Wireless Sensor Network, 2009, 1, 276-283
doi:10.4236/wsn.2009.14034 Published Online November 2009 (http://www.scirp.org/journal/wsn).
Copyright © 2009 SciRes. WSN
Real-Time Automatic ECG Diagnosis Method
Dedicated to Pervasive Cardiac Care
Haiying ZHOU1, Kun-Mean HOU2, Decheng ZUO1
1School of Computer Science & Technology, Harbin Institute of Technology , Har bin, China
2LIMOS Laboratory UMR 6158 CNRS, University of Blaise Pascal, Clermont-Ferrand, France
Email: {haiyingzhou, zdc}@hit.edu. cn, kun-mean.hou@isima.fr
Received May 1, 2009; revised May 25, 2009; accepted May 31, 2009
Recent developments of the wireless sensor network will revolutionize the way of remote monitoring in dif-
ferent domains such as smart home and smart care, particularly remote cardiac care. Thus, it is challenging to
propose an energy efficient technique for automatic ECG diagnosis (AED) to be embedded into the wireless
sensor. Due to the high resource requirements, classical AED methods are unsuitable for pervasive cardiac
care (PCC) applications. This paper proposes an embedded real-time AED algorithm dedicated to PCC sys-
tems. This AED algorithm consists of a QRS detector and a rhythm classifier. The QRS detector adopts the
linear time-domain statistical and syntactic analysis method and the geometric feature extraction modeling
technique. The rhythm classifier employs the self-learning expert system and the confidence interval method.
Currently, this AED algorithm has been implemented and evaluated on the PCC system for 30 patients in the
Gabriel Monpied hospital (CHRU of Clermont-Ferrand, France) and the MIT-BIH cardiac arrhythmias da-
tabase. The overall results show that this energy efficient algorithm provides the same performance as the
classical ones.
Keywords: Pervasive Cardiac Care, Automatic ECG Diagnosis, QRS detector, Rhythm Classifier, Wireless
Sensor Networks
1. Introduction
Due to the increasing occurrence of sudden death events
caused by cardiovascular diseases, there is a need to pro-
vide a long-term, real-time continuous PCC service for
the sudden death high-risk population. The PCC system
has thus been developed for different populations at a
variety of environment, including at home, clinical and
The studies of AED methods focused mainly on the
clinical services. Unlike the clinical applications, the
acquisitions of the PCC system is ambulatory ECG sig-
nal that is non-stationary and easy-disturbed by interfer-
ences. Moreover, the nodes of the PCC system have
strict resource constraints, i.e. the capacities of computa-
tion, storage and power supply. Classical AED algo-
rithms are thus unfit for the PCC system.
This paper presents a real-time and low resource con-
sumption AED algorithm for the PCC system. Section 2
introduces the state-of-the-art of the AED algorithms.
Section 3 describes this algorithm in detail and section 4
presents the performance evaluation. The conclusions are
drawn at the last section.
2. State-of-the-Art
Due to its high potential amplitude, steep slope (R-wave)
and wide duration, QRS complex is generally used for
the cardiac event diagnosis and analysis. Different AED
algorithms are classified by Köhler et al. [1]: 1). Time-
domain analysis can implement a simple and rapid detec-
tion but it is noise-sensitive; 2). Wavelet transform
analysis has high detection performance but has huge
computation overhead; 3). Syntax analysis exposes the
wave pattern elements and their mutual relations, but it is
noise-sensitive and has huge computations; 4). Neural
network analysis needs a large amount of training sample
set and long training time.
Supported by Doctoral Fund of Youth Scholar of Ministry of Education
of China (No.200802131024), French Program of Cooperation with
China (No.20974WG), and Scientific Research Fund of Returned
Oversea Scholars of Harbin city of China (No.RC2009LX010001).
H. Y. ZHOU ET AL. 277
Other classical AED techniques include: template
matching [2], hidden Markov model [3], Hilbert trans-
form [4], mathematical morphology [5] method, etc.
These techniques generally have huge computation
overhead. The new AED algorithms generally integrate
multiple techniques. For example, Oliveira et al. [6] in-
tegrates the Hilbert transfor m and wavelet transform, and
Szilágyi et al. [7] combines the neural network, wavelet
transform and genetic algorithm techniques. Generally,
these hybrid methods can improve the detection accuracy,
but have huge computation overhead, more resource
consumption and less operation efficiency.
