ARRHYTHMIA CLASSIFICATION USING
WAVEFORM ECG SIGNALS
Yakup KUTLU*, Gokhan ALTAN+, Novruz ALLAHVERDI#
*Iskenderun Technical University, Hatay, TURKEY
+Mustafa Kemal University, Hatay, TURKEY
#KTO Karatay University, Konya, TURKEY
Abstract— An electrocardiogram (ECG) is a non-linear and non-
stationary diagnostic biomedical signal that has a great
importance for cardiac disorders. The computer-assisted analysis
of biomedical signals has become an essential tool in recent years.
This study introduces a deep learning application in automatic
arrhythmia classification. The proposed model consists of a
multi-stage classification system in raw ECG using a Deep belief
network (DBN) which has a greedy layer wise training phase.
The multistage DBN model classified the MIT-BIH Arrhythmia
Database heartbeats into 5 main groups defined by ANSI/AAMI
standards. All ECGs are filtered with median filters to remove
the baseline wander. ECG waveforms were segmented from
long-term ECGs using a window with a length of 501 data points
(R peak of the wave is located at the centre of the window). The
proposed DBN-based multistage arrhythmia classification has
discriminated five types of heartbeats with a high accuracy rate
Keywords— Arrhythmia, Deep Belief Networks, Deep Learning,
AAMI, Raw ECG Waveform
According to the World Health Organization surveys ,
heart diseases are one of the most important reasons which
cause death. Heart disease symptoms depend on what type of
heart disease you have. An electrocardiogram (ECG) is a non-
linear and non-stationary diagnostic signal that is important
for cardiac disorders . It is hard to assess a cardiac disorder
using ECG because of long processes that need a control in
detail and infrequent arrhythmias. In order to overcome these
challenges, the computer-assisted analysis of biomedical
signals has become an essential method in recent years. The
computer assisted diagnosis and analysis systems achieve
rapid and advanced assessments in long, and hard to identify
processes. Arrhythmia and many cardiac disorders usually
need to use long-term ECG in inspection controls .
Therefore, computer-based methods and diagnosis systems
provide major simplicity and reliability in the diagnosis and
treatment of the diseases for cardiologists.
Arrhythmia is a problem concerning the abnormal rhythm
and rate of heartbeats. The heart can beat too fast, too slowly,
or inconsistently in different types of arrhythmias, which may
feel like antagonism affection or fluttering. Arrhythmia may
be classified by rate of heartbeats, mechanism (automaticity,
re-entry, triggered) or duration of the heartbeats . Several
types of arrhythmia are harmless, but some of them refer the
cardiac disorders that may cause death. The ECG is a popular
diagnosis tool which is of the primary importance for
There are many studies that are used for detecting
arrhythmias, classifying them and diagnosing cardiac diseases
that occur as a result of arrhythmias. These studies can be
incorporated into two basic feature extractions: fiducial and
non-fiducial methods. The fiducial methods contain the local
features such as temporal, morphological, amplitude, duration,
interval and segments between two selected waves which are
extracted from ECG waveforms. These methods are based on
the time-domain features on the ECG . The non-fiducial
methods are based on the frequency-domain features such as
wavelet transformations, and the other digital signal
processing techniques that extract new signal forms, sub-
bands and coefficients from ECG waveforms .
Deep learning (DL) is an effective and high-performance
machine learning algorithm which is gaining popularity.
Frequently used analyses of the DL are used in image
processing, speech and natural language processing processes.
Actually, DL is a neural network structure which addresses
the deeper feature levels using more hidden layers . In this
study, Deep Belief Networks (DBN), which is an adaptable
DL algorithm, is utilized to classify the heartbeats from
different classes of arrhythmia using ECG waveform as input
of the structure.
The remainder of the paper is structured in the following
manner. The database and the arrhythmia types are defined by
AAMI standards, pre-processing, and feature extraction from
arrhythmia heartbeats are described in detail. The proposed
multistage classification system is explained. The
experimental results that are obtained using the DBN classifier
International Conference on Advanced Technology & Sciences (ICAT’16) September 01-03, 2016 Konya-Turkey
II. MATERIALS AND METHODS
The general management of medical treatment and
assessment systems has become effective and convenient
processes because of the recent technological developments in
integrated circuit systems and computer-aided intelligent
monitoring and diagnosis systems. In this section,
information about ECG waveforms and the DBN classifier are
described in detail.
