2020 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 64, NO.
7, JULY 2015
A Novel Multichannel Wrist Pulse System
With Different Sensor Arrays
Dimin Wang, David Zhang, Fellow, IEEE, and Guangming Lu
Abstract— Wrist pulse has long been recognized as an impor-
tant physiological symbol for judging the health status of patients.
It is regarded as one of the four diagnosis techniques in tradi-
tional Chinese medicine and plays an important role in disease
analysis. However, the pulse diagnosis skill is subjective and needs
decades of experience to master. To make the pulse diagnosis
more objective and easily understood by practitioners, many
pulse-taking platforms were invented, but several disadvantages,
including the simplicity and insufficiency, limit the applications.
This paper proposes a novel wrist pulse system for the pulse diag-
nosis to overcome the shortcomings in previous pulse platforms.
We integrated a pressure sensor with a photoelectric sensor to
make a fusion sensor to feel pulse in different approaches and
designed the multichannel sensor arrays structure to acquire
more pulse information. Finally, pulse analysis algorithm and
classification methods are introduced. Experiments on disease
classification are carried out to test the system performance with
multichannel and different sensor arrays. The results show that
the novel system is not only able to distinguish between healthy
pulse samples and subjects suffering from diabetes, but also good
at obtaining more information than the conventional pulse system
with single channel or simplex-type sensor.
Index Terms— Different sensor arrays, multichannel array
design, pulse waveform analysis, wrist pulse system.
I. I NTRODUCTION
P ULSE diagnosis is known as one of the four examination
methods in traditional Chinese medicine (TCM) diag-
nosis which consists of observation, listening, smelling, and
pulse feeling. In ancient times, Chinese medicine physicians
diagnosed the pathological changes of organs by feeling the
artery pulse [1], [2]. The pulse is transmitted by blood flow
through arteries from the heart. Hence, it is affected not
only by the conditions of the heart beatings, but also by
the conditions of nerves, organs, muscles, skin, arterial walls,
Manuscript received June 21, 2014; revised August 23, 2014; accepted
August 25, 2014. Date of publication March 2, 2015; date of current version
June 5, 2015. This work was supported by the General Research Fund
through the Hong Kong, in part by the Central Fund, Hong Kong Polytechnic
University, Hong Kong, in part by the National Natural Science Foundation
of China under Grant 61332011, Grant 61020106004, Grant 61272292, and
Grant 61271344, in part by the Shenzhen Fundamental Research Fund under
Grant JCYJ20130401152508661, and in part by the Key Laboratory of
Network Oriented Intelligent Computation, Shenzhen, China. The Associate
Editor coordinating the review process was Dr. Domenico Grimaldi.
D. Wang is with the Department of Electronic Engineering, Tsinghua
University, Beijing 100084, China.
D. Zhang is with the Department of Computing, Hong Kong Polytechnic
University, Hong Kong (e-mail: csdzhang@comp.polyu.edu.hk).
G. Lu is with the Biocomputing Research Center, Shenzhen Graduate
School, Harbin Institute of Technology, Harbin 150001, China.
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TIM.2014.2357599
0018-9456 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
blood parameters, etc. [3]. According to this principle, wrist
pulse is usually regarded to give more body information
than the electrocardiogram (ECG), which results in broader
applications in health status analysis [10], [18]–[24].
In TCM, a Chinese medicine practitioner took pulse by
putting fingers on the patients’ wrist, certain position that
is regarded as the position of pulse [1]. However, the pulse
diagnosis skill requires several years experience to practice
and master. Sometimes it may be not accordant among
different practitioners because of the subjective judgment [26].
Compare with the finger-feeling diagnosis, a scientific way of
pulse diagnosis is using the sensing elements to simulate the
functions of fingers and transforming the physical signals into
digital signals [27]. The quantification of the pulse diagnosis
solves the problem of obtaining the objective pulse signals [4].
The pulse information which includes strengths, amplitudes,
fluency, shapes, widths, variations of the rhythm, and so
on are obtained from the digital pulse signals for further
processing [3].
Recently, there are two kinds of pulse devices that have been
reported. One is called pulse oximeter, applied with a probe
attached to the patient’s finger [8], [9]. It is a noninvasive
method for measuring the oxygen saturation of arterial blood.
Because the pulse rate always agreed with the heart rate on the
basis of ECG, the oximeter has been widely used in cardiac
monitoring [50]. A research group of National Dong Hwa
University in Taiwan [10] declared they successfully devel-
oped the artery health prediction system by detecting finger
pulse. Humphreys et al. [22] presented a system capable of
capturing two Photoplethysmography signals at two different
wavelengths simultaneously to give a quick indication of the
cardiac rhythm. There is no doubt that the pulse oximeter is
the big advance in acquiring the physiology blood signals.
However, a falsely high or falsely low reading will occur in
the pulse oximeter when hemoglobin binds to something other
than oxygen [9]. Physiologists also demonstrated that the pulse
waveform changes when the blood moves apart from the heart.
Thus, the pulse waveform of the finger is different from that
of the wrist [49]. As a result, the finger pulse feeling disagrees
with the pulse diagnosis in TCM. It is insufficient to analyze
the health status just through the artery oxygen saturation.
The other kind, operated with probe attached to the
patient’s wrist, is developed for objectifying TCM. Many
sensor types such as polyvinylidene fluoride (PVDF) [7],
optical sensor [8], [9], air pressure sensor [11], and ultrasonic
Doppler sensor [18] have been introduced for the pulse
acquisition. But these systems ignore the static contact
WANG et al.: NOVEL MULTICHANNEL WRIST PULSE SYSTEM
pressure information. According to traditional Chinese pulse
diagnosis (TCPD), Chinese medicine practitioners feel the
pulsations by touching on the wrist vessel with three
fingers and pressing with certain strength. The pressure levels
are then divided into three patterns, Fu, Zhong, and Chen,
corresponding to the levels of pulse depth under skin. It is an
important symbol for judging the patient’s physical conditions
in TCPD [1]. Therefore, it is necessary to get the static contact
pressure in pulse-collecting platforms.
Although the pulse system with pressure sensor is capable
of obtaining the depth information [14], [15], it is short in
acquiring the weak pulsations with high quality. The touch
sensation is easily disturbed by the surrounding noise while
feeling the weak physiological signals under skin. It is a
fatal flaw of the pressure sensors, and the system perfor-
mance needs to be robust under various conditions. Each
sensor type is with its specific characteristics and comple-
ments to each other, whereas previous systems only apply
single sensor type. It is known that single-type sensor cannot
replace the finger-feelings comprehensively and lose pulse
information inevitably [5]. Therefore, we should keep the
advantages of each sensor type and combine them together
to obtain more arterial pulse information. Meanwhile, because
most of the pulse systems are designed with a single-point
transducer [36]–[44], it brings the problem about sampling
location criterion. The three positions, defined as Cun, Guan,
and Chi in TCM, have significant different meanings and pro-
vide a standard for pulse location [1]. The analysis of a wrist
radial artery needs both temporal and spatial dimensions [15].
In this paper, we propose a novel multichannel wrist pulse
system with different sensor arrays to get rid of the prob-
lems of information loss. The pulse probe is composed of
three independent channels, and each channel consists of an
array of nine photoelectric sensors with one pressure sensor.
The photoelectric sensor array is introduced to detect the pulse
width information and to locate the center of radial artery. The
pressure sensor, regarded as the main sensing element, mea-
sures the pulsations with high resolution and the static contact
pressure. Pulse waveforms are acquired from three channels
corresponding to the position of Cun, Guan, and Chi. Step
motors are used to adjust the static contact pressure applied
on the radical artery, imitating the Fu, Zhong, and Chen of
practitioners’ feeling. Then, the pulse waveforms are processed
using amplifying circuit followed by data acquisition circuit,
and finally displayed on liquid crystal display. The pulse
signals collected by pressure sensors reflect the fluctuation of
the vessel, and the signals collected by photoelectric sensors
reflect the changes of blood volume. These two sensor types
with different principles are combined to obtain more pulse
information. We apply our system to the disease diagnosis
and obtain satisfactory results. The experiment results prove
that the proposed system has excellent performance in pulse
acquisition and practical usage for auxiliary diagnosis.
