Available online at www.sciencedirect.

com

Measurement 41 (2008) 463470

www.elsevier.com/locate/measurement

Real-time ECG telemonitoring system design with mobile

phone platform
Cheng Wen a, Ming-Feng Yeh a, Kuang-Chiung Chang
a

a,*

, Ren-Guey Lee

Department of Electrical Engineering, Lunghwa University of Science and Technology, Tauyan 33306, Taiwan
b
Department of Electronic Engineering, National Taipei University of Technology, Taipei 10608, Taiwan
Received 20 August 2006; received in revised form 19 December 2006; accepted 20 December 2006
Available online 17 January 2007

Abstract
In this paper, we propose an ECG (electrocardiogram) telemonitoring system based on a mobile phone platform. It
transmits abnormal heartbeats, which are identied in the patient-worn unit (Holter), in real time by using MMS (multimedia messaging service) on GPRS (general packet radio service) and transfers all ECG data acquired and stored in the
Holter by the Internet. By this way, the Holter can be used outdoors, and the communication cost can be reduced without
losing important information of patients. The GPS (global positioning system) information provided by the Holter can be
used to locate the patient for emergency help. To identify abnormal beats, we develop a real-time ECG classication algorithm that can be executed by the dual-core processor in the Holter. Experimental results show that the proposed system
achieves an ECG classication accuracy of 98.98%.
 2007 Elsevier Ltd. All rights reserved.
Keywords: Telemonitoring; Electrocardiogram (ECG); R wave detection; Real-time classication; Multimedia messaging service (MMS)

1. Introduction
Telemedicine, which refers to the utilization of
telecommunication technology for medical diagnosis, treatment, patient care, and remote monitoring,
is currently a signicant area of research and development. It is seen as an important way of reducing
medical costs by allowing for healthcare to be
administered outside a hospital setting. Telemedicine can also be used within a clinical institution.
One rapid growing area of telemedicine is the
*
Corresponding author. Tel.: +886 2 82093211x5501; fax:
+886 2 82099728.
E-mail address: kcchang@mail.lhu.edu.tw (K.-C. Chang).

long-term monitoring of patients with cardiovascular diseases at home. This is made possible with the
emergence of portable ECG (electrocardiogram)
telemonitoring systems, which are commonly used
to record the arrhythmia when it happens or to
record ECGs for allowing the experts to see various
trends. ECG telemonitoring systems can be divided
into two modes of operations: real-time mode, in
which patient data are available at the server end
immediately after acquisition, and store-and-forward mode, which involves accessing the data at a
later time. In both modes, the ECG data are transmitted via computer networks [13], public telephone networks [4], or cable TV networks [5] to
the server. In these system models, an expert is

0263-2241/$ - see front matter  2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.measurement.2006.12.006

464

C. Wen et al. / Measurement 41 (2008) 463470

expected at places where he/she can use a PC to

access the server for analyzing the ECG data, and
the patient is bounded at a xed place like home
or healthcare center where a PC is equipped for
transmitting the ECG. The use of wired network
connected PCs limits the degree of freedom of both
doctors and patients to move around.
To improve the mobility of the doctor, the GSM
(global system for mobile communication) mobile
telephony network was used for connecting the
server. The mobile phone has been recognized as a
possible tool for telemedicine since it became
commercially available. By using remote portable
devices such as notebooks and personal digital
assistants that wirelessly connected to a computer
network via GSM cellular phone, previous studies
successfully demonstrated the merging of cellular
communication and the Internet in telemedicine
[6,7]. The role of the mobile phone in these designs
was as a wireless modem for downloading data
from a server to a portable computer. A current
trend in telecommunication is the convergence of
wireless communication and computer network
technologies for providing mobile Internet access.
Recently, Hung and Zhang [8] implemented a
WAP (wireless application protocol) based telemonitoring system. It utilized WAP devices as mobile
access terminals and allowed doctors to browse
the monitored data on WAP devices in store-andforward mode.
The improvement on the mobility of the patient
is much less, comparing to the doctor. In many previous ECG telemonitoring systems, the patientworn unit, known as the portable digital Holter,
consisted of an ECG data acquisition circuit, an
A/D converter, and a storage unit with a capacity
being suciently large for 24 h of recording. To
provide a very limited mobility of the patient, the
Holter was equipped with an indoor, wireless transmitter for feeding the monitored data to a network
connected PC [4,8]. A GSM modem was equipped
with a PC for real-time transmission of ECG data
from a moving ambulance vehicle in [9]. The focus
of this design was on transmitting ECG data for
consultation while transporting patients in emergency cases. This is probably because use of a
GSM network is costly, and the data transmission
rate is low. Recently, Rasid and Woodward [10]
suggested a mobile telemonitoring system operating
in store-and-forward mode by using a Bluetoothenabled processor unit, which transmitted the monitored data to a Bluetooth mobile phone and subse-