3. AED Algorithm
3.1. Signal Preprocessing and Conditioning
Due to the non-stationary and easy-disturbed natures of
the ambulatory ECG signals, the acquisitions of PCC
system must be de-noised before making detection. Most
of artifacts, such as baseline shift, electrical noise and
muscle tremor interference, can be effectively eli minated
or reduced by choosing suitable filters. In th is subsection,
we present the filter techniques.
3.1.1. ECG Time Series
There are three ECG signals series, i.e. R(t), AD(t) and
RC(t), in our algorithm. The R(t) series is the raw ECG
signals acquired from electrodes. It’s generally contami-
nated by different kinds of noises. The AD(t) series is the
adaptive differential signals with the processing of the
differential filter and the adaptive filter. The inferences
of the baseline drift and the motion artifacts can be
eliminated in the AD(t) series; hence this series is used to
detect and to localize the QRS complexes. The RC(t)
series is the de-noised ECG signals with the operations
of the band-pass filter and the linear amplifier. Since the
electrical noises and the muscle tremors have been re-
moved from the RC(t) series, hence it is used to extract
the characteristics of the QRS complexes.
3.1.2. Adaptive Filter
The classical filter for the ECG series, e.g. Notch filter,
low-pass filter, and high-pass filter, can effectively re-
move or reduce most of the interferences. But for the
motion artifacts, because of their irregular occurrences
and irregular morphological attributes, these filters can-
not eliminate these disturbances. These artifacts can
make greatly troubles in QRS detection when encoun-
tering QRS-like artifacts.
This algorithm adopts an adaptive filter to reduce mo-
tion artifacts. The resultant signal series, named A(t), are
generated by performing AT operation in the raw series
R(t). The expression of adaptive filter is
s the balance coefficient that is relative to the
signal sample frequency (default 0.95).
Figure 1 shows different ECG series. Figure 1(a) is the
raw signals R(t) which are serious polluted by noises.
Figure 1(b) represents the reconstructed series RC(t)
when filtering the R(t) series by the classical filters, i.e.
Notch filter, low-pass filter and high-pass filter. The
RC(t) series still contain the interferences generally
caused by baseline wandering and motion artifacts.
Figure 1(c) is the adaptive filter signal A(t) when filter-
ing the R(t) series by the adaptive filter, which has better
signal quality than RC(t). Figure 1(d) is the reconstructed
signal RC*(t) based on the adaptive filter signal A(t).
Obviously, in contrast to the previous reconstructed sig-
nal RC(t), the signal RC* (t) has better signal quality in
which the motion artifacts are effectively eliminated.
3.2. QRS Complex Detection
This paper presents a new QRS detector which copes
with noises, artifacts and variability of ECG morphology
by exploiting a self-adaptive threshold method (SAT),
and a particular state transition recognition procedure
(STR). The SAT method is used to estimate the peaks of
ECG sub-segments and the means of contextual thresh-
olds, which allows estimating the optimum thresholds in
segment space. The STR procedure traces the waveform
changes of signal series and identifies QRS complexes
based on the optimum thresholds and the rules of state
3.2.1. Di agnostic Segment Window (DSW)
A short-term redundant data (default 5 seconds) is im-
portant in QRS detection. Firstly, this short-term segment
05001000 1500 2000 250
3500 Data source: MIT−BIH 104
a: R(t)
05001000 1500 2000 250
b: RC(t)
05001000 1500 2000 250
c: A(t)
05001000 1500 2000 250
d: A’(t)
Figure 1. Filtered ECG signal series.
Copyright © 2009 SciRes. WSN
278 H. Y. ZHOU ET AL.
enables the complex contextual correlative analysis and
reduces the interferences of baseline drift. In view of the
low-frequency baseline drift, a short-term segment has
fewer disturbances caused by baseline wandering than
long-term signals. The redundant data enable the QRS
detector to identify current QRS complex by comparing
with the fore-and-aft QRS complexes. Furthermore, in
views of the unpredictability and variability of network
quality, the redundancy is necessary for the data retrans-
mission and the network communication.