There are several arrhythmia databases in the literature. In
this study, the MIT-BIH arrhythmia database (MADB) is
utilized . This database has been used for evaluating
arrhythmia detection and classifying the arrhythmia types.
MADB contains 48 long-term ECGs from 25 men aged 32–89
years, and 22 women aged 23–89 years; each has 11-bit
resolution with 360 Hz sampling frequency. The heartbeats
are labelled as five main arrhythmia types defined by the
Association for the Advancement of Medical Instruments
(AAMI) standard. AAMI standardizations provide an
objective, understanding, and dividual assessments and
monitoring processes of the arrhythmia types for clinical
treatments and an increased capability of testing and training
abilities for supervised learning phases . AAMI classifies
heartbeats into normal beats (N), supraventricular ectopic
heartbeats (S), ventricular ectopic heartbeats (V), fusion
heartbeats (F), and unknown heartbeats (Q). The testing and
training dispersions of the heartbeats from the MADB are
seen in Table I.
DISPERSION OF MADB ACCORDING TO AAMI STANDARDS AND
QUANTITIES OF TEST AND TRAINING SETS
classes MIT-BIH heartbeat classes Train
Left bundle branch block beat
Right bundle branch block beat
Nodal escape beat
Atrial escape beat
Aberrated atrial premature beat
Premature or ectopic
Atrial premature contraction beat
Nodal premature beat
Ventricular flutter wave beat
Ventricular escape beat
Premature ventricular contraction
F Fusion of ventricular and normal
beat 350 350
Fusion of paced and normal beat
Long-term ECG signals can be contaminated by several
types of noise, such as motion during ECG recording,
electromyogram noise, contact noise, clinician artefacts,
coughing, position of patient, baseline wandering, etc. All
ECGs are filtered with two median filters to remove the
baseline wander . 6077 short-term ECGs were segmented
from long-term ECGs using a window with a length of 501
data points (R peak of the wave is located at the center of
window). All data points are normalized to a [0, 1] range.
B. ECG Waveform
The ECG is a method that finds out the regularity or
irregularity of heart beats and heart rates using the electrical
activity of the heart. The recorded electrical activities of the
heart represent for a waveform on the clinical assessments.
These waveforms may have different forms according to the
lead of the ECG , . The use of the ECG in medical
assessment processes is very important in detecting the
different waveforms and various cardiovascular heart diseases.
In the entire body only the heart muscle has the ability to
contract spontaneously. Polarity is the event of discharge of
electrical charge of heart tissue. Depolarization is the positive
charging case of electrical activation in heart tissue .
The ECG has 0 mV to 5 mV amplitude and a frequency
band between 0.5 Hz and 100 Hz , . P, Q, R, S, T and
U waves appear over the baseline in the signal, respectively. If
the amplitude of Q, R and S waves is less than 5 mV, the
wave is referred to using small capitals (q, r, and s). The
remaining portion between the waves is a segment; the
distance between the waves is an interval .
Fig. 1 ECG waveform with P, Q, R, S, and T waves
A P wave occurs as the result of the depolarization of the
atrium. First, the right atrium then the left atrium depolarizes.
Therefore, the first part of the P wave occurs when the right
atrium depolarizes; second part of the P wave occurs when the
left atrium depolarizes. Although it depends on the time of the
year, the duration of the P wave is about 0.11 seconds; the
amplitude of the P wave is between 0.18 mV and 0.22 mV on
a normal lead .
The QRS complex occurs as the result of the depolarization
of the ventricles. One of the waves (R) forming the QRS
complex is positive; the other two waves (Q and S) are
negative. The Q wave represents the first negative wave after
the P wave; the R wave represents the first positive wave after
the P wave and the S wave represents the next negative wave
after the R wave. The QRS monitored complex varied in
different leads. The QRS samples show significant differences
even among normal individuals. R and S waves refer to the
contraction of the myocardium. The QRS complex indicates
the current causing the left and right ventricle contraction .
The QRS complex has the maximum amplitude between the
ECG forms. The duration of the QRS complex does not
International Conference on Advanced Technology & Sciences (ICAT’16) September 01-03, 2016 Konya-Turkey
exceed 0.11 seconds and has an amplitude value up to 2-3 mV
The T-wave occurs as a result of ventricular re-polarization.