II. F RAMEWORK OF THE P ULSE S YSTEM
The proposed system consists of three main components:
pulse sensors, sampling circuit, and user interface (UI) (Fig. 1).
2021
Fig. 1. Brief flow chart of the proposed pulse system.
The sensor arrays first transform the physical pulse beatings
into electric signals, and then the analog circuit follows to
amplify the feeble physiology signals. Next, digital circuit is
employed to implement the analog-to-digital (AD) conversion.
Finally, the digital signals are transmitted to the UI at PC and
stored in open database connectivity (ODBC) database through
universal serial bus (USB).
The pulse from radial artery is usually taken as the TCM
practitioners did for thousands of years. Feeling pulse is not
as easy as it looks like, because pulse may be very weak in
certain subjects such as the obesity group, or at certain parts.
Generally, at Chi position one can hardly felt strong pulsations.
Moreover, the pressure applied on the wrist artery varies
among subjects and also differs in the three positions [1].
Therefore, the pulse diagnosis needs the primary knowledge
to ensure the approximate pulse position before the pulse-
collecting stage. Otherwise, it will be time consuming and
difficult to ensure the right positions of weak pulse signals.
A. Pulse Collecting
In TCM, practitioners took the pulse of patients by putting
their fingers on a specific area of wrist regarded as the
position of pulsation, which is further subdivided into three
adjacent domains named Cun, Guan, and Chi. As described
in previous pulse research, each of the three parts corresponds
to a specific inner body organ and brought different physical
information [1]. Hence, the Chinese medicine practitioners
used to press their index finger, middle finger, and ring finger
upon the wrist of patients to feel the pulse at Cun, Guan,
and Chi positions, respectively. Pulsations from the three body
parts were considered for judging the health status [51].
Fig. 2(a) shows our wrist pulse system and the UI. We use
the Velcro straps to fasten the pulse transducer instead of the
traditional fixed support, because it is more convenient and
flexible for clinical practice. The strap design also meets the
further research trends of wearable devices. When acquiring
subject’s pulse, pulse beatings at the patient’s wrist are first felt
with our fingers and the rough positions of Cun, Guan, and Chi
are searched and ensured. Then, we just twine the Velcro straps
around the wrist, and easily adjust the positions and directions
of the transducer to the right place by aligning each channel
with the corresponding position. Because the pulse waveforms
collected by the sensors are displayed on screen in real time
as ECG, it is convenient and friendly for interaction. Further,
the intervals between channels and the pulse center are subtly
adjusted by the setting knobs to obtain stable pulse signals
at higher amplitudes and less noise. If the optimal sampling
position is ensured by watching the pulse performance in the
interface, then the pulse sensors are applied automatically with
proper pressures by the step motors at the top of each channel.
2022
Fig. 2. (a) Pulse acquisition system. (b) Interaction interface.
TABLE I
F UNDAMENTAL W ORKING PARAMETERS OF P ROPOSED S YSTEM
Therefore, pulse beatings at the wrist are felt by the sensors
straight at the three body parts of Cun, Guan, and Chi. Then,
we preview the collected pulse for a moment to check the
signal stability. The pulse sampling phase starts following the
preview stage.
During the pulse-collecting stage, subjects are required to
hold a sitting position and to keep calm, which would reduce
the noise caused by body movements and breathing actions.
Meanwhile, the parameters and the positions of sensors are
required to stay in a similar manner until the sampling proce-
dure ends. In our experiments, pulse signals are taken from the
left hand of patients uniformly. And the static contact pressures
of sensors are controlled by the step motors instead of the
manual operation, and thus more accurate pressure adjustment
is realized. Fig. 2(b) shows the local enlargement of the pulse
transducer with three independent channels in pulse taking.
The whole collecting period lasts around 60 s for covering
a complete and stable interval of multiperiod pulse. During
this procedure, patients do not feel any discomfort as a result
of the noninvasive approach. When collected pulse data meet
the requirements of timing and quality, we stop the procedure,
take down the straps, and store the database in a hard disk.
The specification parameters of the proposed system are shown
in Table I below. Our previous experiments have shown that
this novel pulse system is robust through the repeatability and
stability test.
B. Pulse Processing and Interaction Design
The sensing elements, which include three pressure sensors
and 27 photoelectric sensors in total, constitute three inde-
pendent sensor channels. Each channel which involves one
pressure sensor and nine photoelectric sensors simulates a
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 64, NO. 7, JULY 2015
Fig. 3. UI of the pulse system.
finger’s function to feel pulse from one of the corresponding
positions marked as Cun, Guan, and Chi. The sensor array in
each channel responses to pulse beatings and blood flowing,
transfers physical changes to quantitative electric signals.
Pulse signals are then processed by signal amplification and
filtering module, AD conversion module, and microdigital
signal processing (DSP) module consequently. The digital
pulse waveforms are finally sent to a computer through the
USB interface for visualized operations and pulse database
management.
The software framework of pulse system involves interface-
driven design, pulse control UI, and database operation. The
proposed system shows digital pulse waveforms on computer
screen (Fig. 3). Pulse signals sampled by the sensor arrays can
be clearly watched in real time. The right part in the UI shows
the pressure pulse signals of the three channels, and the left
region of the interface gives a display of nine photoelectric
signals from the channel selected. Through the operation on
the Graphical User Interface, pulse data are then stored for
further analysis in a database form, the content of which
includes original pulse digital waveforms, subjects’ informa-
tion, channel information, sampling pressure, sampling time,
and sampling rate. Meanwhile, it is also a standard template
for recording quantized pulse information and related labels.
In our database, the characteristic curve of one sample is
taken from the 30 sensing elements (three pressure sensors and
27 photoelectric sensors) in three channels for around 60 s.
Because the sampling frequency is nearly 500 Hz, one sensor
in one sample creates a 500 × 60 = 30 000-D feature vector.
And the total pulse dimension of one subject of the whole
sensor arrays reaches 900 000, it is necessary to reduce the
feature space for extracting the useful information.
C. Data Analysis Methods
The pulse system measures microvoltage changes across
each sensor and converts the raw signal into a digital value.
Then it can be applied to further analysis. The whole analysis
procedure involves three steps: signal preprocessing, feature
extraction, and classification.
1) Pulse Signal Preprocessing: Computerized pulse
signals acquired from our system introduce noise inevitably.
Meanwhile, because of the respiratory movement, the baseline
WANG et al.: NOVEL MULTICHANNEL WRIST PULSE SYSTEM
Fig. 4. Framework of pulse waveform analysis.
Fig. 5. Typical wrist pulse signal. (a) Original pulse signal of certain channel.
(b) Single-period pulse signal segmented from the selected region of (a).
of pulse waveform always wanders during the sampling period.
As a result, pulse preprocessing is needed to improve the
final analysis performance before the feature extraction phase.
Because pulse waveform is a periodic signal and each basic
period represents a breath cycle, the period segmentation
section follows the wander adjustment module. It provides a
local view method to analyze the single-period pulse signal as
a part of the whole. In our experiments, the single periods are
regarded as the basic element of the pulse waveforms, and the
further processing of feature extraction are developed in this
subspace. The whole processing procedure is shown in Fig. 4.
Fig. 5(a) shows a part of typical pulse waveforms collected
from the proposed channels. It includes several complete
cycles and each interval represents a cardiac cycle. Fig. 5(b)
shows a local enlargement of the single period in the rec-
tangular box. All the same periods in the time series are all
processed and segmented at the final stage of preprocessing.
2) Feature Extraction: The aim of feature extraction is to
find a low dimensional subspace that preserves most of the
pulse information in the original time series and removes the
redundancy. In this paper, we define and extract the pulse fea-
tures from three aspects: 1) time-domain features; 2) frequency
features; and 3) wavelet features. Time-domain features
mainly compute the single-period pulse information which
includes pulse shape features, pulse pressure features, and
pulse width features, whereas the frequency features provide
cycle stability information existed in the multiperiod signals.