quently via the GSM/GPRS (general packet radio

services) network to the server. On the other hand,
Engin et al. [11] used a mobile phone to transmit
the measured ECG signal in real-time mode. In
these designs [10,11], the mobility of the patient is
improved. However, the analysis of ECG is not performed in the place where the ECG is acquired, i.e.,
the ECG is analyzed at the server end. In fact, there
is a loss of eciency in the use of the GSM/GPRS
network because normal ECGs are also transferred,
which implies a high cost.
In this paper, we propose an ECG telemonitoring
system operating in both modes. In store-and-forward mode, the Holter records the ECG signal of
the patient continuously up to 48 h. The monitored
data is transmitted to the server through the Internet when a wired network is available. The Holter
also contains a software program performing realtime ECG classication. When an abnormal heartbeat that the doctor concerns is detected, the Holter
transmits it with the GPS (global positioning system) information to the server via MMS (multimedia messaging service) in real-time. The doctor at
the server side could communicate with the patient
also by using MMS if necessary. In the server, a
GIS (geographic information system) is resided for
locating the patient in an emergency case by using
the GPS information packaged in the MMS message. By this way, the Holter can be used outdoors,
and the cost for using the GPRS network is lowered
because only abnormal ECGs are transmitted.
The remainder of this paper is organized as follows. The system is described in Section 2. The proposed system consists of a Holter and a server. The
hardware and software designs of the Holter are
described in Section 3. The system has been implemented and tested. Experimental results are
reported in Section 4. Finally, Section 5 contains
some discussions and conclusions.
2. System description
Fig. 1 illustrates the proposed telemonitoring system. This system consists of a portable Holter, and a
server. The Holter is responsible for ECG data
acquisition, storage, and real-time analysis. In the
Holter, a software program is used to classify the
input ECGs into several beat types. In case that
the program classies a current ECG beat into a
beat type that the doctor concerns, the Holter transmits the ECG data with the GPS information to the
server by using MMS in real time.

C. Wen et al. / Measurement 41 (2008) 463470

Fig. 1. System model.

MMS is a new standard in mobile messaging. It

is a communications technology developed by
3GPP (third generation partnership project) that
allows users to exchange multimedia communications between capable mobile phones and other
devices. It is also possible to send MMS messages
from a mobile phone to an email address. MMS
can include not just text, but also sound, images
and video. An MMS message is a single entity,
not a collection of attachments, and has no size
limit. MMS requires a third generation network to
enable such large messages to be delivered, although
smaller messages can be sent even with second generation networks using GPRS. MMS conrms each
exchange of message by giving a conrmation message, as shown in Fig. 2. This property prevents the