The T-wave may have a pointed or flat view and positive,
negative or biphasic value on various leads. The duration of
the T wave that belongs to a normal subject is between 0.10
and 0.25 seconds. It takes place after about 300ms from the
QRS complex. The positions of these waves vary according to
the heart rhythm. The T wave is closer to the QRS complex
when the heart rhythm accelerates .
C. Deep Belief Networks
This study introduces a deep learning (DL) application for
automatic arrhythmia classification. The proposed model
consists of a multi-stage classification system of raw ECG
using DL algorithms. The DBN is one of the most effective
DL algorithms which has a greedy layer wise training phase
. The DBN is composed of both Restricted Boltzmann
Machines (RBM) or an autoencoder based layer-by-layer
unsupervised pre-training procedure and neural network based
supervised training , . Considering RBM with input
layer activations (for visible units) and hidden layer
activations (hidden units), bias of the visible unit , bias of
hidden unit :
(,) represents the joint distribution of the RBM and
(,) represents the energy function of the distribution.
RBM is used for calculating the conditional distribution of the
visible and hidden units. Each adjacent two layers create an
RBM. The first visible unit is the input feature vector and the
other RBM parameters = (,,) are denoted by
depending on the first visible unit .
In the unsupervised training phase, the sub-network's
hidden layer serves as the visible layer for the next adjacent
layer applying contrastive divergence and the probabilistically
reconstruction of the shared weights is implemented . In the
supervised training phase of the DBN, the calculated shared
weights and the structure of the DBN are unfolded to a neural
network structure for fine-tuning all the parameters of the
deep structure such as the weights and the biases . The
DBN consists of at least two hidden layers (latent variables) in
the neural network. The number of the hidden layers is related
to the deep analysis of the input features in detail , .
III. EXPERIMENTAL RESULTS
The morphological features are the ones most used in
clinical trials for the diagnosis of the arrhythmia types. The
robust and steady detection of arrhythmia is a common need
for all the cardiac diseases. Each arrhythmia type can be
related to different types of cardiac and pulmoner diseases.
That’s why detection and classification of the arrhythmia
types are so important in the early diagnosis and early
treatment processes. Considering the importance of the
classification of the arrhythmia types, a computer-aided
classification of the 5 arrhythmia types is implemented using a
DBN-based multistage classification. Figure 2 depicts the
structure of the arrhythmia classification model.
Fig. 2 Structure of proposed Arrhythmia Classification Method
Two median filters are applied to remove the noise and the
baseline wanders to raw ECGs. Analysis of the long-term
ECGs is a demanding process for clinicians and also for
computer-aided systems. Considering this situation, ECG
waveforms were extracted from long-term ECGs using the
moving window analysis technique. The R peak centred
window with 501 data points was moved to extract ECG
waveforms. 6077 of ECG waveforms were obtained from
long-term ECGs. ECG waveforms with 501 data points were
directly used as features. Having a great number of the feature
dimensionality causes long and deceiving training processes
for the supervised machine learning algorithms. Feature
dimensionality reduction for the provides for the extraction of
more meaningful classification rules, the elimination of the
pointless feature vector for machine learning algorithms, the
improvement of generalization capabilities using fewer
parameters and reduced complexity and run-time and for the
evaluation and prediction of accuracy for classifiers . The
sequential forward feature selection algorithm is utilized in the
proposed method to reduce feature dimensionality . The
algorithm selects a subset of features which are not yet
selected from 501 data points and the best predict the
arrhythmia types by sequentially selecting features until there
is no improvement in the prediction. The highest accuracy is
achieved using 106 features from the ECG waveforms. The
reduced feature vector is normalized to 0-1. The proposed
DBN-based multistage classifier was trained using 106 data
points. Selected data points on the ECG waveform are seen in
Figure 3 with the red asterisk.
International Conference on Advanced Technology & Sciences (ICAT’16) September 01-03, 2016 Konya-Turkey
Fig. 3 Selected data points from ECG waveforms after feature
dimensionality reduction (Red asterisk)
The proposed multistage DBN model separates N, S, V, F,
and Q types of arrhythmias, respectively. 4 of the DBN
models are used in the proposed system. The RBM based
greedy layer-wise pre-training is used in this model at the
unsupervised learning stages of all DBNs with 5 epochs. The
parameters of the RBM were denoted by iterations. The
models were tested with a limited number of the parameters
and the highest classification performances are given. The
learning rate of the model is 3 and the softmax output function
was utilized constantly. The proposed multistage arrhythmia
classification model consists of 4 DBN structures with various
numbers of hidden units. The DBN1 has 2 hidden layers with
100-260 hidden units; the DBN2 has 3 hidden layers with
230-520-210 hidden units; the DBN3 has 2 hidden layers with
120-240 hidden units; and the DBN4 has 2 hidden layers with
70-190 hidden units. The four DBN structures are connected
sequentially and have the ability to separate five classes of
arrhythmia types defined by ANSI/AAMI.