Wavelet features are extracted from the high-level decompo-
sition coefficients of the pulse wavelet transformation (WT).
2023
The feature extraction procedure consists of three stages.
Because pulse waveforms of one subject are simultaneously
sampled from 30 channels, and the redundant information
appears among the neighborhood regions, time series similar-
ity analysis (TSSA) is first employed to optimally select one
representative from the nine photoelectric pulse signals in each
channel of the three channels, respectively. This technique is
similar with the data compression. Pulse width information
and the pressure value are obtained from the curves fitting
of the amplitudes and the baseline at the signal selection
stage. Then, the extreme points is calculated out as the shape
characteristics in the segmented pulse single period and the
period intervals are also obtained from a part of continuous
multiperiod pulse waveforms, whereas the wavelet analysis is
introduced for the energy measurement. Finally, we combine
all the features together in a vector in a serial fashion.
3) Classification: The pattern recognition is proposed for
the disease analysis, and in our experiments we use this
method to distinguish the healthy subjects from the patients
affected with diabetes. Because some of the waveforms
from the three independent channels have a significant linear
correlation with each other, it is necessary to compute the
correlation coefficients and remove the relative feature dimen-
sions. PCA algorithm is then used for the dimensionality
reduction. We calculated the eigenvectors and eigenvalues of
the feature vectors and sort the eigenvectors, by descendant
eigenvalues, and then projected all the feature space onto
the PCA subspace spanned by the selected principal compo-
nents. Finally, the processed feature vectors are classified into
healthy and disease categories through the Linear Discriminant
Analysis (LDA) and Support Vector Machine (SVM) classi-
fier, respectively. LDA is employed for the linear analysis.
We select the SVM with radial basis function (RBF) kernel
for the nonlinear decision. For constructing the classifiers, we
randomly divide the feature samples into five equal groups and
each time we use four of the five groups for the training and
leave one for the verification. The optimum specifications of
the kernel are determined by the grid-searching technique. The
final classification is obtained from the average performance
of the fivefold cross validation.
III. D ESIGN OF THE D IFFERENT S ENSOR A RRAYS
The performance of the pulse system largely depends on the
response of sensors. Thus, the sensor design plays an impor-
tant part and makes our system distinctive compared with
other previous pulse-taking platforms. Multichannel design
and different sensor arrays are employed to obtain more
information and standardize the sampling operations. The three
independent channels correspond to the three body positions
Cun, Guan, and Chi, and each channel is composed of a
pressure sensor fusion with a photoelectric sensor array. The
different sensor arrays complement each other to provide more
features for pulse analysis.
A. Pressure Sensor Array
According to the previous researches, we choose tradi-
tional pressure sensor as the main sensing element in the
2024 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 64, NO. 7, JULY 2015
Fig. 6. Physic model of the pressure sensor. 1: fixed mount. 2: strain gauge.
3: cantilever beam. 4: contactor. 5: skin. 6: blood vessel.
proposed system, which is more similar in function to the
physician’s fingers when taking one’s pulse. They sense the
changes of pulse pressure at the wrist and obtain the pulsation
information in the same way directly. Consequently, pressure
sensor is closer to TCM in essence and easier to be under-
stood and accepted both in psychological and principle. The
previous experiments about the pulse acquisition have shown
that signals taken from these three positions do not carry
identical information, because waveforms from each channel
are obtained under different contact static pressure and they
vary with each other in shapes and amplitudes.
1) Implementation of Pressure Sensor Array: Pressure sen-
sor is a kind of widely accepted sensor to detect human pulse.
It measures the changes of contact pressure upon the detector.
Fig. 6 shows the schematic representation of the pressure
sensor structure. Cantilever beam is selected as the main
measure elements, which presents elastic deformation when it
is acted upon by a force at the free end. A contactor is placed at
the bottom of this end, and an electric resistance strain gauge is
laid on the top of cantilever beam. The electric resistance strain
gauge changes its resistance value when deformation occurs.
With this principle, we can transform the physical signals to
electric signals through the interrelationship between pressure
force and resistance value.
The whole operating procedure of taking pulse involves
three stages. 1) The contactor is placed at the wrist upon
radial artery to feel the pulse beatings with certain pressure.
2) The pulse beatings cause a periodic elastic deformation on
cantilever beam as well as the strain gauge. 3) The deformation
changes the resistance value of gauge and the electric lever of
output signals.
In the proposed system, we select the TP2.6 series semi-
conductor gauge as the electric resistance strain gauge module,
which has the advantage of high sensitivity and quick dynamic
response. Its appearance and corresponding parameters are
shown in Fig. 7(a) and Table II, respectively. It operates on the
principle that the resistance of silicon-implanted piezoresistors
will increase when the resistors flex under any applied force.
The gauge is attached closely to the cantilever beam to sense
its deformation and monitored with the Wheatstone bridge to
transform this specific resistance change to the voltage change.
In the bridge, three high-precision resistors (R2, R3, and R4)
are fixed and the strain gauge is inserted as the active
Fig. 7. (a) TP series semiconductor gauge. (b) Wheatstone bridge.
TABLE II
PARAMETERS OF S EMICONDUCTOR G AUGE
resistor (R1). We add a series resistor (Rm) to compensate
errors caused by temperature variations. Fig. 7(b) shows
the low power, unamplified, compensated Wheatstone bridge
circuit design which provides inherently stable millivolt level
outputs over the force range. Hence, the pulse beatings are
converted into the electric signals accurately eventually.
2) Standardizations of Pressure Sensor: Because the
performance of the pulse system largely depends on the sensor
specification and property, we propose several solutions to
standardize the criterions of pressure sensors, including the
pressure calibration and nonlinear response test. It provides
information from two important aspects, the accuracy and
dynamic region of the static pressure. As the depth information
of pulse, which is completely derived from the contact static
pressure during the collecting phase, makes it necessary to
develop the quantification technique accurately. For ensuring
the dynamic region, we introduce a method to acquire the rela-
tionship between pressure and the signal output. Nevertheless,
the contact pressure and the amplitude of pulse do not match
a simple linear relationship. When contact pressure exceeds
a certain threshold, the amplitude will decrease. The related
specific parameters of the proposed system and experiments
results are given below.
The pressure calibration guarantees the accuracy of the
pressure sensors in the proposed system. The baseline value
of pulse waveform sampled by pressure sensor is influenced
by the contact static pressure directly. Using the characteristic
of the relationship between signal dc voltage and pressure
value, we propose a standard pressure calibration procedure
which is implemented in following steps. First, we fix the
sensor upside down on a platform, and then several weights
WANG et al.: NOVEL MULTICHANNEL WRIST PULSE SYSTEM
TABLE III
P RESSURE S ENSOR S TANDARDIZATION
TABLE IV
S PECIFICATIONS OF THE P RESSURE S ENSOR
are successively placed upon the sensor surface to acquire the
dc voltage, and each stage is carried out for ten times to obtain
the average dc level. Finally, the relationship between contact
static pressures and the dc level are computed using the linear
regression method.
Table III shows the outputs of a pressure sensor under
different sampling pressures, and the fitting curves is obtained
by fitting the pressure with dc voltage. Thus, the amplitudes
of pulse baselines (dc level) are equivalently converted to the
contact pressure through the one-to-one linear correspondence
method.
For ensuring the dynamic linear region, we acquire the
relationship between pressure and the pulse amplitude by
adding contact force gradually for each channel, respectively.
Results show that they do not match a simple linear relation-
ship. When lighter contact pressure is applied, the amplitude
of pulse waveform will increase when the pressure rises.
If contact pressure exceeds certain value, then the amplitude
will decrease. Therefore, the inflection pressure which corre-
sponds to the pulse with highest amplitude is regarded as the
upper limit of the dynamic linear region.
The proper pressure with the highest amplitude is not the
same among the subjects. The experiments show that the
pressure lies from 1.50 to 2.00 N when the pulse waveform
reaches the max amplitude. Therefore, 2.00 N is selected as
the max pressure when sampling the pulse. The amplitude and
the pressure are almost under the linear relationship when the
pressure is below the extreme value. Table IV gives the final
specifications and parameters of the pressure sensors after the
standardizations. It makes the pulse system more robust and
normalized.
