavelet
packet based pulse waveform analysis for cholecystitis and
nephrotic syndrome diagnosis," in IEEE International Conference
on Wavelet Analysis and Pattern Recognition, Hong Kong, China,
2008, pp. 513-517.
[11] S. Charbonnier, S. Galichet, G. Mauris, and J. P. Siche, "Statistical
and fuzzy models of ambulatory systolic blood pressure for
hypertension diagnosis," IEEE Transactions on Instrumentation
and Measurement, vol. 49, pp. 998-1003, 2000.
[12] H.-T. Wu, C.-H. Lee, C.-K. Sun, J.-T. Hsu, R.-M. Huang, and C.-J.
Tang, "Arterial Waveforms Measured at the Wrist as Indicators of
Diabetic Endothelial Dysfunction in the Elderly," IEEE
Transactions on Instrumentation and Measurement, vol. 61, pp.
162-169, 2012.
[13] H. Sorvoja, V. M. Kokko, R. Myllyla, and J. Miettinen, "Use of
EMFi as a blood pressure pulse transducer," IEEE Transactions on
Instrumentation and Measurement, vol. 54, pp. 2505-2512, 2005.
[14] E. Kaniusas, H. Pfutzner, L. Mehnen, J. Kosel, C. Tellez-Blanco, G.
Varoneckas, et al., "Method for continuous nondisturbing
monitoring of blood pressure by magnetoelastic skin curvature
sensor and ECG," IEEE Sensors Journal, vol. 6, pp. 819-828, Jun
2006.
[15] C. Lianyi, H. Atsumi, M. Yagihashi, F. Mizuno, H. Narita, and H.
Fujimoto, "A preliminary research on analysis of pulse diagnosis,"
in IEEE International Conference on Complex Medical Engineering,
Beijing,China, 2007, pp. 1807-1812.
[16] H.-T. Wu, C.-H. Lee, and A.-B. Liu, "Assessment of endothelial
function using arterial pressure signals," Journal of Signal
Processing Systems, vol. 64, pp. 223-232, 2011.
[17] P. Renevey, R. Vetter, J. Krauss, P. Celka, and Y. Depeursinge,
"Wrist-located pulse detection using IR signals, activity and
nonlinear artifact cancellation," in Annual International Conference
of the IEEE Engineering in Medicine and Biology Society, 2001, pp.
3030-3033 vol.3.
[18] P. Wang, W. Zuo, H. Zhang, and D. Zhang, "Design and
implementation of a multi-channel pulse signal acquisition system,"
in IEEE International Conference on Biomedical Engineering and
Informatics, ChongQing, China, 2012, pp. 1063-1067.
[19] C.-S. Hu, Y.-F. Chung, C.-C. Yeh, and C.-H. Luo, "Temporal and
Spatial Properties of Arterial Pulsation Measurement Using
Pressure Sensor Array," Evidence-Based Complementary and
Alternative Medicine, vol. 2012, pp. 1-9, 2012.
[20] A. Zhang and H. Wang, "Real-time detection system For
photoelectric pulse signals," in IEEE International Conference on
Business Management and Electronic Information
Guangzhou,China, 2011, pp. 498-501.
[21] N. Selvaraj, K. H. Shelley, D. G. Silverman, N. Stachenfeld, N.
Galante, J. P. Florian, et al., "A Novel Approach Using Time
Frequency Analysis of Pulse-Oximeter Data to Detect Progressive
Hypovolemia in Spontaneously Breathing Healthy Subjects," IEEE
Transactions on Biomedical Engineering, vol. 58, pp. 2272-2279,
2011.
[22] Q. Guo, K. Wang, D. Zhang, and N. Li, "A wavelet packet based
pulse waveform analysis for cholecystitis and nephrotic syndrome
diagnosis," in IEEE International Conference on Wavelet Analysis
and Pattern Recognition, Hong Kong, China, 2008, pp. 513-517.
[23] I. Wakabayashi and H. Masuda, "Association of pulse pressure with
fibrinolysis in patients with type 2 diabetes," Thrombosis Research,
vol. 121, pp. 95-102, 2007.
[24] N. Arunkumar and K. M. M. Sirajudeen, "Approximate entropy
based ayurvedic pulse diagnosis for diabetics - a case study," in
IEEE International Conference on Trendz in Information Sciences
and Computing, Chennai,India, 2011, pp. 133-135.