3.2.2. Se lf-Adapti ve T h reshold (SA T)
The QRS waveforms in AECG have rapid changes and
high potential amplitudes so that the differential series
D(t) can exactly represent the changes. The QRS signals
have higher absolute amplitudes in a cardiac cycle of D(t)
series. The goal of QRS detection is to search the opti-
mum pair-peak for each QRS complex, i.e. the positive
and negative peaks of a cardiac cycle. In DSW, there are
generally multiple pair-peaks because several heart beats
will occur during the 5s length. These pair-peaks make
up of a pair-peak group in a DSW. Based on the pair-
threshold obtained from the pair-peaks group of a DSW,
the STR procedure is then able to locate QRS complexes.
The absolute amplitude of each peak is generally greater
than the associated absolute threshold in D(t). Further-
more, since the offset of location between the D(t) series
and the A(t) series is constant, we can thus obtain the
positions of QRS complexes in A(t) by locating the com-
plexes in D(t).
The SAT method aims to determine the optimum pair-
threshold, which is estimated from two aspects: the mean
of the pair-peak group of DSW and the pair-threshold of
the previous DSW. The pair-threshold results from the
means of the negative and positive pair-peaks group of
DSW. In order to accurately estimate these pair-peaks,
the diagnostic segment window is divided into 5 sub-
segments with the length of one second (see in Figure 2).
Because the normal heart rate of a healthy adult is
60bmp-100bmp [8], each sub-segment thus contains one
Figure 2. Mean of pair-peak group in diagnostic window.
heart beat. Since the differential signals of QRS complex
have the maximum absolute amplitudes in a cardiac cy-
cle, a pair-peak will represent a QRS complex and then
can be used to estimate the thresholds. Furthermore, the
shorter of sub-segment is, the less interference of base-
line drift the sub-segment has. A sub-segment with the
length of one second can thus be regarded as a stationary
3.2.3. QRS Location: State Tr a ns ition Rec ognition
In view of the QRS morphology properties in D(t) series,
the complexes are categorized into two groups: positive
and negative. Therefore, the different states are defined
to outline the phases of QRS complex in D(t). S2~S9
represent the positive states of QRS complex (see in
Figure 3), corresponding S20-S29 represent negative
states. An adaptive and self-corrected procedure, named
STR (State Transition Recognition), is developed to
automatically track the changes of signal series, to cor-
rect error detection and to record detected complexes.
The states transitions are based on three basic reference
lines: the baseline, the positive threshold and the nega-
tive threshold.
3.2.4 Fe a ture Extraction: Geomet ric Analys is Method
QRS complex has the triangular-alike or triangular-
component morphological characteristics, see in Figure 4.
This paper thus employs the geometric analysis method
(GAM) to extract the features of QRS complexes. GAM
has simple operations and low resource consumption,
being able to predict an d estimate the key points of QRS
complexes under noisy situations, such as R wave peak,
end point of Q wave (Qt) and onset point of S wave (Si).
Therein, R wave peak is obtained from Tpeak1 or
Tpeak2 and it has mono-peak or poly-peaks. The meas-
urement and the detection phases of Qt and Si points are
illuminated as follows.
·Define two-level thresholds for left and right sides of
R wave (LH=1/4*Vpeak1, LL=3/4*Vpeak1, RH=
1/4*Vpeak2 and RL=3/4*Vpeak2).
Figure 3. Positive states of QRS complex in D(t).
Copyright © 2009 SciRes. WSN
H. Y. ZHOU ET AL. 279
Figure 4. Illumination of geometric analysis method.
·Calculate the intersection points between the thresh-
old values and complex signals. The slopes of two
approaching lines represent two characteristics of
QRS complex: SP (Positive Slope) and SN (Nega-
tive Slope).
·Obtain the duration leng th of QRS (LQRS) which is
the distance of two intersection points between the
baseline and two approaching lines.
3.3. Cardiac Arrhythmias Classification
Basing on the features values extracted from ECG sig-
nals, a self-diagnosis expert system is implemented to
classify heart rhythms and interpret cardiac arrhythmias.
The diagnostic rules of the expert system rely on the ex-
periential rules estimated from the self-learning of sys-
tem and the definitions of cardiologists. The diagnosis
system is composed of three phases: a pre-learning ma-
chine, a rhythm classifier and an arrhythmia interpreter,
see in Figure 5.