B. Photoelectric Sensor Array
Photoelectric sensor is selected as auxiliary sensor to
detect changes of radial artery blood volume. It acquires the
2025
Fig. 8. Appearance of the NIR sensor.
Fig. 9. Principle of the photoelectric sensor.
information of hemodynamic parameters and arterial stiffness,
which makes it distinguished from the other sensor types.
Consequently, photoelectric signals are regarded as a different
way to obtain the information and a supplement to the pressure
sensors. Because photoelectric elements are much smaller
(such as Charge-coupled Device) in shape than the other types,
it is easier to develop a high-density-integrated sensor array.
1) Principle of Photoelectric Sensor: Photoelectric sensor
has the advantage of small size, high sensitivity, and less
temperature influence. It has been widely used in the pulse
oximetry system [6]. In the proposed system, a type of
optical sensor known as near-infrared (NIR) light is selected
as it can penetrate skin and reflect information below skin.
According to the physiology foundation and our experiments,
the penetration of NIR around 1000 nm is optimum for the
pulse taking. Therefore, the model of 970 nm NIR sensor is
chosen as the emitter. Fig. 8 shows the NIR sensor appearance.
The inner structure of the photoelectric sensor is shown in
Fig. 9. It is composed of a pair of optical emitter (NIR) and
optical detector (phototransistor). The blood absorbs most of
the NIR light, and the reflected light changes as the blood
flow of the radial artery changes. Thus, we use this feature to
transform the pulse flow into electric signals.
2) Motivation of Photoelectric Sensor Array: We designed
a sensor array to enlarge the sensing area. Each array consists
of nine photoelectric emitters and nine photoelectric detectors,
which makes the number of sensing unit up to nine. As shown
in Fig. 10, the sensor pairs are laid in two longitudinal rows
with a light barrier in the middle. At the left side, we place
nine photoelectric emitters with equidistant intervals of about
1 mm. The arrangement of emitters and detectors are set in a
one-to-one relationship symmetrically.
At the pulse-collecting phase, each photoelectric sensor pair
which involves one emitter and one detector works almost
2026 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 64, NO. 7, JULY 2015
Fig. 10. Top view of photoelectric sensor array and the real object.
Fig. 11. Simple schematic diagram of array placement.
independently with the other sensors in the array. The motiva-
tion of this design belongs to the factor that radial artery varies
in width among subjects and can be covered by nine sensor
pairs in a lateral direction (Fig. 11). We define this feature as
pulse width and add it into the feature space.
Moreover, the design of sensor array is also introduced for
the requirement of ensuring pulse location. A wide coverage
sensor array guarantees that the center of pulse is always
under control and can be easily confirmed by comparing
the amplitudes of pulse signals from the nine sensors of
the array. Then, we adjust sensor probe in the horizontal
direction to make sure that the center of pulse signal with
highest amplitude is collected by the sensor in the middle of
array.
C. Combination of Pressure and Photoelectric Sensor Arrays
Either of these two sensor types has its own characteristic.
The pressure sensor which measures changes of throb force
from the radial artery also acquires the information of pulse
depth. The probe with three channels placed on the corre-
sponding locations of wrist is applied with different external
pressure which varies not only from individual to individual
but also with the three positions to obtain an excellent pulse
waveform. According to TCPD, the depth of pulse was always
an important physiological manifestation and classified into
three patterns named Fu, Zhong, and Chen. Each pattern rep-
resents a certain physical status in TCM. Naturally, this unique
information that cannot be extracted from other nonpressure
sensor types has been introduced into the proposed system to
enrich the feature space. Our experiments have shown that it
improves the final performance.
The photoelectric sensor which has the advantage of small
size and high sensitivity is selected as subchannel in the
proposed system. It provides another visualized way to feel
the pulse in comparison with the pressure sensor as it detects
the change of blood flow at the radial artery. To ensure the right
Fig. 12. Inner structure of the fusion sensors.
Fig. 13. Schematic diagram of the fusion sensors.
position of pulse, we introduce the array design to cover the
region. Besides, the information of pulse width, which reflects
the degree of radial artery stenosis, is also obtained as it
varies in amplitude with each sensor in the array. Considering
these two sensor types with irreplaceable use value, fusion is
adopted for this issue.
1) Different Sensor Arrays Fusion Principle: Fig. 12 shows
the inner structure of the fusion sensors. The pressure and
photoelectric sensors are integrated in a two-lever cantilever
beam to combine their respective merits. There is a light-proof
box containing a photoelectric sensor array at the bottom and
on the upside a strain gauge is attached to acquire the contact
pressure to the skin. These sensor elements are surrounded by
metal shell.
The pulse-sampling principle is shown in Fig. 13. The
photoelectric sensors which lie at the bottom receive reflected
NIR light from the blood vessel directly. Meanwhile, the force
sensor acquires the pressure variation of the radial artery under
skin transferred from the cantilever framework. Consequently,
both of these two sensor types are compatible in the proposed
system and collect pulse simultaneously.
2) Architectural Structure of Multichannel: Fig. 14 shows
the overall structure of the novel pulse probe with three
channels. There are nine adjustable knobs in total, which
are grouped into three directions. And every single channel
is controlled by three of the knobs from each direction,
respectively. We assume that there is a coordinate system. Step
motors are applied in z-axis direction to adjust the height of
sensor arrays which reflects the contact pressure, while the
distance among three channels are controlled by the y-axis
direction knobs to ensure the right positions of Cun, Guan,
and Chi. The knobs in the x-axis direction guarantee that the
WANG et al.: NOVEL MULTICHANNEL WRIST PULSE SYSTEM 2027

Fig. 15. Pulse waveforms taken by the proposed pulse system. (a) Part of
typical oringnal pulse waveform collected by pressure sensor, the horizontal
axis represents the time (s) and the ordinate axis represents the amptitude
of the pulse waveform (mV). (b) Denoised pulse signal. (c) Baseline wander
Fig. 14. Overview of sensor probe framework. removed pulse signal.

center of pulse and the middle of photoelectric sensor array

just coincide.

IV. T ESTING R ESULTS

A. Pulse-Collecting Performance
The proposed system with different sensor arrays collects
pulse waveforms from the three channels known as Cun, Guan,
and Chi. The wrist pulse signals are roughly divided into either
a triple-humped wave or a double-humped wave according
to the shape structure [41]. The proposed system collects the
similar pulse waveforms as the previous pulse-taking platforms
with single channel. Fig. 15(a) shows an original pulse signal
from one of the 30 channels. Fig. 15(b) and (c) shows the
denoising and baseline wander removal results of the raw pulse
signal, respectively. The shape features such as tidal wave,
dicrotic notch, and dicrotic crest also appear in the declining Fig. 16. Fitting diagram of the pulse array from the spatial dimensions.
segments. It proves that our system has the same ability as the
pulse-taking platforms with single channel in pulse acquisition.
The stability and sensitivity of the waveform both meet the The algorithm based on surface fitting is introduced to simulate
requirements of pulse diagnosis. the radial artery blood flow under skin. Fig. 16 shows the
Meanwhile, pulse signals sampled by the photoelectric fitting curves by using spline interpolation algorithm. The
sensor array at each channel bring the pulse width information. pulse signals sampled by the nine array sensors vary
2028 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 64, NO. 7, JULY 2015

Fig. 17. Original photoelectric pulse waveforms from the sensor arrays of the three pulse-taking positions.

Fig. 18. Pressure pulse waveforms collected from the three channels.

in amplitudes, which reflect the distance from the center of
pulse waveforms. On the basis of fitting, the width of the blood
vessel is obtained by setting a proper threshold. The radial
arterial varies in width as it associated with the conditions of
vessel. Different health states usually lead to the change of
vessel form and it is an important symbol for judging the
fullness or tension produced by the fluid content of blood
vessels [15]. On the other hand, the sensor array provides a
criterion for the pulse location. It guarantees that the pulse
waveforms are acquired under the optimal collecting position.