The training set of the DBN-based automatic arrhythmia
classification model includes 4,077 of ECG waveforms from
various types of heartbeat classes distributed homogeneously.
The DBN-based multistage model is tested using 2,000 of
ECG waveforms. The confusion matrix of the classifier is
seen in Table II.
CONFUSION MATRIX OF MULTISTAGE CLASSIFIER
Labels Predicted heartbeats
N S V F Q
N 489 2 8 9 3
S 0 290 7 3 8
V 2 0 472 2 5
F 4 5 9 179 13
Q 5 3 4 7 471
Zhang et al. used inter-beat features, amplitude morphology
and morphological distance features for separating 4 types of
arrhythmia by the Support Vector Machines (SVM)
classification algorithm with 86.66%, 93.81%, and 98.98% for
accuracy, sensitivity, and selectivity, respectively . Melin
et al. utilized cycle features and fiducial features with an
Artificial Neural Network (ANN) and Learning Vector
Quantization based multistage classification algorithm and
classified 15 types of arrhythmia with an accuracy rate of
99.16% . Thomas et al. extracted wavelet based
coefficients from 4th and 5th scale of Wavelet transform, high
order statistics and fiducial features using the QRS complex
from the ECG with ANN and presented an accuracy rate of
94.64% and a sensitivity rate of 94.60% for 5 classes of
arrhythmia types . Batra et al. utilized invariant features
and Principle Component Analysis features using the SVM
classifier with the cross validation technique and achieved an
accuracy rate of 84.82% for 11 classes of arrhythmia types
. Leutheuser et al. compared the Naive Bayes and k-NN
classifier algorithms using statistical and high order statistical
features, heartbeat features and template based features from
segmented ECGs for the real-time classification of 2 types of
arrhythmias on android-based mobile devices with reported
accuracies of 93.30% and 56.10% for k-NN and Naive Bayes
classifiers, respectively . Alajlan et al. extracted
morphological features, high order statistical features and non-
fiducial features applying the Discrete wavelet transform, S
transform and classified arrhythmia types into 2 classes using
the SVM machine learning algorithm with high performances
of 93.49%, and 93.14% for accuracy, and sensitivity,
DL algorithms, especially the DBN, are being effectively
used in ECG analysis. The DBN is utilized at both feature
extraction ,  and classification stages , .
Huanhuan et al. used the DBN-based learning features from
complete waveforms and R-R timing interval features with a
multi-stage (5 stages) SVM classifier model. They achieved
an accuracy rate of 98.82% for 6 classes of arrhythmia types
. Rahhal et al. extracted temporal features, morphological
features and DBN-based features using stacked denoising
autoencoders. They fed the all features to the SVM for
training and classification. They classified 2 classes of
arrhythmia types defined by ANSI/AAMI with an overall
accuracy rate of 98.49% . Yan et al. utilized R-R interval
features, beat features and raw ECG signals from multi-lead to
feed the DBN classifier and achieved a high accuracy rate of
98.82% for 12 classes of arrhythmia types . There are lots
of studies based on different types of arrhythmia classification.
We focused on the arrhythmia types defined by ANSI/AAMI
and the considerable studies are compared in Table 3.
COMPARISON OF THE RELATED WORKS FOCUSED ON
ARHYTHMIA DETECTION DEFINED BY AAMI
Features Classifier Accuracy
Owis et al.
Martis et al.