[25] Q. Y. Lee, G. S. H. Chan, S. J. Redmond, P. M. Middleton, E. Steel,
P. Malouf, et al., "Classification of low systemic vascular resistance
using photoplethysmogram and routine cardiovascular
measurements," in Annual International Conference of IEEE
Engineering in Medicine and Biology Society, Buenos Aires,
Argentina, 2010, pp. 1930-1933.
[26] R. Murata, H. Kanai, N. Chubachi, and Y. Koiwa, "Measurement of
local pulse wave velocity on aorta for noninvasive diagnosis of
arteriosclerosis," in IEEE Annual International Conference of the
Engineering in Medicine and Biology Society, Baltimore, MD, 1994,
pp. 83-84 vol.1.
[27] A. C. Guyton and J. E. Hall, Textbook of medical physiology.
Philadelphia, Pennsylvania: Elsevier, 2006.
[28] ICNIRP, "ICNIRP Statement on Far Infrared Radiation Exposure,"
Health Physics, vol. 91, pp. 630-645, 2006.
[29] W. Schèaberle, Ultrasonography in Vascular Diagnosis: A
Therapy-oriented Textbook and Atlas. Germany: Springer, 2005.
[30] G. E. Mase, Schaum's outline of theory and problems of continuum
mechanics: McGraw-Hill New York, 1970.
[31] M. Kutz, Biomedical Engineering and Design Handbook:
Biomedical Engineering Fundamentals. United States: McGraw-
Hill Professional, 2009.
[32] J. R. Womersley, "Method for the calculation of velocity, rate of
flow and viscous drag in arteries when the pressure gradient is
known," The Journal of physiology, vol. 127, pp. 553-563, 1955.
[33] L. V. Wang and H.-i. Wu, Biomedical optics: principles and
imaging. New Jersey: Wiley & Sons, 2012.
[34] S. M. Finkelstein, W. Feske, J. Mock, P. Carlyle, T. Rector, S. Kubo,
et al., "Vascular compliance in hypertension," in Annual
International Conference of the IEEE Engineering in Medicine and
Biology Society, New Orleans, LA, USA, 1988, pp. 241-242 vol.1.
[35] S. M. Finkelstein, G. E. McVeigh, D. E. Burns, P. F. Carlyle, and J.
N. Cohn, "Arterial Vascular Compliance In Heart Failure," in 12th
Annual International Conference of the IEEE Engineering in
Medicine and Biology Society, Michigan,American, 1990, pp. 548-
549.
[36] J. Melo and J. I. Peters, "Low systemic vascular resistance:
differential diagnosis and outcome," Critical Care, vol. 3, pp. pp.71-
77, 1999.
[37] D. A. Fedosov, W. Pan, B. Caswell, G. Gompper, and G. E.
Karniadakis, "Predicting human blood viscosity in silico,"
Proceedings of the National Academy of Sciences, vol. 108, pp.
11772-11777, 2011.
[38] W. A. N. Dorland, Illustrated medical dictionary: WB Saunders
Company, 2011.
[39] L. Xu, D. Zhang, and K. Wang, "Wavelet-based cascaded adaptive
filter for removing baseline drift in pulse waveforms," IEEE
Transactions on Biomedical Engineering, vol. 52, pp. 1973-1975,
Nov 2005.
[40] L. Xu, M. Q. H. Meng, R. Liu, and K. Wang, "Robust peak detection
of pulse waveform using height ratio," in International Conference
of the IEEE Engineering in Medicine and Biology Society,
Vancouver, BC, Canada, 2008, pp. 3856-3859.
[41] D. Zhang, W. Zuo, D. Zhang, H. Zhang, and N. Li, "Wrist blood
flow signal-based computerized pulse diagnosis using spatial and
spectrum features," Journal of Biomedical Science and Engineering,
vol. 3, pp. 361-366, 2010.
[42] C. Xia, Y. Li, J. Yan, Y. Wang, H. Yan, R. Guo, et al., "Wrist Pulse
Waveform Feature Extraction and Dimension Reduction with
Feature Variability Analysis," in International Conference on
Bioinformatics and Biomedical Engineering, Shanghai, China, 2008,
pp. 2048-2051.
[43] L. Xu, M. Q. H. Meng, K. Wang, W. Lu, and N. Li, "Pulse images
recognition using fuzzy neural network," Expert systems with
applications, vol. 36, pp. 3805-3811, 2009.
[44] Y. Wang, X. Wu, B. Liu, Y. Yi, and W. Wang, "Definition and
application of indices in Doppler ultrasound sonogram," Shanghai
Journal of Biomedical Engineering, vol. 18, pp. 26-29, Aug 1997.