Additionally, the proposed system is able to measure
30 channels of wrist pulse simultaneously at Cun, Guan,
and Chi three pulse-taking positions. The photoelectric pulse
waveforms of the three channels are shown in Fig. 17, and
each column presents a channel and each channel includes
nine signals in the array. It can be seen that the waveforms
are stable and almost the same in shape and frequency. Fig. 18
shows the pressure pulse waveforms from the three channels
with independent contact static pressures. The details and
amplitudes of each period have slight difference among the
whole time series. Generally, the pulse waveforms from the
Chi channel have the least power energies. And the pulsations
even disappear at this channel under certain physiological
condition such as the weakness. It is an important symbol
for the pulse diagnosis.
B. Disease Diagnosis
In the second stage, we test the system ability in disease
diagnosis. The classification between healthy persons and
WANG et al.: NOVEL MULTICHANNEL WRIST PULSE SYSTEM
Fig. 19. Statistical distribution of pulse features between healthy subjects (blue point) and diabetes (red point). (a) Pulse feature distribution. (b) Pulse width
distribution.
TABLE V
D ATA O RGANIZATION
patients known to have suffered from diabetes is selected
to evaluate the performance. Modern clinical studies demon-
strate that premature loss of arterial elasticity and endothe-
lial function exists in patients with diabetes. Such loss will
eventually decrease the flexibility of vasculature and changes
the shape, amplitude, and rhythm of patient wrist pulses in
correspondence with the hemodynamic characteristics of blood
flow [48]. Computerized pulse diagnosis for diabetes has been
studied by many researches [52]–[54]. Therefore, pulse is
considered as an indicator for diabetes.
In our experiments, we select the diabetes diagnosis
as a case study and constitute a pulse dataset containing
125 healthy subjects and 125 diabetes patients. All the patients
were inpatient volunteers from the Hong Kong Prince of Wales
Hospital. All the healthy volunteers were from the physical
examination center of the hospital. The labels of the healthy
subjects were determined according to the recent health check,
whereas the diabetic patients were confirmed and correlated by
comparing their levels with standard clinical blood markers.
The distribution of the pulse database is ensured to avoid the
influence caused by biological interpatient variability. Table V
lists the summary of the dataset. The age, body–mass index,
and gender distribution is similar between healthy and diabetes
patients.
Three pressure pulse signals and 27 photoelectric array
signals are collected for each subject from Cun, Guan, and Chi
three positions. The pulse preprocessing methods composed of
denoising, baseline wanders removal, and period segmentation
are adopted. For the feature extraction, we first choose the
2029
TABLE VI
S TATISTICAL R ESULTS OF THE F EATURE D ISTRIBUTION
TSSA algorithm to select the representive of pulse array in
each channel. Then the features of pulse shape, pulse width,
frequency, and static contact pressure are extracted. We also
extract the wavelet coefficients as energy features, because the
energy distribution in different frequency bands are different
from that of a patient and can be easily obtained by WT.
Finally, we combine all the features as a vector in serial
strategy.
Table VI gives statistical results of pulse feature distribution
between healthy subjects and diabetes. Although there is dif-
ference in average value of energy and frequency feature, the
large standard deviation value leads to the overlap. Fig. 19(a)
shows the distribution of pulse shape features between these
two patterns. The middle line stands for average value of
pulse shape, and the interval at each point is the standard
deviation. It can be seen that the main difference appears
in the descent stage of the pulse waveforms. Especially, in
the period between pulse tidal wave and pulse dicrotic wave
pulse signals have significant difference in average value and
deviation interval. The width distribution shown in Fig. 19(b)
indicates that the width of the healthy subjects is more like
the uniform distribution, while the pulse width of the diabetes
is concentrated in the lower bands.
The statistical distribution of pulse features shows that the
classification between health and diabetes is complicated. The
linear classification algorithm or statistical algorithms are not
able to distinguish the healthy subjects from the diabetes
patients at a high accuracy. A comparison between linear-SVM
classifier and RBF–SVM classifier is implemented.
2030 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 64, NO. 7, JULY 2015
TABLE VII
C LASSIFICATION R ESULTS
Table VII shows the classification rates between healthy
subjects and diabetes by using independent single-pulse
features and fusion features, respectively. It can be seen that
any of these independent pulse features is enough for the pulse
diagnosis between the two patterns. Meanwhile, the relation-
ship between healthy and diabetes is complicated. The linear
classification algorithm used to distinguish the healthy subjects
from the diabetes patients is not able to achieve satisfactory
results as kernel method.
With the fusion features, we select the SVM with RBF
kernel for the classification between health and diabetes
through the comparison. To construct the classifiers, we adopt
the fivefold cross-validation method. The healthy and diabetes
samples are randomly divided into five equal folds, respec-
tively. Each fold contains 25 healthy samples and 25 diabetes
subjects. Then, we select four of them each time to form the
training set and leave one as the testing set for verification.
The specifications of the kernel are determined by the grid-
searching technique. We obtain the final accuracy by averaging
the classification accuracy of the fivefold cross validation as
the indicator of system performance.
To test the performance of fusion sensors, we compare
and list the classification accuracy between single-type pulse
waveforms and the combination pulse waveforms, respectively.
Moreover, the classification results between the traditional
single channel pulse waveforms and the multichannel pulse
waveforms are also presented to test the superiority of the
novel design. Fig. 20 shows a 2-D PCA analysis of the
fusion feature distribution with the first principal component
(PC1) plotted against the second (PC2). The ∗ represents
the diabetes samples and + the healthy samples. The two
dimensions explained 23.7% of the variation in the data, 14.6%
for PC1, and 9.1% for PC2 component. The two classes are
discriminative in spite of some feature overlaps.
To see more clearly which features contribute to PCA1,
we provide insight into the loadings of the PCA1 as shown
in Fig. 21. As mentioned previously, the feature vector is
composed of 1608 elements for one subject from three
channels. At each channel, the feature dimension is 536,
including 268 pressure features and 268 photoelectric features,
respectively. The feature is fusion is fused in a serial strategy.
It can be seen that six parts have obvious influence and
weights for PCA1. The obvious feature interval at the middle
Fig. 20. PCA 2-D plot of the multichannel fusion features for diabetes
diagnosis.
Fig. 21. Loadings of PCA1.
is the pulse principle shape features of Guan channel, the
beginning features interval is the pulse principle shape features
of Cun channel, and the end part is the pulse features of
Chi channel. The loadings of PCA1 are concentrated on pulse
shape features of pressure Guan pulse, pressure Cun pulse,
and photoelectric Guan pulse. The Chi channel contributes
least to the PCA1. Meanwhile, the wavelet energy features
also have effect on the PCA1. To sum up, the pulse shape
features and wavelet energies extracted from Cun and Guan
channel is clearly the most important separation factor between
healthy and diabetes patients. Because modern clinical studies
demonstrate that there is premature loss of arterial elasticity
and endothelial function for patients with diabetes [48].
The selected features of PCA1 mainly reflect the global
shape information, and the detail information such as
secondary shape features are not considered. Therefore, using
the only single PCA1 is not enough to well distinguish the
healthy patients from the diabetes group. The secondary com-
ponents also play an important part in pulse pattern recognition
and improve the classification rates.
These experiments have proved the significance of the
TCPD and verified the theory of Three Parts with Nine
Palpation Sites, which demands that pulse from Cun, Guan,
and Chi should be considered as a whole effect for physiolog-
ical status diagnosis.
V. C OMPARISON AND A NALYSIS
A. Advantages of the Proposed Multichannel Pulse System
We develop a novel wrist pulse system which solves the
problems and disadvantages in the previous pulse-taking plat-
forms. For the sensor selection stage, the traditional strain
WANG et al.: NOVEL MULTICHANNEL WRIST PULSE SYSTEM 2031

gauge fusion with photoelectric sensor is proposed to enrich TABLE VIII

the pulse information. Strain gauge, selected as the pressure P ERFORMANCE C OMPARISONS IN THE L ITERATURES
sensor, senses the changes of dynamic pressure at the wrist
artery directly. And the static contact pressure which reflects
the pulse depth is also obtained. On the other hand, the
photoelectric sensor measures the blood flow at wrist artery
and transforms the volume changes into voltage output. With
the fusion sensors, we then introduce the sensor array design
for the pulse location. A wide coverage of the array guarantees
that the pulse-taking positions satisfy the same evaluation
criterion.