465

loss of information and makes MMS a particularly

appropriate communication tool in the telemonitoring system design. Note that, in this design, the
communication between the MMS center and the
server is through TCP/IP protocol to the communication gateway equipped with the server.
When the server receives the monitored ECG
data, it stores the data and automatically streams
the data to authorized doctors for diagnosis. After
examining the ECG data, the doctor can send a
feedback MMS message to the user. The message
may contain medical advice and/or a list of control
commands to the Holter for resending the abnormal
ECG data, transferring a specied period of ECG
data for further examination, or tuning parameters
of the ECG analysis algorithm. The transmission
of ECG data in this mode can also be operated
manually. Whenever the user feels uncomfortable,
he/she can transfer his/her current ECG data to
the server for advice or a check up. Finally, by
incorporating the GPS information packaged in
the MMS message, the GIS resided in the server
can be used to locate the patient in an emergency
situation.
In store-and-forward mode, the Holter transfers
the entire long-term (up to 48 h) monitored data
package through the Internet to the server for comprehensive analysis. When receiving the package,
the server stores the data and streams the data to
authorized doctors when requested. Because the
design and implementation of the store-and-forward mode is quite easy, the focus will be mainly
on the real-time mode in the following.
3. Portable holter
3.1. Hardware
Fig. 3 shows the hardware block diagram of the
Holter, which is based on a mobile phone platform.

Fig. 2. Information ow of MMS starts from MM1_submit.REQ, the Holter sends a request for submission of an
MMS message to the MMS center, and ends up at MM1_delivery_report.RES, the Holter responds the delivery report from the
center. Information exchange in each step is conrmed by using
request and response messaging.

Fig. 3. Hardware of the Holter.

466

C. Wen et al. / Measurement 41 (2008) 463470

In the ECG signal acquisition module, a one-lead

(Lead II) ECG signal is detected by pre-jelled electrodes, fed into a passive AC coupling with a cuto
frequency of 0.01 Hz that highpass lters the voltage dierence to reduce baseline wander and amplier saturation, amplied by an instrumentation
amplier with a gain of 1000, and ltered by a second-order Bessel low-pass lter with a cuto frequency of 160 Hz for removing high frequency
noise or movement artifacts. The ECG signal is digitized by an A/D converter at 360 Hz with a resolution of 11 bits. After being digitized, the ECG data
is fed into a dual-core processor1 for analysis and
storage. This processor integrates a DSP with an
ARM (advanced RISC machine) and is widely used
in mobile phones. The analysis is done by the software program resided in the processor. When an
input ECG is classied into a beat type that the doctor concerns, the processor packages the ECG with
the information provided by the GPS module into
an MMS message and sends the message to the server through the GSM/GPRS module. The LCD
touch panel is used for the selection of the operation
functions shown on the screen and to display the
message from the doctor. The external ash memory of 1 G bytes is used for storing the ECG data,
record time, location of the R wave, and estimated
ECG beat type. This capacity is sucient for 48 h
of recording. The USB port is for the connection
with the PC, and the Ethernet port is for the connection with the Internet.
3.2. Software
The classication of ECG beats is a dicult
problem even with the aid of a computer, since
ECGs dier signicantly even for the same type
and for the same patient. Due to the limited computation power of the processor, the goal is to develop
a simple and eective software program that can be
executed in real time. Fig. 4 shows a typical ECG
beat, where the R wave is the rst positive (upward)
deection of the QRS complex. The software program resided in the processor consists of an R wave
detection routine and an ECG classication routine.
In the following, we will rst introduce the R wave
detection routine and then the classication routine.
The accurate detection of the R wave is the base
of the automatic ECG recognition. There has been
1

OMAP5910.