DWT, LDA, PCA PNN 99.28%
Kim et al.
feature, DWT, PCA, LDA
Pulse based features LD 93.60%
Ye et al.  Interval Features, Wavelet
Transform, ICA, PCA
Proposed ECG Waveform DBN 95.05%
CWT: Continuous Wavelet Transform, LD: Linear Discriminant, ELM: Extreme Le arning
Machines, PNN: Probabilistic Neural Network, DWT: Discrete Wavelet Transform
It is hard to compare the studies in a stable way, because of
reasons such as the different number of subjects, different
number of the arrhythmia types, different subjects, different
databases and different classification types. High
classification performances are reported in the literature. In
this study, the efficiency of the DL algorithms has been
proven with high classification performances of 95.05%,
93.87%, and 94.51% for accuracy, sensitivity, and selectivity,
PQRS complexes and T waves plots a regular form in
normal sinus rhythm. Any obvious changes occurring in the
PQRST lines indicate the irregularity or arrhythmia in
heartbeats. Since the determination of the features such as
intervals, segment measurements, heart rate, and the
frequency of R waves have great benefits for clinicians to
identify the cardiac diseases and arrhythmias, the physiology
and the morphology of the ECG waveforms have frequently
been used in clinical trials , . The meaningful data
points for arrhythmia classification are thickened between S-T
waves and P-Q waves for the proposed DBN-based multistage
The proposed DBN-based multistage arrhythmia
classification has discriminated five types of heartbeats with a
high accuracy rate of 95.05%. The achievements prove the
success and efficiency of the DBN algorithm in raw ECG
 WHO, “The top 10 causes of death.” [Online]. Available:
http://www.who.int/mediacentre/factsheets/fs310/en/. [Accessed: 07-
 J. G. Webster, Medical Instrumentation, Application and Design, 4th ed.
Boston: Houghtoon Mifflin Company, 1978.
 Y. C. Yeh, C. W. Chiou, and H. J. Lin, “Analyzing ECG for cardiac
arrhythmia using cluster analysis,” Expert Syst. Appl., vol. 39, no. 1, pp.
 P. Hamilton, “Open source ECG analysis,” Comput. Cardiol., vol. 29,
pp. 101–104, 2002.
 D. Ge, N. Srinivasan, and S. M. Krishnan, “Cardiac arrhythmia
classification using autoregressive modeling.,” Biomed. Eng. Online, vol.
1, p. 5, 2002.
 S. A. Israel, J. M. Irvine, A. Cheng, M. D. Wiederhold, and B. K.
Wiederhold, “ECG to identify individuals,” Pattern Recognit., vol. 38,
no. 1, pp. 133–142, 2005.
 K. N. Plataniotis, D. Hatzinakos, and J. K. M. Lee, “ECG Biometric
Recognition Without Fiducial Detection,” in 2006 Biometrics
Symposium: Special Session on Research at the Biometric Consortium
Conference, 2006, pp. 1–6.
 Y. Bengio and O. Delalleau, “Justifying and generalizing contrastive
divergence,” Neural Comput., vol. 21, no. 6, pp. 1601–1621, 2009.
 G. B. Moody and R. G. Mark, “The impact of the MIT-BIH arrhythmia
database.,” IEEE Eng. Med. Biol. Mag., vol. 20, no. 3, pp. 45–50.
 Y. Kutlu and D. Kuntalp, “A multi-stage automatic arrhythmia
recognition and classification system,” Comput. Biol. Med., vol. 41, no.
1, pp. 37–45, 2011.
 U. Rajendra Acharya, J. S. Suri, J. A. E. Spaan, and S. M. Krishnan,
Advances in cardiac signal processing. 2007.
 M. Gabriel Khan, Rapid ECG Interpretation(Contemporary Cardiology),
3rd editio. Humana Press, 2007.
 Dr Patrick Davey, “ECG (electrocardiogram),” NetDoctor, pp. 1–4,
 S. Kara, “Sensing of ECG signals and Imaging at the Computer in Real
Time,” Erciyes University, 1991.
 Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle, “Greedy Layer-
Wise Training of Deep Networks,” Adv. Neural Inf. Process. Syst., vol.
19, no. 1, p. 153, 2007.
 Y. Yan, X. Qin, Y. Wu, N. Zhang, J. Fan, and L. Wang, “A restricted
Boltzmann machine based two-lead electrocardiography classification,”
2015 IEEE 12th International Conference on Wearable and Implantable
Body Sensor Networks (BSN). pp. 1–9, 2015.
 G. E. Hinton, S. Osindero, and Y.-W. Teh, “A fast learning algorithm
for deep belief nets.,” Neural Comput., vol. 18, no. 7, pp. 1527–54, 2006.
 H. S. Lee, Q. lan Cheng, and N. V. Thakor, “ECG waveform analysis by
significant point extraction. I. Data reduction,” Comput. Biomed. Res.,
vol. 20, no. 5, pp. 410–427, 1987.