Additionally, pressure sensor is designed as the main
Wang’s and Hu’s reports make progress in array design and
sensing element in the proposed system. Pulse amplitude
both select the surface analysis. The main difference comes
changes when different contact pressure is applied by pressing
from the number of sensing elements and the arrangement.
the sensors, which reflects the information of pulse depth
However, they are both short in the simplex sensor type.
under skin. Strain gauge, photoelectric sensor, and PVDF
Our proposed system adopts the advantages and improves the
are possible sensor candidates for collecting pulse. However,
performance by combining different sensor types. From the
only strain gauge satisfies the requirements of acquiring the
comparison, the proposed pulse-taking platform is capable of
contact static pressure in sensitivity and accuracy. In TCM, the
acquiring more information and closer to the procedure of
depth of pulsations are classified into three groups namely Fu,
taking pulses in TCM.
Zhong, and Chen according to the contact static pressure felt
To sum up, the proposed wrist pulse system acquires not
by practitioner’s finger. The pulse depth at Cun, Guan, and Chi
only time series as single sensor platform but also the spatial
differs from individual to individual and becomes an important
features from the sensor array. The standard of the optimal
physiological and biochemical characteristic in TCPD.
pulse-taking position is also ensured for operation consistency.
In TCM, it was considered that the pulse taken from Cun,
Moreover, the fusion sensors provide different techniques
Guan, and Chi three places brought distinguishing informa-
to complement pulse information. At last, experiments are
tion correlated to the conditions of different organs inside
conducted to verify the performance of the optimal design
body [51]. The Chinese medicine practitioners judge the health
and to test the ability of our system in disease diagnosis. The
status of patients by analyzing the pulse from the three
classification result is 85.6% by extracting features from Guan
points. Hence, the structure of multichannel is implemented
channel, as shown in Table X. Wang and Zuo [14] proposed
by combining three independent single pulse-taking platforms
a compound pressure pulse acquisition system, and the clas-
as an integrated framework. The three channels, corresponding
sification accuracy is 85.4% for pulse diagnosis by using the
to the three positions, are developed to simulate the tactile sen-
main pulse signals from the Guan channel. Chen et al. [48]
sation of fingers. And the pulsations from the three channels
introduced Gaussian models to extract pulse features from
are all taken for the data analysis.
single channel and the classification accuracy reaches 85.9%
On the basis of the multichannel framework, the proposed
between healthy and diabetes. Compared with these reported
system is then applied with the sensor array in each channel to
platforms, the proposed system can acquire similar classifi-
cover the interesting regions. Photoelectric sensor is selected
cation results and obtain satisfactory pulse quality for pulse
as the sensing elements of array due to the advantages of small
diagnosis.
size and high sensitivity. An array composed of nine photo-
electric sensors laid in an equidistance longitudinal direction
is designed to locate and ensure the right position of pulse B. Verify the System Performance in Disease Diagnosis
center. Besides, the information of pulse width which reflects Although the pulse waveforms taken by the proposed
the degree of radial artery stenosis is also obtained from the system bring similar features as the previous pulse system,
amplitudes of array signals. The motivation of array design and meet the requirements in stability and repeatability, it
also lies in the fact that the wrist pulse analysis has to include is necessary to verify the superiority of the improvements
both temporal and spatial dimensions [15]. Thus, a pulse- and innovations by comparing the performance in pulse
taking platform with a sensor array is required, and the array diagnosis.
acts more like the surface of a finger. Actually, a pulse probe The fusion sensor and multichannel design are introduced
with only a single point needs multiple sampling procedures in our system. We evaluate the performance of fusion sensor
to acquire the spatial information, which turns out to be time through the classification between the healthy samples and
consuming and inconvenience. the diabetes. First, pressure and photoelectric pulse signals
Table VIII lists the pulse systems that have been reported in from each channel are selected as independent features for
detecting wrist radial arterial pulse. Basically, platforms with the pattern recognition, respectively. Then, the combination
single sensor do not provide enough information to measure of these two types of features is applied and the accuracy
the pulse width and the multichannel features. Although nine rise in all three channels. As shown in Table IX, we can
probe-sensing elements are employed in Tang’s report, the see that improvements are significant in the Chi channel by
cruciform arrangement limits the feasibility of surface fitting. a 4% increase, because the pressure pulse at this position
2032 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 64, NO. 7, JULY 2015
TABLE IX
C LASSIFICATION R ESULTS OF THE S ENSOR F USION
TABLE X
C LASSIFICATION R ESULTS OF THE M ULTICHANNEL
are usually missing due to the deep pulse depth and week
energy. Although the photoelectric sensor is hardly influenced
by these factors and contributes to the information supple-
ment, the pulse waveforms at the other two channels bring
enough information in pressure pulse space. Hence, the fusion
does not introduce a significant increase. However, the array
design in photoelectric sensors guarantees the optimal sensing
position and improves the performance of pressure sensor
indirectly.
The multichannel design is another innovation in the pro-
posed system according to the assumptions and rules in TCM.
Both pressure and photoelectric features mentioned previously
are integrated for each channel. Table X gives the classification
results between healthy person and diabetes. Pulse waveforms
from Guan channel achieve the highest accuracy among
the single-channel group. Generally, the pulse platforms with
single taking position are all required to collect the pulsations
from Guan position at the wrist. The classification accuracy
of single channel reaches 85.6% in our system. And the
two-channel combination groups improve the performance
and obtain 90% accuracy at most for the Cun and
Guan channel fusion, whereas the classification with the
three-channel combination reaches 91.6%. The experiments
show that the pulse system with multichannel is more powerful
in acquiring information than the taking platforms with single
channel.
Compared with the traditional pulse from single channel
or simplex sensor, the proposed pulse system achieves higher
classification result by combining the multichannel informa-
tion. The innovations mainly exist in two aspects: multichan-
nel design and sensor fusion. Both are tested and verified
through the disease-recognition experiments. The ability of
the proposed system is satisfactory for practical auxiliary
diagnosis. The novel design results in the increase of clas-
sification accuracy and brings in the standard in pulse-taking
operation.
VI. C ONCLUSION
This paper proposes a novel wrist pulse system with
multichannel design and different sensor arrays, which has
broad applications in medicine diagnosis such as the healthy
condition evaluation and the disease analysis. The framework
of the system, pressure and photoelectric sensor design, novel
array structure, and signal processing procedure are intro-
duced. And the standardizations of the specific parameters
related with the pulse taking are discussed for the operation
normalization. Experiments are then conducted to verify the
system performance and test the ability. The results show that
pulse signals taken from sensor arrays obtain more information
than pulse collected by single-type sensor. And the multi-
channel design brings significant improvement in accuracy.
The proposed system is practical for the disease classification
between healthy subjects and diabetic patients.
The novel design of our system mainly consists of two
aspects: 1) multichannel and 2) sensor arrays. Each part is
well analyzed and brings contributions to the whole system.
Although the current system achieves satisfactory results by
using the pulse database from multichannels, it is also nec-
essary to improve the performance by selecting the optimum
representative. Therefore, pulse feature selection and ranking
will be the next issue. Meanwhile, we will extend the pulse
diagnosis for more diseases and conditions.
ACKNOWLEDGMENT
The authors would like to thank the editor and the anony-
mous reviewers for their help in improving the paper.
R EFERENCES
[1] B. Flaws, The Secret of Chinese Pulse Diagnosis. Boulder, CO, USA:
Blue Poppy Enterprises, Inc., 1995.
[2] M. Broffman and M. McCulloch, “Instrument-assisted pulse evaluation
in the acupuncture practice,” Amer. J. Acupuncture, vol. 14, no. 3,
pp. 255–259, 1986.