Based on the well-known experiential rules of cardiolo-
gists and the results of the training procedure, the pre -
learning machine builds and estimates the diagnostic rules
for every lead ECG signals of a patient. The rhythm cla ssi-
fier classifies each detected heart rhythm into one of two
catalogues: known rhythm or unknown rhythm. For the
known rhythms, they are still classified into two types ac-
cording to the v alues of the RR interva ls: sinu s rhythm and
ventricular rhythm; and for the unknown rhythms, we will
adopt classical met hods t o classi fy, the classification results
will be verified by the cardiologists. In terms of the known
rhythm types and the diagnostic rules, the cardiac ar-
rhythmias interpreter is used to explain cardiac arrhyth-
mias with the symptoms of relative heart diseases.
3.3.1. Automatic Learning Mac h ine
Ten seconds ECG signals are used to calculate the
rhythm template and to estimate the diagn ostic rules. Th e
Figure 5. Illumination of automatic diagnosis system.
initial cardiac status, rhythm type, statistical and mor-
phological features are achieved in this module. The di-
agnostic results will be further fed back to adjust the co-
efficients of diagnostic rules. Unlike resting ECG,
long-term ambulatory ECG has continual tiny changes
with the influences of exterior environments and the pa-
tient’s physical status. The tiny changes are generally
normal and the coefficients of diagnostic rules thus
should be self-updatable to meet the changes.
3.3.2. Rhythm Clas s i f i e r
By adopting the expert system and the confidence inter-
val method, the rhythm classifier can recognize two
kinds of QRS complex rhythms: sinus and ventricular.
The details of signal features (RR interval, QRS duration,
R wave left- & right-sides slopes, R wave amplitude, and
QRS absolute area), the rhythm type and the complex
peaks are used to describe a heart rhythm. Hence, they
can be used to recognize a rhythm and by comparing
with the features of the rhythm template.
The rhythm classifier is based on the features com-
parison and the interval estimation. Since we have ob-
tained the features of current rhythm and the features of
the standard rhythm (rhythm template) in pre-learning
machine, the rhythm classification is thus to estimate the
confidence intervals, the weighed factors and the devia-
tion coefficients of the features. The classification equa-
tion can be expressed as:
 N
Where the i
is the classification factor that is used to
determine the heart rhythm by estimating the confidence
Copyright © 2009 SciRes. WSN
280 H. Y. ZHOU ET AL.
interval that it falls; the i
is the weighed factor that
indicates the contribution of the feature i; the i
is co-
efficient of deviation that associates the variation of the
feature i; the N indicates the number of the features.
3.3.3. Arrhythmia Interpreter
In terms of the rhythm type and heat rate (HR), a heart
rhythm can be recognized and interpreted by the ar-
rhythmia interpreter basing on diagnostic rules. Firstly,
the arrhythmia interpreter classifies the rh ythms into two
catalogues: bradycardia and tachycardia by comparing
current HR with the mean HR of the rhythm template.
In arrhythmia interp reter, the heart rhythms are identi-
fied and classified into two basic categories: normal car-
diac rhythms and cardiac arrhythmias, and the cardiac
arrhythmias can be further divided as two classes: the
known cardiac arrhythmias and the unknown cardiac
arrhythmias. The known cardiac arrhythmias are nor-
mally the cardiac tachycardia events which are caused by
serious heart diseases, including PVC (Premature ven-
tricular complexes), VT (Ventricular tachycardia), VF
(Ventricular Fibrillation), SVT (Supraventricular Tachy-
cardia) and PAC (Premature Atrial Contraction), etc. The
known cardiac arrhythmias have distinctive QRS com-
plexes and rapid heart rates which make them be inter-
preted accurately. The unknown cardiac arrhythmias are
normally the bradycardia events which ar e cau s ed by les s
serious or benign heart diseases. The known cardiac ar-
rhythmias have regular heart rhythms but slow heart
rates, the identifications of which are not reliable when
depending only on the heart rate.
4. Performance Analysis
The algorithm has been assessed on two ECG databases:
MIT-BIH arrhythmia database [9] and CSD database
(Clinic STAR Database). The former contains 48 half-
hour excerpts of two-channel ambulatory ECG re-
cordings, and the latter is obtained from 30 subjects of
the Gabriel Montpied hospital (CHRU de Clermont-
Ferrand, France) by using a PCC system named STAR
[10]. The CSD signals are recorded in the same format
(WFDB) as MIT-BIH Database one.
4.1. STAR System
Currently, a real-time remote continuous cardiac ar-
rhythmia detecting and monitoring system, named STAR
(Système Télé-Assistance Réparti), has been developed
by the SMIR group of LIMOS laboratory of the Blaise
Pascal University and been applied on the CHU de
Gabriel Montpied hospital (Clermont-Ferrand, France).