 D. Ververidis and C. Kotropoulos, “Sequential forward feature selection
with low computational cost,” in Signal Processing Conference 2005
13th European, 2005, vol. 13, pp. 1–4.
 Z. Zhang, J. Dong, X. Luo, K. S. Choi, and X. Wu, “Heartbeat
classification using disease-specific feature selection,” Comput. Biol.
Med., vol. 46, no. 1, pp. 79–89, 2014.
 P. Melin, J. Amezcua, F. Valdez, and O. Castillo, “A new neural
network model based on the LVQ algorithm for multi-class
classification of arrhythmias,” Inf. Sci. (Ny)., vol. 279, pp. 483–497, Sep.
 M. Thomas, M. K. Das, and S. Ari, “Automatic ECG arrhythmia
classification using dual tree complex wavelet based features,” AEU -
Int. J. Electron. Commun., vol. 69, no. 4, pp. 715–721, 2015.
 A. Batra and V. Jawa, “Classification of Arrhythmia Using Conjunction
of Machine Learning Algorithms and ECG Diagnostic Criteria,” Int. J.
Biol. Biomed., vol. 1, pp. 1–7, 2016.
 H. Leutheuser, S. Gradl, P. Kugler, L. Anneken, M. Arnold, S.
Achenbach, and B. M. Eskofier, “Comparison of real-time classification
systems for arrhythmia detection on Android-based mobile devices,”
IEEE Eng. Med. Biol. Soc. Annu. Conf., vol. 2014, pp. 2690–2693, 2014.
 N. Alajlan, Y. Bazi, F. Melgani, S. Malek, and M. A. Bencherif,
“Detection of premature ventricular contraction arrhythmias in
electrocardiogram signals with kernel methods,” Signal, Image Video
Process., vol. 8, no. 5, pp. 931–942, Jul. 2014.
 M. M. Al Rahhal, Y. Bazi, H. AlHichri, N. Alajlan, F. Melgani, and R.
R. Yager, “Deep Learning Approach for Active Classification of
Electrocardiogram Signals,” Inf. Sci. (Ny)., vol. 345, pp. 340–354, Feb.
 M. Huanhuan and Z. Yue, “Classification of Electrocardiogram Signals
with Deep Belief Networks,” Computational Science and Engineering
(CSE), 2014 IEEE 17th International Conference on. pp. 7–12, 2014.
 N. Allahverdi, G. Altan, and Y. Kutlu, “Diagnosis of Coronary Artery
Disease Using Deep Belief Networks,” 2. Int. Conf. Eng. Nat. Sci., The
Book of Abstracts p.9, 2016.
 M. I. Owis, A. H. Abou-Zied, a. B. M. Youssef, and Y. M. Kadah,
“Study of features based on nonlinear dynamical modeling in ECG
arrhythmia detection and classification,” IEEE Trans. Biomed. Eng., vol.
49, no. 7, pp. 733–736, 2002.
 R. J. Martis, U. R. Acharya, C. M. Lim, and J. S. Suri, “Characterization
of ECG beats from cardiac arrhythmia using discrete cosine transform in
PCA framework,” Knowledge-Based Syst., vol. 45, pp. 76–82, 2013.
 J. Kim, S. D. Min, and M. Lee, “An arrhythmia classification algorithm
using a dedicated wavelet adapted to different subjects.,” Biomed. Eng.
Online, vol. 10, no. 1, p. 56, 2011.
 P. Tadejko and W. Rakowski, “Hybrid wavelet-mathematical
morphology feature extraction for heartbeat classification,” in
EUROCON 2007 - The International Conference on Computer as a
Tool, 2007, pp. 127–132.
 M. Llamedo and J. P. Martinez, “Heartbeat classification using feature
selection driven by database generalization criteria,” IEEE Trans.
Biomed. Eng., vol. 58, no. 3 PART 1, pp. 616–625, 2011.
 A. S. Alvarado, C. Lakshminarayan, and J. C. Príncipe, “Time-based
compression and classification of heartbeats,” IEEE Trans. Biomed.
Eng., vol. 59, no. 6, pp. 1641–1648, 2012.
 C. Ye, B. V. K. Vijaya Kumar, and M. T. Coimbra, “Heartbeat
classification using morphological and dynamic features of ECG
signals,” IEEE Trans. Biomed. Eng., vol. 59, no. 10, pp. 2930–2941,