[3] C. T. Lee and L. Y. Wei, “Spectrum analysis of human pulse,” IEEE
Trans. Biomed. Eng., vol. BME-30, no. 6, pp. 348–352, Jun. 1983.
[4] K. Amano, H. Kasahara, H. Ishiyama, and K. Kodama,
“Diagnostic apparatus for analyzing arterial pulse waves,”
U.S. Patent US6 767 329 B2, Jul. 27, 2002.
[5] E. Kaniusas et al., “Method for continuous nondisturbing monitoring
of blood pressure by magnetoelastic skin curvature sensor and ECG,”
IEEE Sensors J., vol. 6, no. 3, pp. 819–828, Jun. 2006.
[6] M. Nitzan and H. Taitelbaum, “The measurement of oxygen saturation
in arterial and venous blood,” IEEE Instrum. Meas. Mag., vol. 11, no. 3,
pp. 9–15, Jun. 2008.
[7] M. Toda and M. L. Thompson, “Contact-type vibration sensors
using curved clamped PVDF film,” IEEE Sensors J., vol. 6, no. 5,
pp. 1170–1177, Oct. 2006.
[8] J. P. Jaeb, D. W. Gilstad, and R. L. Branstetter, “Optical sensor for pulse
oximeter,” U.S. Patent US4 880 304 A, Nov. 14, 1989.
[9] L. S. Lovinsky, “Urgent problems of metrological assurance of opti-
cal pulse oximetry,” IEEE Trans. Instrum. Meas., vol. 55, no. 3,
pp. 869–875, Jun. 2006.
[10] A.-B. Liu, P.-C. Hsu, Z.-L. Chen, and H.-T. Wu, “Measuring pulse wave
velocity using ECG and photoplethysmography,” J. Med. Syst., vol. 35,
no. 5, pp. 771–777, Oct. 2011.
[11] T. T. Selvan and M. S. Begum, “Nadi aridhal: A pulse based auto-
mated diagnostic system,” in Proc. 3rd Int. Conf. Electron. Comput.
Technol. (ICECT), Apr. 2011, pp. 305–308.
[12] P. Wide, F. Winquist, P. Bergsten, and E. M. Petriu, “The human-
based multisensor fusion method for artificial nose and tongue sensor
data,” IEEE Trans. Instrum. Meas., vol. 47, no. 5, pp. 1072–1077,
Oct. 1998.
[13] K. A. Reddy, B. George, N. M. Mohan, and V. J. Kumar, “A novel
calibration-free method of measurement of oxygen saturation in arterial
blood,” IEEE Trans. Instrum. Meas., vol. 58, no. 5, pp. 1699–1705,
May 2009.
[14] P. Wang, W. Zuo, and D. Zhang, “A compound pressure signal acquisi-
tion system for multichannel wrist pulse signal analysis,” IEEE Trans.
Instrum. Meas., vol. 63, no. 6, pp. 1556–1565, Jun. 2012.
WANG et al.: NOVEL MULTICHANNEL WRIST PULSE SYSTEM
[15] 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 Alternative Med., vol. 2012, pp. 1–9,
May 2011, Art. ID 745127.
[16] S. Charbonnier, S. Galichet, G. Mauris, and J. P. Siche, “Statistical and
fuzzy models of ambulatory systolic blood pressure for hypertension
diagnosis,” in Proc. 16th IEEE Instrum. Meas. Technol. Conf. (IMTC),
May 1999, pp. 103–108.
[17] S. Ahmad, M. Bolic, H. Dajani, V. Groza, I. Batkin, and S. Rajan, “Mea-
surement of heart rate variability using an oscillometric blood pressure
monitor,” IEEE Trans. Instrum. Meas., vol. 59, no. 10, pp. 2575–2590,
Oct. 2010.
[18] M. Saito, M. Matsukawa, T. Asada, and Y. Watanabe, “Noninvasive
assessment of arterial stiffness by pulse wave analysis,” IEEE Trans.
Ultrason., Ferroelectr., Freq. Control, vol. 59, no. 11, pp. 2411–2419,
Nov. 2012.
[19] P. Dupuis and C. Eugène, “Combined detection of respiratory and
cardiac rhythm disorders by high-resolution differential cuff pres-
sure measurement,” IEEE Trans. Instrum. Meas., vol. 49, no. 3,
pp. 498–502, Jun. 2000.
[20] J. M. Kang, T. Yoo, and H. C. Kim, “A wrist-worn integrated health
monitoring instrument with a tele-reporting device for telemedicine and
telecare,” IEEE Trans. Instrum. Meas., vol. 55, no. 5, pp. 1655–1661,
Oct. 2006.
[21] H. M. Haqqani, J. B. Morton, and J. M. Kalman, “Using the
12-lead ECG to localize the origin of atrial and ventricular tachycardias:
Part 2—Ventricular tachycardia,” J. Cardiovascular Electrophysiol.,
vol. 20, no. 7, pp. 825–832, 2009.
[22] K. Humphreys, T. Ward, and C. Markham, “Noncontact simultane-
ous dual wavelength photoplethysmography: A further step toward
noncontact pulse oximetry,” Rev. Sci. Instrum., vol. 78, no. 4,
pp. 044304-1–044304-6, 2007.
[23] C.-C. Tyan, S.-H. Liu, J.-Y. Chen, J.-J. Chen, and W.-M. Liang,
“A novel noninvasive measurement technique for analyzing the pressure
pulse waveform of the radial artery,” IEEE Trans. Biomed. Eng., vol. 55,
no. 1, pp. 288–297, Jan. 2008.
[24] L. Xu, D. Zhang, K. Wang, and L. Wang, “Arrhythmic pulses detec-
tion using Lempel–Ziv complexity analysis,” EURASIP J. Adv. Signal
Process., vol. 2006, p. 018268, Mar. 2006.
[25] K. Wang, C. Xu, D. Zhang, and N. Li, “Analysis of Doppler ultrasound
blood flow signals of wrist radial artery for discriminating healthy people
from 3 kinds of patients,” in Proc. 21st IEEE Int. Symp. Comput.-Based
Med. Syst. (CBMS), Jun. 2008, pp. 87–89.
[26] W. C. Tang and H. J. Sun, “The detected method of multipath of pulse
conditions and research of transducer,” Chin. J. Traditional Med. Sci.
Technol., vol. 7, no. 5, pp. 319–320, 2000.
[27] L. Xu, K. Wang, and D. Zhang, “Modern research on traditional Chinese
pulse diagnosis,” Eur. J. Oriental Med., vol. 4, no. 6, pp. 46–54, 2004.
[28] L. Xu, K. Wang, D. Zhang, Y. Li, Z. Wan, and J. Wang, “Objecti-
fying researches on traditional Chinese pulse diagnosis,” Inform. Med.
Slovenica, vol. 8, no. 1, pp. 56–63, 2003.
[29] X. Lisheng, W. Kuanquan, D. Zhang, and S. Cheng, “Adaptive baseline
wander removal in the pulse waveform,” in Proc. 15th IEEE Symp.
Comput.-Based Med. Syst. (CBMS), Jun. 2002, pp. 143–148.
[30] A. J. Joshi, S. Chandran, V. K. Jayaraman, and B. D. Kulkarni,
“Multifractality in arterial pulse,” in Proc. 19th Int. Conf. Pattern
Recognit. (ICPR), Dec. 2008, pp. 1–4.
[31] K.-Q. Wang, L.-S. Xu, D. Zhang, and C. Shi, “TCPD based pulse
monitoring and analyzing,” in Proc. Int. Conf. Mach. Learn. Cybern.,
2002, pp. 1366–1370.
[32] M. Nitzan, “Automatic noninvasive measurement of arterial blood pres-
sure,” IEEE Instrum. Meas. Mag., vol. 14, no. 1, pp. 32–37, Feb. 2011.
[33] Y. Hata, Y. Kamozaki, T. Sawayama, K. Taniguchi, and H. Nakajima,
“A heart pulse monitoring system by air pressure and ultrasonic sensor
systems,” in Proc. IEEE Int. Conf. Syst. Syst. Eng. (SoSE), Apr. 2007,
pp. 1–5.