The STAR system combines the technology advantages
of pervasive computing, AED algorithm and remote
telemedicine system. Figure 6 shows its system architec-
ture, which consists of local wireless ECG sensor (WES)
nodes and remote cardiac surveillance system.
The system description is: a WES device equipped by
the surveillance object, which integrates the AED algo-
rithm, can acquire and analyze the patient’s ECG signals
in real-time. When a cardiac abnormal event is detected,
an alarm message and (or) a segment of ECG signals will
send to the cardiologists via the available wired or wire-
less communication mediums. In the remote cardiac sur-
veillance system, the cardiologists can examine cardiac
abnormal events by employing AED algorithm and make
a respond with the shortest delays. This system aims to
provide a rapid detection and diagnosis method for the
high-risk population of cardiac arrhythmias to prevent
sudden death. It is also used to do long-term heart sur-
veillance for the population who has the history of heart
diseases, or to do periodic heart examination for the
health population.
4.2. QRS Detector Evaluation
Dotsinsky et al. [11] defined four performance parame-
ters to assess the algorithm efficiency (Se: sensitivity and
Sp: specificity): TP (true positive), FP (false positive),
FN (false negative) and shifted SH beats, shown as fol-
Se 
1 (3)
Comparing with the pe rformance results of other algo-
rithms listed in Table 1, the performance results of this
via PSTN
Figure 6. Architecture of STAR system.
At Hospital or
via Wireless Ne tworks
(GSM or Satellite)
Copyright © 2009 SciRes. WSN
H. Y. ZHOU ET AL. 281
Copyright © 2009 SciRes. WSN
Table 1. Performance evaluation of QRS detection algo-
Se(%) Sp(%)
Afonso et al [12] 99.59 99.56
Poli et al [13] 99.60 99.51
Dotsinsky et al [11] 99.04 99.62
Kaiser et al [14] 99.68 99.72
Datex-Ohmeda Corp. [16] 99.86 99.88
Alg 1 94.6 98.0
Millet et al [15] Alg 2 97.3 98.0
Our Algorithm 99.43
MIT 99.25
CSD 98.55
MIT 97.94
detection algorithm, 99.37% sensitivity and 99.68%
specificity on MIT-BIH d atabase, 99.67% sensitivity and
99.74% specificity on CSD database, show the high sen-
sitivity and specificity. This detection algorithm has
minimal beat detection latency, low computational con-
sumption and fast detection ability.
4.3. Rhythm Classifier Evaluation
The rhythm classifier classifies heart rhythms into two
catalogues: non-alarm and alarm-rhythms. The alarm-
rhythms defined in our algorithm are tachycardia, i.e.
PAC, PVC, SVT, VT, and VF. They represent serious
heart diseases which need to be reported immediately.
The non-alarm rhythms include the normal rhythms and
some benign or less serious cardiac arrhythmias, such as
The four parameters are used to assess the algorithm
performance [17]: A true positive (TP) is a serious car-
diac arrhythmia that has been correctly classified as an
alarm- rhythm; A false positive (FP) is an organized
normal rhythm that has been incorrectly classified as an
alarm- rhythm; A true negative (TN) is any normal or
less serious rhythm that has been correctly classified as a
non-alarm rhythm; A false negative (FN) is a serious
cardiac arrhythmia that has been incorrectly classified as
a non-alarm rhythm.
The sensitivity (Se) is the number of true positive ab-
normal rhythms, expressed as a percentage of the total
number of abnormal rhythms. Se is calculated by for-
mula (3). The specificity (Sp, also named positive pre-
dictive accuracy) is the number of organized rhythms
that have been correctly classified as normal rhythms,
expressed as a percentage of the total number of normal
rhythms and computed by formula (4).
Comparing with the performance of other algorithms
listed in Table 2, the performance results of this classifi-
cation algorithm, 90.90% sensitivity and 95.50% speci-
ficity on MIT-BIH database, 95.6% sensitivity and
99.5% specificity on CSD records, show its good per-
formance. Since the features extracted by the detection
algorithm are the time domain characteristics of QRS
complex, this classification algorithm thus can directly
utilize the experiences of cardiologists that reduces the
complexity of rules training and th en improves the accu-
racy of classification. Another advantage is that this al-
gorithm is able to identify various cardiac arrhythmias
comparing to most of other algorithms.