Fig. 4. ECG signal consists of three basic waves: P wave, QRS

complex, and T wave.

much work done in this area by PC-based methods

[1214]. So and Chan [14], developed a real-time R
wave detection algorithm, in which the QRS onset
point was selected by exceeding a self-adaptive slope
threshold in two successive values of slope and the
R wave was found by searching the rst highest
point following the onset. Their algorithm is simple
and eective and can be implemented in a microprocessor. An alternative adaptation rule for nding
the onset point was developed by our group [15]
and is used here. The real-time R wave detection
routine used is summarized as follows.
Routine 3.1
Step 1: In the rst 1500 ms, nd a maximum of the
dierentiated ECG data as M(0). Set the initial
threshold to the slope as T(0) = aM(0).
Step 2: When two successive values of slope that
exceed the threshold, the rst point is recognized
as the QRS onset point.
Step 3: The R wave is detected by searching in the
dierentiated ECG data the rst zero following
the onset point.
Step 4: The self-adaptive threshold is obtained by
T(n) = bM(n) + cT(n  1).
The three constants, 0 < a, b, c < 1, in Routine 3.1
depend on the lters used in the ECG signal acquisition circuit. In general, a small threshold has a
higher sensitivity to the detection of R wave than
a large threshold does, but a lower immunity to
noise. The current maximal slope M(n) is obtained
by searching after 200 ms of the M(n  1) in the
interval where the value of slope is greater than
the threshold T(n  1). Note that no R wave detection is allowed 200 ms after the current one is
detected in Step 3. In this time interval, a new value
T(n) in Step 4 is calculated and the beat classication algorithm described in the below is executed.

C. Wen et al. / Measurement 41 (2008) 463470

We use the QRS part of the ECG signal to

achieve beat classication. After detecting the R
wave, Routine 3.1 extracts the QRS complex out
from the beat by using a 73 samples (about
200 ms) window centered on the R wave and feeds
it to the classication routine. The window length
is selected to ensure that also very wide duration
PVCs (pre-mature ventricular contractions) can be
included. To avoid complicated computations, we
use the waveform of the QRS complex without
any feature extraction in the classication algorithm. The proposed beat classication algorithm
is based on our previous work [16], which followed
from the ART2 (adaptive resonance theory) neural
network initially introduced by Grossberg in 1976
[17,18]. Without preliminary training, ART networks not only allow category templates to adapt
to current circumstances, but also allow on-line creation of categories during clustering sessions. This
twofold exibility is very useful in ECG clustering
problems, because the ECGs may dier from session to session [19]. However, the clustering result
of the ART has to be annotated by a cardiologist
in order to achieve the purpose of classication.
This fact implies that ART networks cannot perform ECG classication in real time.
In order to classify heartbeats in real time, it is
required that a doctor assigns some typical beats
of the patient as initial templates, yj, of the proposed
method. The initialization will not cause any problem in clinical practice because there must exist
some ECG recordings of the patient before the doctor decides to long-term monitor the patient. Then,
input beats are compared with those templates as
follows.

Routine 3.2
Step 1: For an input beat, say x(k) = [x(1), . . . ,
x(73)], compare it with all templates, yj, by using
the similarity measure [20]:
gx; y j

n
X

wk  rxk; y j k

k1

where w(k) P 0 is the weighting function with

P
n
k1 wk 1. In the above equation,
rxk; y j k

d
Dj k d

467

where Dj(k) = jx(k)  yj(k)jand d > 0 is a constant.

The winner, denoted by j*, is the one yielding the
largest measurement maxj{g(x, yj)}.
Step 2: The neuron j* passes the vigilance test if and
only if gx; y j P q; where 0 < q 6 1 is the vigilance
threshold. In this case, the input is classied into the
cluster that the winner belongs, and the weight
vector of the winner j* is adapted by
old

new

y j

old

Numclusterj  y j
old
Numclusterj

where Numclusterj denotes the number of members in clusterj .