[34] K. Wang, L. Xu, Z. Li, D. Zhang, N. Li, and S. Wang, “Approximate
entropy based pulse variability analysis,” in Proc. 16th IEEE Symp.
Comput.-Based Med. Syst., Jun. 2003, pp. 236–241.
[35] G. C. Jin, M. Yu, and N. K. Bao, “Research of multi-point pulse wave
computer measurement system using PVDF,” J. Tsinghua Univ. Sci.
Technol., vol. 39, no. 8, pp. 117–120, 1999.
[36] M. R. Ram, K. V. Madhav, E. H. Krishna, N. R. Komalla, and
K. A. Reddy, “A novel approach for motion artifact reduction in PPG
signals based on AS-LMS adaptive filter,” IEEE Trans. Instrum. Meas.,
vol. 61, no. 5, pp. 1445–1457, May 2012.
2033
[37] K. Q. Wang, L. S. Xu, L. Wang, Z. G. Li, and Y. Z. Li, “Pulse base-
line wander removal using wavelet approximation,” in Proc. Comput.
Cardiol., Sep. 2003, pp. 605–608.
[38] L. Xu, D. Zhang, and K. Wang, “Wavelet-based cascaded adaptive filter
for removing baseline drift in pulse waveforms,” IEEE Trans. Biomed.
Eng., vol. 52, no. 11, pp. 1973–1975, Nov. 2005.
[39] L. Xu, K. Wang, and D. Zhang, “Modern researches on pulse waveform
of TCPD,” in Proc. IEEE Int. Conf. Commun., Circuits, Syst. West Sino
Expo., Jun./Jul. 2002, pp. 1073–1077.
[40] L. S. Xu, K. Q. Wang, and L. Wang, “Pulse waveforms classification
based on wavelet network,” in Proc. IEEE-EMBS 27th Annu. Int. Conf.
Eng. Med. Biol. Soc., Jan. 2006, pp. 4596–4599.
[41] L. Wang, K.-Q. Wang, and L.-S. Xu, “Recognizing wrist pulse wave-
forms with improved dynamic time warping algorithm,” in Proc. Int.
Conf. Mach. Learn. Cybern., Aug. 2004, pp. 3644–3649.
[42] M. F. O’Rourke, A. Pauca, and X.-J. Jiang, “Pulse wave analysis,” Brit.
J. Clin. Pharmacol., vol. 51, no. 6, pp. 507–522, 2001.
[43] L. Wang, K. Q. Wang, and L. S. Xu, “Lempel–Ziv decomposition based
arrhythmic pulses recognition,” in Proc. 27th Int. Conf. IEEE Eng. Med.
Biol. Soc., Shanghai, China, Sep. 2005, pp. 4606–4609.
[44] L. Wang, K. Q. Wang, and L. S. Xu, “Lempel–Ziv decomposition based
arrhythmic pulses recognition,” in Proc. IEEE-EMBS 27th Annu. Int.
Conf. Eng. Med. Biol. Soc., Jan. 2006, pp. 4606–4609.
[45] M. R. Ram, K. V. Madhav, E. H. Krishna, N. R. Komalla, K. Sivani,
and K. A. Reddy, “ICA-based improved DTCWT technique for MA
reduction in PPG signals with restored respiratory information,” IEEE
Trans. Instrum. Meas., vol. 62, no. 10, pp. 2639–2651, Oct. 2013.
[46] K. A. Reddy, B. George, and V. J. Kumar, “Use of Fourier series
analysis for motion artifact reduction and data compression of photo-
plethysmographic signals,” IEEE Trans. Instrum. Meas., vol. 58, no. 5,
pp. 1706–1711, May 2009.
[47] P.-Y. Zhang and H.-Y. Wang, “A framework for automatic time-domain
characteristic parameters extraction of human pulse signals,” EURASIP
J. Adv. Signal Process., vol. 2008, p. 468390, Nov. 2008.
[48] Y. Chen, L. Zhang, D. Zhang, and D. Zhang, “Wrist pulse signal diagno-
sis using modified Gaussian models and fuzzy C-means classification,”
Med. Eng. Phys., vol. 31, no. 10, pp. 1283–1289, 2009.
[49] M. R. Neuman, “Measurement of blood pressure [tutorial],” IEEE Pulse,
vol. 2, no. 2, pp. 39–44, Mar./Apr. 2011.
[50] T. Kageyama, M. Kabuto, T. Kaneko, and N. Nishikido, “Accuracy of
pulse rate variability parameters obtained from finger plethysmogram:
A comparison with heart rate variability parameters obtained from
ECG,” J. Occupat. Health, vol. 39, no. 2, pp. 154–155, 1997.
[51] S. Lukman, Y. He, and S.-C. Hui, “Computational methods for tra-
ditional Chinese medicine: A survey,” Comput. Methods Programs
Biomed., vol. 88, no. 3, pp. 283–294, Dec. 2007.
[52] 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 Trans. Instrum.
Meas., vol. 61, no. 1, pp. 162–169, Jan. 2012.
[53] H.-T. Wu et al., “Arterial stiffness using radial arterial waveforms
measured at the wrist as an indicator of diabetic control in the elderly,”
IEEE Trans. Biomed. Eng., vol. 58, no. 2, pp. 243–252, Feb. 2011.
[54] I. B. Wilkinson et al., “Increased augmentation index and systolic stress
in type 1 diabetes mellitus,” QJM, vol. 93, no. 7, pp. 441–448, 2000.
Dimin Wang received the B.S. degree in
electronic engineering from Tsinghua University,
Beijing, China, in 2009, where he is currently
pursuing the Ph.D. degree with the Department of
Electronic Engineering.
He was a Research Assistant with the Department
of Computing, Hong Kong Polytechnic University,
Hong Kong, from 2010 to 2011. His current
research interests include signal processing, pattern
recognition, machine learning, and computerized
medical diagnosis.
2034 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 64, NO. 7, JULY 2015

David Zhang (F’09) received the M.Sc. degree in Guangming Lu received the B.S. Degree in electri-
computer science from Peking University, Beijing, cal engineering, the master’s degree in control theory
China, in 1982, the Ph.D. degree in computer sci- and control engineering, and the Ph.D. degree in
ence from the Harbin Institute of Technology (HIT), computer science and engineering from the Harbin
Harbin, in 1985, and the Ph.D. degree in electrical Institute of Technology (HIT), Harbin, China, in
and computer engineering from the University of 1998, 2000, and 2005, respectively.
Waterloo, Waterloo, ON, Canada, in 1994. He was a Post-Doctoral Fellow with Tsinghua Uni-
He was a Post-Doctoral Fellow with Tsinghua versity, Beijing, China, from 2005 to 2007. He is cur-
University, Beijing, and then an Associate Professor rently a Professor with the Biocomputing Research
with Academia Sinica, Beijing, from 1986 to 1988. Center, Shenzhen Graduate School, HIT. His current
He has been the Chair Professor with the Hong Kong research interests include pattern recognition, image
Polytechnic University, Hong Kong, since 2005, where he was the Founding processing, and automated biometric technologies and applications.
Director of the Biometrics Research Centre supported by the Hong Kong
Special Administrative Region Government in 1998. He also serves as the
Visiting Chair Professor with Tsinghua University, and an Adjunct Professor
with Peking University, Shanghai Jiao Tong University, Shanghai, China,
HIT, and the University of Waterloo. He has selected as a 2014 Highly
Cited Researcher by Thomson Reuters. He has authored over 10 books,
over 300 international journal papers, and holds around 30 patents from the
U.S./Japan/Honk Kong/China.
Prof. Zhang is a Croucher Senior Research Fellow, Distinguished Speaker
of the IEEE Computer Society, and fellow of the International Association
for Pattern Recognition. He is the Founder and Editor-in-Chief of the Inter-
national Journal of Image and Graphics, a Book Editor of the International
Series on Biometrics (Springer), an Organizer of the International Conference
on Biometrics Authentication, and an Associate Editor of over 10 international
journals, including the IEEE transactions.