5. Conclusion
The objective of our research is to design a real-time
energy efficient Automatic ECG Diagnosis algorithm for
the PCC system. The PCC application is free of the limi-
tations of time and space, that is, this system supports
long-term monitoring (from few days to one month) and
the patient have the freedom of daily actions. Th e results
of the performance evaluation show that our algorithm
satisfies application demands.
Table 2. Performance evaluations of rhythm classification algorithm.
Se Sp Se Sp Se Sp Se Sp Se Sp Se Sp Se Sp
Horácek [18] 90.378.6
Ge et al [19] 93.2 94.4 96.4 96.7 94.896.8100 96.297.798.698.6 97.7 96.8396.73
Ham & Han [20] 99 97
Chen et al [21] 93 96
Minami et al [22] >98>98 >98>98
Chen [23] 95.24 96.00 97.78
Melo et al [24] 93 99
Datex-Ohmeda [16] 94.0897.55
Philips AED [17] 84 91 97*,76** 91
Our Algorithm
*:Ventricular Fibrillation (amplitude > 0.200mv). **: Fine Ventricular Fibrillation (0.100mv<amplitude < 0.200mv)
1Evaluation results on MIT-BIH database 2Evaluation results on CSD clinical records
282 H. Y. ZHOU ET AL.
Copyright © 2009 SciRes. WSN
In this AED algorithm, the QRS detector adopts linear
time-domain statistical analysis and syntactic analysis
methods to locate QRS complex from AECG signals.
The signal preprocessing and conditioning procedure,
adopting adaptive filter and band-pass filter, remove or
reduce various interferences caused by physical and
technical factors. The most serious noisy, such as motion
artifacts, has been effectively eliminated by the adaptive
filter. According to the statistical feature and morphol-
ogic features of QRS complex, i.e. heart rate, steep edges
and sharp amplitude, the QRS complex is located to
mark heart beat by applying SAT method and STR pro-
cedure on sub-segment diagnosis window.
The rhythm classifier classifies rhythms and interprets
cardiac arrhythmias basing upon the diagnostic rules
which are obtained from the experiences of cardiologists
and the training results of pre-learning phase. The initial
ECG signals with the length of 10 seconds are used to
estimate the type of QRS complex and to extract th e fea-
tures of normal rhythm template (the means of LQRS,
RR, etc.). According to the origination of heart beat, the
rhythms are categorized into two classes: sinus rhythm
(atria) and ventricular rhythm (ventricle). According to
the changes of heart rate, cardiac arrhythmias are catego-
rized into two classes: bradycardia and tachycardia. The
cardiac arrhythmias interpretation procedure is adopted
to classify cardiac arrhythmias into various types of bra-
dycardia and tachycardia, based on the features extracted
in the detection algorithm.
Currently, this algorithm has been applied on the
STAR system. The performance evaluations results show
that this algorithm was effective for the QRS detection
and the rhythm classification, and was thus suitable for
PCC services. The simple, fast and efficient features of
this algorithm enable it to be embedded into microproc-
essor system or be implemented on chip.
6. Acknowledgement
This project is supported by OSEO (French research
agency) and the Conseil Régional d’Auvergne (France).
7. References
[1] B.-U. Köhler, C. Hennig, and R. Orglmeister, “The prin-
ciples of software QRS detection,” IEEE Engineering in
Medicine and Biology Magazine, Vol. 21, No. 1, pp.
42–57, 2002.
[2] J. M. Jenkins and S. A. Casw ell, “Detection algorithm in
implantable cardioverter Defibrillators,” Proceedings of
the IEEE, Vol. 84, No. 3, pp. 428–445, 1996.
[3] S. A. Coast, R. M. S tem et al., “An approach to cardiac
arrhythmia analysis using hidden markov models,” IEEE
Transaction On Biomedical Engineering, Vol. 37, No. 9,
pp. 826–835, 1990.
[4] D. S. Benitez, P. A. Gaydecki, A. Zaidi et al., “A new
QRS detection algorithm based on the Hilbert transform,”
IEEE Computers in Cardiology, pp. 379–382, 2000.
[5] P. E. Trahanias, “An approach to QRS complex detection
using mathematical morphology,” IEEE Transaction On
Biomedical Engineering, Vol. 40, No.2, pp. 201–205.