9
[45] M. Costa, A. L. Goldberger, and C. K. Peng, "Multiscale entropy
analysis of biological signals," Physical Review E, vol. 71, pp. 1-18,
Feb 2005.
[46] M. Costa, A. L. Goldberger, and C. K. Peng, "Multiscale entropy
analysis of complex physiologic time series," Physical Review
Letters, vol. 89, pp. 1-4, Aug 5 2002.
[47] L. Liu, N. Li, W. Zuo, D. Zhang, and H. Zhang, "Multiscale sample
entropy analysis of wrist pulse blood flow signal for disease
diagnosis," in Sino-foreign-interchange Workshop on Intelligence
Science and Intelligent Data Engineering, NanJing China, 2012, pp.
pp.475-482.
[48] L. Liu, W. Zuo, D. Zhang, N. Li, and H. Zhang, "Classification of
Wrist Pulse Blood Flow Signal Using Time Warp Edit Distance,"
Medical Biometrics. Springer Berlin Heidelberg, pp. pp. 137-144,
2010.
[49] M. Szafranski, Y. Grandvalet, and A. Rakotomamonjy, "Composite
kernel learning," Machine learning, vol. 79, pp. 73-103, 2010.
[50] A. G. Lalkhen and A. McCluskey, "Clinical tests: sensitivity and
specificity," Continuing Education in Anaesthesia, Critical Care &
Pain, vol. 8, pp. 221-223, 2008.
[51] Q. McNemar, "Note on the sampling error of the difference between
correlated proportions or percentages," Psychometrika, vol. 12, pp.
153-157, June 1947.
David Zhang (F’09) received the Graduate
degree in computer science from Peking University, Beijing,
China. He received the M.Sc. degree in computer science and
the Ph.D. degree from the Harbin Institute of Technology (HIT),
Harbin, China, in 1982 and 1985, respectively, and also the
Ph.D. degree in electrical and computer engineering from the
University of Waterloo, Ontario, Canada, in 1994. From 1986
to 1988, he was a Postdoctoral Fellow at Tsinghua University,
Beijing, and then an Associate Professor at the Academia Sinica,
Beijing. He is currently a Head in the Department of Computing,
and a Chair Professor at the Hong Kong Polytechnic University,
Kowloon, Hong Kong, where he is the Founding Director of the
Biometrics Technology Centre (UGC/CRC) supported by the
Hong Kong SAR Government in 1998. He was also a Visiting
Chair Professor in Tsinghua University, and an Adjunct
Professor in Peking University, Shanghai Jiao Tong University,
HIT, and the University of Waterloo. He is the author of more
than ten books and 200 journal papers. He is a Senior Research
Fellow of Croucher.
Dr. Zhang is the Founder and Editor-in-Chief of International
Journal of Image and Graphics, Book Editor of Springer
International Series on Biometrics, an Organizer of the
International Conference on Biometrics Authentication, and an
Associate Editor of more than ten international journals,
including several IEEE Transactions and Pattern Recognition.
He is a Distinguished Speaker of the IEEE Computer Society
and a Fellow of the International Association of Pattern
Recognition.
Peng Wang received the B.S. degree in
information and computing science from Harbin Normal
University, Harbin, China in 2007. He received M.S. degree in
computer science from the Harbin Institute of Technology
(HIT), Harbin, China, in 2009. He is currently working toward
the Ph.D. degree in School of Computer Science and
Technology, Harbin Institute of Technology. From September
2009 to December 2009, he was a Research Assistant in the
Department of Computing, Hong Kong Polytechnic University.
His research interests include computerized medical diagnosis,
pattern recognition, machine learning, and biometrics.
Wangmeng Zuo (M’09) received the Ph.D.
degree in Computer Application Technology from the Harbin
Institute of Technology in 2007. From July 2004 to December
2004, from November 2005 to August 2006, and from July
2007 to February 2008, he was a Research Assistant at the
Department of Computing, Hong Kong Polytechnic University,
Hong Kong. From August 2009 to February 2010, he was a
Visiting Professor in Microsoft Research Asia. His current
research interests include discriminative learning, image
modeling, low level vision, biometrics, medial image and signal
analysis,. Dr. Zuo has published more than 40 papers in top tier
academic journals and conferences including IEEE T-IP, T-
NNLS, T-IFS, Pattern Recognition, CVPR, ICCV, ICML,
ECCV, and NIPS. Dr. Zuo is an Associate Editor of the IET
Biometrics and the Scientific World Journal, and the Guest
Editor of Neurocomputing.