Step 3: If the winner fails the vigilance test, a new
neuron unit j is created with the input as its template, i.e., yj = x.
Adaptation rule (3) nds the center of gravity of
a cluster and does not suer from selecting an
appropriate learning rate. The most important feature of the similarity measure (1) for the selection
of the vigilance threshold is that we can always
guarantee that 0 < g(xi, yj) 6 1 no matter how large
the input data is. As a consequence, the vigilance
threshold q is limited to the range (0, 1]. In general,
a larger vigilance threshold value may result in
higher classication accuracy but more clusters
(neurons) than a smaller value does. The more the
clusters, the much heavy the computational load.
It is required that Routine 3.2 classies an input
beat within 200 ms, the time interval that Routine
3.1 stops to detect the next R wave, while the processor reads an ECG data every 2.78 ms (360 Hz
sampling rate), calculates T(n) (Step 4, Routine
3.1), stores the classication result, reports an
abnormal beat to the server if necessary, and
receives an MMS message from the server that
may contain a list of control commands. As a consequence, a limit on the number of clusters is needed
to avoid the overload of the processor in the realtime classication procedure.
When a new neuron is on-line created in Step 3,
the template of the new neuron is sent to the doctor.
The doctor labels the beat type of the template and
sends it back for achieving real-time classication.
A false alarm occurs when a normal beat is misidentied as an abnormal beat or an abnormal beat is
misclassied into a cluster dominated by a dierent
abnormal type of beats. In either case, to improve
the classication result, the doctor can on-line
adjust the vigilance threshold value, create a new

468

C. Wen et al. / Measurement 41 (2008) 463470

neuron with an assigned template, or delete an existing neuron if it causes false alarms frequently by
sending an MMS message to the Holter.
4. Experimental results
The design of the Holter has been implemented in
our laboratory by using o-the-shelf components. A
photograph of the prototype, measuring 90 mm
65 mm 35 mm, is shown in Fig. 5. The software
was developed in JAVA language using CLDC
(connected limited device conguration) 1.0 and
MIDP (mobile information device prole) 2.0 on a
Linux OS. For a preliminary test of the real-time
ECG classication, a volunteer carried the Holter
for 24 h. The number of initial templates (neuron
units) selected from the previous recording of the
volunteer was 40 in Routine 3.2. Because the volunteer had no heart diseases, the number of nal templates remained the same. During the 24 h, no
overload was observed. Fig. 6 shows the result of
the real-time R wave detection, in which the R
waves are marked by circles.
Next, we used the MIT/BIH Arrhythmia Database [21], sampled at 360 Hz with 11 bits resolution,
to test the real-time ECG classication algorithm.
This database contains a wide variety of QRS complex morphologies and dierent types of noise and
artifacts. It consists of 48 records of 30 min, and

Fig. 5. The prototype of the Holter is composed of three doubleside PCBs. PCB1 includes an LCD module and a DC-DC yback converter for supplying a higher voltage for the display.
PCB2 includes a port for the connection with the ECG signal
acquisition module (not shown), an ADC, a dual-core processor,
an external NAND ash memory of 1 G bytes, and an Ethernet
chip. PCB3 includes a power-management circuit, a GSM/GPRS
module, and a GPS module.

Fig. 6. R wave detection result.

each of them has two channels (leads) of ECG signals. Accompanying each record is an annotation
le in which the R wave location and beat type of
each ECG beat has been labeled by expert cardiologist annotators. Excluding records containing the
paced beats (record 102, 104, 107, and 217), we used
44 records of the MIT/BIH database. Raw data
from channel 1 (with some beats missing) were used
in this experiment for detecting arrhythmias. To
simulate the eect of the lters in the ECG signal
acquisition module, we used digital lters to lter
the raw data in a PC. Then, the pre-ltered data
was stored in the external memory of the Holter.
The processor read a sampled data every 1/360 s
for real-time analysis.
The AAMI recommended practice [22] was used
to combine the MIT/BIH beat types into four classes. Class V contained ventricular ectopic beats,
class F contained fusion of ventricular and normal
beats, class Q contained unclassiable beats, and
class N contained normal beats, bundle branch
block beats, supraventricular ectopic beats, and
other beats that do not fall into the V, F, or Q classes. Accordingly, in the algorithm, a class might
contain several subclasses (beat types).
The beat classication was done record by
record. Each record had its own initial templates
selected from that record. In addition to the four
classes, a special cluster, named cluster R, was preset without a template, i.e., this cluster was not used
to classify input beats, but for collecting beats
whose R wave were incorrectly located by Routine
3.1. For each record, the number of nal templates
was limited to 40, according to the result of the preliminary test. In the test, when an input beat was
classied into an abnormal (V, F, and Q) class or
a new cluster was created on-line, the Holter
reported it to the server. To simulate the function
of the doctor, the R wave location and beat type
of each ECG beat provided by the MIT/BIH data-

C. Wen et al. / Measurement 41 (2008) 463470

base was used in the server. When the server

received a message, it rst checked the correctness
of the R wave location found by the Holter. In case
that the R wave location was incorrect, the server
labeled the current ECG beat as an R beat and
sent the answer back to the Holter. If the R wave
location was correct and beat type was blank, which
meant that a new cluster was created, the server
labeled the ECG beat according to its annotation
and sent it back for labeling the new cluster in the
algorithm. When the beat type was labeled, the server checked the correctness of the beat type. If the
beat type was right, the server recorded an alarm.
If it was wrong, the server corrected the beat type,
recorded a false alarm, and sent it back to the Holter for creating a new cluster. Note that the on-line
adjustment of the vigilance threshold value was not
tested in this experiment; instead a xed value for
each record was used (see Table 2).
Over the total 101,371 testing beats, there are 514
beats in 44 R clusters (one for each record). The
MIT/BIH database annotation le is used to compare beats outside the R clusters. The accuracy is
100%. This means the overall accuracy of the R
wave detection procedure is about 99.5%. For those
beats outside the R clusters, Table 1 shows the classication result in terms of a confusion matrix
divided into the dierent classes. The main diagonal
shows the number of correctly classied beats.
From Table 1, we see that, in class N, there are
92,456 beats classied correctly, but 470 are classied incorrectly. The reason why there are no beats
in N being misclassied into V or F is that the server
corrects those beats. On the other hand, since the
Holter is designed to report abnormal beats, there
is no chance to correct an abnormal beat when it
is misclassied into N. As a consequence, Routine
3.2 misclassies 469 beats (253 in V, 154 in F, and
62 in Q) into N. There are 684 false alarms during
the entire classication procedure. Among which,
Table 1
Classication result

n
v
f
q

92,412
0
0
0

253
7189
0
0

154
0
642
0

62
0
0
145

The results from the individual records are added together. The
annotation classes, N, V, F, and Q, are depicted on the horizontal
axis, whereas the resulting classes, n, v, f, and q, are shown
vertically.

469

514 false alarms are caused by the incorrect detection of the R wave, 111 are caused by reporting that
new clusters were on-line created, and 59 are caused
by misclassication. The total number of beats in
the algorithm created clusters is 15,281, and number
Table 2
Details of the classication
Record
#

Initial
templates

Final
templates

Accuracy

100
101a
103a
105a
106
108a
109
111a
112a
113
114a
115a
116
117a
118a
119
121a
122
123
124
200a
201a
202
203a
205
207a
208a
209
210a
212a
213
214a
215a
219
220
221a
222a
223a
228a
230a
231
232a
233
234a

0.78
0.78
0.75
0.77
0.82
0.93
0.90
0.82
0.77
0.84
0.88
0.81
0.82
0.80
0.89
0.86
0.76
0.79
0.81
0.89
0.89
0.86
0.84
0.80
0.89
0.92
0.90
0.77
0.88
0.78
0.88
0.88
0.88
0.88
0.79
0.90
0.79
0.87
0.88
0.79
0.87
0.77
0.87
0.79

1N
1N1Q
1N
1N1V
1N2V
7N3V2Q
4N3V1F
4N
1N
2N
1N1V
1N
1N1V
2N
2N1V2Q
1N1V
2N
1N
2N
1N1V1F
1N2V
4N
1N1V
1N1V
1N2V
3N12V
2N3V3F
1N
1N2V
2N
1N2V2F
1N6V
1N4V
2N3V
1N
1N4V
2N
1N4V
1N4V
2N
1N1V2Q
1N
1N4V
1N

1N1F
1N1Q
2N
19N1V2Q
1N2V
12N10V1F2Q
8N6V1F
4N1V
3N
2N
1N1V
2N
3N2V
2N
4N2V2Q
1N2V
5N1V
1N
2N1V
1N2V2F
1N5V
7N4V1F3Q
1N1V1F
8N4V
1N4V
3N13V
6N12V5V
3N1V
3N2V
2N
2N8V14F
10N7V
1N5V
2N3V1Q
1N
1N8V
5N
5N14V3F
1N4V
3N1V
1N1V2Q
5N
1N4V
1N1V

100.00
100.00
100.00
99.53
99.65
99.49
99.96
100.00
100.00
100.00
99.73
100.00
99.96
100.00
99.91
100.00
100.00
100.00
100.00
99.69
98.19
99.75
99.86
95.87
99.59
99.01
97.16
100.00
98.45
100.00
94.37
96.06
99.76
97.20
100.00
99.96
100.00
88.68
99.85
100.00
100.00
100.00
99.45
100.00

This table shows the vigilance threshold value q, the number of

initial templates, the number of nal templates for each selected
record, in which xNyV, for example, denotes that the number of
templates in class N is x, and in V is y, and the classication
accuracy of the entire algorithm including R wave detection.
a
This record is noisy.

470

C. Wen et al. / Measurement 41 (2008) 463470

of beats in the doctor created clusters is 570. The

total number of misclassied beats is 253 + 154 +
62 + 59 = 528, which corresponds a classication
accuracy of 99.48%. Including the number of beats
in R clusters, the entire classication accuracy is
98.98%. Table 2 shows the details of the classication. It is interesting to know the performance of
the proposed design against noise and movement
artifacts. From Table 2, we see that there are 27
noisy records containing 62,708 beats, and 17 clean
records containing 38,663 beats. The algorithm
including Routines 3.1 and 3.2 achieves an overall
accuracy of 99.24% for clean records, and 98.81%
for noisy records. The two accuracy values dier
not much from each other. This shows that the proposed algorithm is robust against noise and movement artifacts.
5. Conclusion
We have presented the design and implementation of an ECG telemonitoring system, in which
the real-time analysis of ECG was performed in the
place where the ECG was acquired. Experimental
results showed that the proposed system could
achieve ECG classication in real time with an overall classication accuracy of 98.98%. During the 22 h
(30 min per record) of the experiment, the Holter
sent 8493 messages (about 8.38% of the testing beats)
including 652 false alarms. When comparing with the
method of using the GSM/GPRS network continuously for real-time transmission of ECG, the communication cost of the proposed design is reduced.
References
[1] A.I. Hernandez, F. Mora, G. Villegas, G. Passariello, G.
Carrault, Real-time ECG transmission via Internet for
nonclinical applications, IEEE Transactions on Information
Technology in Biomedicine 5 (3) (2001) 253257.
[2] D. Reske, Z. Moussavi, Design of a web-based remote heartmonitoring system, in: Proceedings of the Second Joint
EMBS/BMES Conference, Houston, 2002, pp. 18471848.
[3] S. Guillen, M.T. Arredondo, V. Traver, J.M. Graca, C.
Fernandez, Multimedia telehomecare system using standard
TV set, IEEE Transactions on Biomedical Engineering 49
(12) (2002) 14311437.
[4] J. Bai, Y. Zhang, D. Shen, L. Wen, C. Ding, Z. Cui, F. Tian,
B. Yu, B. Dai, J. Zhang, A portable ECG and blood pressure
telemonitoring system, IEEE Engineering in Medicine and
Biology Magazine 18 (1999) 6370.
[5] R.G. Lee, H.S. Chen, C.C. Lin, K.C. Chang, J.H. Chen,
Home telecare system using cable television plantsAn
experimental eld trial, IEEE Transactions on Information
Technology in Biomedicine 4 (1) (2000) 3744.

[6] L.G. Yamamoto, D.R. Williams, A demonstration of instant

pocket wireless CT teleradiology to facilitate stat neurosurgical consultation and future telemedicine implications, American Journal of Emergency Medicine 18 (4) (2000) 423426.
[7] L.G. Yamamoto, L.K. Shirai, Instant telemedicine ECG
consultation with cardiologists using pocket wireless computers, American Journal of Emergency Medicine 19 (3)
(2001) 248249.
[8] K. Hung, Y.T. Zhang, Implementation of a WAP-based
telemedicine system for patient monitoring, IEEE Transactions on Information Technology in Biomedicine 7 (2) (2003)
101107.
[9] S. Pavlopoulos, E. Kyriacou, A. Berler, S. Dembeyiotis, D.
Koutsouris, A novel emergency telemedicine system based
on wireless communication technology AMBULANCE,
IEEE Transactions on Information Technology in Biomedicine 2 (4) (1998) 261267.
[10] M.F.A. Rasid, B. Woodward, Bluetooth telemedicine processor for multichannel biomedical signal transmission via
mobile cellular networks, IEEE Transactions on Information
Technology in Biomedicine 9 (1) (2005) 3543.
[11] M. Engin, E. Caglav, E.Z. Engin, Real-time ECG signal
transmission via telephone network, Measurement 37 (2005)
167175.
[12] J. Pan, W.J. Tompkins, A real-time QRS detection algorithm, IEEE Transactions on Biomedical Engineering 32
(1985) 230236.
[13] G. Friesen, T.C. Jannett, M.A. Jadallah, S.L. Yates, S.R.
Quint, H.T. Nagle, A comparison of the noise sensitivity of
nine QRS detection algorithm, IEEE Transactions on
Biomedical Engineering 37 (1) (1990) 8598.
[14] H.H. So, K.L. Chan, Development of QRS detection
method for real-time ambulatory cardiac monitor, in:
Proceedings of the 19th Annual International Conference
of the IEEE in Engineering in Medicine and Biology society,
Chicago, 1997, pp. 289292.
[15] R.G. Lee, I.C. Chou, C.C. Lai, M.H. Liu, M.J. Chiu, A
novel QRS detection algorithm applied to the analysis for
heart rate variability of patients with sleep apnea, Journal of
Biomedical EngineeringApplications, Basis & Communication 7 (5) (2005) 4448.
[16] M.F. Yeh, S.S. Chiang, GreyART network for data clustering, Neurocomputing 67 (2005) 313320.
[17] S. Grossberg, Adaptive pattern classication and universal
recoding. I: Feedback, expectation, olfaction, and illusions,
Biological Cybernetics 21 (1976) 145159.
[18] S. Grossberg, Adaptive pattern classication and universal
recoding. II: Feedback, expectation, olfaction, and illusions,
Biological Cybernetics 23 (1976) 187202.
[19] M. Femandez-Delgado, S.B. Ameneiro, MART: A multichannel ART-based neural network, IEEE Transactions on
Neural Networks 9 (1) (1998) 139150.
[20] K.C. Chang, R.G. Lee, C. Wen, M.F. Yeh, Comparison of
similarity measures for clustering electrocardiogram complexes, in: Computers in Cardiology, Lyon, 2005, pp. 759762.
[21] MIT/BIH database distributor, Beth Israel Hospital, Biomedical Engineering, Division available in the address:
http://ecg.mit.edu/, USA, 1979.
[22] Recommended practice for testing and reporting performance results of ventricular arrhythmia detection algorithms, Association for the Advancement of Medical
Instrumentation, 1987.