[6] F. I. de Oliveira and P. U. Cortez, “A QRS detection
based on Hilbert transform and wavelet,” Proceedings of
14th IEEE SPSW on MLSP, pp. 481–489, 2004.
[7] S. M. Szilágyi, Z. Benyo, L. Szilagyi, et al., “Adaptive
wavelet transform based ECG waveforms detection,”
Proceedings of 25th Annual IEEE EMBS International
Conference, No. 24, pp. 12–15, 2003.
[8] W. A. H. Engelse and C. Zeelenberg, “A single scan al-
gorithm for QRS-detection and feature extraction, Com-
puters in Cardiology,” No. 6, pp. 37–42, 1979.
[9] G. B. Moody and R. G. Mark, “The MIT-BIH arrhythmia
database on CD-ROM and software for use with it,”
Computers in Cardiology, Vol. 17, pp. 185–188, 1990.
[10] H. Y. Zhou, K. M. Hou et al., “Remote continuous car-
diac arrhythmias detection and monitoring,” 2nd Interna-
tional Conference on E-health in Common Europe, Kra-
kow, pp. 11–12. March 2004.
[11] I. A Dotsinsky and T. V Stoyanov, “Ventricular beat
detection in single channel electrocardiograms,” Bio-
medical Engineering Online, Vol. 3, No. 3, 2004.
[12] V. X. Afonso, W. J. Tompkins, T. Q. Nguyen and S. Luo,
“ECG beat detection using filter banks,” IEEE Transac-
tion on Biomedical Engineering, Vol. 46: pp. 192–202.
[13] R. Poli, S. Cagnoni, and G. Valli, “Genetic design of
optimum linear and nonlinear QRS detectors,” IEEE
Transaction On Biomedical Engineering, Vol. 42, pp.
1137–1141, 1995.
[14] W. Kaiser and M. Findeis, “Novel signal processing
Methods for exercise ECG,” Special issue on Electrocar-
diography in Ischemic Heart Disease, Proceedings of
IJBEM, Vol. 2, 2000.
[15] J. Millet, M. Perez, G. Joseph, A. Mocholi, and J. Chorro,
“Previous identification of QRS onset and offset is not
essential for classifying QRS complex in a single lead,”
Computers in Cardiology, Vol. 24, pp. 299–302, 1997.
[16] Datex-Ohmeda Corp.: Bedside Arrhythmia Monitoring
Quick Guide, Education Report, Internal web journal for
medical professionals (Focus on Cardiovascular), URL:
http://www.clinicalwindow.com, November 2002.
[17] Philips Medical systems: AED Algorithm application
Note Philips HeartStart, Technical Document, URL:
http://www.medical.philips.com, Philips Electronics
North America Corporation, 2003.
[18] M. B Horácek, “Body-surface potential maps can identify
substrate for ventricular arrhythmias,” International
Journal of Bioelectromagnetism,” Vol. 5, No. 1, pp.
331–334, 2003.
[19] D. F. Ge, N. Srinivasan, and S. M. Krishnan, “Cardiac
H. Y. ZHOU ET AL. 283
arrhythmia classification using autoregressive modeling,”
Biomed Eng Online, Vol. 1, No. 5, 2002.
[20] F. M. Ham and S. Han, “Classification of cardiac ar-
rhythmias using fuzzy ARTMAP,” IEEE Trans Biomed
Eng; Vol. 43, No. 4, pp. 425–30, April 1996.
[21] S. W. Chen, P. M. Clarkson, and Q. Fan, “A robust se-
quential detection algorithm for cardiac arrhythmia clas-
sification,” IEEE Trans Biomed Eng, Vol. 43, No. 11,
1120–5, November 1996.
[22] K. C. Minami, H. Nakajima, and T. Toyoshima,
“Real-time discrimination of ventricular tachyarrhythmia
with fourier-transform neural network,” IEEE Trans Bio-
med Eng, Vol. 46, pp. 179–185, 1999.
[23] S. W. Chen, “Two-stage discrimination of cardiac ar-
rhythmias using a total least squares-based prony model-
ing algorithm,” IEEE Trans Biomed Eng, Vol. 47, pp.
1317–1326, 2000.
[24] S. L. Melo, L. P. Caloba, and J. Nadal, “Arrhythmia
analysis using artificial neural network and decimated
electrocardiographic data,” Comp Cardiol, Vol. 27, pp.
7376, 2000.
Copyright © 2009 SciRes